|
|
|
|
|
|
|
|
|
|
|
|
Oral Presentation Progrm |
|
|
|
|
|
|
|
|
BSPC Session 01 |
|
|
|
Time: 14:00~15:30, Saturday, July 18, 2026 |
|
|
|
Session Chair: Wenxing Yang, University of Shanghai for Science and Technology |
|
|
|
|
|
|
|
|
Oral Session: |
BSPC Session 01 (Paper No.6418) |
|
|
|
Title: |
A Behavior-Oriented Feature Time-Series Processing Framework for Long-Term
Anomalous Spectrum Behavior Detection |
|
|
|
Author: |
Qin Li1, Xiang Wang1, Tao Zhi Huang1, Fa Yi Zhang1 and Ya Shu Cao1 |
|
|
|
Abstract: |
Long-term spectrum monitoring is an important foundation for electromagnetic
domain security, spectrum situational awareness, and stable wireless system
operation. One key task is to detect anomalous spectrum behaviors in continuous
observations. Existing spectrum anomaly detection studies mainly focus on
signal-level or image-level representations. They usually treat anomalies as
abnormal components in local spectrum segments and provide limited support for
modeling the persistence and temporal evolution of anomalous spectrum behaviors
under long-term non-stationary monitoring conditions. Meanwhile, continuous
monitoring produces massive complex baseband I/Q data. End-to-end modeling based
on raw I/Q data or high-dimensional spectrum images leads to high storage and
computational costs. To address these issues, this paper proposes a
behavior-oriented feature time-series processing framework for long-term
anomalous spectrum behavior detection. The framework converts continuous I/Q
monitoring segments into a multivariate feature time series. It extracts
time-domain statistical features, frequency-domain structural features, and
cross-time spectral variation features to characterize spectrum states and their
behavioral evolution during long-term monitoring. The constructed feature time
series compresses the data volume while preserving behavior-level information.
It provides a scalable and interpretable representation for downstream anomaly
detection models. Experiments on a long-term wideband spectrum monitoring
dataset show that the proposed framework improves point-level detection
performance and event-level alarm performance compared with power spectral
density image and time-frequency image representations. |
|
|
|
|
|
|
|
|
Oral Session: |
BSPC Session 01 (Paper No.6356) |
|
|
|
Title: |
Few-Shot Open-Set Audio Classification via Transductive Prototype Refinement and
Class Logit Enhancement |
|
|
|
Author: |
Tianyan Deng1, Yanxiong Li2, Rui Gao2 and Jiahao Du2 |
|
|
|
Abstract: |
Few-shot Open-set audio classification requires classifying query samples from
known classes with a few labeled support samples while rejecting query samples
from unknown classes. Transductive inference jointly observes the full unlabeled
query set to improve prototype estimation, yet standard transductive updates do
not distinguish known from unknown query samples, leaving prototypes vulnerable
to open-set contamination. Drawing on latent-inlierness weighting and decoupled
scoring for unknown-class samples, we propose a two-phase transductive method
operating over a frozen audio encoder. First, each query sample is assigned a
latent inlierness score that down-weights likely unknown-class samples, so that
prototype refinement is driven primarily by known-class evidence. The refined
prototypes are then directly optimized on a transductive loss combining support
cross-entropy, inlierness-weighted conditional entropy minimization, and
inlierness-weighted marginal entropy maximization, while open-set rejection uses
a prior-adaptive free-energy score that adjusts its threshold with the prior
proportion of unknown-class samples, decoupling detection from classification.
Experiments on three audio datasets show our method achieves state-of-the-art
results for few-shot open-set audio classification under multiple experimental
conditions. |
|
|
|
|
|
|
|
|
Oral Session: |
BSPC Session 01 (Paper No.6349) |
|
|
|
Title: |
Link-Aware GNN-Based Critical Node Identification for Underwater Acoustic Sensor
Networks |
|
|
|
Author: |
Ziyi Ding1, Yihao Zhao1, Zheyang Chen1, Shenao Tu2, Yougan Chen2 and Xiaomei Xu1 |
|
|
|
Abstract: |
Underwater acoustic sensor networks (UASNs) are indispensable for marine
exploration, yet the failure of critical nodes can easily precipitate network
fragmentation and routing voids. Existing node identification methods rely
primarily on static topological connections, neglecting the dynamic influence of
acoustic links. To address this, we propose a novel critical node identification
algorithm based on a Link-Aware Graph Neural Network (LAGNN), which explicitly
embedding node attributes and link stability metrics into the graph structure.
By designing a joint loss function driven by network Quality of Service (QoS)
degradation and leveraging a message-passing mechanism, LAGNN deeply extracts
the complex correlations between local link condition and global network
topology. Simulation results verify that LAGNN outperforms baselines with a 26%
higher Top-k hit rate. Removing its identified nodes inflicts catastrophic QoS
degradation, plunging the PDR to merely 20% of the benchmark level and spiking
the delay by 30%, which proves its precision in isolating critical bottlenecks. |
|
|
|
|
|
|
|
|
BSPC Session 02 |
|
|
|
Time: 16:00~17:30, Saturday, July 18, 2026 |
|
|
|
Session Chair: Wenxing Yang, University of Shanghai for Science and Technology |
|
|
|
|
|
|
|
|
Oral Session: |
BSPC Session 02 (Paper No.6325) |
|
|
|
Title: |
LiteMeter: Attention-Enhanced Keypoint Detection and Edge Inference for
Industrial Gauge Reading |
|
|
|
Author: |
Huasong Li1, Dongliang Fu2, Jiongmin Yu2, Jiakai Li1, Minghui Ouyang1 and Wei
Gao1 |
|
|
|
Abstract: |
Automatic reading of industrial analog gauges remains challenging in terms of
both detection accuracy and deployment flexibility. This paper presents
LiteMeter, a unified system that integrates an enhanced keypoint detector, a
cross-platform frontend, and an edge inference service. The proposed detector
incorporates the convolutional block attention module (CBAM) into the shallow
stages of YOLOv11 and introduces a P2 high-resolution prediction head to improve
fine-grained feature representation. Gauge values are estimated via a
four-keypoint polar-coordinate mapping strategy. Extensive experiments conducted
on a Hard Test benchmark with 11 types of industrial degradation demonstrate
that the combined use of P2 and CBAM constrains the performance drop to 37.58
pp, limits the generalization gap to 0.22 pp, and achieves a degraded precision
of 85.02%, along with a Grad-CAM high-response ratio of 0.1921. For deployment,
a frontend based on Tauri 2.0 and ONNX Runtime supports both desktop and mobile
platforms, enabling seamless switching between local and remote inference with
automatic fallback. Furthermore, a C++17-based inference server deployed on
Jetson Orin Nano achieves a per-frame latency of 58.6 ms using TensorRT FP16
(85.1 ms end-to-end, corresponding to 11.7 FPS). An EMA-adaptive RTSP streaming
pipeline is further designed to support long-term, unattended operation. |
|
|
|
|
|
|
|
|
Oral Session: |
BSPC Session 02 (Paper No.6317) |
|
|
|
Title: |
Lightweight Spatiotemporal Attention-Based Driver Unsafe Behavior Recognition |
|
|
|
Author: |
Mengqi Liu1, Yang Lai1, Xinzi Wang2, Zhiyuan Xue1, Ben Yang1 and Xuetao Zhang1 |
|
|
|
Abstract: |
Unsafe driver behaviors remain a major cause of traffic accidents. However,
existing methods are often limited by insufficient data diversity and inadequate
temporal modeling. To address these issues, we propose a lightweight
spatiotemporal modeling framework with dynamic attention (LSDA) for driver
behavior recognition. In addition, a driver behavior dataset is constructed by
simulating realistic bus and truck driving scenarios, enabling the
representation of typical operational conditions while maintaining
controllability. Specifically, a lightweight architecture with dense temporal
sampling is designed to effectively capture short-term motion dynamics.
Furthermore, a prototype-driven dynamic self-attention mechanism is introduced
to adaptively emphasize discriminative temporal features, thereby improving the
separability of similar behavior classes. In addition, Mixup and CutMix-based
feature augmentation are employed to enhance data diversity and generalization.
Extensive experiments on the constructed dataset and the SynDD1 dataset
demonstrate that LSDA consistently outperforms state-of-the-art methods,
particularly in fine-grained short-term behavior recognition, validating its
effectiveness and efficiency. |
|
|
|
|
|
|
|
|
Oral Session: |
BSPC Session 02 (Paper No.6472) |
|
|
|
Title: |
Deep Learning-Driven DOA Estimation for Distributed MIMO Radar in UAV-Assisted
Sensing Systems |
|
|
|
Author: |
Chuang Han1, Ke Ding1, Yan dong Sun1, Yue xian Wang1 and Cheng yan He1 |
|
|
|
Abstract: |
Unmanned aerial vehicle (UAV) auxiliary sensing system has been widely used in
the fields of national security, emergency response and intelligent
transportation. Direction-of-arrival (DOA) estimation is necessary for target
localization and trend sensing. However, traditional algorithms such as MUSIC
and SBL often suffer from performance degradation under low SNR conditions. The
spatial-temporal data transmitted by distributed nodes are difficult to fuse,
and the computational complexity is very high. Aiming at these problems, this
paper proposes a deep learning-driven DOA estimation algorithm for UAV-assisted
distributed multiple-input multiple-output (MIMO) radar, and establishes a
modular deep learning model. An end-to-end multi-branch fusion deep neural
network is designed. Each branch processes each subarray data independently, and
then realizes the collaborative processing of multi-node information through
feature fusion, so as to achieve efficient data fusion and high-precision angle
estimation. Its performance is verified by a large number of simulations. The
simulation results show that compared with the traditional algorithm, the
proposed algorithm achieves excellent results in RMSE, accuracy and real-time
performance under low SNR. |
|
|
|
|
|
|
|
|
COMM Session 01 |
|
|
|
Time: 14:00~15:30, Saturday, July 18, 2026 |
|
|
|
|
Session Chair: Long Shi, Southwestern University of Finance and Economics |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 01 (Paper No.6352) |
|
|
|
Title: |
Q-Learning-Based Routing and Caching Optimization for Underwater
Information-Centric Networks |
|
|
|
Author: |
Zheyang Chen1, Yihao Zhao1, Shenao Tu2, Ziyi Ding1, Yougan Chen2 and Xu Xiaomei2 |
|
|
|
Abstract: |
Underwater acoustic networks encounter substantial challenges in long-distance
horizontal transmission, including high propagation delay and excessive energy
consumption, which hinder real-time and energy-efficient operation in the
underwater Internet of Things. To overcome the broadcast storms, excessive
caching overhead, and lack of globally optimal routing in existing underwater
information-centric networking schemes, this paper proposes QL-ICN, a
Q-learning-based information-centric networking framework that integrates hybrid
electromagnetic communication among surface sink nodes with acoustic links to
underwater nodes. The framework models request forwarding as a Markov decision
process and employs Q-learning to select optimal unicast multi-hop paths,
thereby eliminating broadcast overhead. In addition, a trust score based on the
count of node selection as delivery targets is introduced as a priority metric
to guide intelligent cache node selection during the data return phase.
Simulation results demonstrate that as the number of consumers increases, the
network latency decreases significantly, while the total energy consumption
exhibits only sublinear growth. In both ablation studies and comparisons with
baseline algorithms, QL-ICN exhibits superior performance: it maintains the
lowest energy consumption while achieving an average latency that is only 75.85%
of the benchmark algorithm RAOH's. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 01 (Paper No.6367) |
|
|
|
Title: |
DDPG-Based Energy-Efficient Resource Scheduling in Underwater Acoustic
Communications |
|
|
|
Author: |
Tong Li1, xuefeng zhong1, Yuting Xiao1, Hao Zhao2 and Zilong Jiang3 |
|
|
|
Abstract: |
Abstract—In underwater acoustic communications (UAC),
achieving energy-efficient transmission is critical due to the harsh
marine environment and limited battery capacity of underwater
nodes. To overcome this limitation, this paper proposes an
Adaptive Continuous-Action Deep Deterministic Policy Gradient
(AC-DDPG) algorithm. This approach enables decision-making
directly within a continuous action space, achieving fine-grained
joint optimization of transmission frequency and power without
discretization. Different from the DQN-based reinforcement
learning algorithms, AC-DDPG adopts an Actor–Critic architecture
and incorporates an adaptive Ornstein–Uhlenbeck (OU)
noise mechanism to effectively balance exploration and exploitation
during training. Experimental results demonstrate that ACDDPG
achieves energy efficiency close to the optimal solution in
both small and large action spaces. Its performance significantly
outperforms current reinforcement learning algorithms such as
Q-learning, DQN, and TAS-DQN, exhibiting superior optimization
capability and stability, particularly in large-scale continuous
search spaces. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 01 (Paper No.6382) |
|
|
|
Title: |
A Traffic-Aware MAC Protocol for Underwater Acoustic Sensor Networks Based on
DSATUR Graph Coloring |
|
|
|
Author: |
Ruisi Gou1, Xiaohong Shen1, Weiliang Xie1 and Haiyan Wang1 |
|
|
|
Abstract: |
Underwater acoustic sensor networks (UASNs) are characterized by long
propagation delays, limited bandwidth, and strong spatiotemporal variability,
which make it difficult for medium access control (MAC) protocols to
simultaneously achieve collision-free access and efficient resource utilization.
Conventional TDMA suffers from limited spatial reuse, contention-based protocols
experience severe collisions under heavy traffic, and existing
graph-coloring-based MAC protocols adapt poorly to non-uniform traffic loads. To
address these issues, this paper proposes a DSATUR-based traffic-aware MAC
protocol, termed TAGC-MAC. First, a two-hop interference conflict graph is
constructed according to the network topology, and the DSATUR algorithm is
applied to obtain a baseline collision-free slot assignment, thereby reducing
the number of required slots and improving spatial reuse. Based on this
baseline, per-node slot demand is estimated from the packet generation rate of
each node, and additional slots are then iteratively allocated using a greedy
maximal independent-set strategy to alleviate queue buildup at high-load nodes.
Simulation results show that the proposed protocol achieves good performance in
throughput, fairness, and delay, and further improves system throughput under
non-uniform traffic loads, effectively enhancing channel access efficiency in
UASNs. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 01 (Paper No.6430) |
|
|
|
Title: |
A Medium Access Control Protocol for Underwater Acoustic Communication Networks
Based on Deep Reinforcement Learning |
|
|
|
Author: |
Jianmin Yang1, Zhihong Peng1, Yan Lin1, Ming Gong1, Huang Ying1, Hongli Wang2
and Pengyu Du3 |
|
|
|
Abstract: |
Abstract—Underwater Acoustic Communication Networks
(UACNs) play an irreplaceable role in applications such as
marine environmental monitoring, resource exploration, and
disaster warning systems. The performance of the Medium
Access Control (MAC) protocol directly affects network
communication quality, making the design of efficient MAC
protocols a major research focus in UACNs.
This paper proposes a Parameter-Sharing Independent QLearning
(PS-IQL) MAC protocol for UACNs based on a Deep
Q-Network (DQN), referred to as the DQN-MAC protocol. In a
distributed homogeneous network, an external agent is
introduced to conduct centralized training using the DQN
algorithm. During training, all nodes share the same network
parameters, while their learning experiences are stored in a
shared replay buffer. Mini-batch samples are then drawn from
the buffer to update the neural network through the loss
function. After convergence, the learned policy is deployed to
each node and learning is terminated, leaving only inference
during execution.
When a data packet enters the MAC queue, a node
performs inference based on the observed network state and
the learned policy to determine the transmission frame length.
When packet retransmission is triggered by backoff
mechanisms, the node further optimizes its backoff duration
through policy inference, thereby improving the overall
protocol performance. Experimental results demonstrate that
the DQN-MAC protocol achieves stable and effective
convergence during training and performs robustly under
different testing conditions. Moreover, it shows clear
advantages in Packet Delivery Ratio (PDR) and network
throughput. |
|
|
|
|
|
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 01 (Paper No.6434) |
|
|
|
Title: |
Q-Learning and Bayesian Optimization for Energy-Efficient and Self-Adaptive
Routing Protocol in Underwater Wireless Sensor Networks |
|
|
|
Author: |
Jianmin Yang1, Yue Ma1, Yonghui Zhong1, Haohua He1, Qing Liang1 and Jiajing
Chen1 |
|
|
|
Abstract: |
Abstract—Underwater wireless sensor networks (UWSNs)
suffer from limited node energy, unreliable acoustic channels,
and dynamic topologies, which together cause unbalanced
energy consumption and shortened network lifetime. In this
paper, we propose QBEAR, an energy-efficient clustering
routing protocol that combines Q-learning with Bayesian
optimization for UWSNs. In the inner layer, a Q-learning
mechanism drives cluster head (CH) election and inter-cluster
routing decisions: each node perceives multi-dimensional state
features including residual energy, distance to sink, and
neighbor density, and uses a reward function jointly
considering energy retention, topology quality, and link
reliability to update its Q-values. In the outer layer, a Bayesian
optimization framework automatically tunes the Q-learning
hyper-parameters and the cluster-control weight parameters,
allowing the system to autonomously approach globally
balanced performance without manual tuning. Simulation
results show that the proposed QBEAR protocol outperforms
LEACH, EECRAP, and QHUC in terms of packet delivery
ratio, energy-consumption balance, and network lifetime in
large-scale UWSN deployments. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 01 (Paper No.6443) |
|
|
|
Title: |
Graph-Aware MAPPO for Event-Coverage-Oriented Underwater Sensor Deployment under
Meandering Currents |
|
|
|
Author: |
Jianmin Yang1, Ying Huang2, Can Wang3, Ming Gong1, Yan Lin1 and Wenwei Chen1 |
|
|
|
Abstract: |
Event-coverage-oriented underwater acoustic sensor deployment aims to adapt
mobile sensor positions to nonuniform event distributions. In current-affected
underwater environments, this task is complicated by the coupling between
controlled movement and ocean-current-induced drift. Existing heuristic methods
usually rely on handcrafted movement rules or scenario-specific search, which
limits their ability to provide reusable decision policies. This paper presents
a graph-aware multi-agent proximal policy optimization framework for underwater
sensor deployment under meandering currents. The environment is modeled as a
three-dimensional monitoring region affected by the Meandering Current Mobility
model, where each sensor node selects a discrete movement action. For each
agent, a sensor-event graph is constructed to encode acoustic communication
neighborhoods among sensors and distance-dependent sensing relations between
sensors and events. A graph-aware actor extracts the encoded representation of
the current sensor node for decentralized action selection, while training is
performed with decentralized actors and a shared critic. Simulation results show
improved coverage-efficiency performance during training and better final
performance than the compared baseline. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 01 (Paper No.6446) |
|
|
|
Title: |
Q-Learning Routing with Incremental Topology Maintenance for Multi-Sink
Underwater Acoustic Sensor Networks under Ocean-Current Mobility |
|
|
|
Author: |
Jianmin Yang1, Muzi Cui1, Ying Huang1, Haiquan Shi2, Jinwang Luo2, Yonghui
Zhong1 and Yue Ma1 |
|
|
|
Abstract: |
Underwater acoustic sensor networks (UASNs) suffer from severe routing
degradation caused by continuous node displacement driven by ocean currents,
which rapidly invalidates cached topology and link quality. Existing protocols
either ignore current-induced mobility or rely on costly full topology
recomputation, leading to high packet loss and excessive energy consumption.
This paper proposes IR2P-DLMH, integrating four coordinated mechanisms. For a
start, an event-driven incremental topology maintenance scheme updates only
affected neighbor entries and Q-table values when cumulative displacement
exceeds a threshold, avoiding global recomputation while staying synchronized
with ocean-current-driven topology changes. Additionally, a link stability
predictor evaluates candidate links at drift-projected future positions,
proactively bypassing paths about to fail under current influence. Next, a
capability-aware reward coefficient steers relay traffic toward
hardware-superior nodes without centralized control. Ultimately, a multi-sink
load balancing strategy jointly minimizes propagation distance and load
deviation across sinks. Simulation results demonstrate that IR2P-DLMH
significantly reduces packet loss ratio and energy consumption per packet while
achieving lower end-to-end delay compared to VBF and HH-VBF under varying ocean
current velocities and network densities. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 01 (Paper No.6447) |
|
|
|
Title: |
A Delay-Aware Q-Learning-Based Medium Access Control Protocol for Underwater
Acoustic Networks |
|
|
|
Author: |
Jianmin Yang1, Longsen Du2, Zhuoqian Wu1, Ma Yue1, Yonghui Zhong1 and Pengyu Du3 |
|
|
|
Abstract: |
Underwater acoustic networks suffer from long propagation delay, limited
bandwidth, high bit error rate, and severe packet collisions caused by
multi-node channel contention. Conventional ALOHA-based medium access control
(MAC) protocols are simple and easy to deploy, but their random access nature
leads to rapid performance degradation under medium and high traffic loads.
