A Multi-Task BiLSTM-Attention Temporal Network for Simultaneous Activity Classification and Knee Joint Angle Estimation
Authors: He, Z., Wang, Y., Yuan, M., Su, T., Yu, H.
Journal: IFAC Papersonline
Publication Date: 01/01/2025
Volume: 59
Issue: 35
Pages: 413-418
eISSN: 2405-8963
ISSN: 2405-8971
DOI: 10.1016/j.ifacol.2025.12.512
Abstract:Wearable sensors have enabled portable and continuous monitoring of human motion, playing an important role in rehabilitation and biomechanical analysis. However, most existing studies treat activity recognition and joint angle estimation as separate tasks, which overlooks their intrinsic correlation and increases computational cost. To address this limitation, we propose a multi-task BiLSTM-attention temporal network (MT-BiATNet) that jointly performs activity classification and bilateral knee joint angle estimation within a unified framework. The network integrates one-dimensional convolutional layers, bidirectional LSTM, and attention mechanisms to extract multi-scale temporal features and model both global and task-specific dependencies. Using the ENABL3S dataset, subject-dependent 10-fold validation achieved over 0.86 accuracy and less than 3° RMSE (best 1.93°/1.98°), while leave-one-subject-out evaluation reached 0.63-0.76 accuracy, demonstrating reasonable cross-subject robustness.
Source: Scopus