Method and device for tracing quality defects of continuous casting billets, storage medium and electronic equipment

Abstract

The application discloses a continuous casting billet quality defect tracing method and device, a storage medium and an electronic device. It relates to the technical field of quality control and tracing in continuous casting production, wherein the method comprises: using a pre-trained defect recognition model to recognize the macro image of a continuous casting billet to be traced for defects, obtaining a defect type, and the macro image carrying the billet number of the continuous casting billet to be traced for defects; based on the defect type and the billet number, using a pre-constructed continuous casting billet macro quality knowledge graph to query, obtaining the defect range of different process parameters corresponding to the defect type; based on the actual production process parameters of the continuous casting billet to be traced for defects and each defect range, reasoning the cause of the quality defect, and obtaining the quality tracing result. The method improves the performance and accuracy of the continuous casting billet macro quality tracing system, and provides strong protection for the production of high-quality continuous casting billet products by enterprises.

Description

Technical Field

[0001] This invention relates to the field of quality control and traceability technology in continuous casting production, and particularly to a method, device, storage medium, and electronic equipment for tracing quality defects in continuously cast billets. Background Technology

[0002] In the steel production process, quality control of continuously cast billets is crucial. Low-magnification quality defects in billets, such as cracks and inclusions, can significantly impact subsequent rolling processes and the performance of the final product. Traditional methods for detecting and tracing low-magnification quality defects in continuously cast billets mainly rely on manual experience, resulting in low efficiency, strong subjectivity, and difficulty in comprehensively analyzing the causes of defects. Although existing technologies have begun to explore the use of knowledge graphs and vision for quality control, most methods focus on the application of a single technology or fail to fully integrate visual recognition and knowledge graph technologies to achieve efficient quality traceability.

[0003] Existing technologies have made some progress in detecting surface defects in continuously cast billets, using linear array cameras to acquire images and perform feature analysis to achieve surface defect detection. However, limitations remain, primarily targeting surface defects rather than low-magnification structural defects in the billet, and relying on large amounts of sample data, making it difficult to handle sample sparsity issues. Other technologies combine deep learning and knowledge graphs for continuous casting billet defect identification, constructing knowledge graphs using expert knowledge to assist deep learning models in identifying defects, but they do not fully leverage the advantages of both to achieve synergistic optimization.

[0004] In summary, existing technologies have shortcomings in the detection and traceability of low-magnification quality defects in continuously cast billets. They lack deep integration of visual recognition and knowledge graph technologies, making it difficult to achieve efficient and accurate defect identification and root cause tracing. Therefore, there is an urgent need for an innovative method that can fully leverage the advantages of these two technologies to improve the efficiency and accuracy of low-magnification quality control in continuously cast billets. Summary of the Invention

[0005] In view of this, the present invention provides a method, device, storage medium and electronic device for tracing quality defects in continuously cast billets. The main purpose is to solve the problems of excessive reliance on human experience, low analysis efficiency and insufficient accuracy in the existing low-magnification quality defect detection and tracing technologies for continuously cast billets.

[0006] To address the above problems, this application provides a method for tracing quality defects in continuously cast billets, comprising:

[0007] A pre-trained defect recognition model is used to identify the defect type in a low-magnification image of the continuous casting billet to be traced for defects. The low-magnification image carries the billet number of the continuous casting billet to be traced for defects.

[0008] Based on the defect type and the billet number, a pre-constructed low-magnification quality knowledge graph of continuously cast billets is used to query and obtain the defect range of different process parameters that produce the defect type.

[0009] Based on the actual production process parameters of the continuous casting billet to be traced for defects and the range of each defect, the cause of quality defects is inferred, and the quality traceability result is obtained.

[0010] Optionally, before using the pre-trained defect recognition model to identify the low-magnification image of the continuous casting billet to be traced for defects, the method further includes: constructing a defect recognition model;

[0011] The construction of the defect identification model specifically includes:

[0012] Acquire multiple low-magnification images of historical continuous casting billets;

[0013] The historical continuous casting billet low-magnification image is labeled to obtain a label image, which includes the defect type, defect level and defect boundary box;

[0014] The label image is subjected to data augmentation processing to obtain an enhanced image;

[0015] The improved YOLOv10 model is trained using the enhanced image as a training sample to obtain the defect recognition model;

[0016] The improved YOLOv10 model includes a backbone network, a neck network, and a detection head.

[0017] The backbone network includes a first standard convolutional layer, a local and global feature fusion layer, and multiple dynamic serpentine convolutional layers;

[0018] The neck network includes a bidirectional feature pyramid layer and a spatial pyramid pooling layer that enhances the coordination attention mechanism.

[0019] The detection head includes a multi-branch decoupling head and a loss prediction layer.

