Converter slag splashing and protecting intelligent decision method and equipment

Abstract

The present application relates to a kind of converter slag splashing intelligent decision method and equipment, by building converter intelligent decision model, including: input layer, for receiving flame sequence image, audio time series signal, flue gas concentration sequence and furnace lining three-dimensional point cloud;Feature extraction layer is used to extract flame feature, audio feature, flue gas feature and point cloud feature;Feature fusion layer is used to fuse flame feature, audio feature, flue gas feature and point cloud feature, obtain gun height, nitrogen pressure and nitrogen flow rate;Output layer is used to output gun height, nitrogen pressure and nitrogen flow rate;Acquire current flame sequence image, audio time series signal, flue gas concentration sequence and furnace lining three-dimensional point cloud, input trained converter intelligent decision model, obtain gun height, nitrogen pressure and nitrogen flow rate.A kind of new technology of converter slag splashing is provided, which can fuse multi-source heterogeneous data, have end-to-end intelligent decision capability, support adaptive generalization and have predictive maintenance function.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology in iron and steel metallurgy, and in particular to an intelligent decision-making method and equipment for converter slag splashing protection. Background Technology

[0002] Slag splashing technology is a key process in modern converter steelmaking. Its core principle is that after tapping, high-pressure nitrogen gas is blown into the molten pool through an oxygen lance. The impact of the nitrogen jet splashes the molten slag onto the furnace lining surface, forming a dense protective layer rich in MgO and Fe2O3. This protective layer effectively isolates the high-temperature molten steel from direct contact with the furnace lining refractory material, significantly inhibiting lining erosion, thereby extending the furnace lining's service life, reducing refractory material consumption costs, and increasing converter operating rates.

[0003] Although slag splashing furnace protection technology has been widely used in steel companies both domestically and internationally, existing technical solutions still have the following prominent problems: (1) High dependence on experience and low degree of standardization of operation. Key control parameters of traditional slag splashing furnace protection processes, including oxygen lance height, nitrogen blowing pressure, and nitrogen blowing duration, are mainly set based on the operator's personal experience. Significant differences in operating habits between different shifts and operators lead to large fluctuations in furnace protection effects, making it difficult to achieve standardized and stable control of the process.

[0004] (2) The means of perception are limited and the understanding of furnace conditions is incomplete. Existing automated slag splashing control systems mostly employ single-modal detection technologies, such as visual detection systems based on furnace mouth flame image analysis, or sonar slag monitoring systems based on audio signal processing. These single-modal solutions can only capture local features of the furnace interior and are insufficient for comprehensive and accurate perception of complex slag flow, furnace lining erosion, and furnace bottom deposition.

[0005] (3) Lack of predictive ability and lagging control model Existing control systems are mostly based on real-time detection feedback control, which adjusts parameters according to the currently detected furnace condition. This reactive control strategy cannot predict the changing trends of slag condition and the development pattern of furnace lining erosion in advance, making it difficult to achieve proactive optimization and advanced planning of the slag splashing process.

[0006] (4) Insufficient adaptability, making it difficult to cope with complex working conditions. The service life of a converter is divided into three stages: initial, middle, and late, with significant differences in the lining condition at each stage. Furthermore, different furnace types (such as top-and-bottom blown converters and top-blown converters) and different steel grades (such as carbon steel, low-alloy steel, and stainless steel) have different requirements for slag splashing processes. Existing technical solutions lack adaptive adjustment capabilities for different operating conditions, and their model generalization performance is limited.

[0007] In response to the above problems, some domestic patented technologies have attempted to improve them, but significant shortcomings still exist: The patent number is CN115323104A, and the patent title is an invention patent for an automatic slag splashing protection method for converters. It proposes a multi-step progressive slag splashing protection method, but the scheme is based on fixed empirical rules for staged blowing and lacks the ability to intelligently perceive and dynamically adjust the real-time status inside the furnace. The application number is CN202310991292.9, and the patent is titled Automatic Slag Splashing Control Method, Device, Computer Equipment and Storage Medium for Converter. It uses a recurrent neural network (RNN) to predict slag composition, but it only relies on historical process data for modeling and does not integrate multimodal perception information such as flame and sonar, thus limiting the model's generalization ability. The patent number is CN115896387B, and the patent name is a control method and system for automatic slag splashing in a converter. It switches the slag splashing mode based on liquid level detection, but the sensing feature dimension is single and cannot accurately depict the complex state inside the furnace. The patent, with patent number CN111074038A, is titled "A Method for Targeted Slag Splashing Gun Position Control in Converter Slag Splashing Protection." It uses laser thickness measurement technology for furnace lining thickness feedback control, which is a post-compensation mechanism and lacks predictive analysis capabilities for the evolution of the furnace lining state.

[0008] In summary, existing technical solutions have significant shortcomings in data fusion, intelligent decision-making, adaptive generalization, and predictive maintenance, making it difficult to meet the development needs of modern converter steelmaking for intelligence, precision, and efficiency. Therefore, there is an urgent need to develop a new converter slag splashing protection technology that can integrate multi-source heterogeneous data, possess end-to-end intelligent decision-making capabilities, support adaptive generalization, and have predictive maintenance functions. Summary of the Invention

[0009] Based on this, the purpose of this application is to provide a smart decision-making method and device for converter slag splashing protection, so as to solve at least one of the technical problems mentioned in the background art.

[0010] Firstly, this application provides an intelligent decision-making method for converter slag splashing protection, including: A converter intelligent decision-making model was constructed and trained, taking flame sequence images, audio time-series signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining as inputs, and lance position height, nitrogen pressure, and nitrogen blowing flow rate as outputs, including: The input layer is used to receive flame sequence images, audio timing signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining. The feature extraction layer includes a flame feature extraction module, an audio feature extraction module, a flue gas feature extraction module, and a point cloud feature extraction module arranged in parallel. These modules are used to extract flame features, audio features, flue gas features, and point cloud features based on flame sequence images, audio time-series signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining, respectively. The feature fusion layer is used to fuse flame features, audio features, smoke features and point cloud features to obtain multimodal fusion features, so as to make decisions for the next moment based on the multimodal fusion features, including gun position height, nitrogen pressure and nitrogen flow rate; The output layer is used to output the gun position height, nitrogen pressure, and nitrogen flow rate. The system collects current flame sequence images, audio timing signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining. These data are then input into the trained converter intelligent decision-making model to obtain the lance height, nitrogen pressure, and nitrogen flow rate at the next moment.

