Pressure-bearing special equipment weld defect identification method and device and computer equipment

Abstract

The application provides a pressure type special equipment weld defect identification method and device and computer equipment. The method comprises the following steps: based on a pre-established boiler detection database, calling historical boiler detection images, boiler defect images and non-defect boiler images, and constructing a defect classification dataset according to the weld type and the weld defect type; performing data expansion and labeling on the defect classification dataset to obtain a sample set, and dividing the sample set into a training set and a test set; introducing a convolutional neural network based on a YOLOv8 algorithm to establish an initial identification model of different types of weld defects; inputting the training set to train the initial identification model of different types of weld defects, and testing the model through the test set to obtain a trained YOLOv8 defect identification model; in this way, the problem of inaccurate detection caused by incomplete cleaning during existing weld detection can be improved.

Description

Technical Field

[0001] This invention relates to the field of boiler safety management technology, and more specifically, to a method, device, and computer equipment for identifying weld defects in pressure-bearing special equipment. Background Technology

[0002] The overall structure of a boiler consists of two main parts: the boiler body and auxiliary equipment. The boiler body comprises a steam-water system and a combustion system. The steam-water system includes an economizer, steam drum, downcomer, headers, water-cooled walls, superheater, and reheater. Its main function is to effectively absorb the heat released from fuel combustion and heat the feedwater entering the boiler to form superheated steam with a specific temperature and pressure. The combustion system consists of the furnace, flue, burner, and air preheater. Its main function is to ensure proper combustion of fuel within the boiler, releasing heat.

[0003] Boiler welds are prone to cracking and lack of fusion due to long-term cyclic loads and thermal stress, which can eventually lead to structural fatigue failure. Therefore, regular or irregular, effective inspections of boiler welds are necessary to ensure safe, reliable, and stable boiler operation.

[0004] However, due to the large size of the boiler, it is not possible for staff to conduct a comprehensive inspection of the weld seams on the boiler surface. Before the inspection, the large size of the boiler also makes it difficult for inspectors to clean impurities from the boiler surface, resulting in inaccurate inspections. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method, device and computer equipment for identifying weld defects in pressure-bearing special equipment, which can solve the problem that the large size of the boiler makes it difficult for inspectors to thoroughly clean the surface of the boiler before inspection, thus leading to inaccurate inspection.

[0006] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0007] In a first aspect, embodiments of this application provide a method for identifying weld defects in pressure-bearing special equipment, applied to boiler weld defect identification, the method comprising:

[0008] Based on a pre-established boiler inspection database, historical boiler inspection images, boiler defect images, and defect-free boiler images are retrieved, and a defect classification dataset is constructed according to weld type and welding defect type.

[0009] The defect classification dataset is augmented and labeled to obtain a sample set, which is then divided into a training set and a test set.

[0010] Based on the YOLOv8 algorithm, a convolutional neural network is introduced to establish an initial identification model for different types of welding defects;

[0011] The training set is used to train initial identification models for different types of welding defects, and the model is tested using the test set to obtain a trained YOLOv8 defect identification model.

[0012] Acquire multi-source image data of all target recognition areas on the boiler and perform image preprocessing; wherein, the multi-source images include one or more of infrared images, UT images, RT images, PAUT images, ACFM images and TOFD images;

[0013] The multi-source image data is input into the YOLOv8 defect recognition model, and the recognition results of boiler weld type and welding defect type are output.

[0014] In conjunction with the first aspect, in some optional implementations, after augmenting and labeling the defect classification dataset to obtain a sample set, the method further includes:

[0015] Based on the boiler detection database, obtain the anchor frame of the target identification area;

[0016] Using the center of the anchor frame as a reference point, a settling trend model of the attachments on the boiler surface is established. Based on the settling trend model, the weld trace covering parameters are inverted, and foreign object interference features are generated based on the weld trace covering parameters.

[0017] Foreign object noise is removed from boiler defect images based on the aforementioned foreign object interference characteristics;

[0018] Image enhancement processing is performed on boiler defect images after foreign object noise removal.

[0019] In conjunction with the first aspect, in some optional implementations, the data augmentation and annotation of the defect classification dataset to obtain a sample set includes:

[0020] The boiler defect images in the defect classification dataset are cropped to obtain defect region images;

[0021] The defective region image is augmented using the OpenCV algorithm to generate multiple augmented images;

[0022] The augmented image and the original boiler defect image are used together as the dataset for YOLOv8 algorithm feature annotation to obtain a sample set.

[0023] In conjunction with the first aspect, in some optional embodiments, the method further includes: setting an image fusion model, the image fusion model including an encoding network, a fusion network and a decoding network;

[0024] The coding network consists of n groups of hybrid architecture modules of different sizes, where n is a natural number.

