Research On Casting Flash Recognition Algorithm Based On Point Cloud Semantic Segmentation

Today,the sand casting process plays a major role in the manufacture of mechanical components,and similar products can be mass-produced using casting templates.However,due to the positioning accuracy and deformation of the flask assembly during the casting process.There is a flash around the casting,which needs to be processed and polished to remove the flash.Due to the randomness of the position,shape and thickness of the flash edge,it is difficult to locate it effectively.In the existing robot grinding process,it is necessary to align the machining datum of the tool and the part in advance,and use the offline generated machining path to grind the part.Aiming at the problems existing in the automatic casting flash grinding system,such as long offline programming time for medium and small batch casting,high fixture positioning requirements,low efficiency caused by fixed processing paths,and difficult flash positioning.In this thesis a 3D vision-based casting flash recognition technology is proposed.Three-dimensional equipment is used to collect the casting flash point cloud.By constructing a point cloud semantic segmentation network,the points of the flash and the casting body can be segmented.The casting point cloud with semantic information can provide a strong support for the planning of the subsequent processing and grinding trajectory.In order to make up for the defect of fewer samples in the network training process,this thesis uses the adversarial generative network technology to generate flash data to enhance the richness of the training data.The main research contents of this thesis are as follows:（1）Based on graph attention convolution,an algorithm MSF-Net（Max pooling Sampling and feature Fusion Network）is constructed.Because farthest point downsampling algorithm lacks local structure adaptation,MSF-Net proposes a Maximizing activation based Farthest Point Sampling algorithm,MFPS.Because the graph attention convolution is only performed in one dimension,the permutation transformation is fused,and the two-dimensional convolution is integrated into the algorithm model.Aiming at the shortcoming of the lack of global information in upsampling using local linear interpolation,an upsampling algorithm integrating global and local is introduced to expand the field of view in the feature extraction process.By constructing 744 virtual simulation data samples of casting flash segmentation,including 31 different parts,the effects of down-sampling,feature fusion and up-sampling algorithm improvement were explored.The segmentation and intersection ratio of MSF-Net in the test set reaches 95.23%,which is 1.97% higher than the benchmark network,and the improvement effect is obvious.By visualizing the MFPS,verifing that the algorithm can perform adaptive sampling according to the local structure of the casting point cloud.（2）Based on generative adversarial network,a flash point generation algorithm FPG-Net（Flash Point Generation Network）is constructed.Because the aggregated multilayer perceptron lacks local feature extraction,a local aggregated multilayer perceptron is proposed to enhance the local feature extraction ability.Aiming at the difference between point cloud filling and flash point cloud generation tasks,a flash shift algorithm is proposed,and point cloud filling is applied to the flash generation task.In simulation experiment,the chamfering distance between the generated flash edge and the real flash edge is only 0.01735,which is81.81% lower than that of the benchmark network.The resulting casting flash can be used to augment the dataset,enhancing the diversity of deep learning training samples.The prediction speed of MSF-Net is 587 milliseconds,which meets the requirements of robot grinding.（3）Carry out the experiment and exploration of casting flash identification.Using a 3D scanner,a casting point cloud acquisition system was built,and 252 solid data samples were constructed,including 12 different parts.Point cloud segmentation experiments were carried out with global perspective,local perspective and noise points.The MSF-Net algorithm achieves 87.80%,90.37% and 87.25% of the segmentation and intersection ratios in three experimental tasks respectively,which verifies the practicability of MSF-Net in different casting flash recognition tasks.Through the fusion of simulation data and flash generated data,and the introduction of transfer learning technology.In the global perspective segmentation experiment,the segmentation and intersection ratio of MSF-Net has been improved to 90.51%,which verifies the effectiveness of FPG-Net generated samples.