A converter oxygen supply-lance position collaborative control method based on flame visual recognition

Abstract

The present application relates to a kind of converter oxygen supply-gun site collaborative control method based on flame visual identification, comprising: using CNN network to extract the image features of preprocessed furnace mouth flame image;From the light intensity value of target characteristic wavelength extracted from spectral data as flame spectral features;Image features and flame spectral features are spliced and input into BP neural network to obtain molten pool state prediction model by training;The real-time state of molten pool predicted, flame feature vector is used as state input, and the training of reinforcement learning control model is carried out;In the refining process, the flame feature vector collected in real time is input into the trained control model to obtain operating parameters.The present application first realizes the fusion of furnace mouth flame multidimensional features and deep reinforcement learning, breaks through the bottleneck of traditional single parameter control, oxygen supply and gun site disconnection, realizes the dynamic collaborative matching of both, can adapt to the fluctuation of molten iron composition, steel grade switching and other complex working conditions, without manual parameter adjustment, reduce the dependence on the experience of operator.

Description

Technical Field

[0001] This invention relates to the field of iron and steel smelting technology, specifically to a converter oxygen supply-lance position coordinated control method based on flame visual recognition. Background Technology

[0002] The oxygen supply intensity and oxygen lance position in the converter are core process parameters in the blowing process, directly determining the decarburization rate of the molten pool, slag evolution, and the stability of the smelting process. Traditional control methods rely entirely on the on-site experience of operators, manually adjusting the parameters by visually observing the flame state at the furnace mouth. This results in inconsistent operating standards and poor adaptability to different operating conditions. Improper parameter adjustments can easily lead to molten pool splashing and slag drying, resulting in decreased metal yield, prolonged smelting cycle, and fluctuations in molten steel quality.

[0003] In existing technologies, some solutions rely on auxiliary lance detection data for single-parameter control of lance position or oxygen supply. However, auxiliary lance detection suffers from significant time delays and cannot reflect real-time dynamic changes in the molten pool. A few flame recognition-related control technologies only associate flame characteristics with single parameters such as oxygen supply or lance position, failing to achieve coordinated control between the two. Furthermore, they employ traditional machine learning models, lacking self-learning and dynamic optimization capabilities, resulting in control accuracy and process adaptability that still fall short of the demands for efficient and intelligent steelmaking. Overall, existing technologies have not yet achieved real-time coordinated and adaptive control of oxygen supply and lance position, failing to overcome the technical drawbacks of reliance on manual experience and single-parameter control. Summary of the Invention

[0004] To address the aforementioned problems, the purpose of this invention is to provide a converter oxygen supply-lance position coordinated control method based on flame visual recognition.

[0005] A converter oxygen supply-lance position coordinated control method based on flame visual recognition includes:

[0006] Step 1: Acquire images of the furnace flame and radiation spectrum data at the converter furnace opening;

[0007] Step 2: Preprocess the furnace flame image and radiation spectrum data to obtain the preprocessed furnace flame image and radiation spectrum data;

[0008] Step 3: Use a CNN network to extract image features from the preprocessed furnace flame image;

[0009] Step 4: Extract the light intensity values ​​of characteristic wavelengths of 589nm, 656nm, and 760nm from the spectral data as flame spectral characteristics;

[0010] Step 5: Concatenate the image features and flame spectral features to obtain the flame feature vector, and input it into the BP neural network for training to obtain the molten pool state prediction model;

[0011] Step 6: Construct a reinforcement learning reward function with the objectives of maximizing decarbonization efficiency, minimizing splashing and drying risk, and optimizing oxygen consumption. The predicted real-time state of the molten pool and the flame feature vector are used as state inputs, and the oxygen supply intensity and oxygen lance position adjustment are used as action outputs. A deep Q-network is used to train the reinforcement learning control model.

[0012] Step 7: During the blowing process, the flame feature vectors collected and processed in real time are input into the trained reinforcement learning control model, which outputs the optimal oxygen supply intensity and oxygen lance position parameters under the current working conditions, and sends them to the converter PLC control system for execution.

