Converter automatic splashing slag control method and system based on deep learning

Abstract

The application discloses a converter automatic slag splashing control method and system based on deep learning, relates to the technical field of converter steelmaking, and acquires a current converter opening image in a slag splashing process in real time, then inputs the current converter opening image into a trained deep learning model, judges whether the current slag splashing type is in a non-slagging state or a slagging state, returns to step S1 if the current slag splashing type is in the non-slagging state, and determines the current slag amount according to the current converter opening image if the current slag splashing type is in the slagging state, and finally controls the oxygen lance to perform corresponding processing actions according to the current slag amount, so that the problem of slag block misjudgment caused by light and smoke interference in the non-slagging state is avoided, and the accuracy of slag splashing control is improved.

Description

Technical Field

[0001] This application relates to the field of converter steelmaking technology, and in particular to a method and system for automatic slag splashing control in converters based on deep learning. Background Technology

[0002] Converter slag splashing technology is a key process for extending converter life and reducing refractory consumption. Its principle is to use high-speed nitrogen gas to blow residual liquid slag from inside the furnace onto the furnace lining through an oxygen lance, forming a protective layer. During this process, controlling the oxygen lance position is crucial: if the lance is too high, the slag splashing effect is poor; if the lance is too low, it may lead to splashing, furnace bottom erosion, or equipment damage.

[0003] Traditional slag splashing operations rely heavily on manual experience, with operators manually adjusting the lance position by observing images at the furnace opening. This method is not only labor-intensive but also highly dependent on individual skills, making it difficult to guarantee consistency and optimal results. To address this issue, the industry has developed automatic slag splashing technology based on time-series logic, which automatically executes lance pressing and lifting actions according to preset fixed times and steps. However, this open-loop control method cannot detect real-time changes in the furnace's operating conditions, and control deviations can easily occur if the furnace conditions do not conform to the preset model.

[0004] In fact, furnace mouth images can intuitively present the shape, quantity, and trajectory of slag lumps, providing the most direct visual basis for judging the state of the slag splashing process. Therefore, introducing image vision technology to achieve closed-loop control based on image feedback has become an inevitable direction for improving the automation level of slag splashing. However, directly applying image technology to the complex converter slag splashing scene faces severe technical challenges. Due to the intense flames and dense smoke inside the furnace during the initial stage of slag splashing, if slag lump detection is performed directly on the image at this time, it is very easy to misjudge interference sources such as flames and smoke as slag lumps, resulting in distorted detection results. This makes it impossible for the control system to accurately determine when to enter the effective "slag lifting" state, and thus difficult to start precise lance control based on slag quantity feedback at the optimal time. In addition, the timing of lance pressing and lifting directly affects the slag splashing effect. Pressing too early or too late will lead to uneven furnace lining protection, and existing methods lack effective real-time slag condition quantification, making it difficult to achieve precise lance pressing rhythm control. Furthermore, different furnace conditions (such as expanding the furnace bottom and protecting the molten pool) have different requirements for sputtering intensity, and different furnace bases and different shifts have different operating habits and working conditions. This makes it difficult to directly extend the visual model trained for a single furnace base to other scenarios. The model's generalization ability is insufficient, which limits the large-scale application of the technology.

[0005] Slag splashing in converters is a key process for extending converter life and reducing refractory material consumption. Its principle involves using high-speed nitrogen gas through an oxygen lance to splash residual liquid slag onto the furnace lining, forming a protective layer. Currently, some automatic slag splashing control methods employ open-loop control based on time-series logic, automatically executing lance pressing and lifting actions according to preset fixed times and steps. Other methods attempt to directly detect slag blocks or perform multi-state classification and recognition on the furnace mouth image to achieve visual feedback control of the slag splashing process.

[0006] Therefore, there is an urgent need for an intelligent automatic slag splashing and lance position control method that can overcome environmental interference, accurately identify process stages, provide real-time slag condition quantification, and adapt to various furnace type maintenance modes and different furnace conditions. Summary of the Invention

[0007] To address at least one of the aforementioned technical problems, this application provides a deep learning-based method for automatic slag splashing control in converters, comprising: S1: Real-time acquisition of the current furnace mouth image during the converter slag splashing process; S2: Input the current furnace opening image into the trained deep learning model to determine whether the current slag splashing type is a non-slag-forming state or a slