The functional connectivity patterns revealed by emotion recognition closely align with neural pathways involved in emotional processing. Traditional methods often fail to adequately integrate the multidimensional spatiotemporal-spectral characteristics of Electroencephalography (EEG), making it challenging to accurately characterise dynamic emotional processes. Predefined brain regions and fixed thresholds yield coarse functional networks, which limit the accurate identification of critical connections. Furthermore, black-box models lack interpretability, providing little decision support or visualisable evidence of neural circuits for neuroscience and clinical applications. To address these limitations, this study proposes an emotion recognition via channel-feature selection and dynamic brain network (CFSDBN). First, spatial-spectral features are extracted using a residual network, while bidirectional gated recurrent units capture temporal dynamics, thereby improving feature utilisation. Next, these spatio-temporal-spectral features serve as node inputs to a graph attention network, where node attention weights enable adaptive channel selection and sparse functional connectivity learning, thereby overcoming localisation inaccuracies caused by coarse-grained processing. Finally, joint node embeddings and connection weights are used for emotion classification, and key channels and neural circuits are visualised to provide interpretable evidence for emotional neural mechanisms. By deeply coupling multidimensional features with brain network optimisation, CFSDBN achieves significant improvements in classification performance on the DEAP, MODMA, and SEED-V datasets. It enhances hierarchical interpretability from micro-level features to macro-level network interactions, offering a high-performance and explainable solution for EEG-based emotion recognition.

Multimodal sentiment analysis in social video is the foundation of affective computing and artificial intelligence. Due to the heterogeneity of multimodal data including facial expression, speech, and language, multimodal sentiment analysis remains a challenging problem. Existing methods mostly explore delicate fusion strategies to obtain consistent multimodal sentiment representation. There is still the problem of cross-modal bias due to distributional differences in multimodal data. To address this problem, a multimodal sentiment analysis method based on multiview cross-modal fusion (MVCF) is proposed in this article, which significantly improves the performance of extraction between heterogeneous architectures from three perspectives. The unimodal local context module is proposed to implement one to one pattern, reflecting independent change factors combined with specific information adaptively. We propose global union modules to learn the one-to-all pattern by projecting intermediate features into an aligned latent space, where modality-specific information is discarded. To align cross-modal affective prototypes, the all-to-all multimodal fusion module performs adaptive fusion for all modalities and interactions to eliminate perturbing information in modal commonality. Experimental results on multimodal sentiment analysis benchmarks show that our MVCF outperforms baselines in complete and incomplete modality settings, demonstrating the effectiveness and robustness of MVCF. Our code is publicly available onhttps://github.com/zhixingyu/MVCF.

Depression is a critical public health concern characterized by underdiagnosis, often due to stigma, lack of awareness, and reluctance to seek help. Cases of delayed intervention could be alleviated by wearable solutions which enable continuous and unobtrusive monitoring of depression indicators. Compared to electroencephalogram (EEG)-based and video-based depression recognition, speech-based approaches can be performed without deliberate user attention. However, due to the limited accuracy of existing algorithms and constrained resources at edge, performing accurate speech-based depression recognition on wearable platforms remains a challenge. Therefore, to achieve unobtrusive, accurate, and efficient depression recognition at edge, this work presents a hierarchical local–global network (HLG-Net) and optimized processor design for speech-based depression recognition. The proposed HLG-Net integrates convolutional neural networks (CNNs) with multihead attention (MHA) mechanism to simultaneously capture local acoustic features and global utterance-level coherence, enhancing depression stage recognition. For efficient processor design, a cross-layer buffered dataflow is proposed for efficient data handling, reducing data storage by 98.77%. The computing unit (CU) employs layer fusion, operator optimization, and quantization techniques to further improve resource utilization and reduce power consumption while preserving recognition accuracy. System-level low-power techniques such as clock/input gating and near-threshold design for application specific integrated circuit (ASIC) further reduce power consumption. The proposed processor implemented on field-programmable gate array (FPGA) (XC7Z100-2FFG900) achieves the lowest reported mean absolute error (MAE) of 5.13 on AVEC 2014 database. The 180-nm ASIC implementation shows a simulated power consumption of$17.4~\mu $W at 0.4 V. The results demonstrate the feasibility of accurate and efficient speech-based depression recognition on wearables.

