Research On Fault Detection Method Based On Multi-block Modeling Strategy

Guaranteeing stable process operation status and efficient production capacity has always been two important problems that modern industry urgently needs to solve,and process monitoring technology is an effective way to solve the above problems,so it has always been one of the research hotspots in the field of process control.With the vigorous development of sensor technology and distributed control systems,data-driven process monitoring technology has been widely used.Multi-block model,as an integrated model framework proposed in recent years,can effectively monitor industrial processes under distributed and complex working conditions.The thesis carried out research on process monitoring methods under the framework of multi-block modeling.The main research contents are as follows:(1)Aiming at the problem that the traditional Principal Component Analysis(PCA)model is easy to ignore the local information of the process,a sub-block division strategy based on principal component decomposition is proposed.When the process prior knowledge is not available,the process variables are ranked according to their contribution rates on different principal components,and the variables are selected by using the cumulative contribution rate method to form variable combinations of different combinations.These variable subsets contain their respective local information,and it is easier to monitor the local abnormal state of the process by establishing a process monitoring model on this basis.Finally,the Bayesian information criterion is used to fuse the monitoring results of different sub-blocks to obtain a unified BIC monitoring index.The simulation experiment of the TE process shows that the method has improved the overall monitoring performance compared with the PCA model.(2)In order to improve the problem that the sub-block division method in multi-block modeling strategy has ambiguous interpretation and single monitoring information,a process monitoring method based on multiple information extraction is proposed.Different from the previous variable subset division method to extract the local information of the process,this method focuses on further extracting the cumulative error information and the rate of change information from the original observation information to form a separate information sub-block.Cumulative error information is more sensitive to minor and slowly changing faults,and change rate information is easier to extract fault characteristics under abnormal conditions where some variables oscillate violently.Therefore,each sub-block has good monitoring performance for its sensitive fault type.Then the Bayesian method is used to fuse the monitoring results of the sub-blocks,which integrates the advantages of each sub-block model and effectively improves the monitoring accuracy of process monitoring.The contribution graph method is extended to a multi-block modeling strategy,and fault diagnosis is introduced by introducing a weighted form to isolate the relevant variables that caused the fault.(3)Aiming at the problem that the selection of principal elements in the PCA monitoring algorithm is not reasonable,a multi-block PCA process monitoring algorithm based on sensitive principal elements is proposed.The traditional PCA process monitoring method uses principal elements with large variance contribution for monitoring in the principal element space,but principal elements with large variance contribution are not necessarily more favorable for fault detection.Therefore,starting from the performance index of the PCA monitoring algorithm,the fault is divided into several components,and the fault sensitivity coefficient and the fault sensitive principal element are defined for each fault component,and then each sub-block model Sensitive principal components,so as to ensure that each sub-model is most sensitive to its own fault component,and finally uses Bayesian method for fusion to obtain global monitoring results.This method defines the concept of fault-sensitive principal elements without acquiring the fault data set,which effectively improves the performance of process monitoring and improves the generalization ability of the model.