Interpretable attention-based prototype network for UAV fault diagnosis under small sample conditions

Fault diagnosis for Unmanned Aerial Vehicles (UAVs) is a typical ”small data” problem due to the scarcity of failure data. However, existing few-shot methods face challenges in performing UAV fault diagnosis with explicit explanations in real-world flight scenarios. To address this challenge, we propose an interpretable attention-based prototype network (IA-PN) for both UAV few-shot fault classification and unknown fault identification, utilizing flight parameters measured by multiple sensors, while also automatically explaining the diagnostic results. The IA-PN framework constructs an interpretable network structure within the prototype network to mitigate overfitting with small samples. To enhance the model's interpretability, a novel interpretable attention mechanism is employed to mine and locate the most crucial features in both the temporal and parameter domains. This mechanism provides local explanations in the form of parameter importance and identifies anomalous flight parameters for each fault category, offering insights into fault characteristics. Under unsupervised conditions, the proposed interpretable diagnostic framework can identify unknown faults based on relevant known fault data and provide potential explanations. Experimental results on two real UAV flight datasets, along with visualizations of interpretations in the parameter space, demonstrate that our method is both effective and interpretable in UAV fault diagnosis under small sample conditions.