Unsupervised Deep Representation Learning to Remove Motion Artifacts in Free-mode Body Sensor Networks

In body sensor networks, the need to brace sensing devices firmly to the body raises a fundamental barrier to usability. In this paper, we examine the effects of sensing from devices that do not face this mounting limitation. With sensors integrated into common pieces of clothing, we demonstrate that signals in such free-mode body sensor networks are contaminated heavily with motion artifacts leading to mean signal-to-noise ratios (SNRs) as low as -12 dB. Further, we show that motion artifacts at these SNR levels reduce the F1-score of a state-of-the-art algorithm for human-activity recognition by up to 77.1%. In order to mitigate these artifacts, we evaluate the use of statistical (Kalman Filters) and data-driven (Neural Networks) techniques. We show that well-designed methods of representing IMU data with deep neural networks can increase SNRs in free-mode body-sensor networks from -12 dB to +18.2 dB and, as a result, improve the F1-score of recognizing gestures by 14.4% and locomotion activities by 55%.