論文

With the increase in demand for Internet of Things (IoT) applications, the number of IoT devices has drastically grown, making spectrum resources seriously insufficient. Transmission collisions and retransmissions increase power consumption. Therefore, even in long-range (LoRa) networks, selecting appropriate transmission parameters, such as channel and transmission power, is essential to improve energy efficiency. However, due to the limited computational ability and memory, traditional transmission parameter selection methods for LoRa networks are challenging to implement on LoRa devices. To solve this problem, a distributed reinforcement learning-based channel and transmission power selection method is proposed, which can be implemented on the LoRa devices to improve energy efficiency in this paper. Specifically, the channel and transmission power selection problem in LoRa networks is first mapped to the multi-armed-bandit (MAB) problem. Then, an MAB-based method is introduced to solve the formulated transmission parameter selection problem based on the acknowledgment (ACK) packet and the power consumption for data transmission of the LoRa device. The performance of the proposed method is evaluated by the constructed actual LoRa network. Experimental results show that the proposed method performs better than fixed assignment, adaptive data rate low-complexity (ADR-Lite), and e-greedy-based methods in terms of both transmission success rate and energy efficiency.

The rapid growth of Long Range (LoRa) devices has led to network congestion, reducing spectrum and energy efficiency. To address this problem, we propose an energy-efficient reinforcement learning method for distributed LoRa networks, enabling each device to independently select appropriate transmission parameters, i.e., channel, transmission power (TP), and bandwidth (BW) based on acknowledgment (ACK) feedback and energy consumption. Our method employs the Upper Confidence Bound (UCB)1-tuned algorithm and incorporates energy metrics into the reward function, achieving lower power consumption and high transmission success rates. Designed to be lightweight for resource-constrained IoT devices, it was implemented on real LoRa hardware and tested in dense network scenarios. Experimental results show that the proposed method outperforms fixed allocation, adaptive data rate low-complexity (ADR-Lite), and (Formula presented.) -greedy methods in both transmission success rate and energy efficiency.

The rapid changes in high-mobility vehicle environments make it challenging for base stations (BS) to obtain comprehensive channel state information. Furthermore, road and traffic safety require communication with low latency and high reliability, posing significant challenges to spectrum resource allocation in vehicular networks. To address these challenges, this paper proposes a method combining dueling double deep-Q network (D3QN) reinforcement learning (RL) with long short term memory (LSTM) network. By using a Manhattan Grid Layout City Model as the foundational environment, a multi-agent model is constructed, with each vehicle-to-vehicle (V2V) link acting as an individual agent. These agents collaborate and interact with the environment, receiving feedback, and then determining the optimal resource allocation to ensure both superior mobile service and a safe driving environment. The experimental results indicate that our proposed method outperforms the conventional D3QN network in both the vehicle-to-infrastructure (V2I) links and the V2V links.

The increasing demand for spectral efficiency and system capacity in communication networks has driven the integration of Non-Orthogonal Multiple Access (NOMA) technology with Unmanned Aerial Vehicle (UAV) networks. In this paper, we propose an enhanced user clustering and pairing scheme for NOMA-aided UAV networks. Our study aims to maximize the minimum user throughput by proposing an Advanced Balanced K-Means (ABKM) algorithm, based on the traditional K-Means (KM) and Balanced K-Means (BKM) algorithms. The ABKM algorithm addresses the issue in the BKM algorithm where some users are assigned to sub-optimal clusters, resulting in increased distances to the cluster centroids, while simultaneously ensuring balanced clustering. Through extensive numerical simulations, we demonstrate that the ABKM algorithm significantly outperforms KM and BKM algorithms in terms of the minimum user throughput. The results highlight the potential of the proposed ABKM algorithm to enhance system performance and ensure fair resource allocation in NOMA-aided UAV networks, making it a promising solution for future 6G wireless communication systems.

