A novel index for the robustness comparison of classifiers in fault diagnosis

Abstract

Abstract The design of robust data-based fault diagnostic systems can be formulated in terms of classification tasks. A diagnostic classifier designed to effectively minimize the false and missing alarm rates resulting from noise, uncertainty, and unknown disturbances while maintaining a relatively high performance can be defined as robust. This paper presents a novel criterion to compare off-line the robustness of classifiers. The proposed index allows to complement the estimated misclassification rate and to quantify the quality of any data-based diagnostic system more rigorously. In order to evaluate the effectiveness of the proposed index, both Artificial Neural Networks and Support Vector Machines are used as diagnostic classifiers for the Continuous Stirred-Tank Reactor benchmark.