Novelty Detection in Rotating Machinery: Assessment of Unsupervised Machine Learning Models for Medium-Sized Industrial Bearings

Early anomaly detection in industrial rotating equipment is crucial for enhancing predictive maintenance methodologies and preventing unexpected failures. However, real-world industrial settings often lack structured fault-labeled datasets, making supervised classification approaches impractical. In such contexts, unsupervised novelty detection methodologies emerge as the most viable alternative, as they do not require labeled fault data and operate under the assumption that only normal conditions are available during training. This research investigates the application of novelty detection techniques for industrial bearing condition monitoring, comparing machine learning-based methods with fixed threshold-based strategies derived from standards. The analysis is conducted on a dedicated experimental dataset, representing the first contribution of its kind for spherical bearings with localized damage, tested under variable speed and load conditions. The findings reveal that fixed-threshold strategies result in an excessive number of false positives, limiting their practical applicability in industrial monitoring systems. Among the assessed machine learning algorithms, Isolation Forest achieved the highest recall, detecting the largest number of anomalies, while Local Outlier Factor (LOF) demonstrated superior precision, accuracy, F1-Score and precision-recall AUC.