A Study on Diffusion Modelling For Sensor-based Human Activity Recognition

Human activity recognition (HAR) is a core research topic in mobile and wearable computing, and has been applied in many applications including biometrics, health monitoring and sports coaching. In recent years, researchers have focused more attention on sensor-based HAR due to the popularity of sensor devices. However, sensor-based HAR faces the challenge of limited data size caused by the high cost of data collection and labelling work, resulting in low performance for HAR tasks. Data transformation and generative adversarial network (GAN) have been proposed as data augmentation approaches to enrich sensor data, thereby addressing the problem of data size limitations. In this paper, we studied the effectiveness of diffusion-based generative models for generating synthetic sensor data as compared to the other data augmentation approaches in sensor-based HAR. In addition, UNet has been redesigned in order to improve the efficiency and practicality of diffusion modelling. Experiments on two public datasets showed the performance of diffusion modelling compared with different data augmentation methods, indicating the feasibility of synthetic sensor data generated using diffusion modelling.