Automatic modulation classification (AMC) is an essential and meaningful technology in the development of cognitive radio. It can judge the modulation mode according to the signal acquired by the receiver. In recent years, the deep learning (DL) method has been used to take the place of modulation signal recognition based on decision theory and pattern recognition, which has achieved very effective results. The development of the neural network classification model focuses on architectural engineering. Discovering state-of-the-art neural network architectures requires substantial prior knowledge and effort of human experts. Neural architecture search (NAS) can be viewed as a subdomain of automatic machine learning (AutoML), which uses a neural network to automatically adjust the structures and parameters to obtain a network that researchers need by following search strategies that maximize performance. In this paper, we propose a differentiable architecture search (DARTS) based AMC method. In addition, we also consider six other methods, including convolutional neural network (CNN), simple recurrent unit (SRU), a convolutional-recurrent neural network (CRFN-CSS), Residual Networks (ResNet), Inception Modules (Inception) and MobileNet. Simulation results show that the proposed method can achieve the optimal classification accuracy at low parameters and floating-point operations (FLOPs) without manual architecture engineering.

Accurate detection of surface defects on railway tracks is critical for safe railway operation. Most existing models rely solely on red-green-blue (RGB) images, limiting their ability to capture structural information. Incorporating depth features provides richer spatial cues, significantly improving detection accuracy. However, current RGB and depth (RGB-D) dual-stream models suffer from high computational complexity and hardware dependencies, making them impractical for real-world deployment. To address these limitations, we propose DAHNet, an asymmetric knowledge distillation model with a teacher-student architecture. DAHNet-T serves as the teacher network, taking RGB-D inputs and integrating a cross-modal attention feature enhancer (CAFE) module to capture contextual information, along with a depthwise feature interaction block (DFIB) for efficient cross-modal fusion. DAHNet-S is the student network, a lightweight single-stream RGB model employing depthwise separable convolutions to reduce computation. We introduce a multilevel distillation strategy with dynamic temperature scaling to balance coarse-grained and fine-grained knowledge transfer, while incorporating contrastive learning and structural loss to improve pixel-level accuracy. Extensive experiments on the NEU RSDDS-AUG dataset demonstrate that our distilled model DAHNet-KD outperforms state-of-the-art methods. Compared to DAHNet-T, the number of parameters is reduced from 87.72 to 13.97 MParams, and the computational cost decreases from 19.79 to 5.41 GFLOPs. The proposed model achieves superior performance across various evaluation metrics and also generalizes well on other public datasets. Therefore, the model provides a lightweight and high-accuracy solution for deployment on mobile devices in real-world industrial scenarios.

In dynamic and cluttered environments, traditional simultaneous localization and mapping (SLAM) systems often suffer from degraded localization accuracy and unstable map construction due to the presence of moving objects and occlusions. To address these challenges, we propose DFusion-SLAM, a lightweight and robust SLAM framework that integrates an enhanced object detection module into ORB-SLAM3. The detection module is based on an improved D-Fine architecture, in which the original Transformer is replaced with a more efficient PolaLinearAttention mechanism. Furthermore, a MetaFormer-based semantic fusion structure is introduced to strengthen multiscale feature representation. These architectural improvements jointly enhance detection accuracy while reducing model complexity, achieving a performance increase from 42.8% to 43.7% mean average precision (mAP). Experimental evaluations on dynamic RGB-D sequences from the TUM and Bonn datasets demonstrate that DFusion-SLAM significantly improves localization accuracy and mapping stability under dynamic conditions, while maintaining high computational efficiency. These results highlight the framework's strong potential for real-time deployment in IoT-oriented mobile and robotic platforms operating in complex environments.

The deep autoencoder (DAE) framework has turned out to be efficient in reducing the channel state information (CSI) feedback overhead in massive Multiple-Input Multiple-Output (mMIMO) systems. However, these DAE approaches presented in prior works rely heavily on large-scale data collected through the base station (BS) for model training, thus rendering excessive bandwidth usage and data privacy issues, particularly for mMIMO systems. When considering users’ mobility and encountering new channel environments, the existing CSI feedback models may often need to be retrained. Returning back to previous environments, however, will make these models perform poorly and face the risk of catastrophic forgetting. To solve the above challenging problems, we propose a novel gossiping generative adversarial network (Gossip-GAN)-aided CSI feedback training framework. Notably, Gossip-GAN enables the CSI feedback training with low-overhead while preserving users’ privacy. Specially, each user collects a small amount of data to train a GAN model. Meanwhile, a fully distributed gossip learning strategy is exploited to avoid model overfitting, and to accelerate the model training as well. Simulation results demonstrate that Gossip-GAN can: 1) achieve a similar CSI feedback accuracy as centralized training with real-world datasets; 2) address catastrophic forgetting challenges in mobile scenarios, and 3) greatly reduce the uplink bandwidth usage. Besides, our results show that the proposed approach possesses an inherent robustness.

