Published papers

The muti-task Channel State Information (CSI) feedback is a downlink CSI feedback approach. It focuses on Clustered Delay Line (CDL) channel models as the channel environments. There are five different CDL channel models, from CDL-A to CDL-E. In this method, the autoencoder (source model) is pre-trained using the CDL-ALL dataset, which is an equal mixture of each CDL channel dataset (source data) rather than a single CDL channel dataset. The decoder at the BS is subsequently fine-tuned using a small amount of CSI data from the target channel environment (target data) to generate a target model. The source model has a significant influence on the reconstruction performance of the target model, the amount of data, and the number of training epochs required to obtain it. This paper investigates the mixing ratio of each CDL channel dataset used as source data to enhance the CSI reconstruction performance of the source model in the muti-task CSI feedback. Furthermore, we explore the possibility of enhancing the CSI reconstruction performance of the target model by employing the improved source model for fine-tuning the target model. To determine the mixing ratio, two distinct criteria for source model performance are employed. The simulation results identified several mixing ratios of source data in these criteria that improve the reconstruction performance of both the source and the target models compared to the use of CDL-ALL as a source dataset in the multi-task CSI feedback.

Stress has become an increasingly common issue, significantly impacting individuals' health and productivity. Traditional methods for assessing stress levels often rely on physiological signals, such as heart rate and skin conductance. However, these methods can be impractical for continuous, non-invasive monitoring. To address these limitations, we propose a machine learning approach for stress assessment using eye-tracking data. We collected eye-tracking and Electrocardiogram (ECG) data under different stress conditions. From this dataset, we extracted relevant features and conducted preliminary analysis. We further proposed a novel stress index, the Least Absolute Shrinkage and Selection Operator Stress Index (Lasso Stress Index), specifically tailored for stress detection in our dataset. In previous studies, researchers commonly used predefined subjective stress index or Baevsky's Stress Index (BSI) for stress classification. We compared these methods and found that our proposed Lasso Stress Index yielded the highest accuracy, achieving up to 90.6% in stress classification.

Chronic stress has become a significant issue in modern society, negatively affecting individuals' overall well-being and potentially leading to severe physiological and psychological disorders. Most existing methods for automatic stress detection rely on physiological data, which requires direct skin contact with sensors. Not only does this make them intrusive, but it also limits their practicality for everyday use due to accuracy constraints. In this paper, we propose a novel approach using eye data captured by a standard video camera -specifically, eye gaze direction and eye landmarks- for stress detection. We use two types of Recurrent Neural Networks (RNNs), namely a Long Short-Term Memory (LSTM) network and a Gated Recurrent Unit (GRU) network. Through experiments on the UBFC-Phys dataset, we show that our method achieves a stress detection accuracy of up to 83.69% using gaze data, while eye landmarks further improve accuracy to 90.12%. Combined, these two modalities reach accuracy of 90.83%. These results highlight the effectiveness of our approach in enabling accurate and non-intrusive stress detection.

The reconfigurable intelligent surface (RIS) plays a crucial role in enabling sustainable communication infrastructure for 6G networks. This letter presents a long-term net weighted sum-rate maximization problem with queue stability for irregular simultaneous transmission and reflection RIS (ISTAR-RIS)-assisted vehicular networks. The model incorporates time-varying and mobility-induced channel dynamics. A joint optimization problem of the road side unit (RSU) beamforming vector, topological selection and the phase shift matrix of ISTAR-RIS elements is formulated. The simulation results demonstrate the effectiveness of the proposed approach in maximizing the sum-rate while comparing with existing benchmark methods in ISTAR-RIS-aided wireless and device-to-device (D2D) communication.

Synthetic Aperture Radar (SAR) change detection is the process of identifying image changes by comparing SAR images of the same area captured at different time intervals. Considering the interference of speckle noise and the lack of spatial adaptability, this paper proposes a spatial-frequency domain feature fusion cooperative detection framework. This method extracts local structural features in the spatial domain while incorporating high-frequency components from the frequency domain to enhance edge feature representation and improve the discrimination of change regions. Additionally, a multi-region dynamic selection mechanism is designed to adaptively capture multi-scale contextual information through cascaded convolutional layers, overcoming the limitations of fixed receptive fields in complex dynamic scenarios. Furthermore, by integrating an attention mechanism with the Swish-Gated Linear Unit (SWiGLU), the network’s feature representation capability is optimized, effectively enhancing the precision of subtle change detection and significantly reducing the impact of speckle noise. This further improves the stability and reliability of SAR change detection. Through performance evaluation on five datasets, we compare the proposed method with state-of-the-art SAR change detection techniques. The results demonstrate that our method exhibits significant advantages in terms of accuracy and kappa coefficient.

