Optimal Sensor Network Upgrade for Fault Detection Using Principal Component Analysis

The efficiency of a fault monitoring system critically depends on the structure of the plant instrumentation system. For processes monitored using principal component analysis, the multivariate statistical technique most used for fault diagnosis in industry, an existing strategy aims at selecting the set of instruments that satisfies the detection of a given set of faults at minimum cost. It is based on the minimum fault magnitude concept. Because that procedure discards lower-cost feasible solutions, in this work, a new optimal selection methodology is proposed whose constraints are straightaway defined in terms of the principal component analysiss statistics. To solve the optimization problem, a level traversal search with cutting criteria is enhanced taking into account that the fault observability is a necessary condition for fault detection when statistical monitoring techniques are applied. Furthermore, observability and detection degree concepts are defined and considered as constraints of the optimization problems to devise robust sensor structures, which are able to detect a set of key faults under the presence of failed sensors or outliers. Application results of the new strategy to a case study taken from the literature are provided.