Real-Time Fault Detection for UAV Based on Model Acceleration Engine

With the wide applications of the unmanned aerial vehicle (UAV) in the civilian and military fields, its operational safety has drawn much attention. A series of fault detection methods are studied to avoid disasters. Due to the capabilities of strong feature extraction and massive flight data processing, the deep learning-based methods have received extensive attention. However, restricted by UAV airborne size, weight, and power consumption, a significant challenge is posed to deploy these complicated detection methods in the airborne application, which requires to run in real time. In this article, a fault detection model acceleration engine (FDMAE) for UAV real-time fault detection is realized under the airborne constraint. First, a high-performance detection model is designed based on stacked long short-term memory networks, and fault detection is achieved by a statistical threshold in this method. Second, a model pruning method based on principal component analysis is proposed to improve computing efficiency. Finally, the pruned fault detection method is optimized and integrated as a flexible acceleration engine through high-level synthesis and deployed on an airborne embedded computing platform based on a field-programmable gate array. Real UAV flight data are used to verify the proposed FDMAE. By comparing accuracy, the area under the receiver operating characteristic curve, speed, and power consumption, the effectiveness of the FDMAE is proven.