A ladle liquid level online reconstruction method, system and medium

Abstract

This invention discloses an online reconstruction method, system, and medium for ladle liquid level, belonging to the field of steel smelting process detection technology. The method includes: acquiring an infrared thermal image of the ladle's outer wall and a corresponding visible light image; inputting the infrared thermal image and visible light image into a semantic segmentation model to obtain a probability map of the liquid level region; introducing geometric topological constraints during model training, optimizing the model through a hybrid loss function weighted by three loss terms: overlap, easy / difficult sample balance, and edge geometric constraints. The edge geometric constraint loss term applies continuity constraints to the boundary by predicting the gradient difference between the predicted and actual liquid level edges; determining the edge based on the probability map and mapping it to a physical coordinate system fitting baseline; and obtaining the steel liquid level height based on its relative position to the ladle's reference zero point. This invention embeds geometric topological constraints into model training, reconstructing a smooth liquid level contour in low-contrast thermal images, improving detection robustness and accuracy under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of detection technology in the iron and steel smelting process, and in particular to a method, system and medium for online reconstruction of ladle liquid level. Background Technology

[0002] Online reconstruction of ladle level is a critical detection step in the continuous casting process of steel. Its real-time performance and accuracy directly affect the smooth progress of continuous casting, the improvement of steel yield, and the safety of production operations. Due to the high temperature of molten steel in the ladle (above 1600℃) and the harsh environmental factors such as high-temperature radiation, dust, and steam at the work site, developing a non-contact level detection solution that can adapt to these conditions has become a core challenge in this technical field.

[0003] Infrared thermography utilizes the characteristic temperature field distribution formed on the outer wall of a ladle due to the significant temperature difference between molten steel and the upper cavity. By identifying significant inflection points in the temperature distribution on the outer wall, the liquid level position can be determined, providing a non-contact technical approach for ladle level detection. However, in the actual working conditions of high-temperature radiation fields in metallurgy, the non-uniform distribution of the oxide layer on the outer wall of the ladle causes the surface emissivity to exhibit spatiotemporal unsteady characteristics, making the radiation thermometry equation prone to ill-conditioning. Furthermore, existing general-purpose image segmentation algorithms suffer from sub-pixel jitter in edge localization when processing such low-contrast thermal images, making it difficult to meet the actual requirements of accuracy and robustness for liquid level measurement in industrial settings. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system and medium for online reconstruction of ladle liquid level, which embeds geometric topological constraints into model training to reconstruct a smooth liquid level profile in low-contrast thermal images, thereby improving the robustness and accuracy of detection under complex working conditions.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a method for online reconstruction of ladle liquid level, comprising the following steps: Infrared thermal images, visible light images, and corresponding real liquid level annotations of the outer wall of the ladle, collected synchronously from different angles, were used as the training set. A semantic segmentation model is constructed, which is used to perform pixel-level liquid level region segmentation on the input infrared thermal image and visible light image, and output a predicted liquid level probability map. Using the training set, the semantic segmentation model to be trained is iteratively trained based on the hybrid loss function to obtain the trained semantic segmentation model; The hybrid loss function includes a first loss term constructed based on the predicted liquid level probability map and the real liquid level annotation map, a second loss term constructed based on the predicted value of each pixel in the predicted liquid level probability map and the true value of the corresponding pixel in the real liquid level annotation map, and a third loss term constructed based on the difference between the predicted edge gradient map extracted from the predicted liquid level probability map and the real edge gradient map extracted from the real liquid level annotation map. The system acquires real-time infrared thermal image and visible light image of the steel ladle to be tested, inputs them into the trained semantic segmentation model to obtain a predicted liquid level probability map; determines the liquid level edge based on the predicted liquid level probability map, and maps the liquid level edge to the physical coordinate system to calculate the height of the molten steel surface.

[0006] In some optional embodiments, the hybrid loss function is expressed as: ; In the formula, This is the first loss term used to measure the overlap between the predicted liquid level region and the actual liquid level region. This is the second loss term used to adjust the loss contribution of easy and difficult samples. This is the third loss item, which is obtained through... Sobel The operator determines the gradient difference between the predicted edge gradient map and the actual edge gradient map to apply a first-order continuity constraint to the liquid level boundary.

