Defect Identification From Phased Array Ultrasonic Testing Images Based On A Variety Of Convolutional Neural Networks

Phased Array Ultrasonic Testing(PAUT)has become a hotspot in the application of nondestructive testing technology in the industrial field because of its high beam designability,wide sound field coverage and high sensitivity.The result of PAUT testing is presented as 2D images.Although it is intuitive,there is a complex correlation between the content of image and the information of defect type and size.Manual analysis is not only inefficient,but also depended on the subjective factors of the inspector.It is difficult to ensure the accuracy and reliability of defect identification result.Machine vision technology based on neural network has shown good prospects in image analysis of transportation,remote sensing and medical fields.In this paper,aiming at the characteristics of phased array ultrasound images,a machine vision method based on several convolution neural networks was used to optimize the algorithm for confusing sample image classification and small target with fuzzy edge semantics segmentation in phased array ultrasound images.Qualitative and quantitative analysis of typical volumetric and planar defects was achieved.The research contents and main results are as follows:(1)In order to build a dataset which is large enough to support the training and testing of neural network,22 PAUT images of carbon steel plate specimens and 33 BOSS weld PAUT images were collected.The defect types are holes and cracks,crack lengths are 1 mm,2 mm,3 mm and 5 mm,hole diameters are 1 mm,2 mm and 3 mm,and defect depth is in the range of 5 mm to 50 mm.Effective local images were captured from these images,and data enhancement was carried out by means of translation,rotation,and mirroring.Finally,a defect classification dataset containing 10,370 samples and a defect feature restoration dataset containing 7686 samples were established.(2)To accurately distinguish the defects in the PAUT image,an image classification algorithm based on Le Net was established.The algorithm can accurately distinguish the defects in the PAUT image samples of carbon steel plate,but it is not effective for the more complex BOSS welding PAUT image classification.The accuracy of defect classification in BOSS welding PAUT images was greatly improved after balancing the proportion of various samples and using VGG16 which has stronger feature extraction ability as the main body of the algorithm.(3)In view of the small defect target and fuzzy edge in phased array ultrasound images,an improved U-Net semantics segmentation algorithm based on attention mechanism and Focal Loss was adopted.This method can segment phased array ultrasonic testing images of carbon steel plate samples and BOSS weld samples accurately.Moreover,the algorithm has high computational efficiency while guaranteeing the accuracy of semantic segmentation.(4)Combining the defect classification algorithm based on Le Net and VGG16 and the improved U-Net defect feature restore algorithm,the PAUT images were classified and semantically segmented area by area with a sliding window to visually display the category,location,size and orientation of the defects in the image.This method can identify and correctly classify all 24 defects in the PAUT images of carbon steel plate specimens,and the correct ratio of recognition for 33 defects in the BOSS welding PAUT images was 72.72%.Based on the result of image recognition,the defects are further located and quantified.The maximum error of depth quantification was 10.67%,and the maximum error of dimension quantification was 11.00%.