SenTS: A Unified Time-Spectral Modeling Framework with Periodic Regularization for Sensing Behavior Discrimination

Wearable devices are increasingly deployed in applications such as health monitoring, motion analysis, and intelligent interaction, making inertial sensor sequences a key modality for understanding human actions. However, these signals are often affected by noise, pose variations, device heterogeneity, and periodic drift, exhibiting non-stationary local dynamics and unstable periodic structures. Such characteristics make it challenging to robustly characterize motion patterns and hinder cross-scenario generalization. To address these issues, we propose SenTS, a unified time–spectral modeling framework for wearable motion sequences. SenTS jointly models local variations and long-range dependencies via multi-kernel dynamic feature projection and kernelized temporal attention. First, we extract stable rhythmic cues using a learnable spectral filter and a spectral-attention module. Second, we integrate complementary temporal- and spectral-domain representations through dual-domain feature fusion. Finally, we introduce periodic structural regularization to enhance cross-period stability. Based on a large-scale wrist-raise dataset collected from 50 participants using three smartwatch models across standing and walking scenarios, extensive experiments demonstrate that SenTS consistently outperforms advanced methods in single-device and cross-device evaluations. Moreover, SenTS exhibits strong discriminative ability in smartwatch identity authentication under changes in wearing position and orientation, further validating the robustness and real-world transferability of its time–spectral representations.