Fault detection and identification based on dissimilarity of process data

摘要

Multivariate statistical process control (MSPC) has been widely used for process monitoring. When a fault is detected, it is important to identify an actual cause of the fault. Fault identification methods are classified into two groups by availability of historical data sets obtained from faulty situations. When such historical data sets are not available, contributions from process variables to a monitored index can be used for identifying the variables that contribute significantly to an out-of-control value of the index. On the other hand, when historical data sets are available, a fault can be identified by comparing a data set representing the current faulty situation and historical data sets representing past faulty situations. In recent years, a new MSPC method termed "DISSIM," which is based on the dissimilarity of process data, has been developed. In the present work, DISSIM is extended for fault identification with or without historical data sets. The fault detection and identification performance of DISSIM is compared with that of the conventional MSPC using principal component analysis by applying those methods to monitoring problems of a continuous-stirred-tank-reactor (CSTR) process. The simulated results show that DISSIM as well as cMSPC functions well for fault detection and that DISSIM works better than cMSPC for fault identification.