Intelligent identification method of pipeline girth weld radiographic defects based on improved YOLOv5 network

Abstract

The pipeline girth weld X-ray film defect intelligent identification method based on the improved YOL0v5 network of the present invention, through the construction of pipeline girth weld defect data set, YOL0v5 model improvement and model verification, weld defect intelligent identification system design and verification, better solves the problems and shortcomings of traditional weld defect detection technology, including unstable and inaccurate detection results caused by reliance on manual judgment, limited recognition ability of complex defects and low defect contrast; by combining image enhancement and multi-scale feature extraction technology, an end-to-end defect detection model is constructed to realize intelligent recognition of weld defects, thereby improving the accuracy and stability of detection, meeting the needs of industrial production for rapid and accurate detection of welding quality, and is of great significance for improving production efficiency and ensuring product quality.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to an intelligent method for identifying defects in pipeline girth weld radiographic films based on an improved YOLOv5 network. Background Art

[0002] Welding in metal processing and manufacturing is a crucial process used to connect parts into final products. According to statistics, more than 45% of steel and metal materials are processed by welding, and the quality of welding directly determines the performance and service life of the product. However, defects such as cracks, lack of fusion or lack of penetration often occur during the welding process, which directly affect the quality and safety of the product. In scenarios such as pressure pipelines where welding quality requirements are extremely high, it is crucial to quickly and accurately detect the quality of welds, and internal defect detection of welds based on X-ray imaging is a key technology. However, traditional radiographic image detection technology often relies on manual judgment and is easily affected by the judge's experience and subjective factors, resulting in instability and lack of accuracy in the detection results.

[0003] With the development of industrial automation and digitalization, computer vision technology based on deep learning has gradually matured, providing new solutions for the automatic detection and identification of weld defects. With the support of technologies such as big data, cloud computing and 5G Internet of Things, deep learning models can extract features from weld images and automatically identify defects, thereby improving the efficiency and accuracy of detection.

[0004] For example, the Chinese patent with application number CN202410254100.0 and titled "CCTV pipeline image defect target detection method, device and medium based on YOLOv5" discloses a CCTV pipeline image defect target detection method based on YOLOv5, which involves the field of target detection, including: obtaining pipeline defect detection images and classifying them; using multi-scale fusion algorithm to perform data enhancement on pipeline defect detection images, constructing pipeline defect detection data sets and annotating them, and dividing the annotated pipeline defect detection data sets into training sets, validation sets and test sets; improving the algorithm model through partial convolution and global attention mechanism; inputting the image data in the training set into the improved algorithm model for training to obtain the best weight file; saving the best weight file into the algorithm model; using the trained algorithm model to perform defect detection on the image data of the test set, determine the defect location and detection accuracy, and complete the detection of pipeline defects. The precision, recall rate and average precision of the improved model have been significantly improved. However, this method still has room for improvement when facing the recognition of complex defects, irregularly shaped defects, etc.

[0005] Therefore, it is necessary to further improve the existing deep learning weld inspection technology to further improve the accuracy and stability of weld defect detection. Summary of the invention

[0006] In order to solve the above technical problems, the present invention discloses an intelligent identification method for pipeline girth weld radiographic film defects based on an improved YOLOv5 network, which at least comprises the following steps:

[0007] S1. Construction of pipeline girth weld defect dataset:

[0008] S1.1. Digitally image the girth weld area of ​​the pressure pipeline using the DR digital imaging system;

[0009] S1.2. Perform image enhancement processing on the original image by using a denoising algorithm and a contrast adjustment algorithm;

[0010] S1.3. Use Labelimg software to annotate at least seven weld defects in the image, including cracks, lack of fusion, lack of penetration, strip defects, circular defects, root concave and root bite, and construct a data set;

[0011] S2.YOLOv5 model improvement and model verification:

[0012] S2.1. The C2F module is used to replace the CSPDarknet53 module of the traditional YOLOv5 model framework to enhance the extraction capability of high-order features and reduce model parameters to improve feature extraction capability and capture more detailed information;

