Fault detection based on functional relationship among process variables by autoassociative neural networks

Some process variables measured in a plant are strictly constrained by the material and heat balance equations, rate equations and correlations. In this study, we propose a method to judge whether the state of plant operation is normal or not, by examining whether a set of process variables maintains the functional relationship specified at normal operation. The functional relationship at normal operation is identified by an autoassociative neural network (AANN) which approximates the identity mapping for a set of measured values of process variables. An effective method to search for an adequate configuration of the AANN is also presented. Abnormal operation or fault is detected by the magnitude of discrepancy between the input vector and the output vector of the trained AANN. This fault detection method is applied to a continuous flow polymerization process and compared with the conventional 3 sigma fault detection method for a single process variable.