基于神经网络的旋翼无人机空中原位风速反演

旋翼无人机空气动力学模型能够克服无人机旋翼下洗气流的影响，相比携带传感器直接测风具有一定优势。空气动力学模型利用无人机自身的姿态数据可以直接计算得到风向，但是风速的计算还需要开展大量的风洞实验以及复杂的实体建模获取相关参数。为简化流程、规避参数获取过程中存在的误差，采用BP(back propagation)神经网络替代风洞实验和实体建模，研究无人机姿态同自然风速的关系。通过确定风速影响因子，分别建立以两个欧拉角为输入和以两个欧拉角计算得到的倾角为输入的旋翼无人机空中原位风速反演模型EUL-BP(euler back propagation)和INC-BP(incline back propagation)。以无人机悬停等高度风塔上超声风速仪观测自然风速为真值，比较反演模型EUL-BP和INC-BP发现，INC-BP增强了BP神经网络寻优过程的约束，反演效果较EUL-BP更优，其均方根误差大致为0.65 m/s,相关系数为0.86左右；比较反演模型结果与无人机机载自动气象站观测结果发现，当风力条件为三级及以下时，反演风速远优于携带传感器观测的风速。

Wind Velocity Inversion of Hovering UAV Based on Neural Network

The aerodynamic model of UAV can overcome the influence of rotor underwash airflow and has advantage over wind measurement with sensors. The aerodynamic model can directly calculate the wind direction by using attitude data, but the calculation of wind velocity requires complex wind tunnel experiments and physical modeling to obtain relevant parameters. In order to simplify the process and avoid the errors in the process of obtaining parameters, BP neural network was used to replace the wind tunnel experiment and physical modeling to study the relationship between attitude and natural wind velocity. By determining the influencing factors of wind velocity, the wind velocity inversion models EUL-BP and INC-BP were established which two Euler angles as input or the inclination Angle calculated by two Euler angles as input. The natural wind velocity observed by ultrasonic anemometer at hovering height on the wind tower is true. Comparing the inversion model EUL-BP and INC-BP found that INC-BP enhances the constraint of BP neural network optimization process, and the inversion result is better than EUL-BP. Its RMSE is about 0.65m/s, and the correlation coefficient is about 0.86. The comparison between the inversion model results and the observation results of the airborne automatic weather station shows that the inversion wind velocity is much better than that observed by sensors when the wind condition is level 3 or below.