[0007] In some optional embodiments, the semantic segmentation model is an encoder-decoder architecture, specifically including: The encoder is constructed from a ResNet-34 network with the fully connected layers removed, and the decoder is symmetrically configured with the encoder and includes an index-preserving upsampling layer and a skip connection layer. The encoder is used to extract features from the input infrared thermal image and visible light image step by step to obtain a multi-scale feature map. The decoder is used to upsample the obtained multi-scale feature maps step by step to restore the resolution, and during the upsampling process, it performs skip stitching and fusion of the feature maps of the corresponding levels in the downsampling stage and the feature maps of the corresponding levels in the upsampling stage to generate a probability map.

[0008] In some optional embodiments, the semantic segmentation model further includes a cross-modal feature fusion module, specifically including: The cross-modal feature fusion module is located between the encoder and the decoder; The cross-modal feature fusion module is used to dynamically fuse the thermal radiation information represented by the infrared feature map and the texture information represented by the visible light feature map based on the correlation between the infrared feature map and the visible light feature map extracted by the encoder from the infrared thermal image and the visible light image, respectively, to generate a fused feature map, and output the fused feature map to the decoder to replace the infrared feature map and the visible light feature map for subsequent processing.

[0009] In some optional embodiments, before acquiring an infrared thermal image of the outer wall of the ladle, the method further includes temperature calibration of the infrared thermal imager used to acquire the infrared thermal image, as well as environmental interference removal and emissivity correction of the infrared thermal image.

[0010] In some optional embodiments, the specific steps for determining the liquid level edge based on the predicted liquid level probability map and mapping the liquid level edge to the physical coordinate system to calculate the steel liquid level height are as follows: Threshold segmentation and morphological processing are performed on the probability map of the liquid level region to obtain a binary segmentation map; The contour features of the liquid level region are extracted from the binary segmentation map using an edge detection operator to determine the liquid level edge. Based on the pre-calibrated camera intrinsic and extrinsic parameters, the determined liquid level edge is mapped from the image pixel coordinate system to the ladle physical coordinate system; A horizontal straight line is fitted to the mapped set of liquid surface edge points. The fitted horizontal straight line is used as the liquid surface baseline. The height of the steel liquid surface is calculated based on the relative position of the liquid surface baseline and the reference zero point.

[0011] In some optional embodiments, the thickness of the refining slag is also obtained through the edge of the liquid surface, and the specific steps are as follows: Determine the temperature distribution curve in the vertical direction of the liquid level area based on the infrared thermal image; Identify the temperature inflection point in the temperature distribution curve formed by the difference between the temperature of the molten steel and the temperature of the overlying refining slag; The thickness of the refining slag is obtained based on the relative position of the temperature inflection point and the liquid surface baseline.

[0012] In some optional embodiments, the actual liquid level map is obtained by: The system acquires infrared thermal images of the outer wall of the ladle and corresponding visible light images from different angles under different working conditions. It performs pixel-level semantic annotation on the liquid level area in the infrared thermal images or visible light images to distinguish the molten steel area, the inner wall of the ladle and background interference, and generates a real liquid level annotation map to mark whether each pixel in the image belongs to the liquid level area.

[0013] Embodiments of the present invention also provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described ladle level online reconstruction method.

[0014] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described online reconstruction method for ladle liquid level.