We examine a secure wireless communication system that is enabled by unmanned aerial vehicles (UAVs) in this research. In the wireless communication system with an eavesdropping UAV, we deploy a relay UAV to facilitate the transmission of confidential signals from the source station (denoted asS) to ground users. Additionally, we select an idle relay UAV to act as a cooperative jamming UAV, sending interference signals to the eavesdropping UAV. It is quite feasible that the eavesdropping UAV will leverage its mobility to improve the quality of its eavesdropping, making its trajectory unpredictable. First, to address the worst-case scenario for the ground user’s security performance, we assume the eavesdropping UAV approaches at the closest distance.We aim to maximize the worst secrecy rate under perfect CSI via designing the flight trajectory and transmission power of both the relay UAV and the jamming UAV. Second, we investigated the performance of the system’s secrecy outage probability under imperfect CSI. The presence of eavesdropping UAVs and the unpredictable nature of their environment makes traditional convex optimization methods mathematically complex for solving the trajectory optimization problem of the relay and jamming UAVs. To address this, we propose the Multi-Agent joint design trajectory and power (MAJDTP) algorithm based on the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm to optimize the flight trajectory and transmission power of both UAVs. During the design and training process, the relay and jamming UAVs are treated as agents to derive their optimal flight paths and transmission energy. Finally, our approach surpasses the benchmark algorithm, as demonstrated by the simulation results.

Multimodal sentiment analysis recognizes emotions through text, audio, and visual modalities, but data incompleteness is a major challenge. Existing methods often focus on specific types of deficiencies and perform poorly when multiple types of noise are present simultaneously. To address this issue, we propose a noise-prompted adversarial training framework with a multimodal interaction model to enhance the model’s robustness to missing modalities. The model first extracts common and unique features from each modality using a BERT text encoder and a shared-private encoder. Correlation measurements are then used to calculate the similarity between modalities, and a weighting mechanism is applied to the shared features. These features are deeply fused using a Transformer, and adversarial training combined with semantic reconstruction supervision helps the model learn a unified representation of noisy and clean data. Experimental results show that this method significantly improves the performance of multimodal sentiment analysis.

Ensuring signal confidentiality against eavesdroppers is particularly challenging, especially with imperfect channel state information (CSI). To address this, we propose a novel approach leveraging reconfigurable intelligent surfaces (RISs) to enhance security and optimize transmission performance. This paper focuses on secure communication in symbiotic radio (SR) systems by investigating cooperative jamming-assisted transmission with RISs, providing a robust solution to these challenges. RIS-I, acting as a secondary transmitter (STx), multicasts confidential signals from the primary transmitter (Alice) to a primary user (Bob), protecting against eavesdropping by Eve. Additionally, RIS-I transmits its own signals to a secondary user (SU) using backscattering radio technology. Meanwhile, RIS-II serves as a cooperative jammer, converting received confidential signals from Alice into jamming signals by strategically adjusting its reflection coefficients to disrupt Eve’s reception. These RISs can operate cooperatively; when RIS-II transmits as an STx, RIS-I functions as a cooperative jammer. We explore two scenarios: 1. With perfect CSI for the wiretap channel, we propose a joint SDR(Semi-definite relaxation)+MM(Minorization-maximization) optimization algorithm to simultaneously optimize Alice’s beamforming vector and the RISs’ reflection coefficients. 2. With imperfect CSI, we derive the secrecy outage probability formula and evaluate the scheme’s performance across different scenarios. Numerical results demonstrate that RIS-assisted cooperative jamming significantly enhances the secrecy rate and reduces the secrecy outage probability for Bob, outperforming traditional RIS-assisted SR systems.

Accurate cellular network traffic prediction is crucial for intelligent network planning and management in 6G. However, the non-stationary characteristics of cellular network traffic present significant challenges when training deep neural networks for traffic forecasting. To address this issue, we propose a two-stage deep learning framework, JO-DPNet, based on Bayesian joint parameter optimization, which integrates data decomposition techniques with Bayesian joint optimization to effectively mitigate the adverse impacts of non-stationarity and error accumulation on prediction accuracy. In the first stage, a data decomposition module uses Variational Mode Decomposition (VMD) to decompose the original data into network traffic subset series(TSS), thereby alleviating the negative effects of non-stationarity. In the second stage, a prediction and construction module leverages a bi-directional LSTM (Bi-LSTM) network to extract deep spatial-temporal features from the TSS in a bidirectional manner. A fully connected layer then captures the relationships between the TSS and reconstructs the predicted results into the final output. The JO module employs the Tree-structured Parzen Estimator based Bayesian optimization algorithm(TP-BO) simultaneously determines the optimal VMD mode number k and the hyperparameters of the Bi-LSTM network through probabilistic surrogate model. Extensive experiments on three real-world cellular traffic datasets demonstrate that the proposed method significantly mitigates the non-stationary characteristics of the traffic data. Compared to state-of-the-art methods, JO-DPNet achieves reductions in MAE by 29%, 3%, and 19% for three type prediction tasks on the Telecom Italia dataset. The source code is available to the public at: https://github.com/VicentZhang259/JO-DPNet.

