A Smart Identification Method for Iron and Steel Smelting Fume Levels Based on the Allen-Cahn Phase Field Energy Equation

Abstract

The application provides a steel smelting smoke grade intelligent identification method based on an Allen-Cahn phase field energy equation, and comprises the following steps: S1: performing initial classification on an image through GMM clustering to obtain an initial class label; S2: iteratively screening to construct a training set; S3: extracting Allen-Cahn energy features; S4: performing multi-classification through phase field evolution; S5: performing boundary screening to optimize samples; and S6: constructing a model, training, and identifying a smoke grade; the whole process of the application forms a complete data cleaning-feature extraction-classification-verification closed loop, automatically reduces noise, optimizes samples, reduces artificial labeling cost, realizes high-precision, strong anti-interference, fast response, full-adaptive industrial-grade smoke grade identification, and realizes real-time monitoring without manual attendance.

Description

Technical Field

[0001] This invention relates to the field of industrial image recognition technology, and in particular to an intelligent recognition method for the level of smoke from steel smelting based on the Allen-Cahn phase field energy equation. Background Technology

[0002] In blast furnace smelting, smoke concentration is a key indicator reflecting the reaction state inside the furnace, judging the smelting intensity, and assessing environmental emission risks. Excessively high smoke concentration may indicate vigorous reactions inside the furnace, uneven material distribution, or sealing failure, easily leading to flue gas overflow and exceeding environmental standards; excessively low concentration may be related to insufficient smelting load, affecting production efficiency. Therefore, achieving rapid and accurate identification of smoke concentration levels is a core requirement for intelligent blast furnace production and environmental management. However, smoke identification in blast furnace scenarios faces many unique challenges: blast furnace flue gas contains a large amount of dust and water vapor, resulting in irregular smoke morphology and blurred boundaries; complex lighting conditions during smelting (such as strong sunlight during the day and firelight reflection at the furnace mouth at night); high temperatures and airflow disturbances accompanying smoke emissions cause severe deformation of smoke movement; and interference factors such as equipment obstruction around the chimney and furnace mouth exist. These factors lead to the following problems with existing smoke concentration level identification methods: (1) Insufficient capture of complex features of blast furnace smoke: Existing algorithms do not fully consider the dynamic motion deformation, blurred boundaries, and irregular shape of blast furnace smoke, resulting in poor feature extraction and low accuracy of grade recognition.

[0003] (2) Problem of weak anti-interference ability in complex scenes: There are sudden changes in light, equipment obstruction, dust and water vapor interference in the blast furnace scene. Existing algorithms are easily interfered with and may lead to misjudgment.

[0004] (3) The problem of balancing real-time performance and accuracy: Traditional algorithms have strong real-time performance but low accuracy, while deep learning algorithms have high accuracy but insufficient real-time performance, which cannot simultaneously meet the dual requirements of "second-level response" and "high accuracy" in the blast furnace scenario.

[0005] (4) The problem of ambiguous boundary definition of smoke level: The concentration of blast furnace smoke has a continuous transition characteristic. Existing algorithms lack clear basis for the boundary division of light-medium and medium-severe levels, which is prone to misjudgment of level. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent identification method for the smoke level of steel smelting based on the Allen-Cahn phase field energy equation.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent identification of smog levels in steel smelting based on the Allen-Cahn phase-field energy equation includes the following steps: S1: Initial classification of the image is performed using GMM clustering to obtain initial category labels; S2: Iteratively filter and construct the training set; Based on the GMM clustering results, key thresholds were extracted, and three sets were obtained after two rounds of iteration. S3: Extract Allen-Cahn energy features; Calculate the energy features of each sample in the training set at two different scales to form a two-dimensional feature vector; S4: Phase field evolution is classified into multiple categories; S5: Boundary screening to optimize samples; S6: Build and train the model to identify smoke levels; A lightweight classification model was built based on the extracted dual-scale Allen-Cahn energy features. The selected samples were used as the training set to train the lightweight classification model, which was then used for level recognition of actual smoke images.

