TSFA: A Two-Stage Feature Alignment Method for Unsupervised Open-Set Domain Adaptation in Time-Series Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Time Series Analysis

Background:

• Unsupervised open-set domain adaptation (UOSDA) is crucial but challenging for time series due to nonstationarity and distribution shifts.
• These challenges increase the risk of negative transfer in existing UOSDA algorithms.
• Existing methods struggle to effectively adapt models across different operating conditions in time series data.

Purpose of the Study:

• To propose a novel two-stage feature alignment (TSFA) method for UOSDA in time-series classification.
• To address the challenges of nonstationary data and distribution shifts in time series.
• To improve the accuracy and robustness of domain adaptation for time series classification tasks.

Main Methods:

• A time-frequency feature extractor is employed to learn domain-invariant and discriminative representations.
• A two-stage multigranularity feature alignment framework is introduced, including global and local alignment.
• Global alignment uses similarity distribution entropy (SDE) to reduce intraclass distances, while local alignment uses self-supervised learning with target pseudolabels.

Main Results:

• The proposed TSFA method effectively reduces intraclass distances and enhances interclass discriminability.
• An optimal similarity assignment matrix (OSAM) improves pseudolabel accuracy for common samples.
• An adaptive decision boundary effectively rejects private target-domain samples, demonstrating superior performance on real-world datasets.

Conclusions:

• The TSFA method offers a robust solution for UOSDA in time series classification.
• The approach successfully mitigates negative transfer risks associated with nonstationary data.
• Experimental results confirm the superiority of TSFA over state-of-the-art methods in handling complex time-series domain adaptation.