Intelligent Algorithm Research And System Implementation Of Industrial Product Surface Defect Detection

In deep learning applications in industrial scenarios,small samples and imbalanced data sets are common problems.The deep learning application based on neural network model usually requires training based on large-scale label samples,while the small sample problem makes the model overfit on the training set and has poor generalization performance,so it cannot be well applied to the target task.The unbalance of the sample makes the training set different from the actual distribution of the sample,so that it can not be well applied to the target task.The above problems lead to a big difference between the training results of the deep learning model and the actual scene.Although artificial intelligence algorithms based on computer vision have been more and more applied in the industrial field,most of the intelligent algorithms are deployed in the cloud server,unable to timely process the collected industrial data.At the same time,intelligent algorithms have the ability to collect industrial data in the process of operation,and existing systems do not make good use of the value of the above data.Therefore,at present,the industry still lacks a real-time online surface quality detection system for industrial products based on cloud edge collaboration with high timeliness and high availability,and based on industrial data online expansion with high versatility and high expansibility.The topic of thesis is selected from the scientific research cooperation project of enterprises and public institutions "Research on the online detection technology of Intelligent Factory product Quality Defects based on mobile edge intelligence",which solves the technical problems from the perspective of practical application.From the perspective of algorithm optimization,thesis introduces a rebalancing method based on generated data to solve the problem of sample imbalance in industrial scenarios.From the point of system optimization,a real-time on-line quality inspection system for industrial products is designed and implemented.The main work of thesis is as follows:1)The current research on surface defect detection algorithms and systems based on computer vision in industrial scenarios is reviewed.Firstly,the research status of surface defect detection algorithm of industrial products based on computer vision is investigated,and the research status of small sample problem and unbalance problem commonly existed in industrial data set is introduced.Then it introduces the development status of surface defect detection system of industrial products based on computer vision and the basic situation of related system development technology.The summary study lays a foundation for the subsequent algorithm optimization and system design.2)Aiming at the common problems of small samples and unbalance in industrial data sets,a semi-supervised product surface quality detection algorithm based on rebalancing was proposed to improve the accuracy of surface image classification of industrial products.Firstly,a generative adversarial network framework is designed to integrate the rebalancing method into the semi-supervised method,and then a threebranch structure is introduced into the decision to achieve multi-task decoupling,and the average teacher consistent learning is introduced to improve the stability of the training process.Finally,the algorithm is validated on several public data sets and selfbuilt factory copper foil surface defect data sets.3)Aiming at the problems of poor availability and real-time performance of the existing system,lack of universality and extensibility,a real-time online surface quality detection system for industrial products is designed and developed.Based on the modular design idea,the system has designed and realized six system modules:data acquisition module,data processing module,data persistence module,communication cooperation module,system management module and data visualization module.The above modules cooperate with each other to realize the following system functions:algorithm autonomous incremental update,edge cloud collaborative communication and real-time online detection of industrial products.The system realizes the acquisition,processing and analysis of industrial metal surface data,as well as the output,presentation and management of alarm information.Meanwhile,the system enhances the availability and real-time of detection based on the edge-cloud collaborative architecture,and enhances the universality and extensibility of the system based on data expansion and algorithm autonomous incremental update.