Although TDMA-based protocols can reduce collisions through scheduled slot
allocation, they usually require strict time synchronization, centralized
coordination, or prior topology knowledge, which limits their adaptability in
dynamic underwater environments. To address these problems, this paper proposes
DA-ALOHA-Q, a delay-aware Q-learning-based MAC protocol for underwater acoustic
networks. The proposed protocol introduces a lightweight reinforcement learning
mechanism into slotted ALOHA and simplifies the MAC contention problem into a
single-state multi-action slot selection problem. Each node only maintains a
local one-dimensional Q-table and updates its slot preference according to ACK
reception or timeout feedback. In this way, distributed nodes can gradually form
a pseudo-orthogonal slot allocation pattern without centralized scheduling or
explicit signaling exchange. The protocol is implemented in NS-3 with the
Aqua-Sim-NG underwater network module. Simulation results under both ring and
three-dimensional random topologies show that DA-ALOHA-Q can reduce packet
collisions, improve packet delivery ratio, and maintain more stable throughput
under high traffic loads compared with conventional ALOHA and adaptive backoff
ALOHA. |
|
|
|
|
|
|
|
|
COMM Session 02 |
|
|
|
Time: 09:00~10:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Yingke Zhao, Shaanxi University of Science and Technology |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 02 (Paper No.6435) |
|
|
|
Title: |
Change-Point Aware Spatiotemporal Hawkes Process for Dynamic Topology Inference
in Non-Cooperative Underwater Acoustic Networks |
|
|
|
Author: |
Gaoyue Ma1, Xiaohong Shen1, Yuwen Yan1, Bo Geng1 and Haiyan Wang1 |
|
|
|
Abstract: |
The coexistence of multiple underwater acoustic networks (UANs) in ocean
environments leads to highly congested and uncoordinated communication over
bandwidth-limited channels. In such scenarios, inferring communication topology
from passive observations is essential for spectrum sharing and mission
coordination. However, long and uncertain acoustic propagation delays, together
with time-varying communication patterns, make this problem particularly
challenging. In this paper, we propose a Hawkes process model for passive
topology inference in UANs that explicitly accounts for acoustic propagation and
piecewise stationarity. The model incorporates an excitation kernel governed by
a Gamma distribution to characterize propagation delays and temporal uncertainty
in acoustic signal transmission. To model topology evolution over time, we
formulate dynamic topology inference as a change-point detection problem and
develop a candidate generation and dynamic programming segmentation framework
that jointly estimates change points and interaction structures. Experimental
results demonstrate that the proposed method achieves accurate and robust
topology inference under varying observation durations and arrival-time
uncertainty levels, and reliably detects topology changes, outperforming
existing baseline approaches. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 02 (Paper No.6439) |
|
|
|
Title: |
Doppler-Scale Domain Turbo Equalization for Wideband Time-Varying Channels |
|
|
|
Author: |
Yuanyuan Ou1, Zhehan Guo2, Jianchun Xu1 and Yaokun Liang2 |
|
|
|
Abstract: |
Wideband time-varying channels, particularly underwater acoustic channels,
present significant challenges for reliable communication due to
frequency-dependent non-uniform Doppler shifts. While orthogonal time frequency
space (OTFS) has been adopted for time-varying systems, it suffers from
performance degradation in these wideband scenarios. In this paper, an encoded
orthogonal delay scale space (ODSS) scheme integrated with Turbo equalization is
proposed. The proposed scheme leverages the inherent characteristics of the
delay-scale domain in ODSS to suppress wideband Doppler interference and enhance
equalization accuracy. Moreover, the ODSS modulation is extended to higher-order
modulation schemes to improve spectral efficiency for wideband transmission. To
validate the superiority of the proposed Turbo equalization algorithm,
simulations are carried out under typical wideband time-varying conditions,
comparing with existing equalization algorithms under the ODSS framework. The
results demonstrate a BER performance improvement over existing ODSS-based
algorithms, with performance enhancement with the increasing number of Turbo
iterations. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 02 (Paper No.6440) |
|
|
|
Title: |
Transmit Beamforming based on Temporally-Stable Channel Parameters for
Underwater Acoustic OFDM Communications |
|
|
|
Author: |
Pengxiang Zhang1 and Jun Tao1 |
|
|
|
Abstract: |
Due to the long feedback delay, conventional transmit beamforming approaches
relying on instantaneous channel state information (CSI) are generally
inapplicable over underwater acoustic channels. This paper proposed a robust
beamforming method for underwater acoustic orthogonal frequency division
multiplexing (OFDM) systems, by relying on temporallystable channel parameters
in form of direction of arrival (DoA) and large-scale path gain. A path
identification (PI) algorithm is employed to estimate aforementioned channel
parameters, based on which beamforming weights are designed under the minimum
variance distortionless response (MVDR) criterion. To improve numerical
stability, diagonal loading technique is further incorporated. Simulation
results demonstrate that the proposed parametric MVDR (P-MVDR) method
outperforms existing conventional beamforming (CBF) and null-steering (NS)
beamforming both utilizing the DoA knowledge of principal path only. Moreover,
it is less sensitive to DoA estimation error. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 02 (Paper No.6460) |
|
|
|
Title: |
Robust Adaptive Beamforming Based on Sensitivity-Adaptive Gauss-Chebyshev
Quadrature |
|
|
|
Author: |
Han Chuang1, Li Xiang1, Sun Yandong1, Wang Yuexian1 and He Chengyan1 |
|
|
|
Abstract: |
Robust adaptive beamforming (RAB) techniques based on interference-plus-noise
covariance matrix (INCM) reconstruction have demonstrated strong performance
under steering vector mismatch. However, existing methods often suffer from high
computational complexity and sensitivity to inaccurate direction-of-arrival
(DOA) estimation. To address these issues, this paper proposes a novel RAB
algorithm termed sensitivity-adaptive Gauss-Chebyshev quadrature (SA-AGCQ). The
proposed method integrates three key components. First, a centroid-based DOA
estimation approach with power-squared weighting is developed to improve
robustness against noise and angular mismatch. Second, Gauss-Chebyshev
quadrature is employed to efficiently reconstruct the INCM, significantly
reducing the number of integration points compared with conventional uniform
sampling. Third, a sensitivity-adaptive weighting mechanism is introduced to
emphasize dominant interference regions while suppressing noise amplification
during covariance reconstruction. Simulation results demonstrate that the
proposed method performance over a wide range of SNRs, maintains robustness
under steering vector mismatches, and performs effectively even with limited
snapshots. In addition, the proposed algorithm reduces computational complexity
compared with conventional INCM reconstruction approaches, making it suitable
for practical applications. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 02 (Paper No.6473) |
|
|
|
Title: |
DFT-spread OTFS Waveform Design for ISAC-enabled Underwater Acoustic
Communications |
|
|
|
Author: |
Jiaqi Yang1, Tonghui Zheng1, Mingqi Jin1, Chenming Zhu1 and Chengbing He1 |
|
|
|
Abstract: |
Integrated Sensing and Communications (ISAC) is a key trend for future
underwater acoustic (UWA) systems. While Orthogonal Time Frequency Space (OTFS)
modulation is robust against doubly selective fading, its high peak-to-average
power ratio (PAPR) and pilot overhead limit its practical use. This paper
proposes a DFT-spread OTFS (DFT-s-OTFS) waveform design for ISAC-enabled UWA
communications. We employ DFT precoding to suppress PAPR and integrate a
superimposed pilot scheme to enhance spectral efficiency by eliminating
dedicated guard bands. Furthermore, an advanced iterative receiver is developed
to facilitate joint channel estimation, interference cancellation, and
equalization. Simulation results validate the efficacy of the proposed
framework, demonstrating superior error performance and robustness in
challenging UWA channels. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 02 (Paper No.6485) |
|
|
|
Title: |
Joint CFO and Channel Estimation for Underwater Acoustic OCDM Communications |
|
|
|
Author: |
Lingling Zhang1, Hanyu Guo1, Chengkai Tang1 and Lu Ma2 |
|
|
|
Abstract: |
OCDM, by leveraging the diagonal distribution characteristic of Chirp signals in
the time-frequency domain, modulation and demodulation are achieved via the
Discrete Fresnel Transformation (DFnT), endowing the signal with full diversity
gain and inherent resistance to multipath and Doppler interference. This paper
present the joint CFO and channel estimation for underwater acoustic OCDM
communications. It decouples the pilot and data sub-blocks by adopting the
multicarrier OCDM mode, and estimate the CFO and channel multipath effect with
the pilots. Simulation result shows the proposed method could decrease the
channel estimation error and improve the robustness of OCDM communication over
underwater acoustic channel. |
|
|
|
|
|
|
|
|
COMM Session 03 |
|
|
|
Time: 11:00~12:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Yingke Zhao, Shaanxi University of Science and Technology |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 03 (Paper No.6377) |
|
|
|
Title: |
A Combinational Neural Network MIMO-OFDM Uplink Receiver for LEO Satellite
Constellations |
|
|
|
Author: |
Shan Lu1, Chengkai tang1, Yi Zhang1, Lingling Zhang1, Yangyang Liu1 and Zesheng
Dan1 |
|
|
|
Abstract: |
Low-Earth-orbit (LEO) mega-constellations are being deployed as a key
infrastructure for future space-air-ground integrated networks. The conventional
single-TT&C-station/single-satellite uplink, however, is not well suited to
scenarios that require wide coverage, reliable access, and high-rate
transmission. Severe path loss, multipath fading, limited pilot overhead, and
nonlinear interference further complicate channel estimation and signal recovery
in the uplink receiver. To address this problem, we develop MIMO-CN, a
model-and-data-driven MIMO-OFDM receiver for the single-station multi-satellite
TT&C uplink. The receiver consists of a model-initialized channel-estimation
subnetwork followed by a signal-detection subnetwork. LS-based channel
estimation and ZF/MMSE-based detection are used as communication-domain
initialization, and the neural subnetworks are trained to refine the residual
channel and detection errors. Simulation results over a Saleh multipath fading
channel with additive white Gaussian noise show that MIMO-CN achieves a lower
bit-error rate than fully connected DNN and LMMSE-MMSE baselines in both
sufficient-pilot and pilot-limited cases. For the same BER target, MIMO-CN
reduces the required SNR by approximately 3-5 dB and requires much less memory
than a conventional fully connected DNN receiver. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 03 (Paper No.6449) |
|
|
|
Title: |
Research on PMF-FFT Capture Algorithm with Dual segment Adjustable Parameter
Window |
|
|
|
Author: |
Zhang Hang1, Tang Cheng kai2 and Zhang Ling ling3 |
|
|
|
Abstract: |
In response to the problems of severe main lobe attenuation and significant
scallop loss in the PMF-FFT acquisition algorithm for low orbit satellite
signals under conditions of large Doppler frequency shift range and high capture
search dimension, this paper proposes a PMF-FFT acquisition algorithm based on a
dual segment adjustable parameter window. This algorithm divides the capture
process into two stages: PMF and FFT. In each stage, adjustable parameter
windows are introduced for weighted processing, and the window parameters are
adjusted to achieve optimized matching in different stages, thereby achieving
joint suppression of main lobe attenuation and scallop loss. The simulation
results show that this method can significantly suppress main lobe attenuation
and scallop loss, and compared with existing methods, the capture time and
detection probability are significantly improved. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 03 (Paper No.6475) |
|
|
|
Title: |
Research on Anti-Jamming Dynamic Routing and Link Scheduling Algorithms for
Heterogeneous |
|
|
|
Author: |
Jing Cao1, Baowang Lian1 and Xiaoqin Xue2 |
|
|
|
Abstract: |
Abstract—The Internet of Things composed of low-Earthorbit
satellites, high-altitude platform stations, UAV relays,
and ground tactical networks is typically characterized by wide
communication coverage, heterogeneous network nodes, and
dynamically varying network topology. When external UAVs
deploy electromagnetic suppression or localized link jamming,
communication links are further degraded by node mobility,
multipath fading, and transient link disruptions. Traditional
routing protocols relying on static topology information or
shortest-path criteria suffer from escalating end-to-end delay,
declining packet delivery ratio, and frequent route
reconfiguration under such conditions. To address these
challenges, this paper incorporates link quality indicators,
channel availability probability, predicted link expiration time,
navigation positioning uncertainty, and jamming risk into the
routing decision state space, and constructs a four-layer
heterogeneous network model consisting of LEO satellites,
HAPS, UAV relays, and ground mesh nodes. A reward
function jointly optimizing end-to-end delay, link reliability,
hop count, and scheduling conflicts is designed. On this basis, a
Navigation-aided Graph Proximal Policy Optimization with
Link Scheduling algorithm, termed Navi-GPPO-LS, is
proposed. The algorithm introduces anti-jamming navigation
positioning and timing information into the reinforcement
learning state space, extracts dynamic topology features
through graph feature aggregation, and employs jamming risk
prediction and link scheduling conflict penalties to achieve
cooperative interference avoidance for multiple traffic flows.
Simulation results with LQI-Dijkstra, prediction-aided Qrouting,
TD3-like scheduling, and conventional GNN-PPO as
baselines demonstrate that Navi-GPPO-LS achieves superior
comprehensive performance in terms of packet delivery ratio,
end-to-end delay, throughput, outage ratio, and scheduling
conflict control. This work provides algorithmic reference for
enhancing the anti-jamming communication capability of
ground unmanned systems in UAV and counter-UAV
communication countermeasure scenarios. |
|
|
|
|
|
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 03 (Paper No.6478) |
|
|
|
Title: |
Full-Link Simulation of Physical Layer for LEO Satellites |
|
|
|
Author: |
Xinrui Hu1, Wenjin Hou2, Hong Wen3, Ruixiang Yao3, Wendi Ma3 and Yuhui Wu3 |
|
|
|
Abstract: |
Abstract-With the advantages of low latency, wide coverage and large bandwidth,
low-orbit satellite communication has become a core support for the space-earth
integrated information network, and is widely applied in remote area
communication, aviation navigation, emergency disaster relief and other fields.
High dynamic Doppler frequency shift caused by high-speed movement of low-orbit
satellites and additive white Gaussian noise interference seriously affect the
reliable transmission of LEO satellite signals, and existing simulation
platforms are difficult to achieve accurate full-process simulation. Therefore,
this paper constructs a full-link simulation platform for LEO satellite signal
physical layer based on MATLAB, which completely covers signal generation at the
transmitter, high dynamic channel transmission, as well as synchronization,
equalization and decoding processing at the receiver. Experimental results show
that QPSK modulation has outstanding robustness with bit error rate reaching
zero at 7 dB; 16QAM modulation possesses higher spectral efficiency and achieves
zero bit error rate at 11 dB, and LDPC coding brings remarkable error correction
gain. This platform can provide effective support for algorithm verification and
performance evaluation of LEO satellite physical layer. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 03 (Paper No.6480) |
|
|
|
Title: |
LDPC Coding Adaptation and Anti-Jamming Optimization for LEO Satellite
Air-Ground Links |
|
|
|
Author: |
Xiangwei Ren1, Xiayu Chen1, Wenjin Hou2 and Haowen Deng1 |
|
|
|
Abstract: |
Abstract-To support low-Earth-orbit (LEO) satellite communications in
space-based broadband coverage, ubiquitous information access, and highly
reliable dedicated communications, physical-layer coding design for different
mission scenarios has become a key factor affecting air-ground link performance.
Given the significant differences between commercial broadband and dedicate LEO
systems in throughput, reliability, and anti-jamming capability, this paper
focuses on the uplink and downlink air-ground links between user terminals and
satellites, and analyzes their differences in LDPC code construction, parameter
configuration, and encoding/decoding mechanisms. A unified link simulation
platform is built to quantitatively compare bit error rate, block error rate,
coding gain, throughput, and decoding complexity under additive white Gaussian
noise, impulsive interference, and broadband intentional jamming. Results show
that the commercial LDPC code emphasizes high spectral efficiency and low
complexity, achieving over 35% higher throughput than the dedicated code under
ideal channels, whereas the dedicated LDPC code emphasizes reliable transmission
under low-SNR and strong-interference conditions, improving coding gain by
1.2-2.4 dB and reducing block error rate by one to two orders of magnitude. The
findings provide a reference for air-ground link coding optimization,
anti-jamming transmission design, and heterogeneous-constellation compatibility. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 03 (Paper No.6481) |
|
|
|
Title: |
Routing Optimization Models and Enhanced Intelligent Algorithms for
Inter-Satellite Laser Links in LEO Mega-Constellations |
|
|
|
Author: |
Xiangwei Ren1, Wenjin Hou2, Wendi Ma3 and Xingyun Wei1 |
|
|
|
Abstract: |
Abstract-To address the highly dynamic topology of LEO mega-constellations and
the high ATP overhead and frequent handovers of laser inter-satellite links
(LISLs)-which cause routing reconstruction, traffic imbalance, and latency
fluctuations-this paper establishes a multi-objective routing model constrained
by constellation scale and laser physics. The model jointly optimizes end-to-end
latency, link reliability, node load balancing, and handover frequency, while
incorporating constraints such as laser terminal count, line-of-sight
connectivity, minimum link survival time, and handover cost, thereby overcoming
conventional algorithms' weak coupling with laser link characteristics. We
propose an Improved Non-dominated Sorting Genetic Algorithm (INSGA-III) for
optimal routing decisions. Co-simulation on a 24×66 LEO constellation using STK
11.0 and MATLAB R2022b shows that, compared with Dijkstra, MOPSO, and standard
NSGA-III, the proposed method reduces average latency by 19.7%, improves
reliability by 13.2%, optimizes the load-balancing factor by 28.4%, and cuts
handovers by 32.6%, with faster convergence and a more uniform Pareto front. |
|
|
|
|
|
|
|
|
COMM Session 04 |
|
|
|
Time: 14:00~15:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Dongyuan Shi, Northwestern Polytechnical University |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 04 (Paper No.6320) |
|
|
|
Title: |
UWB Non-Contact Respiratory and Heart Rate Monitoring System for 6G ISAC |
|
|
|
Author: |
Yuanhui Cao1, Xingguang Geng1, Fei Yao1, Shuaishuai Hou1, Yitao Zhang1 and
Yunfeng Wang1 |
|
|
|
Abstract: |
Monitoring of respiratory signal and heartbeat signal provides essential
insights into an individual's physiological and psychological conditions.
However, current methods involving direct contact present challenges including
user discomfort and compromised accuracy. This study proposes an advanced
non-contact respiratory and heart rate monitoring system based on channel
impulse response (CIR) using ultra-wideband (UWB) radar technology operating at
6.5 GHz. The system employs a joint signal processing algorithm that integrates
time-domain coherent accumulation (TDCA) and variational mode decomposition
(VMD), followed by multi-algorithm fusion for high-precision estimation of vital
sign frequencies. Experimental results demonstrate that the root mean square
error (RMSE) between the estimated respiratory rate and heart rate and those
measured by a polygraph reference device are 2.98% and 1.54%, respectively. This
approach thus demonstrates significant potential as a reliable, non-invasive
method for accurate detection of physiological vital signs in both clinical and
home environments, which may provide reference value for future Integrated
Sensing and Communication (ISAC) technology in 6G. |
|
<>
|
|
|
|
|
|
|
Oral Session: |
COMM Session 04 (Paper No.6344) |
|
|
|
Title: |
STAR-RIS and RSMA for ISAC: A Joint Optimization Framework |
|
|
|
Author: |
Zhentao Wang1, Wenbin Sun1, Xin Yang1, Lili Chen1, Qian Xu1 and Lin Wang1 |
|
|
|
Abstract: |
This paper proposes a novel integrated sensing and communication system assisted
by a simultaneous transmitting and reflecting reconfigurable intelligent
surface. The framework addresses challenges in serving users with diverse
channel conditions while performing sensing tasks in obstructed environments. By
incorporating rate-splitting multiple access at the base station, the system
manages multi-user interference, improving communication rates and user
fairness. The STARRIS dynamically reshapes the wireless environment through its
transmission and reflection capabilities, creating optimized propagation paths
for both functionalities. We establish a comprehensive system model and
formulate a joint optimization problem for active precoding, passive
beamforming, and rate-splitting parameters to maximize the weighted sum-rate. An
efficient alternating optimization algorithm using fractional programming and
semi-definite relaxation is developed to solve this complex problem. Numerical
results demonstrate the scheme's substantial improvements in spectral efficiency
and sensing accuracy over conventional systems, achieving notable performance
gains under |
|
|
|
practical constraints. |
|
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 04 (Paper No.6345) |
|
|
|
Title: |
Integrated Communication and Jamming Waveform Design for OTFS-Based
Comb-Spectrum Jamming with Spatial Index Modulation |
|
|
|
Author: |
Tianyu Kou1, Wen-Bin Sun1, Qian Xu1, Zhaolin Zhang1 and Ling Wang1 |
|
|
|
Abstract: |
Abstract-We propose a novel integrated communication and jamming (ICAJ) waveform
utilizing a Multiple-Input Single-Output (MISO) Orthogonal Time Frequency Space
(OTFS) framework combined with Spatial Index Modulation (SIM). By exploiting the
delay-Doppler (DD) domain sparsity of wideband comb-spectrum jamming,
communication data are orthogonally embedded into low-energy valleys. To
mitigate the inherent capacity-covertness trade-off, spatial antenna indexing
conveys additional information bits without increasing communication power
allocation. We derive closed-form expressions for the effective
signal-to-interference-plus-noise ratio (SINR), system throughput, and
computational complexity under imperfect channel state information (CSI).
Simulations demonstrate that the MISO-SIM architecture significantly enhances
throughput compared to the Single-Input Single-Output (SISO) baseline, while
strictly preserving macroscopic waveform covertness in terms of time-domain
Pearson Correlation Coefficient (PCC) and in-band Power Spectral Density (PSD)
metrics. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 04 (Paper No.6346) |
|
|
|
Title: |
A 140GHz Array Antenna Based on 4-Bit RF-MEMS Phase Shifters |
|
|
|
Author: |
Haozhe Hou1, Jianming Huang1, Naibo Zhang1, Zilai Wang1, Yansong Cui1, Weizheng
Ren1, Yiran Zhang1, Xinyue Sun1 and Guangbin Dong2 |
|
|
|
Abstract: |
This paper presents a 140 GHz phased array antenna utilizing a 4-bit distributed
micro-electro-mechanical systems (MEMS) transmission line (DMTL) phase shifter.
The antenna element employs a Wilkinson power divider with a 90° phase
difference to feed adjacent edges of a square patch, achieving circular
polarization. Sector-shaped cut corners and rectangular stubs are introduced to
widen the axial ratio and impedance bandwidths, respectively. The phase shifter
adopts a cascaded architecture of 16 unit cells, each loaded with a MEMS bridge
featuring metal-air-metal (MAM) contacts to enhance phase shift per unit length.
Simulation results show the element operates from 133 to 150 GHz with S11 < -15
dB and AR < 3 dB. The 4-bit phase shifter provides a 360° phase range with <4 dB
insertion loss and <5° phase error. The 2×8 array achieves beam steering from
-45° to +45° with a peak gain >13 dB, validating a low-loss beamforming approach
for terahertz applications. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 04 (Paper No.6371) |
|
|
|
Title: |
An analysis of diurnal VLF propagation field versus local time using modified
IRI data |
|
|
|
Author: |
Junke Wang1, Zhewen Chen1, Shitian Zhang2, Kuisong Zheng1, Songming Zou1 and
Qiang Wu1 |
|
|
|
Abstract: |
To precisely obtain the VLF propagation field during 24 hours in the
Earth-ionosphere waveguide, an EM-FDTD method of combing the FDTD and the
electron momentum equation is developed. The absent data of electron density in
the lower height region is supplemented with the log-linear fitting algorithm.
