Contrastive Learning for Zero-Shot Human Activity Recognition Using Labeled Simple Actions on Wearable Devices

Existing Zero-Shot Learning (ZSL) approaches for sensor-based Human Activity Recognition (HAR) often rely on external semantic information—such as attribute annotations or textual descriptions—which are frequently unavailable or ambiguous in real-world scenarios. To address this challenge, we propose a contrastive learning model that combines Temporal Convolutional Networks (TCN), Bidirectional Gated Recurrent Units (BiGRU), and Transformer encoder, without relying on any external semantics. The model learns structured representations directly from raw sensor signals by capturing both temporal and structural similarities among basic activities. The resulting embedding space enables the model to generalize effectively to previously unseen or composite activities in a zero-shot setting, where no target-class examples are seen during training. It achieves high supervised performance with F1 scores of 0.9731 on UCI-HAR, 0.9788 on WISDM, 0.9459 on PAMAP2, and 0.9835 on MHEALTH. Furthermore, the model demonstrates robust inference across domains and performs reliably even in ZSL scenarios without auxiliary labels. These findings suggest a promising direction for building scalable HAR systems that require no predefined semantic information, allowing them to flexibly adapt to complex and evolving real-world activities.