Robust Model Predictive Control And Fault Detection Based On Data Reconstruction

As the scale and complexity of industrial process unceasing enhancement,large volumes of process data are accumulated.Thus,the control of quality and the performance faces the severe challenge in the industrial process.In this situation,the prediction and monitoring gradually become two indispensable means.It is hence important to develop comprehensive tools for process predicting and process monitoring,and then to mitigate the adverse effect on the performance of the industrial process.In this paper,researches are carried out in three aspects,which are robust model predictive control,multiscale nonlinear process monitoring and sparse principal component analysis for process monitoring.Main results are as the following:1 For uncertain polytopic time-delay systems with input constraints and state delay,we propose a delay-dependent memory robust model predictive control(RMPC)algorithm.The memory state feedback control strategy firstly was introduced into RMPC.We present a new sufficient condition of cost monotonicity.The proposed method is based on delay-dependent controllers,which are designed by minimizing the upper bound of the worst-case cost function and by using the sufficient condition.The proposed delay-dependent robust model predictive control law guarantees the stability of the closed-loop systems.The effectiveness of the method is illustrated by a numerical example.2 We demonstrate a novel multi-scale nonlinear process monitoring and fault detection method,called as the scale-sifting multi-scale algorithm(SMA).The SMA includes the scale-sifting benchmark,data decomposition and data reconstruction,dynamic kernel partial least squares.The key innovation feature of SMA is essential scale sifting and scale data reconstruction without prior knowledge of signals monitored compared with state-of-the-art multi-scale monitoring methods.The scale-sifting benchmark is developed to sift out special scales with the essential features of abnormal situations.Then,the data are reconstructed corresponding to selected scales.Finally,dynamic KPLS is applied to analyze data reconstructed for online quality process monitoring and fault detection.The application results illustrate the effectiveness of the proposed method.3 We proposed a new sparse principal component analysis algorithm,called as the compressive sparse principal component algorithm(CSPCA).CSPCA includestwo components: the compressive partial reconstruction of the abnormal signal and the improved sparse principal component analysis.Due to the compressive sensing,the partial reconstruction algorithm was proposed for the abnormal signal without prior information of the sparsity.Using the relationship between principal component analysis(PCA)and the singular value decomposition of data matrices,the convex optimization problem is presented by introducing 2,1L norm into the cost function and the regularization penality for extracting the principal components(PCs)loadings.An iterative algorithm is proposed to solve the convex optimization problem.Meanwhile,CSPCA is capable to takes into account both PCs with large scores and PCs with small scores to promote the sensitivity to the abnormal situations in the process monitoring.Finally,the monotonicity and convergence of CSPCA are demonstrated.Simulation results of Pitprops data and Tennessee-Eastman process illustrate the performance of CSPCA.