Physics-Informed Residual Network (PIResNet) for rolling element bearing fault diagnostics

Abstract

Various deep learning methodologies have recently been developed for machine condition monitoring recently, and they have achieved impressive success in bearing fault diagnostics. Despite the capability of effectively diagnosing bearing faults, most deep learning methods are tremendously data-dependent, which is not always available in industrial applications. In practical engineering, bearings are usually installed in rotating machinery where speed and load variations frequently occur, resulting in difficulty in collecting large training datasets under all operating conditions. Additionally, physical information is usually ignored in most deep learning algorithms, which sometimes leads to the generated results of low compliance with the physical law. To tackle these challenges, a novel Physics-Informed Residual Network (PIResNet) is proposed for learning the underlying physics that is embedded in both training and testing data, thus providing a physical consistent solution for imperfect data. In the proposed method, a physical modal-property-dominant-generated layer is adopted at first to generate the modal-property-dominant feature. Then, a domain-conversion layer is constructed to enable the feasibility of extracting the discriminative bearing fault features under varying operating speed conditions. Lastly, a parallel bi-channel residual learning architecture that can automatically extract the bearing fault signatures is meticulously established to incorporate the bearing fault characteristics. Experimental datasets under variable operating speeds and loads, and time-varying operating speeds are utilized to demonstrate the superiority of the PIResNet under non-stationary operating conditions.