There is a growing interest in high-altitude platform stations (HAPSs) as potential telecommunication infrastructures in the stratosphere, providing direct communication services to ground-based smartphones. Enhanced coverage and capacity can be realized in HAPSs by adopting multicell configurations. To improve the communication quality, previous studies have investigated methods based on search algorithms, such as genetic algorithms (GAs), which dynamically optimize antenna parameters. However, these methods face hurdles in swiftly adapting to sudden distribution shifts from natural disasters or major events due to their high computational requirements. Moreover, they do not utilize the previous optimization results, which require calculations each time. This study introduces a novel optimization approach based on a neural network (NN) model that is trained on GA solutions. The simple model is easy to implement and allows for instantaneous adaptation to unexpected distribution changes. However, the NN faces the difficulty of capturing the dependencies among neighboring cells. To address the problem, a classifier chain (CC), which chains multiple classifiers to learn output relationships, is integrated into the NN. However, the performance of the CC depends on the output sequence. Therefore, we employ an ensemble approach to integrate the CCs with different sequences and select the best solution. The results of simulations based on distributions in Japan indicate that the proposed method achieves a total throughput whose cumulative distribution function (CDF) is close to that obtained by the GA solutions. In addition, the results show that the proposed method is more time-efficient than GA in terms of the total time required to optimize each user distribution.

Local model-agnostic Explainable Artificial Intelligence (XAI), such as LIME or SHAP, has recently gained popularity among researchers and data scientists for explaining black box Machine Learning (ML) models. In the industry, practitioners focus not only on how these explanations can validate their models but also on how they can help maintain trust from end-users. Some studies attempted to measure this ability by quantifying what they refer to as the explainability or interpretability of ML models. In this paper, we introduce a new method for measuring explainability with reference to an approximated human model. We develop a human-friendly interface to strategically collect human decision-making and translate it into a set of logical rules and intuitions, or simply annotations. These annotations are then compared with the local explanations derived from common XAI tools. Through a human survey, we demonstrate that it is possible to quantify human intuition and empirically compare it to a given explanation, enabling a practical quantification of explainability. By relying on this new method, we identified several potential flaws in today's ML selection process. Furthermore, we demonstrate how our method can help to better evaluate ML models.

With the growing adoption of Internet of Things (IoT) devices, ensuring the security of wireless communications has become increasingly critical. Radio frequency fingerprint identification (RFFI) has shown promise in this regard due to its capability of uniquely identifying devices. Although deep learning (DL) approaches have significantly improved RFFI performance, they typically rely on large-scale centralized data. This poses challenges in terms of privacy preservation and heterogeneous data distributions. To address the performance degradation caused by non-independent and identically distributed (non-IID) data in cross-receiver scenarios, this paper proposes a feature alignment strategy based on federated learning (FL) for RFFI. In such scenarios, due to differences in receiver hardware characteristics, deployment locations, and channel conditions, the signals captured by different receivers often exhibit distribution shifts, resulting in misaligned feature spaces across clients. The proposed method guides each client to learn aligned intermediate feature representations during local training, effectively mitigating the resulting adverse impact on model generalization. Experiments conducted on a real-world RF dataset demonstrate that the proposed method achieves higher identification accuracy and improved stability compared with representative federated baselines, including FedAvg and FedProx. The highest identification accuracy reaches 90.83%, and the performance gains are accompanied by generally reduced variance across different client configurations, highlighting the robustness and generalization capability of the proposed approach in heterogeneous wireless environments.

