Bearing Fault Diagnosis under Variable Speed using Machine Learning & Digital Signal Processing Techniques

Abstract

Rolling element bearings are crucial components of rotating machines, and are the leading cause of failure in essential industrial equipment such as rolling machines in paper mills, wind turbines, and induction motors. A faulty bearing may cause equipment breakdown, which can lead to unscheduled and costly downtime for an entire industry; thus, the detection of incipient bearing faults is essential for maintaining production schedules and minimizing costs. Various techniques have been used for fault detection in bearings including the analysis of vibration acceleration signals and the motor stator current. These techniques are useful in diagnosing bearing defects at high rotational speeds; however, incipient bearing defects at low rotational rates, which are characterized by their small energy acoustic emissions (AE), can be more efficiently diagnosed with AE-based methods. These methods can diagnose bearing defects before they appear on the bearing surface. In this thesis report, three AE-based techniques are presented for diagnosing faults in bearing. First is a method for diagnosing incipient bearing defects under variable speed conditions, by extracting features from different subbands of the inherently non-stationary AE signal, and then classifying bearing defects using a weighted committee machine, which is an ensemble of support vector machines and artificial neural networks. Second is a novel method for diagnosing bearing abnormalities under variable operating speeds using convolution neural networks (CNNs). The CNNs use the energy distribution maps of the AE signal spectrum as inputs, and automatically extract the optimal features that can be used to diagnose various single and compound bearing defects under variable speed conditions. In the third, we improve the performance of above CNNs-based method by training CNNs using Stochastic Diagonal Levenberg-Marquardt algorithm, a robust training algorithm that combines the Gauss-Newton and the steepest descent methods to exploit the speed advantage of the former and the stability of the later. It yields better convergence results, even for complex non-quadratic error functions. The proposed methods using data as AE signals are generated by the experimental testbed of Smart HSE Laboratory to validate the presented methodologies. Experimental results demonstrate that the proposed methods yield better diagnostic performance in comparison to state-of-the-art AE-based methods, by can extract optimal features and enhance the performance of classifiers...