Although deep learning has become a mainstream method for Unmanned Aerial Vehicle (UAV) fault diagnosis by virtue of its powerful feature learning capability, most of the existing diagnostic models suffer from insufficient environmental adaptability. Therefore, this paper proposes an Unsupervised Transfer Learning-based Feature Mapping (UTL-FM) for UAV rudder fault diagnosis (UTL-FM), which constructs a cross-domain knowledge transfer framework to efficiently transfer the diagnostic models learnt from the source domain to unknown flight scenarios by deeply mining the fault features and diagnostic knowledge in the known flight environment. The method firstly aligns the source and target domain samples based on the feature space mapping to ensure that the data of the two domains present uniform distribution characteristics in the feature space. Secondly, a minimization process is implemented on the feature distributions of the source and target domains using the Maximum Mean Difference (MMD) metric criterion to minimize the domain variances. Finally, extensive experiments were conducted in real flight cases to verify the superiority of UTL-FM.

UAV, fault diagnosis, transfer learning, MMD