An Adaptive TimeGAN-Augmented Attention for Multi-IMU Human Activity Recognition

Human Activity Recognition (HAR) with wear able multi-Inertial Measurement Unit (IMU) systems is challenged by limited labeled data, sensor instability, and real-time constraints. This paper proposes an adaptive TimeGAN-augmented fully convolutional framework for robust and efficient multi-IMU HAR. TimeGAN is used to generate high-fidelity synthetic data, alleviating data scarcity while preserving temporal and statistical characteristics. Based on augmented data, the Hybrid Activity Recognition Topology (HART) is developed, combining1Dconvolutional layers and Squeeze-and-Excitation (SE) channel attention to achieve robust feature representation under sensor variability. An adaptive online updating mechanism is further introduced to enable continuous model refinement during inference, supporting fast and accurate recognition in dynamic environments. Experimental results on multiple datasets demonstrate that the proposed framework achieves superior accuracy with reduced computational cost and strong robustness under varying conditions.