A method for judging heterogeneous point sources of continuous casting slab based on double-scale target detection

Abstract

The application discloses a continuous casting slab heterogeneous point source judgment method based on double-scale target detection, and the heterogeneous point target in a metallographic picture is positioned and classified and recognized based on a double-scale detection network; the heterogeneous point is classified based on the type, and meanwhile, the source of the heterogeneous point is classified; the source of the heterogeneous point is automatically judged, so that an auxiliary tool is provided for quickly analyzing the defect reason of the heterogeneous point and improving the process. The application is based on visual recognition and is a method for judging the source of the heterogeneous point of the continuous casting slab.

Description

Technical Field

[0001] This invention relates to intelligent manufacturing technology in the metallurgical industry, and more specifically, to a method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection. Background Technology

[0002] Inferior inclusions (such as oxides, MnS, TiN, etc.) in continuously cast steel slabs significantly reduce material strength, affecting subsequent rolling or stamping performance and safety. To control these inclusions, it's crucial to first obtain information on their composition, morphology, size, and distribution. This allows for inclusion classification, and the identification of their origins / phases, leading to the development of optimized continuous casting process schemes to reduce inclusion density in the slab. Process engineers often use EDS / WDS in a scanning electron microscope (SEM) to perform surface scanning of the slab / material, identifying the composition of each inclusion / phase. Based on certain rules, these inclusions / phases are classified (determined to be foreign non-metallic oxides, protective slag, or MnS / TiN formed during solidification) to analyze their origin and develop appropriate process solutions. However, surface scanning in SEM for compositional analysis presents several problems:

[0003] (1) Judgment is based on the human rules of process personnel. Different people have different judgment criteria and different attributions.

[0004] (2) Due to the limitations of component detection accuracy, some components cannot be detected, leading to incorrect judgments;

[0005] (3) Time-consuming, one component surface scan takes more than 1 hour.

[0006] In recent years, the advent of the big data era has led to the rapid development of data-driven machine learning algorithms. Benefiting from their powerful learning capabilities, machine learning is being increasingly widely applied in the steel industry. With the rapid development of convolutional neural networks, machine learning algorithms have demonstrated significant advantages in areas such as object detection, image recognition and segmentation, and sequence analysis, making them highly promising for application in the field of object detection and analysis. In response, steel mills and even mainframe manufacturers are building offline or online heterogeneous / phase visual inspection and analysis systems.

[0007] Existing patent applications, such as patent application number 202210613572.1, propose a method for intelligent identification and rating of non-metallic inclusions in steel using deep learning technology. This method mainly addresses the problems of high workload and low accuracy in manual rating in traditional metallographic analysis, as well as the limitations of traditional grayscale threshold segmentation and feature extraction methods. Through semantic segmentation algorithms in deep learning, it achieves automatic extraction and rating of non-metallic inclusion boundaries, primarily identifying inclusions and extracting their dimensions, without involving inclusion classification. Another example is patent application number 202410656971.5, which proposes an intelligent rating method for non-metallic inclusions in steel based on deep learning. It uses a semantic segmentation algorithm (SAM) to automatically segment non-metallic inclusions and further utilizes a ResNet50 classification neural network to classify and rate the segmented inclusions. This patent can automatically classify and rate four categories of non-metallic inclusions.

[0008] The problem with the above two patented technologies is that:

[0009] 1) In the identification and classification of heterogeneous points in continuous casting slabs, we have only identified and classified non-metallic inclusions. However, in continuous casting slabs, in addition to non-metallic inclusions, there are also metallic inclusions and precipitates in the solid phase. These all affect the subsequent production of continuous casting slabs and need to be identified and classified.

[0010] (2) Failure to determine the source of heterogeneous points. For example, the same TIN phase could be a TIN metallic inclusion formed during solidification or a solid-phase precipitate. Different sources result in different effects of the same physical property, and therefore require different control measures. Therefore, determining the source is crucial and cannot be omitted.

