Multi-Class SVMs for Automatic Detection and Diagnosis in Process Control Loops

Abstract

The aim of this paper is to present a novel framework using Multi-Class Support Vector Machines (MC-SVMs) to classify the performance of closed loop single-input-single-output feedback controllers. A SVM is trained to recognize descriptive statistical patterns originating from an Autocorrelation Function (ACF) of process data vectors. ACF patterns emanating from different closed loop behaviors are used in the feature extraction procedure. Simulation study and application to real world industrial data sets show that the MC-SVM classification tool is capable of detecting and diagnosing problematic control loops with very good accuracy and efficiency.