[0015] The online reconstruction method for ladle liquid level provided by this invention has at least the following beneficial effects: The online reconstruction method for ladle liquid level provided by this invention combines infrared thermal images with visible light images, and uses a semantic segmentation model with a geometric topological constraint training mechanism to obtain a probability map of the liquid level region. Then, it fits the liquid surface baseline in the physical coordinate system of the ladle to obtain the height of the liquid level. This enables the entire detection process to stably reconstruct the liquid level profile that conforms to the expected physical geometry under low-contrast thermal imaging conditions, thereby improving the robustness and accuracy of liquid level detection in complex industrial environments. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0017] Figure 1 This is a flowchart of an online reconstruction method for ladle liquid level according to an embodiment of the present invention; Figure 2 This is a system architecture diagram of an online reconfiguration system for ladle liquid level based on multimodal vision, provided according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the multi-band interference separation and temperature field reconstruction process for steel ladle infrared thermometry according to an embodiment of the present invention; Figure 4 This is an infrared-visible multimodal semantic segmentation model architecture diagram for online reconstruction of ladle liquid level according to an embodiment of the present invention; Figure 5 This is a flowchart of the entire process of an online reconstructing system for ladle liquid level and slag layer thickness according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0019] One embodiment of the present invention relates to a method for online reconstruction of ladle liquid level. The implementation details of the online reconstruction method for ladle liquid level in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0020] The specific process of the online reconstructing method for ladle liquid level in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 101: Obtain the sample infrared thermal image, sample visible light image and corresponding real liquid level annotation map of the outer wall of the ladle from different angles simultaneously, as the training set; High-definition infrared thermal imagers and high-definition visible light cameras are deployed on both sides of the ladle to simultaneously acquire infrared video streams of the ladle's outer wall from different angles, while also acquiring visible light images as auxiliary references. The system's built-in temperature calibration module employs an infrared temperature measurement algorithm with parameter correction. A temperature response model is established through least-squares fitting to dynamically compensate for nonlinear errors caused by changes in the focal plane temperature, ensuring the accuracy of the temperature data.

[0021] As a preferred method, infrared thermal imaging video stream image data is obtained through experimental means, specifically including the following steps: S101.1 Ensure that the ladle body and its auxiliary hardware are in stable working condition, and check whether the resolution of the mid-wave infrared thermal imager (wavelength range 3-5μm) is 640×512 pixels, whether the thermal sensitivity meets NETD≤40mK, whether it reaches the IP65 protection level, and whether it is equipped with a forced air-cooled protective cover and an automatic blowing dustproof window, and whether the resolution of the high-definition visible light camera is not less than 1080p; S101.2. Two mid-wave infrared thermal imagers are installed in a distributed manner, located approximately 3 meters away on both sides of the steel ladle, avoiding areas directly exposed to high-temperature radiation, and ensuring 360° coverage of the outer wall of the steel ladle without blind spots; the imaging parameters of the thermal imagers are set, and the thermal imager program settings are synchronized to the main controller and then transmitted to the host computer. S101.3. The high-definition visible light camera is deployed synchronously next to each infrared thermal imager to acquire visible light images of the outer wall of the steel ladle and to align them with the infrared images in time and space. S101.4 The ladle lining is heated to the target temperature to prevent damage to the lining, and molten steel is filled in it. The temperature field of the outer wall of the ladle is imaged in real time by the thermal imager. S101.5. Temperature calibration is performed using a temperature measurement model based on Planck's radiation law, and the temperature-output response curve is fitted using the least squares method to correct the temperature of the focal plane. S101.6. For strongly scattering media, construct a multi-band spectral interference separation matrix to eliminate environmental interference on infrared radiation and construct a cross-modal physical constraint model. S101.7. An attention gating mechanism is adopted to calculate the correlation weight between the infrared feature map and the visible light feature map, dynamically fuse the feature responses of the two heterogeneous physical fields, and generate a fused feature field that combines temperature information and texture details; by linking the package number of the visible light image recognition with the process database, the historical temperature curve of the ladle is automatically retrieved for cross-validation.

[0022] The architecture diagram of the online reconfiguration system for ladle liquid level based on multimodal vision is as follows: Figure 2 As shown in the figure, by deploying an infrared thermal imager (1) and a visible light camera (2) on both sides of the ladle, multimodal image data of the ladle (3) are collected simultaneously. The data is transmitted to the industrial control computer (4) via the switch (7) for intelligent segmentation and liquid level calculation. The calculation results are output to the monitoring terminal (5) simultaneously. At the same time, the system links the gas control cabinet (6) to adjust the working environment, thereby realizing the accurate online reconstruction of the ladle liquid level and slag layer thickness under high temperature and high dust interference environment.

[0023] The specific implementation of step S101.5 is as follows: Using the temperature measurement model based on Planck's radiation law, the pixel temperature is calculated using the following formula: ; in Radiation intensity, , λ is a constant, and λ is the wavelength.