[0025] The fusion network consists of feature fusion layers, where the number of feature fusion layers is the same as the number of groups in the hybrid architecture module;

[0026] The upsampling settings are configured using a hybrid architecture module to build the decoding network;

[0027] The encoding network, fusion network, and decoding network are connected; a 1*1 convolutional layer is set at the input of the encoding network and a 1*1 convolutional layer is set at the output of the decoding network to form the overall network architecture of the image fusion model.

[0028] In conjunction with the first aspect, in some optional embodiments, the method further includes: fusing multi-source images using a pre-defined image fusion model to obtain a fused image; specifically including:

[0029] The global features of the i-th source image are calculated using the softmax function. Attention weights Global features of the j-th source image Attention weights ; where the i-th source image is one of the multiple source images, and the j-th source image is another of the multiple source images;

[0030] Global features of the i-th source image With attention weight Multiplication, global features of the j-th source image With attention weight Multiply and add them together to obtain the fused features. ;

[0031] The fused features are then subjected to multi-scale decoding and convolution operations to obtain the final fused image.

[0032] In conjunction with the first aspect, in some optional implementations, the hybrid architecture module includes a two-layer network consisting of a normalization layer, a window self-attention mechanism, a residual connection layer, and a multilayer perceptron; the window self-attention mechanism of the two-layer network can perform information interaction between different windows to achieve global modeling of the entire image.

[0033] In conjunction with the first aspect, in some optional implementations, the introduction of a convolutional neural network based on the YOLOv8 algorithm to establish an initial identification model for different types of welding defects includes:

[0034] Convolutional neural networks are constructed using spatial-to-depth layers and stride convolutional layers.

[0035] The YOLOv8 algorithm was selected and introduced into the constructed convolutional neural network;

[0036] The initial recognition model is obtained by adding coordinate attention to the Neck layer of the YOLOv8 algorithm.

[0037] In conjunction with the first aspect, in some optional embodiments, acquiring multi-source image data of all target identification areas on the boiler and performing image preprocessing includes:

[0038] Obtain a three-dimensional model of the boiler, and then calibrate the target recognition area of ​​the three-dimensional model;

[0039] Based on the calibrated target recognition region coordinates, acquire multi-source image data of all target recognition regions;

[0040] Each source image in the multi-source image is cropped, denoised, and aligned in terms of spatiotemporal sequence.

[0041] Secondly, embodiments of this application provide a weld defect identification device for pressure-bearing special equipment, applied to the method described, including:

[0042] The data acquisition module is used to retrieve historical boiler inspection images, boiler defect images, and defect-free boiler images based on a pre-established boiler inspection database, and to construct a defect classification dataset according to weld type and welding defect type.

[0043] The data preprocessing module is used to augment and label the defect classification dataset to obtain a sample set, and then divide the sample set into a training set and a test set.

[0044] The model building module is used to introduce a convolutional neural network based on the YOLOv8 algorithm to build an initial identification model for different types of welding defects.

[0045] The model training module is used to train the initial identification model for different types of welding defects by inputting the training set, and to test the model by using the test set to obtain the trained YOLOv8 defect identification model.

[0046] The real-time image acquisition module is used to acquire multi-source image data of all target recognition areas on the boiler and perform image preprocessing; wherein, the multi-source images include one or more of infrared images, UT images, RT images, PAUT images, ACFM images and TOFD images;

[0047] The defect classification module is used to input the multi-source image data into the YOLOv8 defect recognition model and output the recognition results of boiler weld type and welding defect type.

[0048] Thirdly, embodiments of this application provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, causing the processor to perform the method described.

[0049] The invention employing the above technical solution has the following advantages:

[0050] In the technical solution provided in this application, based on a pre-established boiler inspection database, historical boiler inspection images, boiler defect images, and defect-free boiler images are called, and a defect classification dataset is constructed according to weld type and welding defect type; this provides basic data for model training, ensures that the dataset covers different weld types and defect types, and improves the generalization ability of the model;

[0051] The defect classification dataset is augmented and labeled to obtain a sample set, which is then divided into a training set and a test set. The data augmentation increases the number and diversity of samples, and the labeling clarifies the defect characteristics, providing high-quality samples for model training and avoiding model overfitting.

[0052] Based on the YOLOv8 algorithm, a convolutional neural network is introduced to establish an initial identification model for different types of welding defects. By introducing a convolutional neural network to improve feature extraction capabilities, coordinate attention is added to the Neck layer of the YOLOv8 algorithm to further optimize the model's identification accuracy for weld defects and solve the problem of inaccurate identification of small defects by the traditional YOLOv8 model.

[0053] By acquiring multi-source images and integrating the advantages of different images (such as infrared images that can identify temperature anomalies and UT images that can identify internal defects), image interference is eliminated and image quality is improved through preprocessing (format normalization, noise reduction, time and space alignment, etc.).

