Real-time monitoring system and method for process behavior in RH refining furnace

Abstract

The invention provides a real-time monitoring system and method for process behaviors in an RH refining furnace. The system comprises an image acquisition and monitoring unit in the refining furnace based on an RH top lance, a data transmission unit, a data processing unit, a video display unit and a data display unit. Wherein the image acquisition and monitoring unit is used for acquiring real-time state information and image information in the RH refining furnace. And the real-time monitoring information is sent to the video display unit through the data transmission unit, and real-time pictures in the refining furnace are displayed. The data display unit displays the real-time refining process, the real-time refining temperature, the existence of top lance flames, the top lance height, the molten steel liquid level height and other information. According to the method, real-time online identification and prompt in the RH refining process can be effectively realized, and good support is provided for improving refining safety, refining intelligence, equipment integration, reduction of use of disposable temperature measurement equipment, judgment and improvement of equipment faults, refining process optimization and refining efficiency optimization.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of information technology, and particularly relates to a system and method for visualizing and non-contact detecting of process behavior in an RH refining furnace, including online real-time monitoring of a refining process, temperature, molten steel level, top lance flame and top lance height. BACKGROUND

[0002] As an important link in steel production, steelmaking plays an important role in the national economy. RH refining has the characteristics of short processing cycle, good refining effect and strong processing capacity, and is used for decarburization, desulfurization and alloy composition adjustment of molten steel to improve the quality of steel.

[0003] Currently, the measurement of molten steel temperature in the RH refining process relies on manual measurement, and cannot measure the trend of molten steel temperature change, lacks real-time monitoring of the state in the refining furnace, lacks measurement of the molten steel level in the vacuum chamber, lacks visualization judgment of the top lance flame and the top lance height. This has affected the safety, energy consumption and process optimization of RH refining, and has become a problem that needs to be solved urgently.

[0004] In the prior art, a digital monitoring system is used to monitor the refining furnace, an infrared camera is used to obtain the surface temperature of molten steel, and a multispectral camera is used to obtain the temperature of molten steel and the smelting state. Since the above method does not have the function of all-weather real-time identification of the state in the refining furnace, it leads to no application of refining process visualization identification, top lance flame, top lance height and molten steel level measurement.

[0005] Prior art 1: a method for a digital monitoring system, which can only monitor the refining furnace and does not have all the functions mentioned above such as temperature detection, refining process detection and top lance flame detection.

[0006] Prior art 2: a method for an infrared camera to obtain the surface temperature of molten steel, which can only obtain the surface temperature of molten steel, has limited refining process detection effect and poor visualization effect, and does not have other functions except the above two.

[0007] Prior art 3: a method for a multispectral camera, which describes that the data acquisition part is set at a certain height from the liquid surface inside the RH refining furnace, real-time acquisition of the image information of the molten steel surface inside the furnace body, analysis and confirmation by a slag proportion identification model whether the molten steel surface needs to be blown, and at the same time, the image information of the smelting state and the molten steel temperature in the RH refining furnace is collected without stopping the furnace, and after digital analysis and processing, real-time monitoring and early warning are realized.

[0008] However, the prior art 3 has the following problems:

[0009] 1. The collection and pretreatment unit is placed at a certain height in the refining furnace, and a detection gun needs to be separately arranged for the collection and pretreatment unit, which increases the complexity of the overall equipment; a purging device is added, and an external purging facility needs to be separately provided, which increases the complexity of the equipment;

[0010] 2. It does not have the functions of visualizing and identifying the height of the molten steel surface, the height of the top lance, and the flame of the top lance;

[0011] 3. The collection and pretreatment unit needs to be moved to the required height for detection, which increases the overall complexity of the equipment operation. SUMMARY

[0012] The purpose of the present application is achieved by the following technical solutions.

