論文

Depression is a common mental illness and the second leading cause of disability worldwide. Traditional depression diagnosis requires communication with patients and subjective cooperation of patients, which consumes a lot of manpower, material resources, and time costs. With the accumulation of user data in social media and the development of natural language processing, computer-aided diagnosis is realized, better and objective analysis is provided, and a new idea for the diagnosis of depression is provided. We propose a multimodal model based on EmoBERTa and Transformer and a text preprocessing method for a specific pre-trained model and application context. We use user behavior information and past tweets to detect whether a user has depression. Our model achieves state-of-the-art performance on a publicly available multimodal Twitter dataset.

Federated Learning is now widely used to train neural networks under distributed datasets. One of the main challenges in Federated Learning is to address network training under local data heterogeneity. Existing work proposes that taking similarity into account as an influence factor in federated learning can improve the speed of model aggregation. We propose a novel approach that introduces Centered Kernel Alignment (CKA) into loss function to compute the similarity of feature maps in the output layer. Compared to existing methods, our method enables fast model aggregation and improves global model accuracy in non-IID scenario by using Resnet50.

High Altitude Platform Station (HAPS) is a new airborne communications platform that provides wide-area communications services from the stratosphere and has the potential for coverage extension in 6G networks. In particular, it is a practical scenario where multiple HAPS work together to provide communications in densely populated urban areas and extensive regions. However, changes in ground coverage due to HAPS movement cause frequent handovers, which pose challenges to communication quality and stability. To solve this problem, 'footprint fixation,' which maintains constant ground coverage even when the HAPS moves, is expected to reduce handovers and improve communication quality. How we should design beamforming and trajectory under fixed footprint conditions has not been clarified. In this paper, we propose a beamforming and trajectory planning method under fixed-footprint conditions for multiple HAPS systems. For footprint fixation, beamforming dynamically adapts to HAPS motion to eliminate ground coverage shifts, reduce handover frequency, and improve communication stability. Under fixed-footprint conditions, trajectory planning aims to improve the throughput of UEs with low (5th percentile) and medium (50th percentile) communication quality simultaneously through sequential multi-objective optimization. Simulation results in major Japanese cities with different UE distributions show that the proposed method improves both low and medium communication quality UEs, achieving a 20 - 36% improvement in 5th percentile throughput and an 8-20% improvement in 50th percentile throughput compared to other methods. In addition, the footprint fixation increases coverage stability, reducing outage probability to less than 2% and significantly reducing handover frequency.

High Altitude Platform Stations (HAPS) are instrumental in wireless communications, providing enhanced connectivity and extensive coverage by complementing ground-based infrastructure where its expansion is limited. HAPS enhance wireless communications by employing advanced beamforming technology, traditionally based on 2D with free space path loss models. Addressing the inadequacy of these models in different areas, our research introduces a novel beamforming strategy that incorporates detailed 3D geographic and architectural information. This approach models line-of-sight (LOS) and non-line-of-sight (NLOS) conditions with 3D information and optimizes beamforming patterns using Deep Reinforcement Learning (DRL). Our results indicate that by incorporating 3D information, the beamforming performance in terms of average throughput and SINR is markedly enhanced across all user equipments (UEs), compared to traditional 2D approaches. By taking 3D information into account, beamforming in HAPS systems becomes more equitable.

This paper proposes a novel radar-based framework for non-contact biometric identification through heart signal extraction, targeting secure and privacy-conscious identification scenarios. Traditional biometric methods, such as fingerprint and facial recognition, face challenges including privacy concerns, vulnerability to spoofing, and the requirement for close proximity or direct line-of-sight. Our framework addresses these issues by reconstructing electrocardiogram (ECG) signals from radar-extracted cardiac motion data and implementing an open-set person identification system. Specifically, the framework integrates ECGReconNet, a specialized deep learning model for reconstructing ECG signals from human chest wall displacement, the InceptionTime model enhanced with fixed-Class Anchor Clustering (fixed-CAC) loss for robust feature anchoring, and a hypersphere-based delineation method to differentiate known from unknown individuals. Experimental results on a public dataset demonstrate state-of-the-art performance, achieving 99.61% accuracy in closed-set identification (27 subjects) and 93.97% accuracy under challenging open-set conditions (14 known and 13 unknown subjects). However, the proposed approach exhibits limitations, including sensitivity to abrupt body movements and environmental noise, potential performance degradation under severe cardiac irregularities, and reduced efficacy with increased numbers of unknown identities.

