A New Approach for Fault Diagnosis of Industrial Processes During Transitions

Abstract

This paper presents a new approach for fault diagnosis of industrial processes during transitions. The proposed diagnosis strategy is based on the combination of the nearest-neighbor classification rule and the multivariate Dynamic Time Warping time series similarity measure. The proposal is compared with four different classification methods: Bayes Classifier, Multi-Layer Perceptron Neural Network, Support Vector Machines and Long Short-Term Memory Network which have high performance in the specialized scientific bibliography. The continuous stirred tank heater benchmark is used under scenarios of faults occurring at different moments of a transition and scarce fault data. The proposed approach achieves a classification performance approximately 20% superior compared to the best results of the four instance-based classifiers.