Research On Classification And Detection Method Of Weld Defects Based On Multi-feature Extraction

The occurrence of weld defects in the welding process is an inevitable problem,which will lead to significantly reduced welding quality,pose severe challenges to the reliability of welded products,and even cause serious consequences.Therefore,the research on weld defects is of great significance.At present,there are many deficiencies in traditional weld defect detection and recognition methods.Machine learning methods have been widely used,but there are different standards for different target defects.Research on improving the accuracy of classification and recognition has become a hot spot.This paper takes the steel plate surface weld defect as the research object,studies the multi-feature extraction weld defect identification and classification method,according to the characteristics of the steel plate weld itself,accurately extracts the steel plate weld defect contour,uses the shape geometry feature,the texture feature and the color feature,constructed the multi-parameter eigenvector of the weld,analyzed the relationship between the multi-parameter eigenvector of the weld surface and the defect category,studied the binary tree support vector machine welding defect classification algorithm,introduced the Levy flight strategy,the Circle chaotic map and the simplex reflection operation.The advanced LCSPSO algorithm improves the accuracy of weld defect identification and classification.The specific research contents are:(1)Before detecting,identifying and classifying the surface defect images of steel plate welds,it is necessary to preprocess the weld images to make the characteristics of defects more prominent,which is beneficial for extracting defect information and improving classification accuracy accurately.During processing,image enhancement techniques such as three filtering methods and histogram equalization are used to process defect images.(2)On the basis of preprocessing,the edge of the defect area is extracted completely.Aiming at the shortcomings of the Canny algorithm in removing false edges and protecting edge details in defect image edge detection,and being sensitive to salt and pepper noise,Otsu adaptively selects high and low thresholds,six-directional gradient amplitudes and uses adaptive medians.The filtering method replaces the original algorithm,and the comparison experiment proves that this improved method has a significant improvement compared with the original algorithm.(3)Aiming at the problem of inaccurate detection and extraction of image features of single weld defects,study the characteristic parameters of comprehensive geometric shape features,texture features and color features to ensure the comprehensiveness of defect information;analyze the weld surface characteristic parameters and The relationship between defect categories,in order to prevent the redundancy of input feature parameter information,normalize the parameters before detection,select useful information,and improve the speed of detection tests.(4)Using LCSPSO to optimize BTSVM to classify and detect weld defects.The LCSPSO algorithm is used to train the test samples to find out the two parameters of the optimal binary tree support vector machine.Aiming at the problems that the PSO algorithm is easy to fall into local optimum,insufficient initialization diversity,and low calculation accuracy in the later stage,the LCSPSO algorithm that introduces Levi’s flight strategy,Circle chaotic map and simplex reflection operation is studied,and it is verified by using 8classic complex function tests.The effectiveness of the algorithm;in order to verify whether the extracted multiple feature parameters can improve the classification accuracy,the classification accuracy of only the geometric shape feature input into LCSPSO-SVM and the three types of features are input into LCSPSO-SVM at the same time,and it is verified that multiple feature extraction can improve Classification accuracy;in order to compare the classification effect of the improved algorithm,set three types of features and input LCSPSO-SVM,PSO-SVM,and Levy-SVM at the same time.The test results show that LCSPSO is better than other algorithms,and the classification accuracy is increased by11.3 %,the average classification accuracy rate is 99.2%,and the classification accuracy rate of multi-features is higher than that of single-features.Figure [58] table [10] reference [73]...