Contrastive Weakly Supervised Representation Learning for Robust and Annotation-Efficient Human Activity Recognition with Wearable Sensors

Wearable sensors and Human Activity Recognition (HAR) have turned into the technology of the foundation of healthcare monitoring and smart living environments, analytics in sports, and systems of rehabilitation. Although there are notable improvements in deep learning architecture time-series modeling, the overwhelming majority of the state-of-the-art HAR systems depend on the presence of fully annotated datasets, which are expensive, require human labor, and are not always available on large scale. Moreover, wearable sensor data is prone to noise, subject variability, domain shifts, and inconsistent labelling and thus, in real-world applications, wearable sensor data results in worse generalization performance. The current paper suggests a weakly supervised paradigm of representation learning that is contrastive, and it is expected to be more robust and much less reliant on annotation. By combining contrastive self-supervised tasks with weakly labeled input, the proposed method will learn both discriminative and invariant time-series representations using multivariate time-series signals produced by wearable devices. It has a framework that includes data augmentation policies, consistency regularization, pseudo-label refinement, and noise-sensitive training systems to reduce the effects of small and low-quality annotations. The experimental studies of the proposed method in different label scarcity and simulated noise levels show that the proposed approach has better performance than fully supervised and purely self-supervised baselines. The findings demonstrate significant improvements in the quality of representations, robustness and efficiency of annotations, and the way to scalable and economical systems of HAR. This paper can be used in the further development of annotation-efficient deep learning techniques in sensor-based smart systems.