Digital Twin-driven cross-domain fault diagnosis for axial piston pumps via deep transfer learning under small-sample condition

• Published in: Journal of industrial information integration Vol. 48; p. 100966
• Main Authors: Ma, Chunjie, Yan, Ping, Wang, Bocheng, Gao, Lin, Du, Junlei, Feng, Lianqiang, Zhou, Han
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.11.2025
• Subjects: Axial piston pump; Cross-domain; Digital Twin model; Fault diagnosis; Small sample; Transfer learning; Fault diagnosis; Digital Twin model; Small sample; Axial piston pump; Transfer learning; Cross-domain
• ISSN: 2452-414X
• DOI: 10.1016/j.jii.2025.100966

•A high-fidelity multi-physics Digital Twin model of an axial piston pump is constructed through multi-domain co-simulation integrating multibody dynamics and hydraulic systems.•A Multi-source fusion Gramian Angular Summation Field feature encoding algorithm is proposed to achieve multi-source information fusion, consolidating and enhancing information representation.•A novel multi-generator ACGAN architecture is designed to improve distributional consistency between generated and measured samples while enhancing quality stability.•A Multi-scale Attention Domain Adversarial Transfer Networkdiagnosis model is proposed, leveraging fused multi-source information to enable precise and robust industrial information-integrated diagnostics in the case of small samples. Axial piston pump is a complex and typical thermal-fluid-structural coupled system. Its reliability directly affects the operational stability of the complex hydraulic system. It faces challenges including scarce fault samples and data distribution discrepancies across operating conditions. Regarding the problem that traditional methods fail to effectively integrate and utilize multi-source information, resulting in incomplete description of fault information, this paper proposes an intelligent cross-domain industrial information integration fault diagnosis method that integrates Digital Twin and adversarial transfer. Firstly, a multi-domain coupled Digital Twin model is constructed to generate multi-source fault simulation information data. The model employs co-simulation of multi-body dynamics and hydraulic systems to ensure the physical fidelity of fault information. Multi-source fused Gramian Angular Summation Fields feature encoding is designed to map multidimensional signals into two-dimensional spatiotemporal correlation images, thereby integrating and enhancing the representation of information. Secondly, an improved Auxiliary Classifier Generative Adversarial Network with multiple generators is adopted to align the distributions of simulated and measured data, with a dynamic optimization strategy employed to enhance generation quality. Finally, a Multi-scale Attention Domain Adversarial Transfer Network is constructed, combining a Gradient Reversal Layer and Conditional Maximum Mean Discrepancy to suppress the cross-domain distribution differences between the simulation and the experimental data. The experiment shows that by integrating experimental and simulation data, the proposed method achieves an average accuracy of over 98 % in cross-condition fault diagnosis tasks under unknown conditions, showing significant improvement over traditional transfer learning methods. Ablation studies validate the effectiveness of each module, providing a novel approach for complex hydraulic system fault diagnosis under small-sample scenarios.