Researches On Statistical Performance Monitoring And Control Of Batch Processes Based On Bdkpca Method

The object and motivation of batch process monitoring is to timely and effectively find and restore fault in process, insure the safety of production procedure and consistency of product quality and reduce the productivity loss. Compared with continuous process, batch processes are fairly more complex with more rich data statistical characteristics, such as multiplicity of operation stage, nonlinearities, dynamics, batch and so on. So it is more difficulty to performer process monitoring and controlling on batch processes. Because batch SPMC (statistical performance monitoring and control) dose not need the accurate models of processes and the theory and methods gained from research work can be quickly used, it has become one of the most active research areas in process industry.Based on the further research on data statistical characteristics of batch processes, this dissertation focuses on the nonlinearities and dynamics of batch processes. BDKPCA and FVS-BDKPCA methods have been developed and applied to the on-line statistical performance monitoring of batch processes.Firstly, a survey of the aim, the need and the main methods of the process monitoring is presented. The history and status of SPMC are also introduced, especially the status and challenge of the batch processes SPMC are completely introduced. The basic theory and methods of statistical performance monitoring are introduced and a basic framework for statistical performance monitoring of batch processes is given.For the nonlinearities and dynamics of batch processes, the kernel technique (KPCA) is introduced into the BDPCA method and a BDKPCA-based statistical performance monitoring and controlling algorithm is presented. The new method constructs an auto-regressive model structure for each batch of the three-way training data collected form a batch process and maps the auto-regressive model structures into a feature space with kernel function. In the feature space, it employs the Average Kernel Matrix (AKM) of all batches to define an average or compromise structure across the initial structures, which reflecting the nature of the batch runs. Besides, the BDKPCA model only collects the previous data during the batch run without expensive computations to anticipate the future measurements, enhancing the method's accuracy. Applications on a simple multivariate process and the simulation benchmark of the fed-batch penicillin production demonstrate that the proposed approach can effectively capture dynamics and nonlinearities of batch process and, when used for process monitoring, it shown better performance than MKPCA.However, BDKPCA has a characteristic that when the sample number becomes large, the calculation of eigenvalues and eigenvectors of AKM of all batches becomes time consuming and more storage space will be needed to store historical data on-line that limit its practical applications in process monitoring. So an improved feature vector selection (FVS) scheme, which is improved from FVS according to batch processes data characteristics, is adopted to reduce the computation complexity of BDKPCA and BDKPCA based on FVS is developed. For the three-way training data with a large sample number, the improved FVS is performed to reduce the sample size first, and then BDKPCA is adopted for batch process monitoring. Applications on the simulation benchmark of the fed-batch penicillin production show that FVS has no adverse effect on fault detection rate and reduces the storage cost and computational cost in BDKPCA based on FVS.Finally, there are concluded with a summary and some further research areas in this thesis.