[0011] (3) In terms of methodology, deep learning algorithms based on semantic segmentation are often used to first segment out heterogeneous points (such as inclusions) and then classify them. However, this method based on semantic segmentation is not feasible when source determination is required. Summary of the Invention

[0012] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection, which is a method for determining the source of heterogeneous points in continuously cast slabs based on visual recognition.

[0013] To achieve the above objectives, the present invention adopts the following technical solution:

[0014] A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection;

[0015] A dual-scale detection network is used to locate and classify heterogeneous point targets in metallographic images.

[0016] The classification is based on the type of heterogeneous points, and also combined with the source of the heterogeneous points.

[0017] It automatically identifies the source of heterogeneous points, thus providing an auxiliary tool for quickly analyzing the causes of defects caused by heterogeneous points and making process improvements.

[0018] Preferably, the method for determining the source of heterogeneous points in the continuously cast slab specifically includes the following steps:

[0019] S1, acquire standardized samples and images;

[0020] S2, data acquisition, cropping, filtering and noise reduction preprocessing to form the training dataset D;

[0021] S3, Data annotation: Annotate the heterogeneous points according to their category and shape to generate the labeled dataset D1;

[0022] S4, build and train the dual-scale detection network Y on the framework;

[0023] S5, Construct a heterogeneous point expert knowledge base;

[0024] S6, output the result.

[0025] Preferably, step S1 specifically includes:

[0026] Microscopic image sampling was performed on continuously cast slab samples to obtain a sample image set X under a microscope.

[0027] Preferably, the image sampling range is 0.5mm*0.5mm to 1mm*1mm;

[0028] The image sampling factor is 100 to 500 times.

[0029] Preferably, step S3 specifically includes:

[0030] The target detection and annotation of the heterogeneous points are performed using the EasyData annotation platform. The classification annotation is based on two dimensions: the type and the source of the heterogeneous points.

[0031] Preferably, the types of heterogeneous points include non-metallic inclusions, metallic inclusions, and metallic precipitates;

[0032] The sources of the heterogeneous points include smelting slag, protective slag, high-temperature solidification inclusions in continuous casting, and low-temperature precipitates in continuous casting.

[0033] Preferably, the non-metallic inclusions include Al2O3, SiO2, CaO oxides, and calcium aluminates;

[0034] The metallic inclusions include TiN and MnS;

[0035] The metal precipitates include TiN and AlN.

[0036] Preferably, in step S4, the dual-scale detection network Y includes a shallow detection head and a deep detection head;

[0037] The shallow detection head includes small-scale detection of targets with a size of 1 to 5 mm, where the heterogeneous point size = detection size / magnification.

[0038] The deep detection head includes large-scale detection of targets with a size of 5-20mm, where the heterogeneous point size = detection size / magnification.

[0039] Preferably, in step S5, the heterogeneous point expert knowledge base includes information on the source and process causes of the heterogeneous points, as well as process improvement suggestions.

[0040] Preferably, step S6 specifically includes:

[0041] The dual-scale detection network Y is used to determine the source of inclusions, and outputs the location, classification, source list or visualization image of each heterogeneous point. Then, the heterogeneous point expert knowledge base is integrated to output process factors and improvement suggestions.

[0042] This invention provides a method for determining the source of heterogeneous inclusions in continuously cast slabs based on dual-scale target detection. This method is of great significance in quality inspection in the steel industry, enabling rapid and accurate identification and rating of non-metallic inclusions in steel, thus helping to improve the quality control level of steel products and reduce inspection costs. Simultaneously, this invention also provides new ideas and methods for defect detection in other materials.

[0043] Heterogeneous point classification and origin determination: Unlike existing patents, this invention not only identifies non-metallic inclusions, but also identifies metallic inclusions and precipitated phases. It also includes determining their origin.

[0044] Automation and Intelligence: This invention achieves automatic identification and source determination of heterogeneous points through deep learning algorithms, reducing manual intervention and improving detection efficiency and accuracy. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the method for determining the source of heterogeneous points in continuously cast slabs according to the present invention. Detailed Implementation

[0046] To better understand the above-mentioned technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0047] This invention provides a method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection;

[0048] A dual-scale detection network is used to locate and classify heterogeneous point targets in metallographic images.

[0049] The classification is based on the type of heterogeneous points, and also combined with the source of the heterogeneous points.

[0050] It automatically identifies the source of heterogeneous points, thus providing an auxiliary tool for quickly analyzing the causes of defects caused by heterogeneous points and making process improvements.

[0051] The above-mentioned method for determining the source of heterogeneous points in continuously cast slabs has an accuracy rate of >96% in detecting heterogeneous points in continuous casting, which is significantly better than the traditional manual method and has significant application and promotion value.

[0052] Combination Figure 1 As shown, the method for determining the source of heterogeneous points in continuously cast slabs according to the present invention specifically includes the following steps:

[0053] S1, acquire standardized samples and images;

[0054] Microscopic image sampling was performed on continuously cast slab samples to obtain a sample image set X under a microscope.

[0055] The image sampling range is 0.5mm*0.5mm to 1mm*1mm;

[0056] The image sampling factor is 100 to 500 times.

[0057] To expand the applicability of heterogeneous points in the model training set and increase the model's recognition accuracy and robustness, the steel grade selected for the continuous casting slab samples was a steel defect slab from a certain production base within one year.

[0058] S2, data acquisition, cropping, filtering and noise reduction preprocessing to form the training dataset D.

[0059] S3, Data annotation: The EasyData annotation platform is used to perform target detection annotation of heterogeneous points. The classification annotation is based on two dimensions: the type and the source of the heterogeneous points, and generates the annotation dataset D1.

[0060] The types of heterogeneous points include non-metallic inclusions, metallic inclusions, and metallic precipitates;

[0061] The sources of heterogeneous points include smelting slag, protective slag, high-temperature solidification inclusions in continuous casting, and low-temperature precipitates in continuous casting.

[0062] Non-metallic inclusions include oxides such as Al2O3, SiO2, and CaO, as well as calcium aluminates, etc.

[0063] Metal inclusions include TiN, MnS, etc.

[0064] Metal precipitates include TiN, AlN, etc.

[0065] S4 uses deep learning frameworks such as PyTorch or Huawei Pangu Big Data Model to build and train a dual-scale detection network Y.

[0066] Shallow detection heads include small-scale detection of targets with a size of 1-5 mm; heterogeneous point size = detection size / magnification.

[0067] The deep detection head includes large-scale detection of targets with a size of 5-20mm. The size of the heterogeneous point is equal to the detection size divided by the magnification.

[0068] S5. Construct a heterogeneous point expert knowledge base, including the source of heterogeneous points, process reasons, and process improvement suggestions.

[0069] S6, Output Results: Based on the dual-scale detection network Y, determine the source of inclusions and output the location, classification, source list or visualization image of each heterogeneous point. Then, in the heterogeneous point expert knowledge base, output process factors and improvement suggestions.

[0070] The mechanism and limitations of the method for determining the source of heterogeneous points in continuously cast slabs in this invention are as follows:

[0071] The data selected consists of steel slabs with defects from a specific production base within one year. Unlike existing patents, this invention not only identifies non-metallic inclusions but also metallic inclusions and precipitates, and determines their origin. Therefore, the data source—steel slab samples with defects from a specific production base within one year—ensures a diversity of heterogeneous point types and origins, guaranteeing that the data meets the modeling objectives and the model's accuracy.

