【作者】 李宁；

【导师】 刘莫尘； 王东岳；

【作者基本信息】 山东农业大学 ， 机械（专业学位）， 2022， 硕士

【摘要】 中国是全球第一大苹果生产国,直到目前为止,苹果的选择性收获作业都依靠人工完成,随着劳动力的减少和苹果产业的升级,采摘作业由机器人代替成为了未来发展趋势。本文以苹果为采摘对象,根据苹果的物理特性,着力设计一种结构和控制简单,负载力强,抓取灵活,包络性好的苹果采摘末端执行器并对成熟期苹果的识别进行研究。论文主要完成以下工作:(1)分析了国内外末端执行器的研究现状和传统人工采摘动作模式,在此基础上,提出了现有末端执行器结构设计上存在的不足。以富士苹果为研究对象,对其物理特性与力学特性进行了分析,为末端执行器总体方案的设计提供数据支持。根据物理特性确定了抓取苹果的范围为70～100 mm。根据苹果的力学特性,设定压力传感器的触发值为50 N。(2)基于仿生学原理和欠驱动原理设计了末端执行器的机械结构,结合苹果的物理特性,从实用化的角度对末端执行器进行了具体的结构设计,为了提高手指的自适应能力,手指采用弧形结构,内覆柔性材料,设计弧形手指的弧口直径为100 mm,近指选取弧长圆心角为60°,远指选取弧长圆心角约为33.4°,手指宽度为15 mm。通过在手指添加压力传感器来提高末端执行器的主动感知能力,在采摘过程中能够自行控制抓取苹果的力度。(3)基于Solidworks软件建立末端执行器的三维模型,在ADAMS中建立虚拟样机模拟不同尺寸苹果的实际抓取效果,通过仿真分析评价了苹果与手指不同指节的接触力变化和抓取过程中手指的位移和速度变化,验证了欠驱动弧形机械手指的抓取自适应能力与稳定性。运用ANSYS对传动机构和手指进行有限元仿真分析,得到末端执行器在抓取时最大应力为54.8 MPa,最大变形为0.0086 mm,符合设计要求。利用3D打印技术制作末端执行器样机,测试其合理性和抓取性能,验证表明末端执行器具有良好的实际效果,达到设计目标。(4)本文采集自然环境下苹果的图像建立数据集,并使用旋转、改变亮度和锐度、添加噪声等方法扩充苹果数据集。使用YOLOv5s网络进行苹果识别,并通过添加CBAM注意力机制来提高识别精度,训练后模型的模型比基准模型准确率提高了3.4个百分点,召回率提高了5个百分点,平均均值可以达到95.1%,具有较好的准确性。并通过使用Real Sense D435相机,基于pytoch实现改进型YOLOv5s目标检测,实现返回检测目标相机坐标下的位置信息。更多还原

【Abstract】 China is the world’s largest apple producer,and until now,the selective harvesting of apples has been done by human hands.As the labor force shrinks and the apple industry upgrades,the harvesting of apples will be replaced by robots.In this paper,apple as the picking object,according to the physical characteristics of apple,focus on the design of a simple structure and control,strong load,flexible grasp,good envelope apple picking end-effector and the identification of the mature apple.The thesis mainly completes the following work:(1)The research status of end-effector at home and abroad and the traditional manual picking action mode were analyzed.On this basis,the deficiencies in the structural design of the existing end-effector were proposed.Taking Fuji Apple as the research object,the physical and mechanical properties of Fuji Apple were analyzed,providing data support for the design of the end-effector.According to the physical characteristics,the range of apple capture is determined to be 70 ～100 mm.According to the mechanical properties of apple,the trigger value of the pressure sensor is set as 50 N.(2)based on bionics principle and principle of underactuated mechanical structure design of the end executor,combined with the physical properties of apple,from the Angle of practical end executor has carried on the concrete structure design,in order to improve the adaptive ability of the fingers,fingers using arc structure,with flexible material,the design of curved fingers arc mouth diameter of 100 mm,The central Angle of the near finger is 60°,the central Angle of the far finger is 33.4°,and the finger width is 15 mm.By adding a pressure sensor to the finger,the end-effector can improve its active sensing ability and control the strength of the apple during the picking process.(3)The 3D model of end-effector was established based on Solidworks software,and virtual prototype was established in ADAMS to simulate the actual grasping effect of apples of different sizes.Through simulation analysis,the contact force changes between apples and fingers at different knuckles and finger displacement and speed changes during grasping were evaluated.The grasping adaptive ability and stability of underactuated arc mechanical finger are verified.ANSYS was used to carry out finite element simulation analysis on the transmission mechanism and finger,and the maximum stress of the end-effector was 54.8 MPa and the maximum deformation was 0.0086 mm,which met the design requirements.The endeffector prototype was fabricated by 3D printing technology,and its rationality and grasping performance were tested.The verification shows that the end-effector has good practical effect and achieves the design goal.(4)In this paper,apple images in the natural environment were collected to establish a data set,and the apple data set was expanded by rotation,changing brightness and sharpness,adding noise and other methods.The YOLOv5 s network was used for apple recognition,and the CBAM attention module was added to improve the recognition accuracy.Co MPared with the benchmark model,the accuracy of the trained model increased by 3.4 percentage points,the recall rate increased by 5 percentage points,and the average value could reach 95.1%,showing good accuracy.By using the Real Sense D435 camera,the improved YOLOv5 s target detection based on Pyto CH is realized,and the position information under the camera coordinates of the detected target is returned.更多还原