In B5G and 6G wireless networks, large-scale antenna arrays introduce spherical wavefronts in the near-field region, presenting both opportunities and challenges for communication and positioning. To address this, we propose a novel near-field positioning method using a bio-inspired dendritic liquid neural network (DLNN). The DLNN features a multi-branch dendritic structure, with each neuron processing signals through multiple parallel dendrites and a dynamic gating mechanism to adapt the weighting and fusion of outputs. This enhances its ability to capture complex spatiotemporal features. Additionally, the liquid layer with leaky integration processes time-series data, enabling strong adaptability and robust temporal memory. Simulation results show that DLNN outperforms other deep learning models in localization accuracy and robustness, with RMSE values of 0.1706 meters at 40 dBm and 0.2642 meters at 10 dBm, compared to 0.2836 meters and 0.2986 meters from other models. Its inference latency of 0.0580 seconds demonstrates excellent accuracy, speed, and energy efficiency, making DLNN highly suitable for real-time positioning in resource-constrained environments.

Mental stress is something we experience on a daily basis. However, when it becomes chronic, it has various negative effects on our body. Although stress monitoring methods based on physiological signals are effective, they may not provide a comfortable experience for users, as they require sensors to be worn or attached to the body. As such, the need for non-invasive, non-contact and comfortable methods arises. In this paper, we propose a stress detection method using face-related and emotion-related features, all of which can be acquired by using a camera. Face-related features include action units and face embedding. Emotion related features include valence, arousal and emotion-related labels. Through these features and their fusions we achieved accuracy of 1.000 for stress detection, 0.891 for task recognition and 0.883 for stress level classification, respectively, with the sample sizes of 111, 316, and 319, respectively. These results demonstrate that our proposed method using face-related and emotion-related features and their fusion is effective in stress monitoring and detection.

The rapid increase in wireless devices has raised significant security and privacy concerns, positioning Specific Emitter Identification (SEI) as a crucial physical-layer security technology. While Deep Learning (DL) has been widely applied to SEI, it often requires large amounts of high-quality signal examples, which are laborious and expensive to obtain. Moreover, the DL-enabled SEI models have difficulties in extracting features from the signal examples in the testing process that are consistent with those from the signal examples in the training phase due to the wireless channel variations, further resulting in a significant reduction in identification performance. To address these challenges, we propose a channel-robust Few-Shot SEI (FS-SEI) method based on Meta-Feature Augmentation (MFA). Our approach utilizes datasets from base emitters to construct a meta-feature embedding function that can extract generalizable features from a few signal examples of target emitters. We then calculate and calibrate the statistics of these extracted features to describe the feature distribution of target emitters. A Multi-Layer Perceptron (MLP) is subsequently trained on both original and augmented features derived from this distribution, achieving a robust FS-SEI model. Experiments conducted on a Wi-Fi dataset comprising 16 emitter categories - 10 as base emitters and 6 as target emitters - demonstrate that our method achieves 93.75% identification accuracy with only 5 examples per target emitter, maintaining 92.56% accuracy even under varying wireless channel conditions. Code is available at https://github.com/lovelymimola/MFA-based-FS-SEI.

Non-contact heart rate measurement is anticipated to be used in various scenarios such as medical and disaster sites, driver monitoring, and smart homes. Millimeter-wave radar-based non-contact heart rate measurement leverages the chest wall displacement signal, which is a superposition of heartbeat and respiration. However, the chest wall displacement signals are susceptible to higher harmonics of respiration and other intermodulation harmonics, which can impede heart rate measurement. In this paper, we investigate a heart rate estimation method utilizing MIMO (Multiple-Input Multiple-Output) FMCW (Frequency Modulated Continuous Wave) radar. Conventional methods that utilize mode decomposition are not robust enough, as the accuracy of the results is highly sensitive to the SNR (Signal-to-Noise Ratio) of the selected IMFs (Intrinsic Mode Functions). Therefore, we propose performing chirp correction in addition to the conventional method that combines PA-LI (Phase Accumulation-Linear Interpolation) and ICEEMDAN (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) to improve the accuracy of heart rate estimation. Our experimental results show that this approach enhances the SNR of the decomposed signal containing heart rate information and improves the accuracy of heart rate estimation.

