Bearing Fault Detection and Severity Classification Using a Two-Step Neural Network Model

Abstract

Rolling element bearings are key components in rotating machinery, and their failures often lead to costly downtime and safety risks. Reliable fault diagnosis is therefore essential in predictive maintenance, where early detection and severity assessment can improve system availability. This paper presents a two-step neural framework for bearing fault analysis that combines supervised and unsupervised learning. In the first stage, a feedforward neural network classifies bearing conditions into four fault categories. In the second stage, a self-organizing map refines the diagnosis by distinguishing up to ten severity levels within the detected faults. The method uses features extracted from vibration signals, including statistical indicators and frequency-domain transformations, selected to balance accuracy and computational cost. Experimental results show that the proposed hierarchical approach improves diagnostic precision compared with single-stage classifiers, achieving over 95% accuracy for fault detection and distinguishing up to 10 severity levels based on defect size and location. This two-step framework demonstrates practical potential for robust fault monitoring in industrial environments.