[0011] Furthermore, the flame feature extraction module is used to extract flame features from the flame sequence image; the flame features include flame color features, flame shape features, and flame pulsation features; The audio feature extraction module is used to extract audio features based on the audio time sequence signal; the audio features include foam slag thickness features and slag layer stability features; The flue gas feature extraction module is used to extract flue gas features based on the flue gas concentration sequence; the flue gas features include decarbonization rate features and slag oxidizability features; The point cloud feature extraction module is used to extract point cloud features based on the three-dimensional point cloud of the furnace lining. The point cloud features include the furnace lining thickness distribution features and the location features of erosion hot spots.

[0012] Furthermore, the flame feature extraction module includes: The image segmentation unit is used to divide the flame sequence image into several image blocks to obtain an image block sequence. The position encoding unit, connected to the image block unit, is used to add position encoding information to each image block and preserve spatial positional relationships; The flame feature extraction unit, connected to the position encoding unit, is used to introduce a priori flame attention mask to enhance feature extraction of the central and edge regions of the furnace opening in the image block sequence, thereby obtaining flame color features, flame shape features, and flame pulsation features.

[0013] Furthermore, the flame feature extraction unit includes flame color feature extraction elements, flame shape feature extraction elements, and flame pulsation feature extraction elements arranged in parallel; The flame color feature extraction element is used to extract the color distribution information of each image block in each frame of the flame sequence image to obtain the flame color features; The flame shape feature extraction element is used to extract the flame outline edge and overall morphology information from each frame of the flame sequence image to obtain the flame shape feature. The flame pulsation feature extraction element is used to perform temporal difference operations on the flame color and shape features of a series of consecutive flame sequence images to calculate the amplitude and frequency information of the flame dynamic changes and obtain the flame pulsation features.

[0014] Furthermore, the audio feature extraction module includes a time-frequency transformation unit, a Mel spectrum generation unit, and an audio feature extraction unit connected in sequence; The time-frequency transformation unit is used to perform short-time Fourier transform on the audio signal to obtain a time-frequency representation; The Mel spectrum generation unit, connected to the time-frequency conversion unit, is used to convert the time-frequency representation into a Mel spectrum diagram; An audio feature extraction unit, connected to a Mel spectrum generation unit, includes: Metallurgical noise suppression element, used to enhance the Mel spectrum to suppress electromagnetic interference and mechanical vibration noise, resulting in an enhanced spectrum; The foam slag thickness feature extraction element is connected to the metallurgical noise suppression element. It is used to perform spectral encoding on the enhanced spectrum map, extract the spectral energy attenuation feature, and obtain the foam slag thickness feature. The slag layer stability feature extraction element is connected to the metallurgical noise suppression element. It is used to perform spectral encoding on the enhanced spectrum map, extract the spectral time variance and entropy value change information, and obtain the slag layer stability features.

[0015] Furthermore, the flue gas feature extraction module includes a one-dimensional convolutional feature extraction unit, a temporal position encoding unit, and a flue gas feature extraction unit connected in sequence. A one-dimensional convolutional feature extraction unit is used to extract local features from the flue gas concentration sequence to obtain the flue gas temporal features; the flue gas concentration sequence includes CO, CO2 ... 2、 O2 and N2 concentration sequences; The temporal position encoding unit, connected to the one-dimensional convolutional feature extraction unit, is used to add position encoding to the temporal features of flue gas and preserve the temporal order information; The flue gas feature extraction unit, connected to the time-series position encoding unit, includes a decarbonization rate feature extraction element and a slag oxidation feature extraction element arranged in parallel. The decarbonization rate feature extraction element is used to perform global dependency modeling on the time-series features of flue gas with location encoding, calculate the rate of change of the CO to CO2 concentration ratio, and obtain the decarbonization rate features. The slag oxidation feature extraction element is used to perform global dependency modeling on the temporal features of flue gas with location coding, extract the temporal trend of O2 concentration and N2 concentration compensation information, and obtain the slag oxidation characteristics.

[0016] Furthermore, the point cloud feature extraction module includes: The point cloud sampling unit is used to downsample a set area of ​​the input furnace lining point cloud to obtain several sampling points and their three-dimensional coordinates. The sphere query grouping unit, connected to the point cloud sampling unit, is used to construct a local neighborhood sphere centered on each sampling point; The hierarchical feature learning unit, connected to the ball query grouping unit, is used to perform local-to-global hierarchical feature learning based on each local neighborhood ball to obtain point-level features at each level and global aggregated features. The point cloud feature extraction unit includes parallelly arranged furnace lining thickness distribution feature extraction elements and erosion hotspot location feature extraction elements: The furnace lining thickness distribution feature extraction element is connected to the hierarchical feature learning unit to perform thickness statistical distribution analysis on the global aggregated features, extract the mean and variance information of the thickness of each region of the furnace lining, and obtain the furnace lining thickness distribution features. The erosion hotspot location feature extraction element is connected to the point cloud sampling unit and the hierarchical feature learning unit. It is used to perform local anomaly detection based on each sampling point and its three-dimensional coordinates and each layer of point-level features, and to extract the center coordinates and erosion degree information of the region with a thickness below a set threshold, thereby obtaining the erosion hotspot location features.

[0017] Furthermore, the feature fusion layer includes: The cross-modal attention fusion module is used to map flame features, audio features, smoke features and point cloud features to the same feature dimension through linear projection, and calculate cross-attention weights with flame features as the query and audio features, smoke features and point cloud features as the key to obtain multimodal fused features; The temporal causal reasoning module is used to construct a causal graph based on known knowledge in the metallurgical field, perform counterfactual reasoning based on the causal graph, constrain the information transmission path between features, and obtain causal enhanced features. The decision generation module, connected to the cross-modal attention fusion module and the temporal causal reasoning module, is used to synchronously output the gun position height, nitrogen pressure, and nitrogen flow rate based on multimodal fusion features and causal enhancement features.

[0018] Furthermore, the decision generation module includes: The slag state classification unit is used to map slag state classification based on multimodal fusion features and causal enhancement features; A furnace lining erosion identification unit is used to map the furnace lining erosion stage based on multimodal fusion features and causal enhancement features; The slag splashing mode selection unit is used to map slag splashing mode selection based on multimodal fusion features and causal enhancement features; The decision generation unit is used to obtain the next moment's decision based on multimodal fusion characteristics, slag state classification, furnace lining erosion stage identification, and slag splashing mode selection, including lance position height, nitrogen pressure, and nitrogen blowing flow rate.

[0019] Secondly, this application 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 the intelligent decision-making method for converter slag splashing protection as described in any one of the first aspects.

