Support vector machine approach for model-plant mismatch detection

Abstract

We develop a model-plant mismatch (MPM) detection strategy based on a novel closed-loop identification approach and one-class support vector machine (SVM) learning technique. With this scheme we can monitor MPM and noise model change separately, thus separating the MPM from noise model changes. Another advantage of this approach is that it is applicable to routine operating data that may lack any external excitation signals. Theoretical analysis of the proposed closed-loop identification is provided in this paper, showing that it can give a consistent parameter estimate for the process model even in the case where a priori knowledge about the true noise model structure is not available. A set of normal operation data with satisfactory performance is collected as the training data. We build SVM models based on process and noise model estimates from training data to predict the occurrence of MPM in the test data. The proposed technique can be applied to both single-input-single-output (SISO) and multi-input-multi-output (MIMO) systems. Two examples from paper machine control are provided to verify the effectiveness of the proposed MPM detection framework.