Research On Bearing Defect Detection Technology Based On Machine Vision

Bearing surface defects will seriously affect the quality,practicability,aesthetics and safety of the product.It is very important to automatically identify substandard products in actual production.In an industrial environment,bearing surface defects have the characteristics of random change and lack of fixed shapes.The detection accuracy of artificial recognition and visual methods based on traditional image processing is not high,and the massive image data and label data required for deep learning are in this scene It’s hard to get it.Based on the idea of deep learning,this thesis takes bearing surface defects as the starting point,and quickly obtains training images and labels by synthesizing image data,and uses convolutional neural networks to carry out detection research on bearing surface defects.The main research content of this thesis includes the following directions:(1)A bearing surface defect detection system is designed.Starting from scratches and deformations,analyzed the three-dimensional characteristics of different defects and the corresponding imaging principles,using uneven lighting to highlight the characteristics of deformation defects,and uniform lighting to highlight scratches and text Features lay the foundation for subsequent detection algorithms.(2)A method of scratch detection and text recognition based on deep learning is proposed.A composite data set method based on foreground and background fusion is used to generate a composite image consistent with the real picture style for training.The ROI region conversion module is proposed to combine the detection and recognition modules.In the detection process,the parallel high-resolution network is integrated into the defect detection network to solve the problem of large defect size span.The comparison of experimental results shows the superiority of the method.(3)A method for detecting deformation defects of bearing surfaces based on the category-level adversaries model is proposed.A large number of synthetic images and label data are generated through 3D modeling software for neural network to learn characteristic differences between different datasets.Through synthetic image data sets and part of the real image data sets on joint training,enhance the generalization ability of the model.The results of the example show the effectiveness of the category-level adversaries model.(4)The software system of bearing surface defect detection is designed and implemented.Firstly,the overall design framework and the overall planning of function modules of the bearing defect detection software system are proposed.Then,the database design and model forward propagation acceleration in the system development are introduced respectively.Finally,the function and test results of the bearing defect detection software system are presented to verify its effectiveness and practicability.