Detection Method Of Abnormal Measured Data Based On Hybrid Model For Multimode Batch Process

With the ability to produce products flexibly,batch processes have been widely spread to the fields of semiconductors,food processing,and pharmaceuticals.They could adjust the raw material ratios and production sequence according to the targets,and obtain high quality products.In order to meet the needs in both process safety and product quality,monitoring the performance with the process model that could describe the trend of the process was the most effective and the most widely used techniques.Traditional process model was training with the mode entire measured data,which may lower the accuracy of model and the further abnormal detection results due to the inclusion of the non-similar training data.Therefore,it is of great significance to study on the detection method of abnormal measured data based on the hybrid model for batch process with especially regard to the process multimode characteristic.Based on the study of batch process characteristics and data characteristics,a hybrid model of multimode batch process was constructed and a detection method of abnormal measured data based on hybrid model and support vector data description(SVDD)was proposed.The mechanism of batch process was first analyzed,and the process mechanism model was established after using the particle swarm optimization algorithm to estimate the unknown parameters.Then,a just-in-time learning(JIT)framework was introduced into the relevance vector machine(RVM),and a fusion similarity factor was also defined to evaluate the data similarity in both mode attribution and algebraic distance.After constructing the just-in-time learning and relevance vector machine(JIT-RVM)identification model with the just-in-time training dataset that has the most maximizing data fusion similarity,a mechanism analysis and JIT-RVM based hybrid modeling method for multimode batch process was given,meanwhile,a hybrid model consisting of a mechanism model and an identification model was constructed to make the real-time prediction of the measured data.Finally,the training dataset was constructed based on the difference between the model prediction and the measured data,and the hyper sphere model was built to detect the abnormal measured data of the multimode batch process.The experimental results show that,comparing with the mechanism analysis and RVM hybrid modeling method,the proposed mechanism analysis and JIT-RVM based hybrid modeling method could effectively capture the dynamic characteristics of the batch process,with higher prediction accuracy;and comparing with the traditional SVDD based detection method of abnormal measured data,the proposed detection method of abnormal measured data based on hybrid model and SVDD for multimode batch process was able to capture tiny abnormal data,with higher detection accuracy.