論文

The development of intelligent transportation systems has promoted the application of wireless technology in vehicular communication networks and mobile services. At the same time, electric vehicles are popular due to their environmental-friendliness and cost-saving characteristics. However, charging of electric vehicles without proper and real-time rules may cause serious traffic congestion and high expense, which results in unnecessary charging stations and wasted power generation with high carbon emission. Therefore, it is crucial to design an efficient charging strategy for electric vehicles without causing road congestion, together with an economic dispatch strategy to offer proper power generation and decarbonization. In this paper, we introduce mobile edge computing architecture to enhance real-time response capabilities through edge device positioning data. This approach allows charging strategies to swiftly adapt to varying road conditions and user demands, effectively mitigating traffic congestion. Regarding the forecasting of load demand for charging stations, this paper employs a deep learning model based on informer and deploys it on edge servers, quickly providing accurate demand predictions for economic dispatching and enhancing the precision of scheduling strategies. With the application of deep reinforcement learning models, the system can formulate efficient charging plans based on real-time user data, improving user satisfaction and quality of services. Numerical results show the efficiency of the obtained strategies for electric vehicles charging and power economic dispatching, and the two-stage modeling approach significantly improves the convergence of the model and the quality of the solution.

In simulating synthetic aperture radar (SAR) ship wakes, dynamic wake modeling often uses the linear superposition of sea waves and Kelvin wakes. This method, however, overlooks the alterations in sea surface roughness caused by the nonlinear interaction between waves and wakes, thus failing to accurately capture real sea surface variations. In this letter, we introduce a rapid SAR image simulation technique for ship wakes that incorporates sea waves using fluid velocity potential. Firstly, the computational domain and ship grid are constructed, with the grid scale tailored to the ship's surface structure to satisfy boundary conditions for efficient fluid velocity potential calculations. Next, to enhance boundary calculation accuracy, we employ the Taylor expansion boundary element method to swiftly resolve both steady and unsteady velocity potential components. Additionally, our approach not only depicts the interaction between sea waves and ship wakes but also facilitates the simulation analysis of various sea condition parameters. By treating the ship wake as noise and comparing images containing only background sea waves with the simulation images, the results show that the accuracy of the proposed approach is 0.2 SSIM higher than that of the linear superposition method, and the speed is 3 hours faster than that of CFD method.

In time division duplex (TDD)-based massive multiple-input multiple-output (MIMO) systems, accurate channel state information (CSI) between the base station (BS) and user terminal (UT) is crucial for efficient signal processing, including received signal separation and transmission precoding. However, due to the time-varying nature of wireless channels and the limited coherence time, the pilot signals must be short, and the number of orthogonal pilot sequences is finite. Consequently, pilot signal reuse across neighboring cells leads to pilot contamination, significantly degrading channel estimation accuracy. Traditional methods like minimum mean square error (MMSE) estimation require prior knowledge of the channel covariance matrix, which is often unavailable. This paper proposes a novel deep learning-based channel estimation approach that effectively mitigates the dual challenges of channel aging and pilot contamination. The method leverages two distinct convolutional neural networks (CNNs): one for processing pilot signals to suppress inter-cell interference and another for data signals to address both inter-cell and intra-cell interference. To further enhance estimation accuracy, a smoothing filter is employed to minimize local distortions caused by incorrect symbol detection in the time domain. Simulation results demonstrate the effectiveness of the proposed method, particularly at normalized Doppler frequencies above 0.01, where it significantly outperforms conventional techniques and interpolation-based approaches. The results highlight the proposed method's robustness in maintaining high channel estimation accuracy in dynamic and interference-prone environments.

