Semi-supervised multi-scale attention-aware graph convolution network for intelligent fault diagnosis of machine under extremely-limited labeled samples

Abstract

Labeled data are generally scarce in engineering practice, while data-driven methods fail to mine the correlations between samples to utilize the rich unlabeled data, so they cannot achieve satisfactory performance under limited labeled data. To address this problem, a semi-supervised multi-scale attention-aware graph convolution network (MSA-GCN) is proposed for fault diagnosis under extremely-limited labeled samples. First, available labeled data are transformed with unlabeled data into a graph via determining the k-nearest neighbors in frequency domain to construct the neighbor relations. To obtain the useful structural and feature information of unlabeled samples from different neighborhoods, multi-scale graph convolution is implemented to aggregate multi-scale information for labeled samples. Besides, attention mechanism is utilized and a novel adaptive feature fusing layer is designed to achieve cross-scale information fusion of different neighborhoods. With semi-supervised graph learning, the proposed method can fully utilize topological and feature information from unlabeled samples, resulting in a powerful classifier using only few labeled samples. The proposed method is fully verified on three bearing datasets, experimental results show that MSA-GCN can reach an identification accuracy of above 95 % with even as few as 5 labeled training samples each class, which demonstrates its effectiveness under low-label-ratio data.