A steel dust removal scene smoke concentration intelligent grading method based on contrast learning

Abstract

This invention provides an intelligent classification method for smoke concentration in steel dust removal scenarios based on contrastive learning, comprising the following sub-steps: S1: Training a contrastive learning model; S2: Constructing a standard feature library for smoke concentration; S3: Real-time smoke image feature extraction; S4: Determining the concentration level of a single-frame smoke image; S5: Outputting the final result and connecting it with control. This method combines machine vision and industrial explosion-proof cameras, overcoming the limitations of harsh environments and detection ranges, acquiring and classifying data in real time for a single frame and transmitting the results, eliminating response delays. The model adopts the Swin Transformer-Tiny and SimCLR frameworks to accurately capture smoke features and improve classification accuracy. Relying on unsupervised pre-training to reduce data annotation costs, a four-level refined classification system is constructed, with each level matching a fixed frequency of the fan, supporting gradient frequency conversion control of the dust removal fan.

Description

Technical Field

[0001] This invention relates to the field of metallurgical dust removal engineering technology, and in particular to an intelligent classification method for smoke concentration in steel dust removal scenarios based on comparative learning. Background Technology

[0002] As a pillar industry of the national economy, the steel industry generates a large amount of dust-laden fumes during core smelting processes such as blast furnaces, converters, and refining furnaces. Accurate monitoring and classification of fumes concentration is a key technological prerequisite for achieving frequency conversion control of dust removal fans, improving dust removal efficiency, and reducing industrial energy consumption. Fumes in steel dust removal scenarios are characterized by rapid concentration changes, complex on-site backgrounds (including high-temperature furnaces, various industrial equipment, dust interference, etc.), and harsh lighting conditions (direct strong light, backlighting, furnace radiation interference, etc.). This places extremely high demands on the accuracy, real-time performance, and robustness of fumes concentration classification technology for industrial applications.

[0003] To meet the frequency conversion control requirements of dust collector fans, steel plants urgently need detection technology capable of finely classifying smoke concentration. Currently, the mainstream smoke concentration detection and classification methods in the industry are mainly divided into two categories: traditional sensor detection methods and conventional machine vision analysis methods. Both have insurmountable technical limitations and cannot meet the actual production needs of gradient frequency conversion control for dust collector fans in steel dust removal scenarios. Specific shortcomings are as follows: Traditional sensor detection methods rely on dust concentration sensors, smoke sensors, and other equipment. These methods have problems such as limited detection range, susceptibility to corrosion and interference from high temperature and humidity environments, and large response delays. Furthermore, they can only output a single concentration value and cannot achieve multi-level fine classification, making it difficult to support gradient frequency conversion control of dust removal fans. Conventional machine vision analysis methods often employ manually designed features (such as color, texture, and contour) combined with traditional machine learning (SVM, decision tree) or basic convolutional neural networks (CNN) for density determination, which has the following drawbacks: (1) Insufficient feature extraction capability: Manually designed features cannot adapt to the complex shape and dynamic changes of smoke in the steel smelting scene. Basic CNNs are difficult to capture the fine-grained spatial features of smoke, resulting in low classification accuracy. (2) Poor generalization ability: It does not take into account interference factors such as changes in lighting, furnace body obstruction, and non-dust removal smoke (such as water vapor and steam) in the scene, which may easily lead to misjudgment in actual industrial environment; (3) No self-supervised feature learning ability: It relies on a large amount of labeled smoke concentration image data, but the cost of acquiring labeled data in the steel dust removal scenario is high and the cycle is long, making model training difficult; (4) Coarse grading: mostly binary classification (smoke-containing / smokeless), which cannot meet the fine grading requirements of multi-frequency gradient control of dust removal fans. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent classification method for smoke concentration in steel dust removal scenarios based on comparative learning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent classification of smoke concentration in steel dust removal scenarios based on contrastive learning includes the following steps: S1: Training the contrastive learning model; S11: Prepare the dataset; Beforehand, collect smoke images covering different smoke concentrations, different smelting stages, different lighting conditions, and different background interference. A set proportion of smoke images are extracted from the collected smoke images, and the smoke concentration levels are labeled by the process personnel. The labeled smoke images are used as the validation set, and the remaining collected smoke images are used as the training set. The smoke concentration levels include no smoke, low concentration smoke, medium concentration smoke, and high concentration smoke. S12: Perform data augmentation on the smoke images in the training set and generate augmented views; S121: To enhance the operation, set the corresponding parameter range and execution probability threshold; The enhancement operations include random cropping, random rotation, contrast adjustment, and fine-tuning of smoke blur. Specifically, for random cropping: set the proportion range of the cropping area in the smoke image and the proportion range of the smoke area included in the cropping frame; for random rotation: set the angle range of random rotation; for contrast adjustment: set the contrast adjustment coefficient; for smoke blur fine-tuning: set the Gaussian blur kernel size corresponding to different smoke concentrations. And set a corresponding execution probability threshold for each enhancement operation; S122: Generate an enhanced view for each smoke image; Based on the parameter range and execution probability threshold set in step S121, for each smoke image in the training set, two different enhanced views are generated through two different enhancement operations; each smoke image corresponds to two enhanced views, forming a positive sample pair. S13: Construct and train a contrastive learning model; The contrastive learning model consists of a WinTransformer-Tiny encoder and a SimCLR contrastive learning framework. The lightweight SwinTransformer-Tiny was used as the encoder of the contrastive learning model to extract features from the data-enhanced smoke image. The smoke images in the training set are divided into A batches, each batch containing N smoke images. The N smoke images correspond to N positive sample pairs. The 2N augmented views in the N positive sample pairs are standardized, including size adjustment, pixel adjustment, etc. The standardized 2N augmented views are used as the model input of the contrastive learning model. Here, A and N are positive integers. 2N augmented views are input into the SwinTransformer-Tiny encoder, which outputs 2N high-dimensional feature vectors, denoted as h. The 2N high-dimensional feature vectors h are then input into a two-layer multilayer perceptron (MLP) of the contrastive learning model. The first layer of the MLP performs a non-linear mapping on the high-dimensional feature vectors h using a combination of linear transformation and ReLU activation function. The second layer of the MLP further transforms the high-dimensional feature vectors h after the first layer mapping into fixed-dimensional projections z. Finally, the 2N high-dimensional feature vectors h are mapped one-to-one to the 2N projections z. Let any two projections be ? , The SimCLR contrastive learning framework uses the following cosine similarity calculation formula to calculate the similarity between any two projections z: ; in, For projection , similarity, For projection Length of the mold For projection Length of the mold The adjustable coefficient is set by... Map similarity to a set range (e.g., [-1, 1]); The contrastive loss function is used as the training loss function. The goal is to maximize the similarity of the two projections z corresponding to the same positive sample pair and minimize the similarity of the projections z between different positive sample pairs. All parameters of the SwinTransformer-Tiny encoder and the multilayer perceptron (MLP) are adjusted so that the contrastive learning model can distinguish the features of different smoke concentrations. The test set loss value is obtained through the contrastive loss function. Smoke images from the validation set labeled with smoke concentration levels are preprocessed and then input into the contrastive learning model at fixed training epochs. The contrastive learning model outputs a projection, denoted as... The contrastive learning model calculates each projection within the validation set. The cosine similarity between the pairs is used to obtain the validation set loss value by comparing the loss function. Throughout the process, the changes in the contrastive loss values ​​calculated using the contrastive loss function on the training and validation sets are monitored. If the loss value on the training set continues to decrease but the loss value on the validation set begins to increase, it is determined that the model is overfitting, and training is stopped immediately. If the loss values ​​on both the training and validation sets stabilize for several consecutive rounds without significant decrease, it is determined that the model has converged, and training ends. After training, the Swin Transformer-Tiny encoder part is retained as the contrastive learning feature extraction model; the trained encoder is validated using a validation set. Once the validation is successful, the trained contrastive learning feature extraction model is obtained, and the process proceeds to step S2.

[0006] S2: Construct a standard feature library for smoke concentration; Based on the actual needs of gradient frequency conversion control of steel dust removal fans, a fixed operating frequency of the dust removal fan is set for each smoke concentration level; For example, no smoke corresponds to the standby frequency of the fan, low concentration corresponds to the low frequency of the fan, medium concentration corresponds to the medium frequency of the fan, and high concentration corresponds to the high frequency of the fan. For each smoke concentration level, multiple typical smoke images with different backgrounds and lighting conditions are selected from the collected smoke images as standard images; steel plant process personnel perform cross-calibration to calibrate the smoke concentration level, typical characteristics (such as smoke coverage area, transparency, concentration distribution), and suitable scenarios (such as blast furnace tapping, converter furnace front, strong light backlight conditions, and furnace body shading conditions) of each standard image. The calibrated standard images are preprocessed, including smoke region cropping, size normalization, Gaussian filtering for noise reduction, and pixel value normalization. The preprocessed standard images are then input into a trained contrastive learning feature extraction model, which outputs a high-dimensional feature vector. This high-dimensional feature vector is used as the standard feature vector for the corresponding smoke concentration level of the standard image, including standard feature vectors for no smoke level, low smoke concentration level, medium smoke concentration level, and high smoke concentration level. The standard feature vectors of all smoke concentration levels are compiled to construct a smoke concentration standard feature library. The standard feature vectors are stored in the smoke concentration standard feature library according to the smoke concentration level.