[0013] Preferably, step 2 includes:

[0014] Two-dimensional Gaussian filtering was used to smooth and denoise the furnace flame image to obtain a denoised furnace flame image;

[0015] Histogram equalization is performed on the denoised furnace flame image to obtain the equalized furnace flame image.

[0016] The region of interest in the equalized furnace flame image is extracted to obtain the preprocessed furnace flame image.

[0017] Preferably, step 2 further includes:

[0018] Baseline correction is performed on the radiation spectral data to obtain the corrected radiation spectral data;

[0019] The corrected radiation spectrum data is filtered using a moving average smoothing algorithm to obtain filtered radiation spectrum data.

[0020] The statistical 3σ criterion was used to remove outliers from the spectral data to obtain preprocessed radiation spectral data.

[0021] Preferably, in step 3, the CNN network uses parallel 3×3 and 7×7 dual-scale convolutional kernels to extract local features and global distribution features of the image, respectively.

[0022] Preferably, in step 5, minimum-maximum normalization is used to map the image features and flame spectral features to the same dimension range, and the normalized features are serially concatenated to obtain the flame feature vector.

[0023] Preferably, in step 5, the backpropagation algorithm is used to train the BP neural network with mean squared error as the loss function.

[0024] The present invention also provides an electronic device, including a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor. The transceiver, the memory, and the processor are connected via the bus. The computer program, when executed by the processor, implements the steps in the above-described method for coordinated control of converter oxygen supply and lance position based on flame visual recognition.

[0025] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps in the above-described method for coordinated control of converter oxygen supply and lance position based on flame visual recognition.

[0026] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0027] This invention relates to a converter oxygen supply-lance position coordinated control method based on flame visual recognition. Compared with the prior art, this invention is the first to realize the fusion of multi-dimensional features (image + spectrum) of furnace flame with deep reinforcement learning, breaking through the bottleneck of traditional single parameter control and disconnection between oxygen supply and lance position, realizing dynamic coordinated matching between the two, which can adapt to complex working conditions such as fluctuation of molten iron composition and steel grade switching, without the need for repeated manual parameter adjustment, reducing the dependence on operator experience.

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of a converter oxygen supply-lance position coordinated control method based on flame visual recognition provided by the present invention. Detailed Implementation

[0031] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0032] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0033] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0034] Please see Figure 1 This invention provides a converter oxygen supply-lance position coordinated control method based on flame visual recognition, comprising:

[0035] Step 1: Acquire images of the furnace flame and radiation spectrum data at the converter furnace opening;

[0036] Step 2: Preprocess the furnace flame image and radiation spectrum data to obtain the preprocessed furnace flame image and radiation spectrum data;

[0037] Step 3: Use a CNN network to extract image features from the preprocessed furnace flame image;

[0038] Step 4: Extract the light intensity values ​​of characteristic wavelengths of 589nm, 656nm, and 760nm from the spectral data as flame spectral characteristics;

[0039] Step 5: Concatenate the image features and flame spectral features to obtain the flame feature vector, and input it into the BP neural network for training to obtain the molten pool state prediction model;

[0040] Step 6: Construct a reinforcement learning reward function with the objectives of maximizing decarbonization efficiency, minimizing splashing and drying risk, and optimizing oxygen consumption. The predicted real-time state of the molten pool and the flame feature vector are used as state inputs, and the oxygen supply intensity and oxygen lance position adjustment are used as action outputs. A deep Q-network is used to train the reinforcement learning control model.

[0041] Step 7: During the blowing process, the flame feature vectors collected and processed in real time are input into the trained reinforcement learning control model, which outputs the optimal oxygen supply intensity and oxygen lance position parameters under the current working conditions, and sends them to the converter PLC control system for execution.

[0042] The principles of the present invention will be further explained below with reference to specific embodiments:

[0043] This embodiment is based on a 100t top-and-bottom blown converter, suitable for smelting various steel grades such as carbon steel and low-alloy steel. It integrates with the converter's existing high-temperature industrial vision system, PLC control system, and secondary lance detection system, without any additional large-scale equipment modifications. The specific implementation steps are as follows:

[0044] Step 1: Deployment and debugging of the visual acquisition system

[0045] Two high-speed, high-temperature resistant industrial cameras are symmetrically deployed on both sides of the converter opening. The camera parameters are: resolution 2048×1536, frame rate 50fps, and temperature resistance 800℃. They are equipped with a spectral detection module with a detection wavelength range of 200~1000nm, which enables real-time acquisition of furnace opening flame images and radiation spectrum data without blind spots.

