Comparative Study Based on a Reduced Dynamic Data-Driven Methods of Machine Learning

Chemical and industrial processes are often overly complex and have thousands of process measurements, every day. In this context, data-driven methods are very interesting methods for the diagnosis of dynamic systems. More precisely, kernel methods have attracted attention for several years, for Fault Detection (FD). A performance comparison, in this paper, is presented between two dynamic monitoring methods. However, the two FD methods are based on the Reduced Kernel Partial Least Square (RKPLS). We compare in this case, the dynamic Moving Window RKPLS (MW-RKPLS) with the dunamique Online RKPLS (ORKPLS). In this comparison, we are mainly interested of the False Alarm Rate (FDR) and also the Good Detection Rate (GDR) to increase the FD performance. All this performances are presented through the Tennessee Eastman process.