RH refining furnace control method and related equipment

Abstract

The invention discloses an RH refining furnace control method and related equipment, and relates to the technical field of steel smelting, the method comprises the following steps: obtaining in-furnace color image data of a target RH refining furnace; performing flame state recognition on the color image data in the furnace through a first preset model to obtain a flame detection result; performing material adding action recognition on the color image data in the furnace through a second preset model to obtain blanking event information; performing metallurgy stage feature extraction on the color image data in the furnace through a third preset model to obtain a refining state recognition result; and according to the flame detection result, the blanking event information and the refining state recognition result, interlocking control operation is triggered. Automatic judgment and interlocking control of key working conditions of the RH refining furnace are achieved through multi-model image recognition, the control precision and the system stability can be improved, and energy conservation, consumption reduction and refining quality improvement can be achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel smelting, and more particularly to an RH refining furnace control method and related equipment. BACKGROUND

[0002] With the continuous development of the metallurgical industry towards high efficiency, intelligence and safety, the RH (Ruhrstahl-Heraeus) vacuum refining furnace has become a key equipment in the steel production process due to its significant advantages in deep dehydrogenation, fine alloying and improving the cleanliness of molten steel. The RH refining stage has strict requirements for the stability of the top gun flame, the timing of alloy or slagging agent feeding, and the accurate determination of each metallurgical reaction stage. Any out-of-control link will cause energy consumption to rise, product quality to fluctuate, and even safety accidents. Therefore, how to perceive multiple key working conditions in the furnace in real time, comprehensively and accurately in a harsh environment coexisting with high temperature, strong light, dust and splashing, and execute safety interlocking and process optimization control accordingly, is a core issue that needs to be solved in the intelligent upgrading of steelmaking.

[0003] In the prior art, the flame state of the RH refining furnace mostly depends on ultraviolet or ion flow sensors and manual visual confirmation. The feeding behavior and metallurgical stage are mainly inferred through PLC logic timing, single-point temperature / pressure signals or manual experience estimation. Such methods have the following disadvantages: first, the sensors are easily disturbed by dust and high temperature in the furnace, resulting in high flame misjudgment rate; second, each measurement method is independent of each other, only providing discrete and single process information, and cannot fully reflect the coupling relationship between the flame, feeding and metallurgical reaction; third, relying on manual inspection and experience judgment, it is difficult to meet the requirements of modern steelmaking for continuous monitoring, millisecond-level response and predictive control, leading to lagging or false triggering of interlocking actions, thereby burying safety hazards and restricting energy efficiency and quality improvement. That is, the related art still has the technical problems of inaccurate identification of the flame, feeding and refining stage, and difficulty in coordinating process interlocking and low multivariable control precision. SUMMARY

[0004] A series of simplified concepts are introduced in the summary part of the present application, which will be further described in detail in the specific embodiment part. The summary part of the present application does not mean to attempt to limit the key features and necessary technical features of the claimed technical solution, and even less to determine the protection scope of the claimed technical solution.

[0005] The RH refining furnace control method and related equipment provided by the present application can realize automatic judgment and interlocking control of the key working conditions of the RH refining furnace through multi-model image recognition, which not only can improve the control precision and system stability, but also can realize energy saving and consumption reduction and improvement of refining quality.

[0006] In a first aspect, the application provides a RH refining furnace control method, comprising: obtaining in-furnace color image data of a target RH refining furnace; performing flame state recognition on the in-furnace color image data through a first preset model to obtain a flame detection result; performing material adding action recognition on the in-furnace color image data through a second preset model to obtain material unloading event information; performing metallurgical stage feature extraction on the in-furnace color image data through a third preset model to obtain a refining state recognition result; and triggering interlocking control operation according to the flame detection result, the material unloading event information and the refining state recognition result, wherein the interlocking control operation is used to regulate and control fuel supply or equipment start-stop of the target RH refining furnace.

[0007] In some embodiments, the first preset model is composed of a boundary detection network unit, a first feature extraction unit, a second feature extraction unit and a data fusion unit; the flame state recognition on the in-furnace color image data through the first preset model to obtain the flame detection result comprises: target detection on a single frame of the in-furnace color image data through the boundary detection network unit to obtain a flame position bounding box, wherein the boundary detection network unit is a convolutional neural network of YOLO architecture; extraction of a size change rate of the flame position bounding box in continuous frames of the in-furnace color image data through the first feature extraction unit; extraction of a center point displacement fluctuation value of the flame position bounding box in continuous frames of the in-furnace color image data through the second feature extraction unit; and determination of the flame detection result through the data fusion unit based on the size change rate and the center point displacement fluctuation value.

[0008] In some embodiments, the second preset model is composed of a first spatio-temporal feature extraction network unit, a preset event classifier and a preset material classifier, the material unloading event information includes material unloading event start-stop time points and material type classification results, and the material adding action recognition on the in-furnace color image data through the second preset model to obtain the material unloading event information comprises: extraction of dynamic target trajectory features in the in-furnace color image data through the first spatio-temporal feature extraction network unit, wherein the dynamic target trajectory features include material flow trajectories and equipment movement features; recognition of the material unloading event start-stop time points through the preset material classifier based on the material flow trajectories and the equipment movement features; and recognition of the material type classification results through the preset event classifier based on image region visual features corresponding to the material unloading event start-stop time points.

[0009] In some embodiments, the third preset model comprises a second spatio-temporal feature extraction network unit, a recurrent neural network unit, and a preset metallurgical stage classifier. The metallurgical stage feature extraction of the in-furnace color image data by the third preset model obtains a refining state recognition result, which comprises: multi-scale spatial feature extraction of the in-furnace color image data by the second spatio-temporal feature extraction network unit to obtain molten steel turbulent flow features and slag surface morphology features; time sequence evolution analysis and processing of the molten steel turbulent flow features and the slag surface morphology features by the recurrent neural network unit to obtain a feature evolution vector; and analysis of the feature evolution vector by the preset metallurgical stage classifier to obtain the refining state recognition result, wherein the refining state recognition result comprises decarburization stage identification, deoxidation stage identification, and alloying stage identification.

[0010] In some embodiments, according to the flame detection result, the downloading event information, and the refining state recognition result, an interlock control operation is triggered, which comprises: when the flame detection result indicates that the flame is not ignited in the ignition stage, the fuel supply and the ignition operation of the target RH refining furnace are stopped, and first warning information is issued; when the flame detection result indicates that the flame is extinguished, the oxygen supply and the downloading operation of the target RH refining furnace are stopped; when the downloading event start and end time points, the material type classification result, and the refining state recognition result do not meet the preset metallurgical path requirements, the downloading operation of the target RH refining furnace is stopped, and second warning information is issued; and when the refining state recognition result indicates the end of the first stage, an operation instruction for the second stage is generated, wherein the first stage and the second stage are process stages performed in sequence.

