Switched System Identification Based On Neural Network

In recent years,as innovations in the engineering field continue to grow,modern industrial production processes have become increasingly complex.The hybrid system formed by multiple fields has become more complicated and difficult,with new challenges as well as new opportunities having emerged.As a representative dynamic hybrid system,the switching system has become one of the hottest research areas.The switching system is a system composed of a series of switching rules referring to the relation between a limited number of subsystems and control subsystems.The main responsibility of the switching rules is to implement dynamic switching between control subsystems under certain rules.The dynamic change of the switching system includes a continuous dynamic process and a discrete dynamic process.This combination and switching dynamic characteristics can fit a multi-modal hybrid system.This advantage relies on the accurate mathematical model of the switching system,so the research on the identification and modeling of the switching system is of great theoretical and engineering practical significance.The structure of the switching system is simple and flexible.The sampling data of the system is usually mixed and irregular,making it difficult to find rules from mixed data and build appropriate system models in identification.In this paper,the identification method of switching system is studied.The identification of the switching system is divided into two processes: switching rules identification and system parameter identification.The neural network is introduced to classify and identify the system switching rules of the mixed data,and combined with the discount recursive identification algorithm to complete system parameter identification.The research work of this paper mainly contains the following aspects:(1)For identification of the switching system with prior data,an online identification method of the switching system based on BP neural network is proposed.In identifying the switching rules,first,introduce the priori data of the labeling function to standardize the system,then use the labeled switching rules and system input and output data to construct a data set to train the BP neural network,so as to establish the system's switching rules prediction model.With the input and output data of the online sampling switching system,the switching rules of the system can be predicted;in identification of system parameters,according to the predicted switching rules,a recursive identification algorithm based on discount is further proposed to identify the subsystem ' s parameters online.For parameter identification of switched nonlinear systems,the key term separation method is used to separate the parameter coupling items in the system to obtain the estimated values of the parameters of each subsystem.Finally,the autoregressive output error switching linear system and the Hammerstein-Wiener switching nonlinear system are used for simulation experiments.Comparisons with other methods are made to further verify the effectiveness of the proposed method.(2)For identification of switched linear systems with unknown number of subsystems,a switching linear system identification method based on competitive neural network is proposed.In identification of switching rules,from the perspective of solving underdetermined equations,the Kaczmarz algorithm is used to pre-process the sampled data of the system,so that the collected mixed data can be projected to different hyperplanes.Then the pre-processed data is divided by a competitive neural network to obtain the number of subsystems and the system switching rules;In identification of the system parameters,according to the obtained switching rules,the discount recursive identification algorithm is further used to estimate the parameter value of each subsystem to complete the system parameter identification.Finally,through simulation experiments of the output error switching linear system,the validity of the switching linear system identification method based on the competitive neural network is further verified.(3)For identification of switched nonlinear systems with unknown system order in subsystems,an identification method of switching nonlinear systems based on the improved competitive neural network is proposed.With the number of the subsystems unknown,there are data mixing and parameter coupling problems in switching nonlinear systems,which poses challenges for Kaczmarz algorithm,the convergence speed and accuracy of competitive neural networks.In this paper,the selection strategy is changed to accelerate the convergence of the Kaczmarz method.Meanwhile,according to the characteristics of data aggregation,the data density and Mahalanobis distance are introduced to improve the competitive neural network to further increase the accuracy and efficiency of rules identification.Finally,a simulation experiment is carried out through Hammerstein switching nonlinear system,and the effectiveness of nonlinear system identification method based on the improved competitive neural network switching is further verified.(4)For the temperature control of the semi-batch reactor for esterified olive oil,a switching system model is established.The temperature has a relatively great effect on the quality of esters in the semi-batch reaction of esterified olive oil,so it is necessary to establish a dynamic model to provide a basis for the control of this system.This paper uses competitive neural network and discount recursive identification algorithm to conduct theoretical simulation experiments on the model,and further verifies the effectiveness of switching system identification modeling based on neural network and discount recursive identification algorithm.After proposing the above algorithms and confirming them by corresponding simulations,this thesis summarizes all of the work.In the end,it sorts out the directions worth studying and discussing in the future,and makes some prospects.