[0013] S2.2. Improve the Neck part of the feature fusion module of the YOLOv5 model framework by adding an additional feature layer and fusing it with features from other layers through the feature pyramid network FPN to enhance the detection capability of small objects;

[0014] S2.3. Design an improved YOLOv5 network model loss function, combining the binary cross entropy confidence loss function Lobj, the multivariate cross entropy classification loss function Lcls, and the bounding box regression loss function Lbbox to achieve the evaluation of the specific category, position, and size of the target;

[0015] S2.4. The dataset is divided into a training set and a validation set in a ratio of 7:3, and the original YOLOv5n model and the improved YOLOv5 model are trained separately;

[0016] S2.5. Experimentally verify the trained object detection model on crack, incomplete penetration, and circular defect detection of pipeline girth weld DR image dataset, and evaluate the effect of the improved YOLOv5 model;

[0017] S3. Design and verification of intelligent identification system for weld defects:

[0018] S3.1. Design an intelligent identification system for pipeline annular seam radiographic film defects, which can automatically implement image enhancement processing, weld defect type identification, weld defect location and data storage operations in sequence by importing pictures;

[0019] S3.2. Analyze the effect of the image optimization system through the dataset images.

[0020] Preferably, in S1, due to the structural noise of the imaging plate in the DR imaging system, the electrical noise of the high-voltage radiation source, the diffuse scattering effect and the imaging environment, the image clarity is low, among which the unclearness of circular defects is particularly obvious; Laplace filtering and contrast-limited adaptive histogram equalization algorithm are used to perform image enhancement processing on the original image, so that the clarity of image defects is significantly enhanced while retaining important detail features.

[0021] Preferably, the Laplace filter algorithm formula is:

[0022]

[0023] The formula of the contrast-limited adaptive histogram equalization algorithm is:

[0024] I new (x, y) = CLAHE (I (x, y))

[0025] In S2, the C2F module uses two Bottleneck layers for feature extraction, and through multi-branch flow design and reuse of features at different stages, it improves the feature extraction capability and captures more detailed information.

[0026] Preferably, the additional feature layer is a layer of 4-fold down-sampling feature map added on the basis of the original three-layer feature map obtained by 8-fold, 16-fold and 32-fold down-sampling.

[0027] Preferably, the additional feature layer is fused with features of other levels through a feature pyramid network FPN, and effective fusion of high-level semantic information and low-level spatial information is ensured through top-down and bottom-up information transfer.

[0028] Preferably, the method for fusing the FPN with features of other levels is, on the basis of FPN, fusing the 4-fold downsampled feature map with the 8-fold downsampled feature map, and then passing the fine-grained information upward through the PAN structure to enhance the feature expression capability.

[0029] Preferably, in S2, the improved YOLOv5 network model loss function consists of three parts, namely, a binary cross entropy confidence loss function Lobj, a multivariate cross entropy classification loss function Lcls, and a border regression loss function Lbbox;

[0030] Among them, the binary cross entropy confidence loss function Lobj is expressed as:

[0031]

[0032] The multivariate cross entropy classification loss function Lcls is expressed as:

[0033]

[0034] The multivariate cross entropy classification loss function Lcls is used to determine whether the detected target belongs to a specific category, where: is the probability that the model predicts that the cell in the i-th row and j-th column belongs to category c;

[0035] The border regression loss function Lbbox is expressed as:

[0036]

[0037] The bounding box regression loss function Lbbox is used to accurately predict the location and size of the target, where: λcoord is the weight of the coordinate prediction, w ij and h ij It is the width and height of the target predicted by the model. CIoU is an improved IoU calculation method that takes into account the overlapping area, center point distance and aspect ratio of the bounding box.

[0038] Preferably, in S2, the data set is divided into a training set and a validation set in a ratio of 7:3, and the improved YOLOv5 model and the original YOLOv5n model are trained while ensuring that the training parameters of the model remain unchanged, and the harmonic mean value F is calculated through the accuracy P and recall rate R of the two models during the iteration process for comparative verification.

