Condition monitoring of railway axles by convolutional and temporal neural network processing of vibration signals

Abstract

The railway industry is currently putting a lot of effort into the digitalisation of maintenance and its orientation towards condition monitoring. In the specific case of railway axles, as in the case of other mechanical elements, the main problem to achieve this goal is the lack of predictive models to generalise what will happen under fault conditions in any operating condition. In this work, we propose the use of models based on deep learning, using neural networks for the diagnosis of cracks in railway axles. The features used are vibration signals from four different axle-wheel assemblies, installed on a bogie and driven on a test bench. The signals are obtained at different operating conditions of mounting, load and speed. The neural networks used combine a convolutional neural network (CNN) with a recurrent neural network (LTSM) and finally a multilayer perceptron (MLP) that outputs the axle condition classification. The results are shown using ROC curves, in which a high reliability of the models is observed, which, having been trained only with data from one of the assemblies, are able to generalise and correctly classify the condition of the rest of the assemblies, independently of the mounting and operating conditions.