A New Statistical Features Based Approach for Bearing Fault Diagnosis Using Vibration Signals.

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Condition-based maintenance relies on analyzing machinery signals like vibration to detect faults.
• Current methods often use time, frequency, and time-frequency domain analyses, integrating features with machine learning for fault classification.

Purpose of the Study:

• To detect and localize faults in roller bearings using statistical features from vibration signals.
• To classify bearing conditions into healthy, outer race fault, inner race fault, and ball fault categories.
• To evaluate the efficiency and accuracy of a novel feature extraction and classification method.

Main Methods:

• Statistical features (skewness, kurtosis, average, RMS) were extracted from time and frequency domains of vibration signals.
• Features were derived from the second derivative of time-domain signals and power spectral density.
• Concatenated feature vectors were classified using K-nearest neighbor (KNN), support vector machine (SVM), and kernel linear discriminant analysis (KLDA).

Main Results:

• Achieved over 95% reduction in feature set size, decreasing computational burden and classification time.
• Attained average classification accuracies of 99.13% with KLDA and 96.64% with KNN.
• Demonstrated superior performance compared to traditional Empirical Mode Decomposition (EMD) and Fourier transform features.

Conclusions:

• The proposed method effectively detects, localizes, and classifies roller bearing faults with high accuracy.
• Significant feature reduction enhances computational efficiency for condition-based maintenance systems.
• Statistical feature extraction from vibration signals offers a robust alternative to existing fault diagnosis techniques.