Wearable human activity recognition (HAR) focuses on classifying human activities from multi-sensor data collected by wearable devices and has become increasingly important in pervasive computing. However, existing methods face several challenges: (1) aggregating heterogeneous local models while preserving user-specific data distributions, (2) achieving personalized adaptation of global models to diverse behavioral patterns, and (3) capturing both local and global temporal dependencies inherent in sensor time-series data. To address these challenges, we propose PFLMamba, a personalized federated learning framework integrating an attention-enhanced state-space model (ASSM) for hierarchical temporal feature extraction. PFLMamba employs a server-side personalized attention aggregation mechanism to tailor global models for individual clients, while ASSM captures both local and long-range temporal patterns on the client side. Extensive evaluations on the WISDM and PAMAP2 datasets demonstrate that PFLMamba achieves F 1 scores of 91.83 and 96.01, respectively, outperforming state-of-the-art federated learning baselines such as FCLFD, EFDLS, and FKD. PFLMamba’s effectiveness is further validated on multi-user and heterogeneous-sensor datasets, namely UCI-HAR and UNIMIB-SHAR, confirming its generalization across diverse populations and device types. Beyond predictive accuracy, PFLMamba exhibits a favorable trade-off between efficiency and performance, with lower client-side training overhead than teacher-student-based frameworks and competitive throughput relative to lightweight alternatives. Further experiments on resource-constrained edge devices (i.e., Raspberry Pi 4 and PYNQ-Z2) validate its practical feasibility, highlighting low-latency inference and moderate energy consumption. These results establish PFLMamba as a robust and efficient solution for personalized wearable HAR.

Multivariate time series classification (MTSC) requires a model capable of capturing both localized temporal patterns and long-range dependencies while effectively modeling complex inter-variable relationships. Existing convolutional neural network (CNN)-based capsule models suffer from limited receptive fields, constraining their ability to model long-range dependencies, while transformer-based capsule models rely solely on self-attention, which, despite its effectiveness in capturing global features, struggles with preserving local structures and efficiently processing long sequences. To overcome these limitations, we propose KACapMamba, a Knowledge-Aware State-Space Capsule Network, which integrates three attentive Mamba blocks with a routing layer to achieve hierarchical temporal modeling. Unlike conventional transformer-based methods that primarily depend on self-attention for global dependency modeling, each attentive Mamba block in KACapMamba fuses 1-dimensional CNNs, self-attention, state-space module (SSM), and mutual cross-attention, enabling a more structured and adaptive feature representation. Self-attention ensures effective long-range dependency modeling, while SSM provides a recurrent-state mechanism, inherently better suited for sequential processing compared to purely attention-based architectures, thereby enhancing temporal continuity and long-term pattern retention. Notably, mutual cross-attention addresses the limitations of traditional fusion strategies such as element-wise addition or multiplication, which lack the capacity to selectively enhance relevant features. By dynamically reweighting interactions between features, mutual cross-attention enables more expressive, context-aware representations, leading to improved feature disentanglement and inter-variable modeling. Additionally, the routing layer further enhances hierarchical feature disentanglement by refining capsule activations, reinforcing structural coherence and feature selectivity. Experiments conducted across the UEA benchmark archive demonstrate that KACapMamba consistently achieves the highest ‘win’/‘tie’/‘lose’/‘best’ ratios when evaluated against 10 leading transformer and Mamba architectures under both$Accuracy$and$F_{1}$metrics. Moreover, in comparison with 22 state-of-the-art MTSC models, it again secures the most favorable performance profile, demonstrating a clear and statistically supported advantage across both evaluation measures.

Over the years, various sophisticated deep learning algorithms have surfaced for multivariate time series classification (MTSC), notably the dual-network-based model. This model comprises two parallel networks tailored to time series data: one for local feature extraction and the other for global relation extraction. However, effectively integrating these dual networks poses a significant challenge. To address this, we propose a knowledge aggregation transformer network (KATN) for MTSC. KATN, composed of four aggregation transformer blocks, extracts abundant regularizations and connections hidden within the data. Each block incorporates a modified residual network (MResNet) for local feature extraction and a multi-head attention network for global relation extraction. Initially, the block merges MResNet's output feature with that of the multi-head attention network through an additive operation. Subsequently, it aligns features with a fully connected (i.e., dense) layer and activates neural units using the Gaussian error linear unit function. This strategic feature aggregation allows for capturing long-range dependencies among multiple variables in multivariate time series data. Experimental results demonstrate that KATN significantly outperforms 6 state-of-the-art transformer variants, achieving a ‘win’/‘tie’/‘lose’ record of 9/6/15 and securing the lowest AVG_rank score. Furthermore, when evaluated against 18 existing MTSC algorithms across 13 UEA datasets, KATN consistently delivers superior performance, attaining the lowest AVG_rank score among all compared methods.

