Research On Machine Vision-based Surface Defect Detection Technology For Rare Earth Metal Ingot 1

At present,rare earth elements are widely used in electronic equipment,automobiles,aerospace,new energy and other fields of high-tech products,leading to a sharp rise in the demand for rare earth metal materials,in which rare earth metal ingots play a crucial role.Due to the existence of slag inclusion defects of molten salt and other impurities on the surface of finished metal ingot products affected by the current process,it is necessary to identify and remove them in time before storage and transportation according to the process requirements.In order to improve production efficiency,eliminate manual intervention and stabilize product quality,machine learning was used to realize automatic online identification and cleaning of metal ingot surface defects.The defect image is extracted and processed by using camera nonlinear model optimization,attention mechanism neural network and other methods,and combined with classifier integration and classification classifier and other practical technologies to achieve the classification and processing of defects.Finally,the surface impurity defects are removed by feedback control servo control system,high-speed milling and other auxiliary devices.The aim of this research is to improve the accuracy and efficiency of surface defect detection of rare earth metal ingot,so as to improve product quality and production efficiency.The main research contents of this paper include:(1)To investigate the production characteristics of heavy rare earth electrolytic reduction process in different Chinese regional enterprises,collect conventional information about surface defects of finished products produced by various models of pouring molds under current leading process conditions,and obtain original image sets through data analysis and image processing techniques;The linear model analysis technology of the camera was studied,and the radial constraint algorithm and plane calibration algorithm were combined to realize the logical reasoning of the solution process of the internal and external parameters of the camera,and the information of the number,dimension characteristics and azimuth coordinates of the defects on the surface of the test object was prelim natively determined.The nonlinear camera model L-M(Levenberg-Marquardt)optimization algorithm was used to optimize the calibration parameters of the above conclusions,so as to improve the accuracy.(2)Combined with deep learning neural network and integrated application key technologies based on encoder and decoder architecture,a surface defect detection model based on attention mechanism neural network was established.Combined with the principle and practical technology of machine vision surface defect detection,a higher level feature extraction model of surface defects based on attention was established based on the actual collected image data.With the help of space and channel attention mechanism,the attention mechanism can integrate high and low level feature space to realize the recovery of defect details.By completing the segmentation of defects,it can realize the effective and real-time processing,so that the system can reach the expected computing speed and image recognition ability.(3)This paper summarizes the current popular practical technologies such as classifier integration and hierarchical classifier,and designs a surface defect detection multiclassification combinator based on SLIQ(Supervised Learning In Quest)decision tree by Boosting algorithm.The adaptive weighting technique and weighted voting technique are used to classify surface defects of rare earth metal ingot.On the basis of accurately obtaining the information about the quantity,size,type and orientation of slag inclusion defects on the surface of metal ingot,the image processing system is used to control the corresponding servo control system,high-speed milling and other auxiliary devices to automatically remove the impurity defects on the surface of rare earth metal ingot.The results of the experimental prototype of this research on the finished praseodymium alloy at the production site show that the rate of missed and false detection of surface defects on 8kg class trapezoidal metal ingots is about 2.8%;the average error of dimensional detection of defect area,shape characteristics,orientation coordinates,etc.is about ±1.6%,which meets the expected research objectives and enterprise use requirements.