[0046] The acquisition equipment was calibrated to correct lens distortion and eliminate ambient light interference. The data acquisition frequency was set to 50Hz, and the data was transmitted to the industrial big data pool via 5G industrial Ethernet with the transmission delay controlled within 100ms.

[0047] Step 2: Flame Data Preprocessing and Feature Vector Construction

[0048] Data preprocessing: Gaussian denoising and histogram equalization enhancement were performed on the flame images, and the core area of ​​the flame at the furnace mouth (800×600 pixels) was cropped and retained; baseline correction and smoothing were performed on the spectral data, and abnormal spectral values ​​were removed using the 3σ criterion;

[0049] (I) Flame Image Data Preprocessing Steps + Algorithm Formula

[0050] (1) Gaussian noise reduction

[0051] Two-dimensional Gaussian filtering is used to smooth and denoise the flame image, suppressing high-temperature smoke and salt-and-pepper noise and environmental interference noise.

[0052] Gaussian kernel formula:

[0053]

[0054] The noise is filtered out by performing a convolution operation between the neighborhood of each pixel in the image and a Gaussian kernel, replacing the original pixel value.

[0055] (2) Histogram equalization image enhancement

[0056] By stretching the dynamic grayscale range of the flame image through grayscale transformation, the flame outline and brightness levels are enhanced, highlighting bright areas and edge features.

[0057] Discrete histogram equalization transformation formula:

[0058]

[0059] Where L is the total number of gray levels in the image, P r (r j ) represents the probability of the j-th gray level appearing, thereby enhancing the contrast of the flame image.

[0060] (3) Cropping of region of interest

[0061] Fixed cropping retains the core flame area of ​​the furnace opening at 800×600 pixels, while invalid areas of the furnace wall and background are removed, reducing computational load and improving feature extraction accuracy.

[0062] (II) Spectral data preprocessing steps + algorithm formula

[0063] (1) Baseline correction

[0064] A polynomial baseline fitting algorithm is used to subtract the spectral drift baseline, restore the true characteristic spectral signal, and eliminate the interference of high-temperature radiation background drift.

[0065] (2) Spectral smoothing

[0066] A moving average smoothing algorithm is used to apply neighborhood mean filtering to the wavelength sequence light intensity values ​​to smooth out random fluctuation noise.

[0067] (3) 3σ outlier removal criteria

[0068] The statistical 3σ criterion is used to remove spectral outliers. The formula is:

[0069]

[0070] μ is the mean spectral intensity, and σ is the standard deviation. xi that satisfies the above formula is identified as an outlier and removed to ensure the reliability of the spectral data.

[0071] Multi-scale feature extraction: A CNN network is used to extract flame features. 3×3 small-scale convolutional kernels are used to extract detailed features such as flame edge texture and root shape, while 7×7 large-scale convolutional kernels are used to extract global features such as overall flame brightness distribution and flame height. The light intensity values ​​of characteristic wavelengths of 589nm, 656nm, and 760nm are extracted from the spectral data as flame spectral features.

[0072] Feature vector fusion: The 256-dimensional image features and 3-dimensional spectral features extracted by CNN are normalized and then concatenated to form a 259-dimensional flame feature vector, which serves as the basic input data for the model.

[0073] The specific steps, algorithms, and formulas for constructing a 259-dimensional flame feature vector based on CNN are as follows:

[0074] Step 1: Multi-scale CNN convolutional feature extraction

[0075] (1) Construct a lightweight CNN feature extraction network and use dual-scale convolutional kernels to extract features in parallel:

[0076] 3×3 small-scale convolution kernel: extracts flame edge texture, root details, and subtle contour features;

[0077] 7×7 large-scale convolution kernel: extracts the overall shape of the flame, flame height, and global brightness distribution features.

[0078] (2) General formula for convolution operation:

[0079]

[0080] W mn is the convolution kernel weight, x is the input image pixel, and b is the bias. The convolution sliding traverses the entire flame image and outputs a deep image feature map.