[0011] In some embodiments, the in-furnace color image data of the target RH refining furnace is obtained by an image acquisition device arranged at the bottom of the top gun of the target RH refining furnace.

[0012] In some embodiments, the RH refining furnace control method further comprises: detecting a working state parameter of the image acquisition device, wherein the working state parameter comprises frame rate stability and temperature sensing data; when the frame rate stability is lower than a preset stability threshold, a window cleaning device is activated to perform lens surface dust removal operation; and when the temperature sensing data is higher than a preset temperature threshold, the flow of cooling medium of a camera cooling device is increased and an optical component thermal expansion compensation program is triggered.

[0013] In some embodiments, the RH refining furnace control method further comprises: when it is detected that the proportion of the dust-shielded area in the continuous N frames of the furnace color image data is higher than a preset proportion threshold, generating, by a preset generative adversarial network model, predicted image data for the dust-shielded area; and fusing the current frame of initial furnace color image data with the predicted image data to obtain the current frame of the furnace color image data.

[0014] In a second aspect, the present application further provides an RH refining furnace control device, comprising: a data acquisition unit configured to acquire furnace color image data of a target RH refining furnace; a flame detection unit configured to identify a flame state of the furnace color image data by a first preset model to obtain a flame detection result; a downfeed recognition unit configured to recognize a material addition action of the furnace color image data by a second preset model to obtain downfeed event information; a state recognition unit configured to extract a metallurgical stage feature of the furnace color image data by a third preset model to obtain a refining state recognition result; and an interlock control unit configured to trigger an interlock control operation according to the flame detection result, the downfeed event information and the refining state recognition result, wherein the interlock control operation is configured to regulate fuel supply or equipment start-stop of the target RH refining furnace.

[0015] In a third aspect, the present application further provides an electronic device, comprising: a memory and a processor, wherein the processor is configured to implement the steps of the RH refining furnace control method of the first aspect when executing a computer program stored in the memory.

[0016] In a fourth aspect, the present application further provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the steps of the RH refining furnace control method of the first aspect.

[0017] In a fifth aspect, the present application further provides a computer program product, comprising a computer program or computer executable instructions, wherein the computer program or computer executable instructions are executed by a processor to implement the RH refining furnace control method provided in the embodiments of the present application.

[0018] To sum up, the application realizes automatic identification of the key working conditions of the RH refining furnace by processing the color image data in the furnace, using three preset models of flame state identification, blanking action identification and metallurgical stage feature extraction, breaks away from the traditional mode of relying on manual experience for judgment, and can improve the control accuracy of the process. The flame state identification can timely discover abnormal combustion or extinguishing and the like, automatically adjusts the fuel supply in combination with the interlocking control mechanism, and prevents safety accidents. At the same time, the automatic identification of the blanking event and the metallurgical state effectively reduces the influence of misoperation and equipment abnormalities, and can guarantee the stable operation of the RH refining furnace system. Through accurate control of the flame state and the blanking timing, unnecessary energy waste can be avoided, and the fuel utilization efficiency can be improved. At the same time, it ensures that the metallurgical reaction is carried out at the appropriate stage, which helps to improve product consistency and refining quality, and realizes the goal of reducing consumption and improving quality. To sum up, the RH refining furnace control method provided by the application realizes automatic judgment and interlocking control of the key working conditions of the RH refining furnace through multi-model image recognition, which not only can improve the control accuracy and system stability, but also can realize energy saving and consumption reduction and improve the refining quality. BRIEF DESCRIPTION OF DRAWINGS

[0019] 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 accompanying drawings are included to provide a description of the preferred embodiments and are not intended to limit the scope of the present description. Moreover, like reference numerals designate like parts throughout the several views in the drawings. In the drawings:

[0020] Figure 1 A flowchart of an RH refining furnace control method provided by an embodiment of the application;

[0021] Figure 2 A component structure schematic diagram of an RH refining furnace control device provided by an embodiment of the application;

[0022] Figure 3 A component structure schematic diagram of an electronic device provided by an embodiment of the application. DETAILED DESCRIPTION

[0023] The terms in the specification, claims and drawings of the present application, such as "first", "second", "third", "fourth" and the like (if any) are used to distinguish similar objects, not to describe a particular order or sequence. Therefore, it is understood that these terms can be used interchangeably, as appropriate, to achieve the described embodiments, unless otherwise specifically required by the drawings or description. In addition, the terms "is" and "has" and any variants thereof in the present application are intended to cover non-exclusive inclusion of all possible constituent elements. For example, a process, method, system, product or device including several steps or units does not necessarily limit to only the steps or units explicitly listed, but can also include other steps or units not explicitly listed, or steps or units inherent to the process, method, product or device.

[0024] In the present application, "module" or "unit" refers to a computer program or a part of a computer program with a specific function, and works with other related parts to achieve a predetermined target. These modules or units can be implemented by software, hardware (such as processing circuit or memory) or combination of both. One or more processors or memories can implement one or more modules or units. At the same time, each module or unit can also be part of a larger module or unit.

[0025] The technical solutions in the present application will be described in detail below in combination with the drawings in the embodiments. It should be noted that the described embodiments are only a part of the present application, not all embodiments. In the following description, "some embodiments" mentioned is only a subset of all possible embodiments, which can be the same or different subset, and different embodiments can be combined with each other without conflict.

[0026] Figure 1 is a flowchart of a RH refining furnace control method provided by an embodiment of the present application. For example, referring to Figure 1 The RH refining furnace control method provided by the embodiment of the present application can include the following steps 101 to 105:

[0027] Step 101, obtaining the color image data of the target RH refining furnace;

[0028] In some examples, the target RH refining furnace is a specific RH (Ruhrstahl-Heraeus) refining furnace that is currently being controlled or monitored and is in the position of being identified and controlled object; the RH refining furnace is a kind of vacuum refining equipment used for high-precision refining operations such as degassing, decarburization, alloying, etc. in the steel smelting process, which can include lifting pipe, vacuum tank, heating element and discharging system, etc., and is widely used in high-quality steel, alloy steel, stainless steel production, such as automobile steel, bearing steel, high-strength threaded steel, etc. The color image data in the furnace refers to the RGB color image data collected in the RH refining furnace by an industrial visible light camera, reflecting the current flame state in the furnace, molten steel surface activity, feeding action and other visual information, which can be collected by a high-temperature industrial camera installed on the top gun, furnace mouth or side wall, combined with cooling protection and dustproof window, and output image stream at a fixed frame rate;

[0029] For example, a color camera device with high-temperature resistance and strong light suppression function can be arranged below the top gun of the RH refining furnace, and continuous frame data is transmitted to the image processing module in real time through an industrial image acquisition card; first, the target furnace number and the corresponding camera channel are identified, and the corresponding device is automatically called to obtain the image stream; the collected image is used as the model input, and is used for extracting key visual features such as flame, material and molten steel surface in the subsequent process, to provide visual basis for the whole intelligent identification and control link.

