Research On Fracture Evaluation Of Drop Weight Tear Test Based On Deep Learning

The Drop Weight Tear Test(DWTT)is one of the most important methods to assess the quality of pipeline steel.At present,the use of vernier calipers to measure the percentage of the tough surface regions of the samples is widely used as the standard for assessing the quality of steel.However,manual measurement has the disadvantages of low efficiency,low accuracy and non-reproducibility.At present,machine learning has been used to segment the brittle and tough regions of the sample.However,machine learning requires feature extraction of training data.In the drop weight tear test,the appearance of some specific images is complex,and there are many small and fragile areas discretely distributed,which makes it difficult to fully identify the situation.This brings certain difficulties to machine learning to evaluate the fragile part of images in the drop weight tear test.Deep learning can generate complex features from simple features by automatically learning the features of DWTT samples and the correlation among pixels,which brings more efficient and accurate assessment method to the research of DWTT images with complex texture regions and changeable appearances.This paper is based on the deep convolutional neural network to realize the training of the image data set of the fracture sample of the DWTT.The trained model is used to identify the brittle area in the DWTT fracture samples,and then the steel is evaluated.The specific research work of this paper is as follows:(1)We collect the fracture image of the DWTT and design a method to make the DWTT specimen fracture data set using the annotation tool,then we try preprocessing and enhancement operations on the fracture image of the DWTT specimen.Use Mask R-CNN,ResSegNet that add residual modules to the benchmark network and U-net to train the data set,and optimize the network structure by adjusting the optimizer algorithm and normalization methods.Through the manual observation of the edge contour segmentation effect and the use of image segmentation evaluation indicators to evaluate the segmentation results from subjective and objective perspectives,the experimental results show that the segmentation accuracy of the convolutional neural network in this paper is higher.(2)In order to reduce parameters,reduce memory consumption,this paper proposes a method using DSU-net which uses depthwise convolution and pointwise convolution to replace the original down-sampling stage for extracting the feature of DWTT images.In addition,due to the imbalance of brittleness and toughness caused by the complex texture characteristics of the images in DWTT,we compare the stability of CE loss function,and the Dice loss function,then we choose the better one.(3)In order to evaluate the ductile fracture arrest ability of pipeline steel specimens,an image-based method for calculating the percentage of brittle and ductile area of the shear surface of the specimen fracture is proposed.By comparing with the values measured by engineers,the calculation results of the algorithm model in this paper are compared with the values measured by engineers.The error of the measurement result is only within 5%,which proves the accuracy and effectiveness of this method.The experimental results show that the convolutional neural network model used in this paper achieves a high segmentation accuracy of the DWTT samples,which proves that the optimized network model proposed in this paper can perform pixel-level semantics on the images of DWTT.This method improves the level of intelligence and has a certain practical application value in industrial enterprise projects.