Comparative Evaluation of Classification Methods Used in the Fault Diagnosis of Industrial Processes

Abstract

This article presents a comparative study of the performance of classification techniques used for fault diagnosis in industrial processes. The techniques studied ranging from classifiers based on Bayes theory as Maximum a Posteriori Probability (MAP) and Nearest Neighbor (kNN) classifiers, through minimizing an objective function such as Artificial Neural Networks (ANN) and Support Machines Vector (SVM) and ending with the parameter estimation technique Partial Least Squares (PLS). Comparison of these techniques is based on the capacity of classification of the historical data and the generalization of new observations. Also, a discussion about the robustness of the classifiers against the dimensionality reduction process is presented. The study was conducted using the data from the testing process "Tennessee Eastman Process" (TEP).