Prediction of wheel-rail force and vehicle safety index using genetic algorithm–based backpropagation neural network with physics-based inversion model

Abstract Wheel-rail force is a crucial indicator for vehicle safety and stability in wheel-rail interactions. To predict and display the continuous wheel-rail force and vehicle safety index efficiently and in real time, a multiple-input, multiple-output backpropagation neural network (BPNN) for wheel-rail force and vehicle safety index prediction based on a physical inversion model is developed. The physics-based inversion model calculates wheel-rail forces by using the wheelset inertia force, the primary suspension displacement, and the Nadal derailment criterion. Vehicle safety indices such as wheel derailment coefficient and wheel unloading rate are estimated using the known wheel-rail forces. This physics-based model suggests a nonlinear inversion mapping from the input to the output for constructing the BPNN. Meanwhile, it is a low-cost method to collect training and test samples, and is also used as a training tool for the neural network. A genetic algorithm (GA) is introduced to optimize the initial weight and bias in the BPNN to enhance the network converge speed and prediction performance. The physics-based model is implemented in the field experimental tests conducted on a subway line in China to construct the sampled data. After the BPNN and GA-optimized BPNN (GA-BPNN) are trained, tested, and tuned based on the experimental data, it proves that the BPNN can predict the desired output reliably and that the GA-BPNN performs more accurately than BPNN. The wheel-rail force and vehicle safety index prediction model proposed in this paper can contribute to the development of vehicle intelligent diagnosis and fault warning platforms in the future.