[0054] The multi-source image data is input into the YOLOv8 defect recognition model, which outputs the recognition results of boiler weld type and welding defect type. To further improve recognition accuracy, the multi-source images can be fused using a pre-defined image fusion model. The fused image is then input into the model for recognition, achieving complementary multi-source information and further improving the accuracy of defect recognition. This approach solves the problem that the large size of boilers makes it difficult for inspectors to thoroughly clean the boiler surface before inspection, leading to inaccurate detection. It also addresses the difficulty of fully covering boiler welds during manual inspection, improving the comprehensiveness, accuracy, and efficiency of weld defect recognition. Attached Figure Description

[0055] This application can be further illustrated by the non-limiting embodiments given in the accompanying drawings. It should be understood that the following drawings only illustrate some embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without any inventive effort.

[0056] Figure 1 A flowchart illustrating a method for identifying weld defects in pressure-bearing special equipment, provided in an embodiment of this application;

[0057] Figure 2 A structural block diagram of a weld defect identification device for pressure-bearing special equipment provided in this application embodiment;

[0058] Figure 3 A flowchart illustrating the process of identifying weld defects in pressure-bearing special equipment using a fused image, as provided in this application embodiment.

[0059] Figure 4 This is a structural block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0060] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts are referred to by the same reference numerals in the drawings or description. Implementations not shown or described in the drawings are forms known to those skilled in the art. In the description of this application, terms such as "first" and "second" are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0061] Please refer to Figure 4 This application provides a computer device that may include a processing module and a storage module. The storage module stores a computer program, which, when executed by the processing module, enables the computer device to perform corresponding steps in the task processing method described below. The processing module may include a CPU and a GPU, which can cooperate to achieve collaborative task processing.

[0062] In this embodiment, the processing module can be an integrated circuit chip with signal processing capabilities. The processing module can be a general-purpose processor. For example, the processor can be a central processing unit, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0063] The storage module can be, but is not limited to, random access memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, etc. In this embodiment, the storage module can be used to store datasets, multiple algorithm models, boiler detection databases, etc. Of course, the storage module can also be used to store programs, which the processing module executes after receiving execution instructions.

[0064] Please refer to Figure 1 This application also provides a method for identifying weld defects in pressure-bearing special equipment. This method can be applied to the aforementioned computer equipment or to portable boiler inspection tools, such as portable laptops. The steps of the method are executed or implemented by the computer equipment, and the method may include the following steps 110 to 160:

[0065] Step 110: Based on the pre-established boiler inspection database, call up historical boiler inspection images, boiler defect images, and defect-free boiler images, and construct a defect classification dataset according to weld type and welding defect type;

[0066] Step 120: Augment and label the defect classification dataset to obtain a sample set, and divide the sample set into a training set and a test set;

[0067] Step 130: Based on the YOLOv8 algorithm, introduce a convolutional neural network to establish an initial identification model for different types of welding defects;

[0068] Step 140: Input the training set to train the initial identification model for different types of welding defects, and test the model using the test set to obtain the trained YOLOv8 defect identification model;

[0069] Step 150: Acquire multi-source image data of all target recognition areas on the boiler and perform image preprocessing; wherein, the multi-source images include one or more of infrared images, UT images, RT images, PAUT images, ACFM images and TOFD images;

[0070] Step 160: Input the multi-source image data into the YOLOv8 defect recognition model and output the recognition results of boiler weld type and welding defect type.

[0071] The method in this embodiment establishes a boiler inspection database to provide basic data for model training, ensuring that the dataset covers different weld types and defect types, thereby improving the model's generalization ability. Data expansion increases the number and diversity of samples, and annotation clarifies defect features, providing high-quality samples for model training and avoiding overfitting. A convolutional neural network is introduced based on the YOLOv8 algorithm to establish initial recognition models for different types of welding defects. The introduction of a convolutional neural network enhances feature extraction capabilities, and coordinate attention is added to the Neck layer of the YOLOv8 algorithm to further optimize the model's accuracy in identifying weld defects, solving the problem of inaccurate identification of small defects in traditional YOLOv8 models. By acquiring multi-source images and integrating the advantages of different images (e.g., infrared images can identify temperature anomaly defects, and UT images can identify internal defects), image interference is eliminated through preprocessing (format normalization, noise reduction, and temporal and spatial alignment), improving image quality. The multi-source image data is input into the YOLOv8 defect recognition model, which outputs the recognition results of boiler weld type and welding defect type. If it is necessary to further improve the recognition accuracy, the multi-source images can be fused by the set image fusion model. The fused image is then input into the model for recognition, so as to realize the complementarity of multi-source information and further improve the accuracy of defect recognition.

[0072] The steps of the method will be explained in detail below:

[0073] In step 110, a boiler inspection database is pre-established. This database stores historical inspection images, defective images (including common welding defects such as cracks, lack of fusion, and porosity), and defect-free images of boilers of different models and operating years. It also stores annotation information for weld types (such as butt welds, fillet welds, and T-welds) and welding defect types. It should be noted that for different boiler inspection tasks, the database retrieval range can be limited according to the boiler model to improve data acquisition efficiency.

