Publication

Abstract : The proposed work explores Time Series Classification (TSC) across various applications, including human activity recognition, healthcare, and machine fault diagnosis. These temporal data from disparate domains pose a significant challenge in TSC due to the inherent variability in their distributions. Unsupervised Domain Adaptation (UDA) methods are an effective solution for addressing these distribution disparities. The proposed work initially utilizes a norm-distance and correlation alignment to achieve feature similarity and statistical alignment between the source and target domains. However, these techniques may not capture the domain’s persistent features due to the complex characteristics of the temporal data. Hence, domain and virtual adversarial training are applied to acquire invariant feature representations across domains globally, ensuring local smoothness in the model’s output distribution. Therefore, we put forward a unified approach to Virtual Adversarial and Statistical unsupervised domain adaptation for TSC (VASAD). As a result, the model learns the intrinsic relationships among the temporal data points, enhances the model’s robustness to perturbations, and aligns the domains into a shared subspace, thereby minimizing distribution discrepancies. A thorough experimental analysis is done using different datasets related to human activity, fault diagnosis, and sleep classification. The results are compared against state-of-the-art methods, and the proposed approach consistently outperforms them.