Transforming wearable sensor data for robust feature selection in human activity recognition using reinforcement learning approach.

Abstract

The practical applications of body sensor data in smart healthcare systems have drawn a lot of attention from researchers studying healthcare. Current models have trouble capturing and classifying data, especially when massive datasets are involved. This study makes use of time-sequential data and the deep reinforcement learning technique known as Generative Actor-Critic (GAC). Wearable sensor data collection makes feature selection easier by enhancing inter-class differences and decreasing intra-class variations. For robust activity modeling, deep reinforcement learning and cyclic Generative Adversarial Networks are integrated with GAC and strong temporal-sequential features. This method outperforms traditional deep learning techniques in achieving accurate recognition despite noise, with accuracy of 98.76% on UCI-HAR and 98.84 % on Motion Sense datasets.