[0072] The visual recognition model employs object detection, and the annotation conforms to object detection annotation methods. Classification and origin determination are based not only on the morphology of individual heterogeneous points but also on the morphology of the boundary between the heterogeneous points and the steel matrix, as well as the distribution of the heterogeneous points. Therefore, accurate classification and origin determination cannot be based solely on the morphology of individual points. Consequently, the semantic segmentation model in visual recognition is no longer applicable. Instead, an object detection model is used to delineate individual points and basic boundaries within the annotation box, while continuous / semi-continuous precipitates are labeled as a whole.

[0073] The target detection model employs dual-target detection, involving heterogeneous points ranging in size from 200 micrometers to 2 micrometers, covering a wide scale range. The dual-target detection method can avoid insufficient resolution at small scales.

[0074] By incorporating an expert knowledge base, the system can determine the source of inclusions and output process factors and improvement suggestions. When the source of foreign particles is understood, experts can determine at which process stage they originated in and which aspect of that process led to their formation, thus providing improvement recommendations. Therefore, when the model can identify the source of inclusions, we simultaneously established an expert knowledge base to quickly output possible process factors and improvement suggestions.

[0075] Example

[0076] This embodiment provides a method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection. It is implemented on a scanning electron microscope at a base of a steel plant to perform rapid and automated determination of the source of heterogeneous points in continuously cast slabs, providing process engineers with a closed-loop improvement solution.

[0077] Combined Figure 1 As shown, the method for determining the origin of quality particles in continuously cast slabs using this scanning electron microscope includes the following steps:

[0078] S1, Obtain standardized samples and images:

[0079] We collected typical steel grades and heterogeneous point defects from continuous casting slabs from steel plants, cut and corroded them into 10mm*10mm*5mm metallographic samples, and collected heterogeneous point morphology on a scanning electron microscope.

[0080] S2, data acquisition, cropping, filtering, and noise reduction preprocessing are performed to form the training dataset D:

[0081] 1000 microscope images were acquired at a resolution of 2048*2048 pixels and saved in TIFF format. They were then cropped to a standard size of 1024*1024 pixels according to the scale bar (1 pixel = 0.35 μm). Non-local mean (NLM) filtering was applied with the following parameters:

[0082] The search window is 21×21, the similarity window is 7×7, the noise reduction intensity h=10, and a preprocessed image set D is generated.

[0083] S3, Data annotation: Annotate heterogeneous points according to their category and morphology to generate an annotated dataset D1.

[0084] The EasyData platform was used to classify and annotate heterogeneous points based on their category and origin, including bounding box selection of target boundaries, with an annotation accuracy of ±2 pixels. The categories included non-metallic inclusions (AlO3, SiO2), metallic inclusions (FeS, MnS, TiN), and metallic precipitates (TiN, AlN). Sources included smelting slag, tundish slag, protective slag, high-temperature solidification inclusions from continuous casting, and low-temperature precipitates from continuous casting. The dataset was split into a training set (D1_Train) and a test dataset (D1_Test) in an 8:2 ratio.

[0085] S4. Training and optimization of the object detection network Y model on the PyTorch framework, as detailed below:

[0086] 1) Target detection network Y (improved YOLOv8), with dual-scale detection heads including a shallow detection head (FPN-P2): input resolution of 256×256, anchor frame size of (4×4, 8×8), which detects targets of 0.1~2mm (the size of heterogeneous points is 0.1~2mm / 200) and a deep detection head (FPN-P5): input resolution of 64*64, introducing deformable convolution (DCNv3), which detects targets of 2~10mm (the size of heterogeneous points is 2~10mm / 200).

[0087] 2) Training parameters:

[0088] Optimizer: AdamW, initial learning rate 3e-4, weight decay 0.05.

[0089] Loss function: CIoU Loss + Focal Loss (α = 0.25, γ = 2).

[0090] Batch Size = 16, training epochs = 300, early stopping strategy (patience = 20).

[0091] S5, construct a heterogeneous point expert knowledge base, including its source, process factors, and process improvement information:

[0092] An expert knowledge base for heterogeneous points was built using MySQL. This knowledge base includes information on the source, process causes, and process improvement suggestions. For example, if the heterogeneous source is TiN precipitation, the process cause is excessively high continuous casting cooling rate; the suggested process improvement is to reduce the primary cooling water flow in the crystallizer or reduce the continuous casting speed.

