PIWich: Enhancing Quadrotor UAV Fault Diagnosis Under Imbalanced Data via Physics-Informed End-to-End Learning

Abstract

diagnosis is critical for the safe operation and maintenance of unmanned aerial vehicles (UAVs). As UAV applications expand, both data-driven and model-driven fault diagnosis methods have developed rapidly. However, these two mainstream methods for UAV fault diagnosis have inherent limitations: model-driven methods rely on precise physical models, which have narrow applicability and high research costs; while data-driven methods exhibit poor stability in small-sample and data-imbalanced scenarios due to their strong dependence on large-scale data. To improve the generalization of data-driven models and alleviate the feature learning problem resulting from insufficient physical constraints, this article proposes a novel physics-informed end-to-end fault diagnosis network, PIWich, for quadrotor UAVs. First, during feature generation in the encoder, PIWich introduces the quadrotor UAV’s lift and torques. Then, PIWich constructs dual constraints of physical loss and data loss based on the quadrotor’s dynamic equations. By integrating supervised encoding with dual constraints, PIWich incorporates physical priors into the fault classification framework. Comparative experiments on the public RflyMAD dataset show that PIWich has the capability to effectively learn common physical features across both majority and minority classes. Moreover, it significantly improves the fault diagnosis performance in imbalanced data scenarios.