Uma Nova Abordagem para Classificação de Falhas baseada em Árvores de Decisão Induzidas por Programação Genética

Abstract

Fault detection and diagnosis in industrial sectors are standard practices in many sectors to maintain productivity, ensure safety, and implement efficient maintenance strategies. Therefore, this study proposes a data-driven methodology for fault detection and isolation in dynamic systems, integrating a fuzzy/Bayesian approach for change point detection and a novel multiclass classification technique using decision trees induced by genetic programming. The developed methodology comprises two stages. In the first stage, a combination of fuzzy set theory and the Metropolis-Hastings algorithm is used to monitor changes in sensor signals, identifying faults. In the second stage, which represents the main contribution of the study, the detected faults are classified using decision trees generated through genetic programming. This combination aims to provide a comprehensive and effective solution for fault detection and isolation in dynamic systems. The methodology was applied to a dataset adapted from the Tennessee Eastman process, successfully classifying 21 types of faults with an average accuracy of 81% in the training phase and 80% in the testing phase. These results demonstrate the efficacy and feasibility of the methodology for detecting and classifying faults in complex industrial systems.