Position Calculation Models By Neural Computing And Online Learning Methods For High-speed Train

With the increasing operational speed of high-speed train (HST), high-precision positioning subsystem has become important and significant for HST. Meanwhile, the positioning accuracy also affects the operating capacity and efficiency of the line. Based on the review and the study of the relevant literatures, we analyze the advantages and disadvantages of current train positioning system. Due to the actual data, we found there is still improved room for Chinese train positioning system and then we propose improved algorithms to reduce the positioning error without additional equipment. The main contribution of this paper is described as follows:Firstly, by analyzing the positioning subsystem of CTCS-3 and packet data message used in bidirectional vehicle-ground communication, we formulate a mathematical positioning model based on HST’s positioning data. We found the accumulated positioning error of current positioning system between adjacent balises can reach 2%, which is computed by the field data.Secondly, as a complementary positioning method for HST, neural computing based on back propagation (BP), radical basis function (RBF), and adaptive network based fuzzy inference system (ANFIS) are introduced to estimate the travelling distance of HST. With the six indicators proposed in the thesis, three algorithms are compared with average speed method (ASM) on the basis of the actual operating data collected from Beijing-Shanghai high-speed railway. The computation results show that BP can improve positioning accuracy about 63.52%、34.16% in training set and testing set respectively. RBF can improve about 40.16%、38.61%, and ANFIS can improve 63.11%、39.11%. We also propose online learning methods (OLM) to further reduce positioning error. The positioning accuracy in testing set has been increased respectively by 38.12%、39.60%、42.58% by using BP-OLM、RBF-OLM、ANFIS-OLM. After comprehensive analysis, the ANFIS model with online learning algorithm has the best performance among all the proposed methods.Finally, we realize our proposed positioning algorithms by developing a software tool on MATLAB platform, which is operation-simple and user-friendly. On the designed platform, HST’s position of ASM、BP、RBF、ANFIS、BP-OLM、RBF-OLM、 ANFIS-OLM can be calculated and compared easily and simply. In terms of program design, the software comprises of three parts, namely data loading module, intelligent positioning algorithms module, result displaying module. From interface design perspective, the software contains training set, training result, testing set, testing result. We also present the detail procedures of the software.