| With the rapid development of new generation information technologies such as big data and artificial intelligence,the manufacturing industry is accelerating towards automation and intelligence.At present,the quality control of injection-molded workpieces is still dominated by manual inspection,which has problems such as poor stability and low inspection efficiency.Some enterprises rely on automatic detection methods developed by machine vision technology,relying on subjective experience to extract features,and there are also shortcomings such as poor robustness.Detection methods based on deep learning have the advantages of automatic feature extraction,good robustness,and strong adaptability,and are becoming a new research hotspot in the field of industrial automation quality detection.This thesis focuses on the quality inspection of the injection molded parts of the air conditioner inner plate as the research goal,combined with deep learning technology,studies the surface defect detection algorithm for injection molded parts,and implements a surface defect detection prototype system.The main work and achievements are as follows:(1)Construct a dataset for surface defect detection of injection-molded workpiecesIn this thesis,a dataset of surface defects of injection-molded workpieces for airconditioning inner panel is constructed.First,select the main optical equipment of the image acquisition system and design the image sample acquisition scheme of the industrial site.The constructed injection workpiece dataset contains 5 types of defects: black lines,black spots,gate lines,pits,and scratches.Then,the data is enhanced by digital image processing technology and made into VOC format data set,so as to construct surface defect training and test data sets.These common defect types can fully reflect the real industrial production environment.(2)Research and design a defect detection algorithm based on improved YOLOv4The algorithm uses YOLOv4 as the basic network architecture and uses the lightweight Mobile Net V2 as the backbone network for feature extraction.According to the characteristics of the injection molding workpiece dataset,the K-means++ clustering algorithm is used to recluster the anchor frame size;For defect types with extreme aspect ratios,the EIo U loss function is introduced to separately consider the aspect ratio to solve the problem that the loss function is less sensitive to changes in the long side;then the attention mechanism is introduced to improve the detection rate of small target defects;finally,the detection model is considered It will occupy a lot of memory and computing resources during deployment,and use depthwise separable convolution to lighten the PANet structure of YOLOv4.The improved Mobile Net V2-YOLOv4 model has a detection accuracy of 97.3% m AP and a detection speed of 53 fps,which can meet the needs of industrial production.(3)Design and build a defect detection system platformThe basic software and hardware platform of the defect detection system is constructed by the combination of edge computing and cloud computing.The platform uses CPU and GPU computing resources provided by Alibaba Cloud for training and storage of deep learning algorithm models,dynamically update algorithm weight files to edge server for deep learning inference according to detection requirements,on the premise of ensuring data privacy and security,real-time defect detection in industrial field is realized,and the algorithm has good scalability.(4)Develop the software of prototype system for surface defect detection of injection molded workpiecesBased on the above system platform and the researched defect detection algorithm,a complete surface defect detection prototype system is designed and implemented,and the system function and performance are tested in the edge computing environment.After testing,the data transmission between end-to-end detection has low latency,and the detection algorithm model shows good detection performance and real-time response in the system,which can meet the requirements of industrial real-time detection. |