[0093] S6, Results Output, including information on each heterogeneity, classification, source, process factors, and process recommendations:

[0094] The system identifies heterogeneous points in the sample and outputs a list or visualization of the location, classification, and source of each heterogeneous point, along with process recommendations. For example, if the heterogeneous point originates from TiN precipitation, the process cause is excessively rapid continuous casting cooling; the recommended process improvement is to reduce the amount of cooling water in the crystallizer or to reduce the continuous casting speed.

[0095] This embodiment provides a method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection. After six months, the time required will be reduced from an average of 2 hours to less than 1 minute, with a high accuracy rate. This significantly shortens the time for heterogeneous point identification, source analysis, and process improvement. It also effectively helps reduce the monthly number of customer complaints about stamping cracks caused by heterogeneous point defects in continuously cast slabs by 60%.

[0096] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection, characterized in that: A dual-scale detection network is used to locate and classify heterogeneous point targets in metallographic images. The classification is based on the type of heterogeneous points, and also combined with the source of the heterogeneous points. It automatically identifies the source of heterogeneous points, thus providing an auxiliary tool for quickly analyzing the causes of defects caused by heterogeneous points and making process improvements.

2. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection as described in claim 1, characterized in that, The method for determining the source of heterogeneous points in the continuously cast slab is specifically... Includes the following steps: S1, acquire standardized samples and images; S2, data acquisition, cropping, filtering and noise reduction preprocessing to form the training dataset D; S3, Data annotation: Annotate the heterogeneous points according to their category and shape to generate the labeled dataset D1; S4, build and train the dual-scale detection network Y on the framework; S5, Construct a heterogeneous point expert knowledge base; S6, output the result.

3. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection as described in claim 2, characterized in that, Step S1 specifically includes: Microscopic image sampling was performed on continuously cast slab samples to obtain a sample image set X under a microscope.

4. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection as described in claim 3, characterized in that: The image sampling range is 0.5mm*0.5mm to 1mm*1mm; The image sampling factor is 100 to 500 times.

5. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection as described in claim 2, characterized in that, Step S3 specifically includes: The target detection and annotation of the heterogeneous points are performed using the EasyData annotation platform. The classification annotation is based on two dimensions: the type and the source of the heterogeneous points.

6. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection as described in claim 5, characterized in that: The types of heterogeneous points include non-metallic inclusions, metallic inclusions, and metallic precipitates; The sources of the heterogeneous points include smelting slag, protective slag, high-temperature solidification inclusions in continuous casting, and low-temperature precipitates in continuous casting.

7. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection as described in claim 6, characterized in that: The non-metallic inclusions include Al2O3, SiO2, CaO oxides, and calcium aluminates; The metallic inclusions include TiN and MnS; The metal precipitates include TiN and AlN.

8. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection as described in claim 2, characterized in that, In step S4, the dual-scale detection network Y includes a shallow detection head and a deep detection head; The shallow detection head includes small-scale detection of targets with a size of 1 to 5 mm, where the heterogeneous point size = detection size / magnification. The deep detection head includes large-scale detection of targets with a size of 5-20mm, where the heterogeneous point size = detection size / magnification.

9. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection as described in claim 2, characterized in that: In step S5, the heterogeneous point expert knowledge base includes information on the source and process causes of the heterogeneous points, as well as process improvement suggestions.

10. A method for determining the source of heterogeneous points in continuously cast slabs based on dual-scale target detection as described in claim 2, characterized in that, Step S6 specifically includes: The dual-scale detection network Y is used to determine the source of inclusions, and outputs the location, classification, source list or visualization image of each heterogeneous point. Then, the heterogeneous point expert knowledge base is integrated to output process factors and improvement suggestions.

Images