[0007] S3: Real-time smoke image feature extraction; Industrial explosion-proof cameras are deployed in core dust removal areas such as blast furnaces and converters in steel plants to capture real-time images of smoke areas, obtaining real-time smoke image frames covering different smoke concentrations, different smelting stages, different lighting conditions, and different background interference. Perform the same preprocessing as step S2 on each frame of smoke image, namely smoke region cropping, size normalization, Gaussian filtering for noise reduction, and pixel value normalization; to obtain the preprocessed real-time smoke image frame. The preprocessed real-time smoke image frame is input into the trained contrastive learning feature extraction model. The contrastive learning feature extraction model outputs a high-dimensional feature vector, which is used as the real-time smoke feature vector corresponding to the smoke concentration level of this real-time smoke image frame. S4: Determine the concentration level of a single-frame smoke image; The following cosine similarity function is used to calculate the similarity between the real-time smoke feature vector and all standard feature vectors at each smoke concentration level in the standard feature library of smoke concentration: ; Let x be any real-time smoke feature vector and y be any standard feature vector. Let x be the similarity between vectors x and y, ‖x‖ be the magnitude of the projection onto x, and ‖y‖ be the magnitude of the projection onto y. For each frame of real-time smoke image, calculate the average similarity value between its real-time smoke feature vector and all standard feature vectors at each smoke concentration level in the smoke concentration standard feature library; A threshold for determining the smoke-free level is pre-set. If the average similarity value between the real-time smoke feature vector and the standard feature vector for the smoke-free level is greater than the threshold, it is directly determined to be at the smoke-free level. Otherwise, the smoke concentration level with the highest average similarity value is taken as the single-frame classification result of the real-time smoke image, and the single-frame classification result of each real-time smoke image is output in real time.

[0008] Preferably, a multi-frame decision fusion mechanism is set: based on the known dynamic change rate of steel smoke concentration and the control response speed of the dust removal fan, an adjustable sliding time window (e.g., 3-5 seconds) is set; all single-frame classification results within the sliding time window are statistically analyzed, and the number of times each smoke concentration level appears within the sliding time window is recorded; a majority voting mechanism is used to determine the final classification result. If the proportion of the number of times a certain smoke concentration level appears within the sliding time window exceeds a preset proportion, then that concentration level is taken as the final smoke concentration classification result for the current time window.

[0009] S5: Final result output and control integration; The final classification result within the sliding time window is transmitted in real time to the frequency conversion control system of the dust removal fan in the steel plant. The control system adjusts the operating frequency of the fan according to the classification result, realizing the integration of smoke concentration classification and frequency conversion control of the dust removal fan. For example, the fan is in low-frequency standby mode when there is no smoke and the fan is in high-frequency full-load mode when the concentration is high.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention performs real-time acquisition and classification of single-frame smoke images, and transmits the classification results to the control system in real time, eliminating response delay and matching the real-time requirements of industry; it divides the smoke concentration into multiple levels, with each level corresponding to a fixed operating frequency of the fan, achieving fine classification and directly connecting to the control system to support gradient frequency conversion control of dust removal fans; at the same time, it combines machine vision with industrial explosion-proof cameras, getting rid of environmental interference from high temperature and humidity, dust corrosion and detection range limitations, and adapting to smoke detection in all scenarios of steel smelting. This method uses the Swing Transformer-Tiny encoder and the SimCLR contrastive learning framework to build a model. It captures the fine-grained, multi-scale spatial features of smoke with window attention, and contrastive learning allows the model to learn the essential visual features of smoke autonomously, adapting to the complex shape and dynamic changes of steel smoke, and greatly improving the classification accuracy. The dataset covers smoke patterns under different lighting conditions, background interference, and smelting stages, closely matching the actual working conditions. Scene-specific data augmentation strategies (random cropping, rotation, contrast adjustment, and fine-tuning of smoke blur) simulate interference such as lighting changes and furnace occlusion. A multi-frame decision fusion mechanism is introduced to eliminate misjudgments caused by sudden interference in a single frame through majority voting in a sliding time window, thereby improving the model's generalization ability in real industrial environments. Unsupervised pre-training is achieved based on SimCLR contrastive learning. Model training only requires the collection of on-site smoke images, eliminating the need for extensive manual annotation of concentration levels. Only a small validation set is needed to monitor the training effect, significantly reducing the cost and cycle of data annotation in steel dust removal scenarios and reducing the difficulty of model training. A multi-level classification system is constructed, encompassing smoke-free, low-concentration, medium-concentration, and high-concentration dust collectors. The classification granularity can be flexibly adjusted according to the dust collector fan control requirements. Each level directly corresponds to a fixed operating frequency of the fan, replacing the traditional two-classification mode and precisely matching the refined needs of multi-frequency gradient control of dust collector fans. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the steps of an intelligent classification method for smoke concentration in a steel dust removal scenario based on contrastive learning, as described in this invention. Detailed Implementation

[0012] 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.