The rapid advancement of intelligent perception systems and autonomous navigation technologies imposes increasingly stringent requirements on state estimation algorithms regarding modeling capacity, robustness, and real-time performance. While traditional Kalman Filter (KF) demonstrates computational efficiency and mathematical interpretability for linear Gaussian systems, their effectiveness diminishes in nonlinear, non-Gaussian, and non-stationary dynamic scenarios. Recent developments in neural state space modeling offer new perspectives for recursive estimation, yet existing approaches suffer from structural complexity, training difficulties, and inadequate real-time capability. This paper presents Mamba-KF, a novel state estimation framework integrating deep modeling with recursive filtering. Leveraging the cutting-edge Mamba state space neural network, our method dynamically generates state transition parameters while establishing a differentiable covariance propagation pathway, enabling end-to-end optimization. We further design a lightweight multi-modal fusion architecture tailored for high-frequency sensory data processing in real-time applications. Evaluated on the EuRoC Micro Aerial Vehicle (MAV) dataset, Mamba-KF attains the lowest errors across states, reducing position Root Mean Square Error (RMSE) by up to 33.3% relative to Extended Kalman Filter (EKF) and 24.1% relative to Unscented Kalman Filter (UKF) at 10 dB Signal-to-Noise Ratio (SNR), with an average velocity RMSE of 0.097 m/s and an orientation error of 2.3°. It also achieves superior consistency, with a mean Normalized Estimation Error Squared (NEES) of 6.02 and 93.7% of estimates within the χ2-based 95% confidence interval. Under robustness evaluations, Mamba-KF yields the lowest RMSEs and suppresses divergence observed in EKF. Computationally, it maintains near real-time performance with 0.65 ms per step, offering a favorable accuracy-efficiency trade-off for visual-inertial navigation and autonomous systems, while providing significant theoretical insights and promising engineering applications.

In multi-intersection traffic signal timing (TST) the use of a uniform cycle time reduces efficiency at low-volume intersections and signal changes which can alter vehicle flows and degrade communication quality. This paper presents a solution to this problem introducing a Twin-Timescale model whereby a multi-intersection TST method is jointly considered with communication resource allocation (CRA) strategies. For each intersection, a Twin-Timescale Reinforcement Learning (RL) algorithm is proposed, by considering both CRA and TST as the Short- and the Long- Timescale models, respectively, and jointly formulating their rewards. Based on the characteristics of this Twin-Timescale model, the reward function is constructed as a singularly perturbed model (SPM), and a state feedback controller that can ensure asymptotic stability is designed. Furthermore, a novel asynchronous time Multi-Agent Coordinated RL algorithm is proposed, which allows each intersection to generate each own joint parameters for its decision strategy which are based upon information obtained from neighboring intersections at asynchronous time. In addition, the use of Federated Learning (FL) in conjunction with the proposed algorithms is investigated, and its advantages in enhancing the resistance to malicious attacks are discussed. Using real traffic data, various performance evaluation results obtained by means of extensive computer simulation experiments have confirmed that the proposed TST-CRA joint optimization approach significantly improves traffic indicators and maintains the quality of the overall communication performance. Through numerical calculations, the state feedback controller coefficients have been determined.

With the advances in flexibility, wide coverage, and cost-effectiveness, integrating unmanned aerial vehicles (UAVs) with wireless sensor networks (WSNs) has been regarded as a promising technology in data collection and transmission within complex environments. Due to the limited battery capacity of sensor nodes (SNs) and UAVs, UAV-assisted WSNs face challenges in achieving energy efficiency and optimizing resource allocation. To reduce the total energy consumption of UAV-assisted WSNs, we design a multifactor fitness function which includes the distance between SNs, energy levels of SNs, and UAV trajectory length. Based on that, we propose a balanced clustering and cluster head (CH) selection strategy. As for UAV trajectory design, we employ the branch and bound method according to the selection of CHs. Simulation results show that the proposed algorithm reduces the average energy consumption of all SNs and achieves short UAV trajectory length compared to Low-Energy Adaptive Clustering Hierarchy (LEACH), Hybrid Energy-Efficient Distributed clustering (HEED), Particle Swarm Optimization (PSO), and Stable Election Protocol (SEP).

Heartbeat detection is a promising technology for health care and is in great demand in many applications. For non-contact and non-invasive heartbeat detection, Doppler sensor-based heartbeat detection methods have been investigated extensively. However, since the conventional methods use a single beam Doppler sensor, the accuracy of heartbeat detection tends to degrade, when the SNR (Signal-to-Noise Ratio) of heartbeat components for the beam direction is low. In this paper, we propose a non-contact heartbeat detection method based on beam diversity using a multibeam Doppler sensor. The effects of noises and the corresponding SNRs of heartbeat components over the received signal differ from one beam to another. Based on this fact, the proposed method detects heartbeat by exploiting beam diversity of the received signals. The experimental results showed that compared to the detection method using a single beam, our proposed method using multiple beams detected heartbeat accurately, which indicates that exploiting the beam diversity can lead to the improvement of heartbeat detection accuracy.