[0020] Optionally, the step of using a pre-trained defect recognition model to identify the defect type from a low-magnification image of the continuously cast billet to trace defects specifically includes:

[0021] The backbone network of the defect identification model is used to extract features from the low-magnification image of the continuous casting billet to be traced for defect tracing, thereby obtaining the first target feature image;

[0022] The neck network of the defect recognition model is used to extract features from the first target feature image to obtain the second target feature image;

[0023] The detection head of the defect recognition model is used to detect the second target feature image to obtain the defect type, defect level, target bounding box position coordinates, and confidence score.

[0024] Optionally, the step of using the backbone network of the defect identification model to extract features from the low-magnification image of the continuously cast billet to be traced for defect identification, and obtaining a first target feature image, specifically includes:

[0025] The low-magnification image is downsampled using the first standard convolutional layer of the backbone network to obtain the first feature image.

[0026] The first feature image is subjected to global context feature extraction using the local and global feature fusion layer of the backbone network to obtain the second feature image;

[0027] The second feature image is adaptively processed by the first dynamic serpentine convolutional layer of the backbone network to obtain the third feature image;

[0028] The third feature image is enhanced using the second dynamic serpentine convolutional layer of the backbone network to obtain the fourth feature image.

[0029] The third dynamic serpentine convolutional layer of the backbone network is used to perform deep semantic learning on the fourth feature image to obtain the first target feature image.

[0030] Optionally, the step of using the neck network of the defect recognition model to extract features from the first target feature image to obtain the second target feature image specifically includes:

[0031] The first target feature image is weighted and aggregated using the bidirectional feature pyramid layer of the neck network to obtain the fifth feature image;

[0032] The spatial pyramid pooling layer of the neck network is used to perform defect region activation processing on the fifth feature image to obtain the second target feature image with added global and local features.

[0033] Optionally, the detection head of the defect recognition model is used to detect the second target feature image, specifically including:

[0034] The multi-branch decoupling head of the detection head is used to detect the second target feature image respectively, and the category detection result corresponding to the category branch decoupling head, the bounding box detection result corresponding to the regression branch decoupling head, and the confidence detection result corresponding to the confidence branch decoupling head are obtained.

[0035] The loss prediction layer of the detection head is used to predict the loss based on the category detection results, bounding box detection results, and confidence detection results, and the loss value is obtained.

[0036] Optionally, before querying based on the defect type and the billet number using a pre-constructed low-magnification quality knowledge graph of continuously cast billets, the method further includes: constructing a low-magnification quality knowledge graph of continuously cast billets;

[0037] The construction of the low-magnification quality knowledge graph of continuously cast billets specifically includes:

[0038] A named entity recognition model enhanced with lexical features is used to locate entities in a pre-defined knowledge database to obtain the location of entity keywords.

[0039] The entity keywords in the preset knowledge database are marked;

[0040] The local branch of the residual attention-enhanced bi-branch model captures the relationship keywords between adjacent entity keyword pairs after labeling.

[0041] Based on the adjacent entity key pairs and the relation key, entity relation triples are constructed to obtain an initial set of triples;

[0042] The initial set of triples is filtered to obtain the target set of triples;

[0043] A knowledge graph is constructed based on the target ternary set to obtain a low-magnification quality knowledge graph of the continuously cast billet.

[0044] To address the aforementioned problems, this application provides a continuous casting billet quality defect tracing device, comprising:

[0045] The identification module is used to identify the defect type by using a pre-trained defect identification model to identify a low-magnification image of the continuous casting billet to be traced for defects. The low-magnification image carries the billet number of the continuous casting billet to be traced for defects.

[0046] The query module is used to perform a query based on the defect type and the billet number using a pre-constructed low-magnification quality knowledge graph of continuously cast billets, to obtain the defect range of different process parameters that produce the defect type.

[0047] The reasoning module is used to perform quality defect cause reasoning based on the actual production process parameters of the continuous casting billet to be traced and the range of each defect, and to obtain the quality traceability result.

[0048] To address the aforementioned problems, this application provides a storage medium storing a computer program that, when executed by a processor, implements the steps of the continuous casting billet quality defect tracing method described above.

[0049] To address the aforementioned problems, this application provides an electronic device, comprising at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the continuous casting billet quality defect tracing method described above.

[0050] The beneficial effects of this application are as follows: This application deeply integrates visual recognition technology and knowledge graph technology, realizing intelligent processing of the entire process from accurate identification to efficient traceability of low-magnification quality defects in continuously cast billets. Compared with traditional methods that rely solely on human experience or utilize only one technology, this invention can significantly improve the accuracy of defect identification and the comprehensiveness of traceability analysis, reduce the cost of manual intervention, and improve the level of intelligent quality control in steel enterprises.

