Submerged arc welding pipe weld defect judgment method based on deep learning

Through the multi-scale feature extraction and classification of AEDetNet network, the missed detection and missed detection problem of manual interpretation in the detection of weld defects of submerged arc welds is solved, the generalization ability and detection accuracy of the model are improved, and interpretable detection results are provided.

CN118967609APending Publication Date: 2024-11-15NORTHEASTERN UNIV AT QINHUANGDAO

Patent Information

Application Number
CN202411021666.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

In the detection of weld defects of submerged arc welded pipes, the prior art relies on manual interpretation to risk missing and mis-checking, and the deep learning model has insufficient generalization ability under different welding environments and conditions, making it difficult to provide explainable detection results.

Method used

AEDetNet network is adopted, combining convolutional neural network, feature pyramid network and adaptive dual attention module, multi-scale feature extraction, classification and positioning of weld defects through anchor frames, and training is performed using classification and regression loss functions to improve the model's Generalization capability and detection accuracy.

Benefits of technology

Multi-scale defect classification and positioning are realized, the detection ability of defects of different sizes is enhanced, the accuracy of weld defect recognition and the generalization ability of model are improved, and interpretable detection results are provided.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure BDA0004967457460000036
    Figure BDA0004967457460000036
  • Figure BDA0004967457460000038
    Figure BDA0004967457460000038
  • Figure BDA0004967457460000041
    Figure BDA0004967457460000041
Patent Text Reader

Abstract

The invention belongs to the technical field of welding and computer vision, and discloses a submerged arc welding pipe weld defect judgment method based on deep learning. The collected submerged-arc welding pipe weld defect data in the SEQ, IMG and DCM formats are read in a binary mode and stored as an image file in the JPG format, and a weld image is obtained. The AEDetNet network is used for extracting image features through a convolutional neural network (CNN), constructing multi-scale feature maps by using a feature pyramid network (FPN), and finally performing classification and regression on the feature maps through a series of anchor boxes to realize classification and positioning of weld defects. The AEDetNet network designed by the method enhances feature extraction and representation capabilities, so that the network can better capture and emphasize important features, and the accuracy of weld defect identification is improved. According to the method, defect classification and positioning can be carried out on multiple scales, and the detection capability of the model on defects of different sizes is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of welding technology and computer vision technology, and in particular to a method for determining weld defects of submerged arc welded pipes based on deep learning. Background Art

[0002] Welding is a widely used connection process in the manufacture of various metal structural parts such as automobiles, ships, aircraft, pressure vessels and pipelines. During the welding process, various types of welding defects are inevitably generated due to complex conditions, which seriously damage the welding quality. Therefore, it is necessary to use non-destructive testing (NDT) technology to check the quality of welded joints. Radiographic testing (RT) for detecting internal defects in welds is a key technology for non-destructive testing of welding. X-ray or gamma-ray sources generate weld radiographic images by penetrating the weld structure and exposing photographic films. Weld defects are described by changes in the intensity of the radiographic images.

[0003] In the straight seam welding factory of submerged arc welded pipes, most weld defect detection is done by manual inspection. Most X-ray inspection posts require two inspectors to perform weld defect identification inspection. Each steel pipe needs to be dynamically scanned for the entire pipe body weld, and the scanned video is presented in the imaging room through an industrial camera. The imaging post manually identifies the scanned video, judges the defects, and determines the level. Based on the output of 100 steel pipes within 12 hours of operation, 1,200 meters of video images need to be identified every day. The work intensity of employees is high, there is a risk of missed detection and false detection, and the detection results are easily affected by the subjective factors of the inspectors. Therefore, the research on automatic recognition technology of weld defects in submerged arc welded pipes is of great significance.

