【作者】 张萍；

【导师】 苑进；

【作者基本信息】 山东农业大学 ， 农业工程， 2024， 博士

【摘要】 芦笋富含营养兼具抗肿瘤、抗氧化功效,被誉为“蔬菜之王”。我国芦笋种植面积为150万亩,占世界种植面积的40%,产量约占世界总产量的50%,是芦笋最大的生产和出口国,芦笋是我国加工出口贸易额最大的单一蔬菜品种。白芦笋一次栽植多年收获,笋芽从地下向垄面生长,采收季不断抽发成熟,笋尖冒出垄面时需及时采收,否则见光变紫,品质下降。目前国际上白芦笋采收仍以人工为主,工作量大、效率低,已面临无人采收的严峻形势,实现机械化采收对芦笋产业的可持续发展至关重要。人工采收先识别垄面成熟笋尖,再完成采收机构入土、切割、集箱等过程,白芦笋选择性采收仿生人工采收过程是目前公认的最佳收获方式,而且芦笋脆嫩多汁,易断易损伤,实现机械化智能化采收难度极大。本文以高效低损选择性采收为目标,构建了复杂垄面环境下两类目标笋芽的快速识别模型,设计了拉切式末端执行器,采用DEM-MBD双向耦合仿真方法分析了高速采收中末端执行器-土-笋互作的过程,研究了白芦笋采收“走-看-采”和双臂协同的智能控制方法,搭建了选择性采收机器人平台,开展了田间试验,主要研究内容如下:(1)提出YOLO5-Spear出土笋芽识别方法。针对出土笋尖目标小、笋尖与土块、枯叶等干扰物颜色、纹理、大小相似,田间多场景复杂垄面笋尖识别鲁棒性不足的问题,提出基于重采样的图像增强算法:提取采集的垄面图像中的笋尖斑块,形成多尺度组合图像,通过模拟田间作业场景进行图像增强变换;使用LC3、DWConv、SE模块改进了原模型的网络结构,构建了YOLO5-Spear笋尖识别模型;搭建了田间笋芽检测平台,开展了出土笋芽识别的田间测试试验。结果表明:YOLO5-Spear的识别准确率F1、AP分别为97.5%、97.8%,比YOLOv5提高0.93%、2.41%,模型参数、计算量和权重大小分别减少51.3%、33.7%、50.3%,检测时间仅为9 ms,提高了垄面笋尖检测精度和速度,满足笋尖识别精度和实时性要求;复杂田间测试环境下,出土笋芽的识别成功率达到85.5%。(2)提出HGCA-YOLO不可见笋芽识别方法。为了进一步解决因采收不及时笋尖变紫、品质下降的问题,针对垄面下笋尖顶土出现的土壤裂缝,提出不可见笋芽(裂缝形态)识别方法。建立了田间环境下不可见笋芽图像数据集,并依据其目标特点进行了数据增强;构建一种轻量级的不可见笋芽识别模型HGCA-YOLO,采用超参数进化方法确定基线网络,获取用于笋芽识别模型的最佳超参数;在基线网络中引入Ghost模块和坐标注意力机制,以降低模型的复杂性,提高对目标位置的敏感性;将测试时间增强方法引入到网络推理中,以增强变化环境中目标识别的鲁棒性;开展了不可见笋芽识别的田间测试试验。结果表明:HGCA-YOLO模型识别准确率m AP,m F1分别为95.2%,92.4%,模型大小为7.58 M,满足笋芽识别精度和轻量化要求;复杂田间测试环境下,不可见笋芽识别成功率达到了87.0%,为白芦笋采收机器人提供了支持。(3)设计了拉切式末端执行器,研究“土-笋-器”互作行为。针对土中笋芽高速低损采收难题,提出非原位的网捞采收模式,设计了刚柔耦合的拉切式末端执行器。针对末端执行器-土壤-芦笋的复杂互作行为,对采收关键子过程进行理论受力分析,建立了笋芽和土垄复合体与末端执行器模型,使用多体动力学和离散元双向耦合的方法描述土中网捞采收各个子过程(前行入土、土笋拉切、土笋出土和土笋抛出),从微观上分析了高速采收过程末端执行器的入土、切断、兜土、出土与土笋分离的动力学特性,明确了末端执行器构型变化以及土-笋-部件的互作行为;搭建了笋芽拉切采收土槽试验台,开展了末端执行器采收性能测试试验。结果表明:末端执行器入土最大水平和垂直阻力分别为445 N、530 N,拉切阻力为624 N,试验与模拟相对误差分别为5.3%、11.7%和15.2%;室内土槽试验的采收成功率为90%;入土过程0.8 s、拉切过程1.4 s、出土过程0.6 s、抛笋过程0.4 s,单次平均采收时间3.2 s;采后垄面扰动直径15cm、深度12 cm,扰动较小,验证了白芦笋采收末端执行器的有效性。(4)研究白芦笋选择性采收“走-看-采”和双臂协同控制方法。为了实现白芦笋“不停车”模式下高效、稳定的采收,针对采收区随机分布的多个笋芽,提出了以负载平衡为目标的双臂高效采收协同算法;研究了平台行走、笋芽识别、末端执行器采收“走-看-采”协同的控制策略,研发了基于ROS的多机械臂选择性采收智能控制系统,集成上述软硬件系统,搭建了白芦笋选择性采收机器人试验样机,开展了田间试验和性能评估。采收协同算法仿真分析表明,提出的三种路径规划方案中基于最短距离的先看先采(Shortest Time-First See Pick,ST-FSP)方案是最优的。双臂采收在平衡了工作负载的前提下,与单臂采收相比,时间节省了45.17%,采收成功率提高了3.19%;田间采收试验表明:实际笋芽识别成功率为82.6%,平均检测时间为0.033 s;成功识别笋芽的平均采收成功率为92.3%,机械臂平均运动时间为1.7 s,末端执行器采收时间为5.7 s;芦笋损伤率为7.2%,垄面破坏性较小,采收过程中各机构运行平稳,初步验证了白芦笋选择性采收机器人高效低损采收方案的可行性。更多还原

【Abstract】 Asparagus is rich in nutrients and has anti-tumor and antioxidant effects,and is known as the "king of vegetables".China’s asparagus planting area is 1.5 million mu,accounting for 40%of the world’s planting area,and asparagus production accounts for about 50% of the world’s total output.China is the largest producer and exporter of asparagus,and asparagus is also the largest single vegetable species in China’s export trade.White asparagus is planted once and harvested for many years and mature continuously during the harvesting season.The asparagus need to be harvested timely when they emerge from the ridge,otherwise they will become purple by sunlight irradiation and decrease in quality.At present,the harvesting of white asparagus internationally is still mainly artificial,with heavy workload and low efficiency,and has been faced with the severe situation of unmanned harvesting,mechanized harvesting is crucial for the sustainable development of asparagus industry.Artificial harvesting first identifies the mature spears on the ridge surface,and then complete the process of harvesting mechanism entering the soil,cutting,and collecting.The selective harvesting of white asparagus imitating artificial harvesting process is currently recognized as the best harvesting method.Moreover,asparagus is tender and juicy,easy to break and damage,and it is extremely difficult to realize mechanized intelligent harvesting.In this