CACDT: an approach to cross-machine bearing fault diagnosis

Abstract

Abstract Mechanical fault transfer diagnosis utilizes the acquired diagnostic knowledge of machinery to address diagnostic issues in the target machinery. This approach demonstrates promising results in overcoming the limitations of incomplete fault information and scarce labeled data in the era of big data. However, when confronted with cross-machine fault diagnosis, the significant domain discrepancies pose challenges to traditional fault diagnostic methods, leading to lower accuracy and learning efficiency. To overcome these problems, this work introduces a novel cross-machine bearing fault diagnosis model called Cross-Domain Adaptive Clustering and Dynamic Threshold. The model comprises a feature extraction network and a classifier, and it achieves intra-domain and inter-domain adaptation via adversarial optimization. The feature extraction network minimizes the adversarial adaptive clustering loss, while the classifier maximizes it. Moreover, the model calculates dynamical thresholds for each class in the target domain and generates pseudo-labels for unlabeled samples. This approach increases labeled samples for each category during early training, resulting in a more robust clustering core and improving the learning efficiency of the model. Experimental results show that, in cross-machine fault diagnosis, when the number of labels in the target domain is 5 and 10, the average accuracy reaches 82% and 95.6%, respectively, which is better than the comparison method. The model effectively distinguishes minority samples in class imbalance experiments, and the dynamic thresholds enhance learning efficiency for complex datasets compared to fixed thresholds.