Quadrotor fault diagnosis based on deep domain adaptive network

Fault diagnosis is crucial for quadrotor flight safety, but real data is limited due to complex scenarios and the high cost of fault generation. Transferring data from simulation to real flights offers a cost-effective and safe solution. For the problem of different distribution of simulation data and actual flight data of quadrotors, this paper proposes a cross-domain fault diagnosis method combining multi-metric alignment and uncertainty. The approach integrates Maximum Mean Difference (MMD), Coral, and uncertainty loss to align cross-domain feature distributions. Features of source and target domain data are extracted through a shared feature network. Based on the extracted features, MMD, Coral, uncertainty and cross-entropy losses are computed. These four losses are combined into a weighted total loss function, which is minimized to optimize the network and achieve accurate classification. The experimental results show that the fault classification accuracy of the proposed method in the target domain is much better than that of the traditional Deep Neural Network (DNN) method, the single-metric method, and the method combined single-metric and uncertainty. Multi-metric joint optimization mitigates cross-domain distribution shifts, with the uncertainty constraints further enhancing model robustness.