The non-invasive fetal electrocardiogram (FECG) derived from abdominal surface electrodes has been widely used for fetal heart rate (FHR) monitoring to assess fetal well-being. However, the accuracy of FECG-based FHR estimation heavily depends on the quality of FECG signal itself, which can generally be affected by several interference sources such as maternal heart activities and fetal movements. Hence, FECG signal quality assessment (SQA) is an essential task to improve the accuracy of FHR estimation by removing or interpolating low-quality FECG signals. In recent research, various SQA methods based on supervised learning have been proposed. Although these methods could perform accurate SQA, they require large labeled datasets. Nevertheless, the labeled datasets for the FECG SQA are very limited. In this paper, to address this limitation, we propose an unsupervised learning-based SQA method for identifying high and low-quality FECG signal segments. Specifically, a fully convolutional network (FCN)-based autoencoder (AE) is trained for reconstructing a spectrogram derived from FECG. An AE-based feature related to reconstruction error is then calculated to identify high and low-quality FECG segments. In addition, entropy-based features, statistical features, and ECG signal quality indices (SQIs) are also extracted. The high and low-quality segments are identified by feeding the extracted features into self-organizing map (SOM). The experimental results showed that our proposal achieved an accuracy of 98% in high and low-quality signal classification.

Heartbeat is one of essential vital signs to assess our health condition. Noncontact heartbeat detection is thus receiving a lot of attention in recent years, which motivates many researchers to investigate heartbeat detection via a Doppler radar. In this paper, to detect heartbeat with a high accuracy, we propose a Doppler radar-based heartbeat detection method by the Viterbi algorithm with a fusion of Beat-Beat Interval (BBI) and deep learning-driven Branch Metrics (BM). The Viterbi algorithm is a technique to estimate a sequence with maximum likelihood by using a pre-defined metric, namely, a BM. In the proposed method, we combine two BMs defined based on (i) a difference between two adjacent BBIs and (ii) an output probability of a deep learning model that judges whether a peak is caused by heartbeat or not. We apply the VIterbi algorithm with the fusion of the two BMs to the signal obtained by some signal processing. We experimentally confirmed that our method performed heartbeat detection with small Root Mean Squared Error (RMSE) between the estimated and actual BBIs.

Remote sensing of heartbeat offers various applications in the medical and health care fields. To realize non-contact heartbeat detection, an FMCW (Frequency Modulated Continuous Wave) radar-based heartbeat detection method has been investigated. The conventional FMCW radar-based heartbeat detection method estimates a range from an FMCW radar to a subject and extracts heartbeat components from phase changes for the range. However, the range suitable for extracting heartbeat components can change over time due to respiration and body fluctuation. Thus, when the SNR (Signal-to-Noise Ratio) of heartbeat components over phase changes is low at the estimated range, the accuracy of heartbeat detection tends to degrade. In this paper, we propose a MIMO (Multiple-Input Multiple-Output) FMCW radar-based heartbeat detection method based on space diversity. A MIMO FMCW radar can estimate the range for multiple beam directions and obtain phase changes for a space specified with the range and the beam direction. The SNR of heartbeat components over phase changes differs from one space to another. Taking it into account, the proposed method detects heartbeat by exploiting the space diversity of phase changes. The experimental results showed that compared to the detection method using only one phase change, the proposed method using phase changes for multiple spaces detected heartbeat accurately, which is brought by the diversity effect of phase changes for multiple spaces.

High-Altitude Platform Stations (HAPSs) enable wide-area coverage in 6G networks but introduce interference when sharing spectrum with terrestrial networks (TNs). Null-forming is a technique to mitigate this interference by directing low-power beams (nulls) toward terrestrial users. Traditional nullforming methods, such as Null-Sweeping, rely on changing null directions across the resource blocks (RB) to improve the impact of nullforming. Yet these null directions are predefined for uniform user distributions and may not fully account for nonuniform deployments. We propose a user-aware nullforming approach that leverages K-means clustering to adapt null positions to dense user regions in the terrestrial cells, while time-frequency resources are allocated proportionally to user density. Simulations show that our method reduces HAPS interference for terrestrial users and improves fairness in interference distribution.

