Sensor Location for Enhancing Fault Diagnosis

Multivariate statistical control techniques have been successfully applied to the detection and isolation of process faults. Because those strategies evaluate the current process state using the measurement values and the normal operation model, their performance is strongly influenced by the sensor network installed in the plant. Nevertheless, very few sensor location approaches have been presented to enhance faults isolation, and they do not guarantee that a fault can be distinguished from the other ones before its magnitude reaches a certain critical value. In this work a strategy for updating process instrumentation is presented that aims at detecting and isolating a given set of process failures using statistical monitoring procedures before the fault magnitudes exceed predefined threshold values. In this sense, fault isolation constraints are formulated and incorporated to the instrumentation update optimization problem. The proposed restrictions are expressed in terms of the variable contributions to the inflated statistics. These are used on line to determine the set of observations by which a fault is revealed, but they have not been incorporated into the sensor location problem for fault diagnosis until now. That problem is solved using an enhanced level traversal tree search, which takes advantage of the fact that the structural determinability of a fault is a necessary condition for its isolability. Application results of the methodology to the Tennessee Eastman Process are presented.