Inter-Domain Sensor Alignment for Unsupevised Domain Adaptation of Wearable Multivariate Time Series

Jaehyun Bae, Heesoo Jung, Hankook Lee, Hogun Park

Keywords: Domain Adaptation, Multivariate Time Series

Abstract:

Unsupervised domain adaptation (UDA) for multivariate time-series (MTS) data in the wearable domain transfers knowledge from a labeled source to an unlabeled target, typically with signals collected from multiple body-worn sensors. Although existing UDA methods devote substantial effort to modeling temporal shifts, they often rely on simple spatial alignment across domains, thereby limiting their capacity for effective adaptation.
Real systems in the wearable domain exhibit \emph{sensor-wise domain shift}, including changes in placement or orientation, which necessitates the explicit consideration of inter-domain spatial sensor relations.
Therefore, we introduce \textbf{IDSA}, \textit{\textbf{I}nter-\textbf{D}omain \textbf{S}ensor \textbf{A}lignment for wearable MTS-UDA}, a plug-in module that augments any base UDA loss with two complementary components: (i) an \textit{inter-domain sensor transport} that learns a cross-sensor relation matrix from domain-specific sensor embeddings and transports target channels toward the source, and (ii) a \textit{channel decorrelation} regularizer that sparsifies intra-domain graphs to suppress redundant or noisy couplings. Our sensor transportation loss is shown to be equivalent (up to a constant) to the discrete $1$-Wasserstein objective. When used as a plug-in with Deep CORAL or CLUDA, IDSA achieves consistent gains across five HAR and sEMG benchmarks compared to recent baselines in activity classification accuracy, achieving a performance enhancement in most scenarios.

Primary Area: learning on time series and dynamical systems

Submission Number: 10862