Research Of Multi-phase Batch Process Monitoring Based On Support Vector Data Description

Batch and semi-batch processes have played an important role in modern industrial processes due to the flexibility in the production and high value-added products.Monitoring of batch processes is extremely important to ensure high quality products and safe operation in process control field.The conventional methods such as MPCA and MPLS are widely used for the batch process monitoring.However,the variables are usually assumed to follow a multivariate Gaussian distribution which may constrain the application of above methods.SVDD do not share the above assumption to process variables,but it can't adapt to the different characteristics caused by multi-phase batch process.Therefore,the research of multi-phase batch processes monitoring method based on SVDD is of great theoretical and practical value.Based on the analysis of characteristics of multi-phase batch process data and the characteristics of SVDD monitoring method,a sub-stage separation and monitoring method based on SVDD is introduced in order to accurately describe the data at different phase.The phases of batch process are divided by the change of structure of hypersphere,monitoring sub-models was established respectively which can improve the precision of monitoring;In addition,the static control limit can't express the dynamic characteristics of batch process.To this end,a kernel-similarity based SVDD is developed for monitoring of multi-phase batch process,the kernel function values between support vectors and test sample are chosen to be a weight,which finally utilized to determine local dynamic control limit.The experimental results show the sub-stage separation and monitoring based on SVDD have an improved monitoring result compared with MKPCA and K-means PC A;The proposed kernel similarity based monitoring method focuses on the irregularity of hypersphere,the local distribution characteristics of process dataset in the high dimensional space,and the time-varying of test data samples,so it can realize the batch process monitoring with more reasonable and precise control limit simultaneously.