Prediction of mechanical properties of rail pads under in-service conditions through machine learning algorithms. (January 2021)

• Record Type: Journal Article
• Authors: Ferreño, Diego; Sainz-Aja, Jose A.; Carrascal, Isidro A.; Cuartas, Miguel; Pombo, Joao; Casado, Jose A.; Diego, Soraya
• Is Part Of: Advances in engineering software. Volume 151(2021)
• Publication Date: 2021-01
• Subjects: Railway dynamics -- Sleeper pads -- Machine learning -- Rail service conditions -- Dynamic stiffness; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2020.102927
• Languages: English
• ISSNs: 0965-9978

Abstract

Highlights: This study focuses on rail pads fabricated in EPDM, TPE and EVA. Pads' stiffness was predicted through Machine Learning from the working conditions. The importance of each in-service condition was ascertained. The dependency of each working condition on the stiffness was estimated. An application capable of predicting rail pad stiffness was developed. Abstract: Train operations generate high impact and fatigue loads that degrade the rail infrastructure and the vehicle components. Rail pads are installed between the rails and the sleepers in order to damp the transmission of vibrations and noise and to provide flexibility to the track. These components play a crucial role in maximizing the durability of the railway assets and minimizing maintenance costs. Rail pads can be fabricated with different polymeric materials that exhibit non-linear mechanical behaviours, which strongly depend on the service conditions. Therefore, it is extremely difficult to estimate their mechanical properties, in particular the dynamic stiffness. In this work, several machine learning methodologies (multilinear regression, K nearest neighbours, regression tree, random forest, gradient boosting, multi-layer perceptron and support vector machine) were used to determine the dynamic stiffness of rail pads depending on their in-service conditions (temperature, frequency, axle load and toe load). 720 experimental tests, under different realistic operating conditions, were performed to produce a dataset that was then used for the training and testing of the machine learning methods. The optimal algorithm was gradient boosting for EPDM ( R 2 of 0.995 and mean absolute percentage error of 5.08% in the test dataset), TPE (0.994 and 2.32%) and EVA (0.968 and 4.91%) pads. This model was implemented in an application, available for the readers of this journal, developed on the Microsoft .Net platform that allows the dynamic stiffness of the pads study to be estimated as a function of the temperature, frequency, axle load and toe load.