BiLSTM-AM: A Deep Learning Model for Indirect Identification of Track Irregularity Using In-Service Vehicle Acceleration Responses.

The precise and effective identification of track irregularity is of great significance in ensuring the safety of train operations. The study proposes a novel deep learning model BiLSTM-AM, Bi-directional Long Short-Term Memory (BiLSTM) model incorporating attentional mechanisms (AM), for indirectly identifying track irregularity by utilizing in-service vehicle acceleration responses. The BiLSTM containing a forward LSTM network and a backward LSTM network is employed to capture the dependence of vehicle acceleration and track irregularities at different locations. The joining of AM can further improve the accuracy and robustness of the model by emphasizing the importance of the key information that has a substantial impact on the identification results. The identification precision is verified by comparing the different types of identified track irregularities with the true values, and its applicability and robustness are further investigated by parameter study for the effects of key parameters, including the number of inputting acceleration responses, running speed, noise level and bridge damping ratio. The results indicate that the model applies to the identification of track irregularity in several complex scenarios, even for the speed running at 350 km/h, bridge damping ratios of 0.05, and noise levels up to 20%. Besides, the great significance of the proposed model is that only one vehicle acceleration is required to obtain satisfactory recognition accuracy in practical applications. [ABSTRACT FROM AUTHOR]