論文

In frequency division duplexing (FDD) massive multiple-input multiple-output (MIMO), the downlink channel state information (CSI) feedback method based on deep transfer learning (DTL) has been proposed to obtain the downlink CSI at the Base Station (BS). In the CSI feedback method based on DTL, a target model for one channel environment is obtained by fine-tuning the parameters of a source model trained on a large number of the CSI dataset (source data) of another channel environment. The fine-tuning is done with a small number of the CSI dataset (target data) of the target channel environment. Thus, a target model can be obtained at a low learning cost. However, the performance of the target model could highly depend on the source data. In this paper, we investigate two metrics as criteria for selecting source data to obtain a target model with a high CSI reconstruction performance: (i) Jensen-Shannon Divergence (JSD), which represents the similarity between target and source data, and (ii) entropy, which represents the diversity of source data. The simulation results showed when the target channel model is non line-of-sight (NLOS), the source data with high entropy and low JSD tend to provide higher CSI reconstruction performance of the target model. These results indicate that the JSD and the entropy could be a source data selection metric.

Arrhythmia is one of the causes of sudden infant death, and it is very important to detect fetal arrhythmia for fetal well-being. Fetal electrocardiogram (FECG) is one of the methods to detect a heartbeat. Fetal arrhythmia can be detected based on the heartbeat detection results from FECG signals such as heartbeat intervals. However, the accuracy of arrhythmia detection easily degrades depending on the accuracy of heartbeat detection. In this paper, we propose a deep learning-based fetal arrhythmia detection method using FECG signals. Recently, arrhythmia detection methods using adult ECG signals have achieved a high arrhythmia detection accuracy based on deep learning. Motivated by this fact, in the proposed method, the acquired FECG signals are segmented, and the segments are input into a deep learning model that classifies them into normal or arrhythmia ones. Based on the classification results of multiple segments, a subject is judged as a healthy or arrhythmia subject. Each segment of the training data is divided into three categories based on the estimated heartbeat interval: (i) normal, (ii) arrhythmia, and (iii) a segment that could be both normal and arrhythmia. Only segments labeled as normal or arrhythmia are used for training a deep learning model to achieve a higher classification accuracy of the model. Through these procedures, the proposed method detects fetal arrhythmia with fewer effects of heartbeat detection results. The experimental results showed that the proposed method could outperform the conventional methods with heartbeat detection and feature detection in terms of accuracy, specificity, and recall.

Fetal heart rate monitoring using the abdominal electrocardiograph (ECG) is an important topic for the diagnosis of heart defects. Many studies on fetal heart rate detection have been presented, however, their accuracy is still unsatisfactory. That is because the fetal ECG waveform is contaminated by maternal ECG interference, muscle contractions, and motion artifacts. One of the conventional methods is to detect the R-peaks from the integrated power of the frequency corresponding to the fetal heartbeats. However, the detection accuracy of the R-peaks is not enough. In this paper, we propose a method to generate the candidates of R-peaks using the first derivative of the signal and to pick up the estimated heartbeats by a multiple weighting function. The proposed multiple weighting function is designed by the Gaussian distribution, of which parameters are set from a grid search with the goal of minimizing the standard deviation of RR intervals (neighboring R-peaks intervals). The validation for the proposed framework has been evaluated on real-world data, which got the better accuracy than the conventional method that detects R-peaks from the integrated power and uses the weighting function produced by a fixed parameter of Gaussian distribution [12]. The averaged absolute error (AAE) which compares the estimated fetal heart rate and the reference fetal heart rate has been decreased by 17.528 bpm.

