High-pressure nozzle atomization control method and system for continuous casting secondary cooling section

Abstract

The invention discloses a high-pressure nozzle atomization control method and system for a continuous casting secondary cooling section, and relates to the technical field of metallurgical production. Monitoring images of visible light and near-infrared channels are obtained, a defogging model constructed based on a radiation attenuation physical model and deep learning is input, a fogless image is obtained, real-time temperature distribution of a casting blank is mapped based on the multispectral radiation temperature measurement principle, the real-time temperature distribution is compared with a preset target cooling curve, and a control instruction is generated. And atomization parameters of the high-pressure nozzle are dynamically adjusted. According to the method, the influence of water mist and steam can be effectively inhibited under the working conditions of strong atomization and high interference, the temperature measurement precision and the anti-interference capability are improved, the overall and continuous temperature distribution of the surface of the casting blank is reliably obtained, the cooling strength of a nozzle is adaptively adjusted according to needs, the cooling uniformity and the control precision of a secondary cooling section are improved, and the production efficiency is improved. And surface and internal defects caused by improper cooling are reduced, and the production stability and the casting blank quality consistency are improved.

Description

Technical Field

[0001] This invention relates to the field of metallurgical production technology, specifically to a high-pressure nozzle atomization control method and system for the secondary cooling section of continuous casting. Background Technology

[0002] In continuous casting, the secondary cooling section sprays cooling water onto the billet surface through high-pressure nozzles to control the growth rate and temperature distribution of the solidified shell. This is one of the key processes affecting the internal and surface quality of the billet. In actual production, the secondary cooling section is usually located in a high-temperature, high-humidity, and high-pressure water spraying environment. Under high pressure, a large amount of cooling water forms water mist, steam, and splashing water films, resulting in an extremely complex optical environment that severely interferes with the accurate perception of the billet surface condition.

[0003] In existing technologies, the surface temperature of the billet in the secondary cooling section is typically obtained using infrared thermometers or visible light image analysis methods. Because water mist exhibits significant scattering and absorption of radiation signals across different wavelengths, single-wavelength temperature measurement methods are easily affected by environmental interference, resulting in large deviations and poor stability in the measurement results, making it difficult to accurately reflect the actual temperature distribution on the billet surface. Furthermore, existing temperature measurement methods often focus on single points or small areas, making it difficult to perform continuous temperature mapping of the billet surface based on monitoring images, and thus failing to obtain real-time temperature distribution information covering different areas of the billet. In the absence of a complete temperature distribution, the control of the high-pressure nozzles in the secondary cooling section usually relies on fixed parameters or adjustments based on a single end temperature, making it difficult to adaptively adjust according to the actual temperature state of different areas on the billet surface, thereby affecting the control accuracy and consistency of the secondary cooling process. Summary of the Invention

[0004] The purpose of this invention is to solve the problems mentioned in the background art, such as the interference of water mist on the monitoring image in the strong atomization environment of the secondary cooling section and the difficulty in accurately obtaining the real-time temperature distribution of the billet surface, which makes it difficult for the nozzle to adaptively adjust. Therefore, this invention proposes a high-pressure nozzle atomization control method and system for the secondary cooling section of continuous casting.

[0005] A first aspect of this invention provides a high-pressure nozzle atomization control method for the secondary cooling section of continuous casting, the method comprising: Acquire monitoring images of the second cooling section of continuous casting, including visible light and near-infrared channels; The monitored image is input into a dehazing model constructed based on a radiation attenuation physical model and deep learning to obtain a haze-free image; Based on the fog-free image, temperature mapping is performed using the principle of multispectral radiation thermometry to obtain the real-time temperature distribution of the continuously cast billet. The real-time temperature distribution is compared with the preset target cooling curve to generate control commands; According to the control command, the atomization parameters of the high-pressure nozzle in the corresponding area are dynamically adjusted, and the atomization parameters include at least water flow rate and atomization pressure.

[0006] Optionally, the defogging model includes: The water mist feature extraction module is used to perform deep feature representation on the monitoring image using a U-shaped network to obtain a water mist feature map F1. The prior information fusion module is used to fuse the water mist feature map F1 and the preset transmission prior tensor M1 using an attention mechanism to obtain the fused feature map F2. The transmittance estimation module is used to map the fused feature map F2 to a composite transmittance estimation map t using a dense network; The reverse calculation module is used to reverse calculate the monitoring image based on the radiation attenuation model and the composite transmittance estimation diagram t to obtain a fog-free image.