[0008] Furthermore, step S1 includes: S11: Acquire smoke images; Images of the furnace opening or chimney area are captured using an industrial camera, and then processed to grayscale. The ROI area is fixed and cropped to obtain smoke images. S12: Background subtraction method for calculating image differences; Based on the smoke image, define a smoke-free background image I. bg According to the formula: Calculate the i-th image I i Image I with a smoke-free background bg The difference; where M and N are the height and width of the image; It is the gray value at (x,y) of the smokeless background image; With a preset difference threshold tol2=3, the difference between the current image and the smoke-free background image is compared to the difference threshold. If the difference is less than the threshold, the current image is updated as the new smoke-free background image I. bg Calculate the dissimilarity of all images to obtain a one-dimensional feature sequence; S13: Initial clustering yields initial category labels; Gaussian mixture model clustering is performed on the one-dimensional feature sequence. The number of clusters is set to 3. The posterior probability of each image belonging to each category is calculated, and the category with the highest posterior probability is used as the initial category label of the image. The initial category label of each image is obtained by traversing the sequence.

[0009] Furthermore, step S2 includes: S21: Extract key thresholds; From the GMM clustering results, obtain two breakpoints between the three categories, and extract the breakpoints as two key thresholds flag(1) and flag(2). S22: Set the iteration threshold; Set two iteration thresholds tol1 and tol2, where tol1 takes flag(1) and flag(2) in turn, and tol2 = 0.6 * flag(1); S23: The first round of iterative screening; In the first round of iteration, use flag(1) as the threshold for screening: Compare the difference degree in the feature sequence with flag(1). Images with a difference degree > flag(1) are determined to have smoke and stored in fogStore; images with a difference degree < flag(1) are determined to have no smoke and stored in NofogStore; at the same time, update the no-smoke background image when err < tol2; S24: The second round of iterative screening; In the second round of iteration, use flag(2) as the threshold to further screen the smoky images in fogStore from the first round: Images with a difference degree > flag(2) are determined to be heavy smoke and stored in fs3; the rest are medium smoke and stored in fs2; After two rounds of iteration, finally obtain three sets: the NofogStore set, the fs2 set, and the fs3 set as the samples of the training set.

[0010] Further, step S3 is specifically as follows: For the grayscale image of the sample, according to the formula: Calculate the Allen-Cahn energy; where M is the height of the grayscale image and N is the width of the grayscale image; is the grayscale value on the grayscale image (i,j), h is the grid step size, set two values of h = 1 and h = 2, for each sample in the training set, calculate two different scale energy features E1 and E2 of the sample, and form a two-dimensional feature vector (E1, E2).

[0011] Further, step S4 includes: First, discretize the feature space into a grid of Nx * Ny, and each grid point corresponds to a feature vector (E1, E2); Define three phase field variables u k (i,j), k = 1, 2, 3, respectively represent the probability that the grid point (i,j) belongs to the kth class, and satisfy ; Then, determine the initial phase field distribution according to the samples of the training set ; within a neighborhood with a radius of r around the sample point, set the phase field value of the class to which the sample belongs to 1, and other classes to 0; Then, the Allen-Cahn phase field evolution iteration is performed; according to the formula: Iterative evolution; where , is the derivative of the double-well potential. This is the interface width parameter, a preset hyperparameter; These are the coefficients of the data fitting term, and also the preset hyperparameters; It is the initial phase field distribution; Preset hyperparameters for the time iteration step; It is the phase field value belonging to the k-th class in the nth iteration; It is the phase field value belonging to the k-th class in the (n+1)th iteration; It is the Laplace operator; The iteration proceeds for 260 steps, with intermediate results output every 26 steps. Finally, the category with the largest phase field value at each grid point is taken as the classification result.

[0012] Furthermore, step S5 includes: First, the classification results of the phase field are saved as an image output.jpg. The gray value of each pixel in the image output.jpg represents the category to which that location belongs. Then, for each sample in the training set, the corresponding pixel position in the image output.jpg is found based on the feature vector (E1, E2), and the gray value at that position is read; the initial class label of the sample is obtained, using gray values ​​29, 76, and 150 to represent class 1, class 2, and class 3, respectively; based on the gray value obtained from the initial class label, the absolute value of the difference between the classification result at that position and the gray value corresponding to the initial class label of the sample is calculated. If the absolute value of the difference does not exceed 6, the sample is retained; otherwise, it is discarded, resulting in the filtered samples. Furthermore, step S6 includes: S61: Design the model structure; A multilayer perceptron is used as the classification network. The input is a two-dimensional energy feature vector (E1, E2), and the output is three smoke levels. The network consists of two hidden layers and one output layer, and the activation function is Tanh. S62: Training the model; The samples after phase field filtering are used as the training set, the input features are the Allen-Cahn dual-scale energies (E1, E2) corresponding to the image; the labels are the categories obtained by iterative filtering. The model is trained and saved as a recognition model after training for subsequent real-time inference. S63: Real-time reasoning and classification; First, an industrial camera captures real-time images of smoke from the furnace opening or chimney area as input. The input images are then preprocessed and the Region of Interest (ROI) is extracted. Next, the eigenvectors (E1, E2) are calculated using the Allen-Cahn energy function; The feature vector is then input into the recognition model, which infers the smoke level classification result directly. Save and categorize the images based on the classification results.