[0024] When performing temperature correction on the focal plane, the temperature of the thermal imager's focal plane is monitored in real time. A compensation function is established by fitting the temperature-output response curve using the least squares method. ; coefficient a , b , c The calibration data is dynamically updated using blackbody source calibration.

[0025] Analyzing its nonlinear response mechanism, the infrared focal plane detector exhibits inherent nonlinearity in its response: ; in: The thermal radiation flux received by the target surface (which is exponentially related to temperature T). k 1 , k 2 This refers to the detector gain coefficient; V offset This refers to the dark current voltage.

[0026] The calibration experiment was designed as follows: the blackbody source was a surface source with an emissivity of not less than 0.95 and a temperature control accuracy within ±0.3℃; the temperature point distribution adopted Chebyshev nodes to reduce fitting oscillations; the ambient temperature step change from -10℃ to +70℃ was simulated in a constant temperature chamber, and 100 sets of data were collected at each temperature point and the average value was taken.

[0027] Full-parameter model of the compensation function including environmental coupling terms: ; The steps for solving the problem using the least squares method are as follows: (1) Construct the observation matrix: ; (2) Regularized least squares solution (to suppress overfitting): ; in The Tikhonov regularization coefficient is used to determine the goodness-of-fit validation criteria, which are as follows: (1) Coefficient of determination (2) Residual standard deviation The specific implementation of step S101.6 is as follows: To eliminate environmental interference with infrared radiation, a four-dimensional spectral observation matrix M was constructed, selecting four characteristic bands with center wavelengths of 3.9 μm (sensitive to water vapor), 4.5 μm (sensitive to CO2), 5.0 μm (peak radiation of molten steel), and 8.0 μm (atmospheric window). A four-channel parallel mid-wave infrared detector was used to simultaneously acquire the radiation intensities I1 to I4 of the four bands. The multi-band data acquisition strategy is shown in Table 1 below.

[0028] Table 1 Multi-band data acquisition strategy Mathematical description of the physical model of environmental disturbance: ; in: Atmospheric transmittance, The radiation transmission path length (measured by laser rangefinder 8). Dust emissivity.

[0029] Establish a linear aliasing system of equations Mx=b, which includes the steel radiation term Rs, the water vapor absorption term Aw, and the dust scattering term Sd; solve the overdetermined system of equations by singular value decomposition to reconstruct the undisturbed pure steel radiation field Rs.

[0030] The core algorithms for multi-band filtering include: (1) Construct an interference separation matrix and express the four-band radiation intensity as a linear combination of three terms: steel molten radiation, water vapor absorption and CO2 absorption; (2) An adaptive filtering model is constructed based on the minimum mean square error criterion. The optimal weight matrix WW is solved to balance the signal-to-noise ratio of each band signal. A regularization term is introduced to constrain the norm of the weight matrix to ensure the stability of the solution. (3) Target radiation reconstruction, through the formula To obtain the radiation intensity of pure molten steel.

[0031] Based on the texture gradient features of visible light images, the oxide layer thickness of each pixel on the outer wall of the ladle is calculated using a pre-defined lookup table mapping oxide layer thickness to grayscale. This thickness is then used as a priori parameter input into the infrared radiative transfer equation to dynamically retrieve the surface emissivity correction coefficient at the corresponding location. ϵ cor This is to compensate for the emissivity drift error caused by oxide layer aging and uneven thickness.

[0032] The dynamic environmental adaptation mechanism is as follows: When water vapor concentration fluctuates, the ratio of radiation intensity between Band1 and Band2 is monitored in real time, and the correlation coefficient in the interference separation matrix is ​​updated; when dust accumulates, the sudden increase in radiation intensity of Band4 is monitored, and the regularization intensity of the weight matrix is ​​increased; when the ladle moves, the dynamic environmental adaptation mechanism is adjusted based on feedback from the laser rangefinder. L The atmospheric transmittance compensation term is dynamically calculated.

[0033] The process of multi-band interference separation and temperature field reconstruction for steel ladle infrared thermography is as follows: Figure 3 As shown in the figure, the original infrared signal of the ladle is first acquired through multi-band infrared data acquisition, and then the interference separation matrix is ​​constructed by band characteristic modeling. The target radiation is reconstructed by combining adaptive weight calculation, and finally high-precision temperature field data is output. This eliminates the interference of aerosols such as water vapor and dust in the high-temperature environment and restores the true temperature distribution of the ladle.

