Prediction of rail-wheel contact parameters for a metro coach using machine learning

Abstract

Dynamics of metro coach is complex with frequent acceleration, braking, sharp turns and gradients. Due to sharp turns and twists in track, metro coaches are more prone to the derailment, frictional energy loss and wear. Since contact parameters are difficult to obtain experimentally and are time consuming to determine even computationally through multibody dynamics simulations, in this work, artificial neural network (ANN) and recurrent neural network (RNN) models are used to predict rail-wheel contact parameters. The determined parameters include contact forces, derailment ratio, frictional power, contact location and contact area. These models can be used for reducing derailment tendency, wheel wear and energy consumption. To tackle local extreme values, four meta-heuristic techniques namely Improved Real-coded Genetic Algorithm (IRGA), Success History based Adaptive Differential Evolution (SHADE), Bonobo Optimizer (BO) and Grey Wolf Optimizer (GWO) algorithms are used to optimize the ANN and RNN models. During training, inputs and outputs to the machine learning techniques are generated from a multibody dynamics model built in commercial software Simpack. The multibody dynamics model used is validated using instrumented field trials. Mean absolute percentage error (MAPE) is used for evaluating the performance of the machine learning techniques. Results demonstrate that SHADE is the best optimizer among the four meta-heuristic techniques considered. MAPE in tests, for SHADE-ANN and SHADE-RNN are found to be less than 7%. The developed machine learning models, particularly SHADE-RNN, can be used for onboard monitoring and can help the pilots in running the trains safely and efficiently.