Converter full-automatic blowing method and system based on multi-mode perception and intelligent regulation and control
By employing multimodal sensing and intelligent control methods, real-time converter smelting data is acquired. Data fusion is performed using convolutional neural networks and long short-term memory networks, and dynamic prediction and control are achieved by combining a hybrid intelligent model. This solves the problems of automation and real-time sensing in converter steelmaking endpoint control, and realizes precise control throughout the entire process.
Patent Information
- Application Number
- CN202511648118.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-02-06
AI Technical Summary
Existing methods for controlling the end point of converter steelmaking rely on single-point measurements and experience-based intervention, which cannot achieve full-process automation and lack sufficient perception of the furnace state, resulting in delayed control response and prediction bias.
By employing a multimodal sensing and intelligent control method, multimodal sensing data of the converter smelting process is acquired in real time. Data is fused through convolutional neural networks and long short-term memory networks, and combined with a hybrid intelligent model for dynamic prediction and control, thereby achieving closed-loop control of the blowing process.
It enables real-time, comprehensive perception and dynamic adjustment of the converter steelmaking process, accurately controls the final carbon content, phosphorus content and temperature, and achieves full automation without human intervention.
Smart Images
Figure CN121472508A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of iron and steel smelting technology, specifically to a fully automated converter blowing method and system based on multimodal sensing and intelligent control. Background Technology
[0002] Currently, endpoint control in converter steelmaking primarily relies on single-point measurements taken by the auxiliary lance during the mid-blowing process, combined with static or semi-static material / heat balance models for endpoint prediction. However, this traditional method has several inherent limitations: First, auxiliary lance measurements only provide a snapshot at a specific moment, failing to capture the dynamic changes throughout the entire blowing process. In particular, it lacks the real-time identification and early warning capabilities for key conditions such as slag formation and state evolution, splashing, and re-drying, resulting in a significant lag in control response. Second, calculations based on single-point data are susceptible to fluctuations in the early blowing process and changes in raw material conditions, leading to predictions of endpoint carbon content, phosphorus content, and temperature. Furthermore, existing methods still heavily rely on operator experience for intervention, failing to achieve fully automated, unattended operation.
[0003] Therefore, there is an urgent need for a fully automated blowing method that can perceive the furnace status in real time and make dynamic adjustments accordingly. Summary of the Invention
[0004] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: In a first aspect, embodiments of this application provide a fully automated converter blowing method based on multimodal sensing and intelligent control, including: Real-time acquisition of multimodal sensing data during the converter smelting process; Feature extraction is performed on the multimodal sensing data, and a high-dimensional multimodal feature vector is constructed by combining the static input features; The multimodal high-dimensional feature vector is input into a long short-term memory network for data fusion and the current smelting state is determined. Based on the current smelting status, a hybrid intelligent model is used to dynamically predict the final carbon content, phosphorus content, and temperature, and generate dynamic control commands. The oxygen lance height, oxygen flow rate, and auxiliary material addition are adjusted in real time according to the control commands to achieve closed-loop control of the blowing process.
[0005] In one possible implementation, the real-time acquisition of multimodal sensing data during the converter smelting process includes: Real-time carbon content and temperature data during the middle stage of blowing were obtained using a secondary lance; Visual images of the flames and infrared temperatures at the furnace opening were acquired using a multispectral high-speed camera and an infrared thermal imager installed at the furnace opening, respectively. The audio signal at the furnace opening is acquired by a high-frequency sound pressure sensor installed at the flue location.
[0006] In one possible implementation, feature extraction is performed on the multimodal sensing data, and a high-dimensional multimodal feature vector is constructed by combining the static input features, including: After preprocessing the flame visual image, a convolutional neural network model is used to automatically extract the flame's shape contour, average brightness, color distribution, and texture features to form a flame visual feature vector. Perform a short-time Fourier transform on the furnace mouth audio signal, generate a time-spectrum graph, and then extract the energy value of the preset frequency band. A dynamic correlation model between infrared temperature and actual molten pool temperature is established, and the thermal state of the molten pool in the measurement gap of the sub-gun is continuously estimated using the dynamic correlation model. Based on a unified timestamp, the extracted multimodal features are aligned. The aligned multimodal features are concatenated with the static input features to form a multimodal high-dimensional feature vector.
