Nonlinear process analysis and modeling: Integrating statistical techniques and neural networks

It has been demonstrated by many industrial applications that neural net-works are promising in solving complex engineering problems. However, like many other new techniques, neural network technology is not a universal tool without any shortcoming. In fact, there are a number of problems that occur in neural network applications and the theory of neural networks does not yet have answers for all of them. Fortunately, it has also been demonstrated that by means of integrating neural networks with statistical techniques some of the problems can be solved. Thus, developing integrated methods which can be applied to model, monitor, and optimize industrial processes is a major goal of the dissertation.; Based on principal curve algorithm and neural networks, a nonlinear principal component analysis (NLPCA) method has been proposed. NLPCA can be applied to the same problems as PCA: data reduction, sensor validation, process monitoring, etc. Because of NLPCA's ability to describe nonlinear data more efficiently than PCA, it can enhance the performance of these tasks. Applying the NLPCA method in continuous process monitoring has been discussed. The results of case studies show that the NLPCA monitoring approach has many advantages over existing approaches. Based on the idea of Multi-way PCA, NLPCA has been successfully used in batch tracking. By projecting the data of a batch process down to a low dimensional space defined by nonlinear principal components, the batch can be easily monitored by tracking its progress in this low dimensional space. The application results show that the NLPCA batch tracking is simple and powerful. Using NLPCA in autoassociative neural networks is also discussed. Applying the autoassociative neural networks in sensor data analysis is presented.; Neural Net Partial Least Squares (NNPLS) method is further extended in the dissertation. A Neural Net Multi-way PLS method is proposed to model batch processes and batch-to-batch optimization using an NNMPLS model scheme is presented. Several advantages have been achieved by proposed batch optimization approach. First, a first-principal model and on-line state measurements are not needed. Second, the method can deal with plant-model mismatch. Third, the computation time in optimization is significantly less than that using a method based on a first-principle model. Finally NLPCA and NNPLS are also used together to solve an industrial soft sensor problem, and excellent results are achieved.