Imbalanced Data Driven Intelligent Train Control Algorithms For Heavy Haul Railway

With big load trains,long train marshalling,and complex line conditions,operating mode in heavy haul train systems changes frequently during train driving.Due to this fact,an improper traction or braking operation in the train control system increases the longitudinal impact force of trains and even causes the risk of decoupling.On the Growing downhill,the driver needs to apply air braking to control the speed of train by cyclic braking and the insufficient recharge time may cause the train to lose braking force,which lays hidden dangers for the safe operation.Under the condition that the marshalling mode,running route and load of heavy haul trains are relatively fixed,it is urgent to optimize the train control mode instead of driver control to realize automatic driving.Taking Shuohuang Railway as the research background,based on the analysis of the train operation data of the SS4G locomotive,it is found that the proportion of different operating mode is seriously imbalanced.For the imbalanced characteristic,the classification method of machine learning field is introduced,and the imbalanced data-driven intelligent control model of air braking and traction/electric braking of heavy-haul trains is designed to realize intelligent train driving.The main work includes the following:(1)Reducing the dimension of heavy haul train operation data based on Random Forest algorithm.Based on the standardized data,a random forest model is built to learn the features of heavy-haul train operation data,quantify the importance of different features in intelligent control,and extract the alternative feature set by sequential backward elimination to achieve feature dimensionality reduction.(2)Realizing intelligent control of heavy-haul train air braking based on Adaboost algorithm.The Adaboost-based classifier type is determined by comparing the prediction effects of the two decision tree algorithms,C4.5 and CART,on the dataset.In view of the imbalanced characteristics caused by the serious insufficiency of air braking data,the Adaboost algorithm is optimized from two aspects:the extraction mode of the training sample subset and the voting weight,which increases the F1-Measure value of air braking prediction by 0.0439.(3)Realizing intelligent control of heavy-haul train traction and electric braking based on Support Vector Machine(SVM)algorithm.We optimize the classification of imbalanced data by assigning different penalty factors,C+ and C_,to the majority and minority classes.By introducing kernel functions,the data set is mapped to a high dimension to make it linearly separable.Based on comparison of the performance differences between the Polynomial and the RBF kernel function in different scenarios,we generate a dynamic update factor combined with the speed of train to connect the two kernel functions.The hybrid kernel function optimizes the algorithm and improves the recognition of the data.(4)Building a dynamic model of heavy-haul train to realize the intelligent control model verification.A dynamic model of heavy-haul train is built by combing the output characteristics of heavy-haul train control strategy and train operation data,and the operation scenario is simulated based on the data of line with 408km from the Shenchi South Railway Station to the Suning North Railway Station.To prove the accuracy of-the proposed model,the intelligent control system and the driver's driving results are compared,based on that the safety of the model control is verified from the aspects of speed and airbrake refiling time,the punctuality of the model is verified from the driving time,and the rationality of the model output is verified from the number of working condition switching.