[0013] like Figure 1 As shown, a method for intelligent classification of smoke concentration in a steel dust removal scenario based on contrastive learning includes the following steps: S1: Training the contrastive learning model; S11: Prepare the dataset; The dataset includes a test set and a validation set; Beforehand, collect smoke images covering different smoke concentrations, different smelting stages, different lighting conditions, and different background interference. A set proportion (5%-10%) of smoke images are extracted from the collected smoke images and the smoke concentration level is labeled by the process personnel. The labeled smoke images are used as the validation set, and the remaining collected smoke images are used as the training set. The smoke concentration levels include no smoke, low smoke concentration, medium smoke concentration, and high smoke concentration. The smelting stage includes preheating, smelting, tapping iron / steel, and furnace shutdown. The lighting conditions include natural light during the day (sunny / cloudy), industrial lighting at night, and high-temperature reflections from equipment; The background interference includes the red heat reflection of the furnace body, equipment obstruction, dust drift, and personnel / vehicle movement; S12: Perform data augmentation on the smoke images in the training set and generate augmented views; S121: To enhance the operation, set the corresponding parameter range and execution probability threshold; The enhancement operations include random cropping, random rotation, contrast adjustment, and fine-tuning of smoke blur. Specifically, random cropping: the proportion of the smoke area in the smoke image is statistically analyzed (usually 40%~80%), the proportion of the cropping area in the smoke image is set (60%~95%), and the cropping box must contain at least the set proportion (50%) of the smoke area to avoid the core features of the smoke being cropped out. Random rotation: Based on the actual rotation angle of the furnace body (0°~90°), set the random rotation angle range (e.g. -45°~+45°) to adapt to the spatial distribution characteristics of smoke in the steel smelting scenario; Contrast Adjustment: Based on actual on-site lighting data, set a contrast adjustment coefficient (e.g., 0.6~1.4; when the contrast adjustment coefficient <1, reduce the contrast to adapt to strong light; when the contrast adjustment coefficient >1, increase the contrast to adapt to weak light) to ensure that the enhanced image still retains the core features of smoke concentration. Smoke blur fine-tuning: The corresponding Gaussian blur kernel size is set according to the smoke concentration level (1×1~2×2 for low concentration, 2×2~3×3 for medium concentration, and 3×3~4×4 for high concentration) to match the visual characteristics of different levels of smoke such as no smoke, low concentration, and high concentration, and avoid feature distortion. Furthermore, set execution probability thresholds for each enhancement operation (e.g., random cropping 85%, random rotation 75%, contrast adjustment 65%, and smoke blur fine-tuning 60%). The following combined constraints are set: (1) For a single smoke image, based on the set execution probability, at least one type of enhancement operation is executed and at most three types of enhancement operations are executed; (2) Smoke blur fine-tuning is only performed on the smoke areas detected in the image; S122: Generate an enhanced view for each smoke image; Based on the parameter range and execution probability threshold set in step S121, for each smoke image in the training set, two different enhanced views are generated through two different enhancement operations; each smoke image corresponds to two enhanced views, forming a positive sample pair. S13: Construct and train a contrastive learning model; The contrastive learning model consists of a WinTransformer-Tiny encoder and a SimCLR contrastive learning framework. The lightweight SwinTransformer-Tiny was used as the encoder of the contrastive learning model to extract features from the data-enhanced smoke image. The smoke images in the training set are divided into A batches, each batch containing N smoke images. The N smoke images correspond to N positive sample pairs. The 2N augmented views in the N positive sample pairs are standardized, including size adjustment, pixel adjustment, etc. The standardized 2N augmented views are used as the model input of the contrastive learning model. Here, A and N are positive integers. Specifically, resizing: resize all images to the fixed size required by Swin Transformer-Tiny; Pixel adjustment: Convert the pixel values ​​of the image from integers of 0-255 to floating-point numbers of 0-1, and normalize them according to the mean / variance of the contrastive learning model before training. Input 2N augmented views into the SwinTransformer-Tiny encoder to output 2N high-dimensional feature vectors, denoted as h; input the 2N high-dimensional feature vectors h into the two-layer multilayer perceptron (MLP) of the contrastive learning model, and map each high-dimensional feature vector h to the corresponding projection z; Specifically, the SwinTransformer-Tiny encoder includes a patch embedding layer, a hierarchical window attention layer, and a global average pooling layer; The patch embedding layer divides each enhanced view into image patches of a fixed two-dimensional size (e.g., 7×7). Each image patch corresponds to a three-dimensional pixel matrix containing RGB pixel channels, rows, and columns (e.g., [3, 7, 7], where 3 corresponds to the three RGB channels, the first 7 corresponds to the row, and the second 7 corresponds to the column). For each three-dimensional pixel matrix, all pixel values ​​are traversed sequentially according to the dimension order, and the three-dimensional pixel matrix (e.g., [3, 7, 7]) is flattened into a one-dimensional sequence (e.g., 3×7×7=147 values). Based on the one-dimensional sequence, an initial one-dimensional vector with matching length (a one-dimensional vector of length 147) is constructed. The hierarchical window attention layer rearranges all initial one-dimensional vectors into multiple two-dimensional windows of the same size, according to the fixed two-dimensional size set by the patch embedding layer. Each window attention module assigns attention weights to each vector based on the vector features within the current window (the contrastive learning model learns to assign high weights to vectors in smoke areas and low weights to vectors in background areas such as the furnace / walls). The vectors within the two-dimensional window are weighted and summed using the assigned attention weights to obtain a new fused vector. The iterative operation of dividing the window, assigning attention weights, and weighting and summing is repeated on the new vector. Finally, each enhanced view outputs a two-dimensional feature map, and each feature point in the two-dimensional feature map corresponds to a fused feature vector. The global average pooling layer performs a global average pooling operation on the two-dimensional feature map output by the hierarchical window attention layer, calculates the average value of all pixel values ​​in each channel of the two-dimensional feature map, and compresses the two-dimensional feature map into a one-dimensional high-dimensional feature vector h; each augmented view outputs a unique high-dimensional feature vector h after being processed by this layer, and 2N augmented views output 2N high-dimensional feature vectors h respectively. The 2N high-dimensional feature vectors h output by the encoder are input into the two-layer multilayer perceptron of the contrastive learning model. The first layer of the MLP uses a combination of linear transformation and ReLU activation function to perform non-linear feature mapping on the high-dimensional feature vectors h. The features after the first layer mapping are further converted into fixed-dimensional projections z. Finally, the 2N high-dimensional feature vectors h are mapped one-to-one to the 2N projections z.

