Smeta-LU: A self-supervised meta-learning fault diagnosis method for rotating machinery based on label updating

Abstract

During operation of rotating machinery, collecting high-quality labeled fault samples is difficult, and the corresponding data annotation is time consuming and costly. Therefore, developing novel intelligent diagnostic methods which can extract key information from massive fault data without labeling is of great significance. In this regard, a self-supervised meta-learning fault diagnosis method for rotating machinery based on label updating, called Smeta-LU, is proposed. It eliminates the pre-training phase and generates meta-tasks directly without labeling information during training. A two-branch framework in Smeta-LU is developed using a contrastive learning approach, which involves the application of a dynamic dictionary to construct samples for one branch, represented by an online encoder. The other branch utilizes the parameters of the former to obtain a target encoder through exponential moving average. To dynamically construct diverse meta-tasks during the meta-training process, each sample in the current batch is treated as a query set, while the support set is selected from queues to construct few-shot tasks, thereby generating a larger pool of candidates. The fault diagnosis task is completed by assigning the label matrix with an optimal transport algorithm and identifying the shots closest to each of the prototype centers. Additionally, the iterative properties of the momentum network and dynamic dictionary are implemented for label updating. The outcomes of two validation experiments demonstrate the superiority and scalability of our self-supervised meta-learning approach compared with conventional supervised meta-learning techniques. Better performance in identifying new fine-grained fault categories is also exhibited during our research.