paper,with the goal of efficient and low-loss selective harvesting,a rapid identification model for two types of asparagus spears under the complex ridge environment was constructed.A pull-cutting end-effector was designed,and the DEM-MBD bi-directional coupling simulation method was adopted to analyze the operation process of end-effector-soilasparagus in high-speed harvesting.The intelligent control methods of “walking-lookingharvesting” and dual-arm coordination for selective harvesting of white asparagus were studied.A selective harvesting robot platform was built and field experiments were carried out.The main research contents are as follows.(1)The YOLO5-Spear method for the recognition of unearthed spear tips was proposed.Aiming at the problems of small target of spear tips,the similarity in color,texture and size between spear tips and disturbances such as soil clods and dead leaves,and insufficient robustness in the recognition of spear tips on multi-scenario complex ridge surfaces in the field,an image augmentation algorithm based on resampling was proposed.Extract the spear tip patches from the collected images of the ridge surfaces to form a multiscale combined image,and perform image augmentation transformations by simulating the field harvesting scenarios.LC3,DWConv,and SE modules were used to improve the network structure of the original model,and the YOLO5-Spear spear tip recognition model was constructed.A platform for spear tips detection in the field was built,and field experiments for the recognition of emergent spear tips were carried out.The results showed that the recognition accuracy F1 and AP of YOLO5-Spear were 97.5% and 97.8%,respectively,which were 0.93% and 2.41% higher than that of YOLOv5.The model parameters,computation amount and weight size were reduced by51.3%,33.7%,and 50.3%,respectively,and the detection time was only 9 ms.The model improved the detection accuracy and reduced the identification speed,which met the requirement of identification accuracy and real-time.Under the complex field test environment,the success rate of identifying the emergent spears reached 85.5%.(2)The HGCA-YOLO method for the recognition of invisible spears was proposed.In order to solve the problem of purple spear tips and quality degradation due to untimely harvesting,a method of recognizing invisible spears(leak morphology)was proposed for the soil leaks that appeared as spear tips grew upward against the soil under the ridge surface.The invisible spear image data set in the field environment was established,and data augmentation was carried out based on the characteristics of the targets.A lightweight invisible spear recognition model HGCA-YOLO was constructed,and the baseline network was determined by hyperparameter evolution method to obtain the best hyperparameters for the spear recognition model.Ghost module and coordinate attention mechanism were introduced into the baseline network to reduce the complexity of the model and improve the sensitivity to the target position.The test time augmentation method was introduced into the network inference to enhance the robustness of target recognition in changing environments.Field test experiment was carried out for the identification of invisible spears.The results showed that the recognition accuracy of m AP and m F1 of HGCA-YOLO model were 95.2% and 92.4% respectively,and the model size was 7.58 M,which met the requirements of spear recognition accuracy and lightweight.The success rate of invisible spears recognition reached 87.0% under the complex field test environment,providing support for the harvesting robot of white asparagus.