CN-122023240-A - Method, equipment and storage medium for identifying low-power defects of continuous casting billets

CN122023240ACN 122023240 ACN122023240 ACN 122023240ACN-122023240-A

Abstract

The embodiment of the invention provides a method, equipment and a storage medium for identifying a low-power defect of a continuous casting billet, belonging to the field of continuous casting billet defect identification. The identification method comprises the steps of obtaining a low-power image of a continuous casting billet, converting the image into a target color space to extract color features, extracting spatial features of different levels of the low-power image of the continuous casting billet based on a defect identification model, carrying out multi-mode feature fusion on the spatial features and the color features, carrying out multi-scale feature fusion according to the features after the multi-mode feature fusion to obtain feature images of different scales, and outputting defect identification results and performance prediction results according to the feature images of different scales. The invention improves the detection rate and the distinguishing degree of the defects of the model sensitive to the color by comprehensively utilizing the space texture and the color characteristics, and the output result also comprises a performance prediction result, thereby providing a deeper decision basis for process optimization.

Inventors

LI SHIYU
LIU SHUAI
XU YI
MA JUN
Cheng Shihang

Assignees

中冶京诚工程技术有限公司
中冶京诚数字科技(北京)有限公司

Dates

Publication Date: 20260512
Application Date: 20251217

Claims (10)

1. A method for identifying a low-power defect of a continuous casting billet, the method comprising: acquiring a low-power image of a continuous casting blank, and converting the image into a target color space to extract color characteristics; Based on a defect identification model, extracting spatial features of different levels of the continuous casting billet low-power image, wherein the spatial features comprise first to fourth features, and the first to fourth features are spatial features of different levels with sequentially decreasing spatial resolution and sequentially enhanced semantic information; Carrying out multi-mode feature fusion on the first feature and the color feature; combining the features after the multi-mode feature fusion with the second to fourth features, performing multi-scale feature fusion to obtain a plurality of feature graphs with different scales, and And outputting a defect identification result and a performance prediction result according to the feature diagrams with different scales.
2. The method of claim 1, further comprising, prior to extracting the spatial features of the low-power image of the continuous casting billet based on the defect recognition model: And normalizing the continuous casting billet low-power image, and uniformly adjusting the normalized continuous casting billet low-power image and the color characteristics to a fixed size.
3. The method according to claim 1, wherein extracting spatial features of different levels of the continuous casting billet low power image based on the defect recognition model comprises: based on a main network of the defect recognition model, extracting features of the continuous casting billet low-power image, and carrying out nonlinear transformation on second features through Mlpcon layers of the main network so as to enhance nonlinear characterization capability of the second features on complex defect features; wherein the Mlpcon layers include a plurality of fully-connected layers and a nonlinear activation function.
4. The method of identifying of claim 1, wherein said multi-modal feature fusion of the first feature with the color feature comprises: And carrying out multi-mode feature fusion on the first feature and the color feature in a channel splicing or attention weighted fusion mode.
5. The method according to claim 1, wherein outputting the defect recognition result and the performance prediction result according to the feature maps of the different scales includes: And generating candidate frames by adopting an anchor point mechanism according to the feature diagrams with different scales, and realizing mask prediction in each candidate frame through ROIAlign and a full convolution network to obtain the defect identification result, wherein the defect identification result comprises the form of the defect.
6. The method according to claim 1, wherein outputting the defect recognition result and the performance prediction result according to the feature maps of the different scales includes: carrying out global average pooling on the P4 layer feature images in the feature images with different scales to obtain feature vectors; and inputting the characteristic vector into a fully-connected network to obtain the performance prediction result, wherein the performance prediction result comprises the average carbon content or segregation level of the defects.
7. The identification method according to claim 1, wherein the process of creating the training data set of the defect identification model comprises: Acquiring the continuous casting billet low-power image and carbon content distribution measurement data of a continuous casting billet sample of at least part of the area, and carrying out position registration on the continuous casting billet low-power image and the carbon content distribution measurement data to obtain performance label information; And marking the defects in the continuous casting billet image to obtain defect label information, wherein the defect label information comprises the type, the position and the contour information of the defects.
8. The identification method of claim 7, wherein the process of creating the training data set of the defect identification model further comprises: Performing data enhancement on the continuous casting billet low-power image by adopting a data enhancement strategy, wherein the data enhancement strategy comprises one or more of geometric transformation, color transformation and metal enhancement; randomly adjusting one or more of brightness, contrast, saturation, and hue of the continuous casting billet low power image within a current color space and the target color space, and/or One or more of random noise, gaussian blur, cutout are introduced in the continuous casting billet low-power image.
9. A strand macroscopic defect identification device, characterized in that the strand prediction device comprises: Memory, and A processor configured to execute instructions stored in the memory to perform the strand macro defect identification method according to any one of claims 1-8.
10. A machine-readable storage medium having instructions stored thereon for causing a machine to perform the strand low power defect identification method of any one of the above-described applications.

