Cross-User Activity Recognition Using Deep Domain Adaptation With Temporal Dependency Information

Sensor-based human activity recognition (HAR) is a cornerstone of ubiquitous computing, playing a crucial role in modern consumer electronics such as fitness trackers and smart home systems. Despite significant advancements, traditional sensor-based HAR methods often assume that training and testing data share identical distributions, an assumption invalidated in real-world scenarios by out-of-distribution ( $ \text {o.o.d.}$ ) challenges, including heterogeneous sensors, temporal changes, and individual behavioral variability. This article addresses the cross-user sensor-based HAR problem, where individual behavioral differences lead to varying data distributions. We introduce the deep temporal state domain adaptation (DTSDA) model, an innovative approach tailored for time-series domain adaptation in cross-user sensor-based HAR. Unlike existing domain adaptation methods that assume sample independence, DTSDA leverages temporal dependencies inherent in sensor data. We introduce the concept of “temporal state” to define subactivities within an activity and ensure their temporal sequence through the “temporal consistency” property. The “pseudo-temporal state labeling” method identifies user-invariant temporal dependency relations. Integrating adversarial learning with temporal dependency knowledge, our method enhances classification performance in cross-user HAR. The efficacy of DTSDA is demonstrated through the evaluations of three public datasets, focusing on daily living and sports fitness activities using consumer electronics.