[0007] Optionally, the U-shaped network used in the water mist feature extraction module includes an encoder, a bottleneck layer, and a decoder; wherein: The encoder includes four cascaded coding layers, each of which includes two 3×3 convolutional layers and one downsampling layer; The bottleneck layer comprises a 3×3 convolutional layer, a Transformer module, and a 3×3 convolutional layer connected in sequence. The decoder includes four cascaded decoding layers. Each decoding layer includes an upsampling concatenation layer and two 3×3 convolutional layers. The upsampling concatenation layer is used to upsample the features from the decoding layer and concatenate them with the features from the corresponding encoding layer along the channel before inputting them into the subsequent convolutional layers.

[0008] Optionally, the process of obtaining the prior tensor includes: Multiple sets of foggy and fog-free images were randomly selected from the dataset; The transmittance matrix of a single image is calculated by the relationship between the pixel values ​​of the hazy and foggy images. The transmittance matrix of all selected images is averaged to obtain the global average transmittance tensor, which is used as the transmittance prior tensor.

[0009] Optionally, the fusion process of the prior information fusion module includes: The water mist feature map F1 is convolved by three parallel 1×1 convolutional layers to obtain feature maps Y1, Y2, and Y3. The feature maps Y1 and Y2 are multiplied element by element and normalized using the softmax function to obtain the attention map W1. The feature map Y3 is weighted using the attention map W1 to obtain the feature map F3; The transmitted prior a priori tensor M1 is processed using two 3×3 convolutional layers to obtain the feature map M2; The feature map F3 and the feature map M2 are weighted and fused to obtain the feature map F4; The feature map F4 is subjected to max pooling and average pooling with stride of 2. The two pooling results are concatenated and then passed through a 1×1 convolutional layer and a sigmoid function to obtain the weight map W2. Bilinear interpolation is performed on the weight map W2 to obtain the weight map W3. The feature map F4 is then weighted using the weight map W3 to obtain the fused feature map F2.

[0010] Optionally, the transmittance estimation module includes four 3×3 convolutional layers and one 1×1 convolutional layer; wherein: Each convolutional layer is connected to all the convolutional layers before it; that is, the input of each convolutional layer is composed of the outputs of all the convolutional layers before it. The final 1×1 convolutional layer compresses the high-dimensional dense features into a single channel and maps the output value to (0, 1) through the sigmoid function to obtain the composite transmittance estimation map t.

[0011] Optionally, the processing procedure of the reverse-engineering module includes: The monitoring image I is analyzed using the composite transmittance estimation map t. f Perform element-wise division to obtain the dehazed image I. h : ; in, It is the pixel value of the C channel of the haze-free image at position X, where C∈{R, G, B, NIR}; t(X) is the pixel value of the C channel of the monitored image at position X; t(X) is the pixel value of the composite transmittance estimation map at position X; clip is the clipping function, which outputs 255 when the input value is greater than 255.

[0012] Optionally, the loss function of the dehazing model during training includes L1 norm loss and transmittance smoothing loss between the predicted hazy image and the real hazy image.

[0013] A second aspect of this invention provides a high-pressure nozzle atomization control system for the secondary cooling section of continuous casting, the system comprising: The data acquisition module is used to acquire monitoring images of the secondary cooling section of continuous casting, including visible light channels and near-infrared channels; The image dehazing module is used to input the monitoring image into a dehazing model constructed based on a radiation attenuation physical model and deep learning to obtain a haze-free image; The temperature mapping module is used to perform temperature mapping based on the fog-free image and the principle of multispectral radiation thermometry to obtain the real-time temperature distribution of the continuous casting billet. The instruction generation module is used to compare the real-time temperature distribution with the preset target cooling curve and generate control instructions; The instruction execution module is used to dynamically adjust the atomization parameters of the high-pressure nozzle in the corresponding area according to the control instruction. The atomization parameters include at least water flow rate and atomization pressure.

[0014] The beneficial effects of this invention are: 1. This invention introduces a dehazing model based on a combination of a radiation attenuation physical model and deep learning, which is used to specifically restore images in heavily fogged environments. It can effectively suppress the scattering and absorption effects caused by water mist, steam and splashing water film, and provide stable and reliable input data for subsequent temperature inversion.