[0074] For example, images of crack defects in butt welds, images of non-fusion defects in fillet welds, and images of defect-free T-welds can be categorized separately to ensure that the dataset covers all types of common welds and defects, providing comprehensive data support for subsequent model training. Understandably, the target recognition region may contain only one of the following: butt weld, fillet weld, T-weld, etc. The classification can be flexibly set according to the actual situation, and no specific limitation is made here.

[0075] For example, boiler inspection images, boiler defect images, and defect-free boiler images are all digital images, and can be one of the following: infrared images, UT images, RT images, PAUT images, ACFM images, and TOFD images.

[0076] In step 120, the sample set is divided into a training set and a test set in a 7:3 ratio (the ratio can be adjusted according to actual needs).

[0077] In this embodiment, the defect classification dataset is augmented and labeled to obtain a sample set, including:

[0078] The boiler defect images in the defect classification dataset are cropped to obtain defect region images;

[0079] The defective region image is augmented using the OpenCV algorithm to generate multiple augmented images;

[0080] The augmented image and the original boiler defect image are used together as the dataset for YOLOv8 algorithm feature annotation to obtain a sample set.

[0081] For example, during data augmentation: the boiler defect images in the defect classification dataset are cropped, with the cropping range being the defect area and its surrounding 50-100 pixels, retaining the core features of the defect and removing irrelevant background areas;

[0082] For the cropped defective images, image augmentation is performed using the OpenCV algorithm. Specific methods include: rotation (any angle from 0 to 90°), flipping (horizontal flipping, vertical flipping), scaling (0.8-1.2 times), and brightness adjustment (±20%). Each augmentation method generates 5-10 augmented images to increase the number and diversity of samples.

[0083] The amplified image is stitched together with the original defective and defect-free images in a random stitching manner (to avoid clustering of similar images) to ultimately achieve data augmentation and ensure that the number of samples meets the model training requirements (at least 1000 samples).

[0084] When labeling data:

[0085] The amplified defect images and the original boiler defect images were used together as the dataset for the YOLOv8 algorithm. The LabelImg annotation tool was used to annotate the weld type and welding defect type. The annotation format was txt format supported by the YOLOv8 algorithm, which clearly defined the coordinate location and category of the defects.

[0086] The YOLOv8 algorithm is initially trained using an annotated dataset, with 100-200 training iterations. Then, it is tested using a test set. If the test accuracy is ≥90%, a qualified sample set is determined. If the test accuracy is <90%, the number of samples is increased, the annotation accuracy is optimized, and the annotation and testing are repeated until the requirements are met.

[0087] In some implementation scenarios of this embodiment, after augmenting and labeling the defect classification dataset to obtain a sample set, in order to eliminate the interference of boiler surface deposits on weld defect identification and improve image quality, the method further includes:

[0088] Based on the boiler detection database, the anchor frame of the target identification area is determined according to the weld size and imaging resolution of the target identification area; the anchor frame includes the anchor frame size and position coordinates, which are used to locate the area where the weld is located;

[0089] Using the center of the anchor frame as a reference point, and combining historical boiler operating environment and attachment distribution data, a settling trend model for attachments on the boiler surface is established. The settling trend model is used to characterize the correspondence between the deposition thickness of the attachments and the distance from the center of the anchor frame. Based on the settling trend model, a reverse deduction is performed to obtain weld trace coverage parameters. The coverage parameters include coverage area and coverage degree, which are used to quantify the degree to which the attachments obstruct the weld defect area.

[0090] Foreign object interference features are generated based on the weld trace masking parameters; these foreign object interference features are used to distinguish between actual weld defects and surface deposits.

[0091] Based on the foreign object interference characteristics, a threshold determination method is used to remove foreign object noise from boiler defect images, eliminating image interference caused by attachments such as oil, scale, and rust.

[0092] Histogram equalization algorithm is used to enhance the image of boiler defects after foreign object noise removal, thereby improving the contrast between the defect area and the background and highlighting the contour and texture features of the weld defect, which is convenient for subsequent model recognition.

[0093] The settling trend model can be obtained by fitting historical foreign object image data. The value of the occlusion degree ranges from 0 to 1, where 0 represents no occlusion and 1 represents complete occlusion. When removing foreign object noise, areas with an occlusion degree greater than a preset threshold are identified as foreign object interference and removed.

[0094] For example, the anchor frame size is set according to the actual size of the target recognition area (e.g., 100×100 pixels).

[0095] The mathematical formula for the settlement trend model is:

[0096] (1),

[0097] Where y is the thickness of the deposited material, x is the distance from the center of the anchor frame, and k and b are model parameters that can be obtained by fitting historical foreign object image data in the database.

[0098] Based on the settlement trend model, the weld trace coverage parameters are inverted (reverse deduction). The coverage parameters include the coverage area and the coverage degree. The coverage degree ranges from 0 to 1, where 0 indicates no coverage and 1 indicates complete coverage.

