Research On Metallographic Analysis And Quality Evaluation Technology Based On Convolutional Neural Network

GCr15 bearing steel is one of the widely used high carbon chromium bearing steels around the world.It has important application value in industry,military,aerospace,agriculture,and other fields.The study of its micro-structure has important social and engineering significance for the macro-performance analysis of GCr15 bearing steel.Aiming at the shortcomings of traditional quantitative metallographic analysis methods,such as low efficiency and poor accuracy,this paper qualitatively and quantitatively analyses the micro-metallographic structure of GCr15 bearing steel obtained by high temperature quenching,austenitizing quenching and tempering heat treatment based on machine vision and digital image processing,which overcomes the problem of low efficiency of traditional semi-quantitative method.It improves the accuracy of data results and the way of quantitative analysis.The main contents and the results of this paper are as follows,(1)GCr15 specimens were prepared from tapered roller bearing steel under different heat treatment conditions.The metallographic structure was observed by scanning electron microscopy(SEM),and the micro structure image was collected.The composition of GCr15 specimens was analyzed by X-ray energy dispersive spectrometer.(2)An improved image denoising algorithm based on matrix low rank sparse decomposition is proposed.Block regularization is added to the original LRMR model to denoise the metallographic image of GCr15 bearing steel.It effectively removes the noise in the collected metallographic image and improves the texture distortion and edge blurring in the traditional denoising methods.(3)By comparing the widely used segmentation algorithms and models,an improved U-Net convolution neural network model is introdused to segment carbides in the metallographic image of GCr15 bearing steel,aiming at the problems of low contrast of metallographic image and unclear edge contour of each structure,and combining with the deep learning neural network model,the carbides in the metallographic image of GCr15 bearing steel after pretreatment are segmented,and the results are obtained good segmentation effect.The average DICE coefficient is 0.884.(4)Based on segmented image analysis,the new features and distribution of carbides are analyzed,the area proportion of carbides under quenching and tempering heat treatment is calculated,and the roundness and size distribution of carbide particles are counted.The results show that the Quantitative analysis of metallographic structure based on image processing and deep learning neural network can improve the traditional quantitative analysis methods and improve the accuracy of analysis results;The average area of carbide under quenching and tempering heat treatment is about 7.9%,the average roundness is 1.01,and the diameter of particles is mostly between 0.2 and 0.5 ?m.