Cross-Condition Aerodynamic Control Surface Fault Diagnosis for Fixed-Wing UAV With Dynamic Domain Adaptive Dropout Regularization and Hybrid Alignment

With the widespread use of fixed-wing UAVs (FW-UAVs) in many fields, the flight environment they face when performing tasks is becoming increasingly complex and changeable, which significantly increases the probability of faults occurring during operation, posing a serious challenge to the reliability of UAVs. Deep learning provides a new solution to the problem of UAV fault diagnosis. However, existing methods struggle to maintain stable fault diagnosis performance in such environments. To address this issue, this paper proposes a transfer learning framework (HDMAL) that combines a dynamic domain adaptation dropout regularization (DADR) mechanism with an explicit-implicit hybrid alignment strategy to resolve domain shift problems caused by changes in operating environments. HDMAL consists of a feature extractor, a DADR module, and a hybrid alignment module. First, the feature extractor uses convolutional neural networks (CNN) and long short-term memory networks (LSTM) to extract the spatiotemporal features of faults. Second, DADR evaluates the statistical discrepancies between the Source Domain (SD) and the Target Domain (TD) at the feature channel level and adaptively adjusts the suppression strength of each channel. Based on this suppression strength, channels are randomly dropout and rescaled, thereby suppressing channels that primarily representing domain features and enhancing the relevance and effectiveness of training. Then, hybrid alignment explicitly aligns the global feature distribution of the SD and TD through fixed mathematical constraints. Meanwhile, adversarial domain adaptation promotes the model to adaptively generate alignment strategies by confusing the decision boundary of the domain discriminator, thereby achieving implicit alignment of local nonlinear domain discrepancies and improving the model's cross-domain adaptability. Finally, based on a large number of experimental results from real FW-UAV flight data, the HDMAL method has shown excellent performance in multiple cross-working condition fault diagnosis tasks, significantly outperforming existing mainstream methods, demonstrating strong transferability and excellent application prospects.