Activity recognition is essential in computer vision applications, such as smart homes and healthcare services. While RGB images have been widely used in this area, they pose challenges related to privacy invasion and environmental constraints. To address these issues, some research has explored using 3-D light detection and ranging (3-D LiDAR) to collect 3-D point cloud data for activity recognition. However, the high-computational cost and large model parameters required for processing 3-D point clouds remain major limitations. To overcome these challenges, we propose a novel multiclass activity recognition system based on skeleton extraction from depth images collected by 3-D LiDAR. First, we use 3-D LiDAR to collect depth images of ten distinct activities, such as walking, falling, and squatting. Next, we process these depth images using our proposed multiple dynamic nonlinear mapping (MDNLM) method. The MDNLM method enhances the clarity of human body details by adjusting the color distribution of depth values based on the human position, ensuring that more colors are allocated to specific regions of the human body. This enhancement allows a fine-tuned algorithm to extract skeleton joints accurately from the mapped images. Finally, the extracted skeletons are fed into a convolutional neural network combined with a long short-term memory network (CNN+LSTM) for multiclass activity recognition. Our proposed method achieved 100.0% accuracy for a 2-class classification task (fall detection), 99.0% accuracy for a 7-class classification task, and 94.7% accuracy for a 10-class classification task.

By extracting radio frequency (RF) fingerprints from received signals, specific emitter identification (SEI) becomes a promising technique for physical layer identification of wireless devices. Recently, channel-robust SEI has attracted increasing attention due to the weak robustness exhibited by deep learning (DL)-based SEI methods in cross-channel conditions. To address these limitations, we propose a novel channel-robust SEI framework based on single-source domain generalization (SDG). Initially, we analyze the weak robustness of existing SEI methods from the perspective of the "shortcut learning"phenomenon in DL. Shortcut learning may lead traditional SEI methods to prioritize easily-mined, yet transient, channel characteristics in signal samples, rather than focusing on the more stable RF fingerprints derived from hardware differences. Next, from the perspective of SDG, we outline the optimization goal to rectify the shortcut learning in SEI. Inspired by this optimization goal, we then propose a channel-robust SEI method. This method consists of feature embedding through a multi-scale convolutional attention network (MSCAN), domain expansion using random overlay augmentation (ROA) to generate multiple virtual domains, and dual alignment strategy based on contrastive learning. Specifically, supervised contrastive learning is implemented for category-wise alignment, while supervised contrastive adversarial learning is utilized for domain-wise alignment. This dual alignment strategy can optimize the MSCAN to learn discriminative and domain-invariant feature representations, thereby enhancing the robustness of SEI. Simulation experiments on the ORACLE dataset and the WiSig dataset have demonstrated the superiority of our method compared to state-of-the-art techniques.

A significant challenge that hinders advancements in medical research is the sensitive and confidential nature of patient data in available datasets. In particular, sharing patients’ facial images poses considerable privacy risks, especially with the rise of generative artificial intelligence (AI), which could misuse such data if accessed by unauthorized parties. However, facial expressions are a valuable source of information for doctors and researchers, which creates a need for methods to derive them without compromising patient privacy or safety by exposing identifiable facial images. To address this, we present a quick, computationally efficient method for detecting action units (AUs) and their intensities—key indicators of health and emotion—using only 3D facial landmarks. Our proposed framework extracts 3D face landmarks from video recordings and employs a lightweight neural network (NN) to identify AUs and estimate AU intensities based on these landmarks. Our proposed method reaches a 79.25% F1-score in AU detection for the main AUs, and 0.66 in AU intensity estimation Root Mean Square Error (RMSE). This performance shows that it is possible for researchers to share 3D landmarks, which are far less intrusive, instead of facial images while maintaining high accuracy in AU detection. Moreover, to showcase the usefulness of our AU detection model, using the detected AUs and estimated intensities, we trained state-of-the-art Deep Learning (DL) models to detect pain. Our method reaches 91.16% accuracy in pain detection, which is not far behind the 93.14% accuracy obtained when employing a convolutional neural network (CNN) with residual blocks trained on actual images and the 92.11% accuracy obtained when employing all the ground-truth AUs.