[0030] Through the implementation of step 101, the color image data in the RH refining furnace is collected in real time, which can fully reflect the working condition changes of the flame, material and metallurgical reaction in the furnace, and can provide high-quality and dynamic visual information source for subsequent intelligent identification.

[0031] In step 102, the flame state in the furnace is identified by a first preset model to obtain a flame detection result.

[0032] In some examples, the first preset model is a pre-trained image recognition model specially used for identifying the flame state in the RH refining furnace; the first preset model can be based on a deep convolutional neural network structure, and can fuse boundary detection, time series change analysis and feature fusion modules, and has the ability to automatically identify the existence, stability and intensity of the flame. Flame state identification is to determine whether there is a flame in the current furnace, whether the flame is stable, and how strong the flame is, etc., to reflect whether the combustion is normal. The flame detection result is the structured identification information output by the first preset model, which can include the spatial position (bounding box coordinates) of the flame, the change characteristics (size change rate, position offset fluctuation) in the time dimension and the state classification label (such as “stable combustion”, “flame weakening”, “extinguishing”) etc.

[0033] Through implementation of step 102, the flame state in the image is identified by using the first preset model, whether the combustion is normal can be quickly judged, and problems such as abnormal combustion, unstable flame or flameout can be found in time, real-time feedback can be provided for the fuel system, and production safety can be ensured and energy utilization efficiency can be improved.

[0034] In step 103, the material adding action in the furnace is recognized by the second preset model to obtain the downloading event information.

[0035] In some examples, the second preset model is a pre-trained image recognition model for identifying the material adding (downloading) action in the RH refining furnace, which has the ability to extract the spatiotemporal dynamic features in the color image data in the furnace and judge the downloading behavior. The material adding action recognition refers to identifying whether the behavior of putting material into the furnace occurs in the operation by the second preset model, and judging the flow path of the material, the source equipment, the start and end time of downloading, etc.; for example, whether the material is falling, which way is the material put by, such as human operation, mechanical arm, screw feeder, etc., whether the trajectory direction of the material is aligned with the furnace mouth, whether the action conforms to the process rhythm or there is an abnormality. The downloading event information is the structured data output by the second preset model, which is used to describe the key features of one downloading behavior, and can include: the start time and the end time of downloading (such as represented by frame number or timestamp), the material type classification result (such as aluminum particles, lime, silicon-manganese alloy, etc.), the downloading action identification (whether it is valid / abnormal downloading), the action intensity / frequency analysis (such as the estimated index of the amount of one-time downloading).

[0036] Through implementation of step 103, the material adding behavior is identified by using the second preset model, each downloading event can be accurately recorded, and operation errors caused by manual missed judgment or delayed recognition can be avoided, which helps to optimize the feeding rhythm and ensure the continuity and consistency of the metallurgical process.

[0037] In step 104, the metallurgical stage features are extracted from the color image data in the furnace by the third preset model to obtain the refining state recognition result.

[0038] In some examples, the third preset model is an image intelligent analysis model for identifying different metallurgical stages of the RH refining process; the third preset model integrates spatiotemporal image feature analysis, sequence evolution judgment, and stage classification mechanism, and can accurately judge the current metallurgical stage in the furnace. The metallurgical stage feature extraction refers to identifying key physical phenomenon features that can reflect the current process progress from the color image data in the furnace. The visual differences of the RH refining furnace in different stages include but are not limited to the molten steel turbulent flow characteristics such as bubble rolling intensity, color brightening degree, liquid surface disturbance frequency, the slag surface morphology characteristics such as whether there is white scum, molten slag color change, dynamic change of slag layer thickness, and the flow mode characteristics (such as the rhythm and regional distribution of the flow in the furnace during vacuum rising and alloy stirring process). The refining state recognition result refers to the metallurgical stage label information output by the third preset model after analyzing the image, which can include: the specific refining stage (such as “decarburization”, “deoxidation”, “alloying”) currently in, the time node of stage conversion, and the analysis result of stage stability or whether there is process abnormality.

[0039] Through the implementation of step 104, the metallurgical stage features are identified by the third preset model, which can automatically judge the process stage of the refining process, support dynamic optimization control of key process parameters, and thus improve the product quality stability and the precision of process execution.

[0040] In step 105, an interlock control operation is triggered according to the flame detection result, the charging event information, and the refining state recognition result, wherein the interlock control operation is used to regulate the fuel supply or the start-stop of the equipment of the target RH refining furnace.

[0041] In some examples, the interlock control operation refers to a safety protection and logic control mechanism automatically triggered based on the multi-source perception results of the flame state, the charging behavior, and the metallurgical stage in the automatic control process of the RH refining furnace; when one or more state abnormalities, conflicts, or non-compliance with the preset logic requirements are detected, the corresponding control instructions are immediately executed to prevent equipment damage, process abnormalities, or safety accidents; the control actions can be implemented on the key running units of the RH refining furnace such as the burner, the vacuum pump, the stirring system, and the charging system.

[0042] For example, in the RH refining process, if it is detected that the current is in the ignition stage but the flame is not successfully ignited, the interlock control module will immediately close the gas and oxygen supply and issue a warning sound to avoid the risk of deflagration caused by gas accumulation in the furnace; if it is identified that the charging behavior occurs in the state that the deoxidation stage has not been completed, the charging equipment will be automatically stopped, and the operator will be notified to check, thereby effectively preventing alloy waste or chemical reaction out of control and ensuring the continuity and safety of the entire refining process.

[0043] Through the implementation of step 105, the above identification results are comprehensively applied to the interlocking control mechanism, so that automatic adjustment of fuel supply and equipment start-stop can be realized, manual intervention is reduced, and in turn the response speed of the method is improved, production safety is ensured, and operation efficiency is improved.

[0044] To sum up, the embodiment of the present application realizes automatic identification of the key working conditions of the RH refining furnace by processing the in-furnace color image data and using three preset models of flame state identification, blanking action identification, and metallurgical stage feature extraction, breaks away from the traditional mode of relying on manual experience for judgment, and can improve the control accuracy of the process. The flame state identification can timely discover abnormal combustion or extinguishing, and in combination with the interlocking control mechanism, automatically adjusts the fuel supply to prevent safety accidents. At the same time, by automatically identifying the blanking event and the metallurgical state, the influence of misoperation and equipment abnormalities is effectively reduced, and the stable operation of the RH refining furnace system is ensured. By accurately controlling the flame state and the blanking timing, unnecessary energy waste can be avoided, and the fuel utilization efficiency is improved. At the same time, it ensures that the metallurgical reaction is carried out at the appropriate stage, which helps to improve product consistency and refining quality and achieve the goal of reducing consumption and improving quality. To sum up, the RH refining furnace control method provided by the embodiment of the present application realizes automatic judgment and interlocking control of the key working conditions of the RH refining furnace through multi-model image recognition, which not only can improve the control accuracy and system stability, but also can realize energy saving and consumption reduction and improve the refining quality.

