Generalized Bearing Fault Diagnosis Using Comparative Time-Frequency Scalogram Features and Machine Learning With Advanced Feature Ranking

Bearings are critical components of rotating machinery, ensuring smooth operation by supporting shafts and loads. Even minor defects can severely degrade performance, causing unplanned downtime and economic loss. Accurate and timely fault diagnosis is therefore essential for condition-based maintenance. Conventional time- and frequency-domain methods often fail to capture the non-stationary and transient nature of bearing vibration signals. This work proposes a time–frequency diagnostic framework using Continuous Wavelet Transform (CWT)–based scalograms and machine learning. Coiflet and Morlet wavelets are selected using the Maximum Relative Wavelet Energy (MRWE) criterion, yielding peak MRWE values of 0.1347 for the CWRU dataset and 0.0584 for the Machinery Fault Simulator (MFS) dataset, ensuring optimal fault localization. Scalograms are converted into RGB images, from which 21 texture features are extracted and ranked using the Fisher score. The method is validated on two independent datasets: the benchmark Case Western Reserve University (CWRU) data and an experimental MFS dataset using SKF 6004 bearings with induced inner-race, outer-race, and ball defects. Logistic Regression, SVM, KNN, and Bagged Tree classifiers are evaluated using tenfold cross-validation. On the CWRU dataset, KNN and Bagged Tree achieve accuracies of 98.3% with eight ranked features, while the Bagged Tree reaches 100% accuracy using all 21 features.