Fault Detection of Small Fixed-Wing UAVs Based on Temporal Self-Attention and Long Short-Term Memory Network

To tackle the safety challenges encountered by small fixed-wing unmanned aerial vehicles (UAVs) operating in complex environments, we propose a deep learning model that integrates a temporal self-attention mechanism with a long short-term memory (LSTM) network for high-precision fault detection. The proposed method exploits the self-attention mechanism to explicitly capture long-range dependencies within sequential sensor data, while the LSTM network extracts dynamic temporal features, thereby enhancing the discriminative capability of fault types. Experimental results on real flight datasets demonstrate that the proposed method achieves over 97% accuracy, precision, recall, and F1-score, significantly outperforming conventional machine learning methods such as support vector machines. Moreover, ablation studies validate the effectiveness of integrating LSTM and self-attention, showing that the model substantially enhances robustness and performance in fault detection, thereby providing reliable technical support for the safe operation of small fixed-wing UAVs.