[0099] Based on the weld trace coverage parameters, foreign object features are generated. The feature dimensions are 3-dimensional (coverage area, coverage degree, distance from the center of the anchor frame), which are used to distinguish foreign objects from weld defects.

[0100] Based on the characteristics of foreign objects, a threshold method (with a threshold set to 0.3, areas with an occlusion degree > 0.3 are identified as foreign objects) is used to clean up noise in boiler defect images and remove interference from foreign objects.

[0101] In addition, histogram equalization algorithms can be used to enhance the image of boiler defects after noise removal, improve image contrast, highlight defect features, and facilitate subsequent model recognition.

[0102] In some embodiments of this example, after completing foreign object noise removal and image enhancement processing, to further address the problems of low efficiency, non-standard repair, and difficulty in quality control during manual cleaning of boiler welds, this example provides an automatic foreign object cleaning and intelligent weld repair process in conjunction with foreign object identification. The specific steps are as follows:

[0103] Based on the obtained foreign object interference characteristics and weld trace coverage parameters, foreign objects on the boiler surface are classified. The types of foreign objects include oil stains, scale, rust, welding slag, and dust accumulation. At the same time, according to the coverage degree, foreign object interference is divided into three levels: a coverage degree greater than 0.7 is considered severe interference, a coverage degree between 0.3 and 0.7 is considered moderate interference, and a coverage degree less than or equal to 0.3 is considered mild interference.

[0104] Based on the anchor frame parameters and settlement trend model, the center of the anchor frame is taken as the starting point for cleaning, and a spiral or straight adaptive cleaning path is generated from the periphery to the center of the weld along the direction of decreasing thickness of the attached material. The cleaning power and action time are increased for areas with severe interference, while conventional cleaning parameters are used for areas with moderate and mild interference, so as to ensure the cleaning effect while avoiding damage to the boiler base material and weld body.

[0105] Cleaning is performed using non-contact or micro-contact methods, including at least one combination of high-pressure nitrogen blowing to remove floating dust and loose debris, laser removal of stubborn rust and scale, mechanical micro-grinding to treat protruding foreign objects, and ultrasonic removal of dense scale, adapting to different working conditions and types of foreign objects at the boiler site.

[0106] After cleaning is completed, multi-source image data of the corresponding area are collected in real time, the occlusion parameters are recalculated and input into the YOLOv8 defect recognition model; when the foreign object occlusion degree drops to below 0.1 and no false defects are generated, the cleaning is judged to be qualified; if it is unqualified, the unqualified area is automatically located and a second local cleaning is performed.

[0107] Based on the defect identification results, weld defects are classified and repair strategies are matched accordingly. Repair strategies: micro defects such as porosity and slag inclusions are repaired by filling and welding; defects that are not fused are repaired by grinding and leveling and then remelting and welding; major defects such as cracks are repaired by defect removal, weld reconstruction and post-weld heat treatment to ensure that the strength and sealing of the repaired weld meet the requirements of the pressure equipment.

[0108] After the weld area has been repaired, multi-source image acquisition and defect identification are performed again. If the defect is qualified, the detection data, cleaning record and repair parameters are stored synchronously in the boiler detection database. If the defect is not qualified, the defect location is relocated and the repair operation is performed until the standard is met, forming a closed-loop control of the entire process of identification, cleaning, repair and verification.

[0109] This embodiment achieves adaptive cleaning and graded repair by combining foreign object interference characteristics with a settlement trend model, forming an integrated closed loop of weld defect identification, cleaning, repair, and verification, further improving the automation level and reliability of weld inspection and repair for pressure-bearing special equipment.

[0110] In some embodiments of this example, the step of introducing a convolutional neural network based on the YOLOv8 algorithm to establish an initial identification model for different types of welding defects includes:

[0111] A convolutional neural network is constructed using spatial-to-depth layers and stride convolutional layers. The spatial-to-depth layers are used to convert the spatial features of an image into depth features, while the stride convolutional layers have a kernel size of 3×3 and a stride of 2, and are used to extract deep features of the image.

[0112] The YOLOv8 algorithm was selected and introduced into the constructed convolutional neural network;

[0113] The initial recognition model is obtained by adding coordinate attention to the Neck layer of the YOLOv8 algorithm.

[0114] Among them, the YOLOv8 algorithm can adopt the YOLOv8n lightweight model to adapt to the real-time detection needs of industrial sites; the constructed convolutional neural network is introduced into the Backbone layer of the YOLOv8 algorithm to replace the feature extraction module of the original Backbone layer and improve the feature extraction capability.