[0004] In order to eliminate the limitations of manual interpretation, many computer-aided algorithms have been developed to automatically detect defects in radiographic images. With the improvement of big data and computing power, deep learning algorithms have developed rapidly, providing new solutions for computer-aided defect detection. Deep neural networks (DNNs) can directly process images and output results to achieve end-to-end detection. In addition, DNN models can automatically discover and extract high-level features through training without manually creating features. Compared with traditional detection algorithms, deep learning algorithms have the advantages of strong generalization ability and automatic feature extraction, and have been applied in weld defect detection. Weld defect recognition based on deep learning has obvious advantages in recognition quality and operation efficiency. The defect recognition model established by training a large number of defect images can quickly identify weld defects, thereby replacing manual recognition, achieving job reduction, and achieving high-efficiency and high-quality operations. Although deep neural networks (DNNs) have shown certain advantages in weld defect detection tasks, in actual applications, due to the different forms of weld defects in submerged arc welded pipes, directly migrating and using deep learning-based target detection technology may lead to problems such as target loss. Selecting the most suitable model for weld defect detection and optimizing it requires a lot of experiments and adjustments. In addition, how to improve the generalization ability of the model so that it can perform well in different welding environments and conditions and provide explainable detection results is also a problem that needs to be solved.

[0005] Based on the above difficult problems, further research is needed on how to use deep learning methods to determine the weld defects of submerged arc welded pipes. Summary of the invention

[0006] In order to solve at least one of the technical problems existing in the prior art to a certain extent, the purpose of the present invention is to propose a method for determining weld defects of submerged arc welded pipes based on deep learning.

[0007] The technical solution adopted by the present invention is: a method for determining weld defects of submerged arc welded pipes based on deep learning, which is as follows:

[0008] The collected weld defect data of submerged arc welded pipes in SEQ, IMG, and DCM formats are read in binary mode and saved as image files in JPG format to obtain weld images;

[0009] The AEDetNet network is used to extract image features through a convolutional neural network (CNN) and construct multi-scale feature maps using a feature pyramid network (FPN). Finally, a series of anchor boxes are used to perform classification and regression on these feature maps to achieve the classification and location of weld defects.

[0010] The AEDetNet network includes a residual network (ResNet), an adaptive dual attention (SDA) module, a feature pyramid network (FPN), a classification subnetwork, and a box regression subnetwork;

[0011] The input of the AEDetNet network is a weld image. The features of the images at different stages are extracted through ResNet. The feature maps of different stages are input into the SDA module to enhance the feature representation. The enhanced feature maps are input into FPN to obtain multi-scale feature maps. The multi-scale feature maps are input into the decoder. The decoder is two parallel sub-networks. The probability of each anchor box belonging to each category is obtained through the classification sub-network, and the position adjustment parameters of each anchor are obtained through the box regression sub-network. The position and size of the anchor are adjusted according to the regression results, and the predicted image is output.

[0012] The ResNet includes a 7×7 convolution with a stride of 2, a 3×3 Maxpool layer with a stride of 2, and 4 combination layers of two types of residual modules, wherein the two types of residual modules are divided into a basic residual module (BasicBlock) and a down-sampling residual module (BasicBlock_D).

[0013] The SDA module consists of two branches: spatial attention branch and channel attention branch. Let the input be X in ∈R C ×H×W In order to reduce the amount of calculation of the model, first input X in Perform a maximum pooling operation with a pooling factor of 2, and then reshape the shape of the tensor after pooling to obtain Transpose the reshaped tensor to obtain Right now:

[0014] X r =Reshape(Maxpooling(X in ))

[0015] In the spatial attention branch, X r and Multiply them together to get the similarity matrix between pixels, and normalize the similarity matrix through the Softmax activation function to get the pixel-level attention coefficient K p With X r Multiply them together to amplify the parts of the input feature matrix with high spatial similarity, obtain the feature map after pixel-level attention enhancement, and then reshape and upsample it to its original size to match the dimension of the original input, and obtain X p ∈R C×H×W .

