Research On Hot Runner Parts Recognition And Grasping Based On Deep Learnin

With the continuous development of computer and automation technology,the technology related to robotic arm is improving and the application field is becoming more and more extensive,and also deeper into production practice.This also brings a series of challenges to the performance of robotic arms,which are not only required to adapt to complex working environments,but also need to have the ability of perception,decision making and learning.Among these abilities,target recognition and grasping are the two most important and basic ones.In this paper,we focus on the problem of low efficiency of part sorting in hot runner production shop,and establish a deep learning based target detection model and robotic arm grasping pose estimation model to improve the accuracy of part recognition and the success rate of robotic arm grasping,and then improve the speed of part sorting.The main contents and innovations of this paper are as follows:(1)Build the recognition and grasping system of hot runner parts.Firstly,the overall design of the robotic arm is completed according to the actual workshop environment of hot runner parts and the structural characteristics of the parts;then the reasons for using Kinect v2 depth camera as the image acquisition device of this topic are introduced,the mutual transformation relationship between pixel coordinate system,physical coordinate system,camera coordinate system and world coordinate system is explained,and the preliminary preparation work such as camera calibration and robotic arm hand-eye calibration is completed;Finally,the hardware configuration and software environment of the upper computer in this topic are introduced.(2)Verify the performance of the network model after adding three attention mechanisms.Firstly,the basic principle of convolutional neural network and the overall framework of YOLOv8 algorithm are introduced to train the part recognition network model based on YOLOv8 algorithm;then the network structure of the model is changed by adding channel domain attention mechanism SENet,spatial domain attention mechanism GENet and hybrid domain attention mechanism CBAM to YOLOv8 to try to learn the input The final experimental validation shows that the recognition accuracy of SE-YOLOv8 n model is improved by 2.3%,GE-YOLOv8 n model is improved by 1.7%,and CBAM-YOLOv8 n model is improved by 3.5%.(3)Introducing a bounding box loss function based on a dynamic non-monotonic focusing mechanism.the CIo U Loss used by YOLOv8 in the prediction box regression process,when the width and height of the prediction box are linearly proportional to the real box width and height,the width and height of the prediction box cannot increase or decrease at the same time.To address this problem,this paper introduces a bounding box loss function based on a dynamic non-monotonic focus mechanism,and makes a balance between samples with better and worse regression quality by defining an outlier to represent the quality of the prediction box,fully exploiting the potential of the nonmonotonic focus mechanism,and obtaining a network model with a 4.2% improvement in recognition accuracy.When used in combination with the attention mechanism,the recognition accuracy of the model is improved by up to 6.8%.(4)Generate the grasping poses of robotic arm based on GGCNN2.Firstly,the network structure of GGCNN2 is introduced and the improvement measures of GGCNN2 are explained;secondly,two public datasets,Jacquard and Cornell,are introduced and the grasping dataset of the parts in this paper is built;finally,the GGCNN and GGCNN2 algorithms are used to train and test on different datasets,and the network model with the highest grasping success rate is obtained and then grasped in a real environment.The experiments are conducted and the results are analyzed.