Deep Subdomain Transfer Learning with Spatial Attention ConvLSTM Network for Fault Diagnosis of Wheelset Bearing in High-Speed Trains

Abstract

High-speed trains operate under varying conditions, leading to different distributions of vibration data collected from the wheel bearings. To detect bearing faults in situations where the source and target domains exhibit differing data distributions, the technique of transfer learning can be applied to move the distribution of features gleaned from unlabeled data in the source domain. However, traditional deep transfer learning techniques do not take into account the relationships between subdomains within the same class of different domains, resulting in suboptimal transfer learning performance and limiting the use of intelligent fault diagnosis for wheel bearings under various conditions. In order to tackle this problem, we have developed the Deep Subdomain Transfer Learning Network (DSTLN). This innovative approach transfers the distribution of features by harmonizing the subdomain distributions of layer activations specific to each domain through the implementation of the Local Maximum Mean Discrepancy (LMMD) method. The DSTLN consists of three modules: a feature extractor, fault category recognition, and domain adaptation. The feature extractor is constructed using a newly proposed SA-ConvLSTM model and CNNs, which aim to automatically learn features. The fault category recognition module is a classifier that categorizes the samples based on the extracted features. The domain adaptation module includes an adversarial domain classifier and subdomain distribution discrepancy metrics, making the learned features domain-invariant across both the global domain and subdomains. Through 210 transfer fault diagnosis experiments with wheel bearing data under 15 different operating conditions, the proposed method demonstrates its effectiveness.