[0016]

[0017] In the channel attention branch, With X r Multiply them together to get the similarity matrix between channels, and normalize the similarity matrix through the Softmax activation function to get the channel-level attention coefficient K ch ∈R C×C ; X r With K ch Multiply them to get the channel-level attention-enhanced feature map, which is then reshaped and upsampled to its original size to get X ch ∈R C×H×W By explicitly modeling the correlations between channels, channels with high interdependence are emphasized, thereby improving feature representation ability and attention performance.

[0018]

[0019] Finally, the feature X output by the spatial attention branch p And the feature X output by the channel attention branch ch According to the learnable adaptive weight parameters α and β, weighted fusion is performed and combined with the original feature X in Add together to get the final SDA output X out ∈R C×H×W :

[0020] X out =X in +αX ch +βX p

[0021] The FPN consists of two parts: lateral connection and top-down path. The lateral connection processes the feature map output by the SDA module through a 1×1 convolution layer to unify the number of channels and reduce the amount of computation. The top-down path first uses the nearest neighbor interpolation method to double the spatial resolution to match the resolution of the next level feature map; then, the upsampled feature map is merged with the feature map output by the corresponding lateral connection by element-by-element addition to fuse feature information of different scales; finally, we append a 3×3 convolution to each merged feature map to generate the final feature map to reduce the aliasing effect in the upsampling process and enhance the feature expression capability.

[0022] The decoder consists of two parallel subnetworks: a classification subnetwork and a box regression subnetwork. The classification subnetwork includes multiple 3×3 convolutional layers, each followed by a ReLU activation function to extract and refine features; finally, a 3×3 convolutional layer is used to adjust the number of channels to the number of categories, and the probability of each anchor belonging to each category is output through a Sigmoid activation function. The box regression subnetwork also includes multiple 3×3 convolutional layers, each followed by a ReLU activation function to extract and refine features; finally, a 3×3 convolutional layer is used to adjust the number of channels to 4, and the position adjustment parameters of each anchor are output.

[0023] The AEDetNet network is used to classify and locate weld defects in submerged arc welded pipes. The AEDetNet network training uses two different types of loss functions: classification loss L cls And the regression loss L box ;

[0024] The classification loss uses focal loss to determine the category of the generated defect bounding box; the regression loss uses SmoothL1 loss to evaluate the positioning accuracy.

[0025] Loss total =L cls +L box

[0026] L cls =-α t (1-p t ) γ log(p t )

[0027]

[0028] The beneficial effects of the present invention are:

[0029] 1. The weld defect determination method for submerged arc welded pipe proposed in the present invention can classify and locate defects at multiple scales, thereby enhancing the model's ability to detect defects of different sizes.

[0030] 2. The AEDetNet network designed by the present invention enhances the feature extraction and representation capabilities, enabling the network to better capture and emphasize important features, thereby improving the accuracy of weld defect identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart for determining weld defects of submerged arc welded pipes based on deep learning in an example of the present invention.

[0032] Figure 2 Schematic diagram of the AEDetNet network structure in the example of the present invention.

[0033] Figure 3 It is a schematic diagram of a basic residual module (BasicBlock) in an example of the present invention.

[0034] Figure 4 Schematic diagram of the down-sampling residual module (BasicBlock_D) in the example of the present invention.

[0035] Figure 5 Schematic diagram of the adaptive dual attention (SDA) module in an example of the present invention.

[0036] Figure 6 Schematic diagram of the classification subnetwork structure in an example of the present invention.

[0037] Figure 7 Schematic diagram of the frame regression subnetwork structure in the example of the present invention.

[0038] Figure 8(a)-Figure 8(h) The following is a test result diagram of a submerged arc welded pipe weld image with different types of defects in an example of the present invention. DETAILED DESCRIPTION

[0039] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limitations of the present invention. For the step numbers in the following embodiments, they are only provided for the convenience of explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0040] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., and orientations or positional relationships indicated are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0041] In the description of the present invention, "several" means one or more, "more" means more than two, "greater than", "less than", "exceed" etc. are understood as not including the number itself, and "above", "below", "within" etc. are understood as including the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0042] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0043] At present, the identification of weld defects in submerged arc welded pipes still uses manual identification, which has the disadvantages of lack of objectivity and consistency, low detection efficiency, and possible omission of small defects. Weld defect identification based on deep learning has obvious advantages in recognition quality and operation efficiency. However, how to improve the generalization ability of the model and provide explainable detection results remains challenging.

