A Novel Deeper One-Dimensional CNN With Residual Learning for Fault Diagnosis of Wheelset Bearings in High-Speed Trains

• Published in: IEEE access 2019, Vol.7, p.10278-10293
• Main Authors: Peng, Dandan, Liu, Zhiliang, Wang, Huan, Qin, Yong, Jia, Limin
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Background noise; Bearing strength; Bearings; Convolution; deep learning; Diagnostic systems; Fault diagnosis; Feature extraction; High speed rail; High-speed trains; Kernel; Machine learning; one-dimensional residual block; Performance degradation; Safety; Training; Trains; Undercarriages; Vibrations; wheelset bearings fault diagnosis; Wheelsets; wide convolutional kernel
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2018.2888842

The health condition of a wheelset bearing, the key component of a railway bogie, has a considerable impact on the safety of a train. Traditional bearing fault diagnosis techniques generally extract signals manually and then diagnose the bearing health conditions through the classifier. However, high-speed trains (HSTs) are usually faced with variable loads, variable speeds, and strong environmental noise, which pose a huge challenge to the application of the traditional bearing fault diagnosis methods in wheelset bearing fault diagnosis. Therefore, this paper proposes a 1D residual block, and based on the block, a novel deeper 1D convolutional neural network (Der-1DCNN) is proposed. The framework includes the idea of residual learning and can effectively learn high-level and abstract features while effectively alleviating the problem of training difficulty and the performance degradation of a deeper network. Additionally, for the first time, we fully use the wide convolution kernel and dropout technology to improve the model's ability to learn low-frequency signal features related to the fault components and to enhance the network's generalization performance. By constructing a deep residual learning network, Der-1DCNN can adaptively learn the deep fault features of the original vibration signal. This method not only achieves very high diagnostic accuracy for the fault diagnosis task of wheelset bearings in HSTs under strong noise environment, but also its performance is quite superior when the train's working load changes without any domain adaptation algorithm processing. The proposed Der-1DCNN is evaluated on the dataset of the multi-operating conditions of the wheelset bearings of HSTs. Experiments show that this method shows a better diagnostic performance compared with the state-of-the-art deep learning methods of bearing fault diagnosis, which proves the method's effectiveness and superiority.