According to the Yee cell, the iterative formulas of the electron momentum
equation are deduced in detail. The long-distance propagation field from VTX to
Gwalior are calculated. By comparing the numerical results and the measured
results, a good agreement is reached. It is shown that the proposed method is
validated and corrected. |
|
|
|
|
|
|
|
|
COMM Session 05 |
|
|
|
Time: 16:00~17:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Dongyuan Shi, Northwestern Polytechnical University |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 05 (Paper No.6366) |
|
|
|
Title: |
A Network Traffic Prediction Method Based on STL-Informer-DLinear |
|
|
|
Author: |
Yuze Su1, Haifeng Yang1 and Yong Wang1 |
|
|
|
Abstract: |
Network traffic prediction is a key technology for efficient network management
and optimization. To improve the accuracy of network traffic prediction, a
network traffic prediction method based on STL-Informer-DLinear is proposed.
Firstly, according to the trend, periodicity and randomness characteristics of
network traffic, the Seasonal and Trend Decomposition using Loess (STL)
algorithm is introduced to decompose the network traffic, obtaining trend,
seasonal, and remainder components. Then, the Informer model is used to predict
the trend and remainder components, while the DLinear model is used to predict
the seasonal component. Finally, the predicted results of all components are
combined to obtain the final traffic prediction results. Experimental results
show that the proposed STL-Informer-DLinear method achieves better prediction
performance than other traffic prediction methods on real network traffic
datasets. Specifically, it achieves a 61.5%-96.3% reduction in MSE, a
21.3%-80.7% reduction in RMSE, a 31.2%-83.6% reduction in MAE, and a 26%-89.1%
reduction in MAPE, which verifies the excellent performance of the
STL-Informer-DLinear network traffic prediction method. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 05 (Paper No.6414) |
|
|
|
Title: |
A High-speed Visible Light Coherent Communication System based on Balanced
Conjugate Framing |
|
|
|
Author: |
Zijian Zhou1, Wenting Ju1, Yuhan Hu1, Zengyi Xu1 and Nan Chi1 |
|
|
|
Abstract: |
The rapid growth in data traffic in today's information age places high demands
on throughput of communication systems. Against this background, visible light
communication (VLC) systems have attracted great attention for their ample
spectrum resources, and strong resistance to electromagnetic interference.
However, the performance of intensity-modulation direct-detection (IM/DD) based
VLC systems is restricted by limited device bandwidth and channel nonlinearity.
To address this issue, this paper proposes a balanced conjugate framing (BCF)
scheme for high-speed VLC system based on self-homodyne coherent detection. The
proposed scheme offers a new frame structure, which contains the original signal
block and its conjugated counterpart. Subtraction is applied to the two signal
blocks in the received signal to eliminate the common-mode noise. Experimental
results show that the proposed BCF scheme can effectively improve the received
signal quality and achieve 12 Gb/s visible light coherent transmission. This
scheme provides a new perspective for mitigating noise in visible light coherent
communication (VLCC) systems at the coding level. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 05 (Paper No.6421) |
|
|
|
Title: |
On the SOR Sampling for Discrete Symbol Detection: An Energy-Based Model
Perspective |
|
|
|
Author: |
Zhiheng Zhang1, Le Yang2 and Jun Tao3 |
|
|
|
Abstract: |
This paper revisits the successive over-relaxation (SOR) sampling for discrete
symbol detection from the perspective of energy-based models (EBMs). We first
derive the complex-valued realization of the discrete unadjusted Langevin
algorithm (DULA) using the Wirtinger calculus. This allows us to analytically
establish the equivalence between the SOR sampler and symbol-wise DULA with a
certain stepsize schedule. We further show that the annealed SOR (ASOR)
sampling, a recently proposed enhancement of the SOR sampling, can be viewed as
the simulated annealing (SA) DULA. Simulation experiments corroborate the
theoretical developments. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 05 (Paper No.6459) |
|
|
|
Title: |
Channel Estimation for UAV Swarms Based on Space-Time Autoregressive Models |
|
|
|
Author: |
wang pp1, Lian Baowang1, Guo Huachang2, Zhao Na2, Ma Xingbing2 and Hou Yu2 |
|
|
|
Abstract: |
To address the challenges posed by the high mobility and dynamic topology of
unmanned aerial vehicle (UAV) swarms, which result in complex and highly
variable wireless channels, and the difficulty of model-based channel estimation
methods in accurately describing actual channel fading characteristics, this
paper proposes a Space-Time Autoregressive (ST-AR) channel estimation method
suitable for UAV swarm communication scenarios. First, an ST-AR state-space
model is constructed based on the acquired channel space-time sequence features,
second, the AR coefficients are updated online in an adaptive manner by
combining real-time channel autocorrelation functions with relative geometric
topology; subsequently, a model goodness-of-fit test based on prediction
residuals is introduced to achieve model mismatch detection and order
adaptation; finally, the calibrated ST-AR model is used for channel prediction
and recursive estimation. Simulation results demonstrate that, in a typical UAV
swarm scenario, the proposed method achieves an end-to-end bit error rate (BER)
improvement of approximately 2 dB compared to the MMSE method while reducing
pilot overhead by 50%, validating the method's effectiveness and robustness in
high-dynamic, low-overhead scenarios. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 05 (Paper No.6476) |
|
|
|
Title: |
A WFRFT Secure Communication Method Based on Chaos-Driven Constellation Phase
and Polarity Flipping Hybrid Encryption |
|
|
|
Author: |
Jiayue Li1, Qingwei Meng1, Yanzhi Yun1, Han Wang1, Dan Wang1, Liangsi Zhou1 and
Linghua Su1 |
|
|
|
Abstract: |
To address the issues of limited key space and insufficient constellation
concealment in physical layer encryption, a Weighted Fractional Fourier
Transform (WFRFT) secure communication method based on chaos-driven
constellation phase and polarity flipping hybrid encryption is proposed. This
method utilizes a four-dimensional (4D) hyperchaotic system to generate four
chaotic sequences, which respectively drive row-column scrambling, dynamic phase
shift, and polarity flipping. Finally, the WFRFT is applied to achieve
constellation confusion and diffusion, rendering the encrypted signal with
Gaussian noise-like characteristics. Computer simulation results demonstrate
that while ensuring communication reliability for legitimate users, the proposed
method significantly enhances system security. Under a fixed WFRFT order,
unauthorized users cannot accurately decrypt the encrypted information even if
the key error is merely at the level. The system bit error rate (BER) approach. |
|
|
|
|
|
|
|
|
Oral Session: |
COMM Session 05 (Paper No.6477) |
|
|
|
Title: |
Enhanced Cross-Representation Domain Black-Box Transfer Attack Method for
Automatic Modulation Recognition |
|
|
|
Author: |
xing Yun Wei1, Xiangwei Ren1, Wenjin Hou2, Hong Wen3, Ruonan Jing3 and Wenqi
Tang3 |
|
|
|
Abstract: |
Abstract-As a core technology in cognitive radio and spectrum supervision,
Automatic Modulation Recognition (AMR) is facing severe security challenges
arising from the inherent vulnerability of deep learning models. Existing
adversarial attack studies are mostly limited to a single representation domain,
and the transferability of adversarial samples degrades significantly when
confronted with "representation domain mismatch" in black-box scenarios. To this
end, this paper proposes an attack enhancement method based on
Multi-Representation Loss Ensemble (MRLE). By constructing a unified time-domain
physical perturbation space and introducing a differentiable Fourier transform
layer, the method establishes a gradient backpropagation path between
heterogeneous representations, achieving deep coupling of time-frequency
dual-domain gradients and end-to-end collaborative optimization under physical
constraints. Experimental results demonstrate that the proposed method can
significantly improve the success rate of cross-representation domain transfer
attacks, effectively overcome the transfer asymmetry bottleneck, and control the
adversarial distortion. It provides an important reference for the security
evaluation of AMR systems in black-box scenarios. |
|
|
|
|
|
|
|
|
SPSS Session 01 |
|
|
|
Time: 14:00~15:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Yi Yu, Southwest University of Science and Technology |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 01 (Paper No.6375) |
|
|
|
Title: |
Performance Assessment of Galileo HAS PPP Time Transfer |
|
|
|
Author: |
Xiaofei Liang1 |
|
|
|
Abstract: |
With the rapid evolution of GNSS Precise Point Positioning (PPP) technology,
high-precision remote time transfer using real-time State-Space Representative
(SSR) corrections has become a critical research focus. This study investigates
the comprehensive performance assessment of the PPP time transfer model
utilizing Galileo High Accuracy Service (HAS) real-time corrections.Firstly, an
ionosphere-free (IF) PPP time transfer mathematical model based on Galileo HAS
was constructed. Secondly,Using global IGS station data connected to external
Hydrogen Masers, systematic experiments were conducted across ultra-short,
short, and intercontinental long baselines. Finally, Taking post-processed CODE
products as the reference, the system was assessed using Root Mean Square Error
(RMSE) for external accuracy and Modified Allan deviation (MDEV) for frequency
stability. Results show that Galileo HAS-based time transfer is generally
consistent with post-processed precision series. The average external RMS across
all links is better than 0.3 ns, reaching 0.088 ns for ultra-short baselines. In
terms of frequency stability, the average performance at $10,000$ seconds is
better than the $1 \times 10^{-14}$ for all links, effectively approaching the
level of post-processed products in long-baseline scenarios.The study indicates
that Galileo HAS is fully capable of supporting global-scale real-time precision
time transfer. This provides a robust theoretical and technical foundation for
building low-cost, high-efficiency, and reliable global real-time time transfer
systems. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 01 (Paper No.6389) |
|
|
|
Title: |
An Autonomous Rescue USV System with Hierarchical Search and Exact-Penalty
Planning |
|
|
|
Author: |
Longyun Yuan1, Chengkai Tang1, Lingling Zhang1, Yangyang Liu1, Zesheng Dan1 and
Ding Yuan1 |
|
|
|
Abstract: |
Autonomous man-overboard rescue on water surfaces with Unmanned surface vehicles
requires reliable long-range target acquisition, robust close-range
identification, and safe motion planning under wave-induced disturbances and
dynamic obstacles. This paper presents a closed-loop autonomous USV system for
search-and-rescue that integrates a radio-based long-range search with
near-range multimodal target confirmation and a collision-aware exact-penalty
trajectory planner. In the long-range stage, a VHF radio transceiver returns an
ID and coarse target information, enabling the USV to navigate towards the
target. At closer range, the system switches to sensing using vision, thermal
imagery, and mmWave radar to produce refined target state for the planner. The
exact-penalty planner incorporates obstacle-avoidance constraints as penalty
terms and balances route directness and safety precautions. Field experiments
were conducted in both lake and marine environments. Radio-guided acquisition
together with near-range recognition achieved <2m positioning precision on most
samples, and the repeated tests proved that approach precision reached 2m in
real marine scenario. Compared with A*, DWA, and ACO, the proposed planner
generates smoother trajectories, provides larger safety margins, and achieves
higher path efficiency in mixed static and dynamic obstacle scenarios. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 01 (Paper No.6415) |
|
|
|
Title: |
Multi-Sensor Fusion Positioning Method for UAV Swarms with Improved CPInformer |
|
|
|
Author: |
Ruihan Shen1, Lingling Zhang1, Siyuan Luo2, Yexuan Bai1, Yiheng Chen1 and
Zengrui Zhou1 |
|
|
|
Abstract: |
With the rapid development of the low-altitude economy, UAV swarms have become a
core application platform, and UAV positioning technology is the foundation
supporting swarm applications. To address the issues of poor stability and
difficult scalability in existing combined positioning methods, this paper
proposes a multi-sensor fusion positioning method for UAV swarms based on an
improved Cooperative Positioning Informer(CPInformer). This method first
converts distributed heterogeneous navigation sources into an information
probability model, achieving a unified format for navigation information
parameters; it then designs a lightweight improved CPInformer neural network to
complete the fusion positioning of the UAV swarm. Simulation analysis under both
normal and high-noise conditions, along with validation on an actual UAV swarm,
shows that this method can effectively improve the accuracy and stability of UAV
swarm fusion positioning. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 01 (Paper No.6425) |
|
|
|
Title: |
UWB Based 3D High Precision Cooperative Positioning System Integrating
Information Geometry |
|
|
|
Author: |
Weiming Chen1 and Yangyang Liu2 |
|
|
|
Abstract: |
Ultra-wideband technology has become the mainstream solution to achieve indoor
high-precision positioning by virtue of nanosecond narrow pulses and
centimeter-level ranging accuracy. However, multipath interference,
non-line-of-sight (NLOS) propagation and dynamic noise will significantly reduce
its positioning accuracy. In this paper, a three-dimensional high-precision
cooperative positioning system based on UWB is proposed, which adopts a
five-base station redundant topology. In order to further improve the
positioning optimization performance, this paper proposes a multi-base station
information fusion algorithm based on information geometry, and develops a NLOS
detection model based on residual analysis and a clock synchronization error
correction mechanism using geometric distance back projection. The simulation
results show that the NLOS detection accuracy is more than 75% under both motion
conditions. The results show that after the information geometry combined with
the extended Kalman filter fusion algorithm proposed in this paper, the average
positioning error under the simple motion model is reduced to 0.113m, and the
positioning error under the complex motion model is reduced to 0.173m, which is
76.7% and 76.6% higher than the independent Kalman filter, respectively. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 01 (Paper No.6436) |
|
|
|
Title: |
Fingerprint-Based Visible Light Positioning Using Improved WKNN Algorithm With
Selected Reference Position |
|
|
|
Author: |
Zijian Bai1, Zhonghua Liang1, Wenqian Jiang1, Dongxin Bai1 and Zhuo Zhou1 |
|
|
|
Abstract: |
In indoor visible light positioning (VLP) scenarios, fingerprint-based
localization has garnered considerable attention due to its high accuracy and
broad applicability. Most existing fingerprint-based VLP systems directly
utilize received signal strength (RSS) values from a fixed number of reference
nodes (RNs) and employ the weighted K-nearest neighbor (WKNN) algorithm for
positioning. Such approaches failed to take into account the region-specific
requirements for the number and selection of RNs across different spatial areas,
thereby degrading the overall localization accuracy. To address this issue, this
paper leverages the relationship between RSS and physical distance to develop an
improved weighted K-nearest neighbor algorithm based on selected RN position
information (SRP-WKNN). In the proposed algorithm, the nearest RNs are
adaptively selected by considering both signal similarity and spatial proximity,
with the selection strategy differing for edge and central regions. These
selected RNs are then used to derive the mapping between RSS similarity and
physical distance. Finally, the coordinates of the selected RNs are weighted to
obtain the precise location estimate. Software simulation results show that, at
a signal-to-noise ratio (SNR) of 20 dB and an RN interval of 0.25 m, the average
positioning error (APE) is significantly reduced by at least 11.0\% and at most
41.6\% compared with traditional WKNN and its representative improved versions. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 01 (Paper No.6451) |
|
|
|
Title: |
A Loosely Integrated GBNS/INS Navigation Method for Weak Vertical Geometry
Scenarios |
|
|
|
Author: |
Chengyan He1, Jiawei Liu1, Mingliang Tao1, Zhaolin Zhang1 and Ling Wang1 |
|
|
|
Abstract: |
Ground-Based Navigation Systems (GBNS) offer reliable positioning in Global
Navigation Satellite System (GNSS)-degraded or GNSS-denied environments.
However, due to the near-ground deployment of GBNS transmitters, the
user-transmitter geometry often yields strong horizontal constraints but
extremely weak vertical observability, leading to unreliable height estimates in
GBNS standalone positioning. When such three-dimensional GBNS solutions are
directly fed into a loosely coupled GBNS/inertial navigation system (INS)
integration filter, the contaminated height information inevitably degrades the
inertial error correction and compromises overall navigation stability. This
paper proposes a height-constrained loosely coupled GBNS/INS integration method
for weak vertical geometry. Unlike conventional approaches that feed
unconstrained GBNS positions into the Extended Kalman Filter (EKF), our method
incorporates short-term INS-predicted height as a constraint during GBNS
positioning. The height-constrained GBNS solution-preserving horizontal geometry
while suppressing vertical uncertainty-is then used as the EKF measurement
input. Simulations under typical GBNS deployment (good HDOP, poor VDOP) show
that, compared with conventional loosely coupled integration, the proposed
method reduces height 1σ error from 42.21?m to 27.49?m, while maintaining
comparable horizontal accuracy (0.16?m north, 0.32?m east vs. 0.14?m and
0.33?m). These results demonstrate that the proposed INS-height-constrained
scheme significantly improves vertical stability without sacrificing the
horizontal correction benefit of GBNS. The method provides a practical and
effective solution for reliable GBNS/INS integration in weak height geometry
scenarios, with potential extensions to other ground-based ranging systems. |
|
|
|
|
|
|
|
|
SPSS Session 02 |
|
|
|
Time: 16:00~17:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Zhengqiao Zhao, Northwestern Polytechnical University |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 02 (Paper No.6306) |
|
|
|
Title: |
An Improved YOLOv11-Based Model for Object Detection in Sonar Images |
|
|
|
Author: |
Zeng Li Liu1 and Chao Yang Li1 |
|
|
|
Abstract: |
Abstract: Underwater target detection serves as a pivotal technology for domains
such as marine exploration, national defense, and underwater engineering.
Leveraging its long-range detection capabilities, sonar imaging has emerged as a
critical technological modality. However, sonar imagery is intrinsically
characterized by severe speckle noise, low contrast, and blurred target edges,
which impose significant challenges on deep learning-based detection algorithms.
To address these issues, this paper proposes YOLOCS, an enhanced architecture
based on YOLOv11. First, a C3CFB module is introduced to strengthen the
extraction of multi-scale and multi-directional features from sonar targets.
Subsequently, Space-to-Depth Convolution (SPDConv) is incorporated for
downsampling to mitigate performance degradation caused by the loss of critical
features. Experimental results on the SCTD dataset demonstrate that, compared
with the baseline YOLOv11n, YOLOCS achieves improvements of 0.5%, 2.9%, and 2.6%
in Precision, Recall, and mAP50-95, respectively. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 02 (Paper No.6315) |
|
|
|
Title: |
A Method for Depth Discrimination of Shallow Water Sound Sources Based on
Dynamic Threshold Decision for Broadband Phase Fluctuation |
|
|
|
Author: |
Zhejian Hu1 and Xuanjie Wei1 |
|
|
|
Abstract: |
In shallow water environments, the interference effect of normal modes results
in unique broadband phase fluctuations in the acoustic field. Under typical
negative gradient hydrographic conditions, by dividing the vertical array into
upper and lower arrays for summation and cross spectral analysis, the source
depth exhibits distinct separability. Since this phase fluctuation
characteristic correlates with horizontal distance, fixed decision thresholds
struggle to achieve target depth resolution across varying horizontal distances.
Therefore, this study improves upon fixed decision thresholds by introducing a
horizontal dynamic threshold for phase fluctuation feature discrimination,
enabling source depth resolution at different distances. Based on simulation
data analysis, the depth resolution performance of the horizontal dynamic
threshold selection method is evaluated. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 02 (Paper No.6372) |
|
|
|
Title: |
An Acoustic-Based Side-Scan Sonar Image Simulation Method Incorporating AUV
Attitude Parameters |
|
|
|
Author: |
Zizhuo Liang1, Jianfeng Chen1, Fen Liu1, Yifan Zhang1, Hongyu Wei1 and Yaohui
Wen1 |
|
|
|
Abstract: |
Recent advances in underwater robotics have accelerated the development of
automatic underwater target detection. However, due to environmental complexity
and cost constraints, it remains difficult to obtain sufficient underwater
target images, limiting the performance of detection algorithms. To address this
issue, simulation-based methods provide an effective alternative. In this work,
we propose a Side-Scan Sonar (SSS) image simulation method based on underwater
acoustic propagation modeling. The proposed framework simulates the complete SSS
imaging process, including transmit waveform generation, acoustic propagation,
target backscattering, echo reception, and array directivity, enabling SSS image
simulation under various parameter settings. Meanwhile, the depth and velocity
of the Autonomous Underwater Vehicle (AUV), particularly its attitude variations
during navigation, are incorporated to reproduce SSS image distortions, enabling
more realistic sonar images. To validate the effectiveness of the proposed
method, cross-domain transfer experiments are conducted. Results show that
models pre-trained on the simulated dataset consistently outperform those
pre-trained on COCO in real-world tasks, achieving improvements of 4.7% in
mAP@0.5, 8.6% in mAP@0.5:0.95, and 6.6% in Recall. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 02 (Paper No.6409) |
|
|
|
Title: |
Simulation of Elastic and Geometric Acoustic Scattering from Typical Underwater
Shell Targets |
|
|
|
Author: |
Zihao Shu1, Jianjun Zhu1, Peihong Wang1, Tian Zhou1, B.A Tarasov2 and V. I.