[0081] (3) The depth features of the flame image are output after dimensionality reduction by pooling and fully connected layers.

[0082] Step 2: Spectral Feature Extraction

[0083] Three characteristic wavelengths, 589nm, 656nm, and 760nm, are fixedly selected, and the corresponding light intensity values ​​are extracted to form a 3D spectral feature vector.

[0084] Step 3: Feature normalization + splicing and fusion

[0085] (1) Min-Max Normalization maps image features and spectral features to the same dimension interval [0,1]. Formula:

[0086]

[0087] (2) Dimensional splicing: The normalized 256-dimensional image features + 3-dimensional spectral features are directly spliced ​​in series to generate a 259-dimensional fused flame feature vector, which is used as the standard input for subsequent models.

[0088] Step 3: Training the nonlinear mapping model and the reinforcement learning model

[0089] Nonlinear mapping model construction: Select production data of more than 100,000 heats in Baogang No. 1 converter in the past two years, take 259-dimensional flame feature vector as input, and molten pool decarburization rate, slag state grade and molten pool temperature as output, and use BP neural network to train nonlinear mapping model. The model fit goodness R²≥0.92, realizing the accurate correlation between flame features and molten pool state.

[0090] The specific steps, algorithms, and formulas for establishing a nonlinear mapping model are as follows:

[0091] Step 1: Dataset Construction

[0092] A sample set was constructed by selecting historical production data from over 5000 heats of a 100t converter over the past year:

[0093] Model input: 259-dimensional flame feature vector

[0094] Model outputs: molten pool decarburization rate, slag condition grade, molten pool temperature

[0095] Step 2: Building the BP Neural Network Structure

[0096] A three-layer BP neural network is used: 259-dimensional input layer → hidden layer (with 128~256 neurons) → 3-dimensional output layer (corresponding to 3 melt pool state parameters).

[0097] The neuron activation function uses the Sigmoid function:

[0098]

[0099] Step 3: Model Training and Fit Optimization

[0100] (1) The backpropagation algorithm is adopted, with the mean squared error (MSE) as the loss function:

[0101]

[0102] y i This is the actual value. These are the model's predicted values.

[0103] (2) Iterate the training until the model fits R² > 0.92. The goodness-of-fit formula is:

[0104]

[0105] This is the sample mean.

[0106] (3) Divide the initial, middle and final stages of the blowing process, calibrate the flame characteristic thresholds for each stage, and establish a precise nonlinear mapping relationship between flame characteristics and molten pool state.

[0107] Reinforcement learning reward function design: Construct a multi-objective reward function with the formula: R = 0.4R_1 + 0.3R_2 + 0.3R_3; where R_1 is the decarburization efficiency reward (maximum value when decarburization rate ≥ 0.02% / s), R_2 is the splashing back-drying risk reward (maximum value when there is no splashing back-drying), and R_3 is the oxygen consumption reward (maximum value when oxygen consumption per ton of steel ≤ 55m³); introduce weight coefficients for the blowing stage: initial stage (0~30% oxygen supply) weight coefficient 1.0, middle stage (30%~80% oxygen supply) weight coefficient 1.2, and final stage (80%~100% oxygen supply) weight coefficient 1.1;

[0108] DQN model training: The real-time state of the molten pool (decarburization rate, slag state, temperature) and flame feature vectors are integrated into a 300-dimensional state space. The oxygen supply intensity adjustment (±0.2 m³ / min·t) and oxygen lance position adjustment (±0.5 m) are set as the action space. The actual production process of the converter is used as the training environment. The deep Q-network (DQN) is used to train the model. After 500,000 iterations, the model converges and the loss value stabilizes within 0.05.

[0109] Step 4: Real-time collaborative control and online feedback correction.