[0015] 2. Compared with single-band infrared temperature measurement, this invention uses multispectral information for joint temperature measurement, which can effectively reduce measurement deviations caused by factors such as water mist environment and changes in surface emissivity, improve the accuracy and anti-interference ability of temperature measurement results, and enhance the long-term operational stability of the system under complex working conditions.

[0016] 3. By comparing the real-time temperature distribution of the billet with the preset target cooling curve, control commands are generated based on the temperature deviation in different areas, and the atomization parameters of the high-pressure nozzles in the corresponding areas are dynamically adjusted so that the cooling water volume and atomization intensity can be allocated as needed, avoiding the problem of uneven cooling caused by traditional control methods. Attached Figure Description

[0017] Figure 1 A flowchart of a high-pressure nozzle atomization control method for the secondary cooling section of continuous casting provided by an embodiment of the present invention; Figure 2 A network architecture diagram of a defogging model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a water mist feature extraction module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a priori information fusion module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a transmittance estimation module provided in an embodiment of the present invention; Figure 6 This is an architectural diagram of a high-pressure nozzle atomization control system for the secondary cooling section of continuous casting, provided as an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.

[0019] This invention provides a method for controlling the atomization of high-pressure nozzles in the secondary cooling section of continuous casting. See also... Figure 1 , Figure 1 This is a flowchart illustrating a high-pressure nozzle atomization control method for the secondary cooling section of continuous casting, provided as an embodiment of the present invention. The method includes the following steps: S1, acquire monitoring images of the second cooling section of continuous casting.

[0020] S2 inputs the monitoring image into a dehazing model built based on a radiation attenuation physical model and deep learning to obtain a haze-free image.

[0021] S3. Based on the fog-free image, temperature mapping is performed using the principle of multispectral radiation thermometry to obtain the real-time temperature distribution of the continuously cast billet.

[0022] S4 compares the real-time temperature distribution with the preset target cooling curve and generates control commands.

[0023] S5 dynamically adjusts the atomization parameters of the high-pressure nozzles in the corresponding area according to the control command.

[0024] The monitoring images include visible light channels (R, G, B) and near-infrared channels (NIR). Atomization parameters include at least water flow rate and atomization pressure.

[0025] The high-pressure nozzle atomization control method for the secondary cooling section of continuous casting provided by this invention can significantly improve the accuracy and stability of billet surface temperature acquisition under strong atomization and high interference conditions in the secondary cooling section of continuous casting. It can achieve reliable characterization of the overall and continuous temperature distribution of the billet surface, thereby enabling the cooling intensity of the high-pressure nozzle to be adaptively adjusted according to the actual thermal state of different areas. This effectively improves the uniformity and control accuracy of secondary cooling, reduces surface and internal quality defects caused by improper cooling, and enhances the stability of continuous casting production and the consistency of billet quality.

[0026] In one embodiment, see Figure 2 , Figure 2 This is a network architecture diagram of a dehazing model provided in an embodiment of the present invention. The dehazing model includes a water mist feature extraction module, a prior information fusion module, a transmittance estimation module, and a back-calculation module. Wherein: The water mist feature extraction module uses a U-shaped network to perform deep feature representation on the monitoring images, obtaining the water mist feature map F1. For details, see [link to module]. Figure 3 , Figure 3This is a schematic diagram of a water mist feature extraction module provided in an embodiment of the present invention. The U-shaped network used in the water mist feature extraction module includes an encoder, a bottleneck layer, and a decoder. Wherein: The encoder consists of four cascaded coding layers, each of which includes two 3×3 convolutional layers and one downsampling layer (max pooling).

[0027] The bottleneck layer consists of a 3×3 convolutional layer, a Transformer module, and another 3×3 convolutional layer connected in sequence.

[0028] The decoder consists of four cascaded decoding layers. Each decoding layer includes an upsampling concatenation layer and two 3×3 convolutional layers. The upsampling concatenation layer upsamples (deconvolves) the features from the decoding layer and concatenates them along the channel with the features from the corresponding encoding layer before inputting them into the subsequent convolutional layers.

