Splashing early warning method for converter smelting

Abstract

The invention provides a converter smelting splashing early warning method, which comprises the following steps of: acquiring a current flame image, a current flue gas curve and a current sonar time sequence signal by constructing a splashing early warning model which takes a multi-band flame image, a flue gas curve and a sonar time sequence signal as inputs and a splashing probability, a lance position, an oxygen flow rate and a coolant control quantity as outputs; and inputting the trained splashing early warning model to obtain the current splashing probability, the lance position, the oxygen flow and the coolant control quantity. The problems that in the prior art, accurate early warning cannot be given before splashing occurs, so that metal loss is reduced, the service life of a furnace lining is prolonged, and the smelting efficiency and the intrinsic safety level are improved are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for converter steelmaking processes, and in particular to a method for early warning of splashing during converter smelting. Background Technology

[0002] During converter steelmaking, the carbon-oxygen reaction in the molten pool is highly exothermic. When the reaction rate gets out of control, it often induces a "splashing" phenomenon: high-temperature slag and molten metal droplets instantly gush out from the furnace opening. Splashing not only causes metal loss, but also washes away the furnace lining, pollutes the plant, and endangers personal safety. It has become a common problem that the industry has long struggled to completely solve.

[0003] To suppress splashing, steel mills both domestically and internationally generally employ the following two methods: 1. Manual experience-based early warning: Operators rely on visual observation of the flame brightness, black smoke length, or slag flow pattern at the furnace opening, combined with the detected changes in the CO / CO2 ratio in the flue gas, to make judgments. This method generally has a response time >30s and is affected by factors such as smoke and dust obstruction, limited viewing angle, and sampling lag, resulting in an average false alarm rate of over 20%, and missed alarms also occur frequently.

[0004] 2. Existing suppression methods: When a CO peak or a sudden change in flame brightness is detected, the oxygen lance is immediately raised or the oxygen supply flow rate is reduced to weaken the decarburization rate. However, because this is not coupled with the control of the molten pool temperature and slag condition, "overblowing" or "post-blowing" often occurs, increasing production costs and energy consumption.

[0005] In recent years, some enterprises have attempted to introduce key equipment such as industrial cameras and laser gas analyzers in order to achieve non-contact online monitoring. However, the temperature in the furnace mouth area reaches 800–1000℃, accompanied by high concentrations of dust and strong mechanical vibration. Under these conditions, the equipment lifespan is less than 3 months, requiring frequent shutdowns for maintenance. The system cannot operate continuously for extended periods, making it difficult to meet the long-cycle, high-reliability production requirements of steelmaking.

[0006] Therefore, how to provide accurate early warnings before splashing occurs in order to reduce metal loss, extend furnace lining life, and improve smelting efficiency and intrinsic safety is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] Based on this, the purpose of this application is to provide a method for early warning of spatter in converter smelting, so as to solve at least one of the technical problems mentioned in the background art.

[0008] Firstly, this application provides a method for early warning of splashing in converter smelting, including: A splash warning model is constructed, taking multi-band flame images, smoke curves, and sonar time-series signals as inputs, and splash probability, gun position, oxygen flow rate, and coolant control quantity as outputs, including: The input layer is used to receive multi-band flame images, smoke profiles, and sonar timing signals. The feature extraction layer includes a flame feature extraction module, a smoke feature extraction module, and a sonar feature extraction module, which are used to extract flame features, smoke features, and sonar features respectively based on multi-band flame images, smoke curves, and sonar time-series signals; The feature fusion layer is used to fuse flame features, smoke features and sonar features to obtain a multimodal feature vector, and to obtain the splash probability, gun position, oxygen flow rate and coolant control amount based on the multimodal feature vector; The output layer is used to output splash probability, gun position, oxygen flow rate, and coolant control amount. The system collects current flame images, smoke curves, and sonar timing signals, inputs them into a trained splash warning model, and obtains the current splash probability, gun position, oxygen flow rate, and coolant control quantity.

[0009] Furthermore, the flame feature extraction module is used to extract flame features based on multi-band flame images; flame features include flame brightness features, spark density features, and slag overflow area features. The flue gas feature extraction module is used to extract flue gas features based on the flue gas curve; the flue gas features include CO concentration features, CO2 concentration features, and O2 concentration features; The sonar feature extraction module is used to extract sonar features based on the sonar time-series signal; the sonar features include the thickness of the slag layer.

[0010] Furthermore, the flame feature extraction module includes a flame brightness feature extraction unit, a spark density feature extraction unit, and a slag overflow area feature extraction unit arranged in parallel. The flame brightness feature extraction unit is used to perform band integration based on the pixel values ​​of each pixel in the flame image of each band to obtain the flame brightness features. The Mars density feature extraction unit is used to segment a flame image of any band into several connected regions, so as to identify the Martian pixels in the flame image based on each connected region, count the number of Martian pixels per unit area, and extract the Martian density feature. The overflow area feature extraction unit is used to select a flame image of a certain band and obtain the bright ring area in the flame image to obtain the overflow area feature.

[0011] Furthermore, the flame brightness feature extraction unit includes a brightness acquisition element, a radiation calibration element, a conversion element, a brightness superposition element, and a spatial averaging element connected in sequence: Brightness acquisition element, used to acquire the pixel value of each pixel in flame images of each band; A radiation calibration element is used to calibrate the pixel value to obtain the spectral radiance of each pixel in each band. The conversion element is used to obtain the product of spectral radiance, prior visual function, and prior bandwidth, to obtain the human eye luminance integral; A brightness superposition element is used to sum the human eye brightness of corresponding pixels in each flame image within a selected band range to obtain an integrated brightness map. The spatial averaging element is used to average the pixel values ​​of each pixel in the integrated brightness map to obtain the flame brightness characteristics.

[0012] Furthermore, the Mars density feature extraction unit includes a single-band extraction element, a preprocessing element, a segmentation element, and a density calculation element connected in sequence; A single-band extraction element is used to select a flame image in any band. The preprocessing element is used to perform dark field subtraction, threshold cropping, and median filtering on the selected image to remove noise and obtain a preprocessed image. The segmentation element is used to perform binary segmentation on the preprocessed image, obtain several connected components, and remove connected components with an area greater than a set area threshold, with the remaining connected components as Mars pixels. The density calculation element is used to count the number of Martian pixels per unit area to obtain Martian density characteristics.