To improve the performance metrics of power amplifier (PA) nonlinear modeling while reducing the complexity of neural networks (NN), this paper proposes a multi-channel convolutional long short-term deep neural network (MCLDNN) approach. This method is designed to efficiently extract information from both temporal and spatial dimensions. The input to the model is a two-dimensional (2D) array, which includes envelope-dependent terms and the in-phase and quadrature (I/Q) components of both current and past signals. Additionally, this model splits the 2D array into multiple one-dimensional (1D) array. We utilize 1D and 2D convolution processes on these arrays to capture local and global patterns. Long short-term memory (LSTM) is then applied to process sequential data, enabling accurate and effective extraction of temporal correlations. The experimental evaluation of PA nonlinear behavioral modeling demonstrates that the proposed method outperforms other comparative methods in normalized mean square error (NMSE) and adjacent channel error power ratio (ACEPR).

The rapid and widespread adoption of unmanned aerial vehicles (UAVs) poses significant threats to public safety and security in sensitive areas and subsequently underscores the urgent need for effective UAV surveillance solutions, where UAV classification emerges as a vital technology. Deep learning (DL) methods can autonomously extract implicit features from UAV signals and subsequently infer their types, provided that sufficient signal samples are available. Due to the high mobility of UAVs, it is challenging to ensure continuous monitoring between UAVs and the surveillance system to obtain sufficient samples. Moreover, DL models developed from sufficient but environment-specific datasets tend to be less generalized. This paper proposes a novel federated semantic regularization for learning an UAV classification model and further classifying UAVs across diverse environmental conditions. The approach enhances model generalization by regularizing semantic features during the local model training process on each participant. Subsequently, these local models are aggregated into a robust global model. Extensive testing across multiple environments demonstrates the superior classification performance of our approach compared to existing non-federated and federated approaches. The average classification accuracy of the proposed method in the three environments is 95.68%, which is improved by 13.39% compared to the non-federated methods and by 2.75% compared to the federated methods.

The deployment of deep neural networks (DNNs) in the Internet of Things (IoT) is essential for various AI applications. This trend highlights the growing need for distributed inference (DI) capabilities to process data efficiently and effectively in IoT environments. In lossy networks, data loss can significantly affect inference accuracy. Prioritizing algorithmic optimizations over hardware modifications proves to be a more effective approach for enhancing DNN performance in the IoT. We propose a plug-and-play module called Loss-Adapter, which aims to improve the accuracy of DI on lossy networks. To simulate network packet loss, we design a Gaussian distribution sampling dropout. This method is utilized during the training of the Loss-Adapter to acquire fault-tolerant inference capabilities for operation over unreliable communication links. Moreover, we employ the tree-structured Parzen estimator algorithm to optimize the parameters of the Gaussian dropout layer. We design two IoT packet loss simulation systems to conduct experiments, exploring the impact of the DNN split position and packet size on inference accuracy. Our experiments demonstrate that applying the Loss-Adapter significantly enhances the inference accuracy of DNN models operating on lossy IoT networks.

WiFi-based wireless sensing has emerged as a promising passive sensing technology that is precious for human activity recognition (HAR) across diverse applications. However, achieving robustness across varying scenarios presents a significant challenge, limiting its broader adoption. To address this issue, we propose a robust cross-scenario incremental learning (IL) method for WiFi-based wireless sensing that leverages WiFi channel state information (CSI) and elastic weight consolidation (EWC) loss. Our approach integrates a convolutional neural network and attention-based long short-term memory (CNN-ABLSTM) framework, which effectively captures the spatial and temporal features of CSI data. The IL strategy enhances model adaptability across dynamic environments, while EWC minimizes catastrophic forgetting by preserving critical weights from prior tasks. The method’s integration of a memory set and EWCLoss enables it to balance the retention of learned features with adaptation to new scenarios, effectively mitigating performance degradation across tasks. Experimental results on the MM-Fi dataset demonstrate robust cross-scenario performance: starting with initial training on scene E01, the model achieves incremental recognition in new scenes E02, E03, and E04 with cross-scenario accuracies of 88.01%, 80.16%, and 70.93%, respectively. The proposed approach substantially improves cross-scenario adaptability and test accuracy compared to traditional and cross-domain methods such as transfer learning. This work marks a significant advancement toward robust and scalable WiFi-based wireless sensing for diverse real-world applications.