[0034] Step 102: Construct a semantic segmentation model, which is used to perform pixel-level liquid level region segmentation on the input infrared thermal image and visible light image, and output a predicted liquid level probability map. A semantic segmentation model with an encoder-decoder architecture was trained, specifically using the SegNet semantic segmentation network. The encoder was used for feature extraction, and the decoder was used for upsampling and accurate pixel-level classification. During training, the model learned to segment specific regions representing the molten steel surface from an infrared thermal image that incorporated brightness information from a visible light image.

[0035] The encoder is designed based on a Modified ResNet-34 network with the fully connected layers removed. To better adapt to the low-frequency characteristics of infrared images, the first-layer convolutional kernel is adjusted to 7×7.

[0036] The encoder contains four progressively downsampled coding levels, E1 to E4: Level E1 outputs a feature map with a resolution of 320×320 and 64 channels, representing edge and texture responses; Level E2 outputs a feature map with a resolution of 160×160 and 128 channels, representing local temperature gradients; Level E3 outputs a feature map with a resolution of 80×80 and 256 channels, used for liquid surface region candidates; and Level E4 outputs a feature map with a resolution of 40×40 and 512 channels, representing global context features.

[0037] The decoder and encoder are symmetrically configured. The decoder is designed with index-preserving upsampling operations and reuses the encoder's pooling index to reduce chessboard artifacts. Feature maps from the E3 and D3 layers are fused using skip connections to address the issue of lost deep information.

[0038] The specific operation process for decoding the path is as follows: (1) The feature map resolution is restored by 5×5 transposed convolution (stride of 2); (2) Perform feature concatenation between the upsampled feature map and the feature map of the corresponding coding level; (3) The concatenated features are refined by 3×3 convolution, batch normalization and ReLU activation function; (4) The output layer uses 1×1 convolution and Sigmoid activation function to generate a binary classification probability map.

[0039] In the feature fusion stage, a cross-modal attention mechanism module is constructed. This module calculates features from infrared thermal images. F IR Features of visible light images F Vis Relevance weights α The enhancement intensity of visible light texture information on the infrared feature map is dynamically adjusted to generate a fused feature map. The specific weight calculation formula is as follows: ,in and For learnable parameters, This is the Sigmoid activation function.

[0040] The architecture diagram of the infrared-visible multimodal semantic segmentation model for online reconstruction of ladle liquid level is as follows: Figure 4 As shown in the figure, the infrared thermal image with a resolution of 640×640 and the visible light intensity channel are used as dual inputs. Multi-scale features are extracted by the encoder, and then the feature fusion module realizes the complementary fusion of temperature and texture information. Finally, the decoder outputs the pixel-level classification result, which provides a reliable image recognition basis for the accurate segmentation of the ladle liquid level boundary and the subsequent calculation of liquid level and slag layer thickness under high temperature and dust interference environment.

[0041] Step 103: Using the training set, the semantic segmentation model to be trained is iteratively trained based on the hybrid loss function to obtain the trained semantic segmentation model; wherein, the hybrid loss function includes a first loss term constructed based on the predicted liquid level probability map and the real liquid level annotation map, a second loss term constructed based on the predicted value of each pixel in the predicted liquid level probability map and the ground truth value of the corresponding pixel in the real liquid level annotation map, and a third loss term constructed based on the difference between the predicted edge gradient map extracted from the predicted liquid level probability map and the real edge gradient map extracted from the real liquid level annotation map; A large number of infrared thermal images and corresponding visible light images containing different liquid levels and operating conditions were collected and accurately labeled to generate semantic segmentation ground truth maps of liquid level regions. Augmentation operations such as color jittering and adding Gaussian noise were performed on the dataset, and all images were uniformly randomly cropped to enhance data diversity and model robustness. The dataset was then divided into training, validation, and test sets.

