A fault diagnosis framework using unlabeled data based on automatic clustering with meta-learning

Abstract

With the growth of the industrial internet of things, the poor performance of conventional deep learning models hinders the application of intelligent diagnosis methods in industrial situations such as lack of fault samples and difficulties in data labeling. To solve the above problems, we propose a fault diagnosis framework based on unsupervised meta-learning and contrastive learning, which is called automatic clustering with meta-learning (ACML). First, the amount of data is expanded through data augmentation approaches, and a feature generator is constructed to extract highly discriminative features from the unlabeled dataset using contrastive learning. Then, a cluster generator is used to automatically divide cluster partitions and add pseudo-labels for these. Finally, the classification tasks are derived through taking original samples in the partitions, which are embedded in the meta-learner for fault diagnosis. In the meta-learning stage, we split out two subsets from task and feed them into the inner and outer loops to maintain the class consistency of the real labels. After training, ACML transfers its prior expertise to the unseen task to efficiently complete the categorization of new faults. ACML is applied to two cases concerning a public dataset and a self-constructed dataset, demonstrate that ACML achieves good diagnostic performance, outperforming popular unsupervised methods.