Using Radial Basis Function and Back Propagation to predicate fault in a railway dangerous goods transportation system considering the Markov Correction

Abstract

Nowadays, we can obtain a huge amount of information and data of railway dangerous goods transportation system through various technical documents, accident reports, monitoring technologies etc. For the space composed of massive historical information and data, a data-driven fault predication approach that can better grasp some important local information while ignoring the negative impact of other data in the global might be practical. However, we have the following challenges: (i) How to reduce the impact of unimportant information on prediction results in advance? (ii) What kind of local approximation data-driven approach should be applied, and (iii) How to correct the results and improve the reliability and accuracy after fault predication based on the local approximation data-driven approach? In this paper, two data-driven approaches including Back Propagation with Markov Correction and Radial Basis Function with Markov Correction are proposed to predicate the fault in a railway dangerous goods transportation system. In order to solve challenge (i), we decompose the whole transportation process into multiple sub-processes based on Work Breakdown Structure with clear duration boundary on the time axis, and use Risk Breakdown Structure to detect the possible faults in each sub-process. In order to solve challenge (ii), we use and compare Back Propagation/Radial Basis Function with global approximation/local approximation and nonlinear prediction ability to learn and predicate the fault based on the collected historical data. In order to solve challenge (iii), we correct the prediction results with fluctuating deviation by using Markov Correction due to its non-aftereffect. Finally, a case study is conducted based on the collected historical fault data of railway dangerous goods transportation system in China. The results show that, for the prediction results based on Back Propagation, the average error increases from 1.43 to 1.88 and the Mean Square Error increases from 3.27 to 5.60 after Markov Correction. For the prediction results based on Radial Basis Function, the average error decreases from 1.13 to 1.11 and the Mean Square Error decreases from 1.81 to 1.76 after Markov Correction. The non-aftereffect of Markov chain well corresponds to the local approximation of Radial Basis Function. When the prediction value is determined by some data of the series, Radial Basis Function with Markov Correction can be applied. When the prediction value is determined by all the data (small-scale) of the series, Back Propagation with Markov Correction is better.