[0013] Furthermore, Nx = Ny = 100; determine the initial phase field distribution. When r = 0.3h, h is the grid step size.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention uses the energy functional in the phase field model to characterize the intrinsic structural features of the smoke image, and realizes the unsupervised classification of smoke level through the distribution law of energy. Using the phase field physical model to characterize the smoke structure is more in line with the characteristics of the blurred boundary and irregular shape of blast furnace smoke than artificial features / general CNN, the level classification is more accurate, and the recognition accuracy is high. (2) The present invention adaptively updates the background, ensuring that the background model can adapt to the changes in the environment and adapt to the slow changes in ambient light. It can resist sudden changes in light, dust and water vapor, equipment obstruction, high temperature radiation interference, significantly reduce the misjudgment rate, and has strong anti-interference ability. (3) The present invention only extracts two-dimensional energy features and uses a lightweight MLP, which has a much smaller computational load than general deep learning, fast inference speed, high accuracy and good real-time performance. While ensuring high level classification accuracy, it perfectly solves the contradiction of traditional deep learning models being complex, computationally intensive and lacking real-time performance, and can provide timely and reliable data support for closed-loop control such as induced draft fan regulation and smelting parameter optimization. (4) This invention automatically generates smooth classification boundaries through phase field evolution, which solves the problem of boundary ambiguity caused by continuous concentration transition. The level determination is stable and consistent, and the boundary division is scientific and clear. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of the intelligent identification method for iron and steel smelting fume levels based on the Allen-Cahn phase field energy equation of the present invention. Figure 2 This is a flowchart of the intelligent identification method for iron and steel smelting fume levels based on the Allen-Cahn phase field energy equation of the present invention. Figure 3 This is a schematic diagram of the intelligent identification method for the smoke level of steel smelting based on the Allen-Cahn phase field energy equation of the present invention. Detailed Implementation

[0016] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0017] A method for intelligent identification of smog levels in steel smelting based on the Allen-Cahn phase-field energy equation includes the following steps: S1: Initial classification of the image is performed using GMM clustering to obtain initial category labels; Smoke images were acquired, and preliminary features of the smoke were extracted using the background subtraction method. The difference between all images was calculated to obtain a feature sequence. The feature sequence was then clustered using a Gaussian mixture model (GMM) to obtain the initial class label for each image. S11: Acquire smoke images; Images of the furnace opening or chimney area are acquired using an industrial camera, converted to grayscale, and the Region of Interest (ROI) is fixed and cropped. Irrelevant and interfering areas such as equipment, steel frames, and the sky are removed to obtain smoke images. The ROI is the target area of ​​the smoke from the furnace opening / chimney.

[0018] The smoke image is a grayscale image.

[0019] S12: Background subtraction method for calculating image differences; Based on the smoke images, let the first image be the initial smoke-free image, serving as the smoke-free background image I. bg Then the i-th image I i The difference between the background and the background, err(i), is defined as: Where M and N are the image height (number of pixel columns) and width (number of pixel rows). Height refers to the total number of pixels in the y-direction; width refers to the total number of pixels in the x-direction. It is the grayscale value at (x, y) in the i-th frame of the image. This is the grayscale value at (x, y) in the smoke-free background image. Calculate image I according to the formula. i With background image I bg The degree of difference err(i); The difference err(i) is essentially the root mean square error (RMSE), which measures the overall difference between the current image and the background image. For smoke-free images, err(i) is close to 0; as the smoke concentration increases, err(i) gradually increases. Background subtraction quickly distinguishes between smoke-free and smoke-containing images, with extremely fast computation, ensuring real-time performance.

[0020] With a preset difference threshold of tol2=3, the difference score err(i) between the current image and the smoke-free background image is compared with the difference threshold. If the difference score between the current image and the background image is less than the difference threshold, the image is considered to still belong to the smoke-free background, and the current image is updated as the new smoke-free background image I. bg The difference between subsequent images and the background image is then calculated. After calculating the difference between all images, a one-dimensional feature sequence is obtained. .