[0039] Preferably, the accuracy P is the ratio of the number of correctly predicted samples to the total number of positive samples, where Used to characterize the correct prediction ability of the model. The larger the value, the better the defect detection effect.

[0040] The recall rate R is the ratio of the correctly predicted results to all positive samples, where: It is used to characterize the ability of the model to detect defects. The larger the value, the better the defect detection effect.

[0041] P and R are mutually restrictive. F is introduced to characterize the comprehensive ability of constructing a prediction model with good P and R. F is defined as the harmonic average of P and R, where:

[0042] Preferably, in S3, after the user inputs the image to be optimized, the image enhancement processing, weld defect type identification, weld defect location and data saving operations are automatically implemented in sequence, thereby realizing batch intelligent identification of weld defects.

[0043] The pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network of the present invention has at least the following advantages:

[0044] 1. Aiming at the problems of complex defects, noise, scattering and other image interference of girth welds, Laplace filtering and CLAHE algorithm are used to enhance the original DR images, which effectively improves the signal-to-noise ratio of the image, improves the accuracy and stability of defect recognition, and builds a high-quality DR image dataset;

[0045] 2. The C2F module is introduced for multi-scale feature fusion, and an X-ray DR image defect detection algorithm based on the improved YOLOv5 network is proposed. The experimental results show that the improved model has achieved a high level of detection accuracy for various defect types (lack of fusion, lack of penetration, round defect, strip defect, etc.), with an average detection accuracy of 95.89% and a missed detection rate of 2.1%;

[0046] 3. This method can be used to intelligently identify defects in pipeline girth weld radiographs in batches, which is of great significance for improving production efficiency and ensuring product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a general flow chart of the pipeline girth weld radiographic film defect intelligent identification method based on the improved YOLOv5 network of the present invention;

[0048] Figure 2 It is a comparison diagram of defect image enhancement before and after of the pipeline girth weld radiographic film defect intelligent identification method based on the improved YOLOv5 network of the present invention;

[0049] Figure 3 It is a structural schematic diagram of the pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network of the present invention;

[0050] Figure 4 It is a comparison diagram of the overall F values ​​of the improved YOLOv5 and the original YOLOv5n models of the pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network of the present invention;

[0051] Figure 5 It is the improved YOLOv5 model detection result of the pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network of the present invention. DETAILED DESCRIPTION

[0052] The present invention is further described in detail below through examples so that those skilled in the art can implement the invention with reference to the description.

[0053] It should be understood that the terms such as “having”, “including” and “comprising” used herein do not exclude the existence or addition of one or more other elements or combinations thereof.

[0054] At present, for the radiographic film defect detection technology of pipeline girth welds, traditional radiographic detection technology has low efficiency and insufficient accuracy. Although modern automatic detection methods based on deep learning have improved the detection efficiency, they still have problems such as limited recognition ability of complex defects, low defect contrast, and irregular defect shapes. Therefore, this paper aims to propose a multi-scale defect detection method based on an improved YOLOv5 (You Only Look Once version5) network to solve the problems of traditional radiographic detection technology and modern automatic detection methods. By combining image enhancement and multi-scale feature extraction technology, an end-to-end defect detection model is constructed to realize intelligent recognition of weld defects, thereby improving the accuracy and stability of detection and meeting the needs of industrial production for fast and accurate detection of welding quality.

[0055] The pipeline girth weld radiographic defect intelligent recognition method based on the improved YOLOv5 network proposed in the present invention is specifically implemented as follows: Figure 1 As shown in the figure, it mainly includes three steps: step 1: construction of pipeline girth weld defect dataset; step 2: YOLOv5 model improvement and model verification; step 3: design and verification of weld defect intelligent identification system.