[0020] This application provides a converter slag splashing protection intelligent decision-making method and device. It constructs and trains a converter intelligent decision-making model that takes flame sequence images, audio time-series signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining as inputs, and lance height, nitrogen pressure, and nitrogen blowing flow rate as outputs. The model includes: an input layer for receiving flame sequence images, audio time-series signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining; and a feature extraction layer comprising a flame feature extraction module, an audio feature extraction module, a flue gas feature extraction module, and a point cloud feature extraction module arranged in parallel, used to extract features based on the flame sequence images, audio time-series signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining, respectively. This paper presents a new converter slag splashing protection technology that can integrate multi-source heterogeneous data, possesses end-to-end intelligent decision-making capabilities, supports adaptive generalization, and has predictive maintenance functions. A feature fusion layer is used to fuse these features to obtain multi-modal fusion features, which are then used to make decisions for the next time step, including lance height, nitrogen pressure, and nitrogen flow rate. An output layer is used to output the lance height, nitrogen pressure, and nitrogen flow rate. The technology involves acquiring current flame sequence images, audio time-series signals, flue gas concentration sequences, and the 3D point cloud of the furnace lining, and inputting these into a trained converter intelligent decision-making model to obtain the lance height, nitrogen pressure, and nitrogen flow rate for the next time step. Attached Figure Description

[0021] Figure 1 This is a flowchart of the intelligent decision-making method for converter slag splashing protection according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the intelligent decision-making model for converters according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a preferred structure of the converter intelligent decision-making model according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a preferred structure of the cross-modal attention fusion module according to an embodiment of the present invention; Figure 5This is a schematic diagram of a preferred structure of the decision generation module according to an embodiment of the present invention; Figure 6 This is a preferred flowchart of the intelligent decision-making method for converter slag splashing protection according to an embodiment of the present invention. Detailed Implementation

[0022] 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.

[0023] 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 execution order of the method. Those skilled in the art will understand that anything that does not violate the inventive concept should be included within the scope of protection of the present invention.

[0024] like Figure 1 As shown, this invention provides an intelligent decision-making method for converter slag splashing protection based on a multimodal large model: S1: Construct and train a converter intelligent decision-making model that takes flame sequence images, audio time-series signals, flue gas concentration sequences, and 3D point clouds of the furnace lining as inputs, and lance height, nitrogen pressure, and nitrogen blowing flow rate as outputs. Figure 2 As shown, it includes: The input layer is used to receive flame sequence images, audio timing signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining. The feature extraction layer includes a flame feature extraction module, an audio feature extraction module, a flue gas feature extraction module, and a point cloud feature extraction module arranged in parallel. These modules are used to extract flame features, audio features, flue gas features, and point cloud features based on flame sequence images, audio time-series signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining, respectively. The feature fusion layer is used to fuse flame features, audio features, smoke features and point cloud features to obtain multimodal fusion features, so as to make decisions for the next moment based on the multimodal fusion features, including gun position height, nitrogen pressure and nitrogen flow rate; The output layer is used to output the gun position height, nitrogen pressure, and nitrogen flow rate.

[0025] Specifically, the system can simultaneously receive four different types of multi-source heterogeneous signals through the input layer: flame sequence images, audio time-series signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining. This enables comprehensive and three-dimensional monitoring of the converter slag splashing protection process, improving the comprehensiveness and reliability of detection. It simultaneously captures the furnace state from four dimensions: visual, acoustic, chemical, and geometric, avoiding blind spots caused by a single information source. This expands traditional "process monitoring" to joint monitoring of "equipment status - smelting process," significantly improving the reliability and predictability of slag splashing protection decisions and providing a rich and complementary data foundation for subsequent feature extraction. Then, the feature extraction layer extracts key features of the flame, audio, flue gas, and furnace lining point clouds from the massive and complex raw data, effectively filtering out irrelevant background noise and... Interference signals are significantly reduced, improving the signal-to-noise ratio and decreasing the computational complexity of subsequent processing. A feature fusion layer then deeply fuses the features of the four modes to obtain a multimodal fusion feature that comprehensively characterizes the furnace state. Intelligent decision-making is based on this feature, avoiding the limitations of relying on a single information source. Through mutual verification and supplementation of multimodal information, the impact of single sensor anomalies or data gaps on the final decision is reduced, thereby obtaining optimal control parameters (lance height, nitrogen pressure, and nitrogen flow rate), achieving integration from state perception to intelligent decision-making. Finally, the output layer outputs precise decision commands, providing scientific and quantitative setpoints for slag splashing and furnace protection operations. Parameters with clear physical meaning facilitate operator understanding and intervention, improving the stability of the production process and the furnace lining life.

[0026] In a preferred embodiment, the structure of the converter intelligent decision-making model can be selected as follows: Figure 3 As shown, the system comprises four core layers: a multimodal data acquisition layer, a multimodal feature encoding layer, a cross-modal fusion and inference layer, and a decision output and execution layer. The multimodal data acquisition layer is responsible for acquiring furnace flame images, sonar audio signals, flue gas composition data, and furnace lining laser point cloud data. The multimodal feature encoding layer contains four independent encoders, each processing one of the furnace flame image, sonar audio signal, flue gas composition data, or furnace lining laser point cloud data. The cross-modal fusion and inference layer is the core of the multimodal large model. The decision output and execution layer is responsible for outputting slag splash mode selection, lance position curve generation, and nitrogen parameter settings.

[0027] Preferably, the flame feature extraction module is used to extract flame features from the flame sequence image; the flame features include flame color features, flame shape features, and flame pulsation features; The audio feature extraction module is used to extract audio features based on the audio time sequence signal; the audio features include foam slag thickness features and slag layer stability features; The flue gas feature extraction module is used to extract flue gas features based on the flue gas concentration sequence. These features include decarbonization rate characteristics and slag oxidizability characteristics. For example, the input to the flue gas feature extraction module is the CO, CO2, O2, and N2 concentration sequence output from the flue gas analyzer, with a sampling period of 1 second. Preprocessing uses one-dimensional convolutional feature extraction combined with positional encoding. The core network is a 6-layer temporal Transformer with 256 hidden feature dimensions. The output is the furnace reaction state characteristics, with a feature dimension of 256, representing the decarbonization rate and slag oxidizability.

[0028] The point cloud feature extraction module is used to extract point cloud features based on the three-dimensional point cloud of the furnace lining. The point cloud features include the furnace lining thickness distribution features and the location features of erosion hot spots.

[0029] This embodiment provides specific implementation examples of flame features, audio features, flue gas features, and point cloud features. The specific selection of these four types of features corresponds precisely to the four core dimensions of slag splashing furnace protection effect evaluation and control: Flame characteristics (color, shape, pulsation): During the slag splashing process, the viscosity and fluidity of the slag directly affect the shape of the flame at the furnace mouth. The flame color reflects the temperature distribution inside the furnace. The expansion or contraction of the flame shape is related to the slag layer coverage. The pulsation frequency of the flame is related to the stirring state of the molten pool and the formation and rupture of foam slag. Extracting these three types of characteristics allows for an intuitive and non-contact assessment of the dynamic effect of slag splashing from a visual perspective, making up for the subjectivity and lag of traditional manual visual inspection.

