Transfer learning fault diagnosis of axial piston pumps by fusing knowledge from multisource subdomains and simulation-driven soft labels

Axial piston pumps operate under dynamic conditions in practical applications, leading to cross-domain data distribution discrepancies that pose challenges for their deep-learning based fault diagnosis. While transfer learning has shown promise in mitigating such data distribution discrepancies between different domains, existing single-source transfer learning methods exhibit limited diagnostic accuracy when confronted with substantial domain shift. To address this challenge, this article proposes a novel transfer learning framework that fuses knowledge from multisource subdomains and simulation-driven soft labels. Each source subdomain has an independent domain-specific classifier, and its contribution to the final classification decision is weighted by the distribution discrepancy between source and target data. A computational fluid dynamics model is developed to simulate discharge pressure signals under target operating conditions. These simulated signals are used to compute soft labels through interclass similarity, which enables prior knowledge of target domain to be incorporated during classifier training. Experiments on an axial piston pump at various rotational speeds demonstrate that the proposed model increases the diagnostic accuracy by 28.03% over conventional single-source methods. Furthermore, the integration of simulation-driven soft labels yields an additional 13.59% accuracy gain. As a result, the proposed model achieves a superior average accuracy of 99.02%.