Fault diagnosis for flying-wing UAV sensors based on enhanced ensemble deep auto-encoder

As a kind of unmanned aerial vehicle (UAV) with new layout, the flying-wing UAVs have received increasing attention with unique advantages. Precise fault diagnosis for its critical sensor system can effectively enhance the safety of flight missions. Without the requirement of precise mechanism models, deep learning-based approaches can automatically excavate valuable information and identify the sensor faults intelligently. However, single deep learning model has deficiency in diagnostic precision and stability. In this study, one enhanced ensemble deep auto-encoder (EEDAE) model is proposed to simultaneously combine the advantages of ensemble strategy and deep learning models. First, different structure parameters are generated and diverse training subsets are randomly bootstrapped to increase the diversity of base models and extract critical features from raw data automatically. Meanwhile, to further enhance the effect of model integration, one enhanced weighted voting (EWV) strategy with threshold is designed to realize selective model ensemble through removing the models with poor performance and assigning the voting weights to the remaining models based on their diagnostic accuracy. Finally, the experimental results indicate that the designed EEDAE can realize prominent performance on sensor fault diagnosis.