WiFi-based passive noncontact sensing is widely regarded as a leading technology in wireless sensing, owing to its extensive application scope and favorable growth outlook. Nevertheless, although current WiFi-based sensing techniques attain remarkable accuracy in identifying activities within particular scenarios, they need stronger generalization capabilities across different targets and environments, hindering further commercial development. To address this issue, this article uses convolutional neural network (CNN), BLSTM, and attention layers to propose a cross-subject transfer learning method based on the CNN-ABLSTM algorithm model. This method combines widely used transfer learning methods with deep neural network algorithms in cross-domain sensing. Specifically, this method leverages the performance advantages of the CNN-ABLSTM algorithm model in processing time-series data like channel state information (CSI) and utilizes transfer learning to fine-tune the pretrained model from the source domain for application in the target domain with different subjects. This enables faster and more accurate achievement of cross-subject tasks. The simulated results show that the proposed new approach achieves higher recognition accuracy and shorter training times than traditional transfer learning methods for cross-subject tasks. In testing with the dataset used, it achieves up to around 85% performance of activity recognition accuracy in cross-subject tasks.

High Altitude Platform Stations (HAPS) have emerged as a compelling solution to extend wireless coverage for next-generation vehicular networks, thanks to their ability to hover at altitudes of around 20 km and provide robust Line-Of-Sight (LoS) links. However, the expansive coverage area of a single HAPS introduces intense co-channel interference among vehicles, particularly when traffic is dense or nonuniformly distributed. To address these challenges, this article proposes a joint angle-based user selection (AUS) strategy and an attention-based mean-field actor-critic (MF-A2C) framework for fully digital beamforming. First, AUS partitions angle-near vehicles into different groups to reduce within-group correlation. Then, each group is served by a high-dimensional beamforming vector, where an attention-based mean-field mechanism captures intergroup interference without incurring exponential complexity. Extensive simulations under realistic highway and urban mobility models demonstrate that the proposed scheme not only accelerates the convergence of the reinforcement learning process but also improves system throughput.

The rapid development of wireless communication and Internet of Things (IoT) devices has exacerbated the challenge of spectrum scarcity. Spectrum prediction plays a critical role in enhancing resource utilization and facilitating dynamic spectrum access (DSA). However, existing methods often face difficulties with missing data arising from sensor failures or environmental changes, while prioritizing prediction accuracy over computational efficiency, hindering deployment. To overcome these challenges, this paper introduces the generative augmented cascade broad learning network (GA-CBLN) for spectrum prediction. This approach combines a deep learning (DL)-based pre-trained model for data imputation with an online broad learning (BL) process to ensure efficient prediction. Initially, a DL-based regressor imputes missing values by learning spatiotemporal patterns, generating a complete dataset. Subsequently, a generative adversarial network (GAN) produces synthetic samples to augment the dataset, enhancing the model’s adaptability and generalization. Finally, the cascade broad learning network (CBLN) conducts feature extraction and prediction using both original and augmented data, exploiting the efficient learning capabilities of BL. Simulations on real spectrum datasets from four widely used frequency bands show that GA-CBLN surpasses traditional methods in computational efficiency, maintaining robust performance even under high rates of missing data.

As synthetic aperture radar (SAR) image change detection can continuously acquire target information under all weather conditions, it has been used in the past for various applications. However, it is very challenging to use SAR images for small area change detection due to the presence of noise and the imbalance between classes. To deal with these problems this article proposes a SAR images change detection scheme which is based upon an Attention Mechanism and Convolutional Wavelet Neural Network. First, employing Multiscale Superpixel Reconstructed Difference Image effectively enhances the edge information of the images. Second, a two-stage Weighted Center Constrained Fuzzy C-Means clustering algorithm is proposed. To constrain incorrect migration in the subsequent clustering phase, the algorithm uses a two-stage clustering algorithm. During the first stage weights in the clustering process to improve the accuracy of the resulting clustering centers are introduced while during the second stage, these refined cluster centers can be constrained. Third, a Convolutional Squeeze-and-Excitation Attention Mechanism Wavelet Neural Network (CSWNN) is proposed which more accurately distinguishes uncertain pixels and thus reduces redundant features. The final Denoising Diffusion Probabilistic Model is introduced into the CSWNN to expand pseudolabeled samples. Comparing this algorithm with seven sets of contrast experiments on six real SAR datasets shows that the accuracy of detecting changes in small regions is 2% –9% better than the classical algorithm.

