| Operating and running safety is a very important aspect in current high speed, heavy load, large capacity and high-density railway transportation. Wheel-rail force is a crucial factor which leads to rail failure/damage, train derailment, and vehicle parts damage. Moreover, it is also used to assess the stationary and security of the train, and used as the main basis for speed-limiting and speed-increasing. For such an importance, it has great theoretical and practical significances to predict the wheel-rail force.Artificial Neural Network has been widely used in pattern recognition, complex system modeling and control for its robustness, failure tolerance capability and self-learning ability. In this paper, to take predicting the wheel-rail force as a goal, four predicting method based on ANNs are discussed. At the same time, an improving algorithm to train NARX neural network is put forward.The paper first introduces the knowledge about the rail-wheel interaction system, including the definition of the track irregularity and rail-wheel force and its affect on the train safety operation and then expounds the basic concept, development history, research context and application of artificial neural network. Secondly, the paper discusses the merit and demerit of the static and dynamic neural network, and choosing BP, FFTD, NARX neural network to simulate the models of rail-wheel forces and determine the number of hidden nodes and time delay order with experience. The simulation result shows that dynamic neural network can fit the rail-wheel force predict better. Finally based on the traditional NARX, the paper proposes a training arithmetic for large NARX neural network. Comparing and analyzing with the four kinds of network, it proves that the advances NARX neural network has a higher predict accuracy, verifies the effective and feasible of the neural network used for the rail-wheel force and testifies the superiority and effectiveness of the advanced NARX network. |
PDF Full Text Request