[0007] In one possible implementation, after preprocessing the flame visual image, the flame shape contour, average brightness, color distribution and texture features are automatically extracted using a convolutional neural network model to form a flame visual feature vector. This includes: inputting the acquired flame visual image into the input layer of the convolutional neural network model, normalizing the input image, and scaling the pixel values. Multiple large convolutional kernels are used to quickly reduce the feature map resolution while using large convolutional kernels to capture a large range of the flame's outline and brightness area. The overall shape and outline of the flame are extracted using the macro-morphology branch of the convolutional neural network model, the color and brightness information of the flame are extracted using the color and brightness branch of the model, and the fine texture, edges and local fluctuations of the flame are extracted using the micro-texture branch of the model. The feature vectors output from the macroscopic morphology branch, color and brightness branch, and microscopic texture branch are concatenated and then input into a fully connected layer for advanced feature fusion and nonlinear transformation to form a flame visual feature vector.
[0008] In one possible implementation, the multimodal high-dimensional feature vector is input into a long short-term memory network for data fusion and to determine the current smelting state, including: After receiving the multimodal high-dimensional feature vector, the Long Short-Term Memory Network associates the features of the multimodal high-dimensional feature vector with a specific smelting state to achieve pattern matching. When the flame brightness fluctuates drastically in the multimodal high-dimensional feature vector and the audio energy of the preset frequency band continues to decrease, it is matched with the anti-drying mode and the output shows a high risk of anti-drying. When the flame shape is stable and the color is bright in the multimodal high-dimensional feature vector, and the audio energy of the preset frequency band is stable and moderate, it is matched with the slag removal mode to output good slag removal. When the multimodal high-dimensional feature vector shows a sharp increase in flame brightness, a rapid expansion of flame area, and an abnormal increase in audio energy in a preset frequency band, it is matched with the splashing mode, and the output is "high splashing risk". By analyzing the historical trends in flame area fluctuation frequency and flame color, the decarbonization rate can be determined.
[0009] In one possible implementation, the hybrid intelligent model is a hybrid architecture combining a metallurgical mechanism module and a deep learning module; the metallurgical mechanism module calculates the trajectory based on the mass-energy conservation theory, and the deep learning module is used to learn complex nonlinear relationships and process the time series of multimodal sensing data for state recognition and dynamic prediction.
[0010] In one possible implementation, the step of dynamically predicting the final carbon content, phosphorus content, and temperature using a hybrid intelligent model based on the current smelting state and generating dynamic control commands includes: Based on the current smelting status, incremental learning optimization is performed based on the deviation between the actual and predicted values at the blowing endpoint. The NSGA-II optimization algorithm is used to quickly generate multiple curves with different control strategies from the current moment to the end of the blowing process; For each strategy curve, a virtual refining process is performed to predict the final carbon content, phosphorus content, temperature, oxygen consumption, and time. The multi-objective optimization function is based on high endpoint hit probability, low consumption cost, and short refining time. Forward simulation is performed based on a hybrid model, and the final control strategy is selected from the Pareto optimal solution set as the dynamic control command.
[0011] In one possible implementation, the oxygen lance height, oxygen flow rate, and auxiliary material addition are adjusted in real time according to the control command to achieve closed-loop control of the blowing process, including: The oxygen lance frequency converter, oxygen lance regulating valve, and weighing and feeding system of the batching silo are controlled according to the control instructions. By controlling the oxygen lance frequency converter, the oxygen lance can be precisely positioned and dynamically adjusted in the vertical direction. By controlling the oxygen lance regulating valve, the oxygen supply intensity is adjusted in real time according to the set flow curve to ensure the precise supply of oxygen required for the molten pool reaction. The weighing and feeding system of the linked batching bins enables precise feeding during the blowing process.
[0012] One possible implementation also includes obtaining actual measurements of the final carbon content, phosphorus content, and temperature after the blowing process is completed; The multimodal sensing data, dynamic control commands, and endpoint measurement data of this blow-making process will be stored in the historical database. Initiate the model self-learning process and periodically use new data to incrementally train or fine-tune the Long Short-Term Memory Network and the Hybrid Intelligent Model, so that the Long Short-Term Memory Network and the Hybrid Intelligent Model can adapt to changes in raw material conditions and equipment status.
