GAN-based anomaly detection using fusion of photoplethysmography and accelerometer in a wearable device

Abstract

This dissertation proposes an automatic fall detection system in a wearable device that can reduce risks by detecting falls and promptly alerting caregivers. For this purpose, we propose a novel user-adaptive fall detection system using a fusion of photoplethysmography (PPG) and accelerometer. One of the objectives is the proposed fall detection should be trained by using normal behavior data only to have high applicability in an actual environment. To meet the objective, unsupervised anomaly detection-based methods are attempted. Since the target application is a wearable device, the proposed fall detection system should be considered to have minimum energy consumption and computational complexity because of the heavy computational requirements of deep learning. Therefore, the proposed system is applied adaptive local offloading to save the energy of wearables and to have computational efficiency, which means the traditional machine learning approach for wearables and deep learning for handhelds. For the traditional one, the most effective feature subset of PPG and accelerometer is proposed which is the input of the clustering-based model for the purpose of demonstrating reliable performance and designing a low-complexity model. In addition, from this approach, this study also shows the effectiveness of the user-adaptive method when using both PPG and accelerometer signals and the performance increment of combining a PPG with an accelerometer. For the deep learning method, this study presents the first attempt to applying generative adversarial network (GAN)-based anomaly detection for fall detection by presenting an effective GAN-based model among the recently proposed nine models. In addition, the compelling GAN-based anomaly detection with partially surrounded feature-level data by the user's resting heart rate information (UI-GAN) is introduced. From the UI-GAN, the fall detection performance improvement is proved when using the user heart rate initial information. Moreover, the proposed UI-GAN is exhibited better performance than 20 conventional fall detection studies. Lastly, the proposed fall detection system is proved to be applied in wearables and handhelds through the feasibility test. This study is the first attempt to using initial information of users in GAN, and the effectiveness of the user initial information can be expected in other applications.