Steady-state identification, gross error detection, and data reconciliation for industrial process units

Processing of online data for use with steady-state models requires identification of the existence 14 of a steady state in a process, detection of the presence of gross errors, if any, and data reconciliation to eliminate random measurement noise. The method of Cao and Rhinehart (J. Process Control 1995, 5 (6), 363-374) was modified for steady-state identification by optimizing the filter constants so as to minimize type I and II errors and simultaneously reduce the delay in state detection. A simple algorithm has been presented that makes use of past measurements and the Kalman filter to detect the presence of gross error and estimate its magnitude. This algorithm simultaneously reconciles data for random measurement errors. Another algorithm that makes use of the exponential filter along with the least-squares minimization strategy is also applied for data reconciliation and found to perform equally well. All of these algorithms have been applied to data from an industrial crude distillation unit. The presence of strong autocorrelation in the industrial data was dealt with by adding random noise. This strategy makes the steady-state identification algorithm more suitable for online application.