[0013] The present application first provides a real-time monitoring system for process behavior in an RH refining furnace, comprising:

[0014] an image acquisition and monitoring unit based on an RH top lance, a data transmission unit, a data processing unit, a video display unit, and a data display unit; wherein,

[0015] The image acquisition and monitoring unit is arranged at the bottom of the RH top lance and can be lifted with the RH top lance in the refining furnace to collect real-time state information and image information in the RH refining furnace, and send the real-time monitoring information to the video display unit through the data transmission unit to display the real-time picture in the refining furnace; the image information is transmitted to the data processing unit through the data transmission unit, and the real-time refining process, the real-time refining temperature, whether there is a top lance flame, the top lance height, and the molten steel liquid level height information are displayed on the data display unit after being processed by the data processing unit.

[0016] Secondly, a real-time monitoring method for process behavior in an RH refining furnace is also provided, comprising:

[0017] Step 1. Install the image acquisition and monitoring unit at the bottom of the RH top lance, pass the gas cooling medium, and insert it into the RH refining furnace with the RH top lance to collect real-time images and videos in the RH refining furnace;

[0018] Step 2. According to the image information, use a deep learning model to identify the collected real-time images to determine the state in the refining furnace, and use an in-furnace temperature calculation model to calculate the temperature in the refining furnace;

[0019] Step three, if it is non-refining state, using flame identification model to judge whether there is flame, and using the height of the lance calculation model to calculate the height of the lance; if it is refining state, using deep learning model to judge the specific state of refining, and using liquid level height calculation model to calculate the real-time height information of the liquid level of molten steel in the RH refining furnace;

[0020] Step four, the calculated data is transmitted to the data display unit for display, and the video data is transmitted to the video display unit for display.

[0021] The application has the advantages that:

[0022] 1. The image acquisition and monitoring unit is combined with the RH lance, which is inserted into the refining furnace along with the RH lance, and only nitrogen cooling is used, which can not only observe the situation in the refining furnace in real time, but also collect images in the refining furnace to analyze the state in the refining furnace.

[0023] 2. The image processing method is used for the first time to realize the visualization identification of the RH lance flame in the refining furnace, the measurement and calculation of the lance height, and the calculation and measurement of the molten steel liquid level height.

[0024] 3. The integration of the lance flame visualization detection, refining process visualization detection, molten steel liquid level height visualization detection, lance height visualization calculation, and continuous measurement of the molten steel surface temperature is realized for the first time.

[0025] 4. The video processing module that can access the video network to observe the state in the refining furnace in real time and the image data processing module that can access the image data processing module to realize image data processing in the refining furnace are realized for the first time.

[0026] 5. The real-time online identification and prompt of the RH refining process can be effectively realized, which provides good support for improving refining safety, refining intelligence, equipment integration, reducing the use of one-time temperature measurement equipment, improving equipment fault judgment, optimizing refining process, and optimizing refining efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0027] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustration only and are not considered a limitation of the present application. Moreover, like reference numerals are used to designate identical components throughout the specification. In the drawings:

[0028] Figure 1 A schematic diagram of a real-time monitoring system for process behavior in an RH refining furnace according to an embodiment of the application is shown.

[0029] Figure 2A data processing process schematic diagram of an RH refining furnace in-process behavior real-time monitoring system according to an embodiment of the application is shown.

[0030] Figure 3 A working method flow chart of an RH refining furnace in-process behavior real-time monitoring system according to an embodiment of the application is shown.

[0031] Figure 4 A flame identification and process state monitoring flow chart according to an embodiment of the application is shown.

[0032] Figure 5 A top lance height detection flow chart according to an embodiment of the application is shown.

[0033] Figure 6 A molten steel liquid level height detection flow chart according to an embodiment of the application is shown.

[0034] Figure 7 A deep learning training and application flow chart according to an embodiment of the application is shown.

[0035] Figure 8 A flame identification effect schematic diagram according to an embodiment of the application is shown.