Economic dispatch, as a crucial method for ensuring the normal operation of power systems, is typically modeled as an optimization problem and solved using solvers. The introduction of low-carbon requirements has increased the complexity of the optimization problem, as economic dispatch now needs to consider effectively reducing the emission of CO2. The utilization of renewable energy can mitigate carbon emissions during power system operation, but its inherent uncertainty poses safety risks to transmission lines between microgrids and the main grid. Hence, this paper explores the issue of low-carbon economic dispatch to ensure the reliability of power transmission. We introduce a two-stage low-carbon economic dispatch model grounded in deep reinforcement learning. In the first stage, we incorporate the transmission power limits between microgrids and the main grid as constraints, generate day-ahead dispatch strategies, and determine the weights of factors in the reward function. In the second stage, we employ two methods: adjusting the weight of the penalty for exceeding the transmission limits and adding conservative safety limit constraints, to solve the problem of transmission power exceeding limits caused by uncertainties of renewable energy and prediction errors. Three deep reinforcement learning algorithms, all rooted in the Actor–Critic structure, are employed to implement the two-stage economic dispatch model. Experimental results affirm the efficacy of the proposed model in mitigating the risk of transmission power exceeding its limit, reducing carbon emissions, and minimizing operational costs.

The Deep Transfer Learning (DTL)-based Channel State Information (CSI) approach exploits the Deep Neural Network (DNN) model to provide low-cost CSI feedback for the target channel. It involves offline training on a given CSI dataset, followed by fine-tuning for a new environment using a small amount of collected CSI data, reducing feedback costs. However, in practical scenarios, the fine-tuned model performs worse on the source channel compared to the original source model. This leads to continued use of the poorly performing target model until it's fine-tuned again with newly acquired data. To address this, we propose combining DTL-based CSI feedback with continual learning. We introduce elastic weight consolidation (EWC) into the loss function during fine tuning. Simulations show that our method significantly reduces the degradation of the target model on the source channel, as measured by NMSE, compared to a method without continual learning.

Object detection in remote-sensing imagery plays a crucial role in providing precise geospatial information for urban planning and environmental monitoring. However, real-world remote-sensing scenarios often involve complex conditions, such as varying illumination, weather interference, and low signal-to-noise ratios, which significantly degrade the performance of traditional single-modal detection methods. To overcome these limitations, multimodal object detection has developed, demonstrating great potential by integrating complementary information from multiple modalities. Nevertheless, existing multimodal frameworks still face challenges, such as insufficient cross-modal interaction, limited learning of complementary features, and high computational costs, due to redundant fusion in complex environments. To overcome these challenges, we propose an enhanced multimodal fusion strategy aimed at maximizing cross-modal feature learning capabilities. Our method employs a dual-backbone architecture to extract mode-specific representations independently, integrating a direction attention module at an early stage of each backbone to enhance discriminative feature extraction. We then introduce a dual-stream feature fusion network to effectively fuse cross-modal features, generating rich representations for the detection head. In addition, we embed a local–global channel attention mechanism in the head stage to strengthen feature learning in the channel dimension before generating the final prediction. Extensive experiments on the widely used vehicle detection in aerial imagery multimodal remote-sensing dataset demonstrate that our method achieves state-of-the-art performance, while evaluations on single-modal datasets confirm its exceptional generalization capability.

Remote sensing image (RSI) object detection is crucial for natural disaster management, urban planning, and resource exploration. However, due to significant differences between RSIs and natural images (NIs), existing object detectors designed for NIs are not directly applicable to RSIs. CNN-based models typically require specialized attention modules to relate small targets to global contexts, while Transformer-based models often add modules to enhance detail extraction, increasing computational costs for edge deployment. To address these challenges, we propose DyFormer, a hybrid CNN-Transformer model that combines local and global attention mechanisms for effective feature fusion. Additionally, a feature residual pyramid network (FRPN) is introduced to enhance multi-scale feature integration. We conduct experiments on the open source NWPU VHR-10 dataset to evaluate the proposed DyFormer, achieving a mAP@0.5 of 61.7%, outperforms existing sort-of-the-art methods, demonstrating its effectiveness for RSI object detection tasks.

