Sensor Data Augmentation by Resampling in Contrastive Learning for Human Activity Recognition

While deep learning models have contributed to the advancement of sensor-based human activity recognition (HAR), it usually requires large amounts annotated sensor data extract robust features. To alleviate limitations annotation, contrastive has been applied HAR. One essential factors is augmentation, significantly impacting performance pretraining. However, current popular augmentation methods do not achieve competitive in for Motivated by this issue, we propose a new method resampling, which introduces variable domain information and simulates realistic varying sampling frequency maximize coverage space. The resampling was evaluated supervised [SimCLR HAR (SimCLRHAR) MoCo (MoCoHAR)]. In experiment, use four datasets, UCI-HAR, MotionSense, USC-HAD, MobiAct, using mean F1-score as evaluation metric downstream tasks. experimental results show that outperforms all state-of-the-art with small amount labeled data. also demonstrate positive effects frameworks.