[0056] Step 1: Construction of pipeline girth weld defect dataset

[0057] The original experimental data of this paper comes from the DR images of the pressure pipeline ring weld provided by the Kunshan Branch of Jiangsu Special Inspection Institute, and the image expansion processing is performed to obtain 3120 weld defect images. Affected by various factors such as the structural noise of the imaging plate in the DR imaging system, the electrical noise of the high-voltage radiation source, the diffuse scattering effect, and the imaging environment, the image clarity is low, among which the circular defect is particularly unclear. In response to the above problems, this paper uses Laplace filtering and contrast limited adaptive histogram equalization algorithm (CLAHE) to enhance the original image. The effect of the processed circular defect is as follows: Figure 2 shown.

[0058] 1) Laplace filter algorithm

[0059] Laplace filtering is an edge detection algorithm based on second-order derivatives, which is used to emphasize edges and details in images. Laplace filtering uses the Laplace operator to calculate the second-order derivative of an image and identifies edges by detecting sharp changes in image brightness. The algorithm formula can be expressed as:

[0060]

[0061] 2) Contrast Limited Adaptive Histogram Equalization (CLAHE)

[0062] CLAHE is an algorithm for enhancing image contrast, especially for processing images with large local contrast differences. Unlike ordinary histogram equalization, CLAHE divides the image into small blocks (called "tiles"), applies histogram equalization to each small block separately, and limits the contrast of each small block to avoid noise amplification caused by over-enhancement. The algorithm formula can be expressed as:

[0063] I new (x, y) = CLAHE (I (x, y)) (2)

[0064] After processing, the clarity of image defects is significantly enhanced, and important detail features are retained. Labelimg software is used to classify and annotate the defects in the images. Among the 3120 weld defect images, the defect data set images are obtained: 330 cracks, 670 lack of fusion, 660 lack of penetration, 500 strip defects, 500 circular defects, 260 root concave, and 200 root bites.

[0065] Step 2: YOLOv5 model improvement and model verification

[0066] 1.YOLOv5 model improvements

[0067] This paper improves the YOLOv5 model by improving the CSPDarknet53 module, improving the Neck structure, and designing the loss function structure. The improved YOLOv5 network structure is as follows Figure 3 shown.

[0068] 1) The improved version of the YOL0v5 model uses the C2F (CSP Bottleneck with 2convolutions) module to enhance the extraction of high-order features and reduce model parameters, thereby replacing the traditional CSPDarknet53 module. The C2F module uses two layers of Bottleneck layers for feature extraction. Through multi-branch flow design and repeated use of features at different stages, it effectively improves the feature extraction capability and can capture more detailed information.

[0069] 2) To improve the detection capability of small objects, this study made key improvements to the Neck structure of YOLOv5n, especially by adding an additional feature layer to enhance the detection of small objects. On the basis of the original three layers of feature maps (8x, 16x, and 32x downsampling), an additional layer of feature maps (4x downsampling) was added. This layer of feature maps has a higher resolution and can capture more detailed information, which helps to improve the detection accuracy of small objects.

[0070] The additional feature layer is fused with features from other layers through the feature pyramid network FPN. Through top-down and bottom-up information transfer, the effective fusion of high-level semantic information and low-level spatial information is ensured. Specifically, based on FPN, the 4-fold downsampled feature map is fused with the 8-fold downsampled feature map, and then the fine-grained information is transferred upward through the PAN structure to further improve the feature expression capability.

[0071] 3) In order to accurately identify the specific category of the target and accurately predict the location and size of the target, the loss function of the improved YOLOv5 network model in this paper consists of three parts, namely the binary cross entropy confidence loss function Lobj, the multivariate cross entropy classification loss function Lcls and the bounding box regression loss function Lbbox. Specifically, as shown in formulas (1)-(3):

[0072]

[0073] The multivariate cross entropy classification loss function Lcls is used to determine whether the detected target belongs to a specific category. Among them: is the probability that the model predicts that the cell in row i and column j belongs to category c:

[0074]

[0075] The bounding box regression loss function Lbbox is used to accurately predict the location and size of the target:

[0076]

[0077] Where: λcoord is the weight of coordinate prediction, w ij and h ij It is the width and height of the target predicted by the model. CIoU (Complete Intersection over Union) is an improved IoU calculation method that takes into account the overlapping area, center point distance and aspect ratio of the bounding box.