[0045] In some embodiments, the first preset model comprises a boundary detection network unit, a first feature extraction unit, a second feature extraction unit, and a data fusion unit. Step 102 can include: performing target detection on the single-frame in-furnace color image data by the boundary detection network unit to obtain a flame position bounding box, wherein the boundary detection network unit is a convolutional neural network of a YOLO architecture; extracting a size change rate of the flame position bounding box in the continuous frame in-furnace color image data by the first feature extraction unit; extracting a center point displacement fluctuation value of the flame position bounding box in the continuous frame in-furnace color image data by the second feature extraction unit; and determining the flame detection result based on the size change rate and the center point displacement fluctuation value by the data fusion unit.

[0046] In some examples, the boundary detection network unit is used to identify the position, shape and range of the flame from the in-furnace image, output a rectangular or irregular boundary box, accurately separate the flame area from the background, and provide subsequent analysis. The flame target can be identified on the labeled in-furnace image by training a deep learning model. The convolutional neural network of the YOLO (You Only Look Once) architecture is a high-efficiency target detection network structure that can quickly and in parallel identify multiple targets and their positions in an image. The flame position boundary box is a rectangular area output by the boundary detection network unit, which is used to mark the specific coordinates (such as the center point position, width, height, etc.) of the flame position in the image, providing a spatial reference for subsequent motion and state analysis. For example, a flame position boundary box may be labeled as a center position (240, 160), a width of 100 pixels, and a height of 80 pixels. The first feature extraction unit is used to analyze the size of the boundary box in multiple consecutive frames and extract the volume change trend of the flame over time, such as shrinkage or expansion, which can reflect the activity level of the flame or whether there is an unstable phenomenon. The size change rate is the relative change rate of the size of the flame position boundary box between consecutive image frames, indicating the dynamic amplitude of the expansion or contraction of the flame area. For example, if the size of the flame position boundary box increases from 100 pixels to 120 pixels, the growth rate is 20%. The second feature extraction unit is used to extract the motion trajectory features of the center point of the flame boundary box in consecutive frames to determine whether the flame is stable or has a rapid shift. The center point displacement fluctuation value is a statistical quantity of the displacement difference of the center of the flame boundary box between consecutive frames, which is used to measure the disturbance degree or directional drift of the flame in space. For example, if the center point moves from (200, 150) to (220, 170) in 10 frames, the average displacement and variance are calculated as the center point displacement fluctuation value. The data fusion unit can integrate and analyze multiple features such as size change rate and center displacement, and output the determination result of whether it is an abnormal flame state. For example, if the size change rate is large and the center point fluctuates violently, the data fusion unit determines that the flame is in a "violent fluctuation" state.

[0047] For example, continuous color images of the in-furnace are collected, and the position and size information of the flame boundary box of each frame is extracted in real time by the boundary detection network unit. The first feature extraction unit analyzes the change trend of the flame size over time, and the second feature extraction unit records the motion path of the flame center point in space. After these data are input into the data fusion unit, if the flame area is significantly reduced and the flame is significantly shifted, it can be determined that the flame has been extinguished or is burning abnormally, and the interlocking control mechanism is triggered to automatically adjust the fuel or send an alarm signal, thereby improving the operation safety and automation level of the RH refining furnace.

[0048] Through the implementation of the above embodiments, the YOLO bounding box detection is combined with the time sequence features of the flame size change rate and the centroid displacement fluctuation, so that the three typical states of stable combustion, unstable jitter and extinguishing can be accurately distinguished under the interference of molten steel turbulence, light flicker or background light emission. This fine-grained discrimination enables the combustion system to reduce the lance flow rate in advance or immediately cut off the fuel, reduces the probability of false shutdown, prevents deflagration, and prolongs the service life of the top lance nozzle.

[0049] In some embodiments, the second preset model consists of a first spatio-temporal feature extraction network unit, a preset event classifier, and a preset material classifier. The unloading event information can include unloading event start and end time points and material type classification results. The step 103 can include: extracting dynamic target trajectory features in the color image data in the furnace by the first spatio-temporal feature extraction network unit, wherein the dynamic target trajectory features can include material flow trajectory and equipment movement features; obtaining the unloading event start and end time points by the preset material classifier based on the material flow trajectory and the equipment movement features; and obtaining the material type classification results by the preset event classifier based on the image region visual features corresponding to the unloading event start and end time points.

[0050] In some examples, the first spatio-temporal feature extraction network unit is based on the structure of a deep learning model such as a 3D convolutional network, ConvLSTM, etc., for extracting image motion features changing over time from the in-furnace color image data; for example, the motion trajectory of the material falling from the high-position into the furnace, the moving path of the cantilever device, etc. can be detected by the first spatio-temporal feature extraction network unit. The dynamic target trajectory feature is a feature of the continuous position and morphological change presented by the continuously changing objects (such as the material, the mechanical arm) in the in-furnace color image data, including trajectory, speed, direction, etc.; the target path extracted in the continuous frames by target detection and tracking algorithms such as Kalman filtering, optical flow tracking, etc.; for example, the dynamic target trajectory feature can be the parabolic trajectory formed by the falling of a batch of materials, the mechanical arm trajectory of the swinging of the vibrating hopper, etc. The material flow trajectory is the visual feature of the motion path of the material such as deoxidizer, alloy powder, etc. during the process of entering the furnace from the material port; for example, the falling line of the white powdery material falling from a high position and falling into the surface of the molten steel. The equipment motion feature is the image change feature of the material-related equipment such as the electric hoist, the screw conveyor, the slide, etc. in the motion process, for example, the swinging direction of the mechanical arm, the opening and closing amplitude of the hopper, etc. The preset material classifier is used to identify and determine the start and end time of the observed material falling behavior in the image, analyze the dynamic target trajectory and the equipment action, so as to accurately demarcate the time interval of the material falling event; for example, if a continuous frame of rapidly moving particle flow appears from top to bottom, it is marked as “start of material falling”, and if the target disappears or the equipment is closed, it is marked as “end of material falling”. The preset event classifier is used to analyze the visual features appearing in the image during the material falling, and classify them into specific material types (such as aluminum particles, alloy powder, lime, etc.), to assist subsequent metallurgical control; for example, a yellow fine particle flow is classified as lime powder; a silver-white lump is identified as an aluminum particle. The start and end time points of the material falling event refer to the time stamps corresponding to the start frame and end frame of a material adding operation. These two time points define the whole process time of this material falling behavior, which can be output by the frame number or time stamp of the preset material classifier; for example, an alloy adding event starts from the 320th frame of the image and ends at the 415th frame. The material type classification result represents the material category corresponding to the identified material falling behavior, which is the output result of the event classifier after analyzing the visual features of the image region, such as “lime”, “silicon-manganese alloy”, “aluminum particle”, etc.

