A semi-supervised meta-learning framework with multi-soft labels and cosine similarity for few-shot cross-domain fault diagnosis

Abstract

Intelligent fault diagnosis of mechanical equipment is crucial to ensure safe industrial operation, but the high cost of manual labeling and the scarcity of fault data seriously restrict the performance of traditional methods. In order to identify faults under the condition of few samples and improve the performance of cross-domain classification performance in semi-supervised learning, this paper proposes a meta-learning method based on multi-label cosine similarity (MCML). This method dynamically assigns multiple high-probability pseudolabels to unlabeled data through a multi-label classifier to retain the potential similarity between categories, introduces cosine similarity as a meta-learning metric to effectively alleviate the sensitivity of Euclidean distance in high-dimensional feature space, and designs a dynamic two-stage Adam optimizer to adaptively adjust the learning rate to accelerate convergence. Experiments show that on the rotor rolling bearing, Case Western Reserve University bearing and Southeast University gearbox datasets, MCML has the highest accuracy of 100% in the 10-way-5-shots sample task, and the cross-equipment and cross-fault domain diagnosis accuracy are 100 and 92.38% respectively, which is significantly better than convolutional neural network, Prototype Networks, and other comparison models. This study provides an efficient solution for small-sample fault diagnosis in industrial scenarios.