Unsupervised Sensor-Based Human Activity Recognition via Contrastive Learning of Representation

Human activity recognition (HAR) based on wearable sensors has attracted increasing attention due to its privacy-preserving nature and ubiquitous deployment. However, the scarcity of large-scale labeled sensor data poses a significant challenge for supervised learning, motivating the adoption of unsupervised representation learning methods. Among them, contrastive learning (CL) has shown strong potential, yet its effectiveness critically depends on how self-supervised signals are induced by data augmentation strategies. In this article, we proposeSensorCLR, a CL framework for unsupervised sensor-based HAR that explicitly addresses the limitations of segmentation-based augmentation (SA) designed for sensor data. At the core of SensorCLR, we introduceDualMoCo, a symmetric dual-branch MoCo-style architecture that enables effective cross-segment CL. By sharing the same query encoder across both branches, this design alleviates the misalignment issues caused by the asymmetric query–key structure of conventional MoCo and ensures that meaningful activity patterns in both temporally segmented views are learned in a consistent and balanced manner. Furthermore, we propose a distributional consistency loss that transfers distribution-level supervision from weakly augmented views to their strongly augmented counterparts, thereby mitigating semantic distortion caused by aggressive augmentations such as scaling and jittering. Importantly, the joint design of DualMoCo and the distributional consistency loss enables SensorCLR to recover effective self-supervision signals when SA becomes unreliable for nonrepetitive or weakly structured activities. Extensive experiments on five benchmark datasets, such as OPPO, PAMAP2, UCI-HAR, UniMiB, and WISDM, demonstrate that SensorCLR consistently achieves state-of-the-art performance under both linear evaluation (LE) andk-nearest neighbor (KNN) evaluation protocols. Moreover, on-device deployment experiments on a Raspberry Pi platform confirm that SensorCLR improves representation quality without increasing inference cost, making it suitable for real-world edge deployment. The source code will be available athttps://github.com/KennCoder7/SensorCLR