[0029] The prior information fusion module uses an attention mechanism to fuse the water mist feature map F1 with the preset transmission prior tensor M1 to obtain the fused feature map F2. For details, see [link to module]. Figure 4 , Figure 4 This is a schematic diagram of a priori information fusion module provided in an embodiment of the present invention. The fusion process of the priori information fusion module includes: First, three parallel 1×1 convolutional layers are used to convolve the water mist feature map F1 to obtain feature maps Y1, Y2, and Y3. Then, feature maps Y1 and Y2 are multiplied element-wise (Hadamard product) and normalized using the softmax function to obtain attention map W1. Feature map Y3 is then weighted using attention map W1 to obtain feature map F3. Simultaneously, two 3×3 convolutional layers are used to process the transmission prior tensor M1 to obtain feature map M2.

[0030] Then, feature map F3 and feature map M2 are weighted and fused to obtain feature map F4.

[0031] Next, max pooling and average pooling with stride of 2 are performed on feature map F4 respectively. The two pooling results are concatenated and then passed through a 1×1 convolutional layer and a sigmoid function to obtain weight map W2. Bilinear interpolation is performed on weight map W2 to obtain weight map W3. Weight map W3 is used to weight feature map F4 to obtain fused feature map F2.

[0032] The transmittance estimation module uses a dense network to map the fused feature map F2 to a composite transmittance estimation map t. See details... Figure 5 , Figure 5 This is a schematic diagram of a transmittance estimation module provided in an embodiment of the present invention. The transmittance estimation module includes four 3×3 convolutional layers and one 1×1 convolutional layer; wherein: Each convolutional layer is connected to all the convolutional layers before it, meaning that the input of each convolutional layer consists of the outputs of all the convolutional layers before it. The final 1×1 convolutional layer compresses the high-dimensional dense features into a single channel and maps the output value to (0, 1) using the sigmoid function, resulting in the composite transmittance estimation map t.

[0033] The reverse calculation module is used to calculate based on the radiation attenuation model ( The module uses the composite transmittance estimation map t to back-engineer the monitoring image to obtain a fog-free image. Specifically, this back-engineering module uses the composite transmittance estimation map t to back-engineer the monitoring image I. f Perform element-wise division to obtain the dehazed image I. h : ; in, It is the pixel value of the C channel of the haze-free image at position X, where C∈{R, G, B, NIR}; t(X) is the pixel value of the C channel of the monitored image at position X; t(X) is the pixel value of the composite transmittance estimation map at position X; clip is the clipping function, which outputs 255 when the input value is greater than 255.

[0034] Existing natural image dehazing networks only address atmospheric fog and cannot match the composite attenuation characteristics of water mist and steam in the secondary cooling section of continuous casting, easily leading to dehazing deviations. This embodiment employs a Unet+Transformer feature extraction structure (Unet captures local details of water mist, while Transformer enhances global feature association) and incorporates prior physical information. This not only adapts to the uneven distribution of water mist in the secondary cooling section but also guides network learning with statistical priors on water mist transmittance. Even with changes in steel grade and casting speed, it maintains a stable dehazing effect, significantly improving generalization.

[0035] In one implementation, the process of obtaining the prior tensor includes: Step 1: Randomly select multiple sets of foggy and fog-free images from the dataset.

[0036] Step 2: Calculate the transmittance matrix of a single image set by using the pixel value relationship between the fog-free image and the foggy image.

[0037] Step 3: Average the transmittance matrices of all selected images to obtain the global average transmittance tensor, which is used as the transmittance prior tensor.

[0038] This implementation calculates the pixel value relationship between the foggy image and the corresponding fog-free image to obtain the transmittance matrix of each image set, and then takes the average value to form the global average transmittance tensor. This avoids relying on the local features of a single image, effectively reflects the overall attenuation characteristics of water fog on the image, and ensures the physical reliability of prior information.

[0039] In one implementation, a multispectral camera is installed at the end of the secondary cooling section of the billet where multiple nozzles coexist, and the upper and lower adjacent areas of the billet are captured simultaneously at the same time: the upper nozzle opening area generates dense water mist to form a foggy image, while the lower nozzle closing area remains clear to form a fog-free image. A paired dataset is constructed to train the defogging model.

[0040] The loss function used in the training process of this dehazing model includes the L1 norm loss between the predicted hazy image and the actual hazy image. rec and transmittance smoothing loss L smooth .in: L rec Core loss: L1 loss is a common image reconstruction loss, which will not be discussed in detail here.