[0030] Audio characteristics (foam slag thickness, slag layer stability): When slag is splashed in the converter, the impact of high-pressure nitrogen gas on the molten pool will generate specific acoustic signals. The thickness and stability of the foam slag will significantly change the frequency and energy distribution of the sound. By extracting these characteristics through audio analysis, the physical state of the slag layer can be quantified from an acoustic perspective, providing a key basis for judging whether the slag splashing is uniform and whether the slag layer has reached the ideal thickness.

[0031] Flue gas characteristics (decarburization rate, slag oxidizability): Flue gas composition is a direct product of chemical reactions in the furnace. The decarburization rate reflects the activity level of the molten pool, while the slag oxidizability directly affects the melting point and viscosity of the slag. It is a key chemical factor that determines whether splashed slag can effectively adhere to the furnace lining. By analyzing flue gas composition in real time, the reaction process and slag properties in the furnace can be accurately grasped from a chemical perspective, providing data support for dynamically adjusting the slag splashing process parameters.

[0032] Point cloud features (furnace lining thickness distribution, erosion hotspot location): By performing three-dimensional scanning of the furnace lining, the real-time thickness distribution of the furnace lining and the most severely eroded areas (hotspots) can be obtained directly and accurately. Extracting these features allows for the final verification of the slag splashing effect from a geometric perspective, and precise location of areas requiring key protection, thus achieving closed-loop control from "effect evaluation" to "target-oriented".

[0033] Therefore, the integration and specific selection of these four types of features construct a multi-dimensional, high-precision intelligent perception and decision-making system for slag splashing furnace protection, which integrates "visual, acoustic, chemical, and geometric" features. By comprehensively capturing furnace state information, it achieves precise control and optimization of the slag splashing process, ultimately extending the converter's service life and reducing production costs.

[0034] The following describes preferred embodiments for the flame feature extraction module (extracting flame color, shape, and pulsation features); the audio feature extraction module (extracting foam slag thickness and slag layer stability features); the flue gas feature extraction module (extracting decarburization rate and slag oxidizability features); and the point cloud feature extraction module (extracting furnace lining thickness distribution and erosion hotspot location features). These embodiments are provided but are not limited to these. The key to this invention lies in fusing and analyzing flame, audio, flue gas, and point cloud data based on the slag splashing protection mechanism, and further selecting the aforementioned specific features for finer-grained analysis and segmentation. Any method can be used to extract these features.

[0035] Preferably, the flame feature extraction module includes: The image segmentation unit is used to divide the flame sequence image into several image blocks to obtain an image block sequence. The position encoding unit, connected to the image block unit, is used to add position encoding information to each image block and preserve spatial positional relationships; The flame feature extraction unit, connected to the position encoding unit, is used to introduce a priori flame attention mask to enhance feature extraction of the central and edge regions of the furnace opening in the image block sequence, thereby obtaining flame color features, flame shape features, and flame pulsation features.

[0036] More specifically, the flame feature extraction module takes as input an image of the furnace flame captured by a high-definition industrial camera, with a resolution of 1920 x 1080 and a frame rate of 25 fps. In the preprocessing stage, the image is divided into 16 x 16 pixel blocks, and positional encoding is added. The core network employs a 12-layer Transformer Encoder with a hidden feature dimension of 768 and 12 attention heads. The output is a flame state feature vector with a feature dimension of 768, representing the flame's color, shape, and pulsation characteristics. Furthermore, by introducing a priori attention mask based on flame physics, the feature extraction of the central region of the furnace opening (i.e., the flame core) and the edge region (i.e., the distribution of sparks) is enhanced.

[0037] In this embodiment, the flame image is segmented into a sequence of image blocks by an image segmentation unit, laying the foundation for subsequent steps and reducing computational complexity. Then, a position encoding unit adds positional information to each image block to compensate for the loss of spatial relationships caused by the segmentation operation, ensuring that the model can understand the overall structure of the flame. Furthermore, a priori flame attention mask is introduced by a flame feature extraction unit. This mask, based on metallurgical expert knowledge, guides the model to focus on key areas such as the furnace center (reaction core area) and the edge (slag layer coverage area), suppressing background interference. This allows for more targeted extraction of flame color, shape, and pulsation features closely related to the slag splashing effect, improving the effectiveness of feature extraction and the interpretability of the model.

[0038] Preferably, the flame feature extraction unit includes a flame color feature extraction element, a flame shape feature extraction element, and a flame pulsation feature extraction element arranged in parallel; The flame color feature extraction element is used to extract the color distribution information of each image block in each frame of the flame sequence image to obtain the flame color features; The flame shape feature extraction element is used to extract the flame outline edge and overall morphology information from each frame of the flame sequence image to obtain the flame shape feature. The flame pulsation feature extraction element is used to perform temporal difference operations on the flame color and shape features of a series of consecutive flame sequence images to calculate the amplitude and frequency information of the flame dynamic changes and obtain the flame pulsation features.

[0039] Specifically, the system can optionally analyze the color distribution of image blocks using flame color feature extraction elements to accurately quantify the visual representation of the furnace temperature field and chemical composition (such as FeO content); capture the outline and shape of the flame using flame shape feature extraction elements to determine whether the viscosity and fluidity of the slag are suitable; and perform temporal difference analysis on the features of consecutive frames using flame pulsation feature extraction elements to capture the dynamic changes of the flame, which better reflects the real-time agitation state of the molten pool and the stability of the foamed slag than static image analysis. These three elements, set up in parallel, comprehensively analyze flame information from "static attributes" to "dynamic behavior," providing rich visual evidence for evaluating slag splashing effects.

[0040] Preferably, the audio feature extraction module includes a time-frequency transformation unit, a Mel spectrum generation unit, and an audio feature extraction unit connected in sequence; The time-frequency transformation unit is used to perform short-time Fourier transform on the audio signal to obtain a time-frequency representation; The Mel spectrum generation unit, connected to the time-frequency conversion unit, is used to convert the time-frequency representation into a Mel spectrum diagram; An audio feature extraction unit, connected to a Mel spectrum generation unit, includes: Metallurgical noise suppression element, used to enhance the Mel spectrum to suppress electromagnetic interference and mechanical vibration noise, resulting in an enhanced spectrum; The foam slag thickness feature extraction element is connected to the metallurgical noise suppression element. It is used to perform spectral encoding on the enhanced spectrum map, extract the spectral energy attenuation feature, and obtain the foam slag thickness feature. The slag layer stability feature extraction element is connected to the metallurgical noise suppression element. It is used to perform spectral encoding on the enhanced spectrum map, extract the spectral time variance and entropy value change information, and obtain the slag layer stability features.