[0042] The specific implementation details are as follows: A professional annotation tool is used to generate pixel-level semantic segmentation ground truth maps for the liquid level region in the infrared image. The annotation categories must distinguish between the molten steel region, the inner wall of the ladle, the background, and other interference objects. All images are normalized, mapping pixel values ​​to the [0,1] interval, and uniformly resized to 640×640 pixels to adapt to the network input. The images are randomly divided into training, validation, and test sets in a ratio of 7:1:2. During the division, it is necessary to ensure that each subset covers all working conditions and liquid level ranges. The training and test sets are strictly isolated to avoid data leakage. The validation set is used for hyperparameter tuning, while the test set is only used for final evaluation. Data management employs version control to manage dataset iterations, recording enhancement parameters and partitioning information, and using standard data loading interfaces and mainstream deep learning frameworks.

[0043] Data augmentation strategies are divided into two categories: spatial transformation and radiation perturbation. Spatial transformation includes random rotation of ±15° to simulate ladle tilting, and scaling the image to 0.9 to 1.1 times and cropping ±5% for affine transformation. Radiation perturbation includes adding Gaussian noise with a standard deviation of 0.03, and using random circular masks to simulate steam obstruction.

[0044] The preprocessing method for multimodal input data is as follows: the infrared thermogram is linearly scaled to [0,1] after non-uniformity correction; the brightness channel of the visible light image is normalized to the [-1,1] interval after contrast enhancement.

[0045] This invention proposes a hybrid loss function training strategy that integrates edge geometric topological constraints. Addressing the conflict between the dual optimization objectives of pixel classification and boundary localization in liquid level detection tasks, differential geometric prior knowledge is introduced into a deep learning framework. By constructing an edge-aware loss function, differential geometric constraints are imposed on the liquid level boundary, forcing the solution space to satisfy first-order continuity, thereby reconstructing a physically reasonable liquid level contour even in low-contrast environments. The expression for the hybrid loss function is:

[0046] ; In the formula, This is the first loss term used to measure the overlap between the predicted liquid level region and the actual liquid level region. This is the second loss term used to adjust the loss contribution of easy and difficult samples. This is the third loss item, which is obtained through... Sobel The operator determines the gradient difference between the predicted edge gradient map and the actual edge gradient map to apply a first-order continuity constraint to the liquid level boundary.

[0047] During model training, a hybrid optimization strategy was employed. This strategy combines region overlap loss, easy-sampling balance loss, and edge geometric constraint loss. Specifically, the edge geometric constraint loss guides the model to focus on the geometric continuity of the liquid level boundary by calculating the gradient difference between the predicted edge map and the true edge map.

[0048] The model was trained using the AdamW optimizer for parameter updates, with an initial learning rate set to 3 × 10⁻. 4 A cosine annealing scheduling strategy is adopted; the weight decay coefficient is set to 1×10⁻ 4 To prevent overfitting; the gradient clipping norm is set to 2.0 to avoid gradient explosion.

[0049] The model convergence monitoring employs an early stopping mechanism: training is terminated when the average intersection-union of the validation set improves by less than 0.1% over 10 consecutive training epochs, or when the training loss fluctuation exceeds 15%. Optimal weight selection utilizes an exponential moving average to smooth weight fluctuations.

[0050] In terms of hardware deployment, four NVIDIA A100 80GB GPUs were used for training. Gradient checkpointing technology was employed to reduce memory usage, and a distributed training framework was used to improve training efficiency.

[0051] Step 104: Obtain real-time infrared thermal image and visible light image of the steel ladle to be tested, input them into the trained semantic segmentation model to obtain the predicted liquid level probability map; determine the liquid level edge based on the predicted liquid level probability map, and map the liquid level edge to the physical coordinate system to calculate the height of the molten steel surface.

[0052] The real-time acquired infrared thermal images are input into the trained model, which outputs a binary segmentation map of the liquid level area. By performing edge detection and geometric analysis on the segmented area, and combining the camera calibration parameters, the known inner diameter geometry of the ladle, and the temperature gradient information provided by the infrared temperature measurement data, the real-time steel liquid level height and the thickness of the overlying refining slag are calculated.

