Proposta de metodologia para diagnóstico de falha em rolamentos de baixa velocidade

Abstract

Monitoring low-speed bearings with vibration analysis is more challenging due to the low energy level of the vibration signal that carries the failure data, making it more susceptible to interference from other sources, impairing the interpretation of information. Thus, an alternative is to calculate signal predictors that may be sensitive to pattern changes relative to failure onset and progression. Hence, the present work extracted different types of features, among them two nonlinear features and eleven extracted from the signal in the time domain. Those features were ranked and selected based on their sensibility to class differentiation, which was estimated using the t-Welch statistic value. Among them is the Largest Lyapunov Exponent, which, in this work, had a modification in one of its calculation steps, improving its sensitivity in some cases. In addition, the influence of the vibration signal window size on the class separability of the indicators was evaluated (which is a scarce content in low-speed bearing monitoring literature). After feature selection, the data were subjected to a linear transformation through PCA (Principal Component Analysis), aiming to reduce the data dimensionality to three dimensions and to minimize the redundancy effects of highly correlated features. In sequence, the data represented in the space of principal components were projected on a Hotelling T2 statistic control chart. The chart allowed the detection and rejection of potential outliers, which consisted of points above a limit line estimated based on F statistic distribution. Finally, binary and multiclass Support Vector Machine classification models were trained with experimental data acquired from normal conditions and three levels of incipient fault in bearing. The models performed well, mainly the binary model with test data obtained from belt conveyor pulley bearings in industrial operation.