[0051] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0052] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0053] Figure 1 A flowchart illustrating a method for tracing quality defects in continuously cast billets provided in an embodiment of this application is shown.

[0054] Figure 2 A flowchart illustrating a method for tracing quality defects in continuously cast billets according to another embodiment of this application is shown.

[0055] Figure 3 A structural block diagram of a continuous casting billet quality defect tracing device provided in another embodiment of this application is shown. Detailed Implementation

[0056] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0057] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0058] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0059] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0060] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0061] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0062] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely serve as the basis and representative basis for the claims to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0063] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0064] This application provides a method for tracing quality defects in continuously cast billets, such as... Figure 1 As shown, it includes:

[0065] Step S101: Use a pre-trained defect recognition model to identify the low-magnification image of the continuous casting billet to be traced for defects, and obtain the defect type. The low-magnification image carries the billet number of the continuous casting billet to be traced for defects.

[0066] In this step, the backbone network of the defect recognition model is used to extract features from the low-magnification image of the continuous casting billet to be traced for defects, resulting in a first target feature image; the neck network of the defect recognition model is used to extract features from the first target feature image, resulting in a second target feature image; and the detection head of the defect recognition model is used to detect the second target feature image, resulting in the defect type, target bounding box position coordinates, and confidence score.

[0067] Step S102: Based on the defect type and the billet number, a pre-constructed low-magnification quality knowledge graph of continuously cast billets is used to query and obtain the defect range of different process parameters corresponding to the defect type.

[0068] In the specific implementation process of this step, the target triplet set of the pre-constructed low-magnification quality knowledge graph of the continuously cast billet is queried based on the defect type and the billet number to obtain at least one triplet data corresponding to the defect type and the billet number; based on at least one triplet data, the defect range that generates at least one process parameter corresponding to the defect type is determined.

[0069] Step S103: Based on the actual production process parameters of the continuous casting billet to be traced for defects and the range of each defect, perform quality defect cause reasoning to obtain quality traceability results.

[0070] In the specific implementation process of this step, the actual production process parameters include the actual parameter values ​​of parameters such as superheat, drawing speed, and secondary cooling water volume; based on the actual production process parameters and each of the defect ranges, a matching is performed, and when the actual production process parameter falls within the defect range, the actual production process parameter is determined as the target defect parameter; based on the predetermined weight coefficient of the target defect parameter, a calculation is performed to obtain the quality cause score value of the actual production process parameter that caused the defect type, and the quality traceability result is obtained.

[0071] This application deeply integrates visual recognition technology and knowledge graph technology, realizing intelligent processing of the entire process from accurate identification to efficient traceability of low-magnification quality defects in continuously cast billets. Compared with traditional methods that rely solely on human experience or utilize only one technology, this invention can significantly improve the accuracy of defect identification and the comprehensiveness of traceability analysis, reduce the cost of manual intervention, and enhance the level of intelligent quality control in steel enterprises.

[0072] Another embodiment of this application provides a different method for tracing quality defects in continuously cast billets, such as... Figure 2 As shown, it includes:

[0073] Step S201: Construct a defect identification model;

[0074] In this step, multiple low-magnification images of historical continuously cast billets are acquired. These images include low-magnification images of different types of billets, such as square and round billets. A large amount of image data is collected on the continuous casting production line using a high-resolution camera, ensuring that the images cover different production stages and possible defect types. The collected image data includes, but is not limited to, close-up images of low-magnification quality defects such as cracks and inclusions. The collected image data is preprocessed, including image enhancement, normalization, and cropping, to expand the dataset and improve the model's generalization ability. Data augmentation techniques, such as image rotation, mirroring, translation, random cropping, and grayscale adjustment, are used to expand the original dataset to meet the data requirements of deep learning models. The historical continuous casting billet low-magnification images are labeled to obtain labeled images, which include defect types, defect levels, and defect boundary boxes. The defect types include: crack-like defects such as longitudinal cracks, transverse cracks, and star-shaped cracks; surface irregular defects such as slag inclusions, porosity, scale, depressions, and excessively deep vibration marks; shape defects such as bulging, rhomboid deformation, and bending; internal crack types such as intermediate cracks, central cracks, triangular cracks, and diagonal cracks; segregation defects such as central segregation, V-shaped segregation, and banded segregation; porosity and shrinkage defects such as central looseness and central shrinkage cavities; inclusion defects such as large inclusions, small dispersed inclusions, subcutaneous bubbles, and internal bubbles; and mixed crystals. The defect types include: structural abnormalities such as abnormal organization and abnormal structure; and shape abnormalities such as bulging deformation and out-of-tolerance cross-sectional dimensions. The defect levels can be classified as Level 1, Level 2, etc., based on the actual enterprise casting billet standards or industry casting billet standards. The defect bounding boxes can be annotated using basic annotation methods such as bounding box annotation and polygon annotation, or using annotation tools. The labeled images undergo data augmentation processing to obtain enhanced images. These enhanced images are used as training samples to train the improved YOLOv10 model, resulting in the defect recognition model. To improve model accuracy, reliably evaluate generalization ability, and ensure rigorous model testing, the annotated casting billet defect images are divided into training, testing, and validation sets in an 8:1:1 ratio. To ensure gradient descent reaches the convergence basin and avoids overfitting in later stages, a visual recognition algorithm is used to train the preprocessed image data for epochs of 300-500 to generate the defect recognition model. The improved YOLOv10 model includes a backbone network, a neck network, and a detection head. The backbone network includes a first standard convolutional layer, a local and global feature fusion layer, and multiple dynamic serpentine convolutional layers. The neck network includes a bidirectional feature pyramid layer and a spatial pyramid pooling layer with an added coordinated attention mechanism. The detection head includes a multi-branch decoupling head and a loss prediction layer.