In this paper, we propose a Mean Field reinforcement learning (MFRL) method for dynamic antenna control in High-Altitude-Platform-Station (HAPS) communication system with Multi-Cell Configuration. HAPS works at stratospheric altitudes of about 20 km to provide an ultra-wide coverage area. However, the wind pressure caused HAPS movement leads to the degradation of users' throughput. Considering the multi-antenna arrays in the HAPS, to find the optimal antenna parameters of all antenna arrays for reducing the number of low-throughput users, we formulate the antenna control problem into stochastic game equilibrium. Usually, solving the stochastic to find the equilibrium needs very high computation complexity to calculate the transition probability for getting the Q-value under a certain state and action. Therefore, we use the reinforcement learning (RL) named Deep Q-Network (DQN) to learn the transition probability and predict the Q-value according to the reward fed backed from the environment. Besides, we employ the Mean field Game theory in conjunction with RL during the training phase of DQN to reduce the complexity of the interactions among agents. To evaluate the proposed method, we compare the proposed method with a genetic algorithm (GA) named Particle Swarm Optimization (PSO), Q-learning, Fuzzy Q-learning, and conventional DQN under four realistic user distribution scenarios. The simulation results show that the proposed method achieves comparable throughput performance with a high convergence rate.

Electrocardiogram (ECG) measurement is a fundamental diagnostic and monitoring tool in cardiology and healthcare. It plays a crucial role in identifying and managing cardiac conditions, assessing heart function, and improving patient outcomes through timely interventions and treatment adjustments. Conventional approaches for ECG measurement are often unsuitable for everyday use due to their invasiveness, inaccessibility, or limited accuracy. In this context, we introduce ECGDiff, a novel approach based on conditional denoising diffusion models for ECG signals reconstruction from Doppler signals. ECGDiff provides a non-intrusive and accessible method for non-contact ECG measurement. Our method represents a pioneering use of denoising diffusion models for time series trans-lation. In the proposed framework, ECG signals are iteratively reconstructed from random noise using preprocessed signals derived from a Doppler sensor as input conditions. Extensive experiments have been conducted on a dataset comprising 19 healthy individuals, consisting of pairs of synchronized Doppler and ECG signals. Our experimental results demonstrate that the proposed ECGDiff outperforms other state-of-the-art methods by a large margin for the task of ECG signals reconstruction. Specifically, we achieved a DTW score of 8.32, a Frechet distance score of 3.10, and a Pearson correlation coefficient of 0.93. These results highlight the effectiveness of our approach for reconstructing ECG signals from Doppler signals.

Depression has a significant impact on society. Early detection of depression is essential because timely treatment can mitigate its negative impact. Recent studies have reported that models based on Large Language Models (LLMs), such as Chat GPT-4, can achieve high accuracy in detecting depression when prompt-engineering techniques are applied. In contrast, fine-tuning methods require more memory for training and have yet to demonstrate comparable accuracy to prompting-engineering in detecting depression from conversations. The objective of this study is to achieve higher accuracy than prompt-engineered models with low GPU memory through fine-tuning. To accomplish this, we first quantized the parameters of the LLM to minimize memory requirements during training. Instead of adding a classification layer, we retained the original structure of the LLM and trained it as a generative model. Our evaluation on the DAIC-WOZ dataset demonstrated an F1 score of 84% for depression detection with low GPU memory, which is superior to the performance of prompt-engineered models. Additionally, when tested on the private PROMPT dataset, the proposed model achieved an F1 score of 82%. This result suggests that the model's strong performance is not solely dependent on publicly available data that may have been included in the pre-training.

Unlike conventional techniques relying typically on RGB and depth cameras, millimeter wave (mmWave) radars, especially Frequency-Modulated Continuous Wave (FMCW) radar, have significant advantages in human monitoring. These advantages are more pronounced when detecting human joint positions, as FMCW radars rely solely on collected point cloud data. Compared with camera-based systems, this technology is efficient because it can operate in low-light conditions, and does not present a privacy threat. However, conventional approaches employing FMCW radar, such as voxel-based gathering techniques, still have limitations in optimally capturing spatiotemporal features, making it difficult to detect human joint points with high accuracy. To overcome this obstacle, we propose a novel framework that introduces sequential concatenation of multiple frames in the preprocessing stage to maintain spatiotemporal continuity, thereby improving feature representation. Furthermore, we design a deep learning architecture consisting of Convolutional Neural Network (CNN) for spatial feature extraction, Transformer for capturing long-term dependencies, and Bidirectional Long Short-Term Memory (Bi-LSTM) for temporal sequence modeling. The proposed approach makes optimal use of spatiotemporal data, improves the accuracy and reliability in detecting human joint positions, and accelerates model convergence. Through experimental validation, this method shows a Mean Absolute Error (MAE) of 1.77 cm and a Root Mean Squared Error (RMSE) of 2.92 cm between the actual human joint positions and the predicted ones. This reaffirms the effectiveness of the model in estimating human joint positions.

