Sensor Fault Diagnosis For Flight Control System Based On Neural Network

In the unmanned aerial vehicle (UAV) flight control system (FCS), sensors are used formeasuring flying state parameters. The fight control computer accepts the feedback signal from thesensors and implements the calculation of control law, so as to achieve the automatic flight mission ofUAV. Whether the sensors are working properly directly affects safety and reliability of unmannedflight. Therefore, the research on FCS sensors on-line fault diagnosis is of great significance.The main research content and research results in this thesis are as follows.Firstly, on the basis of understanding the existing fault diagnosis technology and according to thecharacteristics of FCS sensors, a method for sensor fault diagnosis based on the principle of neuralnetwork nonlinear system identification is proposed in this thesis, neural network observer groupswhich estimate the parameters measured by sensors is established.Secondly, aiming at sensor fault type recognition problem, linear regression is adopted to modeland analyze the fault residual information, which identified the type of constant deviation fault andconstant gain fault. And dual-threshold fault detection method is proposed to carry out on-line faultdiagnosis of FCS sensors. This method observes the residual sequences continuously by settingobservation time window, which realized diagnosis of working condition of the sensor.Again, this paper researched the basic and improved algorithm of neural network learning, andall of which are applied to neural network observer. By comparing convergence rate and test error, theresult shows the network learned by LM algorithm has faster convergence and higher precisionestimate, which can satisfy the accuracy requirements of FCS. Meanwhile, the best network structureis chosen from the performance of the neural networks with different hidden layer nodes.Finally, for certain UAV, according to the requirement of the lengthways and transverse controlin different height in matlab/simulink simulation environment, the stuck fault, constant deviation faultand constant gain fault for vertical gyro and rate gyro is diagnosed and analyzed on line in differentflight modes. Simulation results showed: the neural network observer groups can estimate the UAVflight parameters. The type of sensor fault can be recognized via linear regression analyze for outputsample of the faulty sensor and network. And double-threshold fault detection method can eliminateabrupt change of normal signal, reduced false alarming and enhanced fault diagnosis reliability. Withthe presented method, the feedback signal of control law is reconstructed to maintain UAV flight.