[0036] Figure 9 A similar triangle measurement method principle schematic diagram according to an embodiment of the application is shown. DETAILED DESCRIPTION

[0037] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood, and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0038] As Figure 1 shown, the present application provides an RH refining furnace in-process behavior real-time monitoring system, the system comprises an RH top lance-based refining furnace in-image acquisition and monitoring unit, a data transmission unit, a data processing unit, a video display unit, and a data display unit.

[0039] The image acquisition and monitoring unit is arranged at the bottom of the RH top lance and ascends and descends in the refining furnace along with the RH top lance to acquire real-time state information and image information in the RH refining furnace. The real-time monitoring information is sent to the video display unit through the data transmission unit, and the real-time picture in the refining furnace is displayed. The image information is transmitted to the data processing unit through the data transmission unit, and after being processed by the data processing unit, the real-time refining process, real-time refining temperature, whether there is a top lance flame, top lance height and molten steel liquid level height and other information are displayed on the data display unit. As shown in Figure 2 The cooling is performed by nitrogen gas with a pressure of ≥0.6 MPa and a flow rate of ≥15 m 3 / h.

[0040] The data transmission unit is used for transmitting the image and video picture in the RH refining furnace acquired by the image acquisition and monitoring unit to the data processing unit and the video display unit respectively.

[0041] The data processing unit receives the image acquired by the image acquisition and monitoring unit and processes the image data to obtain the refining process, temperature, molten steel liquid level height, top lance flame and top lance height data information in the refining furnace.

[0042] The video display unit receives the video acquired by the image acquisition and monitoring unit and displays the real-time state in the refining furnace online in real time.

[0043] The data display unit displays the data information processed by the data processing unit.

[0044] As shown in Figure 3 Fig. 4 is a working method flowchart of the real-time monitoring system for the process behavior in the RH refining furnace. The method comprises the following steps:

[0045] Step 1: install the image acquisition and monitoring unit at the bottom of the RH top lance, introduce the gas cooling medium, and with the RH top lance into the RH refining furnace, acquire the real-time image and video in the RH refining furnace;

[0046] Step 2: according to the image information, use the deep learning model to identify the acquired real-time image to judge the state in the refining furnace, and at the same time, use the furnace temperature calculation model to calculate the temperature in the refining furnace;

[0047] Step 3: if it is a non-refining state, use the flame identification model and the top lance height calculation model to judge whether there is a flame and calculate the top lance height; if it is a refining state, use the deep learning model to judge the specific state of the refining: molten steel into the furnace, normal refining, oxygen blowing state, adding alloy, etc. At the same time, use the liquid level height calculation model to calculate the real-time height information of the molten steel liquid level in the RH refining furnace;

[0048] Step four: Transmit the data calculated above to the data display unit for display, and transmit the video data to the video display unit for display.

[0049] The methods used in the data processing unit of this application include the following aspects:

[0050] 1. A deep learning-based method for identifying RH top-gun flames and monitoring process status.

[0051] like Figure 4 As shown, this application uses an image acquisition and monitoring unit to acquire images inside the RH refining furnace. After data processing, it identifies the top lance flame and process status. The process status includes the following:

[0052] Molten steel entering the furnace (the process of molten steel entering the RH refining vacuum chamber);

[0053] Adding alloys (the process of adding other solid metals to molten steel during refining);

[0054] Oxygen blowing state (the process of blowing oxygen into molten steel during refining);

[0055] Normal refining (the process of molten steel tumbling in a vacuum chamber during refining).

[0056] Specifically, taking top lance flame recognition as an example, this application collects images of top lance flames in different stages and forms within the refining furnace, including baking, slag melting, melting cold steel, and preheating. 200 images of each flame form are collected and divided into training, testing, and validation sets. The training set accounts for 70%, the testing set for 20%, and the validation set for 10%. A top lance flame recognition model for the RH refining furnace, based on a deep learning model including the YOLOv5 model, is constructed. The trained model is then validated through inference by collecting 2000 images from other datasets not included in the above datasets.

