15 May 2023 in IEEE Transactions on Instrumentation and Measurement
Obtaining robust feature representations from multi-position wearable sensory data is challenging in human activity recognition (HAR) since data from different positions can have unordered implicit correlations. Graph neural networks (GNNs) represent data as structured graphs by mining complex relationships and interdependency via message passing between the nodes of graphs. This paper proposes a novel framework (MhaGNN) that combines GNNs and the multi-head attention mechanism, aiming to learn more informative representations for multi-position HAR tasks. The MhaGNN framework takes the sensor channels from multiple wearing positions as nodes to construct graph-structured data from the spatial dimension. Besides, the multi-head attention mechanism is introduced to complete the message passing and aggregation of the graphs for spatial-temporal feature extraction. The MhaGNN learns correlations among sensor channels that can be used as compensatory features together with the captured features from each single sensor channel to enhance HAR. Experimental evaluations on three publicly available HAR datasets and a ground-truth dataset demonstrate that our proposed MhaGNN achieves state-of-the-art recognition performance with the captured rich features, including PAMAP2, OPPORTUNITY, MHAEATH and MPWHAR.
16 October 2024 in IEEE Sensors Journal
Federated Learning (FL) revolutionizes distributed learning in Human Activity Recognition (HAR) by allowing clients to train models locally and share only model parameters, thus optimizing data usage and mitigating privacy concerns. However, the presence of Non-Independent and Identically Distributed (Non-IID) data across clients impedes FL efficiency between the server and clients, affecting HAR performance and communication efficiency. Acknowledging the significance of post-training parameters as intrinsic representations of knowledge acquired by client models, we introduce FedPA (Federated Parameters Averaging), a novel algorithm for model aggregation. This algorithm strategically assigns varying weight coefficients to clients during aggregation, accurately reflecting each client’s learning contribution and thereby enhancing the efficacy of both global and client models. We conducted empirical analyses using real-world datasets to assess FedPA’s effectiveness. The results demonstrate that FedPA not only preserves accuracy but also improves communication efficiency compared to existing FL aggregation algorithms, like FedAvg, FedMA, and FedCDA. These findings underscore FedPA’s superiority in addressing Non-IID challenges in HAR tasks and highlight its potential to improve overall model performance in FL settings.
15 July 2025 in Biomedical Signal Processing and Control
29 May 2026 in SSRN Electronic Journal
Reliable anaerobic digestion (AD) control must preserve methane productivity while preventing process instability under variable loading and thermal conditions. From an engineering perspective, this requires a controller that remains effective not only under nominal ADM1 operating regimes but also under thermal disturbances that alter microbial kinetics and destabilize reactor behavior. This study presents a safety-first reinforcement learning framework for AD based on a temperature-extended Anaerobic Digestion Model No.~1 (ADM1). The extended model captures thermal disturbances and their influence on closed-loop process behavior, while a Soft Actor-Critic (SAC) controller is trained with a reward that penalizes both the magnitude and occurrence of safety violations. The framework is evaluated under six operating scenarios using both full (13D) and compact (5D) observation settings, and compared against Constant control, PID, Cascaded PID, MPC, and NMPC baselines through a three-dimensional evaluation protocol covering methane yield, safety, and stability. Across the six benchmark scenarios, SAC achieves the strongest average overall performance, although it is not the top-scoring controller in every individual scenario. This advantage is most pronounced under severe thermal stress, where the rule-based baselines considered here lose operability while the learned controller remains operable over the full 30-day horizon. These results indicate that, within this simulator benchmark, safety-first reinforcement learning with a temperature-extended ADM1 environment is a promising engineering approach for safe and effective AD control under both conventional loading disturbances and practical thermal stress.
15 October 2026 in SSRN Electronic Journal
Accurate state-of-health (SOH) estimation is essential for reliable and efficient lithium-ion battery operation. The distribution of relaxation times (DRT) provides a structured view of electrochemical degradation, but existing DRT-based SOH methods commonly reduce the spectrum to a few predefined peak descriptors. Such peak-centric representations may overlook useful information when relaxation responses shift, broaden, overlap, or develop transition structures during aging. To address this limitation, this study proposes the Distribution-of-Relaxation-Time Region Attention Network (DRA-Net), which learns ordered relaxation-region representations directly from the full corrected DRT. First, inductive and broad low-frequency contributions are removed before regularized DRT reconstruction. Then, a dilated encoder extracts local spectral features, and Competitive Region Attention (CRA) combines ordered relaxation-time priors with cross-query competition to form complementary regional representations. Inter-region interaction subsequently models dependencies among the learned regions for SOH regression. The framework is evaluated on three datasets, including an in-house dataset collected under multiple cycling temperatures and C-rates and two public datasets with different chemistries and aging protocols. Under leave-one-cell-out evaluation, DRA-Net achieves RMSEs of 1.50, 1.03, and 1.54 SOH percentage points, reducing RMSE by 27.6--36.1 relative to the strongest baseline in each dataset. Ablation results identify CRA and the ordered relaxation-time priors as the main contributors to the performance gain. Attention analysis further shows that the learned regions preserve a consistent short-to-long organization while adapting to DRT evolution, and that the intermediate region frequently represents transitions between dominant responses rather than another local maximum. This region-centric view of DRT enables more flexible representation of degradation information and more robust battery SOH estimation under evolving electrochemical conditions.