Fault Diagnosis of Wheelset Bearings using Deep Bidirectional Long Short-term Memory Network

Abstract

The wheelset bearing is the core component of a railway bogie, whose health status is critical to guarantee the safety of high-speed trains (HSTs). However, the traditional data-driven fault diagnosis methods are mostly founded on the feature extraction methods requiring extensive domain knowledge and experience, compared with deep learning approaches that can learn hierarchical representations automatically from data. Recurrent neural network (RNN) which can learn features from sequential data directly without any feature engineering has been proved to be effective in the research area of machine failure diagnosis. Therefore, an innovative fault diagnosis method for the wheelset bearings in the HSTs using Deep Bidirectional Long Short-term Memory Network (DBLSTM) is proposed in this paper. Long Short-term Memory Network (LSTM), as an improved framework of Recurrent Neural Network, is able to overcome the gradient vanishing or exploding problem, capturing long-term dependencies effectively. In the DBLSTM, the bidirectional structure is applied to enhance the performance of the LSTM Network by capturing temporal information from both future and past contexts of input sequential data. In addition, by stacking multiple Bidirectional LSTM layers to build the DBLSTM network, for the vibration signals measured under the complicated environment, more complex fault features can be effectively learned. With the increasement of the depth of the network, the regularization method using recurrent dropout technique is introduced to relieve the problem of overfitting and enhance the generalization ability of the deep network. The DBLSTM is assessed on the dataset of the wheelset bearings of the HSTs under different operating states. The experimental results indicate that the DBLSTM can accurately diagnose the fault mode of the wheelset bearings of the HSTs under strong noise, outperforming five existing deep learning methods used in fault diagnosis.