UAV Anomaly Detection Model Based on Integrated Multi-Modal Neural Network and Neural Architecture Search

With the rapid development of UAV technology, UAV anomaly detection has become one of the important tasks to ensure flight safety. However, traditional single-modal anomaly detection models face 3 challenges: 1. difficulty in fully extracting complex sensor features leading to information loss; 2. difficulty in integrating prediction results from heterogeneous neural networks reducing model performance; 3. difficulty in updating model parameters because of heavy reliance on prior knowledge. These limit the performance of UAV anomaly detection. To solve the above problems, this paper proposes a UAV anomaly detection model based on an integrated multi-modal neural network and neural architecture search. Firstly, various heterogeneous neural networks are designed to extract features from UAV sensor data and provide a basis for the subsequent decision-making phase. Secondly, the hybrid soft-voting mechanism combines the prediction results of each model neural network to improve the accuracy of model detection. Finally, neural architecture search technology (NAS) is employed to construct and optimize multiple base classifiers automatically. These improve the detection performance. The experimental results show that the method achieves 98.02% accuracy on the ALFA dataset, about 3.39% higher than the base model.