[0013] Furthermore, the overflow area feature extraction unit includes a single-band extraction element, a positioning element, a background temperature statistics element, a thermal loop extraction element, a threshold segmentation element, and a geometric verification element; A single-band extraction element is used to select a flame image in any band. Positioning elements are used to acquire the center and radius of the furnace opening in the flame image; The background temperature statistical element is connected to the positioning element and the single-band extraction element. It is used to obtain the average pixel value within the radius of the flame image based on the center of the furnace opening, and thus obtain the background temperature. The thermal ring extraction element, connected to the single-band extraction element, is used to obtain the thermal ring response map based on the prior half-ring template and the selected flame image. The threshold segmentation element, connected to the thermal ring extraction element and the background temperature statistics element, is used to filter candidate bright ring pixels based on the thermal ring response map and background temperature, and merge them to obtain several connected components. The geometric verification element, connected to the positioning element and the threshold segmentation element, is used to perform geometric screening of each connected domain based on the furnace opening center and radius to obtain a bright ring mask, which is the slag overflow area feature.

[0014] Furthermore, the flue gas feature extraction module includes a CO concentration feature extraction unit, a CO2 concentration feature extraction unit, and an O2 concentration feature extraction unit arranged in parallel. The CO concentration feature extraction unit is used to extract the CO volume fraction at each time point based on the flue gas curve, and obtain the CO concentration change rate, which is the CO concentration feature. The CO2 concentration feature extraction unit is used to extract the CO2 volume fraction at each time point based on the flue gas curve, and obtain the CO2 concentration change rate, which is the CO2 concentration feature. The O2 concentration feature extraction unit is used to extract the O2 volume fraction at each time point based on the flue gas curve, and obtain the rate of change of O2 concentration, which is the O2 concentration feature.

[0015] Furthermore, the sonar feature extraction module includes; The sonar signal extraction unit is used to extract single-point sonar signals from the sonar timing signal. The slag layer thickness growth rate calculation unit is connected to the output of the sonar signal extraction unit. It is used to obtain the slag layer thickness based on the sonar ranging values ​​of each single point in order to obtain sonar characteristics.

[0016] Furthermore, the specific structure of the feature fusion layer includes, in sequence, a fusion module, a Concatenate node, a Transformer encoder, a global average pooling module, and a multi-task output module: The fusion module includes: The first fusion unit is used to fuse flame mode features based on flame brightness characteristics, spark density characteristics, and slag overflow area characteristics. The second fusion unit is used to fuse the flue gas modal characteristics based on the CO concentration characteristics, CO2 concentration characteristics and O2 concentration characteristics; The third fusion unit is used to fuse sonar modal features based on sonar characteristics to obtain sonar modal features; The Concatenate node is used to concatenate flame modal features, smoke modal features, and sonar modal features to obtain a multimodal feature vector; The Transformer encoder is used to capture global temporal features of multimodal feature vectors through a self-attention mechanism, thereby obtaining global contextual features. The global average pooling module is used to perform average pooling operations on global context features to reduce the computational complexity of the model and obtain global key features. The multi-task output module is used to learn in parallel the mapping relationship between global key features and prior splash risk and control variables, so as to synchronously output splash probability, gun position, oxygen flow increment and coolant increment based on global key features.

[0017] Furthermore, the multi-task output module includes: The splash probability head is used to map the current splash probability based on global key features; The judgment head, connected to the splash probability head, is used to determine whether the splash probability is greater than a set threshold. If so, the global key features are input in parallel to the gun position head, oxygen flow increment head, and coolant increment head. The gun position head, connected to the judgment head, is used to map the gun position based on global key features; The oxygen flow rate increment head, connected to the decision head, is used to map the oxygen flow rate based on global key features; The coolant increment head, connected to the decision head, is used to map the coolant control quantity based on global key features.

[0018] This invention provides a method for early warning of spatter in converter smelting. It constructs a spatter warning model with multi-band flame images, flue gas curves, and sonar time-series signals as inputs, and spatter probability, lance position, oxygen flow rate, and coolant control quantity as outputs. The model includes: an input layer for receiving multi-band flame images, flue gas curves, and sonar time-series signals; achieving comprehensive monitoring and improving detection reliability by simultaneously receiving three different types of time-series signals (flame, flue gas, and sonar); a feature extraction layer connected to the output layer, including a flame feature extraction module, a flue gas feature extraction module, and a sonar feature extraction module, used to extract flame features, flue gas features, and sonar features respectively from the multi-band flame images, flue gas curves, and sonar time-series signals; filtering out irrelevant signal components by extracting key features from a large amount of raw data, thereby improving the signal-to-noise ratio and reducing the complexity of subsequent processing; and a feature fusion layer, connected to the feature extraction layer... The outputs of the first layer are connected to fuse flame, flue gas, and sonar features to obtain a multimodal feature vector. Based on this vector, splash probability, nozzle position, oxygen flow rate, and coolant control parameters are derived. By fusing complementary information from different sensors, the limitations of a single sensor are avoided. Fusion reduces the impact of individual sensor anomalies on the final decision, thereby obtaining the necessary control parameters (splash probability, nozzle position, oxygen flow rate, and coolant control parameters), achieving integrated perception and decision-making. The output layer, connected to the feature fusion layer, outputs the splash probability, nozzle position, oxygen flow rate, and coolant control parameters, directly providing precise setpoints. These parameters, with clear physical meaning, facilitate operator understanding and intervention. The system collects current flame images, flue gas curves, and sonar time-series signals, inputting them into a trained splash warning model to obtain the current splash probability, nozzle position, oxygen flow rate, and coolant control parameters. This solves the problem that existing technologies cannot provide accurate warnings before splashing occurs, thus reducing metal loss, extending furnace lining life, and improving smelting efficiency and intrinsic safety. Attached Figure Description

[0019] Figure 1 This is a flowchart of the converter smelting splash early warning method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the splash warning model according to an embodiment of the present invention; Figure 3 This is another structural schematic diagram of the splash warning model according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the thermal loop extraction element according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the geometric verification element according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the feature fusion layer in an embodiment of the present invention; Figure 7 This is a schematic diagram of the optimization strategy in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that if the embodiments of the present invention involve directional indications, such as up, down, left, right, front, back, etc., these directional indications are only used to explain the relative positional relationships and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly. Furthermore, if the embodiments of the present invention involve descriptions such as "first," "second," "S1," "S2," "step one," "step two," etc., these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance, or implicitly indicating the number of technical features indicated or the order of method execution. Those skilled in the art will understand that anything that does not violate the inventive concept and is within the scope of the present invention should be included in the protection scope of the present invention.