[0013] Secondly, embodiments of this application provide a fully automated converter blowing system based on multimodal perception and intelligent control, including: a dynamic intelligent decision-making module and a multimodal perception module and an execution module connected to the dynamic intelligent decision-making module. The multimodal perception module is used to acquire flame visual images, furnace mouth audio signals, furnace mouth infrared temperatures, secondary lance information, and static parameter data. The dynamic intelligent decision-making module is used to judge and dynamically predict the smelting state based on the multimodal perception data and output dynamic control commands. The execution module is used to perform related operations according to the control commands.
[0014] Compared with the prior art, the beneficial effects of this application are as follows: This application establishes a "digital brain" for multimodal perception fusion and dynamic optimization control. By integrating flame vision, audio spectrum, infrared temperature measurement and secondary lance data, it achieves real-time and comprehensive perception of the converter smelting status. It also utilizes an advanced hybrid intelligent model to dynamically adjust oxygen blowing and feeding strategies, ultimately achieving precise targeting of the final carbon content, phosphorus content and temperature, and unmanned operation throughout the entire process. Attached Figure Description
[0015] Figure 1 A schematic diagram of a fully automated converter blowing method based on multimodal sensing and intelligent control provided for an embodiment of this application; Figure 2 A schematic diagram of the fully automated converter blowing method based on multimodal sensing and intelligent control provided for the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the convolutional neural network model provided in the embodiments of this application; Figure 4 This is a schematic diagram of a fully automated converter blowing method based on multimodal sensing and intelligent control, provided in an embodiment of this application. Detailed Implementation
[0016] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.
[0017] Figure 1A schematic flowchart of a fully automated converter blowing method based on multimodal sensing and intelligent control, provided for an embodiment of this application, is shown below. Figure 1 and Figure 2 This application provides a fully automated converter blowing method based on multimodal sensing and intelligent control, comprising: S101 acquires multimodal sensing data in real time during the converter smelting process.
[0018] In this embodiment, splashing refers to the violent outflow of materials such as slag and molten metal from the furnace. This instantly causes the flame at the furnace opening to become abnormally bright, large, and unstable. Therefore, a rapid and dramatic increase in flame brightness and area is its typical visual sign. The splashing process, accompanied by a large amount of material and furnace gas impacting the furnace opening and flue at high speed, produces specific and intense noise, which manifests as sudden spikes or bursts of energy in specific frequency bands on the audio spectrum. A multispectral visible light and near-infrared high-speed camera is installed at the furnace opening, equipped with a high-temperature resistant, dustproof, and splash-proof protective kit, to acquire flame images at a rate of 100-200 ms / frame.
[0019] High-frequency sound pressure sensors are installed at specific locations in the flue or furnace body to collect the raw audio signal of 20Hz-20kHz at the furnace opening. To ensure the accuracy of the data collection, multiple sensor redundancy can be added to improve the robustness of the system.
[0020] An online infrared thermal imager is installed at the furnace opening to measure the radiation temperature at the flame root or a specific area at the furnace opening. Temperature field distribution data is collected, and its average temperature value is calculated. Real-time carbon content and temperature data during the mid-blowing process are acquired using a secondary lance.
[0021] S102, extract features from multimodal sensing data and construct multimodal high-dimensional feature vectors by combining static input features.
[0022] In this embodiment, after preprocessing the visual image of the flame, a convolutional neural network model is used to automatically extract the shape contour, average brightness, color distribution, and texture features of the flame. Simultaneously, the dynamic fluctuation frequency of the flame area is calculated as an indicator of the severity of decarburization. The calculation formula is as follows: in, The power of the frequency component k, The actual frequency corresponding to each frequency component. The sampling time interval, It is the complex result of the k-th frequency component. It is the nth data point after preprocessing. yes The modulus of N is the sequence length, and k is the frequency component.
[0023] A frequency range of interest [fmin, fmax] is typically set, for example, from 0.1 Hz to 2 Hz. This range excludes extremely slow trend changes and extremely fast noise, focusing on the periodic fluctuations caused by the reaction. Low frequencies, such as 0.2 Hz, indicate slow changes in flame area and a gentle decarburization reaction. High frequencies, such as 1.5 Hz, indicate rapid and violent flickering and jumping of the flame area, meaning that the decarburization reaction in the molten pool is very intense, and carbon monoxide gas is being rapidly generated and released.
[0024] See Figure 3 In this embodiment, the convolutional neural network model includes an input layer, a primary feature extraction layer, a multi-branch refined feature extraction layer, and a feature fusion and output layer. After the acquired 256 x 256 x 3 flame image is input to the input layer, the input image is normalized, scaling the pixel values from [0, 255] to [-1, 1] or [0, 1] to accelerate model training convergence. Subsequently, 64 large 7x7 convolutional kernels with a stride of 2 are used to quickly reduce the feature map resolution, while simultaneously capturing a large area of the flame's outline and brightness region, resulting in an output tensor size of 128x128x64.
