Bearing fault detection and identification using Statistical Features and Feedforward Neural Network

Abstract

Rolling bearings are among the most critical components in rotating machinery, and their reliable condition assessment is essential for preventing unexpected failures and enabling predictive maintenance strategies. In this study, a lightweight and automated methodology is proposed for bearing Fault Detection (FD) and Fault Identification (FI) based on vibration signals. The approach combines statistical feature extraction, Principal Component Analysis (PCA), and an optimized Feedforward Neural Network (FNN), aiming to provide an effective trade-off between diagnostic accuracy, computational efficiency, and robustness to measurement noise. Time-domain indicators are extracted from accelerometer signals and subsequently reduced through PCA to obtain a compact input representation for the classifier. The methodology is validated using the experimental Case Western Reserve University bearing dataset, considering normal operating conditions and three fault classes related to inner race, rolling element, and outer race defects. A cost-benefit analysis is performed to evaluate the influence of the number of principal components on classification performance and model complexity. The optimised FNN architecture achieves high diagnostic performance in both FD and FI tasks, while preserving stable classification metrics when additive white Gaussian noise is introduced into the signals. Furthermore, the reduced computational cost of the proposed feature-based pipeline indicates its suitability for real-time condition monitoring applications. The results highlight the potential of combining compact statistical descriptors, dimensionality reduction, and shallow neural classification as an efficient and interpretable alternative to more computationally demanding diagnostic approaches.