[0051] Through the implementation of the above embodiments, the trajectory features of the material flow and equipment motion are extracted by the first spatio-temporal feature extraction network unit, and the start and end of the material falling and the category are determined by the event and material dual classifier, which can still accurately lock when and what is added when the RH refining furnace material channel is blocked by dust or the light changes drastically, reducing the number of excessive material addition and secondary material addition, and saving alloy cost.

[0052] In some embodiments, the third preset model is composed of a second spatio-temporal feature extraction network unit, a recurrent neural network unit, and a preset metallurgical stage classifier. Step 104 can include: performing multi-scale spatial feature extraction on the in-furnace color image data by the second spatio-temporal feature extraction network unit to obtain molten steel turbulent flow features and slag surface morphology features; performing time series evolution analysis processing on the molten steel turbulent flow features and the slag surface morphology features by the recurrent neural network unit to obtain a feature evolution vector; and analyzing the feature evolution vector by the preset metallurgical stage classifier to obtain a refining state recognition result, wherein the refining state recognition result can include a decarburization stage identifier, a deoxidization stage identifier, and an alloying stage identifier.

[0053] In some examples, the second spatio-temporal feature extraction network unit is a deep learning module for simultaneously extracting spatial features such as local texture, shape, and temporal features such as dynamic changes, flow trends from the in-furnace image sequence. For example, it can identify the intensity of molten steel rolling, the trend of surface color texture change over time, etc. The molten steel turbulent flow features refer to the dynamic motion characteristics of high-temperature liquid steel, such as surface rolling, gushing, vortex, etc., exhibited during the RH refining furnace treatment process, which are important indicators for judging gas stirring effect and reaction intensity. The slag surface morphology features refer to the visual features of the slag layer covering the surface of the molten steel in the furnace, such as color, shape, thickness, or fragmentation degree, etc., which can reflect the progress and stage of metallurgical reactions. The distribution area and texture pattern can be extracted based on image color distribution and edge structure; for example, the deoxidization stage is usually accompanied by white and thickening of the slag layer, and the reaction is quiet when it presents a mirror surface. The recurrent neural network unit is responsible for analyzing the trend of the evolution of molten steel turbulent flow and slag surface morphology over time, capturing the phase transition points and dynamic feature sequences, and is an important structure for time series modeling; the recurrent neural network unit can use LSTM, GRU, etc. Recurrent neural network for modeling the extracted multi-frame image feature vectors. The feature evolution vector is a series of high-dimensional time series feature vectors output by the recurrent neural network unit, representing the evolution trajectory of the in-furnace state in the continuous time period, and is used for classification analysis. The preset metallurgical stage classifier classifies and identifies the current metallurgical treatment stage based on the input feature evolution vector, determines whether the current is in the decarburization, deoxidization, or alloying stage; a supervised classification model such as SVM, fully connected neural network, etc. can be trained as a preset metallurgical stage classifier to map the metallurgical process stage; for example, the evolution vector pattern matches “medium-high turbulent flow + thin slag”, and outputs “decarburization stage identifier”. The decarburization stage identifier indicates that the current is in the reaction stage of converting carbon elements in the molten steel to CO by strong oxygen blowing, and the characteristics are intense turbulent flow and bubble gushing; the deoxidization stage identifier indicates that the current steel mainly undergoes oxygen removal reactions, the turbulent flow is weakened, the reaction is more stable, and aluminum, silicon, and other deoxidizers are often added; the alloying stage identifier indicates that the molten steel starts to add alloying elements such as Mn, Cr, Ni, etc. to meet the composition requirements, which can be performed after the reaction tends to be stable.

[0054] Exemplarily, a continuous image sequence can be acquired from the RH refining furnace top camera, input into the second spatiotemporal feature extraction network to extract spatial morphology features, and then input into the RNN unit for time evolution analysis. If the model identifies that the molten steel gradually changes from violent gushing to stable flow, and the slag surface color gradually becomes white, the "deoxidation stage" label will be output. Subsequently, the corresponding feeding program can be automatically switched to avoid the mis-mixing of decarburizer and deoxidizer, and to ensure process safety and composition accuracy.

[0055] Through the implementation of the above embodiments, the multi-scale spatial features are used to capture the molten steel turbulence and slag surface texture, and the recurrent network is used for time evolution analysis, so that the small flow changes in high-noise video can be converted into clear stage labels such as decarburization, deoxidation, and alloying. The vacuum degree, top and bottom blowing flow, and alloy feeding window can be automatically switched to compress the minute-level error of artificial judgment to seconds, and to ensure the temperature and composition hit rate.

[0056] In some embodiments, step 105 can include: when the flame detection result indicates that the flame is not ignited in the ignition stage, stopping the fuel supply and ignition operation of the target RH refining furnace, and issuing first warning information; when the flame detection result indicates that the flame is extinguished, stopping the oxygen supply and discharging operation of the target RH refining furnace; when the start and end time points of the discharging event, the material type classification result, and the refining state recognition result do not meet the requirements of the preset metallurgical path, stopping the discharging operation of the target RH refining furnace, and issuing second warning information; and when the refining state recognition result indicates the end of the first stage, generating an operation instruction for the second stage, wherein the first stage and the second stage are process stages performed in sequence.

[0057] In some examples, the ignition stage refers to an operation stage in the RH refining furnace start-up process for igniting fuel and establishing an initial flame, which is the starting link of the entire refining cycle; the ignition stage can be identified by the running state or operation instruction of the furnace control system, and is usually automatically entered within a short time after the furnace is started. If the flame detection result does not identify an effective combustion area, i.e., indicates that the flame is not ignited, it means that the ignition attempt has failed or the flame has not been successfully maintained; for example, there is no flame bounding box or the box area tends to zero in 10 consecutive frames of images. When the flame detection result indicates that the flame is not ignited in the ignition stage, the fuel valve and the ignition system power supply can be automatically closed to prevent the risk of fuel accumulation or explosion. The first warning information is an alarm prompt automatically generated after ignition failure, which is used to notify the operator and record the abnormality. The flame detection result indicates that the flame is extinguished, indicating that an abnormal state of sudden disappearance or rapid weakening of the flame is detected in the normal combustion process; for example, the flame area is reduced by more than 90% in consecutive frames and there is no recovery trend. When the flame detection result indicates that the flame is extinguished, the oxygen valve and the material feeding controller are automatically closed to prevent the accumulation of oxygen and combustible materials; for example, the main oxygen pipeline electromagnetic valve is closed within 2 seconds after the combustion is interrupted. The pre-set metallurgical path requirement is a reasonable metallurgical process set based on the process recipe and control strategy, including the order, time period and state condition of material addition; for example, it is stipulated that no deoxidizer or alloy should be added in the decarburization stage. The second warning information is an alarm prompt generated when the metallurgical path is destroyed or the process is abnormal; for example, the prompt “alloy feeding time is abnormal, please check the refining state”. The indication of the end of the first stage is a state signal output by the refining state recognition model, indicating that the current process stage has reached the completion condition. The operation instruction for the second stage is an automatically generated execution instruction for the next process step, such as starting the oxygen lance and adding deoxidizer, to ensure process continuity. In the RH refining process, the first stage can be decarburization (high oxygen and high temperature), and the second stage can be deoxidization or alloying (adding deoxidizer and metal alloy), which are performed in sequence.

