Multi-Rate Vibration Signal Analysis for Enhanced Data-Driven Monitoring of Bearing Faults in Induction Machines

In the realm of data-driven condition monitoring for induction machines, the widespread adoption of machine learning has been notable, particularly for identifying bearing faults from noisy vibration signals in electrical setups. However, the quality of the signals used to train these machine learning algorithms is important to ensure good classification performances. This paper investigates whether optimizing the Fast Fourier Transform of a vibration signal through multi-rate resampling and windowing techniques enhances the quality of the statistical features used for data-driven condition monitoring of ball bearings in induction motors. To do so, the features obtained under different frequency resolutions have been studied along with the performances of two classes of machine learning algorithms trained with them: support vector machines and random forests. The results show that for a given frequency resolution, higher sampling frequencies generally lead to improved machine learning performance. However, they also highlight the trade-off between high-frequency resolution in the Fast Fourier Transform and the number of samples available in the dataset for training. Finally, the use of the Hanning windowing technique was found not to significantly improve the quality of features for machine learning performance.