Dual Learning Architecture Based Accurate Stopping Optimization Control Of Trains In Urban Rail Transit

With the rapid development of urban rail transit in China,the realization of the accurate train stopping in a complex environment is facing the challenge of changing conditions of lines and trains.Complicated factors and variable conditions will further increase the difficulty of fine-tuning the parameters of the accurate stopping controller.Learningbased algorithms,such as Iterative Learning Control(ILC),are feasible ways to solve the above problems,which have the potential to improve the manual parameter tuning process of traditional control methods,e.g.,PID control.However,the convergence of traditional ILC algorithms depends on the Identical Initial Condition(IIC),which limits the application of ILC.The existing ILC algorithms of the accurate train stopping problem are usually based on the IIC hypothesis,otherwise,they can only guarantee the bounded convergence of error,which can not meet the requirements of the stopping accuracy.In this thesis,the initial state distribution of the stopping process is analyzed by using the actual train operation data.The controllers are designed based on the ILC algorithms which aim at solving the initial state shift problem during the train stopping process.Specifically,the main contributions of this thesis are as follows:(i)The initial state shift problem of the train stopping process is introduced by extracting the initial state distribution of a train at the same section from the actual train operation data.Then,based on the traditional P-type ILC control law,the simulation scenarios were designed accordingly to analyze and discuss the influence of the initial state shift type on the convergence of ILC,which lays the foundation for the design of ILC controllers.(ii)Based on the Adaptive Iterative Learning Control(AILC)and the error trajectory tracking approaches,an ILC Control law is designed to solve the ILC initial state shift problem during the train stopping process.The applicability of the proposed algorithm is analyzed.In addition,considering the difference between the accurate stopping control problem and the traditional control problem,the form of the error decay function is designed when the stopping accuracy is satisfied.The simulation process shows that the tracking effect of the algorithm depends on the decay function with intangible type,and the feasibility of the algorithm is low.(iii)Considering the limitations of the existing relaxed initial condition method and the advantages of the optimization methods to handle the ILC initial value constraints,the norm optimal method is introduced into the ILC stopping controller design.The stopping control problem is transformed into a constrained speed tracking problem.Combined with the principle of train position correction during the stopping process,a point-to–point norm optimal iterative learning stopping control method is proposed.Finally,a simulation example is designed to verify the effectiveness of the method.(iv)Since the design of the decay function and the realization of the norm optimal method are highly theoretical and hard to implement,a dual learning architecture of "machine learning-based clustering + ILC-based stopping control" is proposed based on the analysis of ILC convergence under bounded arbitrary initial states.Combining the methods of Canopy clustering and terminal ILC,a Partitioned Terminal Iterative Learning Control(PTILC)algorithm is proposed.Rigorous convergence analysis and simulation experiments verify the effectiveness of the proposed method.Through the above research,this thesis analyzes and identifies the key factors that prevent the existing ILC methods from being used for accurate stopping control problems,and proposes three schemes to solve the IIC issue for train stopping control.Among them,the proposed PTILC algorithm meets the general requirements for accurate stopping of urban rail parking,i.e.,±0.3m,and shows a 50% improvement in the stopping accuracy.Moreover,the algorithm is relatively simple to implement,enhancing its feasibility in engineering,and provides theoretical support for the practical application of learningbased stopping control methods.39 figures,6 tables and 92 references.