A Contrastive Learning-Based Fault Diagnosis Method for Rotating Machinery With Limited and Imbalanced Labels

Deep learning, as a good way of learning discriminative features automatically from original data, has been widely used for fault diagnosis in machinery. However, the excellent performance of most deep learning algorithms relies heavily on the characteristics of data for training, including but not limited to a rich set of labeled samples and class-balanced dataset, which are not typical cases in practical applications due to factors such as tagging cost, fault pattern (typical pattern and atypical pattern of machine fault). In this article, a contrastive learning-based fault diagnosis method was developed for rotating machinery with limited and class-imbalanced samples. An improved dual-view wavelet feature attention fusion embedded contrastive network is designed to enhance the diagnostic knowledge extraction capability of the contrastive learning network and also properly make good use of the available fault information contained in a large number of unlabeled datasets. Furthermore, a dynamic weighted loss is designed by matching datasets of various class cases adaptively and dynamically to alleviate the impacts of class imbalance. Compared with other methods with which to diagnose planetary gearbox fault, the experimental results show that the method proposed achieves higher diagnostic accuracy under conditions of limited and imbalanced labels (LIIM).