[0053] The specific implementation of this stage can be broken down into the following steps: S104.1 Input the infrared thermal image of the fused visible light image brightness channel acquired in real time into the trained semantic segmentation model, classify each pixel, and output a probability map with the same resolution as the input image, where each pixel value represents the probability that it belongs to the liquid level area. S104.2. The Otsu adaptive thresholding algorithm is used to convert the probability map into a binary segmentation map. In this map, white pixels are identified as liquid level areas and black pixels are identified as non-liquid level areas. S104.3 Perform a series of digital morphological operations on the binary segmentation image to eliminate small noise, fill in the internal holes of the region, and enhance the continuity of the liquid level region outline. S104.4 After performing morphological processing on the binary segmentation image, the contour features of the liquid level region are extracted using the edge detection operator; S104.5. Based on pre-calibrated camera intrinsic parameters such as focal length and principal point coordinates, and extrinsic parameters such as spatial position and attitude, establish a mapping transformation model between the image pixel coordinate system and the steel ladle physical coordinate system, and perform geometric correction on the extracted contour. S104.6 In the corrected image, curve fitting is performed on the extracted liquid surface edge to determine the precise position of the liquid surface baseline; S104.7. Combining the known geometric dimensions such as the inner diameter of the ladle and the reference zero point height, the absolute height of the molten steel surface is calculated using a pre-established mapping model between pixels and physical dimensions. S104.8 Analyze the temperature distribution curve in the vertical direction of the liquid level area using synchronously acquired infrared thermography data. Due to the significant temperature difference between the molten steel and the overlying refining slag, the curve will show a clear inflection point. By identifying this temperature inflection point and combining it with the boundary extracted from the liquid surface, the interface between the molten steel and the slag layer can be accurately located. Based on the pixel distance between the interface and the boundary on the liquid surface in the image, the thickness of the refining slag is calculated using the mapping model between pixels and physical dimensions. S104.9 Result Output: The calculated real-time absolute height of molten steel and the thickness of refining slag are output to the upper-level control system or monitoring screen through the industrial interface.

[0054] As a preferred solution, all the aforementioned infrared thermal imagers and visible light cameras achieve millisecond-level synchronous acquisition via GPIO synchronization pulses, ensuring consistent timestamps for multi-angle images. The acquisition frame rate is set according to process requirements: 30fps for infrared video streams and 25fps for visible light. Image data is transmitted in real-time to the edge computing server via a gigabit fiber optic network. The server has a built-in redundant storage array that supports continuous 24-hour caching of raw data such as infrared radiation data, visible light images, and corresponding temperature calibration parameters.

[0055] The complete workflow diagram of the online reconstructing system for ladle liquid level and slag layer thickness is as follows: Figure 5 As shown in the figure, the process starts with hardware deployment and synchronous data acquisition. Data reliability is ensured through temperature calibration and quality control. Model training and verification are completed through data standardization and dataset construction. After the model passes the test, it is used for real-time image prediction. Then, through post-processing, edge extraction and camera calibration, the physical coordinate system of the ladle is established. Finally, the liquid level and slag layer thickness are calculated and the results are output, forming a complete closed loop from data acquisition to result output. This enables online, accurate and reliable detection of ladle liquid level under high temperature and harsh conditions.

[0056] Through the above process, the online ladle level reconstruction method of this embodiment can accurately and stably detect the ladle level and slag layer thickness in harsh industrial environments with high temperature and high dust, significantly improving the detection robustness and accuracy under complex working conditions.

[0057] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0058] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0059] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for on-line reconstruction of a ladle level, characterized by, The method includes: Infrared thermal images, visible light images, and corresponding real liquid level annotations of the outer wall of the ladle, collected synchronously from different angles, were used as the training set. A semantic segmentation model is constructed, which is used to perform pixel-level liquid level region segmentation on the input infrared thermal image and visible light image, and output a predicted liquid level probability map. Using the training set, the semantic segmentation model to be trained is iteratively trained based on the hybrid loss function to obtain the trained semantic segmentation model; The hybrid loss function includes a first loss term constructed based on the predicted liquid level probability map and the real liquid level annotation map, a second loss term constructed based on the predicted value of each pixel in the predicted liquid level probability map and the true value of the corresponding pixel in the real liquid level annotation map, and a third loss term constructed based on the difference between the predicted edge gradient map extracted from the predicted liquid level probability map and the real edge gradient map extracted from the real liquid level annotation map. The system acquires real-time infrared thermal image and visible light image of the steel ladle to be tested, inputs them into the trained semantic segmentation model to obtain a predicted liquid level probability map; determines the liquid level edge based on the predicted liquid level probability map, and maps the liquid level edge to the physical coordinate system to calculate the height of the molten steel surface.