In this paper, we address the challenging task of estimating the distance between different users in a Millimeter Wave (mmWave) massive Multiple-Input Multiple-Output (mMIMO) system. The conventional Time of Arrival (ToA) and Angle of Arrival (AoA) based methods need users under the Line-of-Sight (LoS) scenario. Under the Non-LoS (NLoS) scenario, the fingerprint-based method can extract the fingerprint that includes the location information of users from the channel state information (CSI). However, high accuracy CSI estimation involves a huge overhead and high computational complexity. Thus, we design a new type of fingerprint generated by beam sweeping. In other words, we do not have to know the CSI to generate fingerprint. In general, each user can record the Received Signal Strength Indicator (RSSI) of the received beams by performing beam sweeping. Such measured RSSI values, formatted in a matrix, could be seen as beam energy image containing the angle and location information. However, we do not use the beam energy image as the fingerprint directly. Instead, we use the difference between two beam energy images as the fingerprint to train a Deep Neural Network (DNN) that learns the relationship between the fingerprints and the distance between these two users. Because the proposed fingerprint is rich in terms of the users’ location information, the DNN can easily learn the relationship between the difference between two beam energy images and the distance between those two users. We term it as the DNN-based inter-user distance (IUD) estimation method. Nonetheless, we investigate the possibility of using a super-resolution network to reduce the involved beam sweeping overhead. Using super-resolution to increase the resolution of low-resolution beam energy images obtained by the wide beam sweeping for IUD estimation can facilitate considerate improvement in accuracy performance. We evaluate the proposed DNN-based IUD estimation method by using original images of resolution 4 × 4, 8 × 8, and 16 × 16. Simulation results show that our method can achieve an average distance estimation error equal to 0.13 m for a coverage area of 60 × 30 m2. Moreover, our method outperforms the state-of-the-art IUD estimation methods that rely on users’ location information.

Wi-Fi-based Human Activity Recognition (HAR) using Channel State Information (CSI) has received significant attention due to its non-intrusive nature and wide applicability in smart cities, healthcare, and smart home systems. Despite its potential, real-world deployment faces a major challenge: ensuring the model's ability to generalize across different users, who exhibit distinct behaviors and encounter diverse environmental factors. This issue, termed Cross-User Generalization (CUG), arises from distribution shifts caused by variations in users' body shapes, postures, movements, and environmental conditions. To address this challenge, we propose a cross-domain generalization framework based on an Attention-based Bidirectional Long Short-Term Memory (ABLSTM) network. The proposed framework enhances the generalization capabilities of HAR models by leveraging multi-source training and incorporating attention mechanisms for effective temporal feature learning. Specifically, the model learns user-invariant features across diverse training users and adapts to new users without requiring direct exposure to their data. Extensive experiments on a multi-user CSI dataset demonstrate that the ABLSTM model improves average test accuracy by 2.0%-6.7% compared to BiLSTM and GRU models, and over 18.0% compared to MLP-based models, while achieving the lowest training and testing loss. This methodology offers a robust and scalable solution for Wi-Fi-based HAR, promoting reliable deployment in complex, multi-user environments encountered in smart cities, healthcare monitoring, and security surveillance.

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, due to a lack of usable dataset. AD patients show different patterns while displaying facial expressions because of Apathy, a phenomenon common in AD patients. This makes facial features non-negligible in dementia detection. The face mesh, comprising 478 landmarks, serves as a rich and informative facial representation. This study concentrates on leveraging face mesh data for AD detection. 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 by rotation, shift, and scaling, 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%. 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.

With the rapid increase in Internet of Things (IoT) devices, packet collision has become a serious problem in Long Range (LoRa) communication. Furthermore, the rising demand for mobile IoT applications necessitates parameter allocation methods that take into account the mobility of End Devices (EDs). However, the existing transmission parameter selection methods considering mobility are almost centralized designs, which may increase the probability of interference and system latency. The decentralized design is few and requires prior information on ED, which may cause scalability issues, e.g., adding ED to an existing network. To solve the problems above, a method through a fully autonomous distributed design supposing the mobility of ED using Double Deep Q Network (DDQN) is proposed in this paper. In the proposed method, ED trains the model based on ACKnowledge (ACK) packets and its location information without any prior information. Performance evaluation results show that the proposed method can achieve a higher packet delivery rate than other autonomous distributed methods in mobile IoT scenarios.

With the advancements in wireless communication and the growth of Internet of Things (IoT), spectrum scarcity has become a key bottleneck. Spectrum prediction enhances resource utilization and supports dynamic spectrum access (DSA). However, current models struggle with missing data and prioritize accuracy over efficiency, limiting practical use. To address this, we propose a generative augmented cascade broad learning network (GA-CBLN) for spectrum prediction, combining a deep learning (DL) pre-trained model with an online broad learning (BL) process. First, a DL-based regressor fills missing values by learning spatiotemporal patterns. Next, a generative adversarial network (GAN) generates synthetic samples to enrich the training set, improving adaptability. Finally, the cascade BL network (CBLN) uses the combined data for efficient prediction. Simulations on real datasets show that GA-CBLN significantly outperforms traditional methods in accuracy and efficiency, even with high missing data rates.

