Safety Risk Prediction Based On Deep Learning For Train-to-Train Communication Train Control System

Although there has been substantial research in system analytics for risk assessment in traditional methods which using static analysis in design phase,little work has been done for safety risk prediction in communication-based train control(CBTC)system,especially intelligently predicting risk caused by the uncertainty in the system operation.Safety risk analysis is a technique to measure system safety using qualitative or quantitative methods.Having safety risk analysis for CBTC system and improving its safety,it is not only the goal pursued by international technical standards for a long time,but also one of the hotspots in this research field.Meanwhile,redundancy in the communication mode of the CBTC system increases the uncertainty risk of system,which prompts the research on next generation train control system.This dissertation proposes an intelligent safety risk prediction method based on a deep learning,from constructing a next generation train control system based on train to train communication.The method establishes safety risk prediction features and designs a predictive model based on recurrent neural network to deeply explores the potential link between system operation features and risk status.The main contributions of the dissertation include:(1)We put forward a next generation CBTC system paradigm based on train-totrain communication,and optimize the moving block principle.The paradigm adds train-to-train communication mode,re-allocate system critical functions,and optimize the moving block principle in the new system.It moves some functions in track-side equipment to vehicle-onboard system,for reducing the track-side equipment.It overcomes the problem that moving block principle has high dependence on the track-side system,improves the utilization rate of the track,and effectively reduces the system operation risk.(2)We design a risk prediction features selection method for train control system.According to the recommended risk analysis considerations in IEEE1474.1 standard of CBTC system,the method selects some risk features preliminary.For the rareprobability features in this field,we propose a solve algorithm using statistical model checking.Based on this,we perform features extraction using principal components analysis.This method provides an effective method for risk features establishment in train control system.(3)We propose an intelligent safety-risk prediction model for train-to-train communication next generation CBTC system.The model is implemented by a deep recurrent neural network(RNN)called a long-short-term memory(LSTM)network,which takes into account safety risk prediction features.It predicts the occurrence probability of a hazard from datasets and classify them,through finding the mapping between system safety-risk features and train operation status.To reduce the influence of unbalanced datasets in training,the prediction model is optimized based on L2 regularization,which effectively improves the prediction accuracy and avoids over-fitting of model training.This model realizes the learning from system uncertainty information and the relevance caption among risk status,and achieves the purpose of predictive classification and identification of train operation risk status.The model proposed in this dissertation has been successfully applied on a real sample dataset from cooperative corporation.Much experimental results verified the effectiveness and practicability of the model in the safety risk prediction of the next generation CBTC system based on train-to-train communication.