Deep Learning-based Segmentation And Grading Of Metallographic Organization Of Carburized Gears

High-speed railway gears are in a long-term high-temperature and high-speed working environment,and the thrust-to-weight ratio and power ratio requirements are constantly increasing,so high-speed railway gears must be subjected to complex stresses,strains,contact fatigue,bending fatigue and wear.The carburising process can substantially improve the surface properties of the gears,giving them a high surface hardness and wear resistance,and making the heart of the gears have a high strength and appropriate toughness.Therefore,the high speed railway gear transmission system mainly adopts high strength aluminium alloy gear housing and high precision carburised and quenched gears to ensure the smoothness and reliability of the gear transmission.In the gear quality and safety assessment,the determination of internal microstructure is one of the important criteria to test the quality of gears.For carburised gears,the identification and assessment of residual austenite and martensite levels directly affects the safety assessment results.At present,from the collection of metallographic images of carburised gears to the comparison of the final experimental results is mainly done manually,which has the disadvantages of low efficiency and subjectivity.Therefore,this paper takes carburised gears as the research object and completes the tissue segmentation and automatic grading of residual austenite and martensite based on a deep learning framework.(1)In order to solve the problem of mixed noise in images caused by external interference,the research is carried out from two aspects: traditional filtering noise reduction and deep learning based Dn CNNs noise reduction.Experiments show that the evaluation index and image quality of Dn CNNs are better than those of traditional filtering noise reduction,and its PSNR is 0.01 to 0.03 higher than that of the filtering algorithm,and the average SSIM value reaches 0.9486,which more accurately recovers the detail information of the original image.For the characteristics of small sample size and difficult acquisition of metallographic data,the number of samples is augmented by data enhancement methods,such as adjusting contrast,brightness or flip,so as to build a carburized gear metallographic database and lay the foundation for subsequent network model training.(2)In order to realise the metallographic tissue segmentation of carburised gears,the image is firstly converted to grey scale,and the research is carried out from thresholding-based,edge detection and region image segmentation algorithms,and the Otsu algorithm and Niblack algorithm are used to segment the image according to the grey scale distribution characteristics of the carburised gear metallographic map.Due to the different features and properties within the metallographic tissue,Roberts,Prewitt,Sobel,Canny and LOG operators were used for the experiments.Finally,the region splitting and merging algorithm and the watershed algorithm were proposed based on the inter-tissue region properties.The overall experimental results show that the Otsu algorithm and the Canny algorithm perform well,but are prone to problems such as over-and under-segmentation for complex feature images.(3)In order to achieve accurate segmentation of carburised gear metallographic tissues,EISeg was first used for manual image annotation,and the CBAM-MV3 UNet network architecture was proposed,which was improved on the basis of the U-Net architecture,and the bneck module in Mobile Net V3-Large was selected to replace the U-Net decoding part to achieve deep-level semantic feature extraction,and the The extracted features are fused with the up-sampled features,and the CBAM attention mechanism is added to this part to discriminate between the focused and suppressed regions,and the loss function is improved and the parameters are tuned by a weighted mixture method.The experiments showed that the average values of MPA and MIOU of the improved model reached 88.19% and 84.71%respectively.(4)In order to achieve automatic grading of metallographic tissues of carburised gears,a metallographic grade recognition model based on the attention mechanism and Res Net50 network is proposed,and the weight parameters that have been pre-trained twice are put into the network model to solve the problem of poor recognition accuracy of small sample size.The problem of model overfitting during training is solved by improving the entropy function to improve the model accuracy,and the improved model is compared with Alex Net,Goog Le Net and Res Net50 models.The experimental results show that,compared with the mainstream network models,the proposed model is more accurate for the identification of residual austenite and martensite,and the classification recognition rate reaches 94.8% and 93.63%,while the recognition time of a single metallographic map is measured to be about 4.5ms,achieving the effective identification of residual austenite and martensite grades.