The unsupervised open-set domain adaptation (UOSDA) in computer vision is widely studied. However, designing UOSDA algorithms for time series remains challenging due to the complex nonstationary property of data and distribution shifts across different operating conditions, which heighten the risk of negative transfer. To address this, a novel two-stage feature alignment (TSFA) method for UOSDA in time-series classification is proposed. First, a time-frequency feature extractor is designed to effectively learn domain-invariant and discriminative interclass representations. Subsequently, a two-stage multigranularity feature alignment framework is proposed. Specifically, global alignment is achieved by minimizing the similarity distribution entropy (SDE), which encourages similar samples from both domains to converge in the feature space, thereby reducing intraclass distances of all target samples. Then, local alignment is performed via self-supervised learning guided by target pseudolabels, enhancing interclass discriminability of target common samples. Furthermore, an optimal similarity assignment matrix (OSAM) is designed to obtain more accurate pseudolabels for common samples, while an adaptive decision boundary is introduced to effectively reject private samples based on target-domain statistics. Finally, extensive experimental results on three real-world datasets indicate that the proposed method outperforms state-of-the-art approaches.
Wang et al. (Thu,) studied this question.