Laplacian dimensionality reduction can effectively achieve feature transformation and preserve the important structure of high-dimensional features. However, the trained model with this method usually require better generalization ability to new samples. Hence, a human activity recognition method based on kernel-supervised laplacian eigenmaps (KSLE) by combining the kernel method, laplacian mapping, and supervised learning is proposed. Firstly, the adjacency distance relationship of original samples features in the high-dimensional kernel space is obtained. Secondly, the category labels in the high-dimensional feature set are incorporated in the manifold learning for dimensionality reduction. Then, kernel trick is utilized to directly solve the low-dimensional embedding structure of the test feature set. Finally, the obtained low-dimensional feature set is input into the classifier for recognition. Extensive experiments conduct on two public datasets confirm the effectiveness of the proposed approach for improving the generalization ability of new samples.

The methods based on multimodal representation learning enhance discriminable sentiment expression for multimodal sentiment analysis(MSA). The modal invariant and specific features serve different purposes in sentiment learning and the diversity of inter-sample and inter-category relationships takes less consideration in previous advances. In this paper, we propose a fine-grained tri-modal interaction model for MSA to refine and enhance the overall affective state at the uni/multi-modal level and label level. Concretely, we simultaneously focus on the contributions of both unimodal and multimodal views to improve holistic affective knowledge. The similarity measurement function and self-supervised learning are introduced to reduce the inherent modality gap and refine the invariant representations. We provide two specific constraints to strengthen the specific uniqueness and discriminative capacity. Moreover, we develop a multi-granularity fusion module to fully integrate two views in a coarse-to-fine form. Experimental results on two datasets show that our method fares better than the state-of-the-art model.

Human activity recognition (HAR) based on wearable sensors is a hot topic in health detection and motion management. Nonetheless, conventional identification methods necessitate substantial labeled datasets, and the acquisition of high-quality labeled data crucial for human activity recognition is both time-consuming and costly. To tackle this problem, transfer learning is used to annotating unlabeled or a few labeled target domains using labeled source domains. Meanwhile, individuals exhibit significant differences in amplitude, angles, and other aspects when performing the same action. Due to the limited expressive capacity of time series, effectively capturing various types of differences simultaneously poses a challenge, impacting the effectiveness of transfer. Therefore, in this paper, we propose a stacked transfer network (STN) for 2D modeling and feature extraction of sensor data in steps. It adaptively decomposes complex temporal variations into multiple intra- and inter-periodic variations using the Fast Fourier Transform (FFT), allowing the model to focus more on the trends of the activty and less on the effects of individual discrepancy. Our comprehensive cross-domain activity recognition (CDAR) experiments on three large public activity recognition datasets (i.e., OPPORTUNITY, PAMAP2, and UCI-DSADS) show that the STN achieves high accuracy in activity recognition ttransfer.

With recent advents of industrial Internet of Things (IIoT), the connectivity and data collection capabilities of industrial equipment have be significantly enhanced, yet bringing new challenges for the remaining useful life (RUL) prediction. To fulfill the RUL predicting demand in multivariate time series, this work proposes an encoder-decoder model termed as dual-scale transformer model (DSFormer), built upon the Transformer architecture. First, in the encoder part, a dual-attention module is designed for the weight feature extraction from both dimensions of the sensor and time series, aiming to compensate for the diverse impacts of different sensors on the prediction. Next, a temporal convolutional network (TCN) module is introduced to capture sequence features and alleviate the loss of positional information incurred by stacking blocks. Then, the feature decomposition module is integrated into the decoder for trend feature extraction from sequences, providing the model with additional sequence information. Finally, compared to existing models, the proposed method can obtain the superior performance in terms of the root mean square error (RMSE) and Score metrics on the FD001, FD002 and FD003 subsets of the C-MAPSS dataset, with an average improvement of 3.2% and 2.5% respectively. In particular, the ablation experiment further validates the effectiveness of proposed modules in handling multivariate time series and extracting features.