KOROCHENTSEV3 |
|
|
|
Abstract: |
The acoustic scattering from spherical and hemispherical-capped cylindrical
shells is investigated through numerical simulations of steel and aluminum
shells, followed by analysis of their acoustic scattering fields. The broadband
acoustic fingerprints and target strength curves reveal how geometry, size,
material, and incident angle dominate the spatial texture, orientation, and
frequency-domain characteristics of elastic scattering regions. For spherical
shells, increasing size produces denser interference fringes and richer modal
structure at low-to-mid frequencies. For cylindrical shells, the acoustic
fingerprint shows strong azimuthal dependence, enabling shape and orientation
identification. For same-sized targets of different materials, backscattering
target strength differs more significantly at low-frequency bands but converges
at high-frequency bands. These acoustic fingerprints and scattering field
patterns provide physical insights and fundamental data for feature extraction
and intelligent classification of small underwater targets. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 02 (Paper No.6448) |
|
|
|
Title: |
A Propagation Acoustic Path Error Correction Method for the Overlapping Phase
Center Algorithm in High-Frequency Synthetic Aperture Sonar |
|
|
|
Author: |
Guijuan Han1, Xiangrui Zeng1 and Yuhang Gao2 |
|
|
|
Abstract: |
When utilizing the Displaced Phase Center (DPC) algorithm for motion error
estimation in high-frequency, large-aperture synthetic aperture sonar (SAS)
systems, the propagation acoustic path error of the displaced phase center pairs
and the correlation peak shift induced by sonar motion errors are of a
comparable order of magnitude. This comparability leads to a significant
degradation in the motion error estimation performance of the DPC algorithm. To
address this issue, this paper analyzes the impact of the equivalent phase
center approximation on the DPC algorithm and proposes a propagation acoustic
path difference correction method for displaced phase center pairs in
dual-transmitter, multi-receiver SAS systems. By employing a dual-transmitter
architecture and optimizing the spatial layout of the transceiver array and the
pulse repetition interval (PRI), the proposed method ensures that the
propagation acoustic path lengths of displaced phase center pairs across
adjacent pings are strictly identical. Consequently, the proposed approach
effectively eliminates the adverse effects of the equivalent phase center
approximation on DPC-based motion error estimation. Simulation results validate
the efficacy of the proposed method. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 02 (Paper No.6465) |
|
|
|
Title: |
A Novel Method for Improving the Efficiency of CFAR Detection of Small Targets
in Side-scan Sonar Images |
|
|
|
Author: |
Qiuju Li1, Yifan Wu1, Peng Xiao2 and Gang Xiao2,3 |
|
|
|
Abstract: |
Abstract-To address the issue of low efficiency in detecting small targets in
side-scan sonar images using CFAR (Constant False Alarm Rate) detection, this
paper proposes a novel method to improve the efficiency of CFAR detection for
small targets in side-scan sonar images. This method employs morphological
reconstruction to generate an indexed sonar image, which guides the subsequent
CFAR detection algorithm and improves its efficiency in detecting small targets.
Experimental results demonstrate that this method can effectively improve the
efficiency of CFAR detection for small targets in side-scan sonar images. |
|
|
|
|
|
|
|
|
SPSS Session 03 |
|
|
|
Time: 08:30~10:00, Monday, July 20, 2026 |
|
|
|
Session Chair: Zhongxin Bai, Harbin Engineering University |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 03 (Paper No.6321) |
|
|
|
Title: |
Physics-Guided Side-Scan Sonar Image Simulation for Enhanced Object Detection |
|
|
|
Author: |
Guoqing Xie1, Ju He1, Haoran Hu1, Jinpeng Xu1, Hu Xu1 and Yang Yu1 |
|
|
|
Abstract: |
Side-scan sonar (SSS) plays an important role in underwater perception for
applications such as seabed mapping, marine inspection, and object detection.
However, the performance of deep learning-based detection methods is often
limited by the scarcity of annotated SSS datasets, as acquiring large-scale
underwater sonar data is expensive and time-consuming. To address this
challenge, this paper proposes a physics-aware side-scan sonar image simulation
framework for data augmentation in object detection tasks. The proposed method
integrates geometry-driven scene construction and acoustic-aware echo modeling
to generate realistic SSS images. Specifically, a geometry-driven mechanism is
designed to construct target structures and their corresponding acoustic shadows
based on the geometric relationship between sonar beams and object shapes.
Experimental results demonstrate that incorporating the proposed synthetic data
significantly improves detection performance compared with training using real
data alone. The proposed framework provides an efficient solution for
alleviating the data scarcity problem in underwater sonar perception tasks. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 03 (Paper No.6362) |
|
|
|
Title: |
LAT: A Lightweight Anti-occlusion Tracker for Hyperspectral Videos |
|
|
|
Author: |
Xu Li1, Fuyuan Ge1, Qing Zhang1, Baoguo Wei1, Zhendong Li1 and Junyin Yu1 |
|
|
|
Abstract: |
Hyperspectral object tracking (HOT) has attracted increasing attention due to
its powerful discriminative capabilities derived from rich spectral information
in challenging scenarios. However, existing HOT methods often face two critical
limitations: poor tracking robustness against occlusions and low computational
efficiency caused by high-dimensional spectral data and complex architectures.
To address these limitations, we propose a Lightweight Anti-occlusion Tracker
(LAT) for hyperspectral videos. Firstly, we design a discriminative band
selection (DBS) module based on the minimum redundancy maximum relevance (mRMR)
criterion, and select few spectral bands to reduce redundant computations while
maintaining object-background separability. Secondly, we introduce an
occlusion-aware branch (OB) that fuses local fine-grained features and global
statistical features to accurately predict the degree of occlusion. Thirdly, we
adopt a lightweight backbone and design a Kalman filtering-based motion
prediction (MP) module to form a hierarchical anti-occlusion tracking strategy.
The extensive experiments conducted on the HOT2024 dataset demonstrate that LAT
has advantages in both overall tracking accuracy and anti-occlusion tracking
accuracy. In addition, LAT only contains 0.69 million parameters, achieving an
excellent balance between tracking accuracy and model complexity. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 03 (Paper No.6363) |
|
|
|
Title: |
Hyperspectral Small Object Tracking Method With Target Region Perception |
|
|
|
Author: |
Xu Li1, Qing Zhang1, Yuchao Wang1, Fuyuan Ge1, Zhendong Li1 and Baoguo Wei1 |
|
|
|
Abstract: |
Hyperspectral Small Object Tracking holds significant application value in
fields such as remote sensing and autonomous driving. However, challenges such
as low resolution of small object, complex background interference, and
redundant hyperspectral data hinder the performance of existing tracking
methods. This study introduces a novel hyperspectral small object tracking
method based on target region perception (HTRP), which enhances tracking
accuracy and robustness. Firstly, a Band Grouping Module (BGM) is designed to
calculate band contributions, reorder and group hyperspectral bands, and then
generate multiple sets of false-color images in order to reduce band redundancy
and preserve discriminative spectral information. Secondly, a Target Region
Perception Module (TRPM) is proposed to generate a saliency map by enhancing the
resolution of the object area, thereby improving the network's ability to
perceive the positions of small object and producing region of interest to guide
tracking. Finally, numerous experiments show that the HTRP tracker maintains
superior performance on the HOT2022 dataset. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 03 (Paper No.6379) |
|
|
|
Title: |
SAR-HyperNet: Spectral-Aware Residual Hyperprior Network for Hyperspectral Image
Compression |
|
|
|
Author: |
Fahad Saeed1, Liu Shumin1, Jie Chen1 and Muhammad Salman Khan1 |
|
|
|
Abstract: |
Hyperspectral images (HSIs) provide rich spectral information for remote sensing
applications, but their high dimensionality creates substantial storage and
transmission burdens. Existing deep learning-based compression methods achieve
promising rate-distortion performance, yet they often insufficiently model
spectral dependencies and adaptive redundancy reduction. To address this issue,
we propose \textbf{Spectral-Aware Residual Hyperprior Network (SAR-HyperNet)}
for lossy HSI compression. SAR-HyperNet integrates \textit{Hybrid Residual
Blocks (HRBs)} and \textit{Spectral Attention Blocks (SABs)} into a
hyperprior-based variational autoencoder to improve feature propagation,
training stability, and spectral channel recalibration. This design enables
compact latent representations while preserving spatial structure and spectral
consistency. Experiments on public datasets show that SAR-HyperNet consistently
outperforms representative methods. It achieves higher PSNR and MS-SSIM with
lower SAM across various bitrates. Ablation results further verify the
complementary contributions of HRBs and SABs, demonstrating the effectiveness of
spectral-aware design for deep HSI compression. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 03 (Paper No.6380) |
|
|
|
Title: |
Spectrally Constrained Feature Matching for UAV-Borne Hyperspectral Images |
|
|
|
Author: |
Yuhe Liu1, Liu Shumin1, Yelin Liu2 and Jie Chen1 |
|
|
|
Abstract: |
Feature matching is essential for UAV-borne hyperspectral image registration,
but most existing matchers rely mainly on grayscale or spatial texture
information and underuse spectral signatures. This may lead to mismatches in
regions with similar spatial structures but different material properties. To
address this problem, this paper proposes a plug-and-play spectral consistency
filtering method for hyperspectral feature matching. Initial correspondences are
first obtained from grayscale representations using existing matchers, and local
spectral descriptors are then extracted from the original hyperspectral data
around matched keypoints. Spectral Angle Mapper (SAM) is used as the primary
consistency measure to reject spectrally inconsistent matches, while Spectral
Information Divergence (SID) is evaluated for comparison. The refined matches
are finally used for RANSAC-based homography estimation. Experiments on the
WHU-Hi-LongKou dataset show that the proposed spectral filtering strategy
effectively improves correspondence quality, reduces geometric errors, and
enhances the robustness of hyperspectral image matching. |
|
|
|
|
|
|
|
|
Oral Session: |
SPSS Session 03 (Paper No.6385) |
|
|
|
Title: |
Motion-Aware OSTrack: Latency Compensation for One-Stream Visual Tracking |
|
|
|
Author: |
Shiduo Zhang1, Cunle Zhang1, Chengkai Tang1, Baowang Lian1 and Dongjia Wang2 |
|
|
|
Abstract: |
Online visual tracking on unmanned aerial vehicle (UAV) platforms is sensitive
to inference latency, since the tracker output may correspond to an earlier
processed frame rather than the target state at output time. Although one-stream
Transformer trackers such as OSTrack achieve efficient target-oriented
representation by jointly modeling template and search features, they do not
explicitly compensate for latency-induced target displacement. This issue
becomes more challenging in UAV scenarios, where apparent motion is affected by
both target movement and platform-induced camera jitter. To address this
problem, we propose a jitter-aware latency compensation method built upon
OSTrack. The proposed framework keeps the original one-stream tracker unchanged
and introduces two lightweight modules: a camera jitter compensation branch that
estimates camera-induced shift from recent target states and produces stabilized
motion history, and a temporal motion predictor that estimates the target-center
displacement during the inference interval. The latest raw OSTrack output is
then refined through a residual center update to obtain a latency-compensated
prediction. Experiments are conducted on UAV123, UAVDT, and DTB70 under both
offline evaluation and a fixed-latency online evaluation setting. The results
demonstrate that the proposed method consistently improves online tracking
performance with only limited computational overhead, indicating its
effectiveness for latency-aware UAV tracking under challenging motion
conditions. |
|
|
|
|
|
|
|
|
SPGT Session 01 |
|
|
|
Time: 14:00~15:30, Saturday, July 18, 2026 |
|
|
|
Session Chair: Fan Zhang, Wuhan University |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 01 (Paper No.6305) |
|
|
|
Title: |
ARMANET: AUTO-REGRESSIVE-MOVING-AVERAGE INSPIRED NEURAL NETWORKS FOR SOURCE
SEPARATION USING MICROPHONE ARRAYS |
|
|
|
Author: |
Yucong Liu1, Chao Pan1, Zhuo Liu1, Jacob Benesty1 and Jingdong Chen1 |
|
|
|
Abstract: |
In sequential signal processing, moving-average (MA) and auto-regressive (AR)
operations are often cascaded construct efficient systems. The AR stage
emphasizes signals from specific frequency bands, while the MA block suppresses
interferences from others. Motivated by this principle, we introduce the
auto-regressive moving-average network (ARMAnet), a neural architecture that
integrates AR and MA functionalities for sequential signal processing. AR
processing is implemented via a recurrent neural network, specifically the
recently developed Mamba network, where as MA processing is realized through a
self-attention mechanism that computes weighted sums over the input sequence. To
assess its effectiveness, we combine ARMAnet with the recently proposed
SpatialNet and apply it to microphone array source separation. Experimental
results demonstrate the great potential of ARMAnet for this task. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 01 (Paper No.6330) |
|
|
|
Title: |
Predictive Step-size Selection Time-frequency-domain Hybrid Filtered-x
Normalized Least Mean Square Algorithm for ANC Systems |
|
|
|
Author: |
Zhiyuan Li1, Yi Yu1, Hongsen He1, Yuyu Zhu1 and Rodrigo C de Lamare1 |
|
|
|
Abstract: |
The filtered-x normalized least mean square (FxNLMS) algorithm is popular in
active noise control systems, but its computational burden increases with the
filter length. To enhance the computational efficiency, the frequency-domain
FxNLMS (FD-FxNLMS) algorithm was also proposed. However, it still suffers from a
trade-off between convergence rate and steady-state residual noise. Therefore,
to overcome this problem, this paper proposes a predictive step-size selection
strategy for the FD-FxNLMS algorithm based on a delayless time-frequency-domain
structure, which adaptively selects an appropriate step-size according to the
time-domain mean-square deviation recursion model. Simulation results
demonstrate that the proposed algorithm achieves faster convergence and lower
steady-state residual noise while preserving computational efficiency. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 01 (Paper No.6395) |
|
|
|
Title: |
On the Maximum Likelihood-Based Wiener Postfilter in the Presence of Unmodeled
Interference |
|
|
|
Author: |
Fan Zhang1, Kang Chen1 and Gongping Huang1 |
|
|
|
Abstract: |
The cascaded structure of minimum variance distortionless response (MVDR)
beamforming followed by single-channel Wiener postfiltering is widely used for
microphone array speech enhancement. In practical implementations, interference
sources are often not explicitly modeled due to the difficulty of direction
estimation, resulting in model mismatch whose impact on the Wiener postfilter
remains unclear.In this letter, we analyze the behavior of maximum likelihood
(ML)-based Wiener postfilter in the presence of an unmodeled interferer. We
first derive closed-form bias expressions for the ML estimators of the desired
signal and noise variances. Based on this result, we derive the expected value
of the estimated postfilter and analyze its monotonicity with respect to the
spatial correlation coefficient (SCC) between sources, input signal-to-noise
ratio (SNR), and interference-to-noise ratio (INR). Our analysis reveals that
the postfilter gain is a monotonic function of both SNR and SCC, while its
monotonicity with respect to INR depends on the SCC. Moreover, a condition on
the SCC is established under which effective interference suppression is
guaranteed. The theoretical findings are validated through numerical examples
and speech enhancement experiments. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 01 (Paper No.6411) |
|
|
|
Title: |
Cross-modal Speech Separation Based on Inter-modal and Intra-modal Consistency |
|
|
|
Author: |
Ke Lv1 and Ying Wei1,2 |
|
|
|
Abstract: |
In complex noisy environments, many studies utilized visual information to
assist speech separation task. However, audio and visual signals belong to two
different modalities and have inherent heterogeneity that is difficult to
mitigate. Directly applying visual information into speech separation may
introduce irrelevant or redundant features that degrade model performance.
Moreover, poor-quality visual information can significantly affect model
performance in real-world scenarios. We propose a cross-modal speech separation
method based on inter-modal and intra-modal consistency. It conducts a more
fine-grained pre-training of the inter-modal consistency feature extraction
network to mitigate modality heterogeneity. Using the prior knowledge learned
during pre-training, it imposes intra-modal consistency constraints among speech
signals to reduce reliance on visual quality. We evaluate the proposed method on
VoxCeleb2 dataset. The results show that the method outperforms state-of-the-art
methods on several evaluation metrics. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 01 (Paper No.6423) |
|
|
|
Title: |
Robust and Computationally Efficient WPE Dereverberation Using Forward and
Backward Beamformers |
|
|
|
Author: |
Kang Chen1, Yujie Zhu1, Gongping Huang1, Jingdong Chen1 and Jacob Benesty2 |
|
|
|
Abstract: |
Reverberation severely degrades speech quality and intelligibility. Although the
weighted prediction error (WPE) method has been widely adopted for speech
dereverberation, it remains sensitive to additive noise and entails high
computational complexity. To address these limitations, this paper presents a
robust and computationally efficient WPE framework based on forward and backward
beamformers. Specifically, a forward beamformer steered toward the target
direction is employed to produce a cleaner reference signal, while a backward
beamformer steered toward the opposite direction is introduced to estimate the
residual late reverberation remaining in the forward beamformer output. In this
way, the proposed framework improves dereverberation robustness in noisy
environments while substantially reducing the computational complexity.
Simulation results demonstrate the effectiveness of the proposed method. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 01 (Paper No.6445) |
|
|
|
Title: |
SUBBAND ADAPTIVE DIFFERENTIAL BEAMFORMING WITH FFT AND MEL FILTER BANKS |
|
|
|
Author: |
Li Du1, Fan Zhang2 and Lijun Zhang1 |
|
|
|
Abstract: |
Due to the limited order, differential beamforming has restricted
ability to suppress multiple noise sources, even when
these sources are spectrally non-overlapping. To address
this limitation, this paper presents subband extensions of our
recently developed adaptive differential (AD) beamforming
framework. Two types of filter banks are investigated: the
fast Fourier transform (FFT) and the Mel filter bank. For both
cases, we derive closed-form solutions for the subband AD
beamformer, with particular emphasis on the implementation
of the Mel-based design. Simulation results with two spectrally
distinct noise sources demonstrate that the proposed
approaches achieve superior noise reduction compared to
the original AD beamformer. Furthermore, the Mel-based
design provides a flexible tradeoff between noise reduction
performance and computational complexity. |
|
|
|
|
|
|
|
|
SPGT Session 02 |
|
|
|
Time: 16:00~17:30, Saturday, July 18, 2026 |
|
|
|
Session Chair: Long Shi, Southwestern University of Finance and Economics |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 02 (Paper No.6340) |
|
|
|
Title: |
Denoising Weak Dolphin Whistles Using a Scale-Normalized Improved Wavelet
Thresholding Function |
|
|
|
Author: |
Jiali Chen1, Zhenquan Hu1, Ru Wu2, Peibin Zhu1, Wen Chen1 and Yougan Chen3 |
|
|
|
Abstract: |
To mitigate the Pseudo-Gibbs phenomenon and the constant-bias effect in
conventional wavelet denoising, we propose a scale-normalized wavelet
thresholding function. The proposed method incorporates a scale-dependent
adjustment coefficient and a constant gain factor. It also utilizes the
Suppression Impulsive and Autocorrelation Function (SI-ACF) to enable adaptive
threshold optimization. Simulation experiments using measured background noise
from Xiamen Bay and chirp-like synthetic whistle signals show that, for input
signal-to-noise ratios from -15 dB to -5 dB, the proposed method effectively
suppresses local oscillations and mitigates amplitude distortion. Results show
that the proposed method improves the output signal-to-noise ratio by up to 0.22
dB and reduces the normalized root mean square error compared with conventional
methods. Additionally, the autocorrelation peak remains above 0.80. These
results demonstrate the method's effectiveness in high-fidelity reconstruction
and robust feature extraction within complex, low-SNR underwater environments. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 02 (Paper No.6419) |
|
|
|
Title: |
Source-Level Evaluation of Harmonic-Aware Fusion for Acoustic UAV Detection |
|
|
|
Author: |
Fuyu Tao1, Xu Yang1 and Qian Zhao1 |
|
|
|
Abstract: |
Random clip-level splits can make acoustic unmanned aerial vehicle (UAV)
detection look easier than it is, especially when many short clips come from the
same recording. In this paper we ask whether a simple harmonic-aware
representation is still useful after clips are grouped by inferred source
recording. The method is deliberately lightweight: local spectral contrast
enhances narrowband structures in the log spectrogram, and a second
convolutional neural network branch keeps the original spectrogram view. We test
entropy-based fusion and a quality-informed fusion variant on two public drone
audio sources containing drone, background, helicopter, and unknown non-drone
recordings. Under three-fold source-level cross-validation, entropy fusion gives
the best mean accuracy and precision, reaching 88.38 percent accuracy, while
quality-informed fusion gives the best recall and a slightly higher mean
F1-score of 85.00 percent. The results are not uniformly positive across
metrics. They suggest that the harmonic view is a useful auxiliary cue, but
helicopter-like periodic negatives can still make a harmonic branch confidently
wrong. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 02 (Paper No.6431) |
|
|
|
Title: |
A Multimodal Interface for Enhancing Communication Accessibility with
Bi-Directional Interaction Among Individuals with Hearing and Speech Impairments
Using Deep Learning |
|
|
|
Author: |
K SHILPA1 and Dr M. HEMALATHA1 |
|
|
|
Abstract: |
The problem of communication barriers lingers particularly in the case of the
hearing and speech impaired, especially in daily communication between the two
people where they do not both adopt a common communication channel. To overcome
this drawback, this paper suggests a multimodal two-way communication system
that incorporates the sign language recognition system, speech recognition
system, and text-based communication into one assistive interface. The system
takes visual input sign (captured by a camera with a resolution of 640?480
pixels) and interprets it through a deep neural network which is trained on
5,240 labeled gesture images of alphabetic hand configurations. The identified
characters are translated into written text as well as audio signals. Speech
recognition are then represented as a sequence of sign representations that are
stored in a repository of 1,320 GIF-based sign representations. Spoken input
received on a microphone with a sampling rate of 16,000 Hz is processed as a
sequence of speech recognition entries which, in turn, produce text and find
related sign representations in a repository. An experimental assessment of
4,860 trials of bidirectional interaction generated 4,417 accurate translation
of communication. It has shown to be a stable interface that allows real-time
communication with hearing and speech impaired people through the proposed
multimodal interface. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 02 (Paper No.6441) |
|
|
|
Title: |
Approach to Closed-Form DOA Estimation with Covariance Matrix Residual
Decomposition |
|
|
|
Author: |
Zhixin Ma1, Chao Pan1, Jing Zhao1 and Jingdong Chen1 |
|
|
|
Abstract: |
Direction-of-arrival (DOA) estimation using spherical microphone arrays is
sensitive to directional interference, which distorts the observation covariance
and degrades the performance of closed-form DOA estimators. This paper proposes
a robust closed-form DOA estimation method based on covariance matrix residual
decomposition. The observation covariance matrix is modeled as the sum of
target, interference, and diffuse noise. Given the spatial coherence priors of
interference and diffuse noise, the target covariance is recovered through
residual decomposition, while variances are estimated to construct a reliability
weight for further fusion. The recovered target covariance is incorporated into
a spherical harmonic domain closed-form DOA estimator without spatial spectrum
search. Simulation results demonstrate that the proposed method achieves
improved DOA estimation accuracy and robustness under directional interference. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 02 (Paper No.6444) |
|
|
|
Title: |
A Two-Stage Collaborative Sorting Framework for Radar Signals in Complex
Electromagnetic Environments |
|
|
|
Author: |
Liu Junming1, Li Tao1, Fan Yifei1, Su Jia1 and Liu Xiangyang1 |
|
|
|
Abstract: |
Abstract-To address the challenges of high-density interleaved pulse streams and
severe parameter overlap in radar signal pre-sorting, this paper proposes a
parameter-adaptive hybrid clustering method that integrates Robust Density
Kernel Estimation (RDKE) and Fuzzy C-Means (FCM). A two-stage "coarse-to-fine"
sorting framework is developed, in which RDKE models the pulse feature
distribution and adaptively determines both the number of clusters and initial
centers based on local sample density, overcoming the limitations of traditional
FCM. Subsequently, a fuzzy membership mechanism is introduced for fine sorting
of signals with ambiguous parameter boundaries. Simulations conducted in complex
scenarios with pulse loss and parameter overlapping demonstrate a sorting
accuracy of 96.8%, outperforming several typical clustering-based methods and
validating the effectiveness of the proposed approach. |
|
|
|
|
|
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 02 (Paper No.6471) |
|
|
|
Title: |
Convergence Analysis of a Robust Distributed Adaptive Strategy for Active Noise
Control |
|
|
|
Author: |
Sauravjyoti Senchowa1 and Bijit Kumar Das2 |
|
|
|
Abstract: |
An analysis of a robust diffusion strategy is addressed in this paper to achieve
improved performance in a distributed active noise control (ANC) system. The
system model is considered to consist of multiple microphones and multiple
loudspeakers. We define a robust diffusion strategy for distributed ANC based on
the exponential hyperbolic cosine cost function proven to be robust to both
Gaussian and non-Gaussian observation noise, referred to as the diffusion
filtered-x leaky exponential hyperbolic cosine adaptive filter (d-FxLEHCAF). The
mathematical analysis shows the convergence criteria of the algorithm. The
simulation results illustrate the efficiency of the algorithm in impulsive noise
environments. |
|
|
|
|
|
|
|
|
SPGT Session 03 |
|
|
|
Time: 09:00~10:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Mou Wang, Institute of Acoustics, Chinese Academy of Sciences |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 03 (Paper No.6355) |
|
|
|
Title: |
A Signal Acquisition Method Based on PMF-FFT and PCA-SVM for High-Dynamic and
Low-SNR Environments |
|
|
|
Author: |
Xinyao Wei1, Wenbin Sun1, Xin Yang1, Lili Chen1, Qian Xu1 and Ling Wang1 |
|
|
|
Abstract: |
This paper addresses the problems of difficulty in setting the acquisition
threshold, high miss rate, and degraded detection performance of the traditional
PMF-FFT algorithm in low-SNR and high-dynamic environments. A signal acquisition
method based on PMF-FFT and PCA-SVM is proposed. The proposed method retains
PMF-FFT as the front-end search framework, extracts statistical features from
the acquisition correlation results, employs principal component analysis (PCA)
for dimensionality reduction, and utilizes support vector machine (SVM) to
perform classification, thereby determining whether the signal has been
successfully acquired. Simulation results show that, compared with the
traditional threshold-based decision method, the proposed method achieves a
lower miss rate, a higher F1-score, and improved acquisition accuracy in low-SNR
and high-dynamic environments. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 03 (Paper No.6398) |
|
|
|
Title: |
Direct Localization of Coherent Sources with Moving Arrays Based on Cyclic
Cross-Correlation Tensor and Sparse Bayesian Learning |
|
|
|
Author: |
Yiquan Zhang1, Yuexian Wang1, Chuang Han1, Yandong Sun1 and Ling Wang1 |
|
|
|
Abstract: |
To address the challenges of limited localization accuracy and strong co-channel
interference in localizing coherent radiation sources within complex
electromagnetic environments, this paper proposes a direct position
determination algorithm combining cyclostationary features, joint Doppler
modeling, and sparse Bayesian learning. First, a spatio-temporal geometric model
for a moving array is established, which accurately maps the Doppler shift as a
nonlinear function of the target coordinates and the array velocity. Second, by
exploiting the cyclostationary characteristics of the target signals, the cyclic
cross-correlation matrix is calculated. This step not only suppresses stationary
noise and non-homogeneous interference but also constructs a multi-delay
spatio-temporal third-order tensor. Subsequently, the spatial subspace is
extracted via higher-order singular value decomposition, and the
forward-backward averaging technique is introduced to restore the rank of the
coherent signal subspace without sacrificing array aperture. Finally, a
hierarchical sparse Bayesian model is constructed to achieve high-precision,
high-resolution direct localization through adaptive hyperparameter updates.
