Blind source separation (BSS) in time-frequency (TF) domain is a versatile framework to recover sources from the recorded mixture signals in a reverberant environment. In general, a two-stage strategy is one of the popular BSS frameworks for the underdetermined BSS case (the number of mixtures is less than the number of sources), which is a tough problem due to the mixing matrix is not invertible. In this paper, we propose a new two-stage scheme combining density-based clustering and sparse reconstruction to estimate mixing matrix and sources, respectively. At the first stage, we transform the mixing matrix estimation as an eigenvector clustering problem based on a particular local dominant assumption. The eigenvectors are first exploited from the rank-one structure of local covariance matrices of mixture TF vectors. These eigenvectors are then clustered and adjusted to give estimated mixing matrix by cooperating density-based clustering and weight clustering. At the second stage, we transform the source reconstruction as a ℓp norm (0>p ≤ 1) minimization by an iterative Lagrange multiplier method. With a proper initialization, the obtained solution is a global minimum for any p in (0,1] with convergence guarantee. The proposed approach is demonstrated to be superior to the state-of-the-art baseline methods in various underdetermined experiments.

Meter data collection and management in smart grid has the potential for underlying security risks, e.g., low-sparsity unobservable attacks. Thus, it is crucial to investigate the vulnerability of smart grid through various exposure tests associated with these unobservable attacks. Recently, much attention has been paid to low-sparsity unobservable attacks with complete knowledge of the system matrix. In this paper, the unobservable attack exposure analysis is based on a relaxed condition, i.e., an incomplete knowledge of the system matrix. Furthermore, a data-driven attack scheme is designed to demonstrate that such knowledge can be learned with a two-stage strategy. In the first stage, a sequence of intercepted meter data is utilized to learn about the incomplete system matrix with a blind identification approach. In the second stage, the estimated system matrix at hand is used for the attack vector construction with a sparsity-exploiting method. Finally, the validity of the proposed data-driven attack scheme is tested through various experiments. The proposed result reveals the potential risk of meter data leakage to the security of the smart grid.

Wearable electroencephalogram (EEG) devices have shown great potential in enabling real-time monitoring of subtle changes in brain activity, providing new possibilities for the assessment and management of chronic pain. However, recognizing pain-related biomarkers from EEG data remains a complex, multi-task challenge. Most existing research focuses on single-task approaches and rarely addresses this issue within a unified framework. In this paper, we propose a novel deep neural network (DNN) model called multiscale frequency domain MetaFormer (MFDFormer), which is designed to simultaneously predict the presence, type, and intensity of chronic pain. The proposed MFDFormer comprises two primary sub-networks: multiscale feature extractor (MFE) and frequency domain MetaFormer (FDFormer) encoder. The MFE extracts diverse EEG features through convolutions with different kernel sizes, while a self-attention mechanism is integrated into MFE to emphasize the importance of inter-dependency among these features. The FDFormer encoder refines the output MFE features using causal convolutions to capture local patterns and projects them into a higher-dimensional representation domain. Additionally, it incorporates spatial and temporal frequency domain learners in parallel to effectively capture the spatial-temporal information of EEG data. Based on the publicly available BFCP dataset, the proposed MFDFormer demonstrates superior performance over state-of-the-art algorithms, achieving accuracies of 97.59%, 96.02%, and 83.48% in pain or non-pain state classification (P/NSC), pain type classification (PTC), and pain intensity classification (PIC) tasks, respectively. This paper proposes a unified, end-to-end DNN-based framework for multi-task chronic pain recognition, providing a reliable solution with the potential to advance pain diagnosis and management in IoT and smart wearable applications.

WiFi signal features, particularly channel state information (CSI), have gained considerable attention in human activity recognition (HAR) due to their non-intrusive and privacy-friendly nature. However, CSI packets are often non-stationary and exhibit fluctuations across various human activities. In this letter, we propose an end-to-end deep neural network called efficient residual packet attention network (ERPANet) to tackle these challenges. In the proposed framework, we introduce the multi-layer residual module (MRM) composed of an attention residual (AR) operation and a downsampling attention residual (DAR) operation to effectively capture spatial-temporal features of CSI packets. In addition, a self-attention mechanism is embedded within AR and DAR to emphasize the importance of inter-relationship among these multi-scale CSI packet features. The proposed ERPANet aims to encode both channel information and long-range dependencies of CSI packet features. Extensive experiments show that ERPANet outperforms state-of-the-art methods, achieving average accuracies of 99.4% and 99.6% on the UT-HAR and NTU-HAR datasets, respectively.

Underwater acoustic target classification (UATC) aims to identify the type of unknown acoustic sources using passive sonar in oceanic remote sensing scenarios. However, the variability of underwater acoustic environment and the presence of complex background noise pose significant challenges to enhancing the accuracy of UATC. To address these challenges, we develop an innovative deep neural network algorithm integrated by multiscale feature extractor and efficient channel attention mechanism. The proposed algorithm leverages tailored representations and deep learning to enhance adaptability and reliability of UATC performance. We employ auditory cofeatures, such as Mel-frequency cepstral coefficients and Gammatone frequency cepstral coefficients, combined with their first-order and second-order differentials, to capture the dynamic variations and contextual information of underwater acoustic signals in time-frequency domain. In addition, we integrate multiscale convolution with an efficient channel attention mechanism to share and exploit the interrelationship of auditory cofeatures. Experimental validations using ShipsEar and Deepship datasets, along with various noise types, have demonstrated the effectiveness of our algorithm in comparison to state-of-the-art methods.

Efficient data sharing schemes are one of the key technologies in the Internet of Vehicles (IoV). However, the insufficient willingness of vehicle users to provide data makes the traditional blockchain-based IoV network have low throughput. The income of IoV providers decreases when the vehicle density increases on the road. In this paper, we investigated a mobile vehicle data sharing scheme based on the consortium blockchain. In detail, the consortium blockchain was used to limit the degree of decentralization and openness, and the optimal revenue strategy approach between vehicles and data-demand devices was obtained through the Stackelberg game. The load test library based on Node.js was used to simulate and compare the data transmission performance of the proposed consortium blockchain with traditional blockchain schemes. Simulation results show that the proposed scheme had higher buyer’s revenue, and the block transmission performance was significantly higher than that of traditional blockchain schemes.