Dual Classifier-Discriminator Adversarial Networks for Open Set Fault Diagnosis of Train Bearings

Fault diagnosis of train bearings is crucial to ensure the reliability and safety of a running train. However, it is a big challenge to recognize new fault types of the train bearings under variable working conditions and heavy noise environments. This article proposes a new domain adaptation model, called dual classifier-discriminator adversarial networks (DCDAN), for open set fault diagnosis of the train bearings. The main contributions of the proposed DCDAN are that a novel weighting strategy is designed by constructing a weighting module with a dual classifier-discriminator structure to separate the new fault types in the target domain from the shared health types between the source and target domains, and a parallel channel attention module (PCAM) is embedded in the feature extractor of the DCDAN to promote feature extraction capability from noisy monitoring data. Specifically, the monitoring data are first input to the feature extractor to extract rich key health state information with the help of the PCAM. Then, the features are input to the weighting module to learn credible weights by unifying the similarity between the samples in the two domains evaluated from different perspectives. Finally, the weights are assigned to the adversarial training between the feature extractor and one of the classifiers for accurate separation of the new fault types and identification of the shared health types. Experimental results on two train bearing datasets verified the effectiveness and superiority of the proposed method, indicating that the proposed method has great potential for application in practical new fault recognition of train bearings.

I. Introduction

Trains usually operate under variable speed and heavy load conditions, resulting in a very harsh service environment of the transmission system [1], [2]. Bearing faults account for a large proportion of the fault types of the train transmission system, which may cause economic losses, or even lead to derailment, subversion, and other malignant accidents if they were not timely diagnosed [3]. To improve the intelligent level of health management of the train bearings, some scholars have applied deep learning methods to diagnose the train bearing faults [4], [5]. However, changes in the actual working conditions (e.g., speed, load, etc.) of the train bearings lead to different data distributions of the monitoring vibration signals, and new types of bearing faults that never appeared would probably happen, which greatly degrades the performance of the traditional deep learning methods in application to fault diagnosis of the train bearings. Hence, how to diagnose bearing new types of faults under variable working conditions is of great significance to guarantee the reliability and safety of the train bearings.