(3)A pull-cutting end effector was designed and the interaction behavior of soil-asparagus-end-effector was studied.Aiming at the problem of high-speed and low-loss harvesting of asparagus in soil,an ex-situ net harvesting mode was proposed,and a rigid-flexible coupled pull-cutting end-effector was designed.Aiming at the complex interactions between soilasparagus-end-effector,the theoretical force analysis of the key sub-processes during harvesting was carried out,and the models of asparagus-soil complex and end-effector were established.The multi-body dynamics and discrete element bidirectional coupling method approach was used to describe the sub-processes of net harvesting in soil(move-penetrating,pull-cutting,lifting-out,and eject-throwing).The dynamics characteristics of entering the soil,cutting,pocketing the soil and asparagus,exiting from the soil and separating of soil and asparagus during the high-speed harvesting process were analyzed microscopically,and the configuration change of the end-effector as well as the interaction coupling mechanism of soilasparagus-end-effector were clarified.A test bench was built for asparagus pull-cutting and harvesting in the soil tank.The results showed that the maximum horizontal and vertical resistances of the end-effector were 445 N and 530 N during move-penetrating,respectively,and the pull-cutting resistance was 624 N.The relative errors between the test and the simulation were 5.3%,11.7%,and 15.2%,respectively.In the soil tank test,the harvest success rate was 90%,and the average single harvesting time was 3.2 s,of which,0.8 s for movepenetrating the soil,1.4s for pull-cutting,0.6 s for lifting-out the soil,and 0.4 s for ejectthrowing the asparagus.The disturbance of the ridge surface after harvesting was 15 cm in diameter and 12 cm in depth,and the disturbance was small,which verified the effectiveness of the end-effector for white asparagus harvesting.(4)The “walking-looking-harvesting” and dual-arm cooperative control methods for selective harvesting of white asparagus were studied.In order to realize efficient and stable harvesting of white asparagus under the “non-stop” mode,for multiple spears randomly distributed in the harvesting area,a dual-arm efficient harvesting collaborative algorithm with the objective of load balance was proposed.The “walking-looking-harvesting” cooperative control strategy of platform movement,asparagus spear identification and end-effector harvesting was studied.The dual robotic arms selective harvesting intelligent control system based on ROS was developed.The experimental prototype of a selective harvesting robot for white asparagus was built by integrating the above software and hardware systems,and field tests and performance evaluation were carried out.The simulation analysis of the harvesting cooperative algorithm showed that the shortest distance based first-see-first-pick(ST-FSP)scheme was optimal among the three proposed path planning schemes.Under the premise of balancing the working load,the time of double-arm harvesting was saved by 45.17% and the success rate of harvesting was increased by 3.19% compared with single-arm harvesting.Field harvesting experiment showed that the success rate of actual spear recognition was 82.6% and the average detection time was 0.033 s.The average harvesting success rate of successfully recognized spears was 92.3%,the average movement time of robotic arm was 1.7 s,and the harvesting time of end-effector was 5.7 s.The damage rate of the asparagus was 7.2%,the ridge surface was less destructive,and the mechanisms operated smoothly during harvesting,which initially verified the feasibility of the selective harvesting robot for white asparagus with high efficiency and low loss.更多还原