Research On Accelerometer Fault Diagnosis Of Multi-rotor Aircraft Based On SVM Decision-tree

Multi-rotor aircraft,which features simple structure and flexible flight with the function of hovering and slow flying,plays an increasingly important role in fields such as power patrol inspection,border surveillance,modern agriculture,forest fire control,flood control and disaster prevention.However,multi-rotor aircraft is prone to flight failures,which can lead to aircraft crash,bringing enormous losses and even casualties.In this case,failures of multi-rotor aircraft become a key factor to restrain its application,among which sensor fault is a familiar failure.Therefore,this thesis tries to explore fault diagnosis methods of accelerometer sensor.This thesis begins with an analysis of the sensor system of multi-rotor aircraft with the construction of attitude estimation system composed by accelerometer and gyroscope.Then,the rationale of accelerometer fault diagnosis is analyzed,and three fault models of accelerometer are constructed according to the attitude estimation system of multi-rotor aircraft,which include constant deviation,constant gain and getting stuck.In addition,overall scheme of accelerometer fault diagnosis is designed,which divides fault diagnosis process into two phases,including residual sequence sovled by filter and classified by support vector machine(SVM).For the nonlinearity problem of observational equation of attitude estimation system,extended Kalman filter,unscented Kalman filter based on singularity decomposition and quaternion Kalman filter are designed to solve residual sequence.Besides,this thesis puts forward the method of adaptive particle filter to solve residual sequence,which is based on the importance sampling of quaternion Kalman filter and improved symmetric KL distance.As for the poor precision of importance sampling through priori probability density,importance sampling of quaternion Kalman filter is proposed in this thesis.For the weak real time performance of particle filter based on the importance sampling of quaternion Kalman filter,adaptive method based on symmetric KL distance is introduced,and improvements are made for the algorithm.According to the simulation results,importance sampling based on quaternion Kalman filter improves static and dynamic precision of particle filter while adaptive method based on improved symmetric KL distance saves operation time of particle filter.SVM is introduced to classify residual sequence of attitude estimation system of multi-rotor aircraft to achieve accelerometer fault diagnosis.When mononuclear SVM multiple classifier based on one-against-one multi-classification methods is used to classify residual sequence of the system,the precison is far from satisfaction;therefore,multinuclear SVM multiple classifier based on one-against-one multiclassification methods is designed in this thesis.As for the weak generalization of multinuclear SVM multiple classifier,SVM decision-tree is introduced to classify residual sequence and achieve accelerometer fault diagnosis.According to the simulation results,the diagnosis precision based on SVM decision-tree is higher than that based on either mononuclear or multinuclear SVM multiple classifier,while the generalization of SVM decision-tree is greater than that of multinuclear SVM multiple classifier.