[0022] like Figure 1 As shown, the present invention provides a method for early warning of splashing in converter smelting, comprising: S1: Collect historical multi-band flame images, flue gas curves and sonar time-series signals of the converter, and label the converter splashing status, lance position, oxygen flow rate and coolant control amount at each historical moment to build a training dataset. Specifically, the acquisition equipment can be installed at a set location and collected multi-band flame images, flue gas curves, and sonar timing signals of the converter within a set time period. It can also be labeled whether splashing occurred in the converter at each time and the corresponding operations taken by the staff, such as lance position, oxygen flow rate, and coolant control amount, in order to build a training dataset. The acquisition equipment includes commonly used equipment such as flame information acquisition devices, flue gas analyzers, and sonar detectors.

[0023] For example, the flame information acquisition device can be installed in front of the converter furnace, directly facing the furnace opening, using a vertical mounting mechanism that allows for adjustment of the acquisition range by rotating up, down, left, and right. It is equipped with a water-cooled protective cover to prevent high-temperature damage to the camera components; the lens is fitted with high-temperature resistant quartz glass to reduce dust adhesion; the bracket needs a shockproof design to avoid image blurring; the cable uses high-temperature resistant shielded wire and is fixed along the side of the furnace body. The flue gas analyzer is installed on the flue duct after the evaporator and condenser to avoid damage to the laser analyzer.

[0024] S2: Based on the training dataset, construct a splash warning model that takes multi-band flame images, smoke curves, and sonar time-series signals as inputs, and splash probability, gun position, oxygen flow rate, and coolant control quantity as outputs. Figure 2 , Figure 3 As shown, it includes: The input layer is used to receive multi-band flame images, smoke profiles, and sonar timing signals. The feature extraction layer, connected to the output of the output layer, includes a flame feature extraction module, a smoke feature extraction module, and a sonar feature extraction module, which are used to extract flame features, smoke features, and sonar features respectively based on multi-band flame images, smoke curves, and sonar time-series signals. The feature fusion layer, connected to the output of the feature extraction layer, is used to fuse flame features, smoke features, and sonar features to obtain a multimodal feature vector, and to obtain the splash probability, gun position, oxygen flow rate, and coolant control amount based on the multimodal feature vector. The output layer, connected to the output of the feature fusion layer, is used to output splash probability, gun position, oxygen flow rate, and coolant control amount.

[0025] Specifically, the system can simultaneously receive three different types of signals—multi-band flame images, smoke curves, and sonar timing signals—through the input layer to achieve comprehensive monitoring and improve detection reliability. Then, the feature extraction layer extracts key features such as flame characteristics, smoke characteristics, and sonar characteristics from a large amount of raw data (including multi-band flame images, smoke curves, and sonar timing signals), thereby filtering out irrelevant signal components, improving the signal-to-noise ratio, and reducing the complexity of subsequent processing. Next, the feature fusion layer fuses the flame, smoke, and sonar features to obtain a multimodal feature vector, and obtains the mapping relationship between the multimodal feature vector and splash risk and control variables, avoiding the limitations of a single sensor. The multimodal feature vector reduces the impact of single sensor anomalies on the final decision, thereby obtaining the parameters required for control (splash probability, nozzle position, oxygen flow rate, coolant control quantity), achieving integrated perception and decision-making. Finally, the output layer outputs the results predicted by the feature fusion layer, providing accurate setpoints for subsequent steps. Parameters with clear physical meaning facilitate operator understanding and intervention.

[0026] Preferably, the flame feature extraction module is used to extract flame features based on multi-band flame images; the flame features include flame brightness features, spark density features, and slag area features. The flue gas feature extraction module is used to extract flue gas features based on the flue gas curve; the flue gas features include CO concentration features, CO2 concentration features, and O2 concentration features; The sonar feature extraction module is used to extract sonar features based on the sonar time-series signal; the sonar features include the thickness of the slag layer.

[0027] This embodiment provides specific examples of flame characteristics, smoke characteristics, and sonar characteristics. The specific selection of these three types of characteristics corresponds precisely to the core precursors before a splash occurs: Flame characteristics (brightness, spark density, slag overflow area): Before splashing occurs, the carbon-oxygen reaction in the furnace intensifies dramatically, causing a sudden increase and fluctuation in flame brightness; the number of unreacted molten metal droplets increases, and the spark density rises significantly; simultaneously, the foaming and surging of molten slag leads to a rapid expansion of the slag overflow area at the furnace mouth. Extracting these three characteristics allows for the direct capture of intuitive precursors to splashing from a visual perspective, avoiding the lag inherent in manual observation.

[0028] Flue gas characteristics (CO, CO2, O2 concentrations): Splashing is directly related to the rate of gas generation in the furnace—when the carbon-oxygen reaction runs out of control, the CO concentration will rise sharply in a short period of time, while the CO2 concentration will fluctuate synchronously; while the O2 concentration will drop abnormally due to the imbalance between oxygen supply and reaction. By extracting the dynamic change characteristics of flue gas concentration, the intensity of the reaction can be quantified from a chemical perspective to determine whether it is approaching the splashing critical state.

[0029] Sonar characteristics (slag layer thickness): One of the key causes of converter splashing is an excessively thick slag layer. After the molten slag foams, the slag layer thickness exceeds the critical value, which can easily trigger splashing. Sonar's non-contact monitoring can penetrate high-temperature and high-dust environments to extract slag layer thickness characteristics in real time and accurately determine the state of the molten slag. Compared with traditional contact measurements (such as probes), it can achieve continuous monitoring and avoid missing splash precursors due to measurement blind spots. Therefore, the fusion and specific selection of these three types of features have constructed a multi-dimensional and highly sensitive splash precursor identification system. By capturing multi-source abnormal signals before splashing occurs, it can achieve early warning and accurate prediction of splashing risks, ultimately reducing the splashing accident rate and ensuring smelting safety and efficiency.

[0030] The following describes preferred embodiments for the flame feature extraction module, including how to extract flame brightness, spark density, and slag area features; the flue gas feature extraction module, including how to extract CO concentration, CO2 concentration, and O2 concentration features; and the sonar feature extraction module, including how to extract slag layer thickness. These embodiments are provided, but are not limited to these. The key to this invention lies in fusing and analyzing flame, flue gas, and sonar data based on the splashing mechanism, and further selecting the aforementioned specific features for finer-grained analysis and segmentation. Any method can be used to extract these features.

[0031] More specifically, the flame feature extraction module includes a flame brightness feature extraction unit, a spark density feature extraction unit, and a slag overflow area feature extraction unit arranged in parallel; The flame brightness feature extraction unit is used to perform band integration based on the pixel values ​​of each pixel in the flame image of each band to obtain the flame brightness features. The Mars density feature extraction unit is used to segment a flame image of any band into several connected regions, so as to identify the Martian pixels in the flame image based on each connected region, count the number of Martian pixels per unit area, and extract the Martian density feature. The overflow area feature extraction unit is used to select a flame image of a certain band and obtain the bright ring area in the flame image to obtain the overflow area feature.

