【摘要】 针对扑翼飞行器面向控制建模时无法直接测量气动参数并精确建立气动模型的问题,传统BP网络辨识法依据扑翼飞行器试飞数据,使用BP网络计算当前气动参数,再结合扑翼飞行器动力学模型计算其飞行状态,与试飞数据比较后,将误差经扑翼飞行器动力学模型反向传播至BP网络来更新网络参数。实验表明传统方法计算精度较低,且动力学模型复杂度高,存在梯度消失问题,为此提出一种基于双BP神经网络的气动参数辨识方法。该方法首先采用一个BP网络对扑翼飞行器动力学模型进行逆向辨识,为后续气动参数辨识提供理想网络计算模型,再结合批量随机梯度下降法用另一BP网络将扑翼飞行器柔性等非线性因素综合到待辨识气动模型中,实现扑翼飞行器气动参数辨识。实验结果表明所提双BP神经网络法在辨识精度、模型复杂度和模型训练时间等方面均优于传统BP网络法。更多还原

【Abstract】 Flapping-wing Micro Aerial Vehicle(FMAV) can not establish aerodynamic parameter model exactly by measuring the aerodynamic parameters during the design of the control system. The traditional BP network identification method is based on the flight test data of the FMAV, and the BP network is used to calculate the current aerodynamic parameters, and then the FMAV dynamics model was combined to calculate the flight states. After comparing flight states with flight data, the deviation is transmitted back to the BP(Back Propagation) network by the dynamics model to update the network parameters. Experiments show that this traditional method has low computational accuracy, and the dynamics model has high complexity that may cause the gradient disappearance problem. For this reason, a method of aerodynamic parameter identification based on Double BP(DBP) neural network was proposed. Firstly, one BP network was used to identify the FMAV dynamics model inversely, which provides an ideal network calculation model for aerodynamic parameter identification later. Combined with the nonlinear fitting characteristics of another neural network, the non-rigid body flexibility of the FMAV was also considered and integrated into the aerodynamic model, the batch random gradient descent algorithm was used to iteratively update the network parameters, and realize the aerodynamic parameter identification of the FMAV. The experimental results show that the proposed DBP neural network method is superior to the traditional BP network method in terms of identification accuracy, model complexity and model training time.更多还原