One major challenge in improving traffic efficiency for multi-intersection signal timing is the synchronous iteration across all intersections. To deal with this problem, in this paper, a novel clock asynchronous traffic signal timing method for multi-intersections based on a joint traffic prediction and control method is proposed and its performance is analyzed. To improve traffic signal timing efficiency, each intersection is viewed as an agent which by following this joint approach designs, in a flexible way, actions, states and rewards. Using traffic prediction and control data, each intersection first computes a value which corresponds to a change in the vehicle position in its lane under control and then adds up this value to other values obtained from all the lanes under control to produce the reward. Traditional methods cannot effectively coordinate multi-intersections with different signal timing time. This has led to the introduction of a novel clock consistency mechanism the operation of which is implemented through a Multi-Agent Coordinated Reinforcement Learning (MACRL) algorithm. For any traffic intersection, the mechanism can estimate, within the duration of the previous action, the impact of adjacent intersections. This is accomplished by considering the previous intersections' reward information and quantify it as its joint reward. The proposed MACRL algorithm optimizes multi-intersections traffic by jointly considering these rewards. Using real traffic data, extensive simulations show that the proposed method outperforms existing schemes under the same conditions in traffic efficiency. The method also delivers more accurate traffic signal timing, reducing congestion and improving overall efficiency.

Meetings are vital for team activities, supporting both decision-making and information sharing, making their quality a critical factor in boosting productivity and reducing costs. Therefore, establishing objective methods for measuring meeting quality is essential to identify specific issues and verify the effectiveness of solutions. However, no well-defined objective features have yet been identified to accurately assess meeting quality. In this study, we explored features that can be extracted from meeting video and audio data to objectively assess the quality of meetings. We conducted online meetings and extracted features from video data, including Action Units, Mel-Frequency Cepstral Coefficients (MFCC) and utterances. We then examined the correlations between these features and participant responses regarding psychological safety and perceived meeting effectiveness. As a result, we found that certain features such as the number of utterances and the Action Units such as Lip Part and Dimpler showed significant correlations with psychological safety and perceived meeting effectiveness.

Millimeter-wave (mmWave) communication, which operates at high frequencies, has gained extensive research interest due to its significantly wide spectrum and short wavelengths. However, mmWave communication suffers from the notable drawbacks as follows: i) The mmWave signals are sensitive to the blockage, which is caused by the weak diffraction ability of mmWave propagation; ii) Even though the introduction of reconfigurable intelligent surfaces (RISs) can overcome the performance degradation caused by serve path loss, the location of users and RISs as well as their densities incur a significant impact on the coverage and rate performance; iii) When the RISs’ density is very high, i.e., the network becomes extremely dense, a user sees several line-of-sight RISs and thus experiences significant interference, which degrades the system performance. Motivated by the challenges above, we first analyze distributed multi-RIS-aided mmWave communication system over Nakagami-m fading from the stochastic geometry perspective. To be specific, we analyze the end-to-end (E2E) signal-to-interference-plus-noise-ratio (SINR) coverage and rate performance of the system. To improve the system performance in terms of the E2E SINR coverage probability and rate, we study the optimization of the phase-shifting control of the distributed RISs and optimize the E2E SINR coverage particularly when deploying a large number of reflecting elements in RISs. To facilitate the study, we optimize the dynamic association criterion between the RIS and destination. Furthermore, we optimize the multi-RIS-user association based on the physical distances between the RISs and destination by exploiting the maximum-ratio transmission. Numerical and simulation results indicate that the deployment of distributed RISs can significantly improve the E2E SINR coverage probability and achievable rate of the system compared to the selected benchmarks.