[0025] The multi-branch refined feature extraction layer, as the core of the convolutional neural network model, aims to extract different types of key features in parallel, including macroscopic morphology, color and brightness, and microscopic texture branches. The macroscopic morphology branch focuses on extracting the overall shape and outline of the flame. A 5x5 convolutional kernel can perceive a large area, suitable for capturing the overall morphological fluctuations of the flame. Finally, global average pooling yields a 128-dimensional feature vector representing the global morphological information of the image. In the color and brightness branch, the 1x1 convolution essentially performs a linear weighted combination of each pixel in the input feature map, equivalent to color adjustment in the channel dimension, effectively learning color distribution and brightness information, such as whether the flame is bright yellow or dark red. This branch is computationally efficient, ultimately outputting a 64-dimensional color and brightness feature vector. In the microscopic texture branch, two consecutive 3x3 small convolutional kernels are combined, increasing the network depth and non-linearity without adding too many parameters, making it ideal for extracting fine textures, edges, and local fluctuations of the flame, such as flickering and vortex structures. Global max pooling highlights the most significant features, ultimately outputting a 256-dimensional texture feature vector. The feature vectors output from the three branches are concatenated to form a comprehensive feature vector. This concatenated 448-dimensional feature vector is then input into a fully connected layer with 512 neurons for advanced feature fusion and nonlinear transformation, learning the complex relationships between different features such as shape, color, and texture. The 512-dimensional feature vector is output through the output layer; this high-dimensional feature vector will serve as input to subsequent models such as the Long Short-Term Memory network.
[0026] The NN structure in this embodiment, through a multi-branch design, selectively and efficiently extracts three types of visual features crucial for determining smelting states such as reflow, splashing, and slagging: the shape and outline of the flame, color brightness, and micro-texture. It transforms a raw flame image into a compact 512-dimensional numerical vector containing rich semantic information, providing reliable visual perception capabilities for the entire intelligent blowing system.
[0027] A short-time Fourier transform was performed on the furnace mouth audio signal to generate a time-frequency spectrum. The energy values in the 200-800Hz frequency band were then focused on, as this band's energy is highly correlated with the degree of slag foaming. The standard deviation and mean of this frequency band's energy were calculated as key indicators for assessing the risk of back-drying or splashing.
[0028] A dynamic correlation model between infrared temperature and actual molten pool temperature is established. This model is used to continuously estimate the thermal state of the molten pool during the measurement gap of the secondary lance. Using a unified timestamp as a benchmark, the extracted multimodal features are aligned to ensure that the observed flame and the heard sound are from the same moment. The aligned multimodal features are then concatenated with the static input features to form a multimodal high-dimensional feature vector. This high-dimensional feature vector is the "digital fingerprint" of the furnace state at the current moment. In this embodiment, the static input features include molten iron weight, composition, and scrap steel ratio. The multimodal high-dimensional feature vector may be [flame brightness: 0.85, flame area: 1200 pixels, audio 800Hz energy: 0.72, infrared temperature: 1580°C, secondary lance C: 0.25%, ..., molten iron Si: 0.4%].
[0029] S103, input the multimodal high-dimensional feature vector into the long short-term memory network for data fusion and judge the current smelting state.
[0030] See Figure 4 In this embodiment, a Long Short-Term Memory (LSTM) network is used to process the temporal sequence of the above features. LSTM can remember the state in the early stage of blowing, thereby more accurately judging the current smelting stage. After receiving the multimodal high-dimensional feature vector, the LSTM network associates the features of the multimodal high-dimensional feature vector with specific smelting states to achieve pattern matching. When the multimodal high-dimensional feature vector shows drastic fluctuations in flame brightness and a continuous decline in audio energy in the 200-400Hz frequency band, it is matched with the dry-return mode, and the output is "high risk of dry-return". When the multimodal high-dimensional feature vector shows a stable flame shape and bright color, and a stable and moderate audio energy in the 500-800Hz frequency band, it is matched with the slag-forming mode, and the output is "good slag-forming". When the multimodal high-dimensional feature vector shows a sharp increase in flame brightness, a rapid expansion of flame area, and an abnormal increase in audio energy in the 200-500Hz frequency band, it is matched with the splashing mode, and the output is "high risk of splashing". Furthermore, by analyzing the historical trends in flame area fluctuation frequency and flame color, the decarbonization rate can be determined.
