Surface Defect Detection And Classification Of Metal Shaft Based On Deep Learning

The shaft is a particularly important part of the machinery industry.As a widely used metal,almost all compositions of mechanical equipments are inseparable from it.The processing quality and the precision of shaft directly affect the equipment performance and life span of the machinery.In the course of processing,the surface of shaft may defect due to machining errors,material or other reasons.In order to reduce the cost,it is necessary for The manufacturer of shaft to carry out different recovery processing for different defects,Therefore,there are important significance and engineering application value for shaft defect detection and classification.However,traditional defect detection methods have the disadvantages of low recognition accuracy and lack of generalization ability.The thesis of detection and classification system is based on the deep learning can achieve high-efficiency recognition of metal shaft surface defects and achieve high-precision classification and positioning.The concrete research contents and achievements are as followed:(1)The shaft defect processing system software is designed according to the structural features of axis.We use industrial CCD cameras to obtain axial images.then,we design shaft defects data acquisition and preprocessing algorithm including axis defect image graying module,denoising module,edge detection module,defect image data enhancement module,multi-thread processing module,etc,based on opencv.(2)We analyzed the traditional target detection algorithm and deep learning target detection method.Finally,We chose the Faster R-CNN as the deep learning target detection model.We changed the structure of traditional Faster R-CNN and compared the performance of models with different structures through experiences,and we chose the right training method and parameter settings.(3)We designed and built a visual inspection platform and designed the structure and layout of the metal shaft surface defect detection platform;The underlying control system of inspection platform is based on the STM32 single-chip,including the design of the metal shaft rotation module,lighting module,communication module circuit and control method.And the upper computer interface is designed based on QT,the upper computer sending instructions through the serial port communication to the lower computer for control.(4)We made a metal shaft surface defect data set,which contains a total of 40,000 image data including 10,000 of steel pits,10,000 of scratches,10,000 of abrasions,10,000 of breaches,and it is divided according to the training set,verification set and test set;Faster R-CNN target detection model was build based on the Tensorflow deep learning framework and through the acquired data set used to train and calibrate the model.we compared the recognition ability and detection speed of the different structures of Faster R-CNN and analyzes the error source to verify the correctness,effectiveness and reliability of the shaft classification system.This dissertation designs a detection system of the defect of metal shaft surface based on the deep learning which can effectively locate and identify various types of shaft defects in this project.The final accuracy rate can reach 94%~96%,and the recognition speed can reach less than 5s,compared with the traditional detection method,the recognition accuracy and generalization ability are greatly improved,but the speed should be further improved in the future by the improvement of performance of the software and hardware.