This article proposes a federated contrastive learning with feature-based distillation (FCLFD) framework tailored for human activity recognition (HAR). The FCLFD system integrates a central server with multiple mobile users to address a diverse range of HAR challenges. The framework encompasses two pivotal elements: a contrastive student--teacher (CST) architecture with feature-based distillation and an average weight scheme (AWS). The CST framework facilitates the transfer of comprehensive knowledge from a teacher model to a student model through feature-based distillation and contrastive learning, with both models sharing an identical architecture. Each participating user periodically uploads the weights of its student model to the central server, where the AWS deployed on the server calculates the average weights based on contributions from all connected users. The aggregated weights are then redistributed to each user, who updates their teacher model accordingly. Experimental evaluations demonstrate that when 50 users are connected, the proposed FCLFD scheme obtains the highestF1values of 89.01 and 94.19, outperforming several state-of-the-art federated learning algorithms on the wireless sensor data mining (WISDM) and PAMAP2 datasets.

This article proposes a semi-supervised contrastive capsule transformer method with feature-based knowledge distillation (KD) that simplifies the existing semisupervised learning (SSL) techniques for wearable human activity recognition (HAR), called CapMatch. CapMatch gracefully hybridizes supervised learning and unsupervised learning to extract rich representations from input data. In unsupervised learning, CapMatch leverages the pseudolabeling, contrastive learning (CL), and feature-based KD techniques to construct similarity learning on lower and higher level semantic information extracted from two augmentation versions of the data, "weak" and "timecut," to recognize the relationships among the obtained features of classes in the unlabeled data. CapMatch combines the outputs of the weak- and timecut-augmented models to form pseudolabeling and thus CL. Meanwhile, CapMatch uses the feature-based KD to transfer knowledge from the intermediate layers of the weak-augmented model to those of the timecut-augmented model. To effectively capture both local and global patterns of HAR data, we design a capsule transformer network consisting of four capsule-based transformer blocks and one routing layer. Experimental results show that compared with a number of state-of-the-art semi-supervised and supervised algorithms, the proposed CapMatch achieves decent performance on three commonly used HAR datasets, namely, HAPT, WISDM, and UCI_HAR. With only 10% of data labeled, CapMatch achieves values of higher than 85.00% on these datasets, outperforming 14 semi-supervised algorithms. When the proportion of labeled data reaches 30%, CapMatch obtains values of no lower than 88.00% on the datasets above, which is better than several classical supervised algorithms, e.g., decision tree and -nearest neighbor (KNN).

Over the years, several deep learning algorithms have been proposed for electroencephalography (EEG) signal classification. The performance of any learning method usually relies on the quality of the learned representation that provides semantic information for downstream tasks such as classification. Thus, it is crucial to improve the model’s representation learning capability. This paper proposes an Efficient Temporal Network with dual self-distillation for EEG signal classification, ETNEEG. It enhances the model’s representation learning by promoting mutual learning between higher-level and lower-level semantic information. The proposed ETNEEG consists of two main components: a parallel dual-network-based feature extractor called MLN-GRN and a dual self-distillation module. MLN-GRN includes a multi-scale local network (MLN) and a global relation network (GRN). MLN pays attention to local features of EEG data, and GRN is designed for learning global patterns of EEG data. Meanwhile, the dual self-distillation module extracts semantic information by mutual learning among the output layer and the low-level features. To evaluate the proposed method’s performance, seven widely used public EEG datasets, i.e., FaceDetection, FingerMovements, HandMovementDirection, MotorImagery, PenDigits, SelfRegulationSCP1, and SelfRegulationSCP2, are applied to a set of experiments. Experimental results demonstrate that the proposed ETNEEG achieves excellent performance on these datasets compared with fourteen existing algorithms.