[0075] Step S202: Use the backbone network of the defect identification model to extract features from the low-magnification image of the continuous casting billet to be traced for defect tracing, and obtain the first target feature image;

[0076] In this step, the first standard convolutional layer of the backbone network is used to downsample the low-magnification image to obtain a first feature image. This first standard convolutional layer can be a standard 2D convolutional layer with 3 input channels, 32 output channels, a kernel size of 3×3, and a stride of 2. It halves the height and width of the input low-magnification image while increasing the number of channels. When the input image size is 640×640×3, the first standard convolutional layer is used to downsample the low-magnification image, resulting in a first feature image of 320×320×32. The first standard convolutional layer is then used to downsample the low-magnification image again, extracting edge textures, halving the resolution, increasing the number of channels, and abstracting edges. This quickly extracts the billet outline and reduces noise. The local and global feature fusion layer of the backbone network is then used to extract global contextual features from the first feature image to obtain a second feature image. This local and global feature fusion layer introduces global context, suppresses the background, and combines local and global extraction; CBAM highlights important regions. Focusing on defect clusters improves sensitivity to small defects. The height and width of the first feature image are halved, while the channel dimension is doubled. Through the local and global feature fusion layer, the 320×320×32 feature image can be converted into a 160×160×64 feature image. The first dynamic serpentine convolutional layer of the backbone network is used to adaptively process the boundary of the second feature image to obtain the third feature image. The first dynamic serpentine convolutional layer includes a dynamic serpentine convolution module, a MobileViTv2 module, and a DeformableConv module. This first dynamic serpentine convolutional layer performs adaptive boundary processing on the second feature image and handles deformation, making the boundary softer and correcting deformation. This improves the accuracy of bending crack detection. A first dynamic serpentine convolutional layer is used to convert the 160×160×64 second feature image into an 80×80×128 third feature image. The second dynamic serpentine convolutional layer of the backbone network then performs multi-scale enhancement processing on the third feature image to obtain a fourth feature image. This second dynamic serpentine convolutional layer includes a dynamic serpentine convolution module, a MobileViTv2 module, a Deformable Conv module, and a CBAM module. It converts the 80×80×128 third feature image into a 40×40×256 fourth feature image, refining features, highlighting high-contrast regions, distinguishing defect types, and reducing false positives. The third dynamic serpentine convolutional layer of the backbone network is then used to perform deep semantic learning on the fourth feature image to obtain the first target feature image.The third dynamic serpentine convolutional layer includes a dynamic serpentine convolution module, a MobileViTv2 module, a Deformable Conv module, and a CBAM module; the third dynamic serpentine convolutional layer transforms the 40×40×256 fourth feature image into a 20×20×512 first target feature image; deep semantic learning emphasizes global context and captures a wide range of defects.

[0077] Step S203: Use the neck network of the defect recognition model to extract features from the first target feature image to obtain the second target feature image;

[0078] In this step, the bidirectional feature pyramid layer of the neck network is used to perform weighted aggregation processing on the first target feature image to obtain a fifth feature image. Specifically, the bidirectional feature pyramid layer of the neck network is used to perform weighted aggregation processing on the first target feature image to obtain multi-scale feature images; for example, generating fifth feature images of different scales such as 20×20×512, 40×40×256, 80×80×128, and 10×10×1024 to handle defects with large scale differences. The spatial pyramid pooling layer of the neck network is used to perform defect region activation processing on the fifth feature image to obtain a second target feature image with added global and local features. The spatial pyramid pooling layer performs horizontal and vertical attention enhancement processing on the multi-scale fifth feature images and performs multi-scale pooling processing to obtain attention-enhanced multi-scale second target feature images. The spatial pyramid pooling layer captures direction dependence and multi-scale context, improving direction sensitivity.