With the rapid advancement of deep learning techniques, object detection on Synthetic Aperture Radar (SAR) images has shown increasing potential in recent years. However, due to the unique microwave imaging mechanism of SAR, targets can easily be confused with background structures like harbors and aircraft aprons, leading to poor recognition accuracy. In this paper, we propose YOLOv10n-ACmix, an enhanced object detection model that integrates the ACmix attention mechanism into the YOLOv10 framework to improve feature extraction and detection precision. Through experiments conducted on the SAR-aircraft and MSAR datasets, YOLOv10n-ACmix consistently outperformed YOLOv5, YOLOv8 and YOLOv10 in precision, recall, and mean Average Precision (mAP). The inclusion of the ACmix mechanism enables the model to focus more effectively on critical features, resulting in more accurate detections. Despite a slight increase in computational complexity, the significant improvements in detection performance make YOLOv10n-ACmix a robust choice for high-precision object detection in SAR images. This work highlights the potential of attention mechanisms to enhance object detection in complex, large-scale SAR image datasets.

Renewable energy has become a viable alternative to fossil fuels owing to its environmental benefits. However, its inherent uncertainty pose significant challenges. Demand response (DR) mechanisms have been developed to address these issues, facilitating renewable energy integration through consumer-side flexible resources. However, these mechanisms often affect consumer satisfaction, necessitating precise measurement and control of these impacts. In this article, we propose a two-stage electricity trading and load dispatch optimization model aimed at reducing carbon emission by promoting renewable energy accommodation, and the proposed optimization model takes into account multicategory energy consumer satisfaction. We begin by classifying consumers into distinct categories and designing tailored satisfaction functions that reflect their unique power consumption pReferences. The electricity trading and load dispatch processes are formulated as a two-stage optimization problem, which is then transformed into Markov decision processes (MDPs). A model-free framework applying two state-of-the-art deep reinforcement learning (DRL) algorithms is proposed to solve the optimization problem without requiring complex environmental modeling and prior knowledge. Numerical results demonstrate that the proposed framework outperforms benchmark algorithms regarding both consumer satisfaction preservation and carbon emission reduction.

Classical Simultaneous Localization and Mapping (SLAM) algorithms, fundamental to autonomous navigation, often assume a static environment, an assumption that fails in many real-world scenarios, resulting in significant trajectory drift. This creates a long-standing trade-off in dynamic SLAM: high-accuracy methods are typically too slow for real-time deployment, while real-time systems often suffer from reduced robustness. To address this challenge, we propose R2D-SLAM, a novel real-time visual SLAM framework that achieves both high accuracy and efficiency in dynamic environments. Our system integrates a lightweight and effective dynamic object detector, Detection Transformer (DETR), with an enhanced matching strategy, termed DEIM_Dyna, into the robust ORB-SLAM3 framework. A dedicated parallel thread performs dynamic region detection concurrently with the main SLAM pipeline, enabling early rejection of contaminated features during tracking. This ensures that pose estimation is conducted on a purified, static feature set, significantly improving robustness and accuracy. Extensive experiments on the challenging TUM and BONN dynamic datasets show that R2D-SLAM achieves new state-of-the-art performance in terms of Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), while maintaining real-time processing speeds exceeding 25 FPS. R2D-SLAM effectively bridges the gap between accuracy and efficiency, offering a practical and robust solution for dynamic scene perception in autonomous vehicles and Unmanned Aerial Vehicles (UAVs).

Radio frequency fingerprinting identification (RFFI) is a technology that uses signal processing techniques to extract the characteristics of collected wireless signals to enable the individual identification of devices. Over the past few years, deep learning (DL) has made great achievements in the field of signal identification, and the research on the RFFI methods has made great progress. To address the issue of fixed-size input data required by many neural networks, such as convolutional neural network (CNN) and multilayer perceptron (MLP), this paper proposes a transformation method that converts variable-length signals into contour stella constellation (CSC). In addition, in the most DL-based the RFFI methods, a global analysis of the entire radio frequency signal is required, which demands a large amount of calculation and is easily affected by changes in the propagation environment. We propose an incremental learning (IL)-based the RFFI approach, which enables dynamic updating of the model and improves its identification and generalization capabilities. Experimental results show that our proposed method shows excellent performance in processing data streams in real time and the identification accuracy finally can reach 89.14%. The codes of this paper can be downloaded from https://github.com/sfuaena/CSC-KD.