Accurate behavioral modeling of power amplifiers (PAs) is crucial for digital pre-distortion and energy-efficient vehicular wireless transceivers, yet remains challenging due to bandwidth-dependent nonlinearities and fast-varying drive conditions in practical RF front-ends. Conventional PA models struggle to capture coupled temporal memory effects and spatial signal dependencies, limiting their effectiveness in broadband vehicular transmitters. To address this issue, we propose a Kolmogorov–Arnold Convolutional Liquid Neural Network (KACLN) that jointly leverages Kolmogorov–Arnold representation learning and liquid neuron adaptive dynamics to model nonlinear distortion with strong temporal generalization. The proposed KACLN integrates one-dimensional and two-dimensional convolutional operator kernels to simultaneously characterize long- and short-term PA memory, frequency-selective distortion patterns, and cross-sample bandwidth interactions. Experiments over diverse vehicular RF operating regimes demonstrate that KACLN improves modeling precision and robustness, achieving 1.789 dB normalized mean square error (NMSE) reduction and 2.639 dB adjacent channel error power ratio (ACEPR) suppression gains compared with the best performed MCLDNN. The proposed KACLN further mitigates efficiency degradation associated with static power back-off, enhancing effective output utilization under fast-varying channel and wideband signal envelopes. These results validate KACLN as a practical, real-time capable nonlinear surrogate suitable for next-generation vehicular PA linearization, DPD-aided waveform shaping, and low-latency wireless transmitter design.

Deep autoencoder (DAE) frameworks have demonstrated their effectiveness in reducing channel state information (CSI) feedback overhead in massive multiple-input multiple-output (mMIMO) orthogonal frequency division multiplexing (OFDM) systems. However, existing CSI feedback models struggle to adapt to dynamic environments caused by user mobility, requiring retraining when encountering new CSI distributions. Moreover, returning to previously encountered environments often leads to performance degradation due to catastrophic forgetting. Continual learning involves enabling models to incorporate new information while maintaining performance on previously learned tasks. To address these challenges, we propose a generative adversarial network (GAN)-based learning approach for CSI feedback. By using a GAN generator as a memory unit, our method preserves knowledge from past environments and ensures consistently high performance across diverse scenarios without forgetting. Simulation results show that the proposed approach enhances the generalization capability of the DAE framework while maintaining low memory overhead. Furthermore, it can be seamlessly integrated with other advanced CSI feedback models, highlighting its robustness and adaptability.

Due to the triple mobility of mobile users (MUs), unmanned aerial vehicle (UAV) relays, and low Earth orbit (LEO) satellites, handover becomes a critical and challenging issue for maintaining the continuity and quality of communication services in UAV-assisted LEO satellite networks. This article proposes a distributed handover decision-making process aimed at improving scalability and reducing communication overhead. The handover problem is modeled as a decentralized Markov decision process (DEC-MDP) with the objective of maximizing the total end-to-end (E2E) throughput. We design an independent proximal policy optimization-based distributed intelligent handover (IPPO-DIH) algorithm within a centralized training with decentralized execution framework to solve the DEC-MDP. To analyze the theoretical optimal E2E throughput, we eliminate the correlation between handover decisions at different time steps. A three-sided matching algorithm with theoretical convergence guarantees is designed to obtain a stable matching among MUs, UAV relays, and LEO satellites at each time step. These stable matchings are combined to provide a theoretical performance benchmark for the handover algorithms. Simulation results validate the convergence of the proposed IPPO-DIH and three-sided matching algorithms. Additionally, the total E2E throughput achieved by the IPPO-DIH algorithm approaches the theoretical performance benchmark and outperforms typical handover algorithms.