[0014] The SimCLR contrastive learning framework uses the cosine similarity formula to calculate the similarity between any two projections z, denoted as z = z + z + z. , The formula for calculating the cosine similarity is as follows: ; in, For projection Length of the mold For projection Length of the mold The adjustable coefficient is set by... Map similarity to a set range (e.g., [-1, 1]); The contrastive loss function is used as the training loss function. The goal is to maximize the similarity of the two projections z corresponding to the same positive sample pair and minimize the similarity of the projections z between different positive sample pairs. All parameters of the SwinTransformer-Tiny encoder and the multilayer perceptron (MLP) are adjusted so that the contrastive learning model can distinguish the features of different smoke concentrations. The test set loss value is obtained through the contrastive loss function. Smoke images from the validation set labeled with smoke concentration levels are preprocessed and then input into the contrastive learning model at fixed training epochs. The contrastive learning model outputs a projection, denoted as... The contrastive learning model calculates each projection within the validation set. The cosine similarity between the pairs is used to obtain the validation set loss value by comparing the loss function. Throughout the process, the changes in the contrastive loss values ​​calculated using the contrastive loss function on the training and validation sets are monitored. If the loss value on the training set continues to decrease but the loss value on the validation set begins to increase, it is determined that the model is overfitting, and training is stopped immediately. If the loss values ​​on both the training and validation sets stabilize for several consecutive rounds without significant decrease, it is determined that the model has converged, and training ends. After training, the Swin Transformer-Tiny encoder part is retained as the contrastive learning feature extraction model; the trained encoder is validated using a validation set. Once the validation is successful, the trained contrastive learning feature extraction model is obtained, and the process proceeds to step S2.

