Researches On The Method Of Process Modeling And Monitoring Based On NMF-SVM

With the wide application of computer technology in industry process control, large volumes of data are sampled and collected. However, these data which contain information about process conditions are not well exploited, so that there exists "rich data, poor information" in process industry. To mine the key information hid in process data and realize effective control of producting process and product quality, dealing with data by multivariate statistical analysis is needed. Extracting the low-dimension essential component and eliminating the redundant information, error, noise also can be realized. This thesis researches in the background of Multivariate Statistical Process Control, to mine deep-level information of these collected data about process operation and make full use of the information for process modeling and monitoring.In process modeling, a new multivariate statistical project technique—Non-negative Matrix Factorization (NMF) and Least Square Support Vector Machine (LSSVM) which have been proposed in recent years are used for regression modeling. First, a Non-negative Component Regression model is developed based on the features which are extracted from process information by Non-negative Matrix Factorization. Moreover, the non-negative components which are extracted by NMF are used as the input of LSSVM. Based on this, a new model—NMF-LSSVM regression model is developed. Both of two models showed good performation when they are used in modeling of discharged slab temperature.In process monitoring, an integrated framework for system performance monitoring is presented, which combines Non-negative Matrix Factorization (NMF) with Support Vector Machine (SVM). It includes two aspects:①Extracting the features through NMF, it reduces the dimensions of the monitoring system and obtains the main feature statistical variables. Then the control limit is made by kernel density estimating method, thus the on-line monitoring model is set up;②Training the multi-fault classifiers through SVM, when the fault is detected by NMF on-line monitoring model, the fault classifier can be used for fault recognising and diagnosing and confirmed which kind of faults it is. Through the research and simulation of three-tank system, it is proved that this system framework makes a good performance in process monitoring and acquires much more ideal results in fault diagnosing.As a new multivariate statistical project technique, Non-negative Matrix Factorization has been applied in pattern recognition and other areas. NMF is applied in process modeling and monitoring first time in this thesis, and combining with Support Vector Machine, the method of process modeling and monitoring based on NMF-SVM is proposed. In simulation experiment, it showed good performance, and it also supplied a good attempt for the research of multivariate statistical process control.