A multi-task model for failure identification and GPS assessment in metro trains

Abstract

<p>Air and oil leaks are two of the predominant operational failures in metro trains, which can cause severe issues and a lot of downtime. Predictive maintenance on such machinery can be of great use. This work aimed to develop a deep learning algorithm for fault analysis in metro trains. The MetroPT dataset was used for this work. A multi-task artificial neural network was developed for the simultaneous identification of failures and GPS quality assessment. The network had common dense, batch normalization, and Gaussian noise layers, followed by output sigmoid layers for each output. The algorithm was trained for 20 epochs with a batch size of 5000 using the using Adam optimizer. The local interpretable model agnostic explanations (LIME) technique was used to provide explanations for the model predictions. Finally, a dashboard was developed for the same application consisting of the best-trained algorithm for decision-making, along with trend visualizations and explanations. The developed multi-task model produced 98.89$ \% $, 99.12$ \% $, and 99.24$ \% $ accuracies in the testing set for failure type, failure location, and GPS quality predictions, respectively. The model produced 99.56$ \% $, 99.67$ \% $, and 99.84$ \% $ precision in the testing set for failure type, failure location, and GPS quality predictions, respectively. The loss values for the trained model on the testing set were 0.0035, 0.0026, and 0.0033 for the three tasks, respectively. The deep learning model took 43 seconds for training and 1 second for inferencing for test data. The LIME technique produced explanations for each predictive task with feature importance in positive and negative impacts. On the whole, the proposed framework can be effective for fast and accurate fault analysis in metro trains.</p>