[0115] A coordinate attention module (or coordinate attention module) is added to the Neck layer of the YOLOv8 algorithm. The attention dimension of the coordinate attention module is 16, and the activation function is the Sigmoid function. This is used to enhance the model's sensitivity to the location of weld defects and improve the recognition accuracy of small defects. The parameters of the initial recognition model are initialized, with the learning rate set to 0.001 and the weight decay set to 0.0001, thus completing the establishment of the initial recognition model.

[0116] In some implementations of this embodiment, the training set is input into the initial recognition model for model training. During the training process, the stochastic gradient descent (SGD) optimizer is used, the loss function is the CIoU loss function, and the number of training iterations is 300-500 times until the model converges (loss value < 0.1).

[0117] After training, the test set is input into the trained model for testing. The test metrics include recognition accuracy, recall, and F1 score. If the recognition accuracy is ≥95%, the recall is ≥90%, and the F1 score is ≥92%, then the trained YOLOv8 defect recognition model is obtained. If the metrics are not met, the model parameters (such as learning rate and number of iterations) are adjusted, and the model is retrained and tested until the requirements are met.

[0118] In some optional implementations, the acquisition of multi-source image data of all target identification areas on the boiler and the image preprocessing include:

[0119] Obtain a three-dimensional model of the boiler, and then calibrate the target recognition area of ​​the three-dimensional model;

[0120] Based on the calibrated target recognition region coordinates, acquire multi-source image data of all target recognition regions;

[0121] Each source image in the multi-source image is cropped, denoised, and aligned in terms of spatiotemporal sequence.

[0122] The boiler can be comprehensively scanned using laser scanning equipment to obtain a 3D model. The accuracy of the 3D model is set to ±1mm to ensure the accurate positioning of target components. Infrared images of the target identification area of ​​the boiler are obtained using an infrared imager; UT images, RT images, PAUT images, ACFM images, and TOFD images can be obtained using corresponding ultrasonic flaw detectors, X-ray flaw detectors, phased array ultrasonic flaw detectors, etc. It is understood that not all of the above instruments are necessarily required for boiler defect detection; flexible selection can be made according to actual needs.

[0123] Of course, multi-source image data of all target components can be collected simultaneously at a frequency of 1 frame / second and an image resolution of 1920×1080.

[0124] In practice, noise reduction can be achieved using a Gaussian filtering algorithm with a filter kernel size of 5×5; for spatiotemporal sequence alignment, based on the coordinates of the target recognition region, an image registration algorithm is used to align the positions of target components in different source images, with the time error controlled within 0.1 seconds.

[0125] In some embodiments of this example, the method further includes: setting an image fusion model, wherein the image fusion model includes an encoding network, a fusion network, and a decoding network;

[0126] The encoding network consists of n sets of hybrid architecture modules of different sizes, where n is a natural number, for example, n can be 3; the sizes of the three sets of hybrid architecture modules are 3×3, 5×5 and 7×7, respectively, and are used to extract image features at different scales.

[0127] The fusion network consists of feature fusion layers, where the number of feature fusion layers is the same as the number of groups in the hybrid architecture module. The fusion network consists of 3 feature fusion layers and uses a softmax attention fusion mechanism to fuse features from different source images.

[0128] The upsampling settings are used to construct the decoding network using a hybrid architecture module; the upsampling factor is 2x, which is used to restore the fused features to a fused image with the same size as the original image.

[0129] The encoding network, fusion network, and decoding network are connected; a 1×1 convolutional layer (64 output channels) is set at the input of the encoding network and a 1×1 convolutional layer (3 output channels) is set at the output of the decoding network to form the overall network architecture of the image fusion model.

[0130] In some embodiments, the method further includes: fusing multi-source images using a pre-defined image fusion model to obtain a fused image; specifically including:

[0131] The global features of the i-th source image are calculated using the softmax function. Attention weights Global features of the j-th source image Attention weights ; where the i-th source image is one of the multiple source images, and the j-th source image is another of the multiple source images;

[0132] Global features of the i-th source image With attention weight Multiplication, global features of the j-th source image With attention weight Multiply and add them together to obtain the fused features. ;

[0133] The fused features are then subjected to multi-scale decoding (e.g., 3-layer decoding, corresponding to 3 sets of hybrid architecture modules in the encoding network) and convolution operations (3×3 kernel size) to obtain the final fused image. The fused image integrates the advantageous features of multiple source images, improving the accuracy of defect recognition.

[0134] The hybrid architecture module includes a two-layer network consisting of a normalization layer, a window self-attention mechanism, a residual connection layer, and a multilayer perceptron. The window self-attention mechanism of the two-layer network can perform information interaction between different windows to achieve global modeling of the entire image.

[0135] The window self-attention mechanism of the two-layer network uses an 8×8 window size, which allows for information interaction between different windows, enabling global modeling of the entire image and improving the comprehensiveness of feature extraction.

