CN-121982038-A - Machine vision-based aluminum liquid mirror scum monitoring method and system

CN121982038ACN 121982038 ACN121982038 ACN 121982038ACN-121982038-A

Abstract

The invention belongs to the field of aluminum liquid mirror surface monitoring, and particularly relates to a machine vision-based aluminum liquid mirror surface scum monitoring method and system, wherein the method comprises the steps of controlling an industrial camera to acquire an aluminum liquid surface image sequence in real time; extracting a current frame, preprocessing and evaluating the definition, performing self-adaptive image enhancement if the definition is insufficient, inputting the optimized image into a pre-trained AI model, intelligently identifying defects such as scum and oxide film, evaluating the integrity of a mirror surface, calculating the comprehensive surface quality score based on the identification result, and automatically generating technological parameter adjustment instructions such as adding refining agent, stirring intensity and the like according to the defect characteristics if the score is lower than a threshold value, so as to realize closed-loop control. The invention solves the problems that manual observation is low-efficiency, and the traditional sensor cannot accurately detect the tiny defects on the surface in a non-contact manner, realizes high-resolution, real-time and online intelligent monitoring and active regulation and control on the state of the aluminum liquid mirror surface, and improves the purity of the aluminum liquid and the quality consistency of cast-rolling slabs from the source.

Inventors

FENG ZHONGLING
DANG XIN
LI XU
MA WEI
Xi Mengnan

Assignees

南京迅集科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (10)