[0110] The specific steps, formulas, and algorithms for training reinforcement learning control models are as follows:

[0111] Step 1: Define the state space and action space

[0112] (1) State space: molten pool decarburization rate + slag state level + molten pool temperature + 259-dimensional flame characteristics, integrated into a 300-dimensional state vector;

[0113] (2) Action space:

[0114] Oxygen supply intensity adjustment: ±0.2 m³ / (min·t)

[0115] Oxygen lance position adjustment range: ±0.5m

[0116] Step 2: Construction of Multi-Objective Reward Function

[0117] Using a weighted combination reward function:

[0118] R = 0.4R1 + 0.3R2 + 0.3R3

[0119] R1: Decarbonization efficiency bonus, the maximum bonus value is taken if the decarbonization rate is ≥0.02% / s;

[0120] R2: Splash and re-drying risk bonus, the maximum value is taken when there is no splash or re-drying;

[0121] R3: Oxygen consumption bonus, the maximum value is taken when the oxygen consumption per ton of steel is ≤55m³.

[0122] Step 3: Assigning weight coefficients in stages

[0123] Initial stage of refining (0~30% oxygen supply): Weighting coefficient = 1.0

[0124] Mid-stage of refining (30%~80% oxygen supply): Weighting coefficient = 1.2

[0125] Late stage of refining (80%~100% oxygen supply): Weighting coefficient = 1.1

[0126] Adapt to the process focus of different smelting stages.

[0127] Step 4: Training the DQN Deep Q-Network

[0128] (1) DQN objective Q-value update formula:

[0129] Qtarget(s,a)=r+γmaxa'Q(s',a')

[0130] r is the immediate reward, γ is the discount factor, s is the current state, s' is the next state, a is the current action, and a' is the optimal next action.

[0131] (2) Using the actual smelting process of the converter as the interactive environment, iterative training was conducted 100,000 times;

[0132] (3) The training continues until the loss value is stable within 0.05, and the model converges and the optimal network weights are saved.

[0133] Step 5: Online feedback model correction

[0134] Data on carbon content, temperature, and flue gas from the secondary lance are collected every 10 seconds, and the relative deviation between the predicted and actual values ​​is calculated.

[0135]

[0136] when At the same time, the network weights of the BP nonlinear mapping model and the DQN model are dynamically adjusted to achieve adaptive and coordinated control under raw material fluctuations and changes in operating conditions.

[0137] Real-time control and execution: During the blowing process, the real-time flame feature vector is input into the nonlinear mapping model, which outputs the real-time state of the molten pool; the molten pool state is input into the trained DQN model, which outputs the optimal oxygen supply intensity and oxygen lance position parameters in real time, and sends them to the converter PLC control system, with an execution delay of ≤200ms; the parameter control range for each blowing stage is as follows:

[0138] Initial stage of blowing: oxygen supply intensity 3.0~3.5m³ / min·t, oxygen lance position 1.8~2.2m;

[0139] Mid-stage of blowing: oxygen supply intensity 3.5~4.0 m³ / min·t, oxygen lance position 1.5~1.8 m;

[0140] At the end of the blowing process: oxygen supply intensity 2.8~3.2 m³ / min·t, oxygen lance position 2.0~2.5 m;

[0141] Online feedback correction: The carbon content, temperature data and flue gas analysis data of the molten pool detected by the secondary gun are collected every 10 seconds to calculate the deviation between the actual molten pool state and the model prediction value; if the deviation is ≥5%, the weight parameters of the nonlinear mapping model and the DQN model are dynamically corrected to achieve adaptive control of oxygen supply and gun position.

[0142] According to the specific embodiments provided by the present invention, the following technical effects are disclosed:

[0143] 1. Improved control precision and real-time performance: For the first time, the fusion of multi-dimensional features of the furnace flame (image + spectrum) and deep reinforcement learning is realized, breaking through the bottleneck of traditional single parameter control and the disconnect between oxygen supply and gun position, realizing dynamic collaborative matching between the two. Combined with the online feedback mechanism, the delay of secondary gun detection is compensated, and the control response speed and precision are greatly improved.

[0144] 2. Enhanced stability of the smelting process: With "high decarburization efficiency and low splashing risk" as the core optimization goal, it effectively suppresses splashing and dry-out phenomena in the smelting process, improves the stability of the molten pool reaction, increases metal yield by ≥2%, and shortens the smelting cycle by ≥1 min / furnace.