[0058] For example, in an automatic refining process, the flame image is continuously monitored in the ignition stage, and if no effective flame area is identified, the fuel and ignition operation are stopped immediately, and an alarm prompt is popped up to the operation terminal; when the flame is unexpectedly extinguished during refining, the oxygen and material supply are quickly cut off to prevent safety hazards. At the same time, if the operator mistakenly adds alloy material before the decarburization is completed, the identified feeding time and stage identifier can be compared to immediately interrupt the feeding and issue a second warning information; after the model determines that the decarburization is completed, the next stage operation strategy is automatically pushed, which improves the automation level and ensures the process continuity.

[0059] Through the implementation of the above embodiment, the three signals of flame detection, blanking event and refining state are mapped into multi-level safety interlocking, ignition failure stops burning, flameout stops oxygen, and blanking event and refining state mismatch forces locking, which can maximize the avoidance of accidents such as air-fuel ratio imbalance burning loss of furnace tube and secondary refining caused by misadding alloy; at the same time, the next stage process instruction is automatically generated, combining safety protection and beat optimization into one, and shortening the smelting cycle.

[0060] In some embodiments, the foregoing step 101 can include: acquiring the color image data in the furnace by an image acquisition device arranged at the bottom of the top gun of the target RH refining furnace.

[0061] In some examples, the image acquisition device is an industrial-level vision device installed below the top lifting device (i.e. the top gun) of the RH refining furnace, which can be used for continuous or timed shooting towards the interior of the hearth to collect color image data in the furnace during the refining process; due to the high temperature and complex environment in the RH refining furnace, the image acquisition device needs to have high-temperature heat insulation, dustproof, shockproof, anti-glare and other industrial protection performance. The image acquisition device is fixed at the bottom of the central axis area of the top gun (usually the outer cover of the lifting pipe or the exhaust device), and shoots downward towards the center of the hearth through a narrow window to capture key areas such as furnace flame, molten steel fluctuation and blanking behavior.

[0062] Through the implementation of the above embodiment, the high-temperature-resistant color camera is installed at the bottom of the top gun, sharing the interface with the original cooling channel, without the need to open a viewing window in the furnace wall; the top-down angle can cover the flame, the flow and the molten steel surface at the same time, without moving the lens throughout the process, reducing the transformation cost and maintenance downtime, and avoiding the leakage risk caused by additional perforation in the high-temperature area.

[0063] In some embodiments, the foregoing RH refining furnace control method can further include: detecting working state parameters of the image acquisition device, wherein the working state parameters can include frame rate stability and temperature sensing data; when the frame rate stability is lower than a preset stability threshold, activating the window cleaning device to perform lens surface dust removal operation; when the temperature sensing data is higher than a preset temperature threshold, increasing the flow of the cooling medium of the camera cooling device and triggering the optical component thermal expansion compensation program.

[0064] In some examples, the working state parameter is various types of real-time monitoring data reflecting the running status of the image acquisition device, which is used to determine whether the acquisition device is in a good working state and can be automatically collected through embedded sensors, software timers, thermal elements, etc. Frame rate stability represents the stability of the number of image frames output by the image acquisition device per unit time, which is used to evaluate whether the image stream is continuous and available. The time interval between consecutive frames can be recorded, and the frame rate fluctuation coefficient can be calculated. For example, if the expected frame rate is 25 FPS, and the fluctuation is frequent to 10-25 FPS, it is an unstable state. Temperature sensing data is real-time temperature information of the image acquisition device and its peripheral components, which is used to determine whether it is overheating. It can be collected by an infrared thermocouple or a thermal IC module at regular intervals. The preset stability threshold refers to the minimum frame rate fluctuation stability index allowed. Exceeding this threshold means that the image quality is unacceptable. The window cleaning device is an automatic cleaning device installed on the image acquisition device shell or the furnace wall window, which is used to remove obstructions such as dust, slag, or water vapor. The window cleaning device can be in the form of a mechanical arm, a jet device, an ultrasonic vibrator, etc. The lens surface dust removal operation is to use the window cleaning device to remove dust, coking, etc. on the front end lens or glass window of the image acquisition device to ensure clear images. The preset temperature threshold is the upper limit value set to ensure the stable operation of the image acquisition device. Exceeding this threshold will start the cooling and protection measures. The camera cooling device is an active cooling system installed to prevent the camera from overheating, which can use liquid cooling, air cooling, or heat pipe heat conduction, etc. Cooling medium flow rate is the flow rate of cooling medium such as water, air, or refrigerant in the camera cooling device, which directly affects the cooling efficiency. The optical component thermal expansion compensation program is an algorithm module used to offset the imaging distortion or blur caused by the micro-deformation of the lens, lens, or support due to high temperature. For example, after detecting the temperature rise in front of the lens, the focal length parameter is automatically recalibrated.

[0065] Through the implementation of the above embodiments, the frame rate fluctuation and the shell temperature are monitored in real time, the air knife dust removal and cooling flow adjustment are automatically triggered, the clear imaging can be maintained and the lens burst can be prevented when the dust suddenly increases or the cooling water pressure changes during the vacuum stage of the RH refining furnace, ensuring the long-term stable operation of the image acquisition system in high-temperature harsh environments, and reducing false alarms caused by blocked vision.

[0066] In some embodiments, the foregoing RH refining furnace control method can further include: when it is detected that the dust shielding area proportion in the continuous N frames of furnace color image data is higher than a preset proportion threshold, generating predicted image data for the dust shielding area through a preset generative adversarial network model; and performing data fusion on the current frame of initial furnace color image data and the predicted image data to obtain the current frame of furnace color image data.

