Multi-modal semi-supervised approach to domain adaptive human activity recognition

This study presents a deep semi-supervised domain adaptation framework, termed Enhanced SWL-Adapt, designed to address key challenges in cross-user Human Activity Recognition (HAR), namely limited labeled data and significant inter-user distribution shifts. Unlike conventional methods that struggle with hardware noise and individual variability, the proposed approach employs a multi-modal input module that decomposes accelerometer, gyroscope, and magnetometer signals into separate streams and processes them through a dynamic sensor attention mechanism. A three-branch optimization strategy is adopted to learn discriminative and domain-invariant representations. In addition, a meta-learning-based sample weight allocator and a multi-level alignment mechanism are introduced to handle heterogeneous user shifts. Experimental results on the OPPORTUNITY dataset demonstrate that the proposed method consistently outperforms state-of-the-art approaches, achieving a 2.1% improvement in accuracy and a 1.87% gain in Macro-F1 score.