Description

Method, equipment and storage medium for identifying low-power defects of continuous casting billets Technical Field The invention relates to the technical field of continuous casting billet defect identification, in particular to a continuous casting billet macroscopic defect identification method, continuous casting billet macroscopic defect identification equipment and a storage medium. Background The continuous casting billet may form various low-power defects such as shrinkage cavity, center crack, subcutaneous crack, intermediate crack, segregation, nonmetallic inclusion and the like during the production process. These defects directly affect the quality and performance of the material, so that low-power defect detection is a key link of quality control in steel production. Traditional detection methods rely mainly on manual visual inspection for comparison and rating. However, the manual detection has the problems of strong subjectivity, low detection efficiency, inconsistent standard execution and the like, often causes high defect misjudgment rate, has long single sample detection time, and is difficult to meet the requirements of modern steel production on efficient and accurate quality detection. In order to overcome the defects of manual detection, the prior art provides an automatic detection method based on image processing and machine learning. For example, in one scheme, a weighted adaptive median filtering algorithm is adopted to preprocess a continuous casting billet low-power image, a suspected defect area is extracted through gray average value threshold segmentation, manual characteristics such as circularity, target gray standard deviation and the like are extracted, and finally a single hidden layer BP neural network is used for classification. The identification accuracy of the scheme for certain defects such as center segregation can reach 83% in a laboratory environment, but manual characteristic design depends on expert experience, so that the scheme is difficult to adapt to special defect forms of different steel grades, a shallow neural network is difficult to learn complex nonlinear characteristics, the confusion rate of defects such as cracks and scratches in an actual production line is high, fine defect details are easy to blur by median filtering, and the small-size defect detection rate is low. The other scheme is to build a knowledge mask matrix of the technological parameters to enhance the characteristics of the image and to use Vision Transformer model for identification. The method has the advantages that higher accuracy is obtained in the corner crack identification task, but the ViT model has extremely high calculation complexity on high-resolution images, the real-time detection requirement of a production line is difficult to meet, a fixed knowledge mask cannot adapt to dynamic changes of process parameters, the model robustness is insufficient, and in addition, a large number of labeling samples are needed for training, so that the method is not suitable for the condition of limited sample quantity in actual production. Therefore, the prior art still has the problems of limited detection precision, insufficient feature utilization, weak model generalization capability, difficulty in realizing the correlation analysis of defect morphology and material performance and the like, and a precise, efficient and comprehensive-information intelligent continuous casting billet low-power defect identification method is needed. Disclosure of Invention An object of an embodiment of the present invention is to provide a method, apparatus and storage medium for identifying a low-power defect of a continuous casting billet, which are used for solving the problems in the background art. In order to achieve the above purpose, the embodiment of the invention provides a continuous casting billet low-power defect identification method, which comprises the steps of obtaining a continuous casting billet low-power image, converting the image into a target color space to extract color features, extracting spatial features of different levels of the continuous casting billet low-power image based on a defect identification model, wherein the spatial features comprise first to fourth features, the first to fourth features are spatial features of different levels with sequentially decreasing spatial resolution and sequentially enhanced semantic information, carrying out multi-mode feature fusion on the first feature and the color features, carrying out multi-scale feature fusion on the features after the multi-mode feature fusion, combining the second to fourth features, so as to obtain feature images of different scales, and outputting a defect identification result and a performance prediction result according to the feature images of different scales. Optionally, before the spatial features of the continuous casting billet low-power image are extracted based on the defect identification