[0079] Step S204: Use the detection head of the defect recognition model to detect the second target feature image to obtain the defect type, defect level, target bounding box position coordinates and confidence score;

[0080] In this step, the multi-branch decoupling head of the detection head is used to detect the second target feature image, obtaining the category detection result corresponding to the category branch decoupling head, the bounding box detection result corresponding to the regression branch decoupling head, and the confidence detection result corresponding to the confidence branch decoupling head. The category detection result is the defect type and defect level. The bounding box detection result is the target bounding box position coordinates. The confidence detection result is the confidence score. The loss prediction layer of the detection head is used to predict the loss based on the category detection result, the bounding box detection result, and the confidence detection result to obtain the loss value.

[0081] Step S205: Construct a low-magnification quality knowledge graph for continuously cast billets;

[0082] In this step, a named entity recognition model with enhanced lexical features is used to locate entities in a pre-defined knowledge database to obtain the positions of entity keywords. The entity keywords in the pre-defined knowledge database are then labeled. This pre-defined knowledge database can be composed of enterprise continuous casting process manuals, metallurgical textbooks, and quality analysis reports from enterprise experts. For example, in the sentence "Excessive casting speed causes surface cracks," the entity pairs are "excessive casting speed" and "surface cracks." Positional coding is added to these entities, for example: "[E1]Excessive casting speed[ / E1] causes [E2]surface cracks[ / E2]", etc. A residual attention-enhanced dual-branch model is then used. The method employs local branches to capture relational keywords between adjacent entity keyword pairs after tagging; specifically, it uses residual attention-enhanced local branches to capture short-distance dependencies between entities; for example, "trigger" is used as a relational keyword; it uses residual attention-enhanced global branches to understand the overall semantics of the sentence; for example, the indirect association between "pulling speed" and "crack"; it uses residual attention weighted fusion of dual-branch features; for example, "excessive" is a key description of the "pulling speed" anomaly; it obtains relational keywords between adjacent entity keyword pairs; based on the adjacent entity keyword pairs and the relational keywords, it constructs entity relation triples to obtain an initial triple set; for example, the initial triple set includes [overheating 30℃, The initial set of triplets is filtered to obtain a target set of triplets. Triplet sets without significant relationships are deleted to obtain the target set of triplets. For example, the triplets with no belonging relationship are deleted to obtain the target set of triplets. Based on the target set of triplets, a knowledge graph is constructed to obtain the low-magnification quality knowledge graph of the continuous casting billet. A knowledge graph construction tool (such as Neo4j) is used to import the labeled data into a graph database to construct the low-magnification quality knowledge graph of the continuous casting billet. During the construction process, query languages ​​for graph databases (such as Cypher) are used for data querying and relational reasoning to ensure the integrity and accuracy of the knowledge graph.

[0083] Step S206: Based on the defect type and the billet number, a pre-constructed low-magnification quality knowledge graph of continuously cast billets is used to query and obtain the defect range of different process parameters corresponding to the defect type.

[0084] In this step, the process involves querying the low-magnification quality knowledge graph of the continuously cast billet based on the billet number to obtain the target continuous cast billet low-magnification quality knowledge graph corresponding to the billet number; searching the target continuous cast billet low-magnification quality knowledge graph for triplet data related to the defect type of the input billet to obtain at least one triplet data containing the defect type; and obtaining the defect range of different process parameters that produce the defect type; for example, if the defect type is "center segregation of continuously cast billet", querying the target continuous cast billet low-magnification quality knowledge graph yields the following query results: [Superheat > 30℃, resulting in center segregation], [Stretching speed > 1.5m / min, resulting in center segregation], and [Electromagnetic stirring current > 180A, resulting in center segregation]; and determining the defect range of each different process parameter from the triplet data obtained from each query. For example, superheat > 30℃, pulling speed > 1.5m / min, and electromagnetic stirring current > 180A are the defect ranges of different process parameters that form the defect type "center segregation".

[0085] Step S207: Based on the actual production process parameters of the continuous casting billet to be traced for defects and the range of each defect, perform quality defect cause reasoning to obtain quality traceability results.

[0086] In this step, the real-time production process parameters of continuous casting are matched with the defect range to analyze and deduce the root cause of the billet defects. A quality traceability report is generated, providing defect causes and improvement suggestions. Specifically, the real-time production process parameters are compared with the defect range. When the parameter value corresponding to the actual production process parameter falls within the defect range, the actual production process parameter is determined to be the cause of the defect type. Based on the quality traceability results, a detailed analysis report is generated, including information such as defect type, cause, relevant production parameters, and responsible links. The analysis report is fed back to production management personnel to provide them with decision support, enabling timely adjustments and optimizations to the process. Through continuous application and optimization, the performance and accuracy of the low-magnification quality traceability system for continuously cast billets are continuously improved, providing a strong guarantee for enterprises to produce high-quality continuously cast billet products.