Simulation results demonstrate that the proposed method exhibits excellent
localization robustness in environments characterized by low signal-to-noise
ratios and strong interference. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 03 (Paper No.6399) |
|
|
|
Title: |
Angle Estimation for IRS-Aided Bistatic EMVS-MIMO Array in the Presence of
Mutual Coupling |
|
|
|
Author: |
Zijian Luo1, Yuexian Wang1, Yandong Sun1, Chuang Han1 and Ling Wang1 |
|
|
|
Abstract: |
To address the problem of angle estimation in bistatic polarized MIMO arrays
assisted by Intelligent Reflecting Surfaces (IRS) under mutual coupling effects,
this paper proposes a Reduced-Dimension Rank-Reduction (RD-RARE) estimation
algorithm based on tensor decomposition. First, the factor matrices are
recovered using Tensor Train Decomposition (TTD). Next, the inter-mediate matrix
is reconstructed through a transformation in the angle domain and the properties
of the Kronecker product, enabling angle estimation using the recovered
imperfect array manifolds. The mutual coupling coefficients are then obtained
based on the properties of eigenvectors and eigenvalues. This method transforms
a two-dimensional search into two one-dimensional searches, significantly
reducing the computational complexity of the algorithm. Furthermore, simulation
results show that after self-correction of the mutual coupling coefficients, the
accuracy of the proposed algorithm is comparable to that of the self-corrected
Rank-Reduction (RARE) algorithm. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 03 (Paper No.6400) |
|
|
|
Title: |
Distributed Multi-Dimensional Interference Optimization Method Based on Game
Theory |
|
|
|
Author: |
Yuquan Shu1, Xunbin Zhou2, Fangli Tian1, Lingling Zhang3, Peilin Chen1 and
Chengkai Tang4 |
|
|
|
Abstract: |
As Low-Earth Orbit (LEO) satellite constellations are increasingly utilised in
key sectors such as military communications, navigation and positioning, and
intelligence and reconnaissance, their strategic importance is becoming
increasingly evident. To address the issues of low coordination efficiency,
insufficient resource utilisation and poor dynamic adaptability in traditional
distributed jamming methods for low-Earth orbit satellite communications, this
paper proposes a game-theory-based distributed multi-dimensional jamming
optimisation method. By treating each primary jammer as a non-cooperative game
participant and the allocation of relative jammers as a pure strategy, we
construct a composite utility function that integrates delay, power and Doppler
shift, and introduce a benefit-sharing coefficient to address resource-sharing
conflicts. We employ mixed-strategy games and the Newton iteration method to
solve for the Nash equilibrium, thereby obtaining the optimal allocation scheme
for relative jammers. Experimental results demonstrate that this method achieves
100% single-path coverage under multi-path interference conditions, with a bit
error rate (BER) consistently maintained at the 10?? level and an average
utility value of 3.40, significantly outperforming genetic algorithms, dynamic
programming and gradient descent algorithms. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 03 (Paper No.6413) |
|
|
|
Title: |
Noncircular_Source_Enumeration_via_Pseudo_Covariance_Lifting_and_Ritz-Gap_Detection |
|
|
|
Author: |
Chuang Han1, Feixiang Li1, Yandong Sun1 and Yuexian Wang1 |
|
|
|
Abstract: |
This paper addresses the problem of estimating the number of noncircular sources
from complex-valued array observations. Unlike conventional covariance-based
enumeration, extending the Lanczos/Ritz framework to pseudo-covariance is
nontrivial, since the pseudo-covariance matrix is complex symmetric rather than
Hermitian. To overcome this difficulty, a structured pseudo-covariance lifting
method is proposed. By exploiting the anti-diagonal structure of the
pseudo-covariance matrix under a uniform linear array, the matrix is first
converted into a smoothed one-dimensional sequence through anti-diagonal
averaging and then lifted into a Hankel matrix to form a Hermitian positive
semidefinite matrix. This transformation converts non-circular source
enumeration into a structured rank-determination problem compatible with Lanczos
tridiago-nalization and a practical logarithmic Ritz-gap detector. Under mild
conditions, the rank of the lifted matrix is shown to equal the number of
noncircular sources. Simulation results demonstrate that the proposed method
achieves high detection probability in mixed-signal scenarios with significantly
reduced runtime compared to full-decomposition-based alternatives. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 03 (Paper No.6428) |
|
|
|
Title: |
Direct Calibration of Sensor Position and Clock Errors Based on
Frequency-Weighted Cooperative Signals |
|
|
|
Author: |
XU Jun Nian1, Liu Xiang Yang1, Guo Zi Xun1, Fan Yi Fei1, Xie Jian1 and Tao Ming
Liang1 |
|
|
|
Abstract: |
To address the degradation in localization performance caused by receiver
position errors and clock synchronization biases in multi-station passive
localization systems, a direct calibration method based on cooperative signals
is proposed. Different from conventional two-step methods that first extract
time-difference-of-arrival measurements and then estimate the error parameters,
the proposed method directly exploits the received signals at all observation
stations. After matched filtering, the cross-spectrum between the reference
station and each secondary station is constructed. Based on the maximum
likelihood criterion, an objective function with respect to the receiver
position errors and clock biases is established, and the joint estimation of
these two types of error parameters is achieved by optimizing the objective
function. Simulation results show that the proposed direct calibration method
can effectively improve the calibration accuracy. Especially under low-SNR
conditions, compared with conventional two-step methods, the calibration
accuracy can be improved by more than 10 times. |
|
|
|
|
|
|
|
|
SPGT Session 04 |
|
|
|
Time: 11:00~12:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Mou Wang, Institute of Acoustics, Chinese Academy of Sciences |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 04 (Paper No.6319) |
|
|
|
Title: |
Parameter Optimization Design for Combined Frequency Modulation Signals |
|
|
|
Author: |
Bo You1, Baozhu Chen2, Fangyong Wang1 and Wenbo Ma1 |
|
|
|
Abstract: |
This paper derives the wideband ambiguity function ridge-line equations of
combined frequency modulation signals and their linear approximations under
multi-target conditions. A method for optimizing waveform parameters to avoid
spurious intersections is investigated and validated through simulations. Based
on an analysis of the anti-reverberation performance and time-delay resolution,
the signal combination is extended to non-equal-period N-pulse pairs. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 04 (Paper No.6339) |
|
|
|
Title: |
Joint Optimization of Waveform, Mismatched Filter, and Classifier for Radar
Target Recognition |
|
|
|
Author: |
Jiahang Wang1, Wentao Zhu1, Yifan Wu1, Tao Wang1 and Junli Liang1 |
|
|
|
Abstract: |
The matched illumination theory suggests that designing waveforms matched to
target scattering characteristics can actively enhance discriminative features
beneficial for classification from the returned echoes. However,
classification-driven waveforms often compromise the low-sidelobe properties
essential for pulse compression, creating a conflict between radar detection and
recognition. To address this, this paper proposes a joint
waveform-filter-classifier (JWFC) optimization framework that integrates the
transmit waveform, the mismatched filter (MMF), and a classifier into a
differentiable end-to-end architecture. Within this framework, the waveform acts
as a learnable feature encoder, jointly optimized with the classifier to align
with its inductive bias. Concurrently, an MMF is co-optimized to suppress range
sidelobes induced by the waveform, preserving detection performance. A
knowledge-guided two-stage training strategy is employed to ensure stable
convergence. Experiments on electromagnetic simulation data demonstrate that the
proposed JWFC framework achieves higher classification accuracy than traditional
waveforms across various code lengths, while maintaining peak and integrated
sidelobe levels better than -30 dB. The results confirm that the collaborative
"transmitter-receiver" design achieves functional compatibility between
recognition enhancement and detection requirements. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 04 (Paper No.6342) |
|
|
|
Title: |
Domino Group Sparsity Subarray Beampattern Synthesis |
|
|
|
Author: |
Yifan Wu1, Yubin Zhang1, Yitong Li1, Ziyi Liu1, Junli Liang1 and Jiahang Wang1 |
|
|
|
Abstract: |
Subarray level array beampattern synthesis is an active area of research. To
ensure beampattern performance while enabling the centralized arrangement of
subarrays to save antenna space, this paper proposes a domino-inspired sparse
pattern synthesis method. First, the centralized arrangement of subarrays is
equivalently formulated as a specially designed sparse optimization problem,
where the indices of subarrays are sorted according to their distances from the
center of the panel. Then, inspired by the domino toppling chain reaction, a
composite norm objective function is constructed. By imposing appropriate
constraints, the sparsity mechanism of "toppling the first domino" is triggered,
leading to an NP-hard optimization problem. The alternating direction method of
multipliers (ADMM) is employed to solve this problem. Theoretical analysis and
solution steps demonstrate that the proposed model can effectively force the
selected subarrays to concentrate near the panel center, achieving a domino-like
centralized sparse distribution. As a result, it achieves significant savings in
antenna layout space while maintaining the desired beampattern performance. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 04 (Paper No.6368) |
|
|
|
Title: |
A method of adaptive beamforming based on the cross correlation convex
combination optimization |
|
|
|
Author: |
Bingjie Yin1, Zhongyong Li2, Xi Zhang2 and Ji Xu2 |
|
|
|
Abstract: |
The performance of adaptive beamforming method is affected by the low
signal-to-noise ratio and signal steering vector mismatch. A method of adaptive
beamforming based on the cross correlation convex combination optimization is
proposed in this paper. This method exploits the influence of the conjugate
diagonal loading level to the signal to obtain the signal power of the different
mutual conjugate power vector in the closed range. Then, the maximum value of
the matrix is obtained, and the diagonal loading level corresponding to maximum
power is found. Finally, the robustness performance of the algorithm in the
complex environment is improved via the diagonal loading technique. This method
requires no user parameters, which can greatly save the operation cost in the
closed range, and has better anti-interference ability in the lost environment. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 04 (Paper No.6383) |
|
|
|
Title: |
A Fourth-Order Cumulant-Based Method for Joint Angle-Doppler Resolution
Enhancement |
|
|
|
Author: |
Mengyang Shui1, Yubin Zhang1, Yitong Li1, Qi Guo1 and Junli Liang1 |
|
|
|
Abstract: |
This paper proposes a fourth-order-cumulant-based (FOC-based) space--slow-time
virtual extension method to improve the resolution of joint angle--Doppler
estimation in radar systems. Different from conventional processing that
directly operates on the original array--pulse data, the proposed method employs
a one-step FOC operation to simultaneously synthesize virtual spatial and
slow-time degrees-of-freedom. By introducing the reference transmit waveform
into the cumulant computation, the proposed method incorporates
matched-filter-consistent waveform matching and enables joint array--pulse phase
combination. The resulting cumulant outputs are rearranged into a structured 2D
virtual angle--Doppler matrix with enlarged effective apertures in both
dimensions. Under the assumption that targets are separable in the range domain,
the proposed method yields a separable virtual manifold for angle--Doppler
representation. Therefore, additional structured degrees-of-freedom can be
obtained without increasing the number of physical sensors or transmitted
pulses. Simulation results show that the proposed method improves joint
angle--Doppler resolution compared with conventional processing on the original
data. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 04 (Paper No.6412) |
|
|
|
Title: |
Surrogate-Assisted Multi-Objective Optimization of Distributed Arrays Based on
Incremental Stratified Sampling |
|
|
|
Author: |
Chuang Han1, Chi Cheng1, Dong Yan Sun1 and Xian Yue Wang1 |
|
|
|
Abstract: |
Distributed array geometry optimization involves high-dimensional non-convex
search, multi-objective performance trade-offs, and expensive sidelobe
evaluation. To address these challenges, this paper proposes a
surrogate-assisted multi-objective optimization method based on incremental
stratified sampling. To reduce the cost of generating training data for the
surrogate model, an incremental data generation strategy is developed that
combines stratified uniform sampling with local subarray perturbations. This
approach effectively limits the number of expensive radiation pattern
evaluations while maintaining uniform coverage of the sample space. A gradient
boosting decision tree (GBDT) surrogate model is then constructed to approximate
the peak sidelobe level and integrated into the multi-objective optimization
framework to accelerate the search process. Simulation results demonstrate that
the proposed method significantly reduces computational time while achieving
optimization performance comparable to that of the high-fidelity pattern model,
validating its effectiveness for distributed array optimization problems. |
|
|
|
|
|
|
|
|
SPGT Session 05 |
|
|
|
Time: 14:00~15:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Xuehan Wang, Guangzhou University |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 05 (Paper No.6384) |
|
|
|
Title: |
Interleaved Radar Pulse Sorting via Physics-Guided PA/DTOA Dual-Task Modeling |
|
|
|
Author: |
Huaifeng Tang1, Weixuan Wang1, Hong Luo1, Ming Zhang1, Taoran Qi1 and Junfeng
Qiu1 |
|
|
|
Abstract: |
In dense electromagnetic scenes, intercepted radar pulses rarely appear as clean
single-emitter sequences. Pulses from different emitters are mixed along the
time axis, while jitter, missing pulses, spurious pulses, and measurement errors
further weaken the regular PRI structure. This paper treats radar pulse sorting
as two physics-related recognition problems instead of one general PDW
classification task. Pulse amplitude (PA) is used to describe envelope activity
and beam-related energy changes, whereas difference of time of arrival (DTOA) is
used to describe PRI timing rhythm. Based on this separation, PA-envelope
classes and DTOA/PRI rhythm classes are defined according to radar operating
mechanisms. |
|
|
|
|
|
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 05 (Paper No.6408) |
|
|
|
Title: |
Enhancing Ant-Jamming Perfomance Against DRFM Jammers in Mobile Surveillance
Radars Using Chebyshev Chaotic Encryption-Based Pulse Authentication |
|
|
|
Author: |
Denis Betwell Temba1, Qingwei Meng1 and Wang Han1 |
|
|
|
Abstract: |
Mobile surveillance radars working in contested environment are increasingly
vulnerable to Digital Radio Frequency Memory (DRFM) threats, which produce
coherent deceptive signal by presenting a false target or conceal a real one.
