Research On 3D Point Cloud Segmentation Algorithm For Casting Riser

Casting is widely used in the manufacture of rough mechanical parts,and its production process requires a large number of manual participation.The foundry industry is a labor-intensive industry with high working intensity and poor working environment.China is in urgent need of changing from "foundry power" to "foundry power".How to use robot to realize automatic casting production has become a research hotspot.At present,the casting robot visual recognition methods are mainly cloud segmentation recognition,image segmentation recognition,3D reconstruction recognition,laser sensor recognition and other recognition methods.Point cloud deep learning segmentation and recognition has obvious advantages in casting robot visual recognition due to its features of high recognition accuracy,learning ability,rich data information and no shielding.In view of the existing problems in the process of deep learning point cloud segmentation and recognition,such as large demand for training data,low identification accuracy of sparse geometric features,disorder of point cloud,and large amount of data,a casting blank data set for deep learning training was made in this paper,and a casting blank point cloud segmentation network model FLA-NET was proposed.The network model is optimized according to the characteristics of casting blank point cloud data.It is sensitive to the recognition of geometric features of casting blank casting riser and other target objects,and has the invariance of spatial transformation,which can accurately and efficiently segment and recognize casting blank point cloud data.The main research contents of this paper are as follows:(1)Set up the point cloud data acquisition and processing platform,complete the collection of casting blank point cloud data and the design and production of the data set in the foundry workshop,and develop the semantic annotation plug-in of point cloud data based on the secondary development of Rhinoceros 3D software to meet the production requirements of casting blank data set.(2)Through casting riser design method,the formation law of geometric characteristics of casting blank after casting riser is formed is theoretically analyzed.A coding algorithm based on cosine similarity for casting riser feature segmentation and recognition was proposed.By integrating neighborhood graph construction and feature coding algorithm,the ACSE local feature coding module was formed,which was combined with the frontier point cloud segmentation network,and the initial version Rand LA-NET(ACSE)was proposed.In this paper,the segmentation and recognition results of Rand LA-NET(ACSE)in large scale public dataset of field scenic spot cloud are studied,which foreshadows the further optimization of network model structure.(3)In order to deal with the disorder of point cloud,the F-CONV local spatial feature pre-coding layer is proposed to endue the spatial transformation invariance of the point cloud segmentation network model.,The furthest point sampling algorithm is used to subsample the features extracted from the network model because the amout of the point cloud data is too huge for the network.The attention pooling module and the residual connection module are introduced,and the complete casting casting riser segmentation network model FLA-NET is proposed,which lays a foundation for further experiments(4)The casting blank part segmentation and identification experiment was carried out,and the influence of the number of point cloud division in the learning interval on the accuracy was studied.Through parameter optimization,under the experimental conditions that the number of point clouds divided by learning interval is 12288,the learning rate is0.001,and 250 Epochs are studied,the accuracy rate of gate identification reaches 98.1% of the test set,and the overall average accuracy rate of the test set reaches 93.3%.The feasibility and effectiveness of FLA-NET segmentation and identification of casting blank point cloud parts were verified by analyzing the different types of parts segmentation and the whole experimental results.The cloud segmentation experiment of multi-casting blank field was carried out,and the ability of FLA-NET to deal with large scale point cloud was studied.Under the condition of optimal parameters,the overall segmentation accuracy of test set reached 77.2%.Further comparison verifies the advantages of FLA-NET for casting point cloud data segmentation and recognition and the application prospect of casting robot multi-task.