Research On Defect Classification Of Stainless Steel Welds Based On Deep Learning

Stainless steel materials have been widely used in storage,transportation,processing and mechanical manufacturing due to their good oxidation resistance,corrosion resistance and easy welding.Stainless steel workpiece is to realize the combination of two parts by welding.Because of the local invisibility of welding,cracks,porosity,slag inclusion,incomplete fusion and incomplete penetration are often produced in the weld joints as the weakest area of the whole component,so it is difficult to realize the reliable quality of the two welding.At the same time,with the increase of service life and the influence of operation environment,the weld of steel plate will also produce new defects.It is very important to strengthen the quality inspection of stainless steel weld defects,and to alarm timely when defects are found.It is very important to repair defects or stop production in time,minimize economic losses and ensure the safety of people's lives and property.This paper studies five kinds of welding defects in the process of stainless steel material use,and it is divided into three parts as follows.1.In view of the problems of less recognition difficulties in ultrasonic time domain image data set of stainless steel weld defects,Alex Net model is adjusted and migration learning technology is integrated;The accuracy of weld defect identification model classification is improved to 95.12%.The problem of low classification accuracy of stainless steel weld defects under small sample is solved through the above research.2.A new classification method of Lack Net weld defects based on deep and shallow multi feature fusion is proposed.By extracting the features of time-frequency domain,the paper constructs the Lack Net weld defect identification network by using the deep and shallow multi feature fusion idea.The shallow features are constructed by using Statistical features and Gabor texture features.The VGG11 network is improved as the deep feature extraction network;The high precision recognition of five kinds of weld defects is achieved by fusion of deep and shallow multi-features.Compared with the machine learning method and depth learning method adopted by previous research,this algorithm has the characteristics of high efficiency and time saving,high precision and stability,and has good performance in the processing of weld defect classification.3.Through the integration of the first two parts,and MATLAB programming language,the image detection system of stainless steel weld defects is designed.The system can input the ultrasonic sequence signal of stainless steel weld defects and preprocess the data,and input the image data into the neural network for training the network model,and test the classification model by batch or single data,Finally,the system automatically generates the stainless steel weld defect detection report to assist the inspection personnel team to identify the weld defects.