Traditional Anti-jamming techniques are insufficient because DRFM systems can
replicate and retransmit them, leaving the radar unable to distinguish from
authorized echoes and deception echoes which are copied. This paper proposes
Chebyshev chaotic encryption based-pulse Authentication technique to enhance
Radar Anti-jamming performance against DRFM deception. The method exploits
Chebyshev polynomial chaotic map which offers strong ergodicity, sensitivity to
initial condition and excellence correlation properties, we generated dynamic
pseudo random phase code and embedded it on each transmitted pulse. At the
receiver, synchronization using a regenerated expected chaotic code and verifies
pulse authenticity through high correlation with genuine echoes while rejecting
DRFM retransmitted signal. Simulation results in mobile radar scenario show
improved probability of detection under deception jamming reduced false alarms
from repeater attack over 85% to below 5% and rejects over 95% of deceptive
pulse. Compared to non-chaotic and other map alternatives, Chebyshev approach
offers superior autocorrelations, large space and robust performance suitable
for mobile platform including Unmanned Aerial Vehicle (UAV) and ground based
radars. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 05 (Paper No.6429) |
|
|
|
Title: |
Distributed Gated Generalized Likelihood Ratio Detection Fusion for Underwater
Active Detection |
|
|
|
Author: |
Jinhu Zhao1, Xi'an Feng2 and Lu Qiao3 |
|
|
|
Abstract: |
Underwater active detection is critical for long-range surveillance, yet
single-sensor performance is often degraded by propagation loss, noise
uncertainty, multipath, and target scattering fluctuation. Distributed detection
fusion mitigates these effects by combining information from spatially separated
sensors. Classical hard-decision fusion has low communication requirements, but
its optimal sensor weights depend on detection probabilities that are determined
by unknown target-present signal amplitudes, making weight specification
difficult for underwater active echoes. To remove this dependence, this paper
proposes a gated generalized likelihood ratio test (GLRT) detection-fusion
criterion for unknown single-sensor normalized signal amplitudes. Under a
normalized Gaussian model, the GLRT estimates the unknown signal amplitude from
the current observation, yielding a statistic in which stronger
target-consistent observations contribute more without requiring preassigned
per-sensor signal-to-noise ratios. Each sensor gate is set through a prescribed
sensor false-alarm probability, equivalently a noise-only transmission
probability, so that only sensors whose statistics exceed the resulting
threshold transmit real-valued contributions to the fusion center. The mixed
distribution of the gated statistic is derived in closed form, and the global
distribution is computed by fast Fourier transform based convolution for
Neyman-Pearson threshold design under a false-alarm constraint. Simulations show
accurate global false-alarm control and demonstrate that, at a detection
probability of 0.85, the proposed method reduces the required input
signal-to-noise ratio by about one decibel and increases the detection-range
gain factor from 1.19 to 1.26 relative to hard-decision fusion, without using
preassigned sensor signal-to-noise ratios in fusion weighting. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 05 (Paper No.6437) |
|
|
|
Title: |
Distributed MVI-CFAR detection method based on fuzzy logic |
|
|
|
Author: |
Yujun Hou1 and Qunfei Zhang1 |
|
|
|
Abstract: |
To address the limitations of traditional CA-CFAR detection, which performs
poorly under multi-target interference and clutter edge backgrounds, a
multi-strategy CFAR detection method based on fuzzy logic with adaptive
threshold selection is proposed. This method incorporates fuzzy logic and
combines the advantages of the CA, GO, and ACCA detection algorithms, enabling
the selection of the most appropriate detection strategy by effectively
discriminating the current background. Building on this, a distributed CFAR
detection system employing this method as the local detector is further
developed, with the fusion center utilizing three fusion rules: fuzzy algebraic
product, fuzzy maximum, and fuzzy minimum. Simulation results demonstrate that,
compared to existing methods, the proposed approach achieves superior detection
performance in homogeneous environments, multi-target interference scenarios,
and clutter edge backgrounds. Among the three fusion rules, the distributed
system based on the fuzzy algebraic product fusion rule delivers the best
detection performance. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 05 (Paper No.6469) |
|
|
|
Title: |
Method of Helicopter Target Detection for Anti-rotor False Flicker Jamming |
|
|
|
Author: |
Jie Jiang1, Junqiang Wu1, Shuangshuang Li1, Xuezhou Zhao1, Shuwen Wang1 and
Zhenkai Wu1 |
|
|
|
Abstract: |
Abstract-The detection and identification of helicopter rotor signals has always
been an important research issue in the field of low altitude defense. In this
paper, the method of helicopter target detection is proposed, and the target
segmentation, data smoothing, and quadratic clustering algorithm are used to
detect the target of the suspension helicopter in the complex clutter
background. The experimental results show that the method has greatly improved
the accuracy of the target detection. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 05 (Paper No.6483) |
|
|
|
Title: |
DOA estimation via acoustic vector sensor array under time-varying element
position errors |
|
|
|
Author: |
Yu Chen1, Xinshi Zhang1, Weidong Wang1, Hui Li1, Lingling Zhang2 and Wentao Shi2 |
|
|
|
Abstract: |
To mitigate the degradation in direction of arrival (DOA) estimation performance
caused by time-varying element position errors in an acoustic vector sensor
array (AVSA), a joint calibration iterative minimization algorithm (JCIMA) is
proposed in this paper. Initially, a correction source is introduced, and
eigenvalue decomposition is applied to derive an initial correction matrix. The
corrected received data are subsequently utilized to provide a preliminary
estimate of the signal azimuth. Following this, a cost function is constructed
based on both the corrected received data and the ideal array signal derived
from the estimated azimuth. Through iterative updates of the correction matrix
and the DOA aimed at minimizing this cost function, the algorithm achieves an
accurate DOA estimation. Simulation results indicate that, in comparison to
existing methods, the proposed JCIMA algorithm significantly enhances the DOA
estimation performance under the time-varying element position errors in AVSA. |
|
|
|
|
|
|
|
|
SPGT Session 06 |
|
|
|
Time: 16:00~17:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Xuehan Wang, Guangzhou University |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 06 (Paper No.6350) |
|
|
|
Title: |
Multi-Layer Coupled Transition Network based on SSA and Inter-Layer Association
Representation for Underwater Weak Target Signal Detection |
|
|
|
Author: |
Zihang Xue1, Haiyan Wang1, Bo Geng1 and Xiaohong Shen1 |
|
|
|
Abstract: |
Complex network-based nonlinear time series analysis can directly reveal
intrinsic target characteristics, thereby enabling underwater target detection
without prior target information. However, most existing complex network-based
methods are primarily designed for ideal noise-free time series and cannot
effectively characterize target signals in complex ocean environments. Regarding
this issue, we propose a multi-layer coupled transition network for weak target
detection based on singular spectrum analysis and inter-layer association
representation (SSA-ILAR-MLCTN). First, a progressive residual embedding matrix
is constructed through singular spectrum analysis to enhance the representation
of subtle dynamic differences under low signal-to-noise ratio (SNR) conditions.
Next, multi-channel time series are mapped into a multi-layer coupled network
through an inter-layer association representation, and a multi-layer weighted
cross clustering coefficient entropy is proposed to distinguish the noisy target
signal from the background noise. Comparative experiments were conducted using
two types of data: simulated Chen chaotic signals under different SNR conditions
and measured ocean trial data at different ranges. The results show that the
proposed method is superior to the traditional transition network method and the
narrowband cross spectrum detection method. The proposed method provides a new
approach to weak target detection using vector hydrophones when no prior target
information is available. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 06 (Paper No.6391) |
|
|
|
Title: |
Space-Time Division Cluster-Based DOA Estimation for Underwater Acoustic
Spread-Spectrum Signals |
|
|
|
Author: |
Biao Kuang1, Yibo Ma2, Yuxi Liu3 and Feng Zhou4 |
|
|
|
Abstract: |
Aiming at the problems of insufficient direction-of-arrival (DOA) estimation
accuracy, dependence on prior knowledge of source number, and poor wideband
direction-finding performance of traditional algorithms for underwater acoustic
spread-spectrum communication signals in multipath, noise and interference
environments, this paper conducts a systematic study on the space-time division
cluster (STDC) DOA estimation algorithm. The algorithm integrates spatial
multi-beamforming and time-domain matched filtering, directly utilizes the
excellent autocorrelation characteristics of spread-spectrum signals, and can
realize joint estimation of incident angle and propagation delay of multipath
signals without knowing the number of sources in advance, which can effectively
suppress noise interference and improve coherent multipath resolution.
Simulation experiments are carried out to compare the proposed algorithm with
spatial smoothing MUSIC, generalized almost-cyclostationary MUSIC and the
space-time division cluster algorithm. The results show that the STDC algorithm
has higher direction-finding accuracy and robustness under low signal-to-noise
ratio, strong interference and dense multipath conditions, and is naturally
compatible with the underwater acoustic spread-spectrum communication system.
The research results can provide a reliable direction-finding method and
theoretical support for array signal processing of underwater acoustic
spread-spectrum communication systems. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 06 (Paper No.6396) |
|
|
|
Title: |
Optimal Detection Range Analysis of a Rectangular Parametric Acoustic Array |
|
|
|
Author: |
Peihong Wang1, Jianjun Zhu1, Zihao Shu1, Chuan Liu1, Gouqing Zhao1 and ?.?.
Tarasov2 |
|
|
|
Abstract: |
Parametric acoustic array can generate low-frequency narrow beams with a compact
aperture, but the difference-frequency sound pressure level and beamwidth
exhibit distinct range-dependent behaviors, making the detection performance
non-monotonic along the propagation axis. This paper establishes a numerical
model for a rectangular-aperture parametric array based on the KZK equation and
calculate the axial evolution of the difference-frequency sound pressure level
and ?3 dB beamwidth. A figure of merit (FoM) combining signal strength and beam
footprint area is proposed to quantify the optimal detection range, defined as
the range maximizing FoM, with the effective detection zone bounded by a 3 dB
drop. Simulation results show that the FoM curve exhibits a clear single-peak
structure, with the optimum located near both the sound pressure level peak and
the beamwidth convergence region. The proposed method provides a quantitative
basis for working-range selection in underwater target detection and other
applications such as sub-bottom profiling. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 06 (Paper No.6417) |
|
|
|
Title: |
Acoustic Signal Characteristic Analysis of Marine Targets Based on Distributed
Acoustic Sensing |
|
|
|
Author: |
Wenlong Wei1, Haitao Wang1 and Xiangyang Zeng1 |
|
|
|
Abstract: |
This paper investigates the application of Distributed Acoustic Sensing (DAS)
for wide-area marine target identification. Leveraging the Ocean Observatories
Initiative (OOI) deep-sea dataset, we effectively suppress common-mode system
noise and random noise through two-dimensional (2D) median filtering and
bi-directional median subtraction. Our research demonstrates that distinct
target characteristics can be identified using spatio-temporal maps and
frequency-wavenumber (f-k) spectra. Specifically, whale vocalizations manifest
as clear V-shaped trajectories, whereas vessel signals present as broadband
energy shadows characterized by distinctive displacement patterns. Furthermore,
f-k spectral analysis confirms that differences in axial phase velocity,
represented by the slope of energy distribution, serve as the critical physical
basis for sound source identification. These results highlight the significant
advantages of DAS in wide-area acoustic field reconstruction, suggesting that
its complementary integration with traditional hydrophone arrays will provide
robust support for digitized marine monitoring. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 06 (Paper No.6426) |
|
|
|
Title: |
VMD-Assisted Adaptive Frequency-Band Partitioning and Cross-Band Attention
Fusion for Underwater Acoustic Target Recognition |
|
|
|
Author: |
Lin Zhiyang1, Wang Depeng2 and L? Yaohui1 |
|
|
|
Abstract: |
To fully exploit the multi-band discriminative information of ship-radiated
noise and improve the recognition accuracy and noise robustness of underwater
acoustic targets, this paper proposes a VMD-assisted adaptive frequency-band
partitioning and cross-band attention fusion method. The proposed method employs
variational mode decomposition to obtain the center frequency and energy
proportion of each modal component, and then uses a clustering algorithm to
adaptively divide the modal components into low-, middle-, and high-frequency
sub-bands. According to the time-frequency structural differences among
different frequency bands, Mel spectrograms with different window lengths are
constructed as multi-resolution inputs. Furthermore, a Dual-path CBA Network is
designed, in which the main path extracts global multi-band features, while the
CBA auxiliary path models the correlations among different frequency-band
features, thereby achieving multi-band information fusion. Experimental results
on the DeepShip dataset show that the proposed method achieves a recognition
accuracy of 99.30%, outperforming the compared methods. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 06 (Paper No.6438) |
|
|
|
Title: |
A Multi-Frame 2D Autoregressive Prewhitening Method with Adaptive Order
Selection for Reverberation Suppression |
|
|
|
Author: |
Yuan Cao1, Tianjun Zhou1 and Qunfei Zhang2 |
|
|
|
Abstract: |
Detection performance in active sonar is severely degraded by spatially and
Doppler-spread reverberation under limited sample support. To address this
issue, a multi-frame two-dimensional autoregressive (2DAR) prewhitening approach
is proposed. 2DAR parameters are jointly estimated across adjacent data frames.
In addition, a multi-frame averaging-based model order selection scheme is
developed, whose advantage lies in improved robustness under limited snapshots.
Based on this framework, a generalized likelihood ratio test detector is
derived. Simulation results demonstrate that the method achieves superior
whitening and improves detection performance compared with conventional 1DAR and
single-frame 2DAR approaches, particularly in low signal-to-reverberation ratio
conditions. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 06 (Paper No.6466) |
|
|
|
Title: |
Acoustic-Band Guided Multi-Feature Fusion for Underwater Acoustic Target
Recognition |
|
|
|
Author: |
Weijie Ning1, Zhe Jiang1, Bingbing Zheng1 and Bo Geng1 |
|
|
|
Abstract: |
Underwater acoustic target recognition (UATR) relies on discriminative acoustic
structures in target-radiated noise, such as narrowband tonal components,
spectral envelopes, and broadband energy distributions. However, these acoustic
patterns are unevenly distributed across frequency bands and can vary with
target types and observation conditions. Existing multi-feature methods usually
stack different acoustic features in a fixed order, which may cause the model to
over-rely on a specific feature position and underuse complementary acoustic
information. To address this problem, this paper proposes an acoustic-band
guided multi-feature fusion method. Acoustic Band Adaptive Weighting (ABAW)
adaptively enhances informative acoustic bands and suppresses noise-dominated
bands according to band content and band position. Feature De-correlation Mixing
(FDM) further reduces fixed feature-order dependence in a shared latent feature
space through random feature permutation and cross-feature local exchange.
Experiments on the ShipsEar and DeepShip datasets show that the proposed method
achieves recognition accuracies of 96.3% and 88.5%, respectively. Ablation
studies and qualitative analyses further verify the effectiveness of ABAW and
FDM. |
|
|
|
|
|
|
|
|
SPGT Session 07 |
|
|
|
Time: 08:30~10:00, Monday, July 20, 2026 |
|
|
|
Session Chair: Fan Zhang, Wuhan University |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 07 (Paper No.6335) |
|
|
|
Title: |
IoT-Based Multimodal Early Stroke Detection System Using Edge Computing |
|
|
|
Author: |
Pedamala Bharath1 |
|
|
|
Abstract: |
Stroke is a significant cause of death and long term disability in most sections
of the world with timely response within the golden hour being a game changer on
patient survival. However, the standard diagnostics such as the computed
tomography (CT) and magnetic resonance imaging (MRI) is reactive,
infrastructural and cannot be used in nonclinical environments where continuous
monitoring is needed. The present paper suggests an IoT-based, multimodal, early
stroke detection platform, which functions on the Raspberry Pi edge computing
platform. The proposed system is a combination of gait analysis, facial
asymmetry detection, and speech impairment analysis to monitor the early
neurological symptoms. A camera, microphone, and inertial measurement unit (IMU)
are used to capture the data and are then processed locally with lightweight
machine learning models that do not need a high latency, which ensures data
privacy and provides them with the ability to work offline. The gait, facial,
and speech features are fused on a feature-level to provide a composite stroke
risk score. The proposed multimodal approach, according to the results of the
experiments, has accuracy of 93% t and it is superior to single-modality
systems. It is not invasive and is affordable, and may be applicable in
home-based, resource constrained health care settings. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 07 (Paper No.6348) |
|
|
|
Title: |
DNCDSEM：A novel detection network with color dithering,
squeeze and excitation modules |
|
|
|
Author: |
Yilong Niu1, Xiao Yuan1, Tianhao Shi1, Qi Wang1, Yingmin Wang1 and Yi Wang1 |
|
|
|
Abstract: |
Object detection in remote sensing images aims to identify and localize targets
accurately. With the rapid development of deep learning, numerous detection
networks have been proposed for this task. Although these methods have achieved
satisfactory results, some useful information may still be irretrievably lost
during feature extraction, which can lead to false positives and false
negatives. To address these issues, this paper proposes a novel Detection
Network with Color Dithering and Squeeze-and-Excitation Modules (DNCDSEM) based
on Real-Time Models for Object Detection (RTMDet). Specifically, the color
dithering module is introduced into the data augmentation stage, while the
squeeze-and-excitation module is integrated into the neck to enhance feature
representation. These two modules enable the network to capture more
discriminative and fine-grained feature information, thereby improving its
feature extraction capability and detection performance. Extensive qualitative
and quantitative experiments demonstrate the effectiveness of the proposed
network and verify that it achieves promising performance. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 07 (Paper No.6357) |
|
|
|
Title: |
Robust Surround-View Extrinsic Calibration for Water-Surface Platforms via
ArUco-Constrained Bundle Adjustment |
|
|
|
Author: |
Changsong Pang1, Xiaomin Zhang1 and Yang Yu1 |
|
|
|
Abstract: |
Accurate multi-camera extrinsic calibration is essential for surround-view
perception in intelligent waterborne systems, yet remains challenging in
water-surface environments due to weak textures, strong reflections, dynamic
illumination, and vessel motion. To address this problem, this paper proposes an
ArUco bundle adjustment framework for surround-view extrinsic calibration on
marine platforms. Four floating ArUco boards are placed around the vessel as
shared geometric references, and a global optimization problem is constructed to
jointly refine camera extrinsics and board poses. Besides the reprojection term,
a board-coordinate prior constraint and a coplanarity constraint are introduced
to improve robustness and enforce the physical structure of the water-surface
deployment. Experiments on a real platform with four fisheye cameras show that
the proposed method yields geometrically consistent stitched top-view results in
representative near-shore scenes and remains stable across different vessel
poses and surrounding layouts. The proposed framework provides an effective and
practical solution for surround-view calibration in realistic maritime
environments. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 07 (Paper No.6361) |
|
|
|
Title: |
Al-Driven Innovations to Enhance Tea Production and Quality in Sri Lanka's Mid
Country |
|
|
|
Author: |
Nipuna Abeysekara1, Dinithi Wijerathne1, Harsha Waduthanthri1 and Nimesh
Bandaranayake1 |
|
|
|
Abstract: |
The Sri Lankan tea industry faces several operational challenges, including
inconsistent leaf grading, inefficient yield forecasting, unpredictable auction
demand, and delayed disease/pest detection. This research presents a novel
multi-modal AI framework aimed at addressing these challenges by automating key
decision-making processes. The system leverages deep learning techniques for
four tasks: (1) freshness grading and tea-type classification from
mobilecaptured leaf images, (2) yield forecasting using estate records and
scanned logbooks, (3) auction demand prediction based on historical transaction
data, and (4) disease/pest detection from images with seasonal risk prediction
linked to climate data. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 07 (Paper No.6452) |
|
|
|
Title: |
Low-Light Object Detection via Feature Refinement and Semantic Enhancement |
|
|
|
Author: |
Baoguo Wei1, Xinyu Wang1 and Xu Li1 |
|
|
|
Abstract: |
Object detection in low-light conditions remains a challenging problem due to
the loss of fine details and global semantics caused by poor illumination. In
this paper, we analyze two key factors limiting detection performance: the
scarcity of local detail in shallow feature maps and the degradation of global
semantic information in deep feature maps. To address these issues, we propose a
unified detection framework with two plug-and-play modules: a Feature Refinement
Encoder (FRE) and a Semantic Enhancement Module (SEM). The FRE bridges adjacent
feature maps through a cross-feature attention mechanism, compensating for
missing edge and texture information in early stages. The SEM employs grouped
dilated convolutions with adaptive pooling to expand the receptive field,
thereby recovering contextual cues lost due to dark occlusion. We evaluate our
method on the ExDark dataset. Without any illumination priors or pretrained
weights, our approach improves mAP by 5.8% over the YOLOX baseline,
demonstrating the complementary benefits of synergizing local detail restoration
and global semantic aggregation. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 07 (Paper No.6454) |
|
|
|
Title: |
Multi-temporal Frame Difference Fusion and EfficientNet Ensemble Model for Man
Overboard Action Recognition |
|
|
|
Author: |
Kaixuan Tian1 |
|
|
|
Abstract: |
In complex marine scenarios, man overboard action recognition faces challenges
such as insufficient motion feature representation, missed detection and false
alarms caused by imbalanced sample categories. This paper proposes a man
overboard action recognition method combining multi-temporal frame difference
fusion and EfficientNet ensemble model. Firstly, YOLOv8 is adopted to detect and
track personnel in surveillance videos, and continuous image sequences
containing floating, struggling and swimming actions are obtained by cropping
detection bounding boxes. Then, four temporal intervals of 3, 6, 9 and 12 are
selected to extract multi-scale motion features via inter-frame difference.
After eliminating the redundant model with poor performance, the three optimal
frame difference features are mapped to RGB three channels to construct fused
motion feature maps. Subsequently, multiple EfficientNet-B4 sub-models are
trained respectively, and a weighted voting ensemble strategy is applied to
complete sequence-level action classification. Experimental results show that
the proposed fusion method greatly improves the recall rate of struggling
actions to 84.44%, effectively reduces false alarms of swimming recognition. The
overall accuracy exceeds 90% with an obvious improvement of average F1-score.
The proposed method has strong robustness in complex marine environments, and it
can provide technical support for intelligent maritime search and rescue. |
|
|
|
|
|
|
|
|
SPGT Session 08 |
|
|
|
Time: 10:30~12:00, Monday, July 20, 2026 |
|
|
|
Session Chair: Qian Wang, North Electro-Optic Co., Ltd. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 08 (Paper No.6334) |
|
|
|
Title: |
Fine-Grained Radio Frequency Fingerprint Open-Set Recognition using
Discriminative Loss |
|
|
|
Author: |
Zitai Liu1, Yulan Zhang1, Weishi Chen2, Jun Hu1 and Zengping Chen1 |
|
|
|
Abstract: |
Wireless devices recognition based on radio frequency (RF) fingerprint is
promising for physical layer authentication. Many existing methods mostly adhere
to the closed-set assumption, and may be unsuitable for complex and dynamic
electromagnetic environments. In this letter, we propose a deep learning based
open-set recognition method to extract fine-grained RF fingerprints, and design
a discriminative loss to reduce the intra-class variation and increase the
inter-class variation to expand the decision boundary. Experimental results on
real-world RF datasets demonstrate that our proposed method can significantly
improve the open-set recognition performance of RF fingerprints, while the
proposed loss function does not introduce additional parameters. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 08 (Paper No.6406) |
|
|
|
Title: |
Amplitude-Aware Dual Encoding in Ordinal Partition Transition Networks for
Robust Signal Classification |
|
|
|
Author: |
Bo Geng1, Haiyan Wang1, Gaoyue Ma1, Weijie Ning1 and Xiaohong Shen1 |
|
|
|
Abstract: |
Ordinal partition transition networks (OPTNs) map continuous time series into
discrete topological graphs but fundamentally discard relative amplitude
information. This structural blindness inherently causes state degeneracy, where
distinct local fluctuations are mapped to identical ordinal patterns, severely
limiting their discriminative capacity for complex signals. In this paper, we
propose Amplitude-Aware Ordinal Partition Transition Networks (AOPTNs), a novel
dual-encoding framework that systematically integrates quantized relative
amplitude differences into the topological permutation process. By enriching the
state space symbols, AOPTN seamlessly captures both the ordinal structure and
the local metric dynamics of the underlying attractor. This provides a richer
symbolic state representation than traditional methods, effectively resolving
structural blindness. Through extensive experiments on synthetic chaotic
systems, we demonstrate that AOPTN significantly enhances bifurcation detection
sensitivity and provides higher resolution in distinguishing dynamical regimes.