Affective analysis is a technology that aims to understand human sentiment states, and it is widely applied in human–computer interaction and social sentiment analysis. Compared to unimodal, multimodal sentiment analysis (MSA) focuses more on the complementary information and differences from multimodalities, which can better represent the actual sentiment expressed by humans. Existing MSA methods usually ignore the problem of multimodal data ambiguity and the uncertainty of influence redundant features on the sentiment discriminability. To address these issues, we propose a fuzzy attention fusion-based MSA method, called FFMSA. FFMSA alleviates the heterogeneity of multimodal data through shared and private subspaces, and solves the ambiguity using a fuzzy attention mechanism based on continuous value decision making, in order to obtain accurate sentiment features for downstream tasks. The private subspace refines the latent features within each single modality through constraints on their uniqueness, while the shared subspace learns common features using a nonparametric independence criterion algorithm. By constructing sample pairs for unsupervised contrastive learning, we use fuzzy c-means to model uncertainty to constrain the similarity between similar samples to enhance the expression of shared features. Furthermore, we adopt a multiangle modeling approach to capture the consistency and complementarity of multimodalities, dynamically adjusting the interaction between different modalities through a fuzzy attention mechanism to achieve comprehensive sentiment fusion. Experimental results on two datasets demonstrate that our FFMSA outperforms state-of-the-art approaches in MSA and emotion recognition. The proposed FFMSA achieves sentiment binary classification accuracy of 85.8% and 86.4% on CMU-MOSI and CMU-MOSEI, respectively.

Sidelobe suppression is a crucial preprocessing technique for Synthetic Aperture Radar (SAR) image analysis. The presence of strong scattering targets generates sidelobes that can interfere the targets with relatively weak scattering. The overlapping of multiple strong cross-shaped sidelobes may even generate fake targets, significantly influencing the accuracy of SAR target detection and recognition. Among existing sidelobe suppression methods, the Spectral Reshaping Sidelobe Reduction (SRSR) method has shown promising results. It separates the mainlobe and sidelobes through altering the sidelobe direction while preserving the SAR image resolution. However, this method exhibits limitations in effectively suppressing strong cross-shaped sidelobes. It also introduces additional sidelobes, blurring the surroundings of the scattering points. This paper proposes an improved SRSR method to resolve this disadvantage. It constructs a feature image representing the superposition of sidelobes. This is achieved by analyzing the statistical differences of complex data between orthogonal and non-orthogonal sidelobe regions before and after spectral reshaping. Further modulus selection ensures that the feature image only contains the sidelobe information that needs to be eliminated. The proposed method successfully resolves the drawback of introducing new sidelobes in the original SRSR while achieving better suppression of strong cross-shaped sidelobes. Experimental results on airborne and spaceborne SAR images demonstrate that the proposed method outperforms other state-of-the-art techniques. Improved peak sidelobe ratio (PSLR) and integrated sidelobe ratio (ISLR) in both range and azimuth directions and smaller image entropy can be achieved by our method. Due to its superior sidelobe suppression capability, the SAR images processed by our method exhibit significantly improved accuracy in target detection.

Human activity recognition (HAR) can be exploited to great benefits in many applications, including elder care, health care, rehabilitation, entertainment, and monitoring. Many existing techniques, such as deep learning, have been developed for specific activity recognition, but little for the recognition of the transitions between activities. This work proposes a deep learning based scheme that can recognize both specific activities and the transitions between two different activities of short duration and low frequency for health care applications. In this work, we first build a deep convolutional neural network (CNN) for extracting features from the data collected by sensors. Then, the long short-term memory (LTSM) network is used to capture long-term dependencies between two actions to further improve the HAR identification rate. By combing CNN and LSTM, a wearable sensor based model is proposed that can accurately recognize activities and their transitions. The experimental results show that the proposed approach can help improve the recognition rate up to 95.87% and the recognition rate for transitions higher than 80%, which are better than those of most existing similar models over the open HAPT dataset.

Many institutions and researchers have conducted in-depth research on mobile/tele-health monitoring in recent decades. This paper designs and implements a high concurrent data communication server, improving the concurrent processing ability of health monitoring gateway, based on I/O Completion Port (IOCP) communication model. Meanwhile, improve the communication efficiency by elevating the aspects of duplex communication, session connection pool, dynamic cache. Finally, the paper analyzes the optimal server packet size and the best number of concurrent services, through the response time, throughput and other indicators.