Using a Multibody Dynamic Simulation Code With Neural Network Technology to Predict Railroad Vehicle-Track Interaction Performance in Real Time

Recently, there has been a large demand for predicting, in real time, the performance of multiple railroad vehicle types traversing existing track as the geometry of the track is being measured. To accurately predict a railroad vehicle’s response over a specified track requires the solution of nonlinear equations of motion and extensive calculations based on the suspension characteristics of the vehicle. To realize the real time goal, codes are being implemented that use linear approximations to the fully nonlinear equations of motion to reduce computation time at the expense of accuracy. Alternatively, neural network technology has the ability to learn relationships between a mechanical system input and output, and, once learned, give quick outputs for given input. The training process can be done using measured data or using simulation data. In general, measured data is very expensive to gather due to the instrumentation requirements and is most often not available. In this paper, the use of multibody simulation code as a training tool for a neural network is presented. The example results estimate the vertical and lateral forces at the wheel-to-rail interface as a function of the geometry of the track and the suspension characteristics of the vehicle type by using a multibody code with neural network technique.