Furthermore, applying AOPTN to a highly non-stationary real-world 5-class
underwater acoustic dataset yields an exceptional peak classification accuracy
of 94.33%, substantially outperforming standard OPTN and permutation entropy
methods. Crucially, by unfolding entangled feature spaces and resolving state
degeneracy, AOPTN accelerates optimal hyperplane convergence in downstream
support vector machines. This enables the proposed method to boost
classification accuracy and stability without compromising computational
efficiency. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 08 (Paper No.6442) |
|
|
|
Title: |
VLM-NCD: Words as Anchors - Retrieval-Augmented Multimodal Discovery of Novel
Classes |
|
|
|
Author: |
Baoguo Wei1, Yuetong Su1, Xinyu Wang1, Xu Li1 and Lixin Li1 |
|
|
|
Abstract: |
Novel Class Discovery (NCD) aims to transfer knowledge from labeled known
classes to partition unlabeled data that may contain both known and novel
categories. Existing vision-only methods suffer from limited feature
discriminability and severe degradation under long-tailed distributions. In this
paper, we propose VLM-NCD, a multimodal framework that breaks through these
bottlenecks by fusing visual and textual semantics for prototype-guided
clustering. The core innovations include: (1) a joint representation that aligns
image features with retrieved text descriptions to construct semantic prototypes
for known classes, and (2) a dual-phase discovery mechanism that separates known
and novel samples via semantic affinity thresholds, followed by adaptive
semi-supervised k-means clustering. Experiments on CIFAR-100 and ImageNet-100
show that VLM-NCD improves novel-class accuracy by up to 25.3% over the state of
the art while exhibiting strong robustness to long-tailed data, a rarely
addressed issue in previous NCD literature. With only minimal fine-tuning of a
frozen CLIP backbone, our framework provides an efficient and effective solution
for open-world recognition. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 08 (Paper No.6450) |
|
|
|
Title: |
STAR-Beat: A Spatio-Temporal Attention Recurrent Framework for Lightweight and
Interpretable Atrial Fibrillation Detection using Wearable Photoplethysmography |
|
|
|
Author: |
Mengcheng Hu1, Jiarong Chen2, Guoxing Wang2 and Cheng Chen2 |
|
|
|
Abstract: |
Wearable Photoplethysmography (PPG) enables continuous, non-invasive, and
low-cost screening for real-time Atrial Fibrillation (AF) detection. However,
deploying robust AF detection on edge devices remains challenging. While
existing deep learning methods attempt to mitigate motion artifacts, they
fundamentally treat time-series as local spatial data, lacking the capability to
capture the long-range temporal rhythm dependencies crucial for identifying AF.
Furthermore, their severe over-parameterization and limited black-box
interpretability hinder real-world edge deployment and clinical trust. To
address these critical gaps, this paper proposes STAR-Beat, an
ultra-lightweight, temporally-aware, and highly interpretable multi-task AF
detection framework. Building upon a convolutional denoising autoencoder (CDAE)
for morphological prior extraction, STAR-Beat introduces Squeeze-and-Excitation
(SE) attention to adaptively suppress noise channels, and innovatively utilizes
a Bidirectional GRU (BiGRU) to explicitly capture long-range irregular rhythm
dependencies. Validated on the Stanford and MIMIC PERform datasets, STAR-Beat
achieves outstanding F1 scores of 0.962 and 0.946, respectively. Moreover,
extensive post-hoc interpretability analyses verify that the model's attention
accurately targets pathological hemodynamic characteristics, effectively
bridging the trust gap for clinical application. Finally, to overcome hardware
constraints, an edge-oriented structural ablation strategy compresses model
parameters to merely 8.6% of the baseline, reducing inference delay to an
ultra-low 1.885ms. The code of this article is available at
https://github.com/04-07-08/STAR-Beat. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 08 (Paper No.6455) |
|
|
|
Title: |
Fuzzy Measure Enhanced Unscented Kalman Filtering for Robust Range-Bearing
Maneuvering Target Tracking |
|
|
|
Author: |
Junkai Wang1, Yongsheng Yan2 and Zhuying Wang1 |
|
|
|
Abstract: |
This paper proposes a fuzzy-measure-enhanced unscented Kalman filter (FMUKF) for
robust range-bearing maneuvering target tracking. FMUKF embeds fuzzy innovation
reliability into model-conditioned unscented Kalman filtering, so unreliable
measurements are downweighted before they distort the posterior state estimate.
The implementation combines interacting model mixing, analytic prediction, an
unscented range-bearing update, fuzzy reliability evaluation, and
reliability-driven covariance inflation for glint-contaminated observations. A
fixed-rate interacting multiple model unscented Kalman filter (IMM-UKF) and a
posterior Cramer--Rao lower bound (PCRLB) provide the comparison baselines.
Monte Carlo results show that FMUKF maintains competitive Gaussian accuracy
while improving robustness under non-Gaussian glint noise. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 08 (Paper No.6464) |
|
|
|
Title: |
A Fast Variational Dirichlet Process Mixture Tracking Filter for Online
Multi-Target Tracking with Unknown Target Births |
|
|
|
Author: |
Sihang Zhang1, Yongsheng Yan2 and Wang Xiong1 |
|
|
|
Abstract: |
Multi-target tracking (MTT) requires joint estimation of the number and
kinematic states of targets from noisy measurements with unknown data
association. Generalized labeled multi-Bernoulli (GLMB) filtering provides a
principled random-finite-set solution, but its performance and computational
cost can be affected by the birth model when target-birth times and locations
are not available a priori. This paper presents a fast online variational
Dirichlet process mixture model (DPMM) tracking filter based on a truncated
stick-breaking representation and mean-field variational inference. At each
scan, existing tracks are represented by active mixture components, while
additional free components serve as a measurement-driven birth pool. The
resulting soft association probabilities are used for Kalman state update,
Bernoulli existence recursion, candidate birth confirmation, and track
management. The proposed method is compared with an adaptive-birth GLMB filter
under the same dynamic and measurement models. Simulation experiments show that
the variational DPMM tracking filter achieves competitive tracking accuracy and
improved cardinality estimation while substantially reducing computational cost. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 08 (Paper No.6470) |
|
|
|
Title: |
CIAIC_MED8: A Dual-Stage Multi-Modal Emotion Database |
|
|
|
Author: |
Qian wang1, Jun Zhu2, Mou Wang3, Jingdong Chen4, Yan Yang4 and Jieling Liu1 |
|
|
|
Abstract: |
This paper presents a newly developed multi-modal emotion database, namely CIAIC
Multi-modal Emotion Database (CIAIC_MED8).This database is designed to serve as
a standardized reference for affective computing research and will be released
to the public. It contains multi-modal emotional data from 100 participants
recorded in an anechoic chamber. Four target emotions, i.e., neutral, sad,angry
and happy are induced by carefully selected, evaluated and validated video
clips. Signals are captured in two successive stages: emotion perception while
watching clips, and emotion expression while reading scripts aligned with target
emotions. Synchronously recorded modalities include facial videos,
ElectroEncephalo-Graph(EEG), electrocardiogram(ECG), electrodermalactivity(EDA),
photo-plethysmography(PPG) and respiration(RSP). Inaddition, air-conduction and
bone-conduction speech are collected during the expression stage. Subjective
ratings of emotional experience and expression are annotated under both
categorical andvalence-arousal dimensional models.This database aims to support
fair evaluation and innovation in multi-modal emotion recognition research. |
|
|
|
|
|
|
|
|
SPGT Session 09 |
|
|
|
Time: 10:30~12:00, Monday, July 20, 2026 |
|
|
|
Session Chair: Yingwei Tian, Wuhan University |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 09 (Paper No.6390) |
|
|
|
Title: |
Denoising and Prediction of Time-Varying Underwater Acoustic Channels in the
Delay-Doppler Domain |
|
|
|
Author: |
Yuxin Liu1, Yudong Chen1 and Fujun Lin1 |
|
|
|
Abstract: |
The complex time-space-frequency variability of underwater acoustic (UWA)
channels poses a fundamental challenge to reliable underwater communications. To
address the severe channel state information (CSI) outdating caused by the low
speed of sound in adaptive modulation systems, this paper proposes a
comprehensive processing and prediction framework for UWA channels in the
delay-Doppler (DD) domain. First, measured channel impulse response (CIR) data
are amplitude-normalized and segmented via a dual-layer sliding window, then
transformed to the DD domain via fast Fourier transform (FFT); the 32 most
energetic multipath/Doppler components are retained for dimensionality
reduction. Second, a deep denoising convolutional neural network (DnCNN)
tailored to the highly sparse DD-domain representation is proposed,
incorporating per-sample RMS normalization, complex orthogonal decomposition,
and an L1+MSE hybrid loss to suppress background noise while preserving sparse
path peaks. Finally, a Multi-Scale Convolutional LSTM (MultiScale-ConvLSTM)
prediction network with a parallel multi-branch architecture captures channel
evolution across different temporal scales, enabling DD-domain channel
prediction and inverse mapping to the time domain. Experiments on the Trex04,
Unet06, and AUVFest07 datasets demonstrate that the denoising module reduces the
noise floor to approximately 41.7% of the noisy energy, and an output SNR of
8.4-10.2 dB; the prediction module converges to a training loss below 0.002,
validating the effectiveness of the proposed approach. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 09 (Paper No.6393) |
|
|
|
Title: |
Direction of Arrival Estimation Techniques for Vector Arrays Embedded in
Viscoelastic Materials |
|
|
|
Author: |
Pengfu Ding1, Yu Zhang1, Ke Li2, Chenglong Xia2 and Jun Fan1 |
|
|
|
Abstract: |
Vector sensors have incomparable advantages over traditional acoustic pressure
sensors in detection applications. The ideal working environment for vector
sensors is a free field, but they typically need to be installed on platforms,
operating under non-free field conditions. Therefore, many scholars both
domestically and internationally have successively begun researching vector
sensor applications under non-free field conditions. However, the non-free field
conditions studied here are all based on rigid materials, with very little
research involving the effects of other material properties on vector sensors.
Addressing the problem that vector sensors are affected by absorption and
scattering from viscoelastic materials, which causes phase shifts leading to
inaccurate azimuth estimation, this paper conducts research on robust azimuth
estimation for vector arrays under these conditions. This paper applies an
adaptive phase correction maximum likelihood azimuth estimation method for
vector arrays to adaptively compensate for the phase of vector sensor arrays
under viscoelastic material conditions, thereby achieving target azimuth
estimation. Through comparative analysis of simulation experiments and lake
trials, the excellent azimuth estimation performance of this method under
viscoelastic material conditions has been verified, laying a technical
foundation for enhancing the application of vector sensors in broader scenarios. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 09 (Paper No.6401) |
|
|
|
Title: |
Design and Analysis of a High-Frequency LLC Resonant DC-DC Converter |
|
|
|
Author: |
Ankita Ramanna Katapur1, Dr.Prashant V Joshi1 and Dr. Sudharshan K M1 |
|
|
|
Abstract: |
This paper discusses the design and analysis of an LLC resonant half-bridge
converter. It operates under both open-loop and closed-loop control for a wide
input voltage range of 370 to 430 volts, delivering a regulated 48 volts output
at 1 kW. First, the converter is analyzed in open-loop configuration to examine
the resonant tank characteristics, operating regions, and voltage gain using the
First Harmonic Approximation (FHA) method. The converter is set to operate in
the inductive region to achieve Zero-Voltage Switching (ZVS). This approach
reduces switching losses and improves efficiency. To maintain stable output
regulation, a closed-loop control scheme with a PI controller is used. In this
scheme, the output voltage is compared to a reference voltage, and the error
signal is processed to adjust the switching frequency. The resulting switching
frequency generates gate signals with a fixed 50% duty cycle for the half-bridge
switches. Simulation results show improved voltage regulation, reduced switching
losses, and stable operation across the wide input voltage range, confirming the
effectiveness of the proposed LLC resonant converter. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 09 (Paper No.6427) |
|
|
|
Title: |
Fusing Time Frequency Representations and Orbital Parameters via Self Attention
for LEO Satellite Interference Optimization |
|
|
|
Author: |
Chengkai Tang1, Aomi Chen2, Zesheng Dan2 and Yangyang Liu2 |
|
|
|
Abstract: |
Low?Earth orbit (LEO) satellites have become a cornerstone of integrated
space?ground communication networks, yet their high mobility, narrow beams and
rapid frequency hopping pose severe challenges to military communication
security. Traditional jamming methods struggle to maintain accurate frequency
tracking and rapid synchronization in such dynamic scenarios. To address this,
we propose an optimized interference method based on a multimodal learning
transformer model (OI?MLT). The framework fuses time?frequency spectrum images
with numerical signal?state parameters and introduces a composite positional
encoding that captures both global orbital periodicity and local fast
variations. A standard transformer encoder then models cross?modal dependencies
through its symmetric self?attention mechanism, enabling end?to?end prediction
of interference effectiveness indicators. Extensive experiments under simulated
LEO environments demonstrate that OI?MLT achieves frequency tracking within
±12?kHz in 97% of trials, a mean synchronization time of 0.85?ms, and an
interference success rate of 95%, representing an average improvement of 27.52%
over existing methods. The results validate the superiority of multimodal fusion
and transformer?based optimization for LEO satellite countermeasures. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 09 (Paper No.6453) |
|
|
|
Title: |
Design of an Analog Front-End using Differential Current Signal Technic for
Electric Air-Soft Gun Shooting Detection |
|
|
|
Author: |
Pattapong Sripho1 and Mr.Jedsada Kraikhow1 |
|
|
|
Abstract: |
This research presents the design of an analog front-end circuit using
differential current signal processing techniques for shooting detection in an
electric air-soft gun system. A Hall-effect current sensor was used to measure
the motor current waveform, while analog filtering and differential processing
were applied to extract automatic firing transient characteristics. |
|
|
|
|
|
|
|
|
Oral Session: |
SPGT Session 09 (Paper No.6457) |
|
|
|
Title: |
A New Method for Forward Scanning Imaging of Missile-Borne Radar |
|
|
|
Author: |
Yuanyuan Sha1, Liling Niu2, Di Wu*3, Daiyin Zhu3 and Xiangjun Xu3 |
|
|
|
Abstract: |
Missile-borne radar can obtain high-resolution images of wide-area scenes
through scanning and imaging signal processing. However, in the forward-looking
region, the resolution of Doppler beam sharpening (DBS) degrades rapidly because
of the reduced Doppler gradient. Array super-resolution provides higher spatial
resolution in the forward-looking region, but its performance in the squint
forward-looking region is inferior to that of DBS. To overcome these
limitations, this paper proposes a scanning imaging method of missile-borne
radar that improves the resolution of both the forward-looking and squint
forward-looking regions. By decomposing the echo signal, the high-resolution
imaging regions of DBS and array super-resolution are partitioned according to
the effective array aperture, and Doppler super-resolution is introduced to
further enhance DBS resolution. Simulation and measured results demonstrate that
the proposed method satisfies the requirements of forward-looking scanning
imaging and significantly improves image clarity in complex scenes. |
|
|
|
|
|
|
|
|
CPT Session 01 |
|
|
|
Time: 16:00~17:30, Saturday, July 18, 2026 |
|
|
|
Session Chair: Kaili Yin, Qingdao University of Science and Technology |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 01 (Paper No.6351) |
|
|
|
Title: |
Genetic Algorithm and Q-Learning Based Channel-Aware Cluster Routing Protocol
for Underwater Acoustic Networks |
|
|
|
Author: |
Yihao Zhao1, Zheyang Chen1, Ziyi Ding1, Shenao Tu2, Yougan Chen2 and Xiaomei Xu1 |
|
|
|
Abstract: |
To address the challenges of energy consumption constraints and channel
dynamicity in underwater acoustic networks (UANs), this paper proposes a Genetic
Algorithm and Q-Learning Based Channel-Aware Cluster Routing Protocol
(GA-QL-RP). The protocol establishes a hybrid cluster routing framework
featuring centralized decision-making and distributed adjustment: the Sink node
employs genetic algorithms to periodically perform global cluster head (CH)
selection, base on node residual energy, local link quality, and Sink link
quality; meanwhile, each node conducts online channel-state learning and dynamic
Q-value updates via stateless Q-learning, enabling distributed adaptive
fine-tuning of the clustering structure based on real-time Q-values. By
synergistically combining the global optimization capability of genetic
algorithms with the local dynamic adaptability of Q-learning, the protocol
effectively balances routing reliability and energy efficiency. Simulation
results demonstrate that, compared with the LEACH protocol and the
genetic-algorithm-only clustering scheme GA-RP, GA-QL-RP achieves average
improvements of approximately 172.20% and 23.00% in packet delivery ratio,
respectively, and average reductions of approximately 72.06% and 25.95% in the
energy tax metric, respectively, thereby verifying the significant performance
advantages of the proposed protocol in dynamic underwater acoustic environments. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 01 (Paper No.6369) |
|
|
|
Title: |
Numerical and Experimental Study of a Built-in Pendulum-based Wave Energy
Converter |
|
|
|
Author: |
Shi Yihong1, Yangyang Cui2, Tao Wang3, Dongyang Chen3 and Siya Jin3 |
|
|
|
Abstract: |
This paper investigates a built-in pendulum-based wave energy converter (PWEC)
to address the power constraints of marine Internet of Things (IoT) nodes,
through theoretical modeling, numerical simulation and experimental testing. A
unified mathematical model considering coupled-decoupled state switching is
established and validated experimentally. Experiments are performed on a 1:4
scale PWEC model mounted on a six-degree-of freedom shaking table to reproduce
surge motions representative of real ocean waves. The full-scale device is
designed with a resonance pitch period of approximately 3 s (i.e., 1.5 s in the
scaled model). Results indicate that increasing the arm length of the pendulum
increases the device's resonance period: a 0.335 m arm length yields a resonance
period of 1.25 s in the scaled model. Additionally, the system achieves a
maximum output power of 0.992 W under an optimal load matching condition. These
findings advance the understanding of PWEC dynamics and support the design and
application of self-powered marine equipment using wave energy. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 01 (Paper No.6370) |
|
|
|
Title: |
UAVs-PPSim: A Simulator for Cooperative Path Planning of UAV Swarms |
|
|
|
Author: |
Cunle Zhang1, Shiduo Zhang1, Haonan Wang1, Chengkai Tang1, Baowang Lian1 and
Lingling Zhang1 |
|
|
|
Abstract: |
Reinforcement learning (RL) has demonstrated immense potential in UAV swarm path
planning, yet its training heavily relies on efficient simulation environments.
Based on the underlying framework of Isaac Sim, this paper proposes a simulation
platform specifically tailored for UAV swarm path planning tasks. By decoupling
low-level flight control from high-level decision-making, it facilitates the
easy design and experimentation of various path planning scenarios on top of
GPU-parallel simulations. It is equipped with 2 drone models, 5 sensor
modalities, 4 control modes, and a selection of widely-used RL baselines. To
showcase the capabilities of the platform, we provide preliminary results on
several benchmark planning tasks. We hope this platform will facilitate future
Sim-to-Real research on complex UAV swarms |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 01 (Paper No.6386) |
|
|
|
Title: |
Performance assessment of a 6-float M4 wave energy converter for offshore
autonomous power supply in the East China Sea |
|
|
|
Author: |
Wei Li1, Dongyang Chen1 and Siya Jin1 |
|
|
|
Abstract: |
The 6-float M4 wave energy converter (WEC) is a representative multi-body system
with potential for broadband energy capture. In this study, a time-domain
numerical model is developed in WEC-Sim and validated against experimental data
to investigate its hydrodynamic response and power performance. The results show
that the relative hinge response exhibits two distinct resonance peaks,
corresponding to different coupled motion modes. Consistently, two high-power
regions are identified, indicating that energy capture is governed by the
relative hinge motion. The maximum average power of 2.38 W is achieved at a wave
period of 1.15 s and wave height of 0.04 m with a power take-off (PTO) damping
coefficient of 3 N·m·s/rad. Using Froude similarity, the results are
extrapolated to full scale and applied to the East China Sea. A tuned 6-float M4
device (~50 m) is predicted to achieve about 100 kW under the representative
wave condition of the East China Sea, showing reasonable frequency compatibility
with the regional wave climate. These findings highlight the importance of
multi-body coupling and resonance tuning in enhancing energy capture and
adapting WEC systems to site-specific environments and provide a useful
reference for offshore autonomous power supply applications, such as ocean
monitoring and Internet of Things (IoT) systems. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 01 (Paper No.6388) |
|
|
|
Title: |
Multi-modal Fusion based UAV Data Acquisition System for Maritime Search and
Rescue |
|
|
|
Author: |
Xiaodong Zheng1, Longyun Yuan2, Lingling Zhang2, Peilin Chen1, Hongjun Ye1 and
Fangli Tian1 |
|
|
|
Abstract: |
Maritime person-overboard search and rescue faces critical challenges including
low visibility, small target size, and difficulty in assessing survivor
physiological states. This paper presents an integrated UAV-based perception and
positioning system that fuses visible and thermal infrared imagery. The system
uses a cascaded detection framework combining YOLOv26 and YOLO-World-S that
achieves 79.5% mAP@0.5 with only 3.98M parameters. Vital sign assessment is
performed through a six-dimensional fusion model that integrates thermal, pose,
motion, physiological, consistency, and signal-to-noise features via entropy
weighting, The assessment accuracy reaches 97.5%, a 23.5% improvement over
single-modality baselines in lake trials and maintains 92.1% under complex sea
state. Relative positioning between the UAV and a moving vessel is realized
through an RTK-based dynamic scheme with centimeter-level accuracy. Lake and sea
trial experiments validate the system's effectiveness and robustness for
real-world maritime search and rescue operations. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 01 (Paper No.6432) |
|
|
|
Title: |
A Double Deep Reinforcement Learning Approach for Adaptive Access Control in
Underwater Acoustic Networks |
|
|
|
Author: |
Jianmin Yang1, Yan Lin2, Can Wang3, Zhihong Peng4, Ming Gong5 and Ying Huang6 |
|
|
|
Abstract: |
Abstract-To address challenges such as extreme propagation delays, high
bit-error rates (BER), and stringent bandwidth constraints in Underwater
Acoustic Sensor Networks (UASNs), this paper proposes DDQN-CA, an intelligent
adaptive MAC protocol based on Double Deep Q-Networks (DDQN) and
multi-dimensional feature perception. By integrating non-stationary
features-including transmission success rates, collision frequencies, and node
queue status-the protocol achieves precise prediction and proactive regulation
of channel contention, breaking through the limitations of the traditional
Binary Exponential Backoff (BEB) mechanism. To mitigate the stochasticity of
underwater environments, the protocol introduces experience replay, target
network, and safety circuit-breaker mechanisms, significantly enhancing policy
convergence speed and robustness in complex dynamic scenarios. Simulation
results demonstrate that, compared to the Pure ALOHA protocol, DDQN-CA maintains
a stable Packet Delivery Ratio (PDR) exceeding 90%, reduces end-to-end latency
by 18.7%-34.2%, and achieves more than a two-fold leap in peak throughput.
