Instrument Surveillance and Calibration Verification: A Case Study Using Two Empirical Modeling Paradigms

The development of Instrument Surveillance and Calibration Verification (ISCV) systems for complex processes requires an empirical model to provide estimations of the process measurements. The residual differences between the estimations and measurements are then evaluated to determine the proper operation of the process sensors. This work presents the results of applying two different empirical modeling strategies to a set of 55 process sensors from a nuclear power plant. The application of the Neural Network Partial Least Squares (NNPLS) algorithm for ISCV systems has been developed at the University of Tennessee, and the Process Evaluation and Analysis by Neural Operators (PEANO) system utilizing autoassociative artificial neural networks (AANN) has been developed at the Institutt for Energiteknikk in Halden, Norway. The case study presented illustrates the performance of both systems on historical data which lacks the common features of high correlations, normally present in highly redundant signal sets. While redundant information typically causes numerical instabilities, due to a rank deficient predictor variable matrix, the NNPLS model first performs an orthogonal transformation to combat these instabilities. Moreover, the NNPLS results show a direct relationship between the maximum signal correlations, and the prediction accuracy of the empirical model. With a direct neural network approach, other methods are employed during training to stabilize the solution. Three months of data were available, from which the models were developed and evaluated. The average error of the NNPLS ISCV system predictions was 1.11%, calculated as the average absolute difference with respect to the mean value of the measured signals. Direct comparisons between the NNPLS and PEANO systems are presented for 5 sensors, including an example of a faulted steam flow sensor.