[0057] YOLOv5 is a deep learning algorithm for object detection that can quickly and accurately identify target objects in images or videos. The basic principle of this algorithm is to segment the image into grids of different sizes and predict the target object for each grid. The YOLOv5 algorithm can be divided into three main steps: input processing, feature extraction, and target prediction. First, the input image is resized to the size required by the model and normalized. Then, through a series of convolution and pooling operations, the features in the image are extracted. Finally, by classifying and regressing the feature maps, the category and bounding box of the target object in each grid are predicted.

[0058] In the input processing stage, YOLOv5 divides the image into grids of different sizes, each grid is called an anchor point. Each anchor point is responsible for detecting one target object, and different size anchor points are responsible for detecting different size targets. This multi-scale design enables YOLOv5 to detect target objects of different sizes, thereby improving the accuracy of detection. In the feature extraction stage, YOLOv5 uses the backbone network of CSPDarknet53 to extract features from the image. CSPDarknet53 is a lightweight network structure that uses residual blocks and skip connections to improve the representation ability of features. This network structure can effectively extract semantic information from the image while maintaining low computational complexity. In the target prediction stage, YOLOv5 uses a structure called YOLOv3 head to realize the prediction of the target. In this application, YOLOv3 head is composed of a series of convolutional layers and fully connected layers, which are used for classification and regression of feature maps. Specifically, the classification layer is used to predict the category of the target object, and the regression layer is used to predict the bounding box of the target object. To improve the accuracy of detection, YOLOv5 also introduces an objective function called IoU loss in this application. IoU loss is used to measure the degree of overlap between the predicted box and the true box, thereby optimizing the prediction results of the model. By minimizing the IoU loss, YOLOv5 can make the predicted box more accurately match the true box, improving the accuracy of target detection.

[0059] In this application, YOLOv5 also has some other optimization strategies, such as data augmentation, learning rate decay, and model fusion, etc. Data augmentation increases the generalization ability of the model by rotating, scaling, and translating the training data. The learning rate decay strategy can gradually reduce the learning rate during the training process, thereby improving the convergence speed of the model. The model fusion technology can fuse the prediction results of multiple YOLOv5 models, further improving the accuracy of detection.

[0060] Specifically, the training, inference and post-processing process of the top gun flame identification model of the application includes:

[0061] (1) Data preprocessing

[0062] Data cleaning: remove irrelevant data, incorrectly labeled data, etc.

[0063] Data augmentation: rotate, scale, flip, etc. operations on images to increase data diversity.

[0064] Data labeling: draw bounding boxes for targets on each image and label class information.

[0065] Batch normalization: normalize the input data to meet the requirements of model training.

[0066] (2) Model training

[0067] Define model structure: Use YOLOv5 model structure, including convolutional layers, pooling layers, upsampling layers, etc.

[0068] Define loss function: Used to optimize model parameters and reduce prediction errors.

[0069] Weight initialization: Randomly initialize model weights to start the training process.

[0070] Forward propagation: Input training data into the model and calculate the predicted value.

[0071] Loss calculation: Calculate the loss based on the predicted value and the true value.

[0072] Backpropagation: Adjust the model weights based on the loss value.

[0073] Optimizer update: Copy the weights to the optimizer for use in the next iteration.

[0074] Learning rate adjustment: Dynamically adjust the learning rate based on the number of training rounds or changes in loss.

[0075] (3) Inference stage

[0076] Load trained model weights.

[0077] Preprocess input images to meet model input requirements.

[0078] Input images into the model for forward inference to get the bounding box, class, and other information of each target.

[0079] Post-process the image based on the inference results, such as non-maximum suppression, etc.

[0080] Output the final target detection results.

[0081] (4) Post-processing

[0082] Non-maximum suppression (NMS): Remove bounding boxes with high overlap and keep the best detection results.

[0083] Result visualization: Draw the detection results on the original image for easy analysis of the results.

[0084] Result saving: Save the detection results to a file or database for subsequent processing and analysis.

[0085] (5) Result output

[0086] Output detection results: Output the final target detection results to the console, file, or visualization interface.