Future work will focus on evaluating the protocol's performance in large-scale
networks to further verify its robustness in complex oceanic environments. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 01 (Paper No.6484) |
|
|
|
Title: |
SitePrompt-PV: Metadata-Aware Cross-Site Few-Shot Forecasting for Cold-Start
Photovoltaic Power |
|
|
|
Author: |
Qing Wang1, Guohong Li2, Yun Wang3, Luoxiao Yang1 and Yue Zhao4 |
|
|
|
Abstract: |
Short-term photovoltaic (PV) power forecasting is commonly trained for each site
independently, yet newly deployed sites often have only a few days of reliable
local observations while related sites already contain transferable PV
generation patterns. This paper formulates this setting as cross-site cold start
PV forecasting and proposes SitePrompt-PV, a metadata-aware few-shot adaptation
framework. SitePrompt-PV first maps site outputs to a per-unit capacity scale,
learns a shared PV prior from pooled non-target sites, represents transferable
site attributes as a metadata-conditioning vector, and then performs shot-aware
lightweight calibration using 0-14 days of local target history. The framework
is evaluated under a strict six target leave-one-site-out protocol on a
two-month UNISOLAR slice containing 41 PV sites across five campuses, and is
further examined on a power-only SKIPP'D external validation slice. On UNISOLAR,
SitePrompt-PV Adaptive reduces daylight nRMSE/nMAE from 15.56%/10.08% with zero
target days to 11.79%/7.95% with 7 days and 11.59%/7.94% with 14 days. At 7
days, it outperforms persistence, which obtains 16.44%/10.69%, and target-only
PatchTST, which obtains 15.84%/10.19%. On SKIPP'D, daylight nRMSE decreases from
18.63% to 13.03% after 14 days of adaptation despite the external domain shift.
Ablation results show that the source-site prior is the largest contributor,
while metadata conditioning, forecast-horizon non-power covariates, target
calibration, and shot-aware calibration weighting each provide additional gains.
These results indicate that a capacity-normalized source prior, metadata
conditioning, and shot-aware target calibration provide a practical solution for
cold-start PV forecasting. |
|
|
|
|
|
|
|
|
CPT Session 02 |
|
|
|
Time: 09:00~10:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Ningning Pan, Southwestern University of Finance and Economics |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 02 (Paper No.6359) |
|
|
|
Title: |
YOLO-DFG: A Lightweight Drone Vehicle Target Detection Algorithm for Remote
Sensing Images |
|
|
|
Author: |
Congjing Wang1 and Yongqing Qian2 |
|
|
|
Abstract: |
Aiming at the problems of small vehicle target size, complex background,
susceptibility to occlusion in remote sensing images, as well as the large
number of parameters of existing detection models and the difficulty of
deployment on resource-constrained platforms, a lightweight and high-precision
remote sensing vehicle target detection model based on an improved YOLOv11 is
proposed. In this paper, a DualPoolDown downsampling module is designed to
replace the original structure and combined with C3k2-PFDConv (Partial Frequency
Convolution C3k2 module) to construct an efficient feature extraction backbone,
enhancing the model's ability to extract features of rotationally varying
targets. Finally, an LSCDECA Gate (Long-Short Convolution Enhanced Coordinated
Attention Gating Mechanism) is introduced in the detection head, which
effectively fuses high-level semantic information and low-level spatial details,
suppresses complex background noise, and improves the localization accuracy of
dense small targets. Experimental results show that, compared with the baseline
YOLOv11n model, the proposed algorithm reduces computational cost while
maintaining detection performance, providing a new solution for real-time
vehicle target detection in drone imagery. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 02 (Paper No.6360) |
|
|
|
Title: |
TEFoley: Text-Semantically Enhanced Text-Video-to-Audio Generation |
|
|
|
Author: |
Zhi Cheng1, Cien Fan1 and Lehui Wei1 |
|
|
|
Abstract: |
Recent advances in text-video-to-audio (TV2A) generation have made it possible
to synthesize high-fidelity audio from multimodal inputs. However, the generated
audio aligns with textual semantics remains a significant challenge.
Specifically, existing methods heavily rely on synchronization and semantic
information provided by visual cues, with text content serving only as a
secondary reference, which compromises the overall user experience. To overcome
this limitation, we propose TEClip, a visual encoder for text semantic
enhancement that captures text-related semantic information from videos. We also
construct the VGGSound-TE dataset to train this encoder. Building upon this
encoder, we introduce TEFoley, a multimodal audio generation framework that
significantly enhances text-based semantic guidance in audio synthesis.
Experiments on the VGGSound test set demonstrate that our method achieves
state-of-the-art performance in audio generation tasks, including improvements
in audio quality, semantic alignment, and temporal synchronization. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 02 (Paper No.6364) |
|
|
|
Title: |
Improving Instruction Encoding for Vision-and-Language Navigation via
Similarity-Aware Contrastive Learning |
|
|
|
Author: |
Chengyan He1, Jia'ao sun1, Zhaolin Zhang1, Yandong Sun1 and Ling Wang1 |
|
|
|
Abstract: |
Vision-and-Language Navigation (VLN) requires agents to understand natural
language instructions and align them with vision environments for effective
decision-making. However, existing methods often struggle to capture semantic
consistency under diverse linguistic expressions, limiting retrieval
performance. To address this issue, this paper proposes a BERT-based contrastive
learning framework for semantic similarity modeling in VLN. A dual-encoder
architecture is adopted to obtain contextualized sentence representations, and a
pseudo-label generation strategy combining paraphrase augmentation and in-batch
pseudo-labeling is introduced to enhance data diversity without additional
manual annotation. Furthermore, a contrastive loss is designed to jointly
encourage positive alignment and negative separation, enabling the model to
learn more discriminative embeddings. Experiments on the R2R dataset demonstrate
that the proposed method achieves superior retrieval performance compared with
conventional approaches, validating its effectiveness for semantic
representation and instruction retrieval in VLN tasks. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 02 (Paper No.6376) |
|
|
|
Title: |
CNN-Transformer Super-resolution Network for Hyperspectral Image Reconstruction |
|
|
|
Author: |
Yuxiao Li1, Yukun Zhang1, Jinhang Yan1, Chang Liu1, Shuo Li1 and Yifan Zhang1 |
|
|
|
Abstract: |
To address the issues of insufficient spatial detail recovery, spectral
distortion, and low computational efficiency in hyperspectral image
super-resolution reconstruction, CNN-Transformer Super-resolution Network (CTSN)
integrating CNN and Transformer architectures, is proposed in this paper. The
model employs a multi-layer CNN for initial feature extraction, followed by a
spatial self-attention mechanism for deep feature extraction, and the high
spatial resolution image is then reconstructed through feature fusion.
Reconstruction performance employing CNNs with 1 to 5 layers are compared. It is
illustrated that a 3-layer CNN exhibits the best overall performance and
strongest generalization ability, while a 1-layer CNN performs better in certain
scenarios and would be a preferable option for practical applications.
Furthermore, a hybrid loss function with weighted L1 loss with spectral angle
loss (SAM) is employed to achieve better balance between spatial detail recovery
and spectral fidelity. Finally, comparison with current mainstream models
illustrate that, the newly proposed model has a smaller reconstruction parameter
scale than the mainstream methods and significantly improves computational
efficiency, providing an efficient and feasible solution for hyperspectral image
super-resolution applications aimed at real-time inference. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 02 (Paper No.6378) |
|
|
|
Title: |
Lite-RP-DETR: Lightweight Transformer-Based Rice Pest Detection via Pruning and
Distillation |
|
|
|
Author: |
Rongfu Chen1, Canyang Zhou2, Fanlong Zhang2, Caifeng Zou3 and Jianqi Liu3 |
|
|
|
Abstract: |
We propose Lite-RP-DETR, a lightweight transformer-based detector for rice pest
detection in resource-constrained environments. The framework combines
dependency-aware structured pruning with multi-level knowledge distillation to
reduce model complexity while preserving accuracy. Experiments show that the
proposed method significantly improves efficiency with minimal performance loss,
making it suitable for real-time agricultural applications. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 02 (Paper No.6416) |
|
|
|
Title: |
Training-Time SAM Knowledge Injection for Deployment-Efficient Metallic Surface
Defect Segmentation |
|
|
|
Author: |
Jie Xu1, Zongfang Ma1 and Yun Wang2 |
|
|
|
Abstract: |
Metal surface defect segmentation in industrial scenarios faces the challenge of
learning stable and generalizable defect representations under limited
annotation. Different from existing methods that improve model recognition
ability by expanding datasets, introducing transferable segmentation knowledge
from vision foundation models and injecting it into classical domain-specific
segmentation models provides a promising and low-cost alternative. To this end,
this paper proposes Prompt-Consensus Teacher Distillation (PCTD), which
leverages the generic region, shape, and boundary priors embedded in SAM to
enhance the training of metal surface defect segmentation models. Specifically,
PCTD integrates multiple promptconditioned SAM predictions into a consensus
teacher map through learnable prompt fusion, and transfers reliable region and
contour information to the student model via confidenceaware distillation and
boundary prior learning. Experiments on FSSD-12 and Surface Defects-4i show that
PCTD improves the strongest plain baseline by 3.11/7.89 and 7.08/9.94 percentage
points in IoU/B-F1, respectively, while keeping SAM only in the training stage. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 02 (Paper No.6458) |
|
|
|
Title: |
SAR-ShipDetNet: A Multi-Scale Speckle-Robust Oriented Detection Network for Ship
Detection in SAR Images |
|
|
|
Author: |
Yida Yang1, Jiewen Tian2 and Yifei Zhang1 |
|
|
|
Abstract: |
Ship detection in synthetic aperture radar (SAR) imagery is an important task
for maritime surveillance, port monitoring, traffic management, and ocean
security. Compared with optical ship detection, SAR ship detection is robust to
illumination and weather changes, but it remains challenging due to speckle
noise, complex sea--land clutter, small target size, dense harbor distributions,
and arbitrary ship orientations. To address these issues, this paper proposes
\textit{SAR-ShipDetNet}, a multi-scale speckle-robust oriented detection network
for SAR ship detection. The proposed network contains four key components: a
SAR-aware stem for shallow speckle-resistant feature extraction, a large-kernel
convolutional backbone for elongated ship modeling, a P2--P5 multi-scale feature
pyramid for small and medium ship representation, and a decoupled anchor-free
oriented detection head for predicting ship center, size, angle, objectness, and
category confidence. An optional auxiliary foreground mask branch is introduced
to enhance localization and suppress sea--land background interference when
segmentation annotations are available. The overall loss combines focal
classification loss, objectness loss, rotated-box regression loss, angle
regularization, and optional mask supervision. |
|
|
|
|
|
|
|
|
CPT Session 03 |
|
|
|
Time: 11:00~12:30, Sunday, July 19, 2026 |
|
|
|
Session Chair: Zhengqiao Zhao, Northwestern Polytechnical University |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 03 (Paper No.6310) |
|
|
|
Title: |
AI-Driven IoT Framework for Real-Time Predictive Degradation Modeling in Solar
Photovoltaic Systems |
|
|
|
Author: |
Atul Anand1, Sudhakar Kumar1, Sunil K. Singh1, Varsha Arya2, Kwok Tai Chui2 and
Brij B. Gupta3 |
|
|
|
Abstract: |
Solar Photovoltaic (PV) systems experience inefficiency due to environmental
factors such as dust, thermal stress, and degradation of materials. This paper
develops an IoT-based machine learning prediction framework for the reliability
of solar PV systems. By leveraging a dataset of 68,378 records from 2020,
including DC power output, solar irradiance, module temperature, and ambient
temperature, the framework uses the Decision Tree, Random Forest, and Gradient
Boosting techniques provided by the scikit-learn library. It has achieved 96.2%
accuracy, 95.8% precision, 94.7% recall, 95.2% F1 score, and 0.99 AUC-ROC values
for the Random Forest model. Unlike other studies, which have utilized a limited
number of data records, the proposed framework has the ability for real-time
monitoring of standalone and grid-connected solar PV systems, reducing the cost
of maintenance for the sustainable management of energy resources. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 03 (Paper No.6365) |
|
|
|
Title: |
Design, Verification, and ASIC Implementation of a Five-Stage Pipelined RV32I
RISC-V Processor with Dynamic Branch Prediction |
|
|
|
Author: |
Hoi Lam Huang1, Hao Yang1, Muhammad Irfan2 and Ray Chak-Chung Cheung1 |
|
|
|
Abstract: |
This paper presents the design, verification, and ASIC implementation of a
five-stage pipelined RV32I RISC-V processor through an open-source RTL-to-GDSII
flow. The architecture incorporates a dynamic branch prediction unit featuring a
16-entry direct-mapped branch target buffer (BTB) and a 2-bit saturating counter
branch history table (BHT). Functional verification was conducted using
Verilator-based RTL simulation, including directed tests and CoreMark
benchmarking, which demonstrated a performance improvement from 0.83 to 1.02
CoreMark/MHz relative to a non-predicting pipeline. The design was implemented
to GDSII using OpenLane 1.0.2 and the SkyWater 130 nm (SKY130) process node,
integrating two 1 KB SRAM macros for instruction and data memory. Physical
implementation achieved timing closure at 100 MHz with a signoff area of 1.34
mm? and an estimated power of 50.1 mW. This work demonstrates that open-source
RTL-to-GDSII flows are viable for implementing pipelined processors with
microarchitectural enhancements, while providing physical signoff PPA metrics
that are absent from prior FPGA-based open-source processor works. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 03 (Paper No.6402) |
|
|
|
Title: |
Cooperative Dual-UAV Target Tracking via Hybrid Supervised Pretraining and
Residual MADDPG Fine-Tuning |
|
|
|
Author: |
Lina Zeng1, Zhe Zhang1, Haoyu Wang1 and Hang Xu1 |
|
|
|
Abstract: |
This paper presents a three-phase hybrid learning framework for cooperative
dual-UAV target tracking. The task requires two unmanned aerial vehicles to
maintain a prescribed right-angle observation geometry relative to a moving
ground target while satisfying inter-UAV safety constraints. Behavior cloning
suffers from distribution shift, whereas conventional reinforcement learning
suffers from high sample complexity and unstable convergence in multi-agent
settings. To address these limitations, the proposed framework first uses a
nonlinear model predictive controller to automatically generate expert
demonstration data. A Transformer encoder is then pretrained through behavior
cloning to provide a stable baseline policy. Finally, a lightweight residual
multi-agent reinforcement learning module fine-tunes the baseline policy by
learning corrective actions rather than replacing the pretrained controller.
Three complementary smoothness mechanisms, including action-rate reward shaping,
actor temporal regularization, and test-time exponential moving average
filtering, are introduced to suppress high-frequency control jitter without
sacrificing tracking responsiveness. Experiments over ten randomized episodes
indicate that the proposed method achieves a 100 percent success rate and a mean
steady-state angle error of 2.18 degrees, outperforming behavior-cloning-only
control, which achieves 40 percent success with a 32.2 degree mean error, and
multi-agent reinforcement learning trained from scratch, which achieves 80
percent success with a 6.2 degree mean error using twice the training budget.
Ablation studies further confirm the individual contribution of each smoothness
mechanism. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 03 (Paper No.6420) |
|
|
|
Title: |
NADST-Net: Noise-Adaptive Dual-Domain Shrinkage Transformer Network for Robust
HRRP Target Recognition |
|
|
|
Author: |
Songyan XSY XIE1 and Meng YM Yu1 |
|
|
|
Abstract: |
Abstract—High-resolution range profile (HRRP) recognition
is an important technique for radar automatic target recognition
because HRRP preserves the one-dimensional scattering
distribution of a target along the radar line of sight. However,
recognition performance degrades significantly in low signal-tonoise
ratio (SNR) environments, where weak scattering centers
are easily submerged by noise and discriminative HRRP structures
are distorted. To address this problem, this paper proposes
a noise-adaptive dual-domain shrinkage transformer network,
termed NADST-Net, for robust HRRP target recognition under
complex noise conditions. The proposed method contains four
core components. First, a lightweight noise estimation head
extracts a noise-state embedding from the input HRRP. Second, a
noise-conditioned multi-scale shrinkage front-end adaptively suppresses
irrelevant components while preserving salient scattering
structures. Third, a dual-domain encoder jointly models local
scattering patterns in the range domain and global dependencies
in the frequency domain. Finally, a noise-aware fusion module
dynamically balances the two domains according to the noise
state. To further enhance noise invariance, a multi-objective
loss is introduced by combining classification loss, reconstruction
loss, supervised contrastive loss, and prediction-consistency
regularization. The experiments evaluate evaluates not only
conventional recognition accuracy under multiple SNR levels, but
also robustness to mismatched noise types, module effectiveness,
and computational efficiency. The proposed framework provides
a unified solution that integrates adaptive denoising, dual-domain
feature learning, and noise-invariant representation learning for
robust HRRP recognition. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 03 (Paper No.6422) |
|
|
|
Title: |
Adaptive State Evolution for Ionospheric TEC Forecasting in Emergency
Communication Planning |
|
|
|
Author: |
Rongchang Lu1, Guoming Yuan2, Yongjun Cheng3 and jiayu Sheng4 |
|
|
|
Abstract: |
Emergency communication planning after earthquakes, floods, or other sudden
disasters often depends on GNSS, satellite messaging, and high frequency radio
as fallback channels. These links still traverse the ionosphere, where Total
Electron Content, or TEC, affects group delay, phase advance, positioning
accuracy, and link reliability during space weather disturbances. This paper
presents the Adaptive State Evolution (ASE) recurrent forecasting framework for
TEC spatial sequence prediction under this combined terrestrial and ionospheric
risk. ASE treats hidden state propagation as an input dependent discretization
of a continuous dynamic system, allocating short memory steps to storm
transients and longer steps to quiet intervals. For deployment, the evaluated
pipeline exports forecasts to PostgreSQL/PostGIS as a spatial database and GIS
service layer. The evaluation covers Quiet, Storm, Solar Wind Forced, and IONEX
Regional TEC scenarios. After discarding models whose accuracy is below 60
percent and Pearson correlation is below 0.5, ASE obtains the best retained OMNI
forced result, with 97.40 percent accuracy, 0.9979 Pearson correlation, 0.4668
RMSE, 36.93 dB PSNR, and 0.9893 SSIM. These results indicate that adaptive
temporal evolution can support fast TEC nowcasting for emergency communication
planning. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 03 (Paper No.6424) |
|
|
|
Title: |
Knowledge Distillation for Efficient Underwater Acoustic Classification |
|
|
|
Author: |
Yintao Li1, Zhengqiao Zhao1, Haoxiang Wu1 and Jie Chen1 |
|
|
|
Abstract: |
Underwater acoustic monitoring systems require models that achieve high accuracy
while remaining lightweight enough for deployment on edge devices. This paper
investigates knowledge distillation (KD) from convolutional neural network
(CNN)-based teachers to a MobileNetV2 student using a five-class underwater
target recognition dataset, ShipsEar. Experimental results show that a
lightweight MobileNetV2 model trained from scratch on this dataset attains only
42.94% mean test accuracy, limiting its practical applicability. To improve
performance, we evaluate three CNN-based teacher models for knowledge
distillation, namely, CNN14, CNN10, and CNN6, and show that teacher selection
plays a critical role in knowledge transfer effectiveness. Although CNN14
achieves the highest mean validation accuracy, CNN6 yields the best test
performance and produces the most effective distilled student model. Using the
response-based knowledge distillation, the CNN6-to-MobileNetV2 framework
achieves a mean test accuracy of 55.40% and outperforms all other
teacher-student configurations. Performance gains are consistently observed
across all five folds, together with a notable reduction in cross-fold variance.
These results suggest that a shallower yet architecturally compatible teacher
transfers structural knowledge more effectively to a compact student model,
demonstrating the potential of MobileNetV2 for real-time underwater acoustic
monitoring on resource-constrained devices. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 03 (Paper No.6456) |
|
|
|
Title: |
Dual-Branch Image Transmission Network with Nested Vector Quantization for
Extremely Low Bandwidth |
|
|
|
Author: |
Shaosong Cao1 and Xuechen Chen1 |
|
|
|
Abstract: |
High-fidelity image transmission under extremely low channel bandwidth ratio
remains challenging, as digital schemes suffer from cliff effects, while
existing deep JSCC methods are limited in preserving fine-grained textures under
severe bandwidth constraints. This paper proposes a dual-branch image
transmission network with nested vector quantization for low-bandwidth wireless
channels. The base branch transmits a low-resolution image to preserve global
structure and provide graceful degradation. The detail branch extracts
multi-scale residual features and quantizes them with a coarse-to-fine nested
vector quantization module, enabling compact transmission of high-frequency
information. At the receiver, a CRC-guided multi-stage fusion decoder adaptively
incorporates reliable residual features and suppresses corrupted digital
information to prevent error propagation. Experiments on the Kodak dataset over
AWGN and Rayleigh fading channels demonstrate that the proposed method
outperforms DeepJSCC and SwinJSCC under extremely low bandwidth while
maintaining robustness when the digital branch fails. |
|
|
|
|
|
|
|
|
Oral Session: |
CPT Session 03 (Paper No.6474) |
|
|
|
Title: |
SolarPatchTST: A Linear-Anchored Patch Transformer for Short-Term Photovoltaic
Power Forecasting |
|
|
|
Author: |
Ting Kang1, Yue Zhao2 and Luoxiao Yang3 |
|
|
|
Abstract: |
Short-term photovoltaic (PV) power forecasting must handle daily regularity,
irradiance-driven ramps, and physically nonnegative outputs. Generic Transformer
forecasters can model nonlinear temporal patterns, but direct use on compact PV
slices may be unstable. This paper proposes SolarPatchTST, a lightweight
PV-aware framework that combines a linear PV anchor, a gated PatchTST residual
branch, and a nonnegative power projection. A conservative SafeEns inference
rule averages the nonnegative anchor stream and residual stream to improve
robustness. On public PVDAQ and UNISOLAR datasets with chronological splits,
SolarPatchTST-SafeEns reduces MSE by 20.58% and 42.01%, respectively, relative
to the strongest compared neural baseline. |
|
|
|
|
|
|