[0041] More specifically, the audio feature extraction module takes as input the furnace audio signal collected by a sonar sensor, with a sampling rate of 16kHz and a 10-second sliding window. In the preprocessing stage, a Mel spectrogram is generated using a short-time Fourier transform. The core network is an AST-based spectrogram encoder with an 8-layer Transformer and a hidden feature dimension of 512. The output is the slag state audio features, with a feature dimension of 512, characterizing the thickness of the foamy slag and the stability of the slag layer. Furthermore, a spectrum enhancement module specifically designed for metallurgical noise is used to suppress electromagnetic interference and mechanical vibration noise.

[0042] In this embodiment, the original audio signal can be converted into a Mel spectrum using a time-frequency conversion unit and a Mel spectrum generation unit to simulate the nonlinear perception characteristics of human ears for sound frequencies. This allows for more effective extraction of acoustic features related to the furnace state. Then, metallurgical noise suppression elements are used to specifically filter out electromagnetic and mechanical noise specific to the converter workshop, significantly improving the signal-to-noise ratio of the audio signal. Furthermore, the attenuation mode of spectral energy is analyzed using a foam slag thickness feature extraction element. Since foam slag of different thicknesses has different absorption and reflection characteristics for sound waves, non-contact slag thickness measurement is achieved. The temporal variance and entropy of the spectrum are calculated using a slag layer stability feature extraction element to quantify the degree of fluctuation in the slag layer state. The larger the variance and entropy, the more unstable the slag layer, and the potentially poorer the slag splashing effect.

[0043] Preferably, the flue gas feature extraction module includes a one-dimensional convolutional feature extraction unit, a temporal position encoding unit, and a flue gas feature extraction unit connected in sequence; A one-dimensional convolutional feature extraction unit is used to extract local features from the flue gas concentration sequence to obtain the flue gas temporal features; the flue gas concentration sequence includes CO, CO2 ... 2、 O2 and N2 concentration sequences; The temporal position encoding unit, connected to the one-dimensional convolutional feature extraction unit, is used to add position encoding to the temporal features of flue gas and preserve the temporal order information; The flue gas feature extraction unit, connected to the time-series position encoding unit, includes a decarbonization rate feature extraction element and a slag oxidation feature extraction element arranged in parallel. The decarbonization rate feature extraction element is used to perform global dependency modeling on the time-series features of flue gas with location encoding, calculate the rate of change of the CO to CO2 concentration ratio, and obtain the decarbonization rate features. The slag oxidation feature extraction element is used to perform global dependency modeling on the temporal features of flue gas with location coding, extract the temporal trend of O2 concentration and N2 concentration compensation information, and obtain the slag oxidation characteristics.

[0044] More specifically, the input to the point cloud feature extraction module is a 3D point cloud of the furnace lining scanned by a laser thickness gauge, approximately 100,000 points per furnace. Preprocessing employs farthest point sampling combined with sphere query grouping. The core network uses hierarchical feature learning combined with global feature aggregation. The output is the furnace lining erosion state feature, with a feature dimension of 1024, representing the furnace lining thickness distribution and erosion hotspot locations.

[0045] In this embodiment, a one-dimensional convolutional feature extraction unit can be used to capture the fluctuation pattern of flue gas concentration within a short period of time (local window), such as rapid increases or decreases in concentration. A temporal position encoding unit preserves the temporal sequence information, enabling the model to understand the causal relationship of "what happens first and what happens later." Then, the rate of change of the CO / CO2 ratio is calculated using a decarbonization rate feature extraction element. This ratio is a classic indicator for measuring the intensity of the decarbonization reaction, and its rate of change can more sensitively reflect the dynamic changes in the reaction rate. The long-term trend of O2 concentration is analyzed using a slag oxidation feature extraction element, and N2 concentration is used for compensation (excluding interference from factors such as system air leakage), thereby more accurately determining the oxidative level of the slag and providing a basis for controlling slag viscosity.

[0046] Preferably, the point cloud feature extraction module includes: The point cloud sampling unit is used to downsample a set area of ​​the input furnace lining point cloud to obtain several sampling points and their three-dimensional coordinates. The sphere query grouping unit, connected to the point cloud sampling unit, is used to construct a local neighborhood sphere centered on each sampling point; The hierarchical feature learning unit, connected to the ball query grouping unit, is used to perform local-to-global hierarchical feature learning based on each local neighborhood ball to obtain point-level features at each level and global aggregated features. The point cloud feature extraction unit includes parallelly arranged furnace lining thickness distribution feature extraction elements and erosion hotspot location feature extraction elements: The furnace lining thickness distribution feature extraction element is connected to the hierarchical feature learning unit to perform thickness statistical distribution analysis on the global aggregated features, extract the mean and variance information of the thickness of each region of the furnace lining, and obtain the furnace lining thickness distribution features. The erosion hotspot location feature extraction element is connected to the point cloud sampling unit and the hierarchical feature learning unit. It is used to perform local anomaly detection based on each sampling point and its three-dimensional coordinates and each layer of point-level features, and to extract the center coordinates and erosion degree information of the region with a thickness below a set threshold, thereby obtaining the erosion hotspot location features.

[0047] Specifically, the original point cloud can be preprocessed and local regions can be divided using point cloud sampling units and sphere query grouping units to provide structured input for subsequent feature learning. Local neighborhood information can be aggregated layer by layer through hierarchical feature learning units to obtain point cloud features that can reflect both local details and global structure. Then, the global features can be statistically analyzed using furnace lining thickness distribution feature extraction elements to obtain the overall wear status (mean) and uniformity (variance) of the furnace lining, evaluate the overall effect of slag splashing protection, and perform local anomaly detection using erosion hotspot location feature extraction elements to accurately locate the weakest area of ​​the furnace lining, providing a "key protection" target for the next slag splashing operation, and realizing a personalized and precise furnace protection strategy.

[0048] Preferably, the feature fusion layer includes: The cross-modal attention fusion module maps flame features, audio features, smoke features, and point cloud features to the same feature dimension through linear projection. It then calculates cross-attention weights using flame features as the query and audio, smoke, and point cloud features as keys to obtain multimodal fused features. For example, such as... Figure 4 As shown, the cross-modal attention fusion module first maps the features of each modality to a unified feature dimension of 512 through linear projection. Then, using visual features as queries and other modalities as keys, it calculates the correlation weights between modalities. Specifically, it calculates the attention weights from visual to audio, visual to smoke, and visual to point cloud. Finally, it performs multimodal feature fusion, adding the weighted features of visual features to those of other modalities. Through 8 layers of cross-modal Transformer iterative optimization, it outputs fused features with a feature dimension of 512.