[0021] By introducing a background update mechanism, the background model can be adaptively adjusted to follow environmental changes and adapt to slow variations in ambient lighting. This avoids the failure of a fixed background and improves long-term stability.

[0022] S13: Initial clustering yields initial category labels; For one-dimensional feature sequences Gaussian Mixture Model (GMM) clustering is performed, with the number of clusters set to 3 (corresponding to three smoke levels). The posterior probability of each image belonging to each category is calculated, and the category with the highest posterior probability is used as the initial category label for that image. This process is repeated for each image to obtain its initial category label automatically, reducing the need for extensive manual annotation and lowering labor costs.

[0023] The mathematical principle of Gaussian Mixture Model (GMM) clustering is based on the assumption that the data is generated by a mixture of several Gaussian distributions. The Expectation-Maximization (EM) algorithm is used to estimate the parameters of each Gaussian distribution (the mean μ of the three Gaussian distributions). k σ k (k=1,2,3)), and then determine the posterior probability of each sample belonging to each class. Based on the magnitude of the three posterior probabilities, select the class with the largest posterior probability as the initial class label of the image.

[0024] Posterior probability is the confidence that an image belongs to a certain cluster.

[0025] S2: Iteratively filter and construct the training set; Based on the GMM clustering results, key thresholds were extracted, and after two rounds of iteration, three sets were obtained.

[0026] S21: Extract key thresholds; From the GMM clustering results, two boundary points between the three categories are obtained, and these boundary points are extracted as two key thresholds, flag(1) and flag(2). Based on the central mean values ​​μ1, μ2, and μ3 of the three categories, the midpoint of the mean values ​​of adjacent categories is taken as the boundary threshold. ; .

[0027] S22: Set the iteration threshold; Set two iteration thresholds, tol1 and tol2, where tol1 takes flag(1) and flag(2) in sequence, and tol2 = 0.6 * flag(1).

[0028] S23: First round of iterative screening; In the first round of iteration (j = 1), screening is performed with flag(1) as the threshold: compare the difference degree in the feature sequence with the value of flag(1). Images with a difference degree err > flag(1) are determined to have smoke and are stored in fogStore; images with a difference degree err < flag(1) are determined to have no smoke and are stored in NofogStore; all images are divided into two major categories: no smoke / low smoke and having smoke. At the same time, when err < tol2, the no-smoke background image is updated.

[0029] S24: Screening in the second round of iteration; In the second round of iteration (j = 2), further screening is performed from the images with smoke in the first round with flag(2) as the threshold: images with a difference degree err > flag(2) are determined to have heavy smoke and are stored in the new fogStore; the rest are medium smoke.

[0030] After two rounds of iteration, three sets are finally obtained: the NofogStore (no smoke / low smoke) set, the fs2 (medium smoke) set, and the fs3 (heavy smoke) set. These three sets are used as samples for the training set. Through the threshold screening in two rounds of iteration, a high-purity and high-reliability sample set of no smoke / low smoke, medium smoke, and heavy smoke is obtained, improving the quality of training data from the source, avoiding interference caused by incorrect labeling to subsequent feature extraction and model training, and enhancing the overall stability and classification accuracy of this method.

[0031] Through the adaptive background update mechanism and iterative sample screening, it can automatically adapt to dynamic disturbances such as slow changes in the illumination of the blast furnace scene, working condition switching, and equipment occlusion, overcome the defects of poor dynamic adaptability and limited scene adaptation of existing methods, and meet the needs of long-term unattended industrial monitoring.

[0032] S3: Extract Allen-Cahn energy features; For the grayscale image u(x, y) of the sample, according to the formula: Calculate the Allen-Cahn energy; where M is the height of the grayscale image and N is the width of the grayscale image; is the grayscale value at the image (i, j), 、 、 、 Similarly; h is the grid step size, used to control the nonlinear degree of the double-well potential, and two values of h = 1 and h = 2 are set; two energy features E1 and E2 at different scales are obtained respectively; for each sample in the training set, calculate its two energy features to form a two-dimensional feature vector (E1, E2).

