Application of the asymptotic detection method for disturbance identification in industrial plant test data

Industrial model predictive controllers are ubiquitous in refining and chemical processing facilities. Usually these controllers are based on empirical models identified from plant data. If a disturbance is present within this training data, the identified models will have reduced fidelity to the true plant, and control performance will suffer. To identify and remove disturbances, the asymptotic detection method has been proposed in literature, but it has previously only been applied to case studies with simulations. Because the method is known to be sensitive to nonlinearity and noise, this study was undertaken to see if the asymptotic detection method could be applied effectively to industrial plant test data.

This study includes an evaluation of how both the user-selected sampling numbers and the model form impact the detection results of the asymptotic detection method. The primary focus is processing two separate data sets taken from industrial plant tests. These results show that the raw results of the asymptotic detection method are too sensitive for industrial applications, but the results are still useful for identifying the most significant disturbances. The slices suggested by the asymptotic detection method are compared against the slices made by the control engineers who developed the industrial controller. Although the true disturbances are not and can not be known, the relative locations of disturbances suggest the results are reliably identifying real disturbances. Suggestions are made about several ways the asymptotic detection algorithm could be used to greatest effect in industrial applications.