Diagnóstico de falhas em processos industriais usando classificadores locais avaliados com diferentes características

Abstract

Faults in industrial process lead to equipment malfunction, which can cause permanent damage, risking plant personnel safety and reducing profits. In this context accurate fault diagnosis is fundamental. This work presents an approach for fault diagnosis in industrial process. The diagnostic is performed using classifiers and multivariate data analysis techniques. To improve diagnosis accuracy, faults are clustered by the influence of the variables. Therefore, a single classifier is replaced by multiple local classifiers. Using a single classifier for all faults can make the task of classification more complex and reduce the accuracy of the diagnosis, while local classifiers may be less complex and have a greater power of discrimination among different faults. In addition, to simplify the data to be analyzed by the classifiers, the proposed approach uses feature extraction to analyze the behavior of the process during the occurrence of a fault. The activities are carried out using MATLAB and the approach is applied to three case studies: the classification of time series available in databases in the literature, simulations of the Tennessee Eastman Process plant and simulations of a continuous stirred tank reactor