Representing and diagnosing dynamic process data using neural networks

Abstract A technique for representing and diagnosing dynamic process trend data using neural networks is presented. The approach employs a feedforward neural network with backpropagation as the learning algorithm. Two methods of presenting symptom patterns to the network, one using raw time-series values of measured process variables and another using a moving average value of the same time-series data, are described. Two methods of discretization of the desired output of the networks during training—a linear method and an exponential method—are also discussed. Networks with various numbers of hidden units were tested and compared with respect to their performance in recall and generalization. The results show that accurate recall and generalization behavior was observed in the diagnosis of single-fault measurement patterns. The networks trained using raw time-series data were able to diagnose untrained single-fault patterns sampled earlier in the fault-induced transient, than the ones trained using moving average data. It was possible to diagnose untrained single fault patterns in the transient stage, within 0.25 h of the occurrence of the fault, although the new steady state was reached only after 2 h. For patterns sampled in the later stages of the transient, the moving-average network performed as well as the time-series network.