Abstract 

Fault diagnosis of rotating machinery is essential to minimize damage and downtime in industrial fields. With the development of artificial intelligence, deep-learning-based fault diagnosis has gained significant attention. However, changes in the data distribution from machinery operating under different conditions have led to insufficient diagnostic accuracy. Additionally, the lack of labeled data in industrial settings hampers the performance of these deep-learning algorithms. To address these issues, unsupervised domain adaptation (UDA)-based fault diagnosis methods have been increasingly explored for robust diagnosis under varying conditions. Traditional UDA methods, however, struggle to adapt to hard-to-adapt classes as they focus only on reducing global distribution discrepancies, leading to misclassification and reduced performance for these classes. In this paper, we propose a latent space alignment based domain adaptation (LSADA) approach to overcome this limitation. LSADA reduces local distribution discrepancies by sequentially aligning minority regions and minimizing the distance between source and target data in high-dimensional latent space. Additionally, the feature extractor and predictor in LSADA are synchronized by generating reliable pseudo labels from unlabeled target data. The proposed method is validated using both open-source and experimental datasets, demonstrating that LSADA outperforms existing UDA-based fault-diagnosis algorithms. Moreover, a physical analysis of the method addresses the black-box issue, a common limitation of deep-learning approaches.