The Fault Detection Of Batch Process Based On Process Data

As an important production mode,batch processes ha s been widely used in fine chemical,biopharmaceutical industry etc.In recent years,the market demand for more products of batch process,which makes batch processes moving in the direction of the large-scale and integrated.However,the faults in large-scale production process often cause chain reaction,which will affect the quality of product and even endanger the safety of life.Process monitoring is effective measure to guarantee the safe and stable operation of production process,so the monitoring of batch process is of great significance.The complexity of batch processes hinders the development of process monitoring method based on analytical model and knowledge.The improvement of the computer and simplicity of collecting data provide a chance to process monitoring method based on data-driven.At present,data-driven control method has been widely used in the batch process,but which always set some idealized assumptions,such as Gaussian distribution,Uneven-length data,these assumptions limit the monitoring performance.According to the characteristics of batch process,this paper puts forward two kinds of process monitoring method based on principal component analysis(PCA),and their effectiveness was verified in the process of penicillin(a typical batch process).The process monitoring method based on traditional single model has been used in batch process,which takes all the data of an intermittent operation as a sample.Although this method can monitor the status of batch process,which ignores the local features.Aiming at the shortcomings of the traditional methods,in combination with multi-phase characteristic,time-varying characteristic and Uneven-length data characteristic,a new principal component analysis is proposed,which is based on just-in-time learning(JITL-PCA).Considering the local characteristic and multi-phase characteristic,this method takes the way of online modeling and has been applied to penicillin process.Considering the lack of JITL-PCA method in real time and the complexity of multiple modeling,a new data processing method for multi-distribution is proposed,which based on local neighborhood.The method Standardize process data by local information and make the process data obey Gaussian distribution,which makes process data satisfy the requirements of PCA.This paper applies the method into penicillin simulation and verified the method is effective by comparing this method with JITL-PCA.