Fault identification of ball bearings using Fast Walsh Hadamard Transform, LASSO feature selection, and Random forest classifier

Abstract

To reveal the machinery health condition, time-frequency analysis is an effective tool when signals are non-stationary. To identify bearing faults, numerous techniques have been proposed by various researchers. However, little research focused on image processing-based texture feature extraction for the identification of faults. The time-frequency image contains many sensitive fault information regarding bearing conditions, which can be extracted in the form of features. Therefore, in this paperwork, a methodology is proposed based on Fast Walsh Hadamard Transform (FWHT) time-frequency spectrogram, gray level co-occurrence matrix (GLCM), and machine learning techniques. A feature vector is constructed which consists of one dimension and two-dimension features extracted from Fast Walsh Hadamard Transform coefficients. To identify the fault conditions, LASSO-based feature ranking is applied to determine the suitable features. Finally, classifiers like Support vector machine (SVM), Random forest, and K-nearest neighbors (KNN) are evaluated for identifying bearing faults. Training, Testing, five-fold cross-validation performed on fusion feature vector. Results indicate that ranked fusion features are effective to diagnose bearing faults with good accuracy.