A modified bidirectional long short-term memory neural network for rail vehicle suspension fault detection

Faults of railway suspension units may endanger vehicle operation safety. Therefore, timely fault detection of railway suspension is necessary. The reported fault detection method that uses Bidirectional Long Short-Term Memory (BiLSTM) neural network in a denoising autoencoder fashion suffers from a shortcoming that the condition monitoring data under faulty states needs to be ultilized to determine the added noise level. Such data is not always available in practice. In this paper, we modified the BiLSTM such that the output data point is excluded from the input when predicting this specific output data point. The requirement of using condition monitoring data under faulty states is therefore relaxed, as the modified BiLSTM is free from determining the level of added noise. A case study on railway suspension fault detection was presented. We used the railway vehicle dynamics model to generate car body vibration under healthy, three levels of reduced first suspension stiffness states, and three levels of reduced first suspension damping states. We compared the performance of the modified BiLSTM with the linear autoregression model, LSTM, and BiLSTM-DAE models. The results have shown that the modified BiLSTM-based fault detection method performs the best, having the highest area under the receiver operating curve.

Automated summary

This study proposes an improved method for fault detection of railway suspension by modifying the BiLSTM neural network, which solves the problem of determining the noise level using monitoring data under faulty states. The experimental results demonstrate that the modified BiLSTM-based method outperforms other models in fault detection.