Ambiente de monitoramento e análise interpretativa dos estágios de degradação de rolamentos: uma abordagem com algoritmo SHAP e classificação utilizando floresta aleatória

Abstract

Bearing degradation represents a critical challenge in rotating systems, potentially compromising performance and causing catastrophic failures. Monitoring the condition of bearings is essential to ensure the reliability of rotating systems. Given the high variability of physical and operational characteristics of bearing supports, such as the number of rolling elements, the type of element (spherical, cylindrical, tapered), the type of applied load (radial, axial or combined), the material of the components, the type of lubrication, the internal clearance, the assembly conditions and the rotational regime, the use of a single curve to represent the degradation process may not be sufficiently representative. Therefore, based on a literature review about bearing degradation stages, this work introduces an approach that allows a broader analysis of degradation over time. This work proposes, based on a public database of vibration signals from bearing supports, a methodology for automatic classification of degradation stages in bearing supports, using Fourier transform, Principal Component Analysis (PCA) and decision tree-based classifiers. The model achieved accuracy above 96%, demonstrating high effectiveness in identifying degradation stages. Comparisons with studies from the literature indicate superior performance in classifying degradation stages. Interpretative analysis with SHAP values identified the most relevant stages, their behaviors and respective contributions for each failure stage. These results validate the potential of the approach for real applications in predictive maintenance.