Recursive Modeling, Online Monitoring And Quality Prediction For Industrial Processes Based On Canonical Variate Analysis

As the scale of industrial processes has gradually become large and the complexity has increased,effective process modeling and monitoring methods are critical to ensure the safe operation of the industrial process,to ensure product quality and to improve the economic benefit.Multivariate data sampled from the complex industrial process often has nonlinear,dynamic and time-varying characteristics.The traditional canonical variate analysis algorithm cannot effectively deal with the data characteristics.In this dissertation,aim at the data characteristics of industrial processes,one of subspace identification algorithms canonical variate analysis was adopted,proposed a series of recursive modeling,online monitoring and quality prediction algorithm for industrial process.Main work and contributions are the following several aspects:(1)Because the traditional canonical variate analysis modeling method cannot adapt to the normal change of industrial process,a state space modeling method without the input based on stable recursive canonical variable analysis was researched.First of all,an exponential weighted moving average algorithm was combined with variable forgetting factor to update the past and the future observation vector covariance and the covariance matrix.In order to timely track changes of the process,a variable forgetting factor adjustment strategy was proposed based on subspace angle between continuous measurement variables.Secondly,in order to guarantee the stability of the recursive model,the combination of constraint weighted recursive least squares method was designed to ensure the stability of the recursive model.Finally,the proposed method is applied to the Tennessee Eastman chemical process simulation.(2)Considering state space model with inputs,as the traditional canonical variate analysis cannot adapt the variation of processes,a state space modeling method based on stable recursive canonical variate analysis was researched for modeling of the time-varying process.Firstly,the exponential weight moving average method was adopted to update the covariance and the covariance matrix and the variable forgetting factor can be adjusted based on residual error between the model output and the actual measurement values.Secondly,the recursive update coefficient matrix and the nonlinear constrained optimization are introduced to ensure the stability of the identification model.The third order state space model open-loop numerical example was used to verify the algorithm whether can get consensus estimation of system matrix eigenvalue.Finally,the proposed method was applied to continuous stirred tank heating system.(3)In view of the conventional canonical variate analysis algorithm cannot effectively monitor industrial process with time varying characteristics,a recursive canonical variate analysis method based on the first-order perturbation theory was researched for monitoring time-varying processes.First of all,based on the offline modeling,the exponential weight moving average method was adopted to update the covariance matrix when collected new measurements.Secondly,the first-order perturbation theory was introduced to realize recursive singular value decomposition of the Hankel matrix,and update the state and residual vectors and their statistics.Thirdly,the on-line fault identification method based on weighted variables contribute was also proposed.Finally,the proposed method was applied to continuous stirred tank reactor system and Tennessee Eastman chemical process.(4)For effectively dealing with the nonlinear and time-varying characteristics of data from industrial processes,an online fault monitoring method was proposed based on efficient recursive kernel canonical variate analysis.Firstly,based on the KCVA model,data in low dimensional space was mapped by kernel function to high dimensional linear feature space when collected new measurements.Secondly,an exponential weighted moving average method was adopted to update the covariance matrix in linear feature space.And we introduced the first-order perturbation theory to realize recursive singular value decomposition of the Hankel matrix to reduce computational load.Finally,the proposed method is used for continuous stirred tank reactor in the system the simulation research.(5)Aiming at quality variables are hard to measure on-line in industrial processes,an adaptive efficient recursive canonical variate analysis was proposed for online quality prediction of multimode continuous process and multi-phase batch process.Firstly,establish an offline model based on part of the normal data and get the initial value of covariance matrix.Secondly,we combined the exponential weight moving average method with varialbel forgetting factor based on the norm of outputs residual for updating the covariance matrix.Finally,system matrices were updated by using recursive least square method,and predicted the quality-related variables using the new identified state space model.It was successfully applied in Tennessee Eastman chemical process of multimode operation conditions and multi-phase penicillin fermentation process,The data from different simulation procedures of industrial processes was adopted to verify the proposed method for recursive modeling,online monitoring and quality prediction method.The simulation results show that the proposed recursive modeling method has higher precision of modeling than traditional canonical variate analysis and can ensure the stability of recursive model;the proposed online monitoring method can not only adapt to the process of natural changes,but also can effectively detect the four types of sensor faults;the quality prediction method not only has the less computational load of the online algorithm compared with the conventional singular value decomposition,but also can get higher prediction precision.