Person identification based on radar-extracted vital signs has become increasingly popular due to its non-contact measurement capabilities. This paper introduces a novel deep learning-based person identification algorithm leveraging radar- extracted vital signs. While current studies mainly focus on closeset conditions with consistent training and testing categories, real-world scenarios often involve open-set circumstances, in which there are more data categories in the testing data. The algorithm involves extracting heart pulse signals from Doppler radar echoes, training two Convolutional Neural Network (CNN)-based models using transfer learning, and utilizing a distribution model for calibration. By combining the models' outputs through a strategic decision-making process, we achieve superior person identification results. Experimental results on a public radar vital signs dataset demonstrate an identification accuracy of 99.61% in close-set conditions and 94.35% in openset conditions, surpassing existing approaches.

To reduce the outage of the connection between unmanned aerial vehicles (UAVs) and cellular networks in complex real-time channel state, and reduce the energy consumption of UAV during flight mission, an online trajectory optimization scheme of UAV based on outage probability knowledge map reconstruction is proposed. The outage probability knowledge map is a database that simulates the connection between UAV and the cellular network during real hovers. The UAV first samples sparsely from the target area and calculates the outage probability of the sampling point, and then uses the Kriging algorithm to reconstruct the outage probability knowledge map. Based on the reconstructed outage probability knowledge map, with the goal of minimizing the energy consumption of UAV task execution, the UAV trajectory optimization problem is established, and a trajectory optimization algorithm based on deep reinforcement learning (DRL) is proposed to solve it. Numerical results show that the proposed online trajectory optimization scheme based on outage probability knowledge map can obtain great returns in terms of maintaining connectivity, reducing task completion time and energy consumption.

Deep steganalyzer combined with neural networks has achieved great success in image classification over recent years. However, it suffers from the following persistent challenges: i) Deep steganalyzer is extremely vulnerable and has the risk of being attacked via adversarial steganography when performing the image classification tasks; ii) Pre-processing based methods aiming to remove adversarial perturbations from cover images jeopardize the accuracy performance, as the involved steganographic signal will be wiped off as well. In this context, to defend against adversarial attacks by an adversary, we propose an adversarial steganography detection scheme based on the pre-processing and feature migration. In brief, sub-images are sampled to obtain the dimensionality of the extracted features, which are usually used to expand them while reducing the effect brought by adversarial perturbations. In particular, by computing statistical features together with normalizing the features, our approach can improve the classification accuracy of the samples. Our experimental results show that the proposed approach is capable of detecting adversarial steganographic image with an accuracy gain of up to 35.9% over the state-of-the-art methods.

The emerging technology of simultaneous transmitting and reflecting reconfigurable intelligent surfaces (STAR-RIS) enables 360° full-space signal coverage, making it a promising approach for resilient and sustainable smart cities in next-generation wireless networks. This work investigates long-term downlink communication in STAR-RIS-assisted vehicular networks, aiming to maximize the sum-rate while ensuring queue stability under high vehicle mobility. To model the stochastic variations arising from vehicle mobility and random data arrivals, we employ a Lyapunov optimization framework, which reformulates the long-term stochastic objective into a sequence of per-slot deterministic problems. This method is effective in vehicular networks, as it adapts decisions to rapidly changing channel conditions while guaranteeing queue stability without requiring mobility predictions. A joint optimization strategy is developed, wherein the roadside unit (RSU) beamforming vectors and the STAR-RIS phase-shift coefficients are alternately optimized using zeroforcing beamforming and projected gradient descent (PGD), respectively. Extensive simulations demonstrate that the proposed framework significantly improves the long-term throughput, fairness, and queue stability compared with five benchmark STAR-RIS schemes. The results remain robust across different mobility speeds, transmit powers, and STAR-RIS element configurations, confirming the effectiveness of the proposed approach in dynamic Vehicle-to-Everything (V2X) environments.