[0049] The temporal causal reasoning module constructs a causal graph based on known knowledge in the metallurgical field. It then performs counterfactual reasoning based on this graph, constraining the information transmission paths between features to obtain causal-enhanced features. For example, the module introduces a causal inference mechanism to model the temporal impact of process parameters. First, it constructs a causal graph based on metallurgical knowledge, including causal chains from slag conditioner to slag viscosity to slag splashing adhesion rate, and from nitrogen pressure to slag drop velocity to furnace lining impact. Then, it uses a structural causal model to perform counterfactual reasoning, evaluating the potential outcomes of different decision paths. Finally, a causal Transformer is employed to ensure the model focuses on causally relevant features rather than spurious correlations.

[0050] The decision generation module, connected to the cross-modal attention fusion module and the temporal causal reasoning module, is used to synchronously output the gun position height, nitrogen pressure, and nitrogen blowing flow rate based on multimodal fusion features and causal enhancement features. For example, ... Figure 5 As shown, the decision generation module realizes furnace condition recognition and decision generation based on fused features. Furnace condition recognition includes three sub-tasks: slag state classification into four categories: normal foamy slag, excessively thin, excessively sticky, and re-drying; furnace lining erosion stage identification into initial furnace service, mid-term furnace service, and late furnace service; and slag splashing mode selection into normal maintenance, furnace bottom repair, and furnace bottom expansion treatment. Decision generation adopts an autoregressive generation method to output a sequence of control parameters. The inputs are fused features, the current time, and the generated decision sequence, and the outputs are the decisions for the next time step, including lance height, nitrogen pressure, and nitrogen blowing flow rate. The generation process uses a 6-layer Transformer Decoder, with an output feature dimension of 3.

[0051] Specifically, a cross-modal attention fusion module can be used to achieve deep fusion of features from four modalities. Flame features are used as the "query" because the flame is the most direct reflection of the furnace's state. The model learns how other modalities (audio, flue gas, point cloud) "explain" or "supplement" the state presented by the flame, thus obtaining more correlated fused features. A temporal causal reasoning module introduces prior knowledge from the metallurgical field (such as "a fast decarburization rate leads to increased flame pulsation"), constructs a causal graph, and performs counterfactual reasoning. This effectively avoids the model learning spurious correlations in the data, making the decision-making process more consistent with physical and chemical laws, enhancing the model's robustness and interpretability. Finally, the decision generation module generates optimal control parameters based on the fusion of data-driven and knowledge-driven features, achieving truly intelligent decision-making.

[0052] Preferably, the decision generation module includes: The slag state classification unit is used to map slag state classification based on multimodal fusion features and causal enhancement features; A furnace lining erosion identification unit is used to map the furnace lining erosion stage based on multimodal fusion features and causal enhancement features; The slag splashing mode selection unit is used to map slag splashing mode selection based on multimodal fusion features and causal enhancement features; The decision generation unit is used to obtain the next moment's decision based on multimodal fusion characteristics, slag state classification, furnace lining erosion stage identification, and slag splashing mode selection, including lance position height, nitrogen pressure, and nitrogen blowing flow rate.

[0053] Specifically, the complex furnace state can be decomposed into several key sub-tasks with clear physical meanings through a slag state classification unit, a furnace lining erosion identification unit, and a slag splashing mode selection unit. These sub-tasks include whether the slag is "sparse" or "thick," whether the furnace lining is in the "initial erosion" or "severe erosion" stage, and whether to select the "large flow rate coverage" or "small flow rate point-filling" mode. This step-by-step decision-making strategy makes the entire decision-making process clearer and more organized. Finally, the decision generation unit integrates the judgment results of all sub-tasks and the original fusion features to generate the final, refined control commands (gun position, pressure, flow rate), realizing a complete decision-making chain from macroscopic state judgment to microscopic parameter execution.

[0054] S2: Collect the current flame sequence image, audio timing signal, flue gas concentration sequence and furnace lining 3D point cloud, input them into the trained converter intelligent decision-making model, and obtain the decision for the next moment, lance position height, nitrogen pressure and nitrogen blowing flow rate.

[0055] Specifically, at set intervals, the flame sequence image, audio timing signal, flue gas concentration sequence, and three-dimensional point cloud of the converter lining in the previous time interval can be collected and input into the trained converter intelligent decision model to obtain the parameters that each device needs to adjust in the next time interval, i.e., the decision for the next moment, including lance height, nitrogen pressure, and nitrogen blowing flow rate.

[0056] In a preferred embodiment of the present invention, the intelligent decision-making method for converter slag splashing protection is as follows: Figure 6 As shown, it includes: 1. Large-scale pre-training A pre-trained dataset for converter metallurgy was constructed. The dataset includes 5 million furnace flame images from 3 years of production data from 10 converters; 200,000 hours of sonar audio covering multiple furnace types and steel grades; 1 billion time-series records of flue gas composition from the L2 system historical database; 500,000 furnace lining point clouds from laser thickness gauge archives; and 1 million pairs of process parameters and effects from MES system labeled data.

[0057] The pre-training task design includes three tasks: Masked modality modeling, which randomly masks a certain modality input and trains the model based on reconstructions from other modalities; Cross-modal contrastive learning, which brings features from different modalities closer together at the same time step and pushes features from different time steps further apart; and Temporal prediction, which predicts the furnace condition state in the next few steps.

[0058] 2. Downstream task fine-tuning Efficient parameter fine-tuning is performed for specific converters. The adapter fine-tuning method inserts lightweight adapter modules between Transformer layers, freezing the backbone network. The cue learning method learns continuous cue vectors to guide the pre-trained model to adapt to specific converter types. The dynamic expert hybrid method trains dedicated expert networks for different converter service periods and dynamically schedules them through a gating network.

[0059] 3. System Workflow The system workflow consists of five steps.

[0060] The first step is data acquisition and preprocessing, which is performed 5 minutes before the end of tapping. This involves acquiring static parameters such as the composition, temperature, and scrap ratio of the molten iron for the current heat; acquiring dynamic data such as flame images, sonar signals, and flue gas composition during the blowing process; and acquiring laser point clouds of the furnace lining to generate a current erosion status map.

[0061] The second step is multimodal feature extraction, which is performed 2 minutes before the steel tapping is completed. Each modal encoder extracts features in parallel to generate a fused furnace condition representation vector.

[0062] The third step is the intelligent selection of slag splashing mode, which is executed one minute before steel tapping is completed. Based on the fusion characteristics, the slag condition and the degree of furnace lining erosion are judged, and the slag splashing mode is selected, including normal maintenance, furnace bottom repair, and furnace bottom expansion treatment, and the initial lance position and nitrogen parameters are generated.

