Research On UAV Online Fault Detection Based On Flight Data Of Flight Control System

In recent years,Unmanned Aerial Vehicle(UAV)has been widely used in both military and civilian fields.Intelligence and swarm have become the main trend of UAVs in the future.However,the UAV accidents caused by system failures still account for a large proportion,which seriously affects the flight safety and mission success rate of UAVs.Meanwhile,the existing UAVs mainly rely on the ground platform to perform telemetry data analysis or flight data offline analysis,and fault detection based on expert experience thresholds or rules is realized,which has issues such as dependence on telemetry links,weak detection capabilities for complex fault modes,and insufficient timeliness of detection.This means that there are still deficiencies in UAV airborne flight data analysis and complex fault detection capabilities.Therefore,the research on online fault detection of UAV should be carried out to timely and effectively detect the UAV fault or potential safety hazard through the analysis of flight data,and provide decision support information for the UAV to take fault mitigation measures,which is of great value and significance for improving the autonomous health monitoring capability of the UAV,ensuring flight safety and mission success rate.Flight data from the UAV flight control system(FCS)are an important part of UAV flight data.It not only reflects the status of the FCS,but also indirectly reflects the influence of the status of the structure,power,and other systems on UAV flight status,which provides a data basis for conducting research on UAV online fault detection.At present,research on UAV online fault detection based on flight data still has the following challenges,such as lack of adaptability to multiple flight modes,insufficient detection capabilities for low-amplitude faults,and the mismatch of airborne computing resources with the computing requirements of the detection model.To address the above issues,the online fault detection of UAV is focused on in this dissertation from two aspects: the improvement of fault detection capability and the computing acceleration of the fault detection model.The main research and contributions are as follows:(1)The ability of automatic labeling and online recognition of the flight mode is insufficient,which affects the performance of UAV fault detection.To address this issue,a two-stage flight mode online recognition(TFMOR)method combining unsupervised labeling and supervised recognition is studied.First,an unsupervised labeling method based on the recursive Gaussian mixture model is established.It uses labeling task decomposition,recursive clustering,and interval detection of labeling results to realize automatic labeling of flight modes.Then,a supervised online flight mode recognition method based on a sliding window improved deep neural network(DNN)is constructed.It uses the labels to establish a DNN-based recognition model and introduces a sliding window detection method to suppress short-term recognition errors.Experimental results based on the data from the simulated flight and real flight show that compared with the other flight mode labeling and recognition methods,the TFMOR method can obtain better labeling and online recognition accuracy.It provides effective flight mode switching information for flight data analysis and UAV fault detection modeling.(2)The regression-based UAV fault detection method has deficiencies in low-amplitude fault detection and adaptability to dynamic flight modes.To address the above issues,a flight mode adaptive fault detection(FMAFD)method based on regression and detection strategy optimization of the prediction residual is studied.First,a regression model based on stacked LSTM is established to automatically capture the spatiotemporal features of flight data,which realizes the prediction of the detected flight parameter.Second,the moving average filter is introduced to enhance the fault characteristics in the prediction residuals,so that the detection of the low-amplitude fault can be achieved.Then,a threshold adjustment factor for fault detection based on the flight mode recognition results of the TFMOR is established,which makes the detection threshold automatically adapts to the flight modes.Finally,the predicted value of the regression model is used to replace the observed value as the recovery data of the fault data,which can support the stable control of the UAV flight attitude.Experimental results based on the data from the simulated flight and real flight show that compared with other regression-based fault detection methods,the FMAFD method has better regression and fault detection performance.It provides a basis for UAV online fault detection.(3)Airborne computing resources and power are limited,making it difficult for UAV fault detection models to operate online.To address this issue,a computing acceleration method of the fault detection model based on model reduction and customizable computing is studied.First,a model reduction method based on structured pruning is established.It takes the performance loss of prediction and fault detection as constraints and reduces the dimension of the weight matrices in the FMAFD model to prune the FMAFD model.Then,a computing optimization method based on customizable computing for airborne multiple detection tasks is proposed.It optimizes the inference structure of the detection model by customizable computing,and further,under the constraints of FPGA resources,multiple detection tasks are scheduled based on detection model scales clustering and greedy algorithm,and multi-scale customizable computing units with the ability of parameter dynamic configuration are constructed,which achieves the online operating of multiple detection tasks.Experimental results based on the data from the simulated flight and real flight show that compared with other available airborne computing platforms,the proposed computing acceleration method achieves high operating efficiency while the performance loss of the model is controllable,and has the advantage of airborne energy-efficiency computing.It provides a feasible solution for the online application of the UAV fault detection model.(4)The hardware-in-the-loop simulation(HILS)platform is used to test and verify the UAV online fault detection method.First,the HILS platform of a real UAV and the online fault detection platform are constructed.This realizes the simulation of the airborne bus,monitoring data stream,service delay,etc.in the real airborne operating environment.Second,the proposed online fault detection method is tested and verified by dynamically replaying the real flight data of the UAV.The testing and verification results show that the proposed online fault detection method can be used for the airborne online fault detection of the real UAV.The data interaction of the airborne bus,flight data stream processing,and other functions work correctly.The detection performance is consistent with the results obtained in PC,and it has the ability to achieve energy-efficient online operation of multiple fault detection tasks during the flight control cycle.This work lays the foundation for the proposed UAV online fault detection method to achieve airborne deployment in the future.