Machine Learning-based Bearing Fault Classification Using Higher Order Spectral Analysis

In the defense sector, where mission success often hinges on the reliability of complex mechanical systems, the health of bearings within aircraft, naval vessels, ground vehicles, missile systems, drones, and robotic platforms is paramount. Different signal processing techniques along with Higher Order Spectral Analysis (HOSA) have been used in literature for the fault diagnosis of bearings. Bispectral analysis offers a valuable means of finding higher-order statistical associations within signals, thus proving to detect the nonlinearities among Gaussian and non-Gaussian data. Their resilience to noise and capacity to unveil concealed information render them advantageous across a range of applications. Therefore, this research proposesa novel approach of utilizing the features extracted directly from the Bispectrum for classifying the bearing faults, departing from the common practice in other literature where the Bispectrum is treated as an image for fault classification. In this work vibration signalsare used to detect the bearing faults. The features from the non-redundant region and diagonal slice of the Bispectrum are used to capture the statistical and higher-order spectral characteristics of the vibration signal. A set of sixteen machine learning models, viz., Decision Trees, K-Nearest Neighbors, Naive Bayes, and Support Vector Machine, is employed to classify the bearing faults. The evaluation process involves a robust 10-fold cross-validation technique. The results reveal that the Decision Tree algorithm outperformed all others, achieving a remarkable accuracy rate of 100 %. The naive Bayes algorithm also demonstrated the least performance, with an accuracy score of 99.68 %. The results obtained from these algorithms have been compared with those achieved using Convolutional Neural Network (CNN), revealing that the training time of these algorithms is significantly shorter in comparison to CNN.