Machine learning pipeline for bearing fault diagnosis using statistical and spectral features

Bearing fault diagnosis is critical for predictive maintenance in rotating machinery. This paper presents a comprehensive approach combining signal processing techniques and machine learning models aimed at accurate fault classification. We extract time, frequency, and time-frequency domain features from vibration signals, including statistical parameters, wavelet coefficients, and envelope spectra. Both traditional machine learning models (Decision Trees, Random Forest, KNN, Naive Bayes, XGBoost) and deep learning architectures (MLP, CNN, ResNet1D, TCN, WaveNet, Temporal Attention) are evaluated for fault classification. The methodology comprises data preprocessing, feature extraction, and model training on a dataset containing healthy, inner race fault, and outer race fault bearing signals. Experimental results demonstrate that all deep learning models as well as traditional methods were able to classify the fault accurately. The proposed comparative framework provides an automated, robust solution for industrial bearing fault detection, reducing maintenance costs and downtime.