[0044] The collected weld defect data is read in binary mode to obtain a weld image, and the defective part in the weld is identified and located through AEDetNet, including the following steps:

[0045] Step 1: Obtain the weld image. The collected submerged arc welded pipe weld defect data in SEQ, IMG, and DCM formats are read in binary mode and saved as image files in JPG format.

[0046] Step 2: Construct an AEDetNet network, which includes a residual network (ResNet), an adaptive dual attention (SDA) module, a feature pyramid network (FPN), and a decoder (classification subnetwork and box regression subnetwork).

[0047] The ResNet includes a 7×7 convolutional layer, a 3×3 Maxpool layer, and a combination layer of a basic residual module (BasicBlock) and a down-sampling residual module (BasicBlock_D).

[0048] The SDA module includes two branches: spatial attention branch and channel attention branch. This module can assign different importance weights in spatial and channel dimensions, thereby improving the model's attention to important features and improving overall performance.

[0049] The FPN consists of two parts: a lateral connection and a top-down path, wherein the lateral connection is a 1×1 convolutional layer, and the top-down path includes upsampling and a 3×3 convolutional layer.

[0050] The decoder consists of two parallel subnetworks: a classification subnetwork and a box regression subnetwork. The classification subnetwork includes multiple 3×3 convolutional layers, ReLU activation functions, and Sigmoid activation functions; the box regression subnetwork includes multiple 3×3 convolutional layers and ReLU activation functions.

[0051] Step 3: Set the loss function for network training. The loss function has two parts: classification loss and regression loss. The classification loss uses focal loss, and the regression loss uses Smooth L1 loss.

[0052] Step 4: Preprocess the weld image read in step 1 and resize it to a uniform size. This example can use but is not limited to images of 512×512 size. Use labelimg and other labeling tools to manually label it. Take the upper, lower, left, and right endpoints of the defect area to draw a marking box, where the coordinates of a pair of diagonal points are used as position labels.

[0053] Step 5: Pair the original image and location label and encapsulate them into a data set, iterate the training until the network converges, and obtain the trained AEDetNet network.

[0054] Step 6: Load the trained AEDetNet network weights, input the weld image, and the network automatically processes the image and generates a prediction result.

[0055] like Figure 1 As shown, this embodiment provides a method for determining weld defects of a submerged arc welded pipe based on AEDetNet, which specifically includes the following steps:

[0056] S1. Obtain weld images and save the collected weld defect data of submerged arc welded pipes in SEQ, IMG, and DCM formats as image files in JPG format.

[0057] The SEQ format data is read using binary reading, and the standard header information of the JPEG file is used as a delimiter to split the data content into multiple parts, thereby extracting the JPEG image data. The standard header byte sequence and image data of the JPEG file are written into the file to complete the saving of the JPG image.

[0058] For IMG format data, use binary reading method to read, move the file pointer to the image data position to be read from the end of the file (a total of width × height × 2 bytes), then read the image data and use the reshape method to reconstruct the one-dimensional array into a two-dimensional array, whose shape is the height and width of the image. Next, convert the 16-bit integer image data to 8-bit integers to meet the requirements of the JPG format. Finally, save the converted image data as a JPG file to achieve the conversion from IMG format to JPG format.

[0059] Use binary reading to read DCM format data, move the file pointer to the pixel data position to be read from the end of the file, read the image data and reshape it into a two-dimensional array. Next, convert the 16-bit pixel data into an 8-bit unsigned integer. Finally, save the converted image data as an image file in JPG format.

[0060] S2, build the AEDetNet network, such as Figure 2 The AEDetNet network includes a residual network (ResNet), an adaptive dual attention (SDA) module, a feature pyramid network (FPN), and a decoder (classification subnetwork and box regression subnetwork).