[0015] S2: Construct a standard feature library for smoke concentration; Based on the actual needs of gradient frequency conversion control of steel dust removal fans, smoke concentration levels are set, including no smoke, low concentration smoke, medium concentration smoke, and high concentration smoke. Each smoke concentration level corresponds to a fixed operating frequency of the dust removal fan. For example, no smoke corresponds to the standby frequency of the fan, low concentration corresponds to the low frequency of the fan, medium concentration corresponds to the medium frequency of the fan, and high concentration corresponds to the high frequency of the fan. For each smoke concentration level, multiple typical smoke images with different backgrounds and lighting conditions are selected from the collected smoke images as standard images; steel plant process personnel perform cross-calibration to calibrate the smoke concentration level, typical characteristics (such as smoke coverage area, transparency, concentration distribution), and suitable scenarios (such as blast furnace tapping, converter furnace front, strong light backlight conditions, and furnace body shading conditions) of each standard image. The calibrated standard images are preprocessed, including smoke region cropping, size normalization, Gaussian filtering for noise reduction, and pixel value normalization. The preprocessed standard images are then input into a trained contrastive learning feature extraction model, which outputs a high-dimensional feature vector. This high-dimensional feature vector is used as the standard feature vector for the corresponding smoke concentration level of the standard image, including standard feature vectors for no smoke level, low smoke concentration level, medium smoke concentration level, and high smoke concentration level. The standard feature vectors of all smoke concentration levels are summarized to construct a smoke concentration standard feature library. The standard feature vectors are stored in the smoke concentration standard feature library according to the smoke concentration level. S3: Real-time smoke image feature extraction; Industrial explosion-proof cameras are deployed in core dust removal areas such as blast furnaces and converters in steel plants to capture real-time images of smoke areas, obtaining real-time smoke image frames covering different smoke concentrations, different smelting stages, different lighting conditions, and different background interference. Perform the same preprocessing as step S2 on each frame of smoke image, namely smoke region cropping, size normalization, Gaussian filtering for noise reduction, and pixel value normalization; to obtain the preprocessed real-time smoke image frame. The preprocessed real-time smoke image frame is input into the trained contrastive learning feature extraction model. The contrastive learning feature extraction model outputs a high-dimensional feature vector, which is used as the real-time smoke feature vector corresponding to the smoke concentration level of this real-time smoke image frame. S4: Determine the concentration level of a single-frame smoke image; The following cosine similarity function is used to calculate the similarity between the real-time smoke feature vector and all standard feature vectors at each smoke concentration level in the standard feature library of smoke concentration: ; Let x be any real-time smoke feature vector and y be any standard feature vector. Let x be the similarity between vectors x and y, ‖x‖ be the magnitude of the projection onto x, and ‖y‖ be the magnitude of the projection onto y. For each frame of real-time smoke image, calculate the average similarity value between its real-time smoke feature vector and all standard feature vectors at each smoke concentration level in the smoke concentration standard feature library; A threshold for determining the smoke-free level is pre-set. If the average similarity value between the real-time smoke feature vector and the standard feature vector for the smoke-free level is greater than the threshold, it is directly determined to be at the smoke-free level. Otherwise, the smoke concentration level with the highest average similarity value is taken as the single-frame classification result of the real-time smoke image, and the single-frame classification result of each real-time smoke image is output in real time.

[0016] Preferably, a multi-frame decision fusion mechanism is set: based on the known dynamic change rate of steel smoke concentration and the control response speed of the dust removal fan, an adjustable sliding time window (e.g., 3-5 seconds) is set; all single-frame classification results within the sliding time window are statistically analyzed, and the number of times each smoke concentration level appears within the sliding time window is recorded; a majority voting mechanism is used to determine the final classification result. If the proportion of the number of times a certain smoke concentration level appears within the sliding time window exceeds a preset proportion, then that concentration level is taken as the final smoke concentration classification result for the current time window. Specifically, the majority voting mechanism uses the following formula to determine the final classification result: ; Where M represents the total number of smoke concentration levels; k represents the smoke concentration level, with a value range of [value missing]. [1, M]; argmax represents the number of times the single-frame classification result of the k-th smoke concentration level occurs within the sliding time window, and argmax is the concentration level corresponding to the maximum number of occurrences.

[0017] S5: Final result output and control integration; The final classification result within the sliding time window is transmitted in real time to the frequency conversion control system of the dust removal fan in the steel plant. The control system adjusts the operating frequency of the fan according to the classification result, realizing the integration of smoke concentration classification and frequency conversion control of the dust removal fan. For example, the fan is in low-frequency standby mode when there is no smoke and the fan is in high-frequency full-load mode when the concentration is high.