[0031] S104 uses a hybrid intelligent model to dynamically predict the final carbon content, phosphorus content, and temperature based on the current smelting status and generates dynamic control commands.
[0032] In this embodiment, the hybrid intelligent model is a hybrid architecture combining a metallurgical mechanism module and a deep learning module. The metallurgical mechanism module calculates the trajectory based on the law of conservation of mass and energy, ensuring that the prediction results do not violate basic physicochemical laws. The deep learning module is used to learn complex nonlinear relationships and process the time series of multimodal sensing data for state recognition and dynamic prediction. The two modules work together to output a probability distribution, not just a single value. For example, it predicts the final carbon content to be 0.06%, but also provides a confidence interval (e.g., 0.04%~0.08%), which better reflects the uncertainties of the metallurgical process.
[0033] Subsequently, dynamic optimization of multiple objectives is performed. Based on the current smelting state, incremental learning optimization is carried out based on the deviation between the actual value and the predicted value at the blowing endpoint. The NSGA-II optimization algorithm is used to quickly generate multiple control strategy curves from the current time to the end of blowing. Virtual blowing is performed on each strategy curve to predict the final carbon content, phosphorus content, temperature, oxygen consumption, and time. The multi-objective optimization function is based on high endpoint hit probability, low consumption cost, and short blowing time. Forward simulation is performed based on a hybrid model, and the final control strategy is selected from the Pareto optimal solution set as the dynamic control command.
[0034] S105, the oxygen lance height, oxygen flow rate and auxiliary material addition are adjusted in real time according to the control command to realize closed-loop control of the blowing process.
[0035] In this embodiment, the oxygen lance frequency converter, oxygen lance regulating valve, and weighing and feeding system of the batching silo are controlled according to the control command. By controlling the oxygen lance frequency converter, the oxygen lance is accurately positioned and dynamically adjusted in the vertical direction. By controlling the oxygen lance regulating valve, the oxygen supply intensity is adjusted in real time according to the set flow curve to ensure the accurate supply of oxygen required for the molten pool reaction. By linking the weighing and feeding system of the batching silo, the precise feeding of materials in the blowing process is achieved.
[0036] After the blowing process is completed, the actual measured values of the final carbon content, phosphorus content and temperature are obtained. The multimodal sensing data, dynamic control commands and final measurement data of this blowing process are stored in the historical database. The model self-learning process is started. New data is used regularly to incrementally train or fine-tune the long short-term memory network and the hybrid intelligent model, so that the long short-term memory network and the hybrid intelligent model can adapt to changes in raw material conditions and equipment status, and achieve self-evolution that becomes more and more accurate with each refining process.
[0037] The process in this embodiment is a dynamic, continuous, closed-loop intelligent cycle. It is not a one-time event but repeats every second, reassessing the state, re-predicting the endpoint, and re-optimizing the strategy based on the latest perceived information. Moreover, the decision-making is based on the history of the entire refining process, not just the present moment. Its core is "foreseeing the future" and formulating the best action for the current moment accordingly. It is also self-validating and self-learning. The final result of each batch is compared with the predicted value, and the deviation is used for iterative optimization of the model, making this "digital brain" like a tireless, continuously learning super expert, becoming more and more proficient with each refining process. This embodiment fundamentally solves the pain points of traditional methods, such as "lagging, inaccuracy, and reliance on manual labor," and achieves true full-process automation and intelligence.
[0038] Corresponding to the fully automated converter blowing method based on multimodal sensing and intelligent control provided in the above embodiments, this application also provides a fully automated converter blowing system based on multimodal sensing and intelligent control, comprising: The system includes a dynamic intelligent decision-making module, a multimodal sensing module, and an execution module connected to the dynamic intelligent decision-making module. The multimodal sensing module is used to acquire visual images of the flame, audio signals from the furnace opening, infrared temperatures at the furnace opening, information from the secondary lance, and static parameter data. The dynamic intelligent decision-making module is used to judge and dynamically predict the smelting state based on the multimodal sensing data and output dynamic control commands. The execution module is used to perform related operations based on the control commands.
