Human Activity Recognition (HAR) from wearable sensor data faces significant challenges due to limited labeled data. We propose a novel weakly self-supervised framework that leverages domain knowledge and large volumes of unlabeled data to learn discriminative and robust activity representations …