A GAN-Based Data Augmentation Approach for Sensor-Based Human Activity Recognition

Recently, deep learning has emerged as a powerful technique and been successfully employed for various tasks. It has also been applied to human activity recognition and showed better performance than traditional machine learning algorithms. However, the success of deep learning always comes with large labeled datasets when the learning model goes deeper. If the training data is limited, the performance of the classification model may not generally perform well due to overfitting of the networks to the training data, which can be alleviated through data augmentation. Generative adversarial networks (GANs) can be used as a technique to produce data artificially. GAN-based approaches have made rapid progress in generating synthetic data, but they are mostly studied for image data. Comparatively little research has been conducted to examine the effectiveness of generating sensor data using GANs. This study aims to investigate the data scarcity problem by using conditional generative adversarial networks (CGANs) as a data augmentation method. The proposed approach was experimentally evaluated on a benchmark sensor dataset for activity recognition. The experimental results showed that the proposed approach can boost the model accuracy and has better performance when compared with existing approaches.