High-altitude platform stations (HAPSs), which directly deliver communication services to smartphones on the ground, are attracting much attention as novel mobile communication platforms for ultrawide coverage areas. A multi-cell configuration is needed to enhance communication capacity in wide areas. Assuming the use of a phased-array antenna, we proposed an optimization method for controlling antenna parameters, such as beamwidth and beam direction, to improve spectral efficiency. HAPSs are also promising disaster-resilient networks, as they can recover coverage immediately when coverage holes appear due to the breaking down of terrestrial base stations (BSs). Users can recover connections immediately during disasters through HAPSs that use the same frequency band as terrestrial networks. However, in this scenario, HAPSs become a source of interference in terrestrial networks, so network operators should avoid affecting existing terrestrial networks through HAPSs. However, cell optimization in this coexistence scenario between HAPSs and terrestrial networks has not been investigated thoroughly. In this paper, we propose a design scheme that optimizes antenna parameters for each cell to maximize HAPS coverage in this coexistence scenario using a genetic algorithm (GA). The proposed method can effectively protect terrestrial networks through optimization by setting the initial direction of each beam to avoid terrestrial BSs. Using this scheme, the GA can start optimization with candidate combinations that satisfy the interference constraints of terrestrial networks, which can expedite optimization for better coverage. Simulation results show that the proposed scheme can achieve better coverage than the conventional scheme, which only controls beam direction. Furthermore, they show that the initial beam direction setting effectively reduces the number of performed combinations required for GA to converge.

High altitude platform station (HAPS) is a platform that provides direct communication to wireless devices such as smart phones by equipping unmanned aerial vehicles (UAVs) flying in the stratosphere with cellular base station functions. In this paper, we propose a HAPS trajectory optimization method based on a throughput heat map, where the HAPS trajectory is designed considering the trajectory constraints of HAPS as an aircraft. The HAPS trajectory is optimized according to the throughput characteristics to be improved, such as the average throughput of each UE (user equipment). Computer simulations using actual UE distributions for Tokyo, Osaka, and Nagoya showed that the sub-optimized figure-8 trajectory improved the 50th percentile throughput by 15-16% and the 5th percentile throughput by 3-20% for each city, respectively, compared to the non-optimized circular trajectory. Trajectories aimed at improving the 50th, 75th, and 95th percentiles improved throughput characteristics around the 50th percentile, around the 75th-95th percentile, and above the 96th percentile, respectively.

Accurate heart rate estimation using Doppler radar and Frequency Modulated Continuous Wave (FMCW) radar is highly valued for privacy protection and the ability to measure through clothing. Conventional methods struggle to isolate the heartbeat from respiration and body motion. This paper introduces a novel heart rate estimation method using Variational Mode Decomposition (VMD) via Multiple-Input Multiple-Output (MIMO) FMCW radar. The proposed method first estimates human positions within the radar’s coverage, reducing noise by focusing on signals from these positions. The signal is then decomposed into multiple Intrinsic Mode Function (IMF) signals using VMD, and the heartbeat-specific IMF is extracted based on its center frequency. The heart rate signal is reconstructed using weighted addition of IMF signals for each radar cell, with cells defined by specific angles and distances within the coverage area. Peak detection is used to estimate heart rate from these reconstructed signals. To ensure accuracy, the method selects the heart rate estimate with the highest energy and periodicity for the first four time windows. From the fifth time window onward, it selects the estimate closest to the average of the previous four, minimizing extraneous variations. Experiments conducted with one and two subjects showed promising results. In case 1, with one subject, the method achieved a Mean Absolute Error (MAE) of 2.54 BPM and an exclusion rate of 0.94% using MIMO FMCW radar, compared to 4.72% with Doppler radar. In case 2, with two subjects, the method achieved an MAE of 2.28 BPM, confirming accurate simultaneous heart rate estimation.

Remote sensing change detection (CD) has been widely studied, and the CD of heterogeneous images based on cross-sensor acquisition has significant research significance. However, the scarcity of heterogeneous data and the difficulty in obtaining high-quality change maps remain significant challenges. To address these issues, we propose a deep image translation-based feature refinement-aggregation change detection network (FRAN) designed for heterogeneous images, such as optics and SAR images. First, we use data augmentation to increase the number of available images and a no-independent-component-for-encoding GAN (NICE-GAN) to translate the features from the optical domain to the SAR image domain, enabling direct comparison of images from different domains. Finally, we introduce feature refinement module and feature aggregation module to extract more accurate change information and obtain an accurate change region. Our experiments on two public datasets demonstrate that the proposed FRAN's re-detection accuracy is superior to that of four other heterogeneous detection methods.