Millimeter-wave (mmWave) communication, which operates at high frequencies, has gained extensive research interest due to its significantly wide spectrum and short wavelengths. However, mmWave communication suffers from the notable drawbacks as follows: i) The mmWave signals are sensitive to the blockage, which is caused by the weak diffraction ability of mmWave propagation. ii) The mmWave communication incorporating with reconfigurable intelligent surfaces (RISs) is efficient in overcoming the performance degradation caused by severe path loss. Nevertheless, the location of users and RISs as well as their densities impose a significant impact on the coverage and rate performance. Motivated by the challenges above and based on the stochastic geometry concept, we first construct distributed multi-RIS-aided mmWave communication system over Nakagami-m fading channel. Afterward, we analyze the end-to-end (E2E) signal-to-interference-plus-noise-ratio (SINR) coverage and rate performance of the system. To improve the system performance in term of the E2E SINR coverage probability and achievable rate, we study the optimization of the phase-shifting control of the distributed RISs and optimize the E2E SINR coverage particularly when installing a large number of reflecting elements in RISs. To facilitate the study, we optimize the dynamic association criterion between the source and destination. Our simulation results demonstrate that the deployment of distributed RISs can significantly improve the E2E SINR coverage probability and achievable rate of the system compared to the selected benchmarks.

This paper studies user fairness of an integrated sensing and communication (ISAC) system adopting both simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS) and non-orthogonal multiple access (NO-MA). Note that optimizing user fairness of an ISAC system is critical, since the performance trade-off in terms of spectrum resource utilization between communication users and sensing users is hard to balance and improve, particularly at the low signal-to-noise ratio (SNR) regimes. However, such optimization problem is a coupled, non-convex, and is NP-hard in general. To solve this challenging problem, a low-complexity algorithm based on both the successive convex approximation and the semi-deterministic programming techniques is proposed. Notably, the proposed approach can maximize the signal-to-interference-plus-noise ratio (SINR) for the weakest communication user, while guaranteeing the superior sensing performance characterized by the cramér-rao lower bound (CRLB) for the sensing user. Simulation results demonstrate that our approach is capable of enabling superior sum-rate performance than the STAR-RIS with orthogonal multiple access (OMA), the conventional-RIS with NOMA, and the conventional-RIS with OMA schemes.

Antenatal fetal health monitoring primarily depends on the signal analysis of abdominal or transabdominal electrocardiogram (ECG) recordings. The noninvasive approach for obtaining fetal heart rate (HR) reduces risks of potential infections and is convenient for the expectant mother. However, in addition to strong maternal ECG presence, undesirable signals due to body motion activity, muscle contractions, and certain bio-electric potentials degrade the diagnostic quality of obtained fetal ECG from abdominal ECG recordings. In this paper, we address this problem by proposing an improved framework for estimating fetal HR from non-invasively acquired abdominal ECG recordings. Since the most significant contamination is due to maternal ECG, in the proposed framework, we rely on neural network autoencoder for reconstructing maternal ECG. The autoencoder endeavors to establish the nonlinear mapping between abdominal ECG and maternal ECG thus preserving inherent fetal ECG artifacts. The framework is supplemented with an existing blind-source separation (BSS) algorithm for post-treatment of residual signals obtained after subtracting reconstructed maternal ECG from abdominal ECG. Furthermore, experimental assessments on clinically-acquired subjects' recordings advocate the effectiveness of the proposed framework in comparison with conventional techniques for maternal ECG removal.