[0087] Performance evaluation: evaluate the performance of the model according to the preset indicators, such as accuracy, recall rate, etc.

[0088] Result analysis: analyze the test results, identify the advantages and disadvantages of the model, and provide basis for subsequent optimization.

[0089] 2. RH top lance height measurement method based on pixel width recognition.

[0090] As Figure 5 shown, during the non-steelmaking period, the image of the immersion tube at the bottom of the refining furnace is collected by the image acquisition and monitoring unit. During the non-refining period, the immersion tube at the bottom of the refining furnace is circular in cross-section and has a constant size. When the top lance with the image acquisition and monitoring unit changes position, the number of pixels occupied by the circular cross-section diameter of the immersion tube in the collected image will change. The height data of the top lance at the highest point and the lowest point is known, and the cross-section circular diameter of the immersion tube image can be obtained by calculation. By the above method, 500 images of the bottom of the refining furnace during the non-refining period are collected, and the relationship between the height of the top lance and the cross-section circular diameter of the immersion tube is fitted by the least squares method. The height position of the top lance can be calculated by calculating the number of pixels occupied by the cross-section circular diameter of the collected image.

[0091] 3. Refining furnace molten steel liquid level height measurement method based on monocular image measurement.

[0092] Monocular distance measurement is a commonly used measurement technique that measures the distance between an object and an image acquisition unit through a single image acquisition unit. With the development of modern science and technology, monocular distance measurement is constantly developing and being applied in various fields. Monocular distance measurement mainly relies on the geometric information of objects in images to estimate distance. As Figure 6 shown, for the measurement of the molten steel liquid level height in the refining furnace, during the refining stable period, the image acquisition and monitoring unit is first calibrated by Zhang Zhengyou calibration method, and then the distance from the molten steel liquid surface to the image acquisition and monitoring unit is calculated by similar triangle measurement method. Since the height position of the top lance is fixed during the refining stable period, the distance from the image acquisition and monitoring unit to the bottom of the refining furnace is known, and by calculating the distance from the molten steel liquid surface to the image acquisition and monitoring unit, the height value of the molten steel liquid surface during the refining stable period and its change can be obtained by subtracting the former from the latter.

[0093] The Zhang's calibration method used in the present application is a classical camera calibration method based on a planar template. The core process includes: using a planar checkerboard calibration board with known size, acquiring a plurality of images by multi-angle shooting; automatically extracting image corner points and establishing their correspondence with physical corner points of the calibration board; calculating a homography matrix corresponding to each image to describe the projection mapping of the plane to the image; innovatively using the constraints between the homography matrices and the orthogonality of the rotation matrix, linearly solving the initial estimate of the camera intrinsic parameters (focal length, principal point, axis tilt); further solving the extrinsic parameters (rotation matrix and translation vector) of each image based on the initial value of the intrinsic parameters and the homography matrix; introducing a radial distortion model (which can be extended to tangential distortion), and using a nonlinear optimization algorithm (such as Levenberg-Marquardt) to jointly optimize the intrinsic parameters, extrinsic parameters and distortion coefficients by minimizing the reprojection error of all corner points. This method is known for its simple operation (only a planar target is required), high precision (sub-pixel level), and strong robustness, and has become a widely used standard calibration process in computer vision.

[0094] In the present application, the Zhang's calibration method process is as follows:

[0095] 1. Data acquisition

[0096] Use a planar checkerboard calibration board with known size to acquire images through multi-view (≥3 groups, recommended 10-20 groups), ensuring that the pose of the calibration board covers significant rotation and translation transformation.

[0097] 2. Feature point extraction

[0098] Automatically detect the pixel coordinates of the checkerboard corner points in each image, and establish their accurate correspondence with the two-dimensional plane corner points (Z=0) in the world coordinate system.

[0099] 3. Homography matrix estimation

[0100] Based on the plane projection constraint, calculate a homography matrix for each image to describe the affine mapping relationship between the calibration board plane and the image plane.