[0145] 3. Significantly improved adaptability to working conditions: The model has self-learning and dynamic correction capabilities, and can adapt to complex working conditions such as fluctuations in molten iron composition and changes in steel grades. There is no need for repeated manual parameter adjustments, reducing reliance on the operator's experience.

[0146] 4. Easy to promote and integrate: This method can be directly connected to existing converter industrial vision systems and PLC control systems without large-scale equipment modification. It is easy to promote and apply in existing converter production lines of steel enterprises. It can also be integrated with converter endpoint prediction, furnace condition identification and other systems to form a complete intelligent steelmaking control system.

[0147] 5. After the application of this embodiment, the occurrence rate of splashing and re-drying in the converter blowing process decreased from 8% to 1.5%, the metal yield increased by 0.8 percentage points, and the smelting cycle was shortened by 1-2 minutes / furnace; the standard deviation of oxygen supply intensity and oxygen lance position parameter fluctuations decreased by 60% and 55% respectively, realizing precise coordinated control of oxygen supply and lance position.

[0148] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A converter oxygen supply-lance position coordinated control method based on flame visual recognition, characterized in that, include: Step 1: Acquire images of the furnace flame and radiation spectrum data at the converter furnace opening; Step 2: Preprocess the furnace flame image and radiation spectrum data to obtain the preprocessed furnace flame image and radiation spectrum data; Step 3: Use a CNN network to extract image features from the preprocessed furnace flame image; Step 4: Extract the light intensity values ​​of characteristic wavelengths of 589nm, 656nm, and 760nm from the spectral data as flame spectral characteristics; Step 5: Concatenate the image features and flame spectral features to obtain the flame feature vector, and input it into the BP neural network for training to obtain the molten pool state prediction model; Step 6: Construct a reinforcement learning reward function with the objectives of maximizing decarbonization efficiency, minimizing splashing and drying risk, and optimizing oxygen consumption. The predicted real-time state of the molten pool and the flame feature vector are used as state inputs, and the oxygen supply intensity and oxygen lance position adjustment are used as action outputs. A deep Q-network is used to train the reinforcement learning control model. Step 7: During the blowing process, the flame feature vectors collected and processed in real time are input into the trained reinforcement learning control model, which outputs the optimal oxygen supply intensity and oxygen lance position parameters under the current working conditions, and sends them to the converter PLC control system for execution.

2. The converter oxygen supply-lance position coordinated control method based on flame visual recognition according to claim 1, characterized in that, Step 2 includes: Two-dimensional Gaussian filtering was used to smooth and denoise the furnace flame image to obtain a denoised furnace flame image; Histogram equalization is performed on the denoised furnace flame image to obtain the equalized furnace flame image. The region of interest in the equalized furnace flame image is extracted to obtain the preprocessed furnace flame image.

3. The converter oxygen supply-lance position coordinated control method based on flame visual recognition according to claim 2, characterized in that, Step 2 also includes: Baseline correction is performed on the radiation spectral data to obtain the corrected radiation spectral data; The corrected radiation spectrum data is filtered using a moving average smoothing algorithm to obtain filtered radiation spectrum data. The statistical 3σ criterion was used to remove outliers from the spectral data to obtain preprocessed radiation spectral data.

4. The converter oxygen supply-lance position coordinated control method based on flame visual recognition according to claim 3, characterized in that, In step 3, the CNN network uses parallel 3×3 and 7×7 dual-scale convolutional kernels to extract local and global features of the image, respectively.

5. The converter oxygen supply-lance position coordinated control method based on flame visual recognition according to claim 4, characterized in that, In step 5, minimum-maximum normalization is used to map image features and flame spectral features to the same dimension range, and the normalized features are serially concatenated to obtain the flame feature vector.

6. The converter oxygen supply-lance position coordinated control method based on flame visual recognition according to claim 5, characterized in that, In step 5, the backpropagation algorithm is used to train the BP neural network with mean squared error as the loss function.

7. An electronic device comprising a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the transceiver, the memory, and the processor are connected via the bus, characterized in that, When the computer program is executed by the processor, it implements the steps in the converter oxygen supply-lance position coordinated control method based on flame visual recognition as described in any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps in the converter oxygen supply-lance position coordinated control method based on flame visual recognition as described in any one of claims 1-6.

Images