In recent years, scientists are paying much attention to the research on automatic dementia detection that could be applied to the speech samples of dementia patients. In a related context, recent research has seen the fast development of Deep Learning (DL) and Natural Language Processing (NLP). The techniques developed for text classification or sentiment analysis have been applied to the field of early dementia detection by many researchers. However, text classification and sentiment analysis are different tasks from dementia detection, which makes us believe that for dementia detection, some adjustments would help improve the performance of the machine learning models. In this work, we implemented experiments with various language models including traditional n-gram language models, Average stochastic gradient descent Weight-Dropped Long Short-Term Memory (AWD-LSTM) models, and attention-based models to evaluate the speech data of dementia patients. Unlike traditional works where the text is stripped from stop words, we propose the idea of exploiting the stop words themselves, since they offer non-context information which helps to identify dementia. As a result, 3 different language models are prepared in this work: a model processing only context words, a model processing stop words and Part-of-Speech (PoS) tag sequences, and a model processing both of them. By performing the aforementioned experiments, we show that both grammar and vocabulary contribute equally to classification: The 3 models achieve an accuracy equal to 70.00%, 76.16%, and 81.54%, respectively.

Previous works proposed deep learning models to estimate blood pressure from electrocardiogram (ECG) signals. However, they can only estimate max, min, and mean arterial blood pressures and cannot estimate arterial blood pressure (ABP). This paper presents the ABP estimation method from ECG signals using the deep learning model of U-Net. Through the performance evaluation with signals of about 185 hours, we observed that the proposed method estimated ABP with high accuracy. Furthermore, the accuracies of the calculated max\min and mean ABPs were comparable to those in the previous works, even though our method also estimated ABP. In the end, we discussed the subject-overfitting problem and future work toward practical use of daily blood pressure monitoring.

Specific emitter identification (SEI) plays a more and more important role to recognize unknown radio frequency (RF) devices in the field of physical layer security. Existing deep learning (DL)-based SEI methods have been proposed for solving problems that enough RF dataset can be obtained at the model training stage. In many real application scenarios, however, only the limited dataset can be collected which causes the conventional DL methods unstable. To solve this problem, in this paper, we propose a robust SEI method using a self-attention mechanism and convolutional neural network (attention-CNN). To be specific, we collect a real RF dataset from the automatic dependent surveillance-broadcast (ADS-B) signal. Here, the ADS-B dataset is limited in some special cases and data augmentation is used to improve the features of the dataset. Simulation results show that the proposed method achieves better identification performance than the benchmark.

In the pilot-based channel estimations, a large number of pilot signals enable an improvement in the channel estimation accuracy but force a decrease in the data transmission efficiency. Therefore, the pilot pattern design schemes on the resource grid have been researched to achieve high channel estimation accuracy with a small number of pilot signals. In the conventional scheme for pilot pattern design, the autoencoder which enables discrete feature selection is utilized to design pilot patterns for Vehicular-A channels. However, the pilot pattern for other channels is not reported, and it is not clear whether this scheme has good performance compared to uniform pilot patterns. In this paper, we derive the pilot patterns for Clustered Delay Line (CDL) channels specified in the 5th Generation Mobile Communication System (5G) standard using the autoencoder-based pilot pattern design scheme. We also derive the pilot patterns for CDL-A and CDL-D, which model respectively NLOS (Non-Line-of-Sight) and LOS (Line-of-Sight) environments, with different three speeds of User Equipment (UE). Through computer simulation, we show that the autoencoder-based pilot patterns improve the channel estimation accuracy compared to the uniform pilot pattern.

Over the past few years, there has been a significant rise in the number of fall accidents occurring among elderly individuals, a problem that has been accentuated with to the aging population. Researchers and developers have focused their efforts on investigating and creating various fall detection methods that utilize an accelerometer. However, conventional fall detection methods typically target specific positions where accelerometers are placed. In addition, they suffer from low accuracy which can be attributed to the fact that the classification algorithms commonly employed, such as the support vector machine (SVM) and the random forest (RF), are not specialized in making predictions based on time series data. In this paper, we propose the fall detection method based on a long short-term memory (LSTM) neural network, using an accelerometer. In the proposed method, four kinds of possession positions are set: (i) in hand, (ii) inside a chest pocket, (iii) inside a waist pocket, and (iv) in a bag. The acceleration data collected are classified using the LSTM classifies into one of four classes: (i) standing, (ii) walking, (iii) falling, and (iv) lying down. The results of the multi-class classification are further reclassified into two classes, i.e., fall and non-fall. The experimental results demonstrate that our approach outperforms the conventional methods in terms of fall detection accuracy.