Mobile edge computing is an emerging computing paradigm in the Internet of Things. Task offloading is a critical method in mobile edge computing to alleviate computational resource constraints. Nowadays, the rising number of tasks is placing greater demands on computing resources. The increasing consumption of computing resources leads to high carbon emission. Achieving environmentally friendly mobile edge computing while effectively managing low-carbon task offloading poses a significant challenge. Recently, deep reinforcement learning has made certain progress in many research fields. However, there are few deep reinforcement learning methods that consider the carbon emission in task offloading. In this paper, we propose a deep reinforcement learning based low carbon emission task offloading algorithm for minimizing carbon emission in mobile edge computing. Firstly, since different base stations exist in the mobile edge computing environment, we consider the mobile edge computing environment with multiple heterogeneous agents. Secondly, to minimize carbon emission, we consider the carbon intensity of the base station as an optimization factor. We conclude the task offloading strategy to minimize carbon emission, consequently achieving the minimization of carbon emission. Moreover, our proposed algorithm allows user devices to decide their own preference for task offloading. Based on the specific requirements and preferences of user devices, our proposed algorithm can dynamically adjust the weights of delay, energy consumption, and carbon emission, respectively. Experiments indicate that the proposed algorithm can accurately and quickly conclude the task offloading strategy to minimize carbon emission.

In this paper, we introduce an innovative Radio Frequency Fingerprinting (RFF)-based device authentication scheme for the Internet of Things (IoT), a network marked by extensive interconnections and interactions among various entities. Our approach, designed for an open and dynamic communication environment, not only identifies devices encountered during training but also effectively rejects those not previously seen. The scheme employs a hypersphere projection for feature embedding, strategically avoiding the need to optimize intra-device variations in the radial direction. It uses a K-Means-based binary classifier for initial device assessment based on cosine similarity scores, followed by a SoftMax classifier for precise identification of known devices. Our extensive numerical analysis confirms that this method delivers superior performance, setting a new benchmark in RFF authentication for IoT security.

This paper introduces activities of AI and Digital Twin working group (AI-DT WG) in XGMF 6G Radio Technology Project. Especially in the fiscal year 2024, AI-DT WG has been engaged in activities for getting the beyond 5G white paper "AI/ML and Digital Twin Technologies"ready for the world. The AI-DT WG also launched the discussion for establishing a shared dataset for AI training. This paper provides an overview of the white paper and the status of the discussion for shared dataset establishment.

Millimeter-wave (mmWave) radar has emerged as a powerful sensing modality for human activity recognition (HAR) owing to its capability to capture 3D point cloud sequences without privacy concerns. However, effectively modeling the sparse, irregular, and non-uniform nature of radar data remains a major challenge. Existing approaches often rely on highly complex network architectures to improve accuracy, which leads to excessive computational overhead and poor scalability. To overcome these limitations, this paper proposes a compact hybrid feature extraction network for voxelized radar point cloud classification, which performs early-stage spatio-temporal fusion and is termed STFusionNet (Spatial-Temporal Fusion Network). The STFusionNet comprises (i) a lightweight 3D convolutional front-end, which treats consecutive temporal frames as input channels to encode motion dynamics into a compact volumetric representation and further aggregates features along the depth axis; (ii) a minimalist 2D convolutional backbone after a Depth-to-Channel Folding operation, which captures spatial features with minimal computational cost. Extensive experiments on the public MMActivity and MiliPoint datasets demonstrate that our model achieves competitive accuracy (e.g., 92.20% on MMActivity and 72.86% on MiliPoint) with only about 50K parameters and 46 MMac, outperforming or matching representative spatio-temporal baselines under a much lower computational budget. Extensive ablation experiments verify the contribution of key components, while statistical significance tests confirm the reliability of the performance improvements. These results confirm that STFusionNet offers a robust, efficient, and generalizable solution for mmWave radar-based HAR in Internet of Things (IoT) applications.

With the rapid advancement of wireless communication technologies, WiFi signals have become essential for modern connectivity across diverse applications. However, their widespread deployment introduces significant security vulnerabilities, including unauthorized access, data leakage, and interference. Accurate identification of WiFi transmitters is crucial for mitigating these threats. While existing methods perform well in ideal conditions, their effectiveness degrades in real-world scenarios, particularly in environments with low signal-to-noise ratios (SNRs). To address this limitation, we propose a novel transmitter identification framework that integrates blueprint separable convolution (BSC) and a squeeze-and-excitation (SE) module. The BSC extracts critical features efficiently, while the SE module dynamically enhances feature representations. Simulation results demonstrate that the proposed approach achieves competitive or superior accuracy compared to state-of-the-art models. Moreover, the framework exhibits strong robustness, maintaining high recognition performance even in challenging transmission conditions with low SNRs.