[0039] Furthermore, the fully automated converter blowing system based on multimodal perception and intelligent control is equipped with a visual human-machine interface. In this embodiment, the visual human-machine interface is used to display in real time the information perceived by the system, the status judgment results, the predicted endpoint trajectory, and the recommended control commands. Confirmation and rejection mechanism: When the system provides a suggestion, the operator should have the option to "confirm" or "reject" it with one click. All manual intervention operations, including the reasons, must be fully recorded and fed back as valuable data to the self-learning system.
[0040] Alarm classification: The system differentiates between warning, alarm, and emergency levels, specifying corresponding audible and visual alerts and notification methods for each level. Initially, the system provides decision-making suggestions through the HMI interface, with experienced furnace operators making the final decisions and executing manually. This verifies the system's accuracy, allows furnace operators to participate in optimization, and provides optimal safety redundancy. Once the system performs stably and gains operator trust, fully automated closed-loop control can be tested on individual work shifts or specific steel grades.
[0041] In terms of safety and fault tolerance mechanism design, multi-layered protection must be implemented. At the sensor level: The rationality of each sensor's data is verified; upon detection of anomalies, an immediate flag is issued, and a backup data source or model is activated. At the model level: Confidence levels are set for the model's predicted output. When the confidence level is too low, the system automatically degrades to a rule-based conservative control strategy and triggers an alarm requesting manual intervention. At the execution level: Deep integration with the PLC is implemented, setting soft and hard limits for all automatic commands and ensuring that manual override signals have the highest priority.
[0042] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.
[0043] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0044] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.
Claims
1. A fully automated converter blowing method based on multimodal sensing and intelligent control, characterized in that, include: Real-time acquisition of multimodal sensing data during the converter smelting process; Feature extraction is performed on the multimodal sensing data, and a high-dimensional multimodal feature vector is constructed by combining the static input features; The multimodal high-dimensional feature vector is input into a long short-term memory network for data fusion and the current smelting state is determined. Based on the current smelting status, a hybrid intelligent model is used to dynamically predict the final carbon content, phosphorus content, and temperature, and generate dynamic control commands. The oxygen lance height, oxygen flow rate, and auxiliary material addition are adjusted in real time according to the control commands to achieve closed-loop control of the blowing process.
2. The fully automated converter blowing method based on multimodal sensing and intelligent control according to claim 1, characterized in that, The real-time acquisition of multimodal sensing data during the converter smelting process includes: Real-time carbon content and temperature data during the middle stage of blowing were obtained using a secondary lance; Visual images of the flames and infrared temperatures at the furnace opening were acquired using a multispectral high-speed camera and an infrared thermal imager installed at the furnace opening, respectively. The audio signal at the furnace opening is acquired by a high-frequency sound pressure sensor installed at the flue location.
3. The fully automated converter blowing method based on multimodal sensing and intelligent control according to claim 1, characterized in that, Feature extraction is performed on the multimodal sensing data, and a high-dimensional multimodal feature vector is constructed by combining the static input features, including: After preprocessing the flame visual image, a convolutional neural network model is used to automatically extract the flame's shape contour, average brightness, color distribution, and texture features to form a flame visual feature vector. Perform a short-time Fourier transform on the furnace mouth audio signal, generate a time-spectrum graph, and then extract the energy value of the preset frequency band. A dynamic correlation model between infrared temperature and actual molten pool temperature is established, and the thermal state of the molten pool in the measurement gap of the sub-gun is continuously estimated using the dynamic correlation model. Based on a unified timestamp, the extracted multimodal features are aligned. The aligned multimodal features are concatenated with the static input features to form a multimodal high-dimensional feature vector.
4. The fully automated converter blowing method based on multimodal sensing and intelligent control according to claim 3, characterized in that, After preprocessing the flame visual image, a convolutional neural network model is used to automatically extract the flame's shape contour, average brightness, color distribution, and texture features, forming a flame visual feature vector, including: After the acquired visual image of the flame is input into the input layer of the convolutional neural network model, the input image is normalized and the pixel values are scaled. Multiple large convolutional kernels are used to quickly reduce the feature map resolution while using large convolutional kernels to capture a large range of the flame's outline and brightness area. The overall shape and outline of the flame are extracted using the macro-morphology branch of the convolutional neural network model, the color and brightness information of the flame are extracted using the color and brightness branch of the model, and the fine texture, edges and local fluctuations of the flame are extracted using the micro-texture branch of the model. The feature vectors output from the macroscopic morphology branch, color and brightness branch, and microscopic texture branch are concatenated and then input into a fully connected layer for advanced feature fusion and nonlinear transformation to form a flame visual feature vector.