[0018] The core technical process of this method consists of five key steps: First, through dataset preparation, scene-specific data augmentation, and model architecture construction, unsupervised pre-training of the SimCLR+Swin Transformer contrastive learning model is completed, allowing the model to learn the general robust features of smoke. Second, features of standard smoke images at different concentration levels are extracted to construct a standard feature library as the basis for classification. Subsequently, smoke images acquired in real time by industrial cameras are preprocessed and feature extracted to obtain real-time smoke feature vectors. Through cosine similarity calculation, concentration classification of a single frame smoke image is achieved. Finally, a multi-frame decision fusion mechanism is introduced to make majority voting decisions on the single-frame results within the sliding time window, outputting a stable and reliable final classification result, and connecting to the frequency conversion control system of the dust collector fan to achieve the integration of classification and control.

[0019] This invention effectively overcomes the adaptation deficiencies of existing smoke concentration classification methods in steel dust removal scenarios. Through an innovative combination of comparative learning and the Transformer architecture, it achieves refined, high-precision, and real-time classification of smoke concentration. Model training can be completed without a large amount of labeled data, reducing industrial deployment costs. The classification results can be directly integrated with the frequency conversion control system of dust removal fans, effectively improving dust removal efficiency and reducing industrial energy consumption. It has strong industrial practicality and promotional value, and can be widely applied to various steel dust removal scenarios such as blast furnaces, converters, and refining furnaces.

[0020] 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 classification of smoke concentration in steel dust removal scenarios based on contrastive learning, characterized in that: Includes the following steps: S1: Training the contrastive learning model; S11: Prepare the dataset; The dataset includes a training set and a test set; Smoke images are collected in advance, and a set proportion of smoke images are extracted from the collected smoke images. The smoke concentration levels are then labeled by process personnel. The labeled smoke images are used as the validation set, and the remaining collected smoke images are used as the training set. The smoke concentration levels include no smoke, low concentration smoke, medium concentration smoke, and high concentration smoke. S12: Perform data augmentation on the smoke images in the training set and generate augmented views; S121: To enhance the operation, set the corresponding parameter range and execution probability threshold; S122: Generate an enhanced view for each smoke image; S13: Construct and train a contrastive learning model; A contrastive learning model is constructed, which consists of a WinTransformer-Tiny encoder and a SimCLR contrastive learning framework. The contrastive learning model was trained using the test set and the training set. After training, the SwinTransformer-Tiny encoder part was retained as the contrastive learning feature extraction model. S2: Construct a standard feature library for smoke concentration; For each smoke concentration level, multiple smoke images are selected from the collected smoke images, and the collected smoke images are cross-labeled to obtain a standard image. The standard image is then input into the contrastive learning feature extraction model, and the output high-dimensional feature vector is used as the labeled feature vector and saved in the smoke concentration standard feature library. S3: Real-time smoke image feature extraction; Real-time images of the smoke area are captured to obtain real-time smoke image frames; Each smoke image frame is preprocessed to obtain a preprocessed real-time smoke image frame; The preprocessed real-time smoke image frame is input into the trained contrastive learning feature extraction model. The contrastive learning feature extraction model outputs a high-dimensional feature vector, which is used as the real-time smoke feature vector corresponding to the smoke concentration level of this real-time smoke image frame. S4: Determine the concentration level of a single-frame smoke image; S5: Final result output and control integration; The final classification results within the sliding time window are transmitted in real time to the frequency conversion control system of the dust removal fan in the steel plant. The control system adjusts the operating frequency of the fan according to the classification results.

2. The intelligent classification method for smoke concentration in steel dust removal scenarios based on contrastive learning as described in claim 1, characterized in that: The specific details of step S12 are as follows: S121: To enhance the operation, set the corresponding parameter range and execution probability threshold; The enhancement operations include random cropping, random rotation, contrast adjustment, and fine-tuning of smoke blur. Specifically, for random cropping: set the proportion range of the cropping area in the smoke image and the proportion range of the smoke area included in the cropping frame; for random rotation: set the angle range of random rotation; for contrast adjustment: set the contrast adjustment coefficient; for smoke blur fine-tuning: set the Gaussian blur kernel size corresponding to different smoke concentrations. And set a corresponding execution probability threshold for each enhancement operation; S122: Generate an enhanced view for each smoke image; Based on the parameter range and execution probability threshold set in step S121, for each smoke image in the training set, two different enhanced views are generated through two different enhancement operations; each smoke image corresponds to two enhanced views, forming a positive sample pair.

