Machine learning-driven fault detection in rotating machinery: A predictive maintenance framework
Abstract. The traditional condition monitoring methods, such as threshold-driven vibration warning systems and periodic check-up programs, do not have the discriminative ability needed to differentiate between incipient faults signatures and normal operational variation. This paper proposes a new predictive maintenance model that incorporates multi-domain feature extraction, and hybrid feature selection approach and optimized ensemble classifier to facilitate high-fidelity fault detection in four operational modes, which are normal operation, inner-race bearing fault, outer-race bearing fault, and gear tooth spalling. Vibration signals are acquired at 12 kHz using tri-axial accelerometers, pre-processed through adaptive band-pass filtering and Z-score normalization, and subsequently transformed into a 47-dimensional feature space spanning time-domain statistics (RMS, kurtosis, skewness, crest factor), fast Fourier transform (FFT) spectral amplitudes, and discrete wavelet transform (DWT) energy coefficients. A two-stage feature selection pipeline—combining mutual information ranking with recursive feature elimination reduces the feature space to 18 optimal descriptors. An ensemble of random forest with Bayesian-optimization in cross validation of 10 folds and stratified cross-validation reach a mean accuracy of 98.7 percent, which is statistically better than support vector machine (SVM) and Artificial Neural Network (ANN) baselines by 4.1 and 2.9 percentage points respectively. The proposed framework demonstrates strong generalisability across variable operating speeds and load conditions, affirming its suitability for deployment in industrial cyber-physical monitoring systems.