1. The machine vision-based aluminum liquid mirror scum monitoring method is characterized by comprising the following steps of: S1, controlling an industrial camera to acquire images of the surface of an aluminum liquid melt in real time at a preset sampling frequency to obtain an original image sequence; S2, extracting a current frame image from the original image sequence, performing preliminary image preprocessing on the current frame image to obtain a first optimized image, and calculating an image definition evaluation value of the first optimized image; S3, judging whether the image definition evaluation value is larger than or equal to a preset definition threshold value; if the image is smaller than the first optimized image, performing enhancement processing on the first optimized image to generate a second optimized image, and taking the second optimized image as an image to be analyzed; If the first optimized image is larger than or equal to the first optimized image, the first optimized image is taken as an image to be analyzed; S4, inputting the image to be analyzed into a pre-trained mirror detection evaluation model to generate a surface defect identification result and a mirror state evaluation result; S5, calculating the current surface quality comprehensive score based on the surface defect identification result and the mirror surface state evaluation result; S6, judging whether the current surface quality comprehensive score is lower than a preset quality threshold value, if so, generating a control instruction set containing a refining process parameter adjustment amount according to the type and the distribution of the surface defect identification result, and if not, outputting a surface state qualified signal and returning to S1 to acquire a next frame of image; And S7, outputting the control instruction set to a control subsystem of the refining equipment so as to drive the control subsystem to execute preset refining operation, and returning to S1 to execute the next monitoring period after the operation is completed.
2. The machine vision-based molten aluminum mirror scum monitoring method of claim 1, wherein performing preliminary image preprocessing on the current frame image comprises: Dividing the current frame image into a plurality of non-overlapping image sub-blocks, and calculating the gray variance of each image sub-block; Taking the median value of the gray variance of all the image sub-blocks as a global noise intensity estimation value; determining the size and standard deviation parameters of a Gaussian filter kernel according to the global noise intensity estimated value and a preset mapping relation; And taking the current frame image as an input, performing two-dimensional convolution operation by using the Gaussian filter check image with the size and standard deviation parameters, and outputting the processed image as a noise reduction image.
3. The machine vision-based molten aluminum mirror scum monitoring method of claim 2, wherein performing preliminary image preprocessing on the current frame image further comprises: calculating a global gray histogram of the noise reduction image; constructing a gray level cumulative distribution function mapping table based on the global gray level histogram; and according to the gray level cumulative distribution function mapping table, carrying out nonlinear reassignment on the gray value of each pixel in the noise reduction image so as to expand the dynamic range of the gray level of the image and improve the overall contrast, and outputting the processed image as the first optimized image.
4. The machine vision based molten aluminum mirror scum monitoring method of claim 3 wherein calculating the image sharpness evaluation value of the first optimized image comprises: calculating a gradient amplitude image of the first optimized image in a space domain based on the first optimized image; calculating the total value of high-frequency energy of the gradient amplitude image based on the gradient amplitude image; calculating to obtain a first optimized image definition evaluation value through a preset mapping function based on the high-frequency energy total value; Judging whether the first optimized image definition evaluation value is greater than or equal to a preset definition threshold value or not comprises the following steps: if the definition evaluation value of the first optimized image is larger than or equal to a preset definition threshold value, the first optimized image is used as an image to be analyzed; And if the definition evaluation value of the first optimized image is smaller than a preset definition threshold value, performing enhancement processing on the first optimized image to generate a second optimized image, and taking the second optimized image as an image to be analyzed.
5. The machine vision based molten aluminum specular dross monitoring method of claim 4, wherein performing enhancement processing on the first optimized image comprises: Determining a required enhancement processing level based on the image definition evaluation value and a preset enhancement level threshold; And according to the enhancement processing stage number, performing one-stage or multi-stage serial image enhancement operation, enhancing the first optimized image, and outputting the processed image as the second optimized image.
6. The machine vision-based aluminum liquid mirror scum monitoring method of claim 5, wherein the mirror detection evaluation model comprises a shared feature extraction network, a defect segmentation sub-network, a mirror evaluation sub-network and an output network, wherein the shared feature extraction network is used for carrying out multi-level feature encoding on an input sample image and outputting a shared feature map, the defect segmentation sub-network is connected to the tail end of the shared feature extraction network and is used for carrying out pixel level classification based on the shared feature map and outputting a defect segmentation prediction map, the mirror evaluation sub-network is connected to the tail end of the shared feature extraction network and is used for carrying out global feature aggregation and classification based on the shared feature map and outputting a mirror state level prediction vector, and the output network is respectively connected to the defect segmentation sub-network and the tail end of the mirror evaluation sub-network and is sequentially used for carrying out post-processing on the defect segmentation prediction map to generate a structured surface defect identification result and carrying out analysis on the mirror state level prediction vector to generate a mirror state evaluation result.
7. The machine vision-based molten aluminum mirror scum monitoring method of claim 6, wherein generating the surface defect identification result and the mirror state evaluation result comprises: inputting an image to be analyzed obtained in an online monitoring process into the mirror detection evaluation model, and performing forward calculation on the input image to be analyzed to generate an intermediate result; The intermediate result comprises a defect segmentation prediction graph output by the defect segmentation sub-network and a mirror state grade prediction vector output by the mirror evaluation sub-network; performing defect decoding and structured output flow based on the defect segmentation prediction graph to generate a surface defect recognition result; and executing a grade analysis and grading flow based on the mirror state grade prediction vector to generate a mirror state evaluation result.
8. The machine vision based molten aluminum mirror scum monitoring method of claim 7 wherein the defect decoding and structured output flow comprises: Performing maximum value index operation of channel dimensions on the defect segmentation prediction graph to generate a pixel category index graph; carrying out connected domain analysis on the pixel class index map, and extracting each independent defect instance; Calculating a class label, a spatial position boundary and a pixel area of each defect instance, and summarizing the class label, the spatial position boundary and the pixel area into a structured defect list, wherein the defect list is the surface defect identification result; the grade analysis and scoring process comprises the following steps: performing maximum value index operation on the mirror state grade prediction vector, and determining a corresponding mirror integrity grade index; And converting the mirror surface integrity grade index into a quantized mirror surface integrity grade according to a preset grade-grade mapping relation, wherein the mirror surface integrity grade index is the evaluation result of the mirror surface state.
9. The machine vision based molten aluminum mirror scum monitoring method of claim 8 wherein said determining the number of enhancement treatments required includes: Acquiring a first enhancement level threshold and a second enhancement level threshold, wherein the first enhancement level threshold is smaller than the second enhancement level threshold; If the image definition evaluation value is smaller than the first enhancement level threshold, determining that the number of enhancement processing stages required is a first number of stages, wherein the first number of stages represents that two stages of image enhancement operation connected in series are required to be executed; If the image definition evaluation value is greater than or equal to the first enhancement level threshold and less than the second enhancement level threshold, determining that the required enhancement processing level is a second level, wherein the second level indicates that one-level image enhancement operation needs to be executed; If the image definition evaluation value is greater than or equal to the second enhancement level threshold, determining that the required enhancement processing level is zero; when the enhancement processing stage number is the first stage number, performing two-stage series image enhancement operation, including: performing limited contrast self-adaptive histogram equalization processing on the first optimized image to enhance the local contrast of the image and outputting a first-stage enhanced image; And a second stage operation of performing a guided filter process on the first stage enhanced image to maintain and enhance the target edge while smoothing the background region, and outputting the second optimized image.
10. A machine vision-based molten aluminum mirror scum monitoring system for implementing the machine vision-based molten aluminum mirror scum monitoring method of any one of claims 1-9, comprising: The acquisition module is used for controlling the industrial camera to acquire images of the surface of the molten aluminum melt in real time at a preset sampling frequency to obtain an original image sequence; the image processing module is used for extracting a current frame image from the original image sequence, performing preliminary image preprocessing on the current frame image to obtain a first optimized image, and calculating an image definition evaluation value of the first optimized image; The definition judging module is used for judging whether the image definition evaluation value is larger than or equal to a preset definition threshold value; if the image is smaller than the first optimized image, performing enhancement processing on the first optimized image to generate a second optimized image, and taking the second optimized image as an image to be analyzed; If the first optimized image is larger than or equal to the first optimized image, the first optimized image is taken as an image to be analyzed; The detection module is used for inputting the image to be analyzed into a pre-trained mirror detection evaluation model so as to generate a surface defect identification result and a mirror state evaluation result; the evaluation module is used for calculating the current surface quality comprehensive score based on the surface defect identification result and the mirror surface state evaluation result; The control instruction module is used for judging whether the current surface quality comprehensive score is lower than a preset quality threshold value, if so, generating a control instruction set containing the adjustment quantity of the refining process parameters according to the type and the distribution of the surface defect identification result, and if not, outputting a surface state qualified signal and returning to the acquisition module to acquire the next frame of image; And the execution feedback module outputs the control instruction set to the control subsystem of the refining equipment so as to drive the control subsystem to execute preset refining operation, and after the operation is completed, the control instruction set returns to the acquisition module to execute the next monitoring period.

