TCPPN: A physics-informed prototypical network with temporal convolution for UAV fault diagnosis with small samples

The widespread application of unmanned aerial vehicle (UAV) systems has imposed increasingly stringent demands on their reliability and safety. Deep learning provides a new technical approach for UAV fault diagnosis. However, the supervised learning paradigm is severely limited by the scarcity of labeled fault data. This scarcity leads to diagnostic models constructed in high-dimensional feature spaces that are prone to overfitting and exhibit insufficient generalization capabilities. To address these challenges, an end-to-end prototype learning framework incorporating physical information embedding is proposed. By incorporating UAV dynamic equations, flight data is enhanced into flight state feature data. This provides additional prior knowledge under small-sample conditions. Then, combined with raw UAV flight data, deep feature extraction is achieved by a temporal convolutional network to fully uncover fault characteristics. Next, a Physical Perception Prototypical Network (PPN) is designed. Physical information constraints and physical subspace metrics are embedded into the prototypical network, so that each prototype category contains more physical information. Finally, the model is validated and analyzed on a flight dataset. Results indicate that the proposed model achieves up to 99.71% classification accuracy and consistently exceeds 97.5% across different datasets and metrics, outperforming existing small sample and temporal fault diagnosis methods. SHAP analysis illustrates the effectiveness of the physical information embedding framework.