[0061] S2.1. In order to capture features at different depths, this example uses ResNet to extract image features, including 1 7×7 convolution with a stride of 2, 1 3×3 Maxpool layer with a stride of 2, and 4 combination layers (C2-C5) consisting of a basic residual module (BasicBlock) and a downsampled residual module (BasicBlock_D).

[0062] The detailed architecture of BasicBlock is as follows Figure 3 As shown in the figure, the module consists of two 3×3 convolutional layers with a stride of 1, each of which is followed by a batch normalization layer (Batch Normalization) and uses the ReLU activation function. The input data is then directly added to the output after two layers of convolution and batch normalization (this structure is called a "residual connection"), and finally the ReLU activation function is used. Figure 4 As shown in the figure, BasicBlock_D is similar to BasicBlock in structure, except that its first convolutional layer is a 3×3 convolution with a stride of 2, and a 1×1 convolution with a stride of 2 is added to the residual connection to adjust the number of channels and spatial dimensions of the input.

[0063] S2.2. In order to improve the model's attention to important features and enhance feature representation, this example introduces an SDA module at the connection between ResNet and FPN, that is, an SDA module is connected after each C3-C5 layer, and the feature maps output by each layer are recorded as CS3-CS5. Figure 5 As shown in the figure, the SDA module consists of two branches: spatial attention branch and channel attention branch. This module takes the feature X output by the spatial attention branch p And the feature X output by the channel attention branch ch According to the learnable adaptive weight parameters α and β, weighted fusion is performed and combined with the original feature X in Add together to get the feature map X after attention enhancement outAmong them, α and β (α, β ≥ 0) are two trainable parameters used to adaptively weigh the contributions of pixel-level and channel-level attention mechanisms so that the model can focus on more representative and relevant parts of the input data. By combining spatial attention and channel attention, the network can better capture and emphasize important features, thereby improving the accuracy of weld defect recognition.

[0064] X out =X in +αX ch +βX p

[0065] S2.3. In order to classify and locate defects at multiple scales and improve the model's ability to detect defects of different sizes, this example uses FPN to achieve multi-scale feature fusion. FPN combines low-resolution, semantically strong features with high-resolution, semantically weak features through lateral connections and top-down paths, forming a feature pyramid with rich semantics at all levels, thereby significantly improving the performance of weld defect detection tasks.

[0066] The lateral connection processes the feature map output by the SDA module through a 1×1 convolution layer to unify the number of channels and reduce the amount of computation. The top-down path first uses the nearest neighbor interpolation method to double the spatial resolution to match the resolution of the next-level feature map; then, the upsampled feature map is merged with the corresponding feature map output by the lateral connection by element-by-element addition to fuse feature information of different scales; finally, we append a 3×3 convolution to each merged feature map to generate the final feature map to reduce the aliasing effect during upsampling and enhance feature expression capabilities.

[0067] Specifically, CS5 first performs a 1×1 convolution to reduce the dimension to obtain P5_1, and then uses a 3×3 convolution with a step size of 2 to obtain the third feature layer P5 of the decoder; CS4 performs a 1×1 convolution on the feature map obtained by upsampling P5_1 and fuses it to obtain P4_1, and then uses a 3×3 convolution with a step size of 2 to obtain the second feature layer P4 of the decoder; CS3 performs a 1×1 convolution on the feature map obtained by upsampling P4_1 and fuses it to obtain P3_1, and then uses a 3×3 convolution with a step size of 2 to obtain the first feature layer P3 of the decoder.

[0068] In addition, C5 is convolved with a stride of 2 by 3×3 to obtain the fourth feature layer P6 of the decoder; ReLU is applied to P6, and then a 3×3 convolution with a stride of 2 is performed to obtain the fifth feature layer P7 of the decoder.