[0087] This application deeply integrates visual recognition technology and knowledge graph technology, realizing intelligent processing of the entire process from accurate identification to efficient traceability of low-magnification quality defects in continuously cast billets. Compared with traditional methods that rely solely on human experience or utilize only one technology, this invention can significantly improve the accuracy of defect identification and the comprehensiveness of traceability analysis, reduce the cost of manual intervention, and enhance the level of intelligent quality control in steel enterprises.

[0088] Another embodiment of this application provides a device for tracing quality defects in continuously cast billets, such as... Figure 3 As shown, it includes:

[0089] Identification module 1 is used to identify the defect type by using a pre-trained defect identification model to identify a low-magnification image of the continuous casting billet to be traced for defect tracing. The low-magnification image carries the billet number of the continuous casting billet to be traced for defect tracing.

[0090] Query module 2 is used to perform a query based on the defect type and the billet number using a pre-constructed low-magnification quality knowledge graph of continuously cast billets, to obtain the defect range of different process parameters corresponding to the defect type;

[0091] The reasoning module 3 is used to perform quality defect cause reasoning based on the actual production process parameters of the continuous casting billet to be traced and the range of each defect, and to obtain the quality traceability result.

[0092] In the specific implementation process, the device further includes a model building module, which is specifically used to acquire multiple low-magnification images of historical continuous casting billets; to annotate the low-magnification images of the historical continuous casting billets to obtain label images, the label images including defect type, defect level, and defect bounding box; to perform data augmentation processing on the label images to obtain enhanced images; and to train the improved YOLOv10 model using the enhanced images as training samples to obtain the defect recognition model; wherein, the improved YOLOv10 model includes a backbone network, a neck network, and a detection head; the backbone network includes a first standard convolutional layer, a local and global feature fusion layer, and multiple dynamic serpentine convolutional layers; the neck network includes a bidirectional feature pyramid layer and a spatial pyramid pooling layer with an added coordinated attention mechanism; and the detection head includes a multi-branch decoupling head and a loss prediction layer.

[0093] In the specific implementation process, the identification module 1 is specifically used to: extract features from the low-magnification image of the continuous casting billet to be traced for defects using the backbone network of the defect identification model to obtain a first target feature image; extract features from the first target feature image using the neck network of the defect identification model to obtain a second target feature image; and detect the second target feature image using the detection head of the defect identification model to obtain the defect type, target bounding box position coordinates, and confidence score.

[0094] In the specific implementation process, the recognition module 1 is further configured to: perform downsampling processing on the low-magnification image using the first standard convolutional layer of the backbone network to obtain a first feature image; perform global context feature extraction on the first feature image using the local and global feature fusion layer of the backbone network to obtain a second feature image; perform adaptive boundary processing on the second feature image using the first dynamic serpentine convolutional layer of the backbone network to obtain a third feature image; perform multi-scale enhancement processing on the third feature image using the second dynamic serpentine convolutional layer of the backbone network to obtain a fourth feature image; and perform deep semantic learning on the fourth feature image using the third dynamic serpentine convolutional layer of the backbone network to obtain the first target feature image.

[0095] In the specific implementation process, the recognition module 1 is further used to: perform weighted aggregation processing on the first target feature image using the bidirectional feature pyramid layer of the neck network to obtain a fifth feature image; and perform defect region activation processing on the fifth feature image using the spatial pyramid pooling layer of the neck network to obtain a second target feature image with added global and local features.

[0096] In the specific implementation process, the recognition module 1 is further used to: use the multi-branch decoupling head of the detection head to detect the second target feature image respectively, and obtain the category detection result corresponding to the category branch decoupling head, the bounding box detection result corresponding to the regression branch decoupling head, and the confidence detection result corresponding to the confidence branch decoupling head; use the loss prediction layer of the detection head to perform loss prediction on the category detection result, the bounding box detection result, and the confidence detection result to obtain the loss value.

[0097] In specific implementation, the device further includes a knowledge graph construction module, which is specifically used for: using a named entity recognition model with enhanced lexical features to locate entities in a preset knowledge database and obtain entity keyword positions; marking the entity keywords in the preset knowledge database; using a local branch of a bi-branch model with enhanced residual attention to capture relational keywords between adjacent entity keyword pairs after marking; constructing entity relation triples based on the adjacent entity keyword pairs and the relational keywords to obtain an initial triple set; filtering the initial triple set to obtain a target triple set; and constructing a knowledge graph based on the target triple set to obtain the low-magnification quality knowledge graph of the continuously cast billet.

