KECA Similarity-based Monitoring And Diagnosis Of Fault In The Multi-phase Batch Process

Batch production process is one of important production modes in modern industry.It is widely used in manufacturing of small batch,multi-variety and high value-added products,such as microbial fermentation,genetic engineering pharmaceutical,fine chemical industry,etc.It plays an important role.Therefore,ensuring production safety,low-carbon environmental protection and product quality of batch process has become focus of increasing attention.Batch process generally has characteristics of nonlinearity,nongaussianity,gaussianity and multi-stage,and complexity of production process is much higher than that of continuous process.Quality of product is easily affected by external conditions such as equipment,environment and raw materials.In order to ensure safety of production process,monitoring and fault diagnosis technology for production process has been paid more and more attention by academia and industry.Fermentation process is a typical batch process.In view of multi-stage and nonlinear problems in fermentation process,contents of this paper are as follows:(1)Phase partitioning algorithm based on KECA similarityAiming at multi-stage problem of fermentation process,a phase partition algorithm based on KECA similarity is proposed.After preprocessing simulation data of penicillin fermentation process,it is divided into each time slice matrix.Then according to the similarity index of KECA,the sub-period is divided into stable sub-period and transition sub-period.The time slice in each sub-period will have the same process characteristics,and the data in the same time period can be set up a unified model.(2)Fault Monitoring of fermentation process based on KECAKECA and sliding weighted KECA monitoring models were established for each stable stage and transition stage respectively.Based on angle structure of principal component selected by KECA,a divergence measure statistic(CS)is introduced to express angle structure,which is used as statistical quantity of monitoring model.Control limit of CS statistics is calculated by kernel density estimation,and fault is detected if control limit is higher than control limit.(3)Fault diagnosis algorithm based on SV-KCDBecause it is impossible to find an inverse mapping from high-dimensional feature space to low-dimensional input space and to deduce contribution expression of corresponding statistics,traditional contribution plot can not be applied to kernel space plotting method.In order to solve above problems,this paper presents a SV-KCD method based on standard vectors.This method not only retains traditional contribution graph calculation simple,intuitive,no fault samples,but also does not need to calculate derived contribution expression.Theoretically,this method can be applied to any kernel plotting method.(4)Application of Industrial process dataFinally,research method is applied to data test of Escherichia coli.Application of this method to production process of E.coli shows that method can better reflect diversity of characteristics of each stage,and can effectively extract data information of production process,detect occurrence of fault rapidly and judge fault source accurately,which has a certain practical valuefor solving multi-stage intermittent process of fault monitoring and diagnosis problems.