Uma Análise Comparativa do Desempenho de Algoritmos de Classificação no Diagnóstico de Falhas em Rolamentos de Motores Elétricos a Partir de Sinais de Vibração

Abstract

Electric motors are the most common equipment in industrial plants, and bearings are the most vulnerable components to failure. Considering the importance of electric motors operating in good conditions to maintain the continuity of the production process in industrial plants, this work investigates and compares the performance of several classification architectures when applied to the diagnosis of bearing failures. To build the models 13 statistical descriptors were extracted from the vibration signals available in the Paderborn data set. The models were built in the time domain and time-scale domain using the wavelet transform, and the k-Nearest Neigbour (k-NN), Support Vector Machine (SVM) and Decision Tree (DA) algorithms were applied. The performances of the models were evaluated using the metrics of accuracy, precision, sensitivity, specificity, and F1-score. The average result obtained in all classifier configurations was around 98%.