[0032] Specifically, the flame brightness feature extraction unit can be used to obtain the radiation energy integral of multiple bands such as visible and near-infrared, which is monotonically positively correlated with the heat release of combustion and can be directly used for combustion intensity assessment to obtain flame brightness features. Then, the spark density feature extraction unit can be used to remove connected regions with an area larger than a set area threshold, and the remaining connected regions can be used as spark pixels, thereby filtering out false targets such as large reflectors and furnace walls, reducing false detections. Finally, the overflow slag area feature extraction unit can be used to obtain the tilt angle change of the slag surface, providing "physical boundary" evidence for visual and chemical signals.

[0033] Preferably, the flame brightness feature extraction unit includes a brightness acquisition element, a radiation calibration element, a conversion element, a brightness superposition element, and a spatial averaging element connected in sequence. Brightness acquisition element, used to acquire the pixel value of each pixel in flame images of each band; A radiation calibration element is used to calibrate the pixel value to obtain the spectral radiance of each pixel in each band. The conversion element is used to obtain the product of spectral radiance, prior visual function, and prior bandwidth, to obtain the human eye luminance integral; A brightness superposition element is used to sum the human eye brightness of corresponding pixels in each flame image within a selected band range to obtain an integrated brightness map. The spatial averaging element is used to average the pixel values ​​of each pixel in the integrated brightness map to obtain the flame brightness characteristics.

[0034] Specifically, a brightness acquisition element can be used to acquire the pixel values ​​of each pixel in the flame image of each band, and a radiation calibration element can be used to perform radiation calibration on the pixel values ​​to obtain the spectral radiance of each pixel in each band. This converts the raw digital quantity into an absolute radiance quantity with physical meaning, eliminating measurement differences between different batches and different probes and improving data consistency. Then, a conversion element multiplies the spectral radiance by the prior visual function and the prior band width to obtain the human eye brightness integral, so that the extraction result directly corresponds to the "whitening" or "bluish" perception of the worker's naked eye, enhancing the interpretability of the feature and making it easier for operators to understand and intervene. Subsequently, a brightness superposition element sums the human eye brightness integrals of the corresponding pixels in each flame image within the selected band range to obtain an integral brightness map, realizing information compression of the band dimension and reducing the complexity of subsequent calculations. Finally, a spatial averaging element averages the pixel values ​​of each pixel in the integral brightness map to obtain the flame brightness feature, compressing the two-dimensional matrix into a one-dimensional scalar, further reducing the data dimension and improving real-time performance.

[0035] The Mars density feature extraction unit includes a single-band extraction element, a preprocessing element, a segmentation element, and a density calculation element connected in sequence. A single-band extraction element is used to select a flame image in any band. The preprocessing element is used to perform dark field subtraction, threshold cropping, and median filtering on the selected image to remove noise and obtain a preprocessed image. The segmentation element is used to perform binary segmentation on the preprocessed image, obtain several connected components, and remove connected components with an area greater than a set area threshold, with the remaining connected components as Mars pixels. The density calculation element is used to count the number of Martian pixels per unit area to obtain Martian density characteristics.

[0036] Specifically, the system can select any one band of flame image using a single-band extraction element to avoid multi-band registration errors, reduce reliance on hardware synchronization, and improve system robustness. Then, a preprocessing element performs preprocessing steps on the selected image, such as dark field subtraction, threshold cropping, and median filtering, to remove interference from probe thermal noise, window dust accumulation, and random pulse noise, thus improving the signal-to-noise ratio. Next, a segmentation element performs binary segmentation on the preprocessed image, obtaining several connected components. Connected components with areas larger than a set area threshold are removed, and the remaining connected components are used as Martian pixels, thereby filtering out false targets such as large reflectors and furnace walls, reducing false detections. Finally, a density calculation element counts the number of Martian pixels per unit area to obtain Martian density features, compressing two-dimensional spatial information into a one-dimensional density index, simplifying subsequent model input.

[0037] Preferably, the overflow area feature extraction unit includes a single-band extraction element, a positioning element, a background temperature statistics element, a thermal loop extraction element, a threshold segmentation element, and a geometric verification element; A single-band extraction element is used to select a flame image in any band. Positioning elements are used to acquire the center and radius of the furnace opening in the flame image; The background temperature statistical element is connected to the positioning element and the single-band extraction element. It is used to obtain the average pixel value within the radius of the flame image based on the center of the furnace opening, and thus obtain the background temperature. The thermal ring extraction element, connected to the single-band extraction element, is used to obtain the thermal ring response map based on the prior half-ring template and the selected flame image. The threshold segmentation element, connected to the thermal ring extraction element and the background temperature statistics element, is used to filter candidate bright ring pixels based on the thermal ring response map and background temperature, and merge them to obtain several connected components. The geometric verification element, connected to the positioning element and the threshold segmentation element, is used to perform geometric screening of each connected domain based on the furnace opening center and radius to obtain a bright ring mask, which is the slag overflow area feature.

[0038] Specifically, a flame image of any band can be selected using a single-band extraction element to avoid segmentation errors caused by differences in multiple bands and simplify the algorithm process. Then, a positioning element is used to obtain the furnace opening center and radius in the flame image, providing a unified coordinate reference for subsequent geometric verification and improving positioning accuracy. Next, a background temperature statistics element is used to calculate the average pixel value within the radius of the furnace opening center in the flame image to obtain the background temperature, allowing the segmentation threshold to adaptively adjust with furnace condition drift and enhancing environmental adaptability. Then, a thermal ring extraction element is used to obtain the thermal ring response map based on a priori semi-ring template and the selected flame image, transforming the irregular two-dimensional arc-shaped target into a one-dimensional response peak, reducing computational complexity. Subsequently, ... The threshold segmentation element selects candidate bright ring pixels based on the thermal ring response map and background temperature, and merges them to obtain several connected regions, achieving adaptive segmentation and reducing manual parameter tuning. Then, the geometric verification element performs three-level geometric screening of each connected region based on the furnace center and radius, considering area, roundness, and position, to obtain a bright ring mask, eliminating false targets such as probe reflection and local bright spots, and improving the accuracy of slag overflow area extraction. Finally, the flame feature fusion unit fuses the slag overflow area features with the flame brightness features and spark density features, enabling the system to have a three-dimensional visual perception capability of "whole-particle-region", avoiding the one-sidedness of single visual features in splash risk assessment, thereby improving the reliability of all-round monitoring.

