Development Of The Surface Defect Detection System Of Carbon Fiber Stereoscopic Braid Based On Improved Faster RCNN

Carbon fiber composite material has high strength,small friction factor and other excellent properties.The carbon fiber composite material has a three-dimensional structure,which has the advantages of low manufacturing cost,excellent mechanical properties,and has been widely used in aerospace,architecture,energy,and medical and health care and many other fields.However,carbon fiber yarn will cause linting and fracture during the weaving process,resulting in unnecessary waste in the batch automated production of carbon fiber three-dimensional braid.Therefore,the identification and detection of surface defects in carbon fiber three-dimensional woven fabrics is a crucial part of the production process.This paper analyzes the current situation of carbon fiber braid defect detection at home and abroad,combines the implementation principle of traditional defect detection algorithm and the characteristics of carbon fiber three-dimensional braid surface defects,introduces a deep learning model to realize defect feature detection,and designs a deep learning-based defect detection system for carbon fiber braid defect characteristics and the actual detection needs of factories,so as to realize efficient and automatic detection of surface defects of carbon fiber threedimensional braid.Based on the actual production survey of the factory,this paper collects and arranges the defects produced by the factory carbon fiber stereoscopic weaving process,and analyzes all the original images statistically,and takes the three defects with the highest frequency in the actual production as the research objects.Combined with the characteristics of carbon fiber three-dimensional braid surface defects,factory production environment and other factors,the image recognition module and software control module of the defect detection system were designed,and suitable cameras,lenses,light sources and other hardware were selected to build a carbon fiber braid defect detection platform.Through image preprocessing and data enhancement technology,we expand the defective data set.This paper uses the label Img image annotation tool to complete the defective image annotation,and generates the PASCAL VOC2007 format data set for deep learning network training.Through the characteristics of carbon fiber stereo woven surface defect image,and the analysis of common flaw detection algorithm at home and abroad,this paper comprehensive selected in the field of small target detection effect better two-stage target detection network Faster RCNN model,as the basis of the carbon fiber woven flaw detection algorithm,and the network improvement and adaptation of Faster RCNN.The Res Net50 network based on residual fusion is used as the backbone feature extraction network to reduce the distortion of defective features and improve the multi-scale feature extraction ability of the network for multi-layer defect special maps.The improved nonmaximum suppression algorithm(Soft-NMS)is used to improve the original nonmaximum suppression operation,reduce the phenomenon of repeated detection frame error deletion of defect features,and optimize the problem of missing detection of defect detection.The RoI Align algorithm is used to improve the original region of interest pooling algorithm,which eliminates the regression box position bias caused by the two quantization rounding operations,and improves the accuracy of defect regression positioning.Finally,based on VS studio platform,pytorch framework,Open CV machine learning library,etc.,the C# programming language was used to complete the construction of the carbon fiber three-dimensional woven surface defect detection system,which realized the development of defect annotation,network training,defect recognition,data transmission and other functions.At the same time,by building a software,hardware and software environment that meets the experimental requirements,and testing the PASCAL VOC2007 format of carbon fiber braid flaw sample test set on the Py Torch framework,we verify that the proposed deep learning flaw detection method based on improved Faster RCNN has 92.7% accuracy in carbon fiber braid flaw detection.Moreover,the surface flaw detection method of carbon fiber stereoscopic braid based on deep learning does not need to design the corresponding feature extraction algorithm for the defect types,and has the advantages of good model generalization ability,high robustness,fast detection speed and high detection accuracy,which meets the needs of automatic detection of carbon fiber stereoscopic braid surface defects.