Application of neural networks for the prediction of railway bearing failures

Bearing overheating and anomalous accelerations are two principal failure modes for this safety component. The supervision of bearing's behaviour is essential to ensure a safe and reliable operation. A safety component's failure may cause a speed limitation or even a non-available train for operating, so a predictive maintenance for bearings and other critical components is mandatory for the manufacturers, operators and maintainers in the railway sector. Bearing temperature, exterior temperature, train speed and other variables are measured every second in real time. From all the data collected and stored in the last years some algorithms and models are designed and trained in this paper to detect bearing anomalies 2 days before a real failure is detected and the safety alarm is enabled. The methodology for obtaining the optimal algorithm is exposed. Different artificial neural networks based on different optimization models such as the Mini-batch Gradient Descent (MGD) or Adam optimizer are compared. A final neural network with 10 hidden layers to detect bearing failure is proposed reaching 99% of accuracy, 95% of precision and 90% of sensitivity. The objective of predicting a bearing anomaly with some days in advanced is reached with high precision level, which will lead also to cost savings and a contribution for the sustainability because many inspections could be reduced and the energy cost associated to them.