[0063] The fourth step is dynamic decision generation, which begins slag splashing at time T0. Model inference is performed every 2 seconds to generate the gun position height, nitrogen pressure, and nitrogen flow rate for the next moment, which are then output to the PLC actuator.

[0064] The fifth step is online learning and optimization, which is performed after the slag splashing is completed. This involves collecting slag splashing effect evaluation data, including slag thickness and uniformity, calculating the deviation between the actual effect and the prediction, and updating the model parameters online.

[0065] 4. Summary of Innovation Points First, it is the first time that a multimodal large model has been introduced into the field of converter slag splashing protection, realizing unified representation and fusion decision-making of heterogeneous industrial data.

[0066] Second, we propose an industrial-scale large-scale model deployment paradigm that combines a pre-trained base with downstream task fine-tuning to solve the model adaptation problem for different furnace types and service periods.

[0067] Third, we designed a cross-modal attention fusion mechanism to achieve efficient fusion of four types of heterogeneous data: flame, soundprint, smoke, and point cloud.

[0068] Fourth, introduce causal reasoning and time-series prediction capabilities to enable the system to have causal cognition of the impact of process parameters and the ability to plan ahead.

[0069] Fifth, construct an end-to-end decision generation framework to directly output a continuous sequence of control parameters, replacing the traditional segmented control logic.

[0070] 5. Beneficial effects Traditional experience-based control relies on manual observation and a single-modal approach, with decisions depending on empirical rules. This results in poor adaptability, achieving a slag uniformity of 75%, a furnace life extension effect at the baseline level, and nitrogen consumption at the baseline level. Existing automated systems use 2 to 3 types of sensors and preset models, exhibiting moderate adaptability and requiring manual parameter adjustment. This achieves a slag uniformity of 82%, a 15% improvement in furnace life extension, and a 5% reduction in nitrogen consumption. The system of this invention employs deep fusion of four modalities and uses a large model for end-to-end generation. It exhibits strong adaptability through pre-training and fine-tuning, achieving a slag uniformity of 92%, a 25% improvement in furnace life extension, and a 12% reduction in nitrogen consumption. 6.2 Summary of Innovations: First, this is the first time that a multimodal large-scale model has been introduced into the field of converter slag splashing protection, realizing unified representation and fusion decision-making of heterogeneous industrial data; Second, an industrial large-scale model deployment paradigm of pre-trained base plus downstream task fine-tuning is proposed to solve the model adaptation problem for different furnace types and furnace service periods; Third, a cross-modal attention fusion mechanism is designed to achieve efficient fusion of four types of heterogeneous data: flame, acoustic signature, flue gas, and point cloud; Fourth, causal reasoning and time-series prediction capabilities are introduced, enabling the system to have causal cognition and advanced planning capabilities regarding the influence of process parameters; Fifth, an end-to-end decision generation framework is constructed to directly output continuous control parameter sequences, replacing the traditional segmented control logic. 6.3 Economic Benefit Analysis: The implementation of this invention system in a single 300-ton converter can generate significant economic benefits. Regarding refractory material cost reduction, the furnace life is extended from 8500 heats to 10700 heats, a 25.9% increase. Based on a refractory material cost of 3800 yuan per ton and a refractory material consumption of 1.2 tons per heat, this translates to annual cost savings of approximately 3 million yuan. Regarding nitrogen consumption reduction, nitrogen consumption per ton of steel is reduced by 12%. Based on an annual production capacity of 3 million tons and a nitrogen cost of 0.15 yuan per cubic meter, this translates to annual nitrogen cost savings of approximately 500,000 yuan. Regarding quality improvement, slag uniformity is increased from 82% to 92%, reducing the number of furnace repairs due to uneven furnace lining erosion, resulting in approximately 15 fewer unplanned shutdowns per year, with each shutdown incurring a loss of approximately 200,000 yuan, leading to an annual benefit of approximately 3 million yuan. Regarding labor cost savings, the fully automated slag splashing process reduces the number of operators by 2. Based on an average annual cost of 200,000 yuan per operator, this translates to annual labor cost savings of 400,000 yuan. In terms of overall benefits, a single converter generates approximately 6.9 million yuan in annual comprehensive benefits, with an investment payback period of approximately 8 months. 6.4 Social Benefit Analysis: In terms of safe production, it reduces the safety risks of close-range observation of furnace flames and lowers the probability of injury to operators from high temperatures and splashes; in terms of green manufacturing, it optimizes nitrogen usage efficiency, reduces energy consumption, extends furnace lining life, and reduces refractory material consumption and solid waste generation; in terms of knowledge accumulation, it transforms experienced workers' expertise into reusable model parameters, solving the problem of experience inheritance in the metallurgical industry; in terms of industry demonstration, it provides a benchmark case for the application of large-scale models in the iron and steel metallurgical industry, promoting the intelligent upgrading of the industry.

[0071] 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 intelligent decision-making methods for converter slag splashing protection.

[0072] 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 intelligent decision-making methods for converter slag splashing protection.

[0073] 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.

[0074] 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.

[0075] 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.

[0076] 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.

[0077] The aforementioned computer storage medium and terminal equipment are created based on the aforementioned intelligent decision-making method for converter slag splashing protection. 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.

[0078] 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 smart decision-making method for converter slag splashing protection, characterized in that, include: A converter intelligent decision-making model was constructed and trained, taking flame sequence images, audio time-series signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining as inputs, and lance position height, nitrogen pressure, and nitrogen blowing flow rate as outputs, including: The input layer is used to receive flame sequence images, audio timing signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining. The feature extraction layer includes a flame feature extraction module, an audio feature extraction module, a flue gas feature extraction module, and a point cloud feature extraction module arranged in parallel. These modules are used to extract flame features, audio features, flue gas features, and point cloud features based on flame sequence images, audio time-series signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining, respectively. The feature fusion layer is used to fuse flame features, audio features, smoke features and point cloud features to obtain multimodal fusion features, so as to make decisions for the next moment based on the multimodal fusion features, including gun position height, nitrogen pressure and nitrogen flow rate; The output layer is used to output the gun position height, nitrogen pressure, and nitrogen flow rate. The system collects current flame sequence images, audio timing signals, flue gas concentration sequences, and three-dimensional point clouds of the furnace lining. These data are then input into the trained converter intelligent decision-making model to obtain the lance height, nitrogen pressure, and nitrogen flow rate at the next moment.

