Machine Learning Based Intelligent Operation Methods For Heavy Haul Train

Heavy haul railway transportation,which is characterized by the economy of large scale rail transportation and intensive management,is currently one of the most efficient and low-cost traffic methods that provide the services for mass goods transportation.In order to increase haulage capacity,trains now have increased to a maximum of 20,000 tonnes.Due to characteristics of heavy haul trains and the complexity of practical lines,it is a big challenge for manual driving.Thus,the intelligent train operation strategies will need to be developed and refined for heavier and longer trains accordingly.In daily operations,a large amount of historical field data associated with the real-time states of different types heavy haul trains and manual operations are recorded by the on-board recording systems when running on the same route.Based on statistical analysis of these practical information,data features can be obtained and domain ex-pert knowledge of experienced drivers can be discovered.Moreover,employing machine learning techniques,intelligent operation algorithms can be designed to improve safety,punctuality and longitudinal forces of heavy haul train operation.This thesis investigates the intelligent operation problem of heavy haul train consid-ering the specific line condition and transportation management approach of Shuohuang heavy haul railways.First of all,by exploring the on-line running and line static data,deep learning based data feature learning methods are established to reduce the high-dimensional feature space into a low-dimensional feature space.By this way,redundant information are removed for original data,and better control accuracy and less process-ing burden can be guaranteed.Next,allowing for the characteristics of different driving commands,finding suitable air braking pressure reductions to operation is modeled as a multi-class classification problem,while regression algorithms are applied to determine appropriate continuous handles such as traction or electrodynamic braking.Furthermore,the cooperative control strategy of different locomotives in heavy haul combined trains is detailed.Regarding the whole train as different intelligent agents connected by flexible couplers,the stochastic dynamic programming model is formulated with the minimized in-train forces criteria and approximate dynamic programming algorithm with lookup ta-ble representation is then introduced to find the optimal air braking reduction action when running on steep descent.The main innovations of the thesis are list as follows:1.Based on the analysis of data constitution and feature,this study proposes a data preprocessing and feature learning method.Stacked autoencoders deep networks are es-tablished to automatically learn features from heavy haul train operation data without any priori knowledge.Since the running train will inevitably suffer from the uncertain distur-bance from real-world environment and model inaccuracy,incremental feature learning based on-line learning approach is introduced to enhance the robustness and adaptability.2.EasyEnsemble for multi-class with KNN based Denoising(EMKD)algorithm is introduced to determine the exact timing for exerting and releasing air braking as well the feasible air braking pressure reductions.EMKD method utilizes the EasyEnsemle.M1 algorithm to sample subsets from majority air braking datasets,and takes the advantages of AdaBoost.Ml algorithm to ensemble weak classifiers.In addition,K-nearest neigh-bor based denoising(KD)algorithm is elaborated to remove the possible noise data in minority air braking dataset.3.This thesis formulates the continuous handles problem such as traction or electro-dynamic braking as a least square support vector machine(LSSVM)regression model.Multi-swarm cooperative particle swarm optimizer is then adopt to optimize the struc-tural parameters of LSSVM.Moreover,the LSSVM based intelligent control algorith-m is integrated with EMKD method to control the traction or electrodynamic braking and braking pressure reduction accordingly.Hence,the real-time intelligent operation of heavy haul train is realized.4.In order to find the optimal air braking pressure reduction actions for different locomotives in heavy haul combined trains,the value function approximation with fixed bounded lookup table maintenance(VFA-FBLTM)algorithm is proposed based on ap-proximate dynamic programming(ADP)method.The stochastic dynamic programming model is formulated with the minimized longitudinal forces criteria considering practi-cal operational constraints and uncertain disturbance.VFA algorithm with lookup table representation is then introduced around post-decision states.As the size of lookup ta-ble grows exponentially with states,FBLTM algorithm is therefore put forward to reduce storage space and achieve better efficiency.