Progressive PCA modeling for enhanced fault diagnosis in a batch process

Abstract

The conventional process monitoring procedure using principal component analysis (PCA) can show which variable is highly related with the fault by looking at the contribution plots for the monitoring statistics, SPE (squared prediction errors) and T2. However, this procedure is not able to determine if the variable is just affected by the fault or the variable is the cause of the fault. In addition, it is not able to show fault propagation through the process variables during the process time. The proposed progressive PCA modeling procedure can identify all variables related to the fault through progressively removing the identified variables and PCA modeling with the remaining variables. It can also provide timing information of when abnormal behaviors are observed for the identified variables by using time series SPE plots with control limits estimated by weighted chi-squared distribution. Based on the timing information, it is able to build a flow chart showing the fault propagation paths. The proposed method is demonstrated on a benchmark fed-batch penicillin process simulator.