基于PINNs的单旋翼植保无人机下洗流场预测模型  

Prediction model for the downwash flow field of single-rotor plant protection UAV using PINNs

作　　者：王涛[1] 文晟[1,3] 兰玉彬 张海艳 尹选春 张建桃 WANG Tao;WEN Sheng;LAN Yubin;ZHANG Haiyan;YIN Xuanchun;ZHANG Jiantao(College of Engineering,South China Agricultural University,Guangzhou 510642,China;College of Electronic Engineering,SouthChina Agricultural University,Guangzhou 510642,China;National Center for International Collaboration Research on PrecisionAgricultural Aviation Pesticides Spraying Technology,Guangzhou 510642,China;College of Information and Electrical Engineering,Shenyang Agricultural University,Shenyang 110161,China;College of Mathematics and Informatics,South China AgriculturalUniversity,Guangzhou 510642,China)

机构地区：[1]华南农业大学工程学院,广州510642 [2]华南农业大学电子工程学院,广州510642 [3]国家精准农业航空施药技术国际联合研究中心,广州510642 [4]沈阳农业大学信息与电气工程学院,沈阳110161 [5]华南农业大学数学与信息学院,广州510642

出　　处：《农业工程学报》2023年第6期83-91,共9页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金项目(32271985);广东省自然科学基金项目(2022A1515011008);广东省普通高校特色创新类项目(2019KTSCX016)。

摘　　要：植保无人机(unmanned aerial vehicle,UAV)进行喷施作业时,旋翼高速旋转所产生的下洗流场是影响雾滴飘移的重要因素。为了快速准确地预测单旋翼植保无人机下洗流场的速度等流场参数,提升无人机精准施药效果,该研究基于物理信息神经网络(physics-informed neural networks,PINNs)构建了单旋翼植保无人机下洗流场的预测模型。在全连接神经网络结构的基础上,嵌入纳维-斯托克斯(Navier-Stokes,N-S)方程作为物理学损失项来参与训练,减轻网络模型对数据依赖性的同时增强了模型的可解释性。通过最小化损失函数,使得该模型学习到流场中流体的运动规律,得到时空坐标与速度信息等物理量之间的映射关系,从而实现对单旋翼无人机下洗流场的速度等参数的快速预测。最后通过风洞试验验证该预测模型的可行性和准确性。结果表明:没有侧风的情况下,预测模型在旋翼下方0.3、0.7、1.1以及1.5 m共4个不同高度处各向速度的预测值和试验值的误差均小于0.6 m/s,具有较小的差异性;不同侧风风速情况下,水平和竖直方向速度的预测值与试验值的总体拟合优度R2分别为0.941和0.936,表明所提出的模型在单旋翼植保无人机下洗流场预测方面具有良好的应用效果,能够快速准确地预测下洗流场的速度信息。研究结果可为进一步研究旋翼风场对雾滴沉积分布特性的影响机理提供数据支撑。Plant protection drones can rotate at high speed in the process of droplet spraying.The downwash flow field can be generated by the rotors,leading to droplet drift.A rapid and accurate prediction of the velocity in the downwash flow field under the rotor can greatly contribute to improving the effectiveness of the UAV’s precision application.In this study,a prediction model was constructed in the downwash flow field of the single-rotor plant protection UAV using a physics-informed neural network.The prediction model effectively combined fluid dynamics and artificial intelligence(AI).A neural network model also incorporated into the physical equations.The powerful capabilities of the neural network were combined with the disciplinary context.Firstly,a physical model was used with the Lattice-Boltzmann to numerically simulate the flow field of the single-rotor plant protection UAV.The low-resolution flow field was then used to train the prediction model after numerical simulation.Secondly,the Navier-Stokes equations were embedded as the physics loss term in the prediction model,according to the fully connected neural network structure.The physics equations were utilized in the prediction model to learn the fluid flow patterns in the flow field.The interpretability of the model was enhanced to reduce the data dependence of the network model.Thirdly,the trainable parameters were updated iteratively to minimize the loss function during the training.The loss function was composed of both the physics and data loss terms.The training process was then realized to obtain the mapping relationship between physical quantities(such as velocity information)and space-time coordinates.As such,the mapping relationship was used to realize the fast prediction of the downwash flow field in the single-rotor plant protection UAV.Finally,the wind tunnel experiment was carried out to measure the velocity information of the flow field of the single-rotor plant protection drone under different side wind speed conditions.The accuracy and

关 键 词：无人机 模型 神经网络 下洗流场 单旋翼 

分 类 号：S252[农业科学—农业机械化工程]