Research On Surface Defect Detection Method Of Hot Rolled Steel Strip Based On Semi-Supervised Transfer Learning

The steel strip is one of the leading products of China’s modern steel industry,widely used in automobile manufacturing,high-speed rail transportation,aerospace and shipbuilding.However,affected by raw materials,the rolling process,the production environment and other factors,the surface of the steel strip in the rolling process will appear to be defective,affecting product quality.Detecting surface defects in hot-rolled steel strips is essential in modern industrial production.However,due to the difficulty of detecting,the production site environment is harsh,and the traditional detection technology is complex and inefficient.Deep learning technology can detect the surface defects of hot-rolled steel strips more efficiently and accurately.In actual industrial production,obtaining a large number of labelled defect samples is difficult due to the lengthy collection period of surface defect samples of hot rolled steel strips,few industrial defect samples,and heavy manual labelling work.Therefore,the semi-supervised learning method can effectively use unlabeled data to improve prediction performance and reduce labelling costs.It can provide a basis for the state assessment and fault diagnosis of hot-rolled steel strips.This thesis is mainly divided into two parts:The first part is to improve the Faster R-CNN algorithm based on the characteristics of surface defect data of hot-rolled steel strips.In order to solve the problem that the number of samples of hot rolled steel strips is small,the surface defect data set is expanded using data augmentation technology.In order to solve the problem of overfitting and low detection accuracy in deep learning model training due to the small number of hot-rolled steel strip samples,a transfer learning method based on pre-training was adopted.According to the characteristics of complex defects in images,difficult-to-distinguish boundaries,and significant differences in defect shapes and sizes,Res Net-50 is selected as the backbone network in the Faster R-CNN detection algorithm and improved.The Efficient Channel Attention(ECA)module was used to improve the influence of defect features on the network model and strengthen the position of surface defects on hot-rolled strips.Deformable Convolutional Networks(DCN)techniques are used to generate deformable feature images,to extract the features better.Feature Pyramid Grids(FPG)can supplement low-level spatial location features and high-level semantic features so that different-sized objects can be detected at various levels.In the NEU-DET hot-rolled strip surface defect dataset,the mean Average Precision(m AP)of the experiments before and after the improvement increased by 6.9%,indicating that the improved algorithm is able to complete the detection of surface defects more accurately and effectively.In the second part,aiming at the difficulty of surface defect sample labelling of hot rolled steel strip,the improved Faster R-CNN algorithm is taken as the base algorithm.On this basis,a defect detection method based on semi-supervised transfer learning is proposed.The semi-supervised learning framework of Self-Training and the Augmentation driven Consistency regularization(STAC)was improved.Improved STAC framework using both strong and weak data augmentation is applied to the unlabeled data and the weak data augmentation input to the teacher model.In the weak data augmentation,flip and noise are added to the framework,and in the strong data augmentation,Mosaic data enhancement is added to the framework;The Keep Augment method is used to preserve the original critical information.The Exponential Moving Average(EMA)method is used to update the teacher and student model weights.The experimental results show that the m AP of the base model using the improved STAC framework is 16.2%,12.9% and 4% higher than that of the base model when only 5%,10%,and 20% of the NEU-DET hot-rolled steel strip surface defects dataset have labelled data,respectively.It is verified that the improved framework can reduce the labelling cost and improve the model’s accuracy by using a small amount of labelling data.