[0039] Preferably, the thermal ring extraction element, such as Figure 4 As shown, it includes a global radiation calibration component, a temperature inversion component, and a ring template convolution component connected in sequence; The global radiometric calibration component is used to calibrate the pixel values ​​of each pixel in the flame image to obtain the spectral radiance of each pixel and obtain the absolute radiance map. The temperature inversion component is used to independently invert each pixel in the absolute radiance map, obtain the temperature value corresponding to each pixel, and obtain the apparent temperature map. A ring template convolution component is used to perform normalized cross-correlation between a prior semi-ring template and the apparent temperature map to obtain a thermal ring response map.

[0040] Specifically, a global radiometric calibration component can be used to radiometrically calibrate the pixel values ​​of each pixel in the flame image to obtain the spectral radiance of each pixel, thereby converting the original grayscale values ​​into absolute physical quantities, eliminating measurement drift caused by probe aging and window contamination, and improving data consistency. Then, a temperature inversion component can be used to independently invert each pixel in the absolute radiance map to obtain the corresponding temperature value of each pixel, resulting in an apparent temperature map, giving each pixel independent physical meaning and enhancing feature interpretability. Finally, a ring template convolution component can be used to perform normalized cross-correlation between the prior semi-ring template and the apparent temperature map to obtain a thermal ring response map, thereby transforming the two-dimensional irregular arc-shaped target into a one-dimensional response peak, reducing the complexity of subsequent segmentation, improving computational efficiency, and achieving high-precision positioning and quantification of the furnace mouth slag overflow area, providing key spatial information for splash risk assessment.

[0041] For example, the camera is calibrated at two points in a blackbody furnace within the range of 1200-1700℃ at the factory to establish a grayscale-radiance curve. Therefore, the grayscale value of each 950nm image corresponds to the absolute radiance L(λ,T). Then, the "apparent temperature" T_app(x,y) is inverted according to Planck's formula. The normal temperature range at the furnace mouth is T_app≈1450℃. When there are signs of slag overflow, molten slag foam approaches the furnace mouth, and its surface thin steel film radiates secondary radiation, forming a "thermal ring" with T_app 30-80℃ higher than the background. Then, a normalized cross-correlation is performed between the semi-ring templates (inner diameter = 0.9 × furnace mouth radius, outer diameter = 1.1 × furnace mouth radius) in four directions (0° / 45° / 90° / 135°) and the T_app diagram to obtain the thermal ring response diagram R_ring. The template functions are: h(θ) = +1 (ring zone); h(θ) = -1 (center + outside).

[0042] It can suppress uniform bright fields and highlight the annular feature of "bright outer ring and dark center".

[0043] Next, take the set of pixels where R_ring > 0.7 and T_app > T_bg + 25°C, perform 8-connected component merging to obtain candidate bright ring regions; then perform geometric verification: Area A must be ≥ 0.08m² 2 (≈5% of the furnace opening area); Roundness C = 4πA / P 2 ≥0.85; If the offset between the circumscribed circle and the center of the furnace opening is less than 3% of the furnace opening diameter, it is considered a "bright ring".

[0044] Finally, timing verification is performed. Only when a bright ring is detected for three consecutive frames (100ms / frame) is the "overflow precursor" flag triggered. This flag is then packaged with I(t) and n(t) into a 3D vector and sent to the flame control model. This ensures that interference such as high-temperature smoke and local sparks are excluded, based on the complete chain of 950nm radiance → temperature → annular thermal characteristics → geometric + timing verification.

[0045] Preferred geometric verification elements, such as Figure 5 As shown, it includes an area verification component, a roundness verification component, and a position verification component connected in sequence; The area verification component is used to set a verification area threshold based on the furnace opening radius and remove connected components with an area smaller than the verification area threshold to obtain several initial candidate regions. The roundness verification component is used to obtain the roundness of each initial candidate region and remove the initial candidate regions whose roundness is less than the set roundness threshold to obtain the bright ring candidate regions. The position verification component is used to obtain the centroid of each bright ring candidate region, calculate the offset distance between each centroid and the center of the furnace mouth, and remove the bright ring candidate regions whose offset distance is greater than the set offset threshold to obtain the bright ring mask, which is the overflow area feature.

[0046] Specifically, an area verification component can be used to set a verification area threshold based on the furnace opening radius and eliminate connected regions with areas smaller than the verification area threshold, thereby filtering out small false targets such as probe reflections and random bright spots, reducing false detections. Then, a roundness verification component is used to obtain the roundness of each initial candidate region and eliminate initial candidate regions with roundness smaller than the set roundness threshold, retaining the true bright rings of overflow slag in approximately arc shapes, improving shape consistency. Next, a position verification component is used to obtain the centroid of each bright ring candidate region and calculate the offset distance between each centroid and the furnace opening center, eliminating bright ring candidate regions with offset distances greater than the set offset threshold, thereby removing false bright areas far from the furnace opening and improving spatial positioning accuracy. Finally, a bright ring mask is obtained as an overflow slag area feature, providing a highly reliable spatial indicator for subsequent flame modal feature fusion, reducing the risk of misjudgment caused by single geometric attributes, and improving the overall robustness of the system.

[0047] Preferably, the flue gas feature extraction module includes a CO concentration feature extraction unit, a CO2 concentration feature extraction unit, and an O2 concentration feature extraction unit arranged in parallel. The CO concentration feature extraction unit is used to extract the CO volume fraction at each time point based on the flue gas curve, and obtain the CO concentration change rate, which is the CO concentration feature. The CO2 concentration feature extraction unit is used to extract the CO2 volume fraction at each time point based on the flue gas curve, and obtain the CO2 concentration change rate, which is the CO2 concentration feature. The O2 concentration feature extraction unit is used to extract the O2 volume fraction at each time point based on the flue gas curve, and obtain the rate of change of O2 concentration, which is the O2 concentration feature.

[0048] Specifically, the CO concentration feature extraction unit, CO2 concentration feature extraction unit, and O2 concentration feature extraction unit can be selected to extract the volume fraction of the gas at each time point according to the flue gas curve, and calculate its rate of change, thereby converting the original concentration sequence into a dynamic fluctuation index, eliminating the influence of probe baseline drift on the absolute value, and improving feature stability.

[0049] Preferably, only CO concentration features and CO2 concentration features can be fused to obtain flue gas modal features, and only O2 concentration features can be used for logical verification to improve efficiency and accuracy.