5. The fully automated converter blowing method based on multimodal sensing and intelligent control according to claim 1, characterized in that, The multimodal high-dimensional feature vector is input into a long short-term memory network for data fusion and to determine the current smelting state, including: After receiving the multimodal high-dimensional feature vector, the Long Short-Term Memory Network associates the features of the multimodal high-dimensional feature vector with a specific smelting state to achieve pattern matching. When the flame brightness fluctuates drastically in the multimodal high-dimensional feature vector and the audio energy of the preset frequency band continues to decrease, it is matched with the anti-drying mode and the output shows a high risk of anti-drying. When the flame shape is stable and the color is bright in the multimodal high-dimensional feature vector, and the audio energy of the preset frequency band is stable and moderate, it is matched with the slag removal mode to output good slag removal. When the multimodal high-dimensional feature vector shows a sharp increase in flame brightness, a rapid expansion of flame area, and an abnormal increase in audio energy in a preset frequency band, it is matched with the splashing mode, and the output is "high splashing risk". By analyzing the historical trends in flame area fluctuation frequency and flame color, the decarbonization rate can be determined.
6. The fully automated converter blowing method based on multimodal sensing and intelligent control according to claim 1, characterized in that, The hybrid intelligent model is a hybrid architecture that combines a metallurgical mechanism module and a deep learning module. The metallurgical mechanism module is based on the mass-energy conservation calculation theory trajectory, and the deep learning module is used to learn complex nonlinear relationships and process the time series of multimodal sensing data for state recognition and dynamic prediction.
7. The fully automated converter blowing method based on multimodal sensing and intelligent control according to claim 1, characterized in that, The process of dynamically predicting the final carbon content, phosphorus content, and temperature based on the current smelting state using a hybrid intelligent model and generating dynamic control commands includes: Based on the current smelting status, incremental learning optimization is performed based on the deviation between the actual and predicted values at the blowing endpoint. The NSGA-II optimization algorithm is used to quickly generate multiple curves with different control strategies from the current moment to the end of the blowing process; For each strategy curve, a virtual refining process is performed to predict the final carbon content, phosphorus content, temperature, oxygen consumption, and time. The multi-objective optimization function is based on high endpoint hit probability, low consumption cost, and short refining time. Forward simulation is performed based on a hybrid model, and the final control strategy is selected from the Pareto optimal solution set as the dynamic control command.
8. The fully automated converter blowing method based on multimodal sensing and intelligent control according to claim 1, characterized in that, The oxygen lance height, oxygen flow rate, and auxiliary material addition are adjusted in real time according to the control commands to achieve closed-loop control of the blowing process, including: The oxygen lance frequency converter, oxygen lance regulating valve, and weighing and feeding system of the batching silo are controlled according to the control instructions. By controlling the oxygen lance frequency converter, the oxygen lance can be precisely positioned and dynamically adjusted in the vertical direction. By controlling the oxygen lance regulating valve, the oxygen supply intensity is adjusted in real time according to the set flow curve to ensure the precise supply of oxygen required for the molten pool reaction. The weighing and feeding system of the linked batching bins enables precise feeding during the blowing process.
9. The fully automated converter blowing method based on multimodal sensing and intelligent control according to claim 1, characterized in that, Also includes: After the blowing process is completed, the actual measured values of the final carbon content, phosphorus content, and temperature are obtained; The multimodal sensing data, dynamic control commands, and endpoint measurement data of this blow-making process will be stored in the historical database. Initiate the model self-learning process and periodically use new data to incrementally train or fine-tune the Long Short-Term Memory Network and the Hybrid Intelligent Model, so that the Long Short-Term Memory Network and the Hybrid Intelligent Model can adapt to changes in raw material conditions and equipment status.
10. A fully automated converter blowing system based on multimodal sensing and intelligent control, characterized in that, include: The system includes a dynamic intelligent decision-making module, a multimodal sensing module, and an execution module connected to the dynamic intelligent decision-making module. The multimodal sensing module is used to acquire flame visual images, furnace mouth audio signals, furnace mouth infrared temperatures, secondary lance information, and static parameter data. The dynamic intelligent decision-making module is used to judge and dynamically predict the smelting state based on the multimodal sensing data and output dynamic control commands. The execution module is used to perform related operations based on the control commands.