Data-driven Robust System Identification Methodology Investigation

With the rapid development of artificial intelligence and big data technology,industries have ushered in a new era.To accelerate the transformation of industrial production from automation to intelligence,data-driven modeling will play a more important role in intelligent analysis and decision optimization of industrial production.At present,data-driven modeling has achieved remarkable results in the application of industrial process control,and has been widely used in petroleum smelting,chemical production,power control and other fields.The complex dynamic processes in industrial processes are often accompanied by nonlinear and multi-operation modes,which makes it difficult for traditional first-principle models to accurately describe realworld systems.Unlike first principle models,data-driven models have strong fitting ability and adaptability,making them suitable for situations where there is a large amount of data and it is difficult to accurately describe the internal structure and working principles of the system.Considering that the actual industrial environment often suffers from different disturbances,such as information transmission delay,sensor failure and noise,it is necessary to conduct robustness analysis on data-driven models and improve their suitability for industrial data.This thesis mainly studies robust identification methods based on data-driven modeling under the probabilistic framework.The main contents are as follows:(1)The observed data are affected by outliers in the complex industrial environment,so a robust noise model is proposed by combining a hidden Markov model and a mixture of Gaussian distributions.The conventional mixture of Gaussian distribution models noise by assigning Gaussian components with different variances,but it does not consider the time evolution of noise distribution.Hidden Markov models can capture the historical transition information of the operation mode of noise distribution,and provide a more robust parameter estimation strategy.Next,the Linear AutoRegressive eXogenous(ARX)model is combined with the proposed robust noise model to complete the modeling of time series data.Because the operation mode of noise distribution cannot be observed,the parameters of the noise distributions were estimated by Expectation Maximization(EM)algorithm.A numerical simulation example,a continuous stirring reaction process and a DC motor model are used to verify the superiority and robustness of the proposed method.(2)Due to missing data,nonlinearity,and outliers of the collected observation data,a robust identification method is proposed by combining the Gaussian Process Regression(GPR)model with the robust noise model.Industrial data usually has a strong nonlinearity,sensor fault,data transmission failure and network communication loss,which lead to data missing problems.GPR models can model the nonlinear process with the help of fewer parameters.However,the traditional GPR model assumes that the noise distribution is Gaussian,which makes the model sensitive to outliers.To mitigate the impact of outliers on the prediction performance,the noise distribution of the GPR model is replaced with the proposed robust noise model.Under the framework of EM algorithm,the problem of parameter estimation with missing data is also solved.Finally,the performance of the proposed method is verified by two functional simulations and a continuous stirring reaction process.(3)Considering that industrial processes have switching and nonlinear characteristics,it is difficult to model dynamic processes globally by a single model,and the collected data also often have problems such as time delays and outliers.Therefore,a Piecewise Affine Linear AutoRegressive eXogenous(PWARX)model is developed with consideration of a t distribution,and the modeling and identification of this process data are solved based on the Variational Bayesian(VB)algorithm.The multi-model structure of PWARX model can effectively deal with complex systems with switching and nonlinearity.At the same time,the t distribution is used to model the noise,and the length of the distribution tail is adjusted by changing the value of the degree of freedom,which effectively improves the robustness of the model.Data classification and parameter estimation are completed according to the identification results of VB algorithm.Finally,the proposed algorithm is verified by a numerical simulation and a continuous fermentation process simulation.(4)Considering the limitations of switched Linear AutoRegressive eXogenous(SARX)model and the diversity of noise forms,the submodel of SARX model is extended to the Gated Recurrent Unit(GRU)model,and the generalized Gaussian distribution is used to model the noise,and a robust identification method for Switched Gated Recurrent Unit(SGRU)model was proposed.Because the submodels of the SARX model are linear,the performance of the model will deteriorate in the presence of complex nonlinear industrial process.The powerful nonlinear fitting ability of GRU model can effectively solve this problem.At the same time,by adjusting the shape parameter of the generalized Gaussian distribution,it can adapt to various noise forms and ensure the robustness of the model.Finally,the identification problem of the proposed model is solved by using EM algorithm,and the effectiveness of the proposed method is verified by a numerical simulation and a continuous fermentation process.Finally,the thesis summarizes the conducted research work,and points out the future of the research work and the key directions of the follow-up research.