[0050] Preferably, the sonar feature extraction module includes: The sonar signal extraction unit is used to extract single-point sonar signals from the sonar timing signal. The slag layer thickness growth rate calculation unit is connected to the output of the sonar signal extraction unit. It is used to obtain the slag layer thickness based on the sonar ranging values ​​of each single point, so as to obtain the slag layer thickness growth rate characteristics.

[0051] Specifically, a single-point sonar signal can be extracted from the sonar time sequence signal through a sonar signal extraction unit, thereby splitting the multi-channel echo into an independent ranging sequence, providing a single data source for subsequent calculations, and avoiding errors caused by channel aliasing.

[0052] Preferably, the specific structure of the feature fusion layer is as follows: Figure 6 As shown, it includes the fusion module, Concatenate node, Transformer encoder, global average pooling module, and multi-task output module, which are connected in sequence: The fusion module includes: The first fusion unit is used to fuse flame modal features based on flame brightness features, spark density features, and slag area features. For example, the first fusion unit may use a flame CNN with 1-D convolution kernel size=3 and 64 channels, receiving flame brightness features, spark density features, and slag area features, and outputting a 256-dimensional feature vector F_f.

[0053] The second fusion unit is used to fuse the CO concentration features, CO2 concentration features and O2 concentration features to obtain flue gas modal features. For example, the second feature fusion unit can be a flue gas CNN with 1-D convolution kernel size=3 and 64 channels, receiving CO concentration features, CO2 concentration features and O2 concentration features, and outputting a 256-dimensional feature vector F_g.

[0054] The third fusion unit is used to fuse sonar modal features based on sonar features. For example, the third fusion unit can be a sonar CNN with 1-D convolution kernel size=3 and 64 channels. It receives slag slope features and slag thickness growth rate features and outputs a 256-dimensional feature vector F_s.

[0055] The Concatenate node is used to concatenate flame modal features, smoke modal features, and sonar modal features to obtain a multimodal feature vector; The Transformer encoder is used to capture global temporal features of multimodal feature vectors through a self-attention mechanism, thereby obtaining global contextual features. The global average pooling module is used to perform average pooling operations on global context features to reduce the computational complexity of the model and obtain global key features. The multi-task output module is used to learn in parallel the mapping relationship between global key features and prior splash risk and control variables, so as to synchronously output splash probability, gun position, oxygen flow increment and coolant increment based on global key features.

[0056] Specifically, the system can be configured to sequentially fuse flame brightness features, spark density features, and slag overflow area features through a fusion module to obtain flame modal features. This allows the system to simultaneously possess three complementary types of information: energy intensity, high-temperature particle activity, and slag overflow area, avoiding the limitations of a single indicator in splash risk assessment. Furthermore, it can fuse CO concentration features, CO2 concentration features, and O2 concentration features to obtain flue gas modal features, enabling the system to simultaneously possess three complementary types of information: reducing power, oxidizing power, and combustion efficiency, again avoiding the limitations of a single gas indicator in splash risk assessment. Finally, it can fuse sonar features to obtain sonar modal features, reducing the impact of single-point sensor failure on the final decision and improving the accuracy and robustness of splash warnings. Then, the flame modal features, flue gas modal features, and sonar modal features are concatenated using a concatenate node to obtain a multimodal feature vector, thereby achieving cross-modal alignment without losing the original information. The network structure is simplified; then, the Transformer encoder uses a self-attention mechanism to capture global temporal features of multimodal feature vectors, obtaining global context features. This enables the model to perceive non-aligned temporal correlations such as "flame changes first - flue gas changes later - sonar abrupt changes," improving the modeling ability for complex splash evolution processes. A global average pooling module then performs average pooling operations on the global context features, compressing variable-length sequences into fixed-dimensional vectors, reducing model computational complexity and improving edge deployment efficiency. Finally, a multi-task output module learns in parallel the mapping relationship between global key features and prior splash risks and control variables, simultaneously outputting splash probability, nozzle position, oxygen flow increment, and coolant increment, achieving integrated perception and decision-making. This provides setpoints with clear physical meaning for subsequent PLC closed-loop control, facilitating operator understanding and intervention, reducing the frequency of manual intervention, and improving production stability.

[0057] Preferably, the multi-task output module includes: The splash probability head is used to map the current splash probability based on global key features; The judgment head, connected to the splash probability head, is used to determine whether the splash probability is greater than a set threshold. If so, the global key features are input in parallel to the gun position head, oxygen flow increment head, and coolant increment head. The gun position head, connected to the judgment head, is used to map the gun position based on global key features; The oxygen flow rate increment head, connected to the decision head, is used to map the oxygen flow rate based on global key features; The coolant increment head, connected to the decision head, is used to map the coolant control quantity based on global key features.

[0058] Specifically, since measures (such as moving the nozzle position, adjusting oxygen flow rate, and controlling coolant flow) are only required to suppress splashing when it has already occurred or is about to occur, a possible approach is to obtain the current splashing probability using a splashing probability head and then determine whether the splashing probability exceeds a set threshold using a judgment head. If not, it indicates that there is no splashing risk, and therefore no measures need to be taken to suppress splashing, reducing data processing load. If so, it indicates that splashing is about to occur, and measures (such as moving the nozzle position, adjusting oxygen flow rate, and controlling coolant flow) need to be obtained based on global key features to suppress splashing. Therefore, global key features need to be input in parallel into the nozzle position head, oxygen flow rate increment head, and coolant increment head to obtain the nozzle position, oxygen flow rate increment, and coolant increment. Preferably, the splashing probability head, nozzle position head, oxygen flow rate increment head, and coolant increment head can be selected as a fully connected layer or a multilayer perceptron network.

[0059] S3: Collect the current flame image, smoke curve and sonar timing signal, input them into the trained splash warning model, and obtain the current splash probability, gun position, oxygen flow rate and coolant control amount.

[0060] Specifically, the system can selectively acquire current flame images, flue gas curves, and sonar timing signals from the converter, input them into a trained splash warning model, and obtain the current splash probability, lance position, oxygen flow rate, and coolant control quantity. Based on the current splash probability, lance position, oxygen flow rate, and coolant control quantity, the corresponding suppression system can be controlled to take corresponding measures for suppression. The suppression system includes oxygen lances, servo hydraulic cylinders, butterfly valves, high-level silo vibrating feeders, and belt scales, etc.

