Physics-driven cross domain digital twin framework for bearing fault diagnosis in non-stationary conditions

Abstract

Existing methodologies for digital twin-based domain adaptation primarily focus on steady or variable working conditions, frequently encountering limitations in scenarios where operational conditions change over time, such as in the case of wind turbines subjected to fluctuating wind speeds. This paper proposes a novel physics-driven cross domain digital twin framework designed to address the challenges associated with diagnosing bearing faults in non-stationary conditions. The model incorporates a phenomenological bearing model that generates virtual datasets, capturing a diverse range of fault types under non-stationary conditions. Furthermore, it introduces a physics-driven adaptive domain adaptation approach that aims to reduce the disparity between simulated and real-world data. This approach dynamically aligns domain distributions from both global and local perspective, markedly enhancing the accuracy of fault diagnosis under non-stationary conditions using exclusively unlabeled real-world data. The efficacy and robustness of the proposed model are validated through applications on two distinct use cases, involving various bearing types and time-varying working conditions. This study significantly contributes to the field by being among the first to explore digital twin-based domain adaptation in non-stationary conditions.