[0098] This application deeply integrates visual recognition technology and knowledge graph technology, realizing intelligent processing of the entire process from accurate identification to efficient traceability of low-magnification quality defects in continuously cast billets. Compared with traditional methods that rely solely on human experience or utilize only one technology, this invention can significantly improve the accuracy of defect identification and the comprehensiveness of traceability analysis, reduce the cost of manual intervention, and enhance the level of intelligent quality control in steel enterprises.

[0099] Another embodiment of this application provides a storage medium storing a computer program, which, when executed by a processor, implements the following method steps:

[0100] Step 1: Use a pre-trained defect recognition model to identify the low-magnification image of the continuous casting billet to be traced for defects, and obtain the defect type. The low-magnification image carries the billet number of the continuous casting billet to be traced for defects.

[0101] Step 2: Based on the defect type and the billet number, use a pre-constructed low-magnification quality knowledge graph of continuously cast billets to query and obtain the defect range of different process parameters corresponding to the defect type.

[0102] Step 3: Based on the actual production process parameters of the continuous casting billet to be traced for defects and the range of each defect, perform quality defect cause reasoning to obtain quality traceability results.

[0103] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0104] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0105] The specific implementation process of the above method steps can be found in the embodiment of the above arbitrary continuous casting billet quality defect tracing method, which will not be repeated here.

[0106] This application deeply integrates visual recognition technology and knowledge graph technology, realizing intelligent processing of the entire process from accurate identification to efficient traceability of low-magnification quality defects in continuously cast billets. Compared with traditional methods that rely solely on human experience or utilize only one technology, this invention can significantly improve the accuracy of defect identification and the comprehensiveness of traceability analysis, reduce the cost of manual intervention, and enhance the level of intelligent quality control in steel enterprises.

[0107] Another embodiment of this application provides an electronic device, which can be a server. The electronic device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the program is executed by the processor, it implements the functions or steps of a continuous casting billet quality defect tracing method on the server side.

[0108] In one embodiment, an electronic device is provided, which can be a client. The electronic device includes a processor, memory, a network interface, a display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When the program is executed by the processor, it implements the functions or steps of a continuous casting billet quality defect tracing method on the client side.

[0109] Another embodiment of this application provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, performs the following method steps:

[0110] Step 1: Use a pre-trained defect recognition model to identify the low-magnification image of the continuous casting billet to be traced for defects, and obtain the defect type. The low-magnification image carries the billet number of the continuous casting billet to be traced for defects.

[0111] Step 2: Based on the defect type and the billet number, use a pre-constructed low-magnification quality knowledge graph of continuously cast billets to query and obtain the defect range of different process parameters corresponding to the defect type.

[0112] Step 3: Based on the actual production process parameters of the continuous casting billet to be traced for defects and the range of each defect, perform quality defect cause reasoning to obtain quality traceability results.

[0113] The specific implementation process of the above method steps can be found in the embodiment of the above arbitrary continuous casting billet quality defect tracing method, which will not be repeated here.

[0114] This application deeply integrates visual recognition technology and knowledge graph technology, realizing intelligent processing of the entire process from accurate identification to efficient traceability of low-magnification quality defects in continuously cast billets. Compared with traditional methods that rely solely on human experience or utilize only one technology, this invention can significantly improve the accuracy of defect identification and the comprehensiveness of traceability analysis, reduce the cost of manual intervention, and enhance the level of intelligent quality control in steel enterprises.

[0115] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for tracing quality defects in continuously cast billets, characterized in that, include: A pre-trained defect recognition model is used to identify the defect type in a low-magnification image of the continuous casting billet to be traced for defects. The low-magnification image carries the billet number of the continuous casting billet to be traced for defects. Based on the defect type and the billet number, a pre-constructed low-magnification quality knowledge graph of continuously cast billets is used to query and obtain the defect range of different process parameters that produce the defect type. Based on the actual production process parameters of the continuous casting billet to be traced for defects and the range of each defect, the cause of quality defects is inferred to obtain the quality traceability result. The method involves using a pre-trained defect recognition model to identify the defect type from a low-magnification image of the continuous casting billet to trace defects. Specifically, this includes: The backbone network of the defect identification model is used to extract features from the low-magnification image of the continuous casting billet to be traced for defect tracing, thereby obtaining the first target feature image; The neck network of the defect recognition model is used to extract features from the first target feature image to obtain the second target feature image; The detection head of the defect recognition model is used to detect the second target feature image to obtain the defect type, defect level, target bounding box position coordinates, and confidence score; The step of using the neck network of the defect recognition model to extract features from the first target feature image to obtain the second target feature image specifically includes: The first target feature image is weighted and aggregated using the bidirectional feature pyramid layer of the neck network to obtain the fifth feature image; The spatial pyramid pooling layer of the neck network is used to perform defect region activation processing on the fifth feature image to obtain the second target feature image with added global and local features; The defect identification model includes a backbone network, a neck network, and a detection head. The backbone network includes a first standard convolutional layer, a local and global feature fusion layer, and multiple dynamic serpentine convolutional layers; The neck network includes a bidirectional feature pyramid layer and a spatial pyramid pooling layer that enhances the coordination attention mechanism. The detection head includes a multi-branch decoupling head and a loss prediction layer.

