Failure Diagnosis for UAV-Born Inertial Navigation Device Based on BP-Adaboost Model

When a certain UAV conduct a predetermined task, its flying track is planned in advance. The accumulated historical data of position information and respective state of the error can be used to diagnose the real time motion and position information deviation of the UAV-born inertial navigation devices through BP-Adaboost methods. The relationship between the position information deviation and characteristic parameters of UAV-born Inertial Navigation Device are in a high level of nonlinearity. As an efficacious non-linear classifier, BP neural network is capable of delineating the above nonlinear relationship. However, the performance and preciseness of the BP network are sensitively subject to the infuence of its initial weights, which is likely to incur the relatively unstable training outcomes. In order to address the contradiction and attain the training results we want, this paper provides a new approach to the UAV-born inertial navigation devices failure diagnosis based on BPAdaboost model. The first step is utilizing BP neural network as a weak classifier to discern the relationship between the failure states and the parameters. Each of the several BP networks chosen beforehand are applied to the repeated training process to become weak classifier one by one. Secondly, Adaboost algorithm is conducted by amalgamating the volidated weak classifier aiming at a strong classifier to diagnose the failure of UAV-born Inertial Navigation Device. Then, application of the BP-Adaboost method to a practical cases evinces that the failure diagnosis model of the strong classifier gains more satisfied accuracy than the weak classifiers, which can meet the requirements of the practices and offer another reference for the position deviation judgment during some repeating tasks for UAV-born inertial navigation devices, especially under the condition that GPS and inertial navigation devices are not stable and reliable.