| In soybean fields,chemical weeding is the main method of plant protection operation.After long-term use,a large number of malignant weeds have developed obvious resistance to herbicides.Moreover,these weeds are scattered in distribution,large in volume and complex in root system.Secondary weeding operations still need to be carried out manually,with high labor intensity,high labor cost and low weeding efficiency.In response to the above problems and in response to the national policies of “14th Five-Year Plan” for comprehensive green transformation of agriculture and pesticide reduction and efficiency enhancement,a precision weeding device based on Delta parallel mechanical arm was developed by combining target detection technology based on deep learning algorithm.Through the design and selection of key components,simulation motion analysis,structural optimization of genetic optimization algorithm,training and optimization of recognition model,indoor and outdoor experiments and other methods,the designed weeding device can achieve optimal performance.The main contents and conclusions of the research are as follows:(1)The overall plan for the precision weeding device has been determined.The soybean field environment was investigated and field research was conducted,and the types of weeds before and after spraying were compared.It was determined that weeds with developed root systems and large growth volumes would be the target of precision weeding.Based on the principle of traumatic weeding,a weeding plan was determined to first cut weeds with a rotating knife and then spray herbicides on the stem and leaf wounds.Finally,the overall structure and control plan of the precision weeding device were determined.(2)Optimization design of key components of precision weeding device.According to the design index of parallel arm and the overall design scheme,the structural parameters of the traumatic weeding end effector and Delta parallel arm are optimized.Shear force tests are carried out on target weeds to determine the length,thickness,blade angle and motor parameters of the weeding knife;theoretical analysis methods are used to determine the model of the spray nozzle;geometric methods are used to derive the forward and inverse kinematics of the parallel arm,and a virtual prototype is established in ADAMS software to verify the accuracy of its kinematic model;genetic optimization algorithm is used to optimize the structural parameters of the parallel arm to ensure that the weeding space of the parallel arm meets the weeding operation requirements in soybean fields.(3)Establishment of weed recognition and positioning model based on deep learning.After spraying in the seedling stage,field weed images were collected,with a total of 3500 effective images collected and processed for image annotation and dataset production.The YOLO V5 network model was selected for training,and the highest accuracy,recall rate and m AP value of the recognition model were 86.4%,81.9% and 88.6%,respectively,with good recognition performance.The Intel Realsense depth camera was used to locate the recognized weeds,and the coordinate transformation of weed coordinates from camera coordinate system and image coordinate system to world coordinate system was analyzed to complete the real-time acquisition of weed recognition and three-dimensional coordinates.(4)A prototype of the precision weeding device was built and indoor and field tests were carried out.In the indoor environment,the recognition rate of the recognition system at different speeds,the accuracy of weed positioning and the motion accuracy of the parallel arm were tested.The test results showed that the recognition rate of soybean and weed could reach more than 98% at a speed of 0.1-0.6 m/s;the accuracy of weed coordinate positioning based on depth camera reached 94%;the average motion accuracy of the weeding mechanical arm was ±2 mm.After correcting the positioning error and motion error,field weeding performance tests were carried out.The test results showed that at an operating speed of 0.1-0.4 m/s,the average weeding rate was 89.7%,the average seedling injury rate was 2.5%,the average recognition rate of soybean was 92%,and the average recognition rate of weed was 86%. |
PDF Full Text Request