[0061] In this embodiment, a splash warning method for converter smelting according to the present invention is presented. This method constructs a splash warning model with multi-band flame images, flue gas curves, and sonar time-series signals as inputs, and splash probability, lance position, oxygen flow rate, and coolant control quantity as outputs. The model includes: an input layer for receiving multi-band flame images, flue gas curves, and sonar time-series signals; by simultaneously receiving three different types of time-series signals (flame, flue gas, and sonar), comprehensive monitoring is achieved, improving detection reliability; a feature extraction layer connected to the output layer, including a flame feature extraction module, a flue gas feature extraction module, and a sonar feature extraction module, used to extract flame features, flue gas features, and sonar features respectively based on the multi-band flame images, flue gas curves, and sonar time-series signals; by extracting key features from a large amount of raw data, irrelevant signal components are filtered out, improving the signal-to-noise ratio and reducing the complexity of subsequent processing; and a feature fusion layer. Connected to the output of the feature extraction layer, this layer fuses flame, flue gas, and sonar features to obtain a multimodal feature vector. Based on this vector, it calculates the splash probability, nozzle position, oxygen flow rate, and coolant control parameters. By fusing complementary information from different sensors, it avoids the limitations of a single sensor and reduces the impact of individual sensor anomalies on the final decision, thereby obtaining the necessary control parameters (splash probability, nozzle position, oxygen flow rate, and coolant control parameters) and achieving integrated perception and decision-making. The output layer, connected to the output of the feature fusion layer, outputs the splash probability, nozzle position, oxygen flow rate, and coolant control parameters, directly providing precise setpoints. These parameters, with clear physical meaning, facilitate operator understanding and intervention. The system acquires current flame images, flue gas curves, and sonar time-series signals, inputs them into a trained splash warning model, and obtains the current splash probability, nozzle position, oxygen flow rate, and coolant control parameters. This solves the problem that existing technologies cannot provide accurate warnings before splashing occurs, thus reducing metal loss, extending furnace lining life, and improving smelting efficiency and intrinsic safety.

[0062] In a preferred embodiment, the converter smelting splash early warning method of the present invention includes: Step 1: Signal acquisition (including time-series signals such as flame brightness I(t), spark density M(t), slag overflow area S(t), CO(t), CO2(t), O2(t), and slag layer slope K(t)). ① Flame: The average brightness I(t) at 1Hz was taken from a water-cooled multispectral camera 10m from the furnace opening. When I(t) > 200% of the reference value (12000cd / m²), 2 And the number of Martian particles is greater than 300 per 0.1m. 2 At this time, "flame anomalies" are marked, including bright white flames, a surge of sparks, and signs of impending spillage.

[0063] ② Flue gas: The laser analyzer outputs CO and CO2 at 1Hz after the evaporator cooler; if CO increases by more than 8% and CO2 drops by more than 5% within 3 seconds, it is marked as "flue gas abnormal", including sudden increase, sudden drop, violent reaction, etc.

[0064] ③ The thickness of the slag layer is measured by the furnace skirt sonar at 1Hz. If the growth rate is >8mm / s, it is marked as "slag layer steep rise".

[0065] If any of the above three markers appear, the system enters a 50-second sliding window buffer.

[0066] Step 2: Early warning judgment (10–50 seconds in advance) If both "flame anomaly" and "smoke anomaly" exist simultaneously within the buffer window for ≥5s, the Transformer network outputs a splash probability P(t) ≥0.7, and immediately sends an "early warning frame" to the PLC with a timestamp accuracy of 1ms.

[0067] Step 3: According to... Figure 7 The optimization strategy shown generates a suppression command (≤1s). The host computer's MPC (Multi-Process Calculation) uses a rolling optimization approach with three objectives: minimizing metal loss, minimizing furnace lining erosion, and achieving carbon hit at the endpoint. The solution is as follows: ① Gun position ΔH: 1.2m→2.5m (+1.3m); ②Oxygen flow rate increment ΔQ: 48000→35000Nm 3 / h (−13000); ③ Coolant ΔC: 80kg (30% / 70% of 8–12mm and 12–20mm mixed ore).

[0068] The result is encapsulated as an EtherCAT frame and sent.

[0069] Step 4: Perform the action (complete in ≤3 seconds) ① The servo hydraulic cylinder raises the gun to 2.5m in 0.8s; ②The butterfly valve closes linearly to 35000 Nm in 3 seconds. 3 / h; ③ The high-level silo vibrating feeder + belt scale completes 80kg of precise feeding in 4 seconds.

[0070] After the action is completed, the system continues to monitor at 1Hz. If P(t) < 0.3 for 10 seconds, it will exit the suppression mode and resume normal blowing.

[0071] Preferably, the meanings of each parameter and professional name in the above steps are shown in Tables 1, 2, and 3: Table 1: Signal Parameter Definitions

[0072] Table 2: Explanation of Process Parameters

[0073] Table 3: Definitions of Operation Names

[0074] The operator in the central control room only needs to observe the P(t) curve: once the value exceeds 0.7, the system completes the "lifting the lance - reducing oxygen - adding ore" three-step process within 1 second, eliminating splashing on average 32 seconds in advance, reducing metal loss from 1.8 kg / t to 0.25 kg / t, increasing furnace lining life by 18%, and achieving "early, accurate, and stable" fully automatic closed-loop control.

[0075] On the other hand, the present invention also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above-mentioned converter smelting splash early warning methods.

[0076] On the other hand, the present invention also provides a terminal device, including a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to execute any of the above-mentioned converter smelting splash early warning methods.

[0077] For example, the program code can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the program code in the terminal device.

[0078] The terminal device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the terminal device may also include input / output devices, network access devices, buses, etc.

[0079] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0080] The memory can be an internal storage unit of the terminal device, such as a hard drive or RAM. The memory can also be an external storage device of the terminal device, such as a plug-in hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units of the terminal device. The memory is used to store the program code and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output.

[0081] The aforementioned computer storage medium and terminal equipment are created based on the aforementioned converter smelting splash early warning method. Their technical functions and beneficial effects will not be elaborated here. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for early warning of splashing in converter smelting, characterized in that, include: A splash warning model is constructed, taking multi-band flame images, smoke curves, and sonar time-series signals as inputs, and splash probability, gun position, oxygen flow rate, and coolant control quantity as outputs, including: The input layer is used to receive multi-band flame images, smoke profiles, and sonar timing signals. The feature extraction layer includes a flame feature extraction module, a smoke feature extraction module, and a sonar feature extraction module, which are used to extract flame features, smoke features, and sonar features respectively based on multi-band flame images, smoke curves, and sonar time-series signals; The feature fusion layer is used to fuse flame features, smoke features and sonar features to obtain a multimodal feature vector, and to obtain the splash probability, gun position, oxygen flow rate and coolant control amount based on the multimodal feature vector; The output layer is used to output splash probability, gun position, oxygen flow rate, and coolant control amount. The system collects current flame images, smoke curves, and sonar timing signals, inputs them into a trained splash warning model, and obtains the current splash probability, gun position, oxygen flow rate, and coolant control quantity.

