Structure-Regularized Clustering Domain Adaptation for Lying Posture Recognition Under Category Shift

Lying posture recognition plays a critical role in health-aware smart systems, such as smart beds and sleep monitoring devices. In this study, we utilize SleepMatrix®, a piezoelectric-array-based smart sensing platform, to collect high-resolution in-bed posture data across varied deployment conditions. However, practical deployment faces significant challenges due to limited labeled data and distribution shifts across users, particularly when target domains contain only a subset of source domain categories. To address these issues, we propose a Structure-Regularized Domain Clustering method with Soft selection (SRDC-S) for cross-user posture recognition. Structure-regularized clustering dynamically optimizes cluster centers to align feature distributions between domains under category misalignment, while soft source sample selection reduces negative transfer by weighting source samples based on similarity to target clusters. Additionally, we integrate data augmentation with consistency loss to enhance model robustness. Experimental results demonstrate that SRDC-S achieves 85.07% accuracy in cross-user posture recognition tasks with category misalignment, confirming its effectiveness for real-time deployment in embedded healthcare systems.