Target Label-Free Diagnosis of Rotor Bolted Joints via a Hybrid Physical-Data-Driven Partial Domain Adaptation Framework

Abstract

Flange bolted joints are widely used in the rotor systems of aeroengines. Intelligent fault diagnosis of these bolted joints is crucial for ensuring operational safety. However, due to the extreme scarcity of real-world fault cases, deep learning-based fault diagnosis methods are difficult to apply in engineering practice. Although simulation-to-reality (Sim2Real) transfer learning provides a promising solution, its application in rotor systems faces three challenges: 1) bolt looseness at different locations can excite the similar combination resonance frequencies, making feature differences weak and difficult to distinguish; 2) various inevitable uncertainties introduced during rotor system assembly are difficult to fully account for in mechanical modeling; and 3) the types of faults occurring in actual operation rarely cover all categories present in computational simulations, and this asymmetric label space ( $\mathcal {Y}_{t} \subset \mathcal {Y}_{s}$ ) leads to negative transfer. To overcome these difficulties, this article proposes a novel hybrid physical-data-driven framework, called physics-guided domain randomization and adversarial domain adaptation (Phys-DRADA). This method integrates physical model-based domain randomization (Phys-DR) and frequency-based normalization (F-Norm), eliminating absolute amplitude interference and establishing a robust physical baseline. Subsequently, a physics-guided operational-spatial representation network is proposed. It employs a harmonic-aware dilated spectrum extractor (HADS) to match the equidistant comb-like topologies in the vibration spectrum that shift with operating conditions, and utilizes an operational-spatial coupled transformer to extract the implicit rotor mode shape information from the vibration spectrum and characterize its dynamic variations with rotational speed, thereby distinguishing bolt looseness at different locations. Furthermore, a partial adversarial domain adaptation (PADA) mechanism is introduced to dynamically cutoff the feature alignment of source-private classes. Extensive experimental validations demonstrate that, entirely without prelabeled real-world fault data, Phys-DRADA achieves diagnostic accuracies of 99.33 % and 93.71 % under standard and partial domain adaptation (PDA) scenarios, respectively, providing a highly reliable algorithmic paradigm with practical engineering value for complex rotating machinery.