[0101] 4. Linear initialization of intrinsic parameters

[0102] Core innovation: use the correlation of multi-view homography matrices, combined with the orthogonality constraint of the rotation matrix, to directly solve the initial value of the camera intrinsic parameters (focal length, principal point coordinates, axis tilt factor) through linear equations.

[0103] 5. Extrinsic parameter analysis

[0104] Based on the initial value of the intrinsic parameters and the homography matrix, decompose to obtain the rotation matrix and translation vector of the calibration board relative to the camera in each image, and correct the rotation matrix through orthogonalization.

[0105] 6. Distortion modeling

[0106] A radial distortion model (extensible to tangential distortion) is introduced to define a non-linear correction relationship of image point coordinates.

[0107] 7. Global non-linear optimization

[0108] The intrinsic, extrinsic and distortion coefficients are jointly optimized to minimize the re-projection error:

[0109] Project all three-dimensional corner points to the image plane using the initial parameters;

[0110] Calculate the positional deviation of the projected points from the actual corner points;

[0111] Use an iterative algorithm (such as Levenberg-Marquardt) to optimize the parameter set until it converges to sub-pixel level accuracy.

[0112] In this application, the distance from the object or target to the camera is calculated using similar triangles. Assuming there is an object or target with a width of w. Then place the target at a distance of d from the image acquisition and monitoring unit. Take a picture of the object with the image acquisition and monitoring unit and measure the pixel width p of the object.

[0113] Specifically, as shown in Figure 9 w represents the actual width of the diameter of the molten steel liquid surface, d represents the distance from the molten steel liquid surface to the image acquisition and monitoring unit, f represents the focal length of the image acquisition and monitoring unit, and p represents the pixel width of the molten steel liquid surface in the image.

[0114] Thus, the formula for the focal length of the camera is obtained:

[0115] f = (p x d) / w

[0116] Further calculation of the distance of the object from the camera:

[0117] d = (w x f) / p

[0118] The method comprises the following steps:

[0119] Step 1: Calibrate the image acquisition and monitoring unit to obtain the intrinsic matrix and extrinsic matrix of the camera;

[0120] Step 2: Take a single frame image containing the target through the calibrated image acquisition and monitoring unit, process the obtained image, extract the target circular region, and calculate the pixel number of the minimum circumscribed circle diameter of the region;

[0121] Step 3: According to the principle of similar triangles, calculate the distance from the image acquisition and monitoring unit to the target.

[0122] Specifically, when measuring the height of molten steel in the RH refining furnace, the width w of the molten steel surface is known during refining. The focal length f of the image acquisition and monitoring unit can be obtained based on the calibration of the internal and external parameter matrices. The pixel width p of the circular cross-section of the molten steel surface can be calculated from the captured image. The distance d can be calculated using the formula d = (w x f) / p, and then the height of the molten steel surface can be further calculated.

[0123] 4. Key processes in the RH refining process, identified and processed using deep learning models, include:

[0124] Molten steel entering the furnace (the process of molten steel entering the RH refining vacuum chamber);

[0125] Adding alloys (the process of adding other solid metals to molten steel during refining);

[0126] Oxygen blowing state (the process of blowing oxygen into molten steel during refining);

[0127] Normal refining (the process of molten steel tumbling in a vacuum chamber during refining).

[0128] Regarding the identification of the above content, such as Figure 7 As shown, it is divided into two parts:

[0129] The first part describes the model training process, and the second part describes the model application process.

[0130] Part 1: Model Training

[0131] First, images of the steel refining process are acquired using an image acquisition and monitoring unit mounted on the top lance. Five hundred images are captured each of the following stages: steel entering the furnace, alloy addition, oxygen blowing, and normal refining. These images are then subjected to data augmentation operations such as rotation, scaling, and flipping to enhance data diversity. Second, for each of the different processes, salient features that clearly distinguish each process from the others are identified and labeled. After labeling, the data undergoes batch standardization, which involves normalizing the labeled data to meet the requirements of model training.

