Aplicação de SVM na classificação de falhas em rolamentos: uma comparação entre o domínio tempo e tempo-frequência

Abstract

In order to enhance the operational efficiency of mechanical components, predictive maintenance has emerged as a critical technique for detecting potential failures before they lead to severe consequences such as accidents, productivity losses, and unexpected disruptions in production processes. This investigation employed vibration analysis to differentiate faults in spherical bearings, utilizing vibration signals collected from both faultless bearings and those bearing point defects. Statistical descriptors such as standard deviation, root mean square value, and shape factor were employed in the time domain, while the wavelet packet transform, featuring energy values and Shannon entropy in the time-frequency domain, was used to perform the analyses. The data was classified using the Support Vector Machine (SVM) machine learning algorithm. Both approaches demonstrated their effectiveness in discerning the signals, with prediction accuracies of 97.92% and 100%, respectively, highlighting the feasibility of this technique for fault identification in spherical bearings. The time-frequency approach was more effective than the time domain approach, as the latter yielded misclassifications in certain cases.