2. The method according to claim 1, characterized in that, The flame feature extraction module is used to extract flame features from multi-band flame images; flame features, This includes flame brightness characteristics, spark density characteristics, and slag overflow area characteristics; The flue gas feature extraction module is used to extract flue gas features based on the flue gas curve. Flue gas characteristics, including CO concentration characteristics, CO2 concentration characteristics, and O2 concentration characteristics; The sonar feature extraction module is used to extract sonar features based on the sonar time-series signal; Sonar characteristics, including slag layer thickness.

3. The method according to claim 2, characterized in that, The flame feature extraction module includes a flame brightness feature extraction unit, a spark density feature extraction unit, and a slag overflow area feature extraction unit arranged in parallel. The flame brightness feature extraction unit is used to perform band integration based on the pixel values ​​of each pixel in the flame image of each band to obtain the flame brightness features. The Mars density feature extraction unit is used to segment a flame image of any band into several connected regions, so as to identify the Martian pixels in the flame image based on each connected region, count the number of Martian pixels per unit area, and extract the Martian density feature. The overflow area feature extraction unit is used to select a flame image of a certain band and obtain the bright ring area in the flame image to obtain the overflow area feature.

4. The method according to claim 3, characterized in that, The flame brightness feature extraction unit includes a brightness acquisition element, a radiation calibration element, a conversion element, a brightness superposition element, and a spatial averaging element connected in sequence. Brightness acquisition element, used to acquire the pixel value of each pixel in flame images of each band; A radiation calibration element is used to calibrate the pixel value to obtain the spectral radiance of each pixel in each band. The conversion element is used to obtain the product of spectral radiance, prior visual function, and prior bandwidth, to obtain the human eye luminance integral; A brightness superposition element is used to sum the human eye brightness of corresponding pixels in each flame image within a selected band range to obtain an integrated brightness map. The spatial averaging element is used to average the pixel values ​​of each pixel in the integrated brightness map to obtain the flame brightness characteristics.

5. The method according to claim 3, characterized in that, The Mars density feature extraction unit includes a single-band extraction element, a preprocessing element, a segmentation element, and a density calculation element connected in sequence. A single-band extraction element is used to select a flame image in any band. The preprocessing element is used to perform dark field subtraction, threshold cropping, and median filtering on the selected image to remove noise and obtain a preprocessed image. The segmentation element is used to perform binary segmentation on the preprocessed image, obtain several connected components, and remove connected components with an area greater than a set area threshold, with the remaining connected components as Mars pixels. The density calculation element is used to count the number of Martian pixels per unit area to obtain Martian density characteristics.

6. The method according to claim 3, characterized in that, The overflow area feature extraction unit includes a single-band extraction element, a positioning element, a background temperature statistics element, a thermal loop extraction element, a threshold segmentation element, and a geometric verification element; A single-band extraction element is used to select a flame image in any band. The positioning element, connected to the single-band extraction element, is used to acquire the center and radius of the furnace opening in the flame image; The background temperature statistical element is connected to the positioning element and the single-band extraction element. It is used to obtain the average pixel value within the radius of the flame image based on the center of the furnace opening, and thus obtain the background temperature. The thermal ring extraction element, connected to the single-band extraction element, is used to obtain the thermal ring response map based on the prior half-ring template and the selected flame image. The threshold segmentation element, connected to the thermal ring extraction element and the background temperature statistics element, is used to filter candidate bright ring pixels based on the thermal ring response map and background temperature, and merge them to obtain several connected components. The geometric verification element, connected to the positioning element and the threshold segmentation element, is used to perform geometric screening of each connected domain based on the furnace opening center and radius to obtain a bright ring mask, which is the slag overflow area feature.

7. The method according to claim 2, characterized in that, The flue gas feature extraction module includes a CO concentration feature extraction unit, a CO2 concentration feature extraction unit, and an O2 concentration feature extraction unit arranged in parallel. The CO concentration feature extraction unit is used to extract the CO volume fraction at each time point based on the flue gas curve, and obtain the CO concentration change rate, which is the CO concentration feature. The CO2 concentration feature extraction unit is used to extract the CO2 volume fraction at each time point based on the flue gas curve, and obtain the CO2 concentration change rate, which is the CO2 concentration feature. The O2 concentration feature extraction unit is used to extract the O2 volume fraction at each time point based on the flue gas curve, and obtain the rate of change of O2 concentration, which is the O2 concentration feature.

8. The method according to claim 2, characterized in that, The sonar feature extraction module includes: The sonar signal extraction unit is used to extract single-point sonar signals from the sonar timing signal. The slag layer thickness growth rate calculation unit is connected to the output of the sonar signal extraction unit. It is used to obtain the slag layer thickness based on the sonar ranging values ​​of each single point in order to obtain sonar characteristics.

9. The method according to claim 1, characterized in that, The specific structure of the feature fusion layer includes, in sequence, a fusion module, a Concatenate node, a Transformer encoder, a global average pooling module, and a multi-task output module: The fusion module includes: The first fusion unit is used to fuse flame mode features based on flame brightness characteristics, spark density characteristics, and slag overflow area characteristics. The second fusion unit is used to fuse the flue gas modal characteristics based on the CO concentration characteristics, CO2 concentration characteristics and O2 concentration characteristics; The third fusion unit is used to fuse sonar modal features based on sonar characteristics to obtain sonar modal features; The Concatenate node is used to concatenate flame modal features, smoke modal features, and sonar modal features to obtain a multimodal feature vector; The Transformer encoder is used to capture global temporal features of multimodal feature vectors through a self-attention mechanism, thereby obtaining global contextual features. The global average pooling module is used to perform average pooling operations on global context features to reduce the computational complexity of the model and obtain global key features. The multi-task output module is used to learn in parallel the mapping relationship between global key features and prior splash risk and control variables, so as to synchronously output splash probability, gun position, oxygen flow increment and coolant increment based on global key features.

10. The method according to claim 9, characterized in that, The multi-task output module includes: The splash probability head is used to map the current splash probability based on global key features; The judgment head, connected to the splash probability head, is used to determine whether the splash probability is greater than a set threshold. If so, the global key features are input in parallel to the gun position head, oxygen flow increment head, and coolant increment head. The gun position head, connected to the judgment head, is used to map the gun position based on global key features; The oxygen flow rate increment head, connected to the decision head, is used to map the oxygen flow rate based on global key features; The coolant increment head, connected to the decision head, is used to map the coolant control quantity based on global key features.

Images