[0078] 2. Model training and validation

[0079] The dataset is divided into a training set and a validation set in a ratio of 7:3. The improved YOLOv5 model and the original YOLOv5n model are trained while keeping the training parameters of the model unchanged. The harmonic mean value is calculated through the accuracy and recall rate indicators of the two models during the iteration process for comparison and verification. The F value comparison between the two is as follows: Figure 4 shown.

[0080] Accuracy P is the ratio of the number of correctly predicted samples to the total number of positive samples, as shown in formula (4). It is mainly used to characterize the correct prediction ability of the model. The larger the value, the better the defect detection effect.

[0081]

[0082] The recall rate R is the ratio of the correctly predicted results to all positive samples, as shown in formula (5). It is mainly used to characterize the ability of the model to detect defects. The larger the value, the better the defect detection effect.

[0083]

[0084] P and R are mutually constrained. The introduction of F is mainly used to characterize the comprehensive ability of constructing a prediction model with good P and R. The definition of F is the harmonic average of P and R, as shown in formula (6):

[0085]

[0086] Step 3: Design and verification of intelligent weld defect identification system

[0087] During the image processing, after the user inputs the image to be optimized, the image enhancement processing, weld defect type identification, weld defect location and data saving operations are automatically implemented in sequence, realizing batch intelligent identification of weld defects. The image data set is input to the weld defect identification system for batch processing. Some of the processing results are shown in the figure below. Figure 5 shown.

[0088] Judging from the processing results, the pipeline girth weld radiographic film defect intelligent recognition method based on the improved YOL0v5 network of the present invention constructs a high-quality DR image data set, effectively improves the signal-to-noise ratio of the image, and improves the accuracy and stability of defect recognition. The detection accuracy of various defect types (lack of fusion, lack of penetration, round defects and strip defects, etc.) has reached a high level, with an average detection accuracy of 95.89% and a missed detection rate of 2.1%. The method can perform intelligent defect recognition of pipeline girth weld radiographic films in batches, which is of great significance for improving production and quality inspection efficiency and ensuring product quality.

[0089] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the embodiments shown and described herein.

Claims

1. An intelligent identification method for pipeline girth weld radiographic defects based on an improved YOLOv5 network, characterized in that: At least the following steps are included: S1. Construction of pipeline girth weld defect dataset: S1.

1. Digitally image the girth weld area of ​​the pressure pipeline using the DR digital imaging system; S1.

2. Perform image enhancement processing on the original image by using a denoising algorithm and a contrast adjustment algorithm; S1.

3. Use Labelimg software to annotate at least seven weld defects in the image, including cracks, lack of fusion, lack of penetration, strip defects, circular defects, root concave and root bite, and construct a data set; S2.YOLOv5 model improvement and model verification: S2.

1. The C2F module is used to replace the CSPDarknet53 module of the traditional YOLOv5 model framework to enhance the extraction capability of high-order features and reduce model parameters to improve feature extraction capability and capture more detailed information; S2.

2. Improve the Neck part of the feature fusion module of the YOLOv5 model framework by adding an additional feature layer and fusing it with features from other layers through the feature pyramid network FPN to enhance the detection capability of small objects; S2.

3. Design an improved YOLOv5 network model loss function, combining the binary cross entropy confidence loss function Lobj, the multivariate cross entropy classification loss function Lcls, and the bounding box regression loss function Lbbox to achieve the evaluation of the specific category, position, and size of the target; S2.

4. The dataset is divided into a training set and a validation set in a ratio of 7:3, and the original YOLOv5n model and the improved YOLOv5 model are trained separately; S2.