High altitude platform station (HAPS) systems have emerged as a key solution to address the increasing networking demands of the Internet of Things (IoT), providing wide-area coverage, low latency, enhanced network resilience, and cost-effective service delivery, particularly in remote regions. Given that the continuous movement of HAPS and the inherent mobility of user equipments (UEs) often lead to low and unevenly distributed UE throughput, it is crucial for HAPS systems to dynamically control the antenna using beamforming techniques. However, the current reactive approaches to dynamic control fail to effectively minimize the number of low-throughput UEs and achieve low time complexity. To overcome these challenges, we propose a clustering-aided particle swarm optimization (PSO) algorithm to determine the antenna parameters, enabling HAPS to configure multiple cells and dynamically control beams based on UE distribution. This algorithm leverages UE clustering information to redefine the search space, reducing the complexity while enhancing the ability to find the global optimum. Specifically, we propose a novel regulated K-means algorithm that groups UEs into appropriately balanced clusters, precisely reducing the search space for global optimization. Simulations using real-world UE distributions demonstrate that our proposed method outperforms conventional approaches in reducing low-throughput UEs and providing balanced throughput distribution, while maintaining low computational complexity.

Traditional deep learning-based end-to-end (E2E) channel estimation approaches typically require extensive paired data for offline training, which is even more demanding when the signal is quantized by low-resolution analog-digital converters (ADCs), making E2E channel estimation impractical. Motivated by the above, we propose a model-driven self-supervised framework aimed at estimating channel information in massive multiple-input multiple-output systems. By applying the additive quantization noise model (AQNM) and the Bussgang decomposition theorem, we reformulate the nonlinear quantization process of low-resolution ADCs into a representation that resembles additive noise. Afterward, we integrate the self-supervised framework into the minimum mean squared error (MMSE) and Bussgang linear MMSE (BLMMSE) methods to refine channel estimation in low-resolution ADC scenarios. Notably, our approach has the advantage of not requiring true data labels, making it more adaptable to new scenarios compared to traditional E2E methods. Simulation results reveal that our approach using limited data outperforms existing supervised-learning approaches and exhibits an inherent robustness and generalization capabilities.

Alzheimer’s disease (AD) stands as a prevalent form of dementia, primarily impacting the elderly, and presently lacking any known cure. Early screening and intervention have demonstrated efficacy for both current and prospective treatments. With the global population aging, early detection of AD has become imperative. Language-based approaches for dementia detection have received considerable attention in research due to dementia’s nature of hurting cognitive and language functions. Facial features, on the other hand, have been comparatively neglected despite their potential in identifying AD. The face mesh, comprising 478 landmarks, serves as a rich and informative facial representation. This study concentrates on leveraging face mesh data extracted from the PROMPT dataset for AD detection. The dataset comprises 445 voice and video recordings from 117 dementia patients and healthy individuals collected during professionally conducted interview sessions. We explored data augmentation techniques to address the challenges of over-fitting, particularly impactful in training neural networks on a small dataset. Deep learning methods, including transformer, Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks, were employed to classify dementia patients and healthy subjects. By augmenting the dataset, we improved the performance of the deep learning methods. The combined use of 1D CNN networks and data augmentation method reported the highest accuracy of 76%, which is the state-of-the-art accuracy in dementia detection using face mesh. These results confirmed the effectiveness of our data augmentation methods. They also underscore the potential of face-related features in enhancing early detection approaches for AD.