[0033] Multi-scale features of images are extracted using the energy functional of the Allen-Cahn phase-field model. This method can accurately characterize the concentration distribution, texture complexity, and boundary morphology of blast furnace smoke, effectively distinguishing smoke from interference such as dust, water vapor, and flame reflection. It addresses the shortcomings of traditional manual features and general visual features in representing dynamically deformed smoke, providing core support for high-precision classification.

[0034] S4: Phase field evolution is classified into multiple categories; First, the feature space is discretized into an Nx*Ny grid (Nx=Ny=100), with each grid point corresponding to an eigenvector (E1,E2). Three phase field variables u are defined. k (i,j), k=1,2,3, represent the probabilities that grid point (i,j) belongs to the k-th class, satisfying the following conditions: .

[0035] Next, the initial phase field distribution is determined based on the samples in the training set. Within a neighborhood of radius r around a sample point, the phase field value of the sample's class is set to 1, and 0 for other classes. The value of r is related to the grid spacing; we take r = 0.3h, where h is the grid step size. A sample point refers to the grid point corresponding to the sample's feature vector.

[0036] Then, Allen-Cahn phase field evolution iterations are performed. The evolution equations are discretized using an explicit Euler scheme: in, , is the double-well potential function. , is the derivative of the double-well potential. It is the interface width parameter, which is obtained through experiments or analytical fitting. It is a preset hyperparameter used to control the sharpness of the interface. These are the coefficients of the data fitting term, which are also preset hyperparameters, determined through cross-validation or theoretical analysis. It is the initial phase field distribution. The time iteration step can be preset using hyperparameters. It is the phase field value belonging to the k-th class in the nth iteration; It is the phase field value belonging to the k-th class in the (n+1)th iteration; It is the Laplace operator (diffusion term), calculated via numerical difference. Lagrange multipliers ( Ensure that the sum of the three phase fields is always 1.

[0037] The iteration proceeds for 260 steps, outputting an intermediate result every 26 steps. As the iteration continues, the phase-field system gradually evolves to a steady state, and the interfaces between different categories become clearer and smoother. Finally, the category with the largest phase-field value is taken as the classification result for each grid point. Thus, the classification decision boundary is obtained over the entire feature space.

[0038] The classification problem is transformed into a physical evolution process of minimizing phase field energy, which automatically generates smooth and clear category decision boundaries, scientifically solving the problem of ambiguous level definition caused by continuous transition of smoke concentration; the phase field probability distribution satisfies the normalization constraint, the classification results are stable and conflict-free, which greatly improves the classification robustness of the boundary region and reduces level misjudgment.

[0039] S5: Boundary screening to optimize samples; The training samples are optimized and screened based on the phase field classification results, and samples near the classification decision boundary are removed.

[0040] First, the phase field classification results are saved as an image output.jpg. The gray value of each pixel in the image output.jpg represents the category to which that location belongs.

[0041] Then, for each sample in the training set, the corresponding pixel position in the image output.jpg is found based on its feature vector (E1, E2), and the gray value at that position is read. The initial class label for the sample is obtained, using a gray value of 29 to represent class 1 (no smoke / low smoke), 76 to represent class 2 (moderate smoke), and 150 to represent class 3 (heavy smoke). Based on the gray values ​​obtained from the initial class labels, the absolute value of the difference between the classification result at that position and the gray value corresponding to the sample's initial class label is calculated. If the absolute value of the difference does not exceed 6, the sample is retained; otherwise, it is discarded, resulting in high-quality samples after filtering. These retained samples constitute the final training set for subsequent classification model training.

[0042] This screening mechanism performs self-consistency verification on training samples based on the phase field evolution classification results, removes abnormal samples that are inconsistent with the predictions of the physical model, retains samples that are consistent with the predictions of the physical model (Allen-Cahn phase field), purifies the training set, improves the reliability of the training dataset, enhances the quality of the training set, strengthens the generalization ability and industrial scenario adaptability of the final classification model, and ensures the accuracy and reliability of subsequent model training.

[0043] S6: Build and train the model to identify smoke levels; A lightweight classification model is built based on the extracted dual-scale Allen-Cahn energy features (E1, E2). The selected high-quality samples are used as the training set to train the lightweight classification model. After training, it is used for level recognition of actual smoke images.

[0044] S61: Design the model structure; A simple multilayer perceptron (MLP) is used as the classification network. The input is two-dimensional energy features (E1, E2), and the output is three smoke levels (no smoke / low smoke, moderate smoke, and heavy smoke). The network consists of two hidden layers and one output layer, and the Tanh activation function is used to ensure smooth feature mapping.