[0041] L smooth To mitigate losses: . and These represent the gradients of the pixels in the horizontal and vertical directions, respectively. The spatial continuity of t is constrained by the transmittance smoothing loss to avoid image distortion caused by abrupt changes in local transmittance.

[0042] Total loss function L total for: .in, and These are weighting coefficients, which can be optimized by technicians based on experiments.

[0043] In one embodiment, the temperature mapping process includes: First, in the offline stage, radiation data of the billet surface in multi-spectral channels (R, G, B, NIR) at different temperatures are collected through experiments. Combined with blackbody or standard temperature measuring instrument calibration, a physical inversion model (i.e. temperature-radiation relationship model) between multi-channel radiation signal and surface temperature is established.

[0044] Then, in actual online monitoring, the radiation values ​​of each pixel in the defogging multispectral image are input into the pre-built model, which can directly and quickly map out the corresponding temperature values, thereby generating a continuous two-dimensional temperature distribution map of the billet surface.

[0045] This invention provides a high-pressure nozzle atomization control system for the secondary cooling section of continuous casting. See also... Figure 6 , Figure 6 This is an architectural diagram of a high-pressure nozzle atomization control system for the secondary cooling section of continuous casting, provided as an embodiment of the present invention. The system includes: The data acquisition module is used to acquire monitoring images of the secondary cooling section of continuous casting.

[0046] The image dehazing module is used to input monitoring images into a dehazing model built based on a radiation attenuation physical model and deep learning to obtain a haze-free image.

[0047] The temperature mapping module is used to perform temperature mapping based on the fog-free image and the principle of multispectral radiation thermometry to obtain the real-time temperature distribution of the continuously cast billet.

[0048] The instruction generation module is used to compare the real-time temperature distribution with the preset target cooling curve and generate control instructions.

[0049] The instruction execution module is used to dynamically adjust the atomization parameters of the high-pressure nozzles in the corresponding area according to the control instructions.

[0050] The monitoring images include visible light and near-infrared channels. Atomization parameters include at least water flow rate and atomization pressure.

[0051] The high-pressure nozzle atomization control system for the secondary cooling section of continuous casting provided by this invention can significantly improve the accuracy and stability of billet surface temperature acquisition under strong atomization and high interference conditions in the secondary cooling section of continuous casting. It can reliably characterize the overall and continuous temperature distribution of the billet surface, thereby enabling the cooling intensity of the high-pressure nozzle to be adaptively adjusted according to the actual thermal state of different areas. This effectively improves the uniformity and control accuracy of secondary cooling, reduces surface and internal quality defects caused by improper cooling, and enhances the stability of continuous casting production and the consistency of billet quality.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention are within the scope of the claims of the present invention.

Claims

1. A method for controlling the atomization of high-pressure nozzles in the secondary cooling section of continuous casting, characterized in that, The method includes: Acquire monitoring images of the second cooling section of continuous casting, including visible light and near-infrared channels; The monitored image is input into a dehazing model constructed based on a radiation attenuation physical model and deep learning to obtain a haze-free image; Based on the fog-free image, temperature mapping is performed using the principle of multispectral radiation thermometry to obtain the real-time temperature distribution of the continuously cast billet. The real-time temperature distribution is compared with the preset target cooling curve to generate control commands; According to the control command, the atomization parameters of the high-pressure nozzle in the corresponding area are dynamically adjusted, and the atomization parameters include at least water flow rate and atomization pressure.

2. The high-pressure nozzle atomization control method for the secondary cooling section of continuous casting according to claim 1, characterized in that, The defogging model includes: The water mist feature extraction module is used to perform deep feature representation on the monitoring image using a U-shaped network to obtain a water mist feature map F1. The prior information fusion module is used to fuse the water mist feature map F1 and the preset transmission prior tensor M1 using an attention mechanism to obtain the fused feature map F2. The transmittance estimation module is used to map the fused feature map F2 to a composite transmittance estimation map t using a dense network; The reverse calculation module is used to reverse calculate the monitoring image based on the radiation attenuation model and the composite transmittance estimation diagram t to obtain a fog-free image.

