ANN FRAMEWORK FOR BEARING FAULT CLASSIFICATION

Abstract

Bearings are among the most critical mechanical components in rotating machinery. A damaged bearing can have negative consequences for machine integrity and operational safety. Therefore, the ability to detect faulty bearings is important. In this work, the study investigated the use of vibration signals to classify bearing condition. The signals were divided into one-revolution segments. Each segment was converted into the frequency domain to create FFT magnitude features. These features were then fed to a compact MLP classifier using the Paderborn University dataset as a benchmark. Training was performed using data from artificially damaged bearings, while testing was performed on data from previously unseen bearings with real damage. The split was performed at the bearing level to avoid data leakage and support generalization. The study showed that performance increased as more spectral information was retained. Among the contiguous feature sets, the strongest baseline performance was observed with 512 FFT bins. Furthermore, it was found that performance is not only dependent on the number of retained features, but also on which FFT regions are included. Some non-contiguous compact regions outperformed several contiguous ones. With further refinement, the best run achieved 89.43% test balanced accuracy using only 320 bins. In addition, the framework was applied to a different speed condition. In this experiment, training was performed at 1500 rpm and testing was performed at 900 rpm. After frequency alignment, L2 normalization provided the best overall compromise. The study shows that a compact and carefully selected spectral representation can support reliable binary bearing fault classification.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Science and Engineering