3. The intelligent classification method for smoke concentration in steel dust removal scenarios based on contrastive learning as described in claim 1, characterized in that: In step S13, the specific training method is as follows: The smoke images in the training set are divided into A batches, each batch containing N smoke images. The N smoke images correspond to N positive sample pairs. The 2N augmented views in the N positive sample pairs are standardized, including size adjustment, pixel adjustment, etc. The standardized 2N augmented views are used as the model input of the contrastive learning model. Here, A and N are positive integers. 2N augmented views are input into the SwinTransformer-Tiny encoder, which outputs 2N high-dimensional feature vectors, denoted as h. The 2N high-dimensional feature vectors h are input into the two-layer multilayer perceptron (MLP) of the contrastive learning model, and mapped one-to-one to 2N projections z. The SimCLR contrastive learning framework uses the following cosine similarity calculation formula to calculate the similarity between any two projections z: It employs a contrastive loss function as the training loss function, aiming to maximize the similarity between the two projections z corresponding to the same positive sample pair and minimize the similarity between projections z between different positive sample pairs. All parameters of the SwinTransformer-Tiny encoder and the multilayer perceptron (MLP) are adjusted to enable the contrastive learning model to distinguish features of different smoke concentrations; the test set loss value is obtained through the contrastive loss function. Smoke images from the validation set labeled with smoke concentration levels are preprocessed and then input into the contrastive learning model at fixed training epochs. The contrastive learning model outputs a projection, denoted as... The contrastive learning model calculates each projection within the validation set. The cosine similarity between the pairs is used to obtain the validation set loss value by comparing the loss function. Throughout the process, the changes in the contrastive loss values ​​calculated using the contrastive loss function on the training and validation sets are monitored. If the loss value on the training set continues to decrease but the loss value on the validation set begins to increase, it is determined that the model is overfitting, and training is stopped immediately. If the loss values ​​on both the training and validation sets stabilize for several consecutive rounds without significant decrease, it is determined that the model has converged, and training ends. After training, the Swin Transformer-Tiny encoder part is retained as the contrastive learning feature extraction model; the trained encoder is validated using a validation set. Once the validation is successful, the trained contrastive learning feature extraction model is obtained, and the process proceeds to step S2.

4. The intelligent classification method for smoke concentration in steel dust removal scenarios based on contrastive learning as described in claim 1, characterized in that: The specific details of step S2 are as follows: Based on the actual needs of gradient frequency conversion control of steel dust removal fans, a fixed operating frequency of the dust removal fan is set for each smoke concentration level; For each smoke concentration level, multiple smoke images with different backgrounds and lighting are selected from the collected smoke images as standard images; the steel plant process personnel perform cross-calibration to calibrate the smoke concentration level, characteristics, and applicable scenarios of each standard image. The calibrated standard image is preprocessed, including smoke region cropping, size normalization, Gaussian filtering for noise reduction, and pixel value normalization. The preprocessed standard image is input into the trained contrastive learning feature extraction model. The contrastive learning feature extraction model outputs a high-dimensional feature vector, which is used as the standard feature vector for the smoke concentration level corresponding to the standard image. This includes standard feature vectors for no smoke level, low smoke level, medium smoke level, and high smoke level. The standard feature vectors of all smoke concentration levels are compiled to construct a smoke concentration standard feature library. The standard feature vectors are stored in the smoke concentration standard feature library according to the smoke concentration level.

5. The intelligent classification method for smoke concentration in steel dust removal scenarios based on contrastive learning as described in claim 1, characterized in that: In step S4, the similarity between the real-time smoke feature vector and all standard feature vectors at each smoke concentration level in the smoke concentration standard feature library is calculated using the following cosine similarity function: ； Let x be any real-time smoke feature vector and y be any standard feature vector. Let x be the similarity between vectors x and y, ‖x‖ be the magnitude of the projection onto x, and ‖y‖ be the magnitude of the projection onto y. For each frame of real-time smoke image, calculate the average similarity value between its real-time smoke feature vector and all standard feature vectors at each smoke concentration level in the smoke concentration standard feature library; A threshold for determining the smoke-free level is pre-set. If the average similarity value between the real-time smoke feature vector and the standard feature vector for the smoke-free level is greater than the threshold, it is directly determined to be at the smoke-free level. Otherwise, the smoke concentration level with the highest average similarity value is taken as the single-frame classification result of the real-time smoke image, and the single-frame classification result of each real-time smoke image is output in real time.

6. The intelligent classification method for smoke concentration in steel dust removal scenarios based on contrastive learning as described in claim 5, characterized in that: A multi-frame decision fusion mechanism is set up: Based on the known dynamic change rate of steel smoke concentration and the control response speed of dust removal fan, an adjustable sliding time window is set; all single-frame classification results within the sliding time window are statistically analyzed, and the number of times each smoke concentration level appears within the sliding time window is recorded. The final classification result is determined by a majority voting mechanism. If the proportion of occurrences of a certain smoke concentration level within the sliding time window exceeds a preset proportion, then that concentration level is taken as the final smoke concentration classification result for the current time window.

Images