3. The high-pressure nozzle atomization control method for the secondary cooling section of continuous casting according to claim 2, characterized in that, The water mist feature extraction module employs a U-shaped network comprising an encoder, a bottleneck layer, and a decoder; wherein: The encoder includes four cascaded coding layers, each of which includes two 3×3 convolutional layers and one downsampling layer; The bottleneck layer comprises a 3×3 convolutional layer, a Transformer module, and a 3×3 convolutional layer connected in sequence. The decoder includes four cascaded decoding layers. Each decoding layer includes an upsampling concatenation layer and two 3×3 convolutional layers. The upsampling concatenation layer is used to upsample the features from the decoding layer and concatenate them with the features from the corresponding encoding layer along the channel before inputting them into the subsequent convolutional layers.

4. The high-pressure nozzle atomization control method for the secondary cooling section of continuous casting according to claim 2, characterized in that, The process of obtaining the priori tensor includes: Multiple sets of foggy and fog-free images were randomly selected from the dataset; The transmittance matrix of a single image is calculated by the relationship between the pixel values ​​of the hazy and foggy images. The transmittance matrix of all selected images is averaged to obtain the global average transmittance tensor, which is used as the transmittance prior tensor.

5. The high-pressure nozzle atomization control method for the secondary cooling section of continuous casting according to claim 2, characterized in that, The fusion process of the prior information fusion module includes: The water mist feature map F1 is convolved by three parallel 1×1 convolutional layers to obtain feature maps Y1, Y2, and Y3. The feature maps Y1 and Y2 are multiplied element by element and normalized using the softmax function to obtain the attention map W1. The feature map Y3 is weighted using the attention map W1 to obtain the feature map F3; The transmitted prior a priori tensor M1 is processed using two 3×3 convolutional layers to obtain the feature map M2; The feature map F3 and the feature map M2 are weighted and fused to obtain the feature map F4; The feature map F4 is subjected to max pooling and average pooling with stride of 2. The two pooling results are concatenated and then passed through a 1×1 convolutional layer and a sigmoid function to obtain the weight map W2. Bilinear interpolation is performed on the weight map W2 to obtain the weight map W3. The feature map F4 is then weighted using the weight map W3 to obtain the fused feature map F2.

6. The high-pressure nozzle atomization control method for the secondary cooling section of continuous casting according to claim 2, characterized in that, The transmittance estimation module comprises four 3×3 convolutional layers and one 1×1 convolutional layer; wherein: Each convolutional layer is connected to all the convolutional layers before it; that is, the input of each convolutional layer is composed of the outputs of all the convolutional layers before it. The final 1×1 convolutional layer compresses the high-dimensional dense features into a single channel and maps the output value to (0, 1) through the sigmoid function to obtain the composite transmittance estimation map t.

7. The high-pressure nozzle atomization control method for the secondary cooling section of continuous casting according to claim 2, characterized in that, The processing procedure of the reverse propagation module includes: The monitoring image I is analyzed using the composite transmittance estimation map t. f Perform element-wise division to obtain the dehazed image I. h : ； in, It is the pixel value of the C channel of the haze-free image at position X, where C∈{R, G, B, NIR}; t(X) is the pixel value of the C channel of the monitored image at position X; t(X) is the pixel value of the composite transmittance estimation map at position X; clip is the clipping function, which outputs 255 when the input value is greater than 255.

8. The high-pressure nozzle atomization control method for the secondary cooling section of continuous casting according to claim 2, characterized in that, The loss function of the dehazing model during training includes L1 norm loss and transmittance smoothing loss between the predicted hazy image and the real hazy image.

9. A high-pressure nozzle atomization control system for the secondary cooling section of continuous casting, characterized in that, The system includes: The data acquisition module is used to acquire monitoring images of the secondary cooling section of continuous casting, including visible light channels and near-infrared channels; The image dehazing module is used to input the monitoring image into a dehazing model constructed based on a radiation attenuation physical model and deep learning to obtain a haze-free image; The temperature mapping module is used to perform temperature mapping based on the fog-free image and the principle of multispectral radiation thermometry to obtain the real-time temperature distribution of the continuous casting billet. The instruction generation module is used to compare the real-time temperature distribution with the preset target cooling curve and generate control instructions; The instruction execution module is used to dynamically adjust the atomization parameters of the high-pressure nozzle in the corresponding area according to the control instruction. The atomization parameters include at least water flow rate and atomization pressure.

Images