2. The method according to claim 1, characterized in that, The flame feature extraction module is used to extract flame features from flame sequence images; flame features include flame color features, flame shape features, and flame pulsation features. The audio feature extraction module is used to extract audio features based on the audio time sequence signal; Audio characteristics, including foam slag thickness characteristics and slag layer stability characteristics; The flue gas feature extraction module is used to extract flue gas features based on the flue gas concentration sequence. Flue gas characteristics, including decarburization rate characteristics and slag oxidizing characteristics; The point cloud feature extraction module is used to extract point cloud features based on the three-dimensional point cloud of the furnace lining. The point cloud features include the furnace lining thickness distribution features and the location features of erosion hot spots.

3. The method according to claim 2, characterized in that, The flame feature extraction module includes: The image segmentation unit is used to divide the flame sequence image into several image blocks to obtain an image block sequence. The position encoding unit, connected to the image block unit, is used to add position encoding information to each image block and preserve spatial positional relationships; The flame feature extraction unit, connected to the position encoding unit, is used to introduce a priori flame attention mask to enhance feature extraction of the central and edge regions of the furnace opening in the image block sequence, thereby obtaining flame color features, flame shape features, and flame pulsation features.

4. The method according to claim 3, characterized in that, The flame feature extraction unit includes flame color feature extraction elements, flame shape feature extraction elements, and flame pulsation feature extraction elements arranged in parallel. The flame color feature extraction element is used to extract the color distribution information of each image block in each frame of the flame sequence image to obtain the flame color features; The flame shape feature extraction element is used to extract the flame outline edge and overall morphology information from each frame of the flame sequence image to obtain the flame shape feature. The flame pulsation feature extraction element is used to perform temporal difference operations on the flame color and shape features of a series of consecutive flame sequence images to calculate the amplitude and frequency information of the flame dynamic changes and obtain the flame pulsation features.

5. The method according to claim 2, characterized in that, The audio feature extraction module includes a time-frequency transformation unit, a Mel spectrum generation unit, and an audio feature extraction unit connected in sequence. The time-frequency transformation unit is used to perform short-time Fourier transform on the audio signal to obtain a time-frequency representation; The Mel spectrum generation unit, connected to the time-frequency conversion unit, is used to convert the time-frequency representation into a Mel spectrum diagram; An audio feature extraction unit, connected to a Mel spectrum generation unit, includes: Metallurgical noise suppression element, used to enhance the Mel spectrum to suppress electromagnetic interference and mechanical vibration noise, resulting in an enhanced spectrum; The foam slag thickness feature extraction element is connected to the metallurgical noise suppression element. It is used to perform spectral encoding on the enhanced spectrum map, extract the spectral energy attenuation feature, and obtain the foam slag thickness feature. The slag layer stability feature extraction element is connected to the metallurgical noise suppression element. It is used to perform spectral encoding on the enhanced spectrum map, extract the spectral time variance and entropy value change information, and obtain the slag layer stability features.

6. The method according to claim 2, characterized in that, The flue gas feature extraction module includes a one-dimensional convolutional feature extraction unit, a temporal position encoding unit, and a flue gas feature extraction unit connected in sequence. A one-dimensional convolutional feature extraction unit is used to extract local features from the flue gas concentration sequence to obtain the flue gas temporal features. Flue gas concentration sequence, including CO, CO2 2、 O2 and N2 concentration sequences; The temporal position encoding unit, connected to the one-dimensional convolutional feature extraction unit, is used to add position encoding to the temporal features of flue gas and preserve the temporal order information; The flue gas feature extraction unit, connected to the time-series position encoding unit, includes a decarbonization rate feature extraction element and a slag oxidation feature extraction element arranged in parallel. The decarbonization rate feature extraction element is used to perform global dependency modeling on the time-series features of flue gas with location encoding, calculate the rate of change of the CO to CO2 concentration ratio, and obtain the decarbonization rate features. The slag oxidation feature extraction element is used to perform global dependency modeling on the temporal features of flue gas with location coding, extract the temporal trend of O2 concentration and N2 concentration compensation information, and obtain the slag oxidation characteristics.

7. The method according to claim 2, characterized in that, The point cloud feature extraction module includes: The point cloud sampling unit is used to downsample a set area of ​​the input furnace lining point cloud to obtain several sampling points and their three-dimensional coordinates. The sphere query grouping unit, connected to the point cloud sampling unit, is used to construct a local neighborhood sphere centered on each sampling point; The hierarchical feature learning unit, connected to the ball query grouping unit, is used to perform local-to-global hierarchical feature learning based on each local neighborhood ball to obtain point-level features at each level and global aggregated features. The point cloud feature extraction unit includes parallelly arranged furnace lining thickness distribution feature extraction elements and erosion hotspot location feature extraction elements: The furnace lining thickness distribution feature extraction element is connected to the hierarchical feature learning unit to perform thickness statistical distribution analysis on the global aggregated features, extract the mean and variance information of the thickness of each region of the furnace lining, and obtain the furnace lining thickness distribution features. The erosion hotspot location feature extraction element is connected to the point cloud sampling unit and the hierarchical feature learning unit. It is used to perform local anomaly detection based on each sampling point and its three-dimensional coordinates and each layer of point-level features, and to extract the center coordinates and erosion degree information of the region with a thickness below a set threshold, thereby obtaining the erosion hotspot location features.

8. The method according to claim 1, characterized in that, The feature fusion layer includes: The cross-modal attention fusion module is used to map flame features, audio features, smoke features and point cloud features to the same feature dimension through linear projection, and calculate cross-attention weights with flame features as the query and audio features, smoke features and point cloud features as the key to obtain multimodal fused features; The temporal causal reasoning module is used to construct a causal graph based on known knowledge in the metallurgical field, perform counterfactual reasoning based on the causal graph, constrain the information transmission path between features, and obtain causal enhanced features. The decision generation module, connected to the cross-modal attention fusion module and the temporal causal reasoning module, is used to synchronously output the gun position height, nitrogen pressure, and nitrogen flow rate based on multimodal fusion features and causal enhancement features.

9. The method according to claim 8, characterized in that, The decision generation module includes: The slag state classification unit is used to map slag state classification based on multimodal fusion features and causal enhancement features; A furnace lining erosion identification unit is used to map the furnace lining erosion stage based on multimodal fusion features and causal enhancement features; The slag splashing mode selection unit is used to map slag splashing mode selection based on multimodal fusion features and causal enhancement features; The decision generation unit is used to obtain the next moment's decision based on multimodal fusion characteristics, slag state classification, furnace lining erosion stage identification, and slag splashing mode selection, including lance position height, nitrogen pressure, and nitrogen blowing flow rate.

10. A terminal device, characterized in that, It includes a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to execute the intelligent decision-making method for converter slag splashing protection as described in any one of claims 1 to 9.

Images