[0136] Please refer to Figure 3 The preprocessed multi-source image data (or fused images) is input into the trained YOLOv8 defect recognition model. The model extracts features and matches defects, and outputs the recognition results of boiler weld types and welding defect types. The recognition results include: weld type (such as butt weld, fillet weld), defect type (such as crack, lack of fusion, porosity), defect location coordinates, and defect size. The recognition results are stored in the boiler inspection database, and an inspection report is generated for staff to view and process.

[0137] In another embodiment, a weld defect identification device for pressure-bearing special equipment is provided, applied to the method described, including:

[0138] The data acquisition module 210 is used to call historical boiler inspection images, boiler defect images and defect-free boiler images based on a pre-established boiler inspection database, and to construct a defect classification dataset according to weld type and welding defect type.

[0139] Data preprocessing module 220 is used to augment and label the defect classification dataset to obtain a sample set, and divide the sample set into a training set and a test set.

[0140] The model building module 230 is used to introduce a convolutional neural network based on the YOLOv8 algorithm to build an initial identification model for different types of welding defects.

[0141] The model training module 240 is used to train the initial identification model for different types of welding defects by inputting the training set, and to test the model by using the test set to obtain the trained YOLOv8 defect identification model.

[0142] The real-time image acquisition module 250 is used to acquire multi-source image data of all target recognition areas on the boiler and perform image preprocessing; wherein, the multi-source images include one or more of infrared images, UT images, RT images, PAUT images, ACFM images and TOFD images;

[0143] The defect classification module 260 is used to input the multi-source image data into the YOLOv8 defect recognition model and output the recognition results of boiler weld type and welding defect type.

[0144] In one implementation of this embodiment, the data acquisition module 210 can adopt a database interface module. Based on a pre-established boiler inspection database, it calls historical boiler inspection images, boiler defect images, and defect-free boiler images, and constructs a defect classification dataset according to weld type and welding defect type. This module can realize database query and data call through SQL statements to ensure the efficiency and accuracy of data acquisition.

[0145] The data preprocessing module 220 includes a data augmentation unit, a data labeling unit, a noise removal unit, and an image enhancement unit. The data augmentation unit uses the OpenCV algorithm to perform image cropping, augmentation, and stitching. The data labeling unit integrates the LabelImg labeling tool to perform feature labeling and test verification. The noise removal unit and the image enhancement unit respectively remove foreign objects and enhance image contrast, ultimately obtaining a sample set and dividing it into a training set and a test set.

[0146] The model building module 230 uses the Python programming language and is based on the PyTorch framework to build a convolutional neural network, introduce the YOLOv8 algorithm and add coordinate attention to complete the initial recognition model building and parameter initialization. This module can call the open-source YOLOv8 code for secondary development to adapt to the needs of this embodiment.

[0147] The model training module 240 integrates the SGD optimizer and CIoU loss function. It inputs the training set into the initial recognition model for training, verifies the model performance through the test set, and outputs the trained YOLOv8 defect recognition model. This module can monitor the loss value and accuracy in real time during the training process, and realize the automatic convergence and testing of the model.

[0148] The real-time image acquisition module 250 includes a 3D model calibration unit and a multi-source image acquisition unit. The 3D model calibration unit acquires the 3D model of the boiler through a laser scanning device and calibrates the coordinates of the target component. The multi-source image acquisition unit deploys multi-source acquisition devices to synchronously acquire multi-source images of the target component and performs image preprocessing through algorithms such as Gaussian filtering and image registration.

[0149] The defect classification module 260 inputs the preprocessed multi-source images (or fused images) into the trained YOLOv8 defect recognition model to identify weld types and defect types, output the recognition results and generate an inspection report, and store the results in the boiler inspection database for staff to view.

[0150] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the computer equipment described above can be referred to the corresponding steps in the aforementioned method, and will not be elaborated further here.

[0151] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the methods described in the above embodiments.

[0152] Based on the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, computer equipment, or network equipment, etc.) to execute the methods described in the various implementation scenarios of this application.

[0153] In summary, this application provides a method for identifying weld defects in pressure-bearing special equipment, a computer device, and a storage medium. This solution optimizes the YOLOv8 algorithm, introduces convolutional neural networks and coordinate attention, and improves the accuracy and efficiency of defect identification, resulting in increased detection efficiency compared to traditional manual inspection. Multiple image sources can be flexibly selected to adapt to different boiler inspection scenarios. The device and computer equipment can be directly deployed in industrial sites, demonstrating strong practicality and broad applicability.

[0154] In the embodiments provided in this application, it should be understood that the disclosed apparatus, systems, and methods can also be implemented in other ways. The apparatus, systems, and methods embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0155] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for identifying weld defects in pressure-bearing special equipment, characterized in that, The method includes: Based on a pre-established boiler inspection database, historical boiler inspection images, boiler defect images, and defect-free boiler images are retrieved, and a defect classification dataset is constructed according to weld type and welding defect type. The defect classification dataset is augmented and labeled to obtain a sample set, which is then divided into a training set and a test set. Based on the YOLOv8 algorithm, a convolutional neural network is introduced to establish an initial identification model for different types of welding defects; The training set is used to train initial identification models for different types of welding defects, and the model is tested using the test set to obtain a trained YOLOv8 defect identification model. Acquire multi-source image data of all target recognition areas on the boiler and perform image preprocessing; The multi-source image data is input into the YOLOv8 defect recognition model, and the recognition results of boiler weld type and welding defect type are output.

