Application Research Of Steel Pipe Weld Defect Detection Based On Deep Learning

Steel pipes are widely used in high-risk and high-pressure scenarios such as petroleum,chemical,natural gas,shale gas,etc.Steel pipe weld defects are the main factor causing destruction of pipes,and if there is some kind of defect in the steel pipe weld,it will lead to serious adverse consequences.The previous generations have made rapid progress in the field of steel pipe weld defect detection,but still face some limitations and challenges.Two major issues are:(1)Some of the current steel pipe weld defect detection studies only focus on some of the indicators and ignore other indicators that affect steel pipe production efficiency,the precision of the model and the inference speed of the model for a single picture of the weld defects need to be improved;(2)Steel pipe weld defect detection is difficult to automate,and factors such as light changes,temperature and moisture or hardware environment in the real production environment may affect the results of the application of the model on the task of detection of steel pipe weld defects,and the generalization of the model needs to be further improved.To solve the above drawbacks,this paper takes the weld defects of spiral submerged arc welded steel pipes produced in a steel pipe factory in Jingzhou City,Hubei Province as the research object,and combines the industry-leading one stage object detection model YOLOv5 in the field of deep learning to conduct crossdisciplinary fusion application research.An improved genetic algorithm with very strong search capability in high-dimensional space is first proposed,and then this algorithm is used to search and improve part of the network structure of YOLOv5 to make the model more applicable to the real application of steel pipe weld defect detection.The innovations in this paper are as follows:(1)In-depth factory research and sampling,a standard steel pipe weld defect detection dataset containing eight common steel pipe weld defects was built based on the real production conditions of submerged arc spiral welded steel pipe welding.(2)Based on the simple genetic algorithm,a novel improved genetic algorithm is proposed,and its superiority-seeking ability in highdimensional space is analyzed and demonstrated from both qualitative and quantitative perspectives.(3)In the process of applying the improved genetic algorithm to the network architecture search task of the steel pipe weld defect detection model,the idea that the network structure and network super-parameters should "evolve" along with the training of the model is proposed,and that the network structure,network superparameters and model parameters are equally important.It is believed that the network structure,network hyperparameters and model parameters are equally important,and the model should be driven by data as much as possible to reduce the influence of human a priori intervention.Based on this idea,we design a strategy of network structure optimization and network training in parallel in two platforms.The experimental results show that the method and model proposed in this study are very effective in the application of real-time detection of steel pipe weld defects,improving the detection accuracy of defects while ensuring the detection speed.The mAP@0.5 of this paper’s model is 98.9%,the precision is 98.6%,the average inference delay of a single defect picture is 38 ms,and the average confidence level of all types of defects is above 80%.The research of this paper can provide methods and ideas for real-time automatic detection of steel pipe weld defects in an industrial production environment and establish the foundation for industrial automation.