A simulation and experimental study of transfer learning for mode detection of aeroengine fan noise

Abstract

The growing popularity of deep learning can be attributed to its capacity to model intricate nonlinear problems. However, this approach is constrained by its reliance on extensive datasets, making it less applicable in certain real-world situations. In the context of aeroengine applications, we propose a multi-constraint transfer learning framework that employs pressure measurements outside the turbofan duct to identify associated fan noise modes. We validate this methodology through both simulations and practical experiments. Traditional transfer learning methods apply knowledge derived from a single source task but face challenges due to factors like measurement location discrepancies and varying duct geometries. To mitigate these limitations, we establish a dataset comprising multiple source tasks, enabling our framework to focus on transfer learning with these diverse inputs. We contrast our method with traditional transfer learning across three scenarios: distinct measurement locations and heterogeneous geometries using the simulations, and the experimental tests. The results underscore the benefits of employing multiple sources in enhancing the accuracy and robustness of our deep learning model.