[0067] In some examples, N represents a preset positive integer parameter for specifying the number of consecutive image frames, to control the time range or frame number window for judging the dust shielding degree. The value of N can be set by engineers according to the operation rhythm of the steelmaking furnace, the image frame rate, and the dust fluctuation characteristics. For example, if the camera frame rate is 10 fps and N = 10, the image stability is monitored for about 1 second. Reasonable selection of N helps to balance the sensitivity and robustness of the system. If N is too small, the system may be too sensitive to occasional shielding, resulting in frequent triggering of the completion and waste of resources. If N is too large, the system may be too slow to respond, missing the key opportunity for obstruction repair. The dust shielding area ratio is the pixel ratio of the area in the entire image that is visually blurred and dark in color due to the shielding of particles such as smoke and steam in the furnace color image data. An image segmentation model (such as U-Net or Mask R-CNN) can be used to identify and calculate the area ratio of the dust shielding area in the image. For example, if the dust shielding area in a certain frame of image is 80%, it means that most of the details in the image are not visible. The preset coverage threshold is the upper limit of the shielding coverage for judging the image availability or clarity. If it exceeds the threshold, the image is considered invalid. The preset generative adversarial network model is a deep learning model for image generation and restoration, which is trained by a generator and a discriminator. It can infer the content of the obscured area based on historical images or context. The preset generative adversarial network model can be trained on a large dataset of furnace images to learn the structure and color rules under shielding conditions. The predicted image data is the image completion result output by the preset generative adversarial network model after inferring the content of the obscured area. It can be local area image data. The initial furnace color image data of the current frame refers to a certain frame of furnace image data collected by the original camera without processing. It may contain shielding, blurring, or noise. The image content predicted by the preset generative adversarial network model can be fused with the initial furnace color image data of the current frame at the pixel level or regional weight to generate a new image frame that is de-shielded and complete in information. Image fusion algorithms such as alpha blending, edge smoothing splicing, and mask covering can be used to synthesize a clear and continuous image.

[0068] Through the implementation of the above embodiments, after detecting continuous multiple frames of dust shielding, the obscured area is restored in real time using a generative adversarial network, and then fused with the original image to maintain the recognition link uninterrupted during the gray screen generated by high-vacuum slagging or alloy injection. Compared with simple frame dropping or threshold filtering, the preset generative adversarial network model reconstruction ensures that key features such as flames and material flows are continuously visible, enabling no-blind-area monitoring in a shielding scenario, and further improving the reliability and accuracy of the RH refining furnace control method.

[0069] Furthermore, as an implementation of the aforementioned method embodiments, this application also provides an RH refining furnace control device for implementing the aforementioned method embodiments. This device embodiment corresponds to the aforementioned method embodiments. For ease of reading, this RH refining furnace control device embodiment will not repeat the details of the aforementioned method embodiments one by one, but it should be understood that the device in this application embodiment can correspondingly implement all the contents of the aforementioned method embodiments. For example... Figure 2 As shown, the RH refining furnace control device 20 includes: a data acquisition unit 201, a flame detection unit 202, a material feeding identification unit 203, a status identification unit 204, and an interlock control unit 205. The data acquisition unit 201 acquires color image data of the target RH refining furnace. The flame detection unit 202 performs flame status identification on the color image data using a first preset model to obtain flame detection results. The material feeding identification unit 203 performs material feeding action identification on the color image data using a second preset model to obtain material feeding event information. The status identification unit 204 extracts metallurgical stage features from the color image data using a third preset model to obtain refining status identification results. The interlock control unit 205 triggers interlock control operations based on the flame detection results, material feeding event information, and refining status identification results. These interlock control operations are used to regulate the fuel supply or equipment start-up and shutdown of the target RH refining furnace.

[0070] In some embodiments, the first preset model comprises a boundary detection network unit, a first feature extraction unit, a second feature extraction unit, and a data fusion unit; the flame detection unit 202 is further configured to perform target detection on a single frame of furnace-in-color image data through the boundary detection network unit to obtain a flame position bounding box, wherein the boundary detection network unit is a convolutional neural network with a YOLO architecture; extract the size change rate of the flame position bounding box in the continuous frame furnace-in-color image data through the first feature extraction unit; extract the center point displacement fluctuation value of the flame position bounding box in the continuous frame furnace-in-color image data through the second feature extraction unit; and determine the flame detection result through the data fusion unit based on the size change rate and the center point displacement fluctuation value.

[0071] In some embodiments, the second preset model is composed of the first spatio-temporal feature extraction network unit, a preset event classifier and a preset material classifier, the blanking event information includes blanking event start and end time points and material type classification results, and the blanking recognition unit 203 is further configured to extract dynamic target trajectory features in the color image data in the furnace through the first spatio-temporal feature extraction network unit, wherein the dynamic target trajectory features include material flow trajectory and equipment movement features; based on the material flow trajectory and the equipment movement features, the preset material classifier is used to identify the blanking event start and end time points; based on the image region visual features corresponding to the blanking event start and end time points, the preset event classifier is used to identify the material type classification results.

[0072] In some embodiments, the third preset model is composed of a second spatio-temporal feature extraction network unit, a recurrent neural network unit and a preset metallurgical stage classifier, and the state recognition unit 204 is further configured to perform multi-scale spatial feature extraction on the color image data in the furnace through the second spatio-temporal feature extraction network unit to obtain molten steel turbulent flow features and slag surface morphology features; perform time series evolution analysis and processing on the molten steel turbulent flow features and the slag surface morphology features through the recurrent neural network unit to obtain a feature evolution vector; and analyze the feature evolution vector through the preset metallurgical stage classifier to obtain a refining state recognition result, wherein the refining state recognition result includes decarburization stage identification, deoxidation stage identification and alloying stage identification.

[0073] In some embodiments, the interlock control unit 205 is further configured to, when the flame detection result indicates that the flame is not ignited in the ignition stage, stop the fuel supply and the ignition operation of the target RH refining furnace and issue a first warning information; when the flame detection result indicates that the flame is extinguished, stop the oxygen supply and the blanking operation of the target RH refining furnace; when the blanking event start and end time points, the material type classification results and the refining state recognition result do not meet the preset metallurgical path requirements, stop the blanking operation of the target RH refining furnace and issue a second warning information; and when the refining state recognition result indicates the end of the first stage, generate an operation instruction for the second stage, wherein the first stage and the second stage are process stages performed in sequence.

[0074] In some embodiments, the data acquisition unit 201 is further configured to acquire the color image data in the furnace through an image acquisition device arranged at the bottom of the top gun of the target RH refining furnace.

[0075] In some embodiments, the data acquisition unit 201 is further configured to detect working state parameters of the image acquisition device, wherein the working state parameters include frame rate stability and temperature sensing data; when the frame rate stability is lower than a preset stability threshold, activate the window cleaning device to perform lens surface dust removal operation; and when the temperature sensing data is higher than a preset temperature threshold, increase the flow of cooling medium of the camera cooling device and trigger the optical component thermal expansion compensation program.

[0076] In some embodiments, the data acquisition unit 201 is further configured to, when detecting that the proportion of the dust-shielded area in the continuous N frames of in-furnace color image data is higher than a preset proportion threshold, generate predicted image data for the dust-shielded area by using a preset generative adversarial network model; and perform data fusion on the current frame of initial in-furnace color image data and the predicted image data to obtain the current frame of in-furnace color image data.

[0077] The present application also provides a computer-readable storage medium, which stores computer-executable instructions or computer programs. When the computer-executable instructions or computer programs are executed by a processor, the processor is caused to perform any step of the RH refining furnace control method provided by the present application.