[0045] S62: Training the model; High-quality samples selected through phase field filtering are used as the training set. The input features are the Allen-Cahn dual-scale energies (E1, E2) corresponding to the image, and the labels are the categories obtained through iterative filtering. The model is trained using a standard supervised learning process, enabling the model to learn the mapping relationship between energy features and smoke levels. After training, the model is saved as a recognition model for subsequent real-time inference. S63: Real-time reasoning and classification; First, an industrial camera captures real-time images of smoke from the furnace opening or chimney area as input. The input images are then preprocessed and the Region of Interest (ROI) is extracted. Preprocessing includes grayscale conversion.

[0046] The ROI (Region of Interest) is the target area of ​​the smoke. It can be obtained by selecting a preset target area.

[0047] Next, the Allen-Cahn energy function is used to calculate the dual-scale features (E1, E2); Input the feature vectors into the trained MLP model, and directly output the smoke level classification results; Images are automatically saved and categorized based on the classification results, enabling unattended real-time monitoring.

[0048] Employing a lightweight multilayer perceptron that only requires two-dimensional features as input, the model has a simple structure, low computational cost, and short inference time per frame, meeting the real-time requirement of ≤1 second for blast furnace scenarios. After training, it can directly output smoke level results, enabling unattended real-time monitoring of the entire process. It is also convenient to link with systems such as induced draft fan control and smelting parameter optimization, supporting closed-loop industrial control.

[0049] The entire process of this invention forms a complete closed loop of data cleaning, feature extraction, classification, and verification. It automatically reduces noise, optimizes samples, lowers the cost of manual annotation, and achieves high-precision, strong anti-interference, fast response, and fully adaptive industrial-grade smoke level recognition.

[0050] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method for intelligent identification of smog levels in steel smelting based on the Allen-Cahn phase-field energy equation, characterized in that: It includes the following steps: S1: Initially classify the image through GMM clustering to obtain initial class labels; S2: Iteratively screen and construct the training set; Extract key thresholds according to the GMM clustering results, and after two rounds of iteration, obtain three sets; S3: Extract Allen-Cahn energy features; Calculate the energy features of each sample in the training set at two different scales to form a two-dimensional feature vector; S4: Perform multi-classification through phase field evolution; S5: Optimize samples through boundary screening; S6: Construct and train a model to identify the smoke level; Build a lightweight classification model based on the extracted dual-scale Allen-Cahn energy features, use the screened samples as the training set to train the lightweight classification model, and after training, use it for the level identification of actual smoke images.

2. The intelligent identification method for iron and steel smelting fume levels based on the Allen-Cahn phase-field energy equation as described in claim 1, characterized in that: Step S1 includes: S11: Collect smoke images; Collect images of the furnace mouth or chimney area through an industrial camera, perform grayscale processing, and fixedly intercept the ROI area to obtain smoke images; S12: Calculate the difference degree of the image by the background difference method; Based on the smoke image, define a smoke-free background image I. bg According to the formula: Calculate the i-th image I i Image I with a smoke-free background bg The difference; where M and N are the height and width of the image; It is the gray value at (x,y) of the smokeless background image; With a preset difference threshold tol2=3, the difference between the current image and the smoke-free background image is compared to the difference threshold. If the difference is less than the threshold, the current image is updated as the new smoke-free background image I. bg Calculate the dissimilarity of all images to obtain a one-dimensional feature sequence; S13: Obtain initial class labels through initial clustering; Perform Gaussian mixture model clustering on the one-dimensional feature sequence, set the number of clusters to 3, calculate the posterior probability of each image belonging to each category, and use the category with the largest posterior probability as the initial class label of the image, and traverse to obtain the initial class label of each image.

3. The intelligent identification method for iron and steel smelting fume levels based on the Allen-Cahn phase-field energy equation as described in claim 1, characterized in that: Step S2 includes: S21: Extract key thresholds; Obtain two break points between three categories from the GMM clustering results, and extract the break points as two key thresholds flag(1) and flag(2); S22: Set iteration thresholds; Set two iteration thresholds tol1 and tol2, where tol1 takes flag(1) and flag(2) in turn, and tol2 = 0.6 * flag(1); S23: First-round iterative screening; In the first round of iteration, screen with flag(1) as the threshold: compare the difference degree in the feature sequence with flag(1), images with a difference degree > flag(1) are determined to have smoke and are stored in fogStore; images with a difference degree < flag(1) are determined to have no smoke and are stored in NofogStore; at the same time, update the smoke-free background image when err < tol2; S24: Second-round iterative screening; In the second round of iteration, further screen the smoky images in the first-round fogStore with flag(2) as the threshold: images with a difference degree > flag(2) are determined to be heavy smoke and are stored in fs3; the rest are medium smoke and are stored in fs2; after two rounds of iteration, finally obtain three sets: the NofogStore set, the fs2 set, and the fs3 set as the samples of the training set.

