Study On Performance Monitoring For Process Industry Based On Data-driven Technique

Statistical process performance monitoring has been an important means of ensuring the consistency of product quality and the safety of production procedure in the process industry. It is based on the method of multivariate statistical theory, to establish the monitoring model by analyzing and interpreting the input and output data collected, and then used for online fault detection and identification to the anomalies, to reduce the losses caused by the fault and improve product quality.The traditional methods of multivariate statistical theory include principal component analysis (PCA), partial least square (PLS), fisher discriminant analysis (FDA), canonical variate analysis (CVA), and independent component analysis (ICA). Multivariate statistical method is an effective tool for high-dimensional data analysis using multi-dimensional projection method to get less feature vectors and eliminate redundant information. It is a good solution for the industry with large number of process variables which have strong coupling relations. This dissertation focuses on the research of process monitoring method based on partial least squares (PLS), and achieves the goal of the detection of key variables in the process industry. Main results and contributions of this dissertation are as follows:(1) The basic theory about partial least squares (PLS) is deeply studied. In this thesis, the PLS method is theoretically derived using non-linear iterative algorithm, and the geometric significance of selecting feature vectors of PLS model is explained, and the performance of PLS model is analyzed as well. The application of PLS model is introduced in the area of process monitoring and control, and the multivariate statistical process monitoring methods is described based on the PLS model.(2) The multi-way PLS method is proposed and applied to the batch-process, in which the measurement data collected is three-dimensional matrix. The three-dimensional data matrix is expanded into two-dimensional data by cutting the batch and variable data block along the time axis, and the monitoring model is established using PLS method based the two-dimensional data. A MMA polymerization batch process is studied, and the simulation results demonstrate the effectiveness of MPLS method.(3) A recursive partial least square (RPLS) method is proposed to solve the problem that the PLS monitoring model can not be updated real-time. According to this method, the model parameters are updated real-time by using the new data matrix together with the old data matrix, so that the monitoring model can effectively track the time variant characteristic of the systems. The MMA polymerization batch process is studied again, and the monitoring model using MPLS method and RPLS method are established. The simulation results demonstrate that the effect of the RPLS method is better than MPLS method which has no model update.(4) The PLS method and RBF networks are combined to deal with the high nonlinearity of the process. The nonlinear function of the process internal model will be linearized as the result of the universal approximation of RBF network, and the recursive method is applied to adjust the linear parameters of the model. This method effectively solves the problem of the real-time update in nonlinear PLS monitoring model. In this thesis, this method is applied to monitor the melt flow rate (MFR) of polypropylene. The simulation results demonstrate the effectiveness of this method.