5. Experimentally verify the trained object detection model on crack, incomplete penetration, and circular defect detection of pipeline girth weld DR image dataset, and evaluate the effect of the improved YOLOv5 model; S3. Design and verification of intelligent identification system for weld defects: S3.

1. Design an intelligent identification system for pipeline annular seam radiographic film defects, which can automatically implement image enhancement processing, weld defect type identification, weld defect location and data storage operations in sequence by importing pictures; S3.

2. Analyze the effect of the image optimization system through the dataset images; In S2, the improved YOLOv5 network model loss function consists of three parts, namely, a binary cross entropy confidence loss function Lobj, a multivariate cross entropy classification loss function Lcls, and a border regression loss function Lbbox; Among them, the binary cross entropy confidence loss function Lobj is expressed as: The multivariate cross entropy classification loss function Lcls is expressed as: ； The multivariate cross entropy classification loss function Lcls is used to determine whether the detected target belongs to a specific category, where: is the probability that the model predicts that the cell in the i-th row and j-th column belongs to category c; The border regression loss function Lbbox is expressed as: ； The bounding box regression loss function Lbbox is used to accurately predict the location and size of the target, where: λcoord is the weight of the coordinate prediction, w ij and h ij It is the width and height of the target predicted by the model. CIoU is an improved IoU calculation method that takes into account the overlapping area, center point distance and aspect ratio of the bounding box. In S2, the data set is divided into a training set and a validation set in a ratio of 7:

3. The improved YOLOv5 model and the original YOLOv5n model are trained while keeping the training parameters of the model unchanged. The harmonic mean F is calculated through the accuracy P and recall R of the two models during the iteration process for comparison and verification. Accuracy P is the ratio of the number of correctly predicted samples to the total number of positive samples, where , which is used to characterize the correct prediction ability of the model. The larger the value, the better the defect detection effect; The recall rate R is the ratio of the correctly predicted results to all positive samples, where: , used to characterize the model's ability to detect defects. The larger the value, the better the defect detection effect. P and R are mutually restrictive. F is introduced to characterize the comprehensive ability of constructing a prediction model with good P and R. F is defined as the harmonic average of P and R, where: .

2. The pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network according to claim 1 is characterized in that: In S1, the image clarity is low due to the structural noise of the imaging plate in the DR imaging system, the electrical noise of the high-voltage radiation source, the diffuse scattering effect and the imaging environment, among which the circular defect is particularly unclear. The Laplace filter and the contrast-limited adaptive histogram equalization algorithm are used to enhance the original image, which significantly enhances the clarity of the image defects while retaining important detail features.

3. The pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network according to claim 2 is characterized in that: The Laplace filtering algorithm formula is: ； The formula of the contrast-limited adaptive histogram equalization algorithm is: 。 4. The pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network according to claim 1 is characterized in that: In S2, the C2F module uses two Bottleneck layers for feature extraction, and through multi-branch flow design and reuse of features at different stages, it improves the feature extraction capability and captures more detailed information.

5. The pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network according to claim 4 is characterized in that: In S2, the additional feature layer is an additional 4-fold down-sampling feature map on the basis of the original three-layer feature map obtained by 8-fold, 16-fold, and 32-fold down-sampling.

6. The pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network according to claim 5 is characterized in that: The additional feature layer is fused with features of other levels through the feature pyramid network FPN, and the effective fusion of high-level semantic information and low-level spatial information is ensured through top-down and bottom-up information transmission.

7. The pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network according to claim 6 is characterized in that: The method for fusing the FPN with features of other levels is to fuse the 4-fold downsampled feature map with the 8-fold downsampled feature map on the basis of the FPN, and then pass the fine-grained information upward through the PAN structure to improve the feature expression capability.

8. The pipeline girth weld radiographic defect intelligent identification method based on the improved YOLOv5 network according to claim 1 is characterized in that: In S3, after the user inputs the image to be optimized, the image enhancement processing, weld defect type identification, weld defect location and data saving operations are automatically implemented in sequence, thereby realizing batch intelligent identification of weld defects.

Images