Research On Classification And Recognition Method Of Weld Defects Based On Feature Fusion

Welding technology is a common industrial connection technology,widely used in machinery manufacturing,agricultural machinery and equipment production,industrial production and other industries.Due to the complexity of welding environment and the influence of welding technology,it is inevitable to produce local defects in welding process.In order to ensure the reliability and safety of welding,it is very important to accurately detect the materials of welding parts.The traditional quality evaluation method based on Xray welding image adopts manual evaluation,but manual evaluation has the disadvantages of time-consuming and laborious,and depends on the quality of the evaluation personnel.The automatic detection system of weld image defects based on computer has become a new trend and an important direction.The traditional computer-aided assessment of the main defect part segmentation,feature extraction,feature classification.In this paper,morphological processing method was used to obtain the location of weld defects.Then,three weld features are extracted,two manual features and one self-learning feature.The contour information of weld defects was extracted by directional gradient histogram,texture features of weld defects were extracted by gray level co-occurrence matrix,and the deep neural network was preprocessed unsupervised by autoencoder to obtain self-learning deep features of defect images.Considering that the description of weld defect features by single feature is not comprehensive,this paper proposes a method of weld defect image classification and recognition based on feature fusion technology.Firstly,the extracted single features are combined with each other to obtain the fusion features.Considering the different classification effects of extracted features by different methods,different features are given different weights for fusion.Finally,support vector machine is used for classification.The experimental results show that the classification effect of weld defect feature fusion is better than that of single feature classification,which effectively proves the reliability of the proposed method.Moreover,the method can accurately classify weld defect based on a small number of training samples,and the classification effect is better than the traditional method.