2. The method according to claim 1, characterized in that, After augmenting and labeling the defect classification dataset to obtain a sample set, the method further includes: Based on the boiler detection database, obtain the anchor frame of the target identification area; Using the center of the anchor frame as a reference point, a settling trend model of the attachments on the boiler surface is established. Based on the settling trend model, the weld trace covering parameters are inverted, and foreign object interference features are generated based on the weld trace covering parameters. Foreign object noise is removed from boiler defect images based on the aforementioned foreign object interference characteristics; Image enhancement processing is performed on boiler defect images that have undergone noise removal.

3. The method according to claim 1 or 2, characterized in that, The defect classification dataset is augmented and labeled to obtain a sample set, including: The boiler defect images in the defect classification dataset are cropped to obtain defect region images; The defective region image is augmented using the OpenCV algorithm to generate multiple augmented images; The augmented image and the original boiler defect image are used together as the dataset for YOLOv8 algorithm feature annotation to obtain a sample set.

4. The method according to claim 1, characterized in that, The method further includes: setting an image fusion model, wherein the image fusion model includes an encoding network, a fusion network, and a decoding network; The coding network consists of n groups of hybrid architecture modules of different sizes, where n is a natural number. The fusion network consists of feature fusion layers, where the number of feature fusion layers is the same as the number of groups in the hybrid architecture module; The upsampling settings are configured using a hybrid architecture module to build the decoding network; The encoding network, fusion network, and decoding network are connected; a 1*1 convolutional layer is set at the input of the encoding network and a 1*1 convolutional layer is set at the output of the decoding network to form the overall network architecture of the image fusion model.

5. The method according to claim 4, characterized in that, The method further includes: fusing multi-source images using a pre-defined image fusion model to obtain a fused image; specifically including: The global features of the i-th source image are calculated using the softmax function. attention weights Global features of the j-th source image attention weights ; Among them, the i-th source image is one of the multiple source images, and the j-th source image is another of the multiple source images; Global features of the i-th source image With attention weight Multiplication, global features of the j-th source image With attention weight Multiply and add them together to obtain the fused features. ; The fused features are then subjected to multi-scale decoding and convolution operations to obtain the final fused image.

6. The method according to claim 4, characterized in that, The hybrid architecture module includes a two-layer network consisting of a normalization layer, a window self-attention mechanism, a residual connection layer, and a multilayer perceptron; the window self-attention mechanism of the two-layer network can perform information interaction between different windows to achieve global modeling of the entire image.

7. The method according to claim 1, characterized in that, The method of introducing a convolutional neural network based on the YOLOv8 algorithm to establish an initial identification model for different types of welding defects includes: A convolutional neural network is constructed using a spatial-to-depth transform layer and a stride convolutional layer. The YOLOv8 algorithm was selected and introduced into the constructed convolutional neural network; The initial recognition model is obtained by adding coordinate attention to the Neck layer of the YOLOv8 algorithm.

8. The method according to claim 1, characterized in that, The acquisition of multi-source image data of all target recognition areas on the boiler, and the image preprocessing, include: Obtain a three-dimensional model of the boiler, and then calibrate the target recognition area of ​​the three-dimensional model; Based on the calibrated target recognition region coordinates, acquire multi-source image data of all target recognition regions; Each source image in the multi-source image is cropped, denoised, and aligned in terms of spatiotemporal sequence.

9. A device for identifying weld defects in pressure-bearing special equipment, characterized in that, The method applied to any one of claims 1-8 includes: The data acquisition module is used to retrieve historical boiler inspection images, boiler defect images, and defect-free boiler images based on a pre-established boiler inspection database, and to construct a defect classification dataset according to weld type and welding defect type. The data preprocessing module is used to augment and label the defect classification dataset to obtain a sample set, and then divide the sample set into a training set and a test set. The model building module is used to introduce a convolutional neural network based on the YOLOv8 algorithm to build an initial identification model for different types of welding defects. The model training module is used to train the initial identification model for different types of welding defects by inputting the training set, and to test the model by using the test set to obtain the trained YOLOv8 defect identification model. The real-time image acquisition module is used to acquire multi-source image data of all target recognition areas on the boiler and perform image preprocessing; wherein, the multi-source images include one or more of infrared images, UT images, RT images, PAUT images, ACFM images and TOFD images; The defect classification module is used to input the multi-source image data into the YOLOv8 defect recognition model and output the recognition results of boiler weld type and welding defect type.

10. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the method as described in any one of claims 1-8.

Images