Robust Elbow Angle Prediction With Aging Soft Sensors via Output-Level Domain Adaptation

Abstract

Wearable devices equipped with soft sensors provide a promising solution for body movement monitoring. Specifically, body movements like elbow flexion can be captured by monitoring the stretched soft sensors’ resistance changes. However, in addition to stretching, the resistance of a soft sensor is also influenced by its aging, which makes the resistance a less stable indicator of the elbow angle. In this paper, we leverage the recent progress in Deep Learning and address the aforementioned issue by formulating the aging-invariant prediction of elbow angles as a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">domain adaption</i> problem. Specifically, we define the soft sensor data ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e.</i> , resistance values) collected at different aging levels as different domains and adapt a regression neural network among them to learn domain-invariant features. However, unlike the popular pairwise domain adaptation problem that only involves one source and one target domain, ours is more challenging as it has “infinite” target domains due to the non-stop aging. To address this challenge, we novelly propose an output-level domain adaptation approach which builds on the fact that the elbow angles are in a fixed range regardless of aging. Experimental results show that our method enables robust and accurate prediction of elbow angles with aging soft sensors, which significantly outperforms supervised learning ones that fail to generalize to aged sensor data.