Knowledge-aided self-supervised deep representation learning method for few-shot fault diagnosis

Machine learning technology has been widely used in industrial intelligent fault diagnosis, but the prerequisite for its successful application is to obtain enough labeled fault data to train the machine learning model. In actual industrial scenarios, equipment often works in normal state, and the cost of obtaining fault data and marking fault labels is huge. Therefore, the training requirements of machine learning model cannot be guaranteed, and then it is difficult to apply the model to the target equipment. To solve this problem, this paper explores the intrinsic relationship between key temporal dependent features, expert prior knowledge and fault diagnosis tasks, and proposes a knowledge self-supervised deep representation learning method for few-shot fault diagnosis. In this method, a model pre-training strategy combining prior feature prediction and mask signal reconstruction is designed. The feature extractor model of the industrial intelligent fault diagnosis model is pre-trained by using massive historical unlabeled data accumulated by similar equipments. This pre-training strategy can make the model add knowledge guidance of artificial prior features on the basis of mining temporal dependent features of samples, so as to obtain high generalization fault representation ability. After training the industrial fault diagnosis model based on the knowledge self-supervised representation learning method, only a few labeled fault samples of the target device are used to fine-tune the global parameters of the model in the diagnosis process, and the problem of model dependence on a large number of labeled samples will be solved. Finally, a cross-dataset fault diagnosis experiment is conducted to simulate the cross-equipment fault diagnosis scenario with few samples, and the effectiveness of the proposed method in the small-sample scenario is verified. © 2024 Northeast University. All rights reserved.