2. The ladle level online reconstruction method according to claim 1, characterized in that, The hybrid loss function is expressed as follows: ； wherein is a first loss term for measuring the overlap of the predicted liquid level region with the true liquid level region, is a second loss term for adjusting the loss contribution of difficult samples, is a third loss term that imposes a first order continuity constraint on the liquid level boundary by Sobel an operator determines the gradient difference of the predicted edge gradient map and the true edge gradient map.

3. The ladle level online reconstruction method according to claim 1, characterized in that, The semantic segmentation model is an encoder-decoder architecture, specifically including: The encoder is constructed from a ResNet-34 network with the fully connected layers removed, and the decoder is symmetrically configured with the encoder and includes an index-preserving upsampling layer and a skip connection layer. The encoder is used to extract features from the input infrared thermal image and visible light image step by step to obtain a multi-scale feature map. The decoder is used to upsample the obtained multi-scale feature maps step by step to restore the resolution, and during the upsampling process, it performs skip stitching and fusion of the feature maps of the corresponding levels in the downsampling stage and the feature maps of the corresponding levels in the upsampling stage to generate a probability map.

4. The ladle level online reconstruction method according to claim 3, characterized in that, The semantic segmentation model also includes a cross-modal feature fusion module, specifically comprising: The cross-modal feature fusion module is located between the encoder and the decoder; The cross-modal feature fusion module is used to dynamically fuse the thermal radiation information represented by the infrared feature map and the texture information represented by the visible light feature map based on the correlation between the infrared feature map and the visible light feature map extracted by the encoder from the infrared thermal image and the visible light image, respectively, to generate a fused feature map, and output the fused feature map to the decoder to replace the infrared feature map and the visible light feature map for subsequent processing.

5. The on-line ladle level reconstruction method of claim 1, wherein, Before acquiring an infrared thermal image of the outer wall of the ladle, the process also includes temperature calibration of the infrared thermal imager used to acquire the image, as well as environmental interference removal and emissivity correction of the image.

6. The ladle level online reconstruction method of claim 1, wherein, The specific steps for determining the liquid level edge based on the predicted liquid level probability map and mapping the liquid level edge to the physical coordinate system to calculate the steel liquid level height are as follows: Threshold segmentation and morphological processing are performed on the probability map of the liquid level region to obtain a binary segmentation map; The contour features of the liquid level region are extracted from the binary segmentation map using an edge detection operator to determine the liquid level edge. Based on the pre-calibrated camera intrinsic and extrinsic parameters, the determined liquid level edge is mapped from the image pixel coordinate system to the ladle physical coordinate system; A horizontal straight line is fitted to the mapped set of liquid surface edge points. The fitted horizontal straight line is used as the liquid surface baseline. The height of the steel liquid surface is calculated based on the relative position of the liquid surface baseline and the reference zero point.

7. The ladle level online reconstruction method of claim 1, wherein, This also includes obtaining the thickness of the refining slag through the edge of the liquid surface, with the specific steps as follows: Determine the temperature distribution curve in the vertical direction of the liquid level area based on the infrared thermal image; Identify the temperature inflection point in the temperature distribution curve formed by the difference between the temperature of the molten steel and the temperature of the overlying refining slag; The thickness of the refining slag is obtained based on the relative position of the temperature inflection point and the liquid surface baseline.

8. The method for online reconstruction of ladle liquid level according to claim 1, characterized in that, The actual liquid level chart was obtained through the following method: The system acquires infrared thermal images of the outer wall of the ladle and corresponding visible light images from different angles under different working conditions. It performs pixel-level semantic annotation on the liquid level area in the infrared thermal images or visible light images to distinguish the molten steel area, the inner wall of the ladle and background interference, and generates a real liquid level annotation map to mark whether each pixel in the image belongs to the liquid level area.

9. A computer system, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the ladle level online reconstruction method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, is capable of performing the online reconfiguration method for ladle liquid level as defined in any one of claims 1 to 8.

Images