Batch Process Fault Detection Based On Multiway Kernel Entropy Component Analysis

In modern industrial production,batch process is an important manufacturing production mode and has been widely used in biopharmaceutical,fine chemical industry,food,polymer reaction and many other fields.In order to ensure the safety of production,it is of great significance to monitor the batch process in real time and guarantee the human life and property safety.Aiming at some valuable issues such as strong noise pollution,multi-stage,time-varying characteristics and so on,a fault detection method based on multiway kernel entropy component analysis(MKECA)is studied in this paper.The main research contents are as follows:To handle strong noise pollution and the batch change of working condition,an improved multiway kernel entropy component analysis(IMKECA)method is proposed.Firstly,the wavelet transform is used to reduce the strong noise influence upon process variables data.Furthermore,in order to fit the change of working condition in batch operation,the k-Nearest Neighbor(kNN)based Mahalanobis distance,called as M statistic,is constructed to replace the traditionalT~2 statistic.Lastly,simulations on the penicillin fermentation process show that the IMKECA method can detect the fault more quickly than the MKECA method.In order to solve the problem of multi-stage characteristics,this paper proposes a multi-stage fault detection method,referred to as multi-stage multiway kernel entropy component analysis(MsMKECA).This method firstly constructs a matrix similarity based stage division according to the relevancy of time-series kernel entropy principal component,thereby the batch process can be divided into several stages.Furthermore,a batch-variable unfolding is introduced in each sub-stage for building the MKECA monitoring model.Finally,simulations on the penicillin fermentation process show that the MsMKECA method is more sensitive to detecting the fault in batch process.Aiming at the time-varying characteristic of the batch process,this paper introduces Just-in-time learning(JITL)strategy into KECA method to build the JITL-KECA monitoring approach.This method applies the sliding window technique to improve the traditional JITL method,then a number of similar history samples are selected as similarity datasets for the current batch process monitoring samples.Furthermore,the KECA models are built in these similarity datasets for process monitoring.The validity of this method is verified by the simulation of penicillin fermentation process.