[0069] S2.4, the decoder consists of two parallel sub-networks: the classification sub-network and the box regression sub-network. Figure 6 As shown in , the classification subnetwork includes multiple 3×3 convolutional layers, each followed by a ReLU activation function to extract and refine features; finally, a 3×3 convolutional layer is used to adjust the number of channels to the number of categories, and the Sigmoid activation function is used to output the probability that each anchor belongs to each category. Figure 7 As shown in the figure, the box regression subnetwork also includes multiple 3×3 convolutional layers, each of which is followed by a ReLU activation function to extract and refine features; finally, a 3×3 convolutional layer is used to adjust the number of channels to 4, and the position adjustment parameters of each anchor are output.

[0070] The first to fifth feature layers (P3-P7) of the decoder are input into the classification subnetwork and the box regression subnetwork to obtain the probability of each anchor box belonging to each category and the position adjustment parameters of each anchor, and the position and size of the anchor are adjusted according to the regression results, and the predicted image is output. Through the synergy of these two subnetworks, the model can accurately identify and locate the target in the image.

[0071] S2.5. Set the loss function for network training. AEDetNet network training uses two different types of loss functions: classification loss L cls And the regression loss L box Among them, the classification loss uses focal loss to determine the category (true or false) of the generated defect bounding box; the regression loss uses Smooth L1 loss to evaluate the positioning accuracy.

[0072] Loss total =L cls +L box

[0073] L cls =-α t (1-p t ) γ log(p t )

[0074]

[0075] S3. Preprocess the weld image read by S1 and resize it to a uniform size. This example can use but is not limited to 512×512 images. Use labelimg and other annotation tools to manually annotate them. Draw a marking box at the four upper, lower, left and right endpoints of the defect area, with the coordinates of a pair of diagonal points as position labels. Pair the original image and the position label and encapsulate them as a data set. Iterate the training until the network converges to obtain the trained AEDetNet network. Load the trained AEDetNet network weights, input the weld image, and the network automatically processes the image and generates the prediction result, as shown in Figure 8.

[0076] The embodiment of the method for determining weld defects of submerged arc welded pipes based on deep learning of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented by software, or by hardware or a combination of software and hardware. For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative labor.

Claims

1. A method for determining weld defects of submerged arc welded pipes based on deep learning, characterized in that: The details are as follows: The collected weld defect data of submerged arc welded pipes in SEQ, IMG, and DCM formats are read in binary mode and saved as image files in JPG format to obtain weld images; The AEDetNet network is used to extract image features through a convolutional neural network (CNN) and construct multi-scale feature maps using a feature pyramid network (FPN). Finally, a series of anchor boxes are used to perform classification and regression on these feature maps to achieve the classification and location of weld defects.

2. The method for determining weld defects of submerged arc welded pipes based on deep learning according to claim 1, characterized in that: The AEDetNet network includes a residual network (ResNet), an adaptive dual attention (SDA) module, a feature pyramid network (FPN), a classification subnetwork, and a box regression subnetwork; The input of the AEDetNet network is a weld image. The features of the images at different stages are extracted through ResNet. The feature maps of different stages are input into the SDA module to enhance the feature representation. The enhanced feature maps are input into FPN to obtain multi-scale feature maps. The multi-scale feature maps are input into the decoder. The decoder is two parallel sub-networks. The probability of each anchor box belonging to each category is obtained through the classification sub-network, and the position adjustment parameters of each anchor are obtained through the box regression sub-network. The position and size of the anchor are adjusted according to the regression results, and the predicted image is output.

3. The method for determining weld defects of submerged arc welded pipes based on deep learning according to claim 2, characterized in that: The ResNet includes a 7×7 convolution with a stride of 2, a 3×3 Maxpool layer with a stride of 2, and 4 combination layers of two types of residual modules, wherein the two types of residual modules are divided into a basic residual module (BasicBlock) and a down-sampling residual module (BasicBlock_D).