[0078] In some embodiments, the computer-readable storage medium can be a random access memory (RAM), a read-only memory (ROM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM), etc. The computer-readable storage medium can also be various devices including one or any combination of the above storage devices.

[0079] In some embodiments, the computer-executable instructions can be in the form of programs, software, software modules, scripts or codes, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and can be deployed in any form, including being deployed as independent programs or being deployed as modules, components, subroutines or other units suitable for use in a computing environment.

[0080] In some embodiments, the computer-executable instructions can but not necessarily correspond to files in a file system, can be stored in a part of a file storing other programs or data, for example, stored in one or more scripts in a HyperText Markup Language (HTML) document, stored in a single file dedicated to the program in question, or stored in multiple cooperating files (for example, files storing one or more modules, subroutines or code portions).

[0081] In some embodiments, the computer-executable instructions can be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed at multiple sites and interconnected through a communication network.

[0082] As Figure 3As shown, the present application also provides an electronic device 30, comprising a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor, wherein the processor 320 implements any step of the RH refining furnace control method described above when executing the computer program 311.

[0083] The present application also provides a computer program product, comprising a computer program or computer executable instructions stored in a computer readable storage medium. The processor of the electronic device reads the computer program or computer executable instructions from the computer readable storage medium, and the processor executes the computer program or computer executable instructions, so that the electronic device executes any step of the RH refining furnace control method described above.

[0084] The above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for controlling an RH refining furnace, characterized in that, include: Acquire color image data of the inside of the target RH refining furnace; The flame state is identified by the color image data inside the furnace using a first preset model, and the flame detection result is obtained. The material addition action is identified by the second preset model on the color image data inside the furnace to obtain material feeding event information. The metallurgical stage features are extracted from the color image data inside the furnace using a third preset model to obtain the refining state identification result. Based on the flame detection results, the feeding event information, and the refining status identification results, an interlock control operation is triggered, wherein the interlock control operation is used to regulate the fuel supply or equipment start-up and shutdown of the target RH refining furnace.

2. The RH refining furnace control method according to claim 1, characterized in that, The first preset model consists of a boundary detection network unit, a first feature extraction unit, a second feature extraction unit, and a data fusion unit; the step of performing flame state recognition on the furnace color image data using the first preset model to obtain flame detection results includes: The boundary detection network unit performs target detection on a single frame of the furnace color image data to obtain the flame position bounding box, wherein the boundary detection network unit is a convolutional neural network with YOLO architecture; The first feature extraction unit extracts the size change rate of the flame position bounding box in the continuous frames of the furnace color image data; The displacement fluctuation value of the center point of the flame position bounding box in the continuous frames of the furnace color image data is extracted by the second feature extraction unit. Based on the dimensional change rate and the center point displacement fluctuation value, the flame detection result is determined by the data fusion unit.

3. The RH refining furnace control method according to claim 2, characterized in that, The second preset model consists of a first spatiotemporal feature extraction network unit, a preset event classifier, and a preset material classifier. The material feeding event information includes the start and end times of the material feeding event and the material type classification result. The material feeding event information is obtained by recognizing the material adding action through the furnace color image data using the second preset model, including: The dynamic target trajectory features in the furnace color image data are extracted by the first spatiotemporal feature extraction network unit, wherein the dynamic target trajectory features include material flow trajectory and equipment motion features; Based on the material flow trajectory and the equipment motion characteristics, the start and end times of the feeding event are identified by the preset material classifier. Based on the visual features of the image region corresponding to the start and end times of the material feeding event, the material type classification result is obtained by the preset event classifier.

4. The RH refining furnace control method according to claim 3, characterized in that, The third preset model consists of a second spatiotemporal feature extraction network unit, a recurrent neural network unit, and a preset metallurgical stage classifier. The step of extracting metallurgical stage features from the furnace-in-color image data using the third preset model to obtain refining state identification results includes: The second spatiotemporal feature extraction network unit performs multi-scale spatial feature extraction on the color image data inside the furnace to obtain the turbulent features of molten steel and the morphological features of slag surface. The recurrent neural network unit is used to perform time-series evolution analysis on the turbulent characteristics of molten steel and the morphological characteristics of slag surface to obtain the feature evolution vector. The feature evolution vector is analyzed by the preset metallurgical stage classifier to obtain the refining state identification result, wherein the refining state identification result includes the decarburization stage identifier, the deoxidation stage identifier, and the alloying stage identifier.

5. The RH refining furnace control method according to claim 4, characterized in that, Based on the flame detection results, the material feeding event information, and the refining status identification results, an interlocking control operation is triggered, including: When the flame detection result indicates that the flame has not ignited during the ignition stage, the fuel supply and ignition operation of the target RH refining furnace are stopped, and a first warning message is issued. When the flame detection result indicates that the flame is extinguished, the oxygen supply and feeding operation of the target RH refining furnace shall be stopped. When the start and end times of the feeding event, the material type classification result, and the refining status identification result do not meet the preset metallurgical path requirements, the feeding operation of the target RH refining furnace is stopped, and a second warning message is issued. When the refining status identification result indicates the end of the first stage, an operation instruction for the second stage is generated, wherein the first stage and the second stage are process stages performed sequentially.

6. The RH refining furnace control method according to any one of claims 1 to 5, characterized in that, The acquisition of color image data of the target RH refining furnace includes: Color image data of the furnace interior is acquired by an image acquisition device installed at the bottom of the furnace top lance of the target RH refining furnace.

7. The RH refining furnace control method according to claim 6, characterized in that, The RH refining furnace control method also includes: The working status parameters of the image acquisition device are detected, wherein the working status parameters include frame rate stability and temperature sensing data; When the frame rate stability is lower than a preset stability threshold, the window cleaning device is activated to perform a lens surface dust removal operation. When the temperature sensing data is higher than the preset temperature threshold, the flow rate of the cooling medium in the camera cooling device is increased and the thermal expansion compensation program of the optical components is triggered.

8. A control device for an RH refining furnace, characterized in that, include: The data acquisition unit is used to acquire color image data of the target RH refining furnace inside the furnace; The flame detection unit is used to identify the flame state of the color image data inside the furnace using a first preset model, and obtain the flame detection result. The material feeding recognition unit is used to identify the material feeding action by analyzing the color image data inside the furnace using a second preset model, and to obtain material feeding event information. The state recognition unit is used to extract metallurgical stage features from the color image data inside the furnace using a third preset model to obtain refining state recognition results. The interlock control unit is used to trigger interlock control operations based on the flame detection results, the feeding event information, and the refining status identification results. The interlock control operations are used to regulate the fuel supply or equipment start-up and shutdown of the target RH refining furnace.

9. An electronic device, comprising: A memory and a processor, characterized in that the processor, when executing a computer program stored in the memory, implements the steps of the RH refining furnace control method as described in any one of claims 1 to 7.

10. 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 of the RH refining furnace control method as described in any one of claims 1 to 7.

Images