Description

Machine vision-based aluminum liquid mirror scum monitoring method and system Technical Field The invention belongs to the field of aluminum liquid mirror surface monitoring, and particularly relates to a machine vision-based aluminum liquid mirror surface scum monitoring method and system. Background In the casting process of aluminum and aluminum alloy, the real-time monitoring of the surface state of aluminum liquid is a crucial task. The traditional monitoring method mainly relies on two major categories of manual observation and periodic sampling analysis, namely, the manual observation is usually carried out by operators with abundant experiences to judge the mirror surface state and scum condition of the surface of the aluminum liquid by visual observation, the method is low in efficiency and poor in continuity, is extremely easy to influence by subjective experiences of the operators, visual fatigue and on-site light environment, and causes the problems of inconsistent judgment, missed detection of tiny defects and the like, and on the other hand, with the development of an automation technology, a contact or indirect measuring method based on a sensor is gradually applied, for example, the surface state is deduced by detecting the melt resistance, temperature field distribution or an inserted probe. However, such automated measurement methods, although improving the objectivity of the data to a certain extent, have obvious limitations that firstly, the contact sensor needs to work for a long time in a high-temperature and corrosive aluminum liquid environment, so that the requirements on high temperature resistance and erosion resistance of the sensor material are extremely high, the equipment cost is expensive, the service life is limited, the installation and maintenance are inconvenient, secondly, the indirect measurement method often only reflects local or comprehensive parameters, and it is difficult to directly and intuitively capture the overall visual characteristics of the aluminum liquid surface, such as the form and position information of specific defects of tiny scum particle distribution, oxide film cracking form or tiny ripples, and the existing monitoring system generally only can provide simple alarm or numerical feedback, lacks intelligent evaluation and depth analysis on the integrity of the surface state, and cannot provide visual and comprehensive data support for the refined adjustment of the refining process parameters. The problems of the prior art are that in the complex industrial site of molten aluminum casting, due to the interference of strong specular reflection, environmental thermal disturbance and uneven illumination on the surface of the molten aluminum, the directly collected image often has serious quality problems of high light overexposure, fuzzy details, low contrast, dense noise and the like, so that the follow-up algorithm is difficult to reliably distinguish real micro scum, oxide film defects and image artifacts formed by reflection or interference, and the accurate sensing and quantitative analysis of the surface state based on machine vision are restricted. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a machine vision-based aluminum liquid mirror scum monitoring method and a machine vision-based aluminum liquid mirror scum monitoring system, wherein the method comprises the steps of controlling an industrial camera to acquire an aluminum liquid surface image sequence in real time; extracting a current frame, preprocessing and evaluating the definition, performing self-adaptive image enhancement if the definition is insufficient, inputting the optimized image into a pre-trained AI model, intelligently identifying defects such as scum and oxide film, evaluating the integrity of a mirror surface, calculating the comprehensive surface quality score based on the identification result, and automatically generating technological parameter adjustment instructions such as adding refining agent, stirring intensity and the like according to the defect characteristics if the score is lower than a threshold value, so as to realize closed-loop control. The invention solves the problems that manual observation is low-efficiency, and the traditional sensor cannot accurately detect the tiny defects on the surface in a non-contact manner, realizes high-resolution, real-time and online intelligent monitoring and active regulation and control on the state of the aluminum liquid mirror surface, and improves the purity of the aluminum liquid and the quality consistency of cast-rolling slabs from the source. In order to achieve the above purpose, the present invention provides the following technical solutions: The machine vision-based aluminum liquid mirror scum monitoring method comprises the following steps: S1, controlling an industrial camera to acquire images of the surface of an aluminum liquid melt in real time at a preset samp