4. The method for determining weld defects of submerged arc welded pipes based on deep learning according to claim 3, characterized in that: The SDA module consists of two branches: spatial attention branch and channel attention branch. Let the input be X in ∈R C×H×W In order to reduce the amount of calculation of the model, first input X in Perform a maximum pooling operation with a pooling factor of 2, and then reshape the shape of the tensor after pooling to obtain Transpose the reshaped tensor to obtain Right now: X r =Reshape(Maxpooling(X in )) In the spatial attention branch, X r and Multiply them together to get the similarity matrix between pixels, and normalize the similarity matrix through the Softmax activation function to get the pixel-level attention coefficient K p With X r Multiply them together to amplify the parts of the input feature matrix with high spatial similarity, obtain the feature map after pixel-level attention enhancement, and then reshape and upsample it to its original size to match the dimension of the original input, and obtain X p ∈R C×H×W . In the channel attention branch, With X r The similarity matrix between channels is obtained by multiplication, and the similarity matrix is ​​normalized by the Softmax activation function to obtain the channel-level attention coefficient K ch ∈R C×C ; X r With K ch Multiply them to get the channel-level attention-enhanced feature map, which is then reshaped and upsampled to its original size to get X ch ∈R C×H×W By explicitly modeling the correlations between channels, channels with high interdependence are emphasized, thereby improving feature representation ability and attention performance. Finally, the feature X output by the spatial attention branch p And the feature X output by the channel attention branch ch According to the learnable adaptive weight parameters α and β, weighted fusion is performed and combined with the original feature X in Add together to get the final SDA output X out ∈R C ×H×W : X out =X in +αX ch +βX p 5. The method for determining weld defects of submerged arc welded pipes based on deep learning according to claim 4, characterized in that: The FPN consists of two parts: lateral connection and top-down path. The lateral connection processes the feature map output by the SDA module through a 1×1 convolution layer to unify the number of channels and reduce the amount of computation. The top-down path first uses the nearest neighbor interpolation method to double the spatial resolution to match the resolution of the next level feature map; then, the upsampled feature map is merged with the feature map output by the corresponding lateral connection by element-by-element addition to fuse feature information of different scales; finally, we append a 3×3 convolution to each merged feature map to generate the final feature map to reduce the aliasing effect in the upsampling process and enhance the feature expression capability.

6. A method for determining weld defects of submerged arc welded pipes based on deep learning according to claim 5, characterized in that: The decoder consists of two parallel subnetworks: a classification subnetwork and a box regression subnetwork. The classification subnetwork includes multiple 3×3 convolutional layers, each followed by a ReLU activation function to extract and refine features; finally, a 3×3 convolutional layer is used to adjust the number of channels to the number of categories, and the probability of each anchor belonging to each category is output through a Sigmoid activation function. The box regression subnetwork also includes multiple 3×3 convolutional layers, each followed by a ReLU activation function to extract and refine features; finally, a 3×3 convolutional layer is used to adjust the number of channels to 4, and the position adjustment parameters of each anchor are output.

7. A method for determining weld defects of submerged arc welded pipes based on deep learning according to claim 6, characterized in that: The AEDetNet network is used to classify and locate weld defects in submerged arc welded pipes. The AEDetNet network training uses two different types of loss functions: classification loss L cls And the regression loss L box ; The classification loss uses focal loss to determine the category of the generated defect bounding box; the regression loss uses Smooth L1 loss to evaluate the positioning accuracy. Loss total =L cls +L box L cls =-a t (1-p t ) γ log(p t )

Citation Information

Patent Citations

  • Detection method for automatically identifying and detecting weld defects

    CN110570410A

  • X-ray weld defect detection method based on convolutional neural network

    CN113674247A

  • Display screen defect detection method and system based on cascade multilayer feature fusion network

    CN118154603A

Cited By

  • Construction engineering steel structure welding quality detection method based on deep neural network

    CN119963556A

  • Weld joint detection method and device based on deep learning

    CN120598949A

  • Deep learning-based weld inspection methods and devices

    CN120598949B

  • Weld defect automatic identification method and system based on visual inspection

    CN121170459A