Cross-Domain Complementary Dual-Stream Deep Self-Supervised Learning Framework for Human Activity Recognition on Wearable Devices

Abstract To reduce the reliance on large-scale annotated data in Human Activity Recognition, this paper proposes a deep self-supervised learning framework that integrates temporal and wavelet-coefficient information. The framework employs Discrete Wavelet Transform to decompose raw signals collected from wearable sensors into multi-scale time–frequency subbands. Deep features are extracted only from the low-frequency subband, after which the signal is reconstructed via Inverse Discrete Wavelet Transform and combined with the original temporal signal to form a dual-stream input, enabling cross-domain feature complementarity. UMAP is used for dimensionality reduction, and a Hidden Markov Model is applied for clustering to generate pseudo-labels, with a label-consistency weighting mechanism introduced for iterative refinement. Experiments on four public datasets—PAMAP2, UCI-Sm, WISDM-v1, and WISDM-watch—demonstrate that the proposed method significantly outperforms existing unsupervised approaches in terms of Accuracy, Normalized Mutual Information, Adjusted Rand Index, and F1 Score, achieving 94.3% ACC on the PAMAP2 dataset. This work offers a novel technical pathway for self-supervised activity recognition.