4. The intelligent identification method for iron and steel smelting fume levels based on the Allen-Cahn phase-field energy equation as described in claim 1, characterized in that: Step S3 is specifically: For the grayscale image of the sample, according to the formula: Calculate the Allen-Cahn energy; where M is the height of the grayscale image and N is the width of the grayscale image; is the gray value on the grayscale image (i,j), and h is the grid step size. Two values ​​are set: h=1 and h=2. For each sample in the training set, the energy features E1 and E2 at two different scales of the sample are calculated to form a two-dimensional feature vector (E1,E2).

5. The intelligent identification method for iron and steel smelting fume levels based on the Allen-Cahn phase field energy equation as described in claim 4, characterized in that: Step S4 includes: First, the feature space is discretized into an Nx*Ny grid, with each grid point corresponding to an eigenvector (E1, E2); three phase field variables u are defined. k (i,j), k=1,2,3, represent the probabilities that grid point (i,j) belongs to the k-th class, satisfying the following conditions: ; Next, the initial phase field distribution is determined based on the samples in the training set. Within a neighborhood of radius r around a sample point, set the phase field value of the sample's class to 1, and the value of other classes to 0. Then, the Allen-Cahn phase field evolution iteration is performed; according to the formula: Iterative evolution; where , is the derivative of the double-well potential. This is the interface width parameter, a preset hyperparameter; These are the coefficients of the data fitting term, and also the preset hyperparameters; It is the initial phase field distribution; Preset hyperparameters for the time iteration step; It is the phase field value belonging to the k-th class in the nth iteration; It is the phase field value belonging to the k-th class in the (n+1)th iteration; It is the Laplace operator; Iterate 260 steps, output intermediate results every 26 steps, and finally, take the category with the largest phase field value at each grid point as the classification result.

6. The intelligent identification method for iron and steel smelting fume levels based on the Allen-Cahn phase-field energy equation as described in claim 1, characterized in that: Step S5 includes: First, save the classification result of the phase field as an image output.jpg, and the gray value of each pixel point in the image output.jpg represents the category to which the position belongs; Then, for each sample in the training set, the corresponding pixel position in the image output.jpg is found according to the feature vector (E1,E2), and the gray value at that position is read; the initial class label of the sample is obtained, and gray values ​​of 29, 76 and 150 are used to represent class 1, class 2 and class 3 respectively; the gray value is obtained according to the initial class label, and the absolute value of the difference between the classification result at that position and the gray value corresponding to the initial class label of the sample is calculated. If the absolute value of the difference does not exceed 6, the sample is retained; otherwise, it is discarded, and the filtered samples are obtained.

7. The intelligent identification method for iron and steel smelting fume levels based on the Allen-Cahn phase-field energy equation as described in claim 6, characterized in that: Step S6 includes: S61: Design the model structure; A multilayer perceptron is used as the classification network. The input is a two-dimensional energy feature vector (E1, E2), and the output is three smoke levels. The network consists of two hidden layers and one output layer, and the activation function is Tanh. S62: Training the model; The samples after phase field filtering are used as the training set, the input features are the Allen-Cahn dual-scale energies (E1, E2) corresponding to the image; the labels are the categories obtained by iterative filtering. The model is trained and saved as a recognition model after training for subsequent real-time inference. S63: Real-time reasoning and classification; First, an industrial camera captures real-time images of smoke from the furnace opening or chimney area as input. The input images are then preprocessed and the Region of Interest (ROI) is extracted. Next, the eigenvectors (E1, E2) are calculated using the Allen-Cahn energy function; The feature vector is then input into the recognition model, which infers the smoke level classification result directly. Save and categorize the images based on the classification results.

8. The intelligent identification method for iron and steel smelting fume levels based on the Allen-Cahn phase-field energy equation as described in claim 5, characterized in that: Nx=Ny=100; Determine the initial phase field distribution When r = 0.3h, h is the grid step size.

Images