Machine Learning Based Flight State Prediction for Improving UAV Resistance to Uncertainty

Unmanned Aerial Vehicles (UAVs) encounter various uncertainties, including unfamiliar environments, signal delays, limited control precision, and other disturbances during task execution. Such factors can significantly compromise flight safety in complex scenarios. In this paper, to enhance the safety of UAVs amidst these uncertainties, a control accuracy prediction model based on ensemble learning abnormal state detection is designed. By analyzing the historical state data, the trained model can be used to judge the current state and obtain the command tracking control accuracy of the UAV at that instant. Ensemble learning offers superior classification capabilities compared to weak learners, particularly for anomaly detection in flight data. The learning efficacy of support vector machine, random forest classifier is compared and achieving a peak accuracy of 95% for the prediction results using random forest combined with adaboost model . Subsequently, a trajectory planning method leveraging the DWA(Dynamic Window approach) algorithm was designed to mitigate the safety risks associated with uncertain control command tracking. By employing the obtained model of nominal command execution results of UAVs subjected to uncertainty, and by adjusting the original assessment criteria to a probability-weighted comprehensive optimal metric, optimal control commands that factor in uncertainty are derived. The simulation results affirm the effectiveness of the designed method.