2. The method as described in claim 1, characterized in that, Before using a pre-trained defect recognition model to identify the low-magnification image of the continuous casting billet to be traced for defects, the method further includes: constructing a defect recognition model; The construction of the defect identification model specifically includes: Acquire multiple low-magnification images of historical continuous casting billets; The historical continuous casting billet low-magnification image is labeled to obtain a label image, which includes the defect type, defect level and defect boundary box; The label image is subjected to data augmentation processing to obtain an enhanced image; The improved YOLOv10 model is trained using the enhanced image as a training sample to obtain the defect recognition model.

3. The method as described in claim 1, characterized in that, The process of using the backbone network of the defect identification model to extract features from the low-magnification image of the continuously cast billet to be traced for defect identification, thereby obtaining a first target feature image, specifically includes: The low-magnification image is downsampled using the first standard convolutional layer of the backbone network to obtain the first feature image. The first feature image is subjected to global context feature extraction using the local and global feature fusion layer of the backbone network to obtain the second feature image; The second feature image is adaptively processed by the first dynamic serpentine convolutional layer of the backbone network to obtain the third feature image; The third feature image is enhanced using the second dynamic serpentine convolutional layer of the backbone network to obtain the fourth feature image. The third dynamic serpentine convolutional layer of the backbone network is used to perform deep semantic learning on the fourth feature image to obtain the first target feature image.

4. The method as described in claim 1, characterized in that, The detection head using the defect recognition model to detect the second target feature image specifically includes: The multi-branch decoupling head of the detection head is used to detect the second target feature image respectively, and the category detection result corresponding to the category branch decoupling head, the bounding box detection result corresponding to the regression branch decoupling head, and the confidence detection result corresponding to the confidence branch decoupling head are obtained. The loss prediction layer of the detection head is used to predict the loss based on the category detection results, bounding box detection results, and confidence detection results, and the loss value is obtained.

5. The method as described in claim 1, characterized in that, Before querying the continuous casting billet using a pre-constructed low-magnification quality knowledge graph based on the defect type and the billet number, the method further includes: constructing a low-magnification quality knowledge graph for continuous casting billets; The construction of the low-magnification quality knowledge graph of continuously cast billets specifically includes: A named entity recognition model enhanced with lexical features is used to locate entities in a pre-defined knowledge database to obtain the location of entity keywords. The entity keywords in the preset knowledge database are marked; The local branch of the residual attention-enhanced bi-branch model captures the relationship keywords between adjacent entity keyword pairs after labeling. Based on the adjacent entity key pairs and the relation key, entity relation triples are constructed to obtain an initial set of triples; The initial set of triples is filtered to obtain the target set of triples; A knowledge graph is constructed based on the target ternary set to obtain a low-magnification quality knowledge graph of the continuously cast billet.

6. A continuous casting billet quality defect tracing device, used to implement any one of the continuous casting billet quality defect tracing methods as described in claims 1 to 5, characterized in that, include: The identification module is used to identify a low-magnification image of a continuously cast billet to be traced for defects using a pre-trained defect identification model, thereby obtaining the defect type. The low-magnification image carries the billet number of the continuously cast billet to be traced for defects. Specifically, it includes: using the backbone network of the defect identification model to extract features from the low-magnification image of the continuously cast billet to be traced for defects, obtaining a first target feature image; using the neck network of the defect identification model to extract features from the first target feature image, obtaining a second target feature image; using the detection head of the defect identification model to detect the second target feature image, obtaining the defect type, defect level, target bounding box position coordinates, and confidence score; the step of using the neck network of the defect identification model to extract features from the first target feature image to obtain the second target feature image specifically includes: using the bidirectional feature pyramid layer of the neck network to perform weighted aggregation processing on the first target feature image, obtaining a fifth feature image; using the spatial pyramid pooling layer of the neck network to perform defect region activation processing on the fifth feature image, obtaining the second target feature image with added global and local features. The query module is used to perform a query based on the defect type and the billet number using a pre-constructed low-magnification quality knowledge graph of continuously cast billets, to obtain the defect range of different process parameters that produce the defect type. The reasoning module is used to perform quality defect cause reasoning based on the actual production process parameters of the continuous casting billet to be traced and the range of each defect, and to obtain the quality traceability result.

7. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the continuous casting billet quality defect tracing method according to any one of claims 1-5.

8. An electronic device, characterized in that, It includes at least a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the continuous casting billet quality defect tracing method according to any one of claims 1-5.

Images