[0132] Next, the labeled images are divided into a training set, a test set, and a validation set, with respective proportions of 75%, 15%, and 10%.

[0133] Finally, the model training process involves defining the YOLOv5 model structure, including convolutional layers, pooling layers, upsampling layers, defining the loss function, and initializing the weights. Then, the training process begins.

[0134] After multiple rounds of training, the P (accuracy) and R (recall) of each process result are ensured to be above 0.91 (the maximum value is 1).

[0135] Then output the weight file of the training optimal structure, use the optimal result to infer 10000 refined images, and count the inference accuracy of each process to ensure that the inference accuracy of all processes is above 95%. If it does not meet the above requirements, retrain and test until it meets the above requirements.

[0136] Actual training results:

[0137] P (0.954), R (1.0)

[0138] P (0.964), R (1.0)

[0139] P (0.91), R (0.921)

[0140] P (0.985), R (0.98)

[0141] Output the weight file of the above training results for on-site application.

[0142] Second part: application of the model

[0143] In the model application stage, the following processes are experienced:

[0144] First, the real-time image in the RH refining furnace is collected by the image acquisition and monitoring unit, and the image is transmitted to the data processing unit through the data transmission unit. The data processing unit calls the deep learning model of YOLOv5 and the optimal training result weight file to identify the image and determine whether it is the above process. The judgment result is transmitted to the data display unit for display. If it is, display the specific refining process. If not, display waiting for detection, continue to read the next image, and repeat the above processing process.

[0145] 5. Finally, the flame detection of the present application is different from the conventional flame image. The training and application process of the flame detection is as follows:

[0146] The training and application process of the deep learning model for identifying the RH refining furnace top gun flame is consistent with the above, but the processing process is different. The processing process is as follows:

[0147] Collect the top gun flame in different states, such as Figure 8As shown, the image acquisition and monitoring unit described in the present application collects flame images of different intensities in the RH refining furnace and flame images during the melting of cold steel to train, and in the training, to prevent overfitting, images without flame and without molten steel are added for training, and the actual results of the training are as follows:

[0148] Top lance flame: P (0.941), R (0.991)

[0149] No molten steel and no flame: P (0.992), R (1.0).

[0150] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A system for real-time monitoring of process behavior in a RH refining furnace, characterized in that, The system comprises: The image acquisition and monitoring unit is arranged at the bottom of the RH top lance and can ascend and descend with the RH top lance in the refining furnace to acquire real-time state information and image information in the RH refining furnace, send the real-time monitoring information to the video display unit through the data transmission unit, display the real-time picture in the refining furnace, transmit the image information to the data processing unit through the data transmission unit, and display the real-time refining process, real-time refining temperature, presence or absence of top lance flame, top lance height and molten steel liquid level height information on the data display unit after processing by the data processing unit.

2. The real-time monitoring system for process behavior in an RH refining furnace according to claim 1, characterized in that The image acquisition and monitoring unit is cooled by nitrogen. The method comprises:

3. A method for real-time monitoring of the process behavior in an RH refining furnace using the system according to claim 1 or 2, characterized in that Step one, installing the image acquisition and monitoring unit at the bottom of the RH top lance, introducing a gas cooling medium, and collecting real-time images and videos in the RH refining furnace when the RH top lance is inserted into the RH refining furnace; Step two, identifying the collected real-time images by using a deep learning model according to the image information to determine the state in the refining furnace, and calculating the temperature in the refining furnace by using a furnace temperature calculation model; Step three, if it is a non-refining state, determining whether there is a flame by using a flame identification model and calculating the top lance height by using a top lance height calculation model; if it is a refining state, determining the specific state of refining by using a deep learning model and calculating the real-time height information of the molten steel liquid level in the RH refining furnace by using a liquid level height calculation model; Step four, transmitting the calculated data to the data display unit for display and transmitting the video data to the video display unit for display.

