Soft Sensor Modeling Method Based On The Data And Applications

A large number of process parameters, in industrial production and manufacturingprocesses, are often associated with production efficiency and product quality closely.Therefore, they need to be controlled strictly. However, for some technological, techni-cal or economic reasons, these parameters are not able to be measured directly. In orderto solve the problem, soft-sensing technique has emerged and has been developed asone of hot spots for studying on the current field of process control and process detec-tion. Generally speaking, there are three approaches in the soft-sensing modeling. Oneis based on reaction-mechanism modeling method; Another is based on data-driven sta-tistical modeling method; Still another is hybrid modeling combined with the first twoapproaches. The main research work in this paper includes the following aspects:1. Outlines the background of the current soft-sensing technique, and introducesits concept, basic model and commonly used modeling approaches. At last, the paperfocuses on the history and status quo of the partial least squares (PLS).2. Introduces the method of the PLS modeling and its basic characteristics, andthen analyzes the geometric meaning and characteristics of PLS shown that the fittingresult of PLS is better than the one of principal component regression (PCR) with thesame number of feature variables.3. Introduces a new type of nonlinear partial least squares (NLPLS) method,which is more effective to deal with complicated processes with nonlinearity anddata collinearity. The proposed method is applied to the actual plant to estimate thepolypropylene MFR. It is shown that the accuracy and generalization ability of the softsensor using the method proposed are superior to quadratic polynomial partial leastsquares (QQPLS) method in the NLPLS-II model. The results also show that satisfyingperformance is achieved and the accuracy of the model can satisfy the practical demand.4. The fuzzy c-means (FCM) clustering algorithm is highly sensitive to its initialvalue and easy to fall into local optimal solution. To tackle this problem, the subtrac-tive FCM clustering algorithm is introduced for fuzzy structure identification and thennonlinear PLS combined with the Gaussian Kernel Function is proposed for parameteridentification. Thus, a modeling method of fuzzy PLS (FPLS) appears. The proposedmethod is applied to the actual plant to estimate the PH value. It is shown that theaccuracy and efficiency of the soft sensor using the method proposed are superior toRBFPLS and QQPLS method in the NLPLS-II model. 5. Briefly introduces the mechanism of coal gasifier and the NLPLS modelingmethod is applied to establish the soft-sensing model of coal gasifier's synthesis gascompositions. The results show that satisfying performance is achieved and the accu-racy of the model can satisfy the practical demand.6. The thesis summarizes the main conclusions and discusses some issues forfurther research and exploration on the soft-sensing modeling technique.