4. The real-time monitoring method for process behavior in an RH refining furnace according to claim 3, characterized in that The specific state of refining includes one or more of the following: molten steel charging, normal refining, oxygen blowing state and alloy addition.

5. The real-time monitoring method for process behavior in an RH refining furnace according to claim 3, characterized in that The determination of whether there is a flame by using a flame identification model comprises: Collecting the flame of the top lance in the refining furnace in different periods and different forms including baking, slag melting, cold steel melting and preheating, 200 flame images for each form to form a data set, dividing the data set into a training set, a test set and a validation set, constructing a RH refining furnace top lance flame identification model based on a deep learning model, and verifying the trained flame identification model by inference through image data other than the data set.

6. The real-time monitoring method for process behavior in an RH refining furnace according to claim 3, characterized in that The calculation of the top lance height by using a top lance height calculation model comprises: ​ During the non-refining period, the image acquisition and monitoring unit is located at different heights to acquire images of the immersion tube at the bottom of the refining furnace and the image of the bottom of the refining furnace, and the relationship between the height of the top lance and the cross-sectional circular diameter of the immersion tube is fitted by the least square method. The height position of the top lance can be calculated by calculating the number of pixels occupied by the cross-sectional circular diameter of the acquired image.

7. The method according to claim 3, wherein the real-time height information of the molten steel liquid level in the RH refining furnace is calculated by using a liquid level calculation model, and the method comprises the following steps: During the refining stable period, first, the image acquisition and monitoring unit is calibrated by Zhang Zhengyou calibration method, and then the distance from the molten steel liquid level to the image acquisition and monitoring unit is calculated by similar triangle measurement method. The height value of the molten steel liquid level during the refining stable period is obtained by subtracting the distance from the molten steel liquid level to the image acquisition and monitoring unit from the preset distance from the image acquisition and monitoring unit to the bottom of the refining furnace.

8. The method according to claim 3, wherein the specific state of the refining is determined by using a deep learning model, and the method comprises the following steps: The first part: training of the model First, the image acquisition and monitoring unit installed on the top lance acquires process images of the molten steel refining, including 500 images of molten steel charging, alloy adding, oxygen blowing state and normal refining. Then, data enhancement operations including rotation, scaling and flipping are performed on the images. Then, for different processes, the significant features that are obviously different from other processes are found and labeled. After labeling the data, batch standardization processing is performed on the data. Secondly, the labeled images are divided into a training set, a test set and a validation set. Finally, the model training is performed: define the model structure of YOLOv5, including convolution layer, pooling layer, up-sampling layer, define loss function and initialize weight; After multiple rounds of training, the weight file of the optimal structure is output, and the optimal result is used to infer 10,000 refining process images, and the inference accuracy rate of each process is counted. If the inference accuracy rate is not up to standard, retraining and testing are performed. The second part: application of the model First, the real-time image in the RH refining furnace is acquired by the image acquisition and monitoring unit, and the image is transmitted to the data processing unit through the data transmission unit. The data processing unit calls the deep learning model of YOLOv5 and the optimal training result weight file to identify the image, determines whether it is a refining process, and transmits the determination result to the data display unit for display. If it is, the specific refining process is displayed. If not, it displays waiting for detection, and the next image is read, and the above identification process is repeated.

9. The method according to claim 5, wherein the training process of the flame identification model comprises: During the refining period, the top lance flame images under different process states and the flame images during the melting of cold steel are collected for training. In order to prevent overfitting during training, images without flame and without molten steel are added for training. ​ ​ ​ 10. The method of claim 7, wherein the method comprises the following steps: The similar triangle measurement method comprises the following steps: Step one, calibrate the image acquisition and monitoring unit, and obtain the intrinsic matrix and extrinsic matrix of the camera; Step two, take a single frame image containing the target through the calibrated image acquisition and monitoring unit, process the obtained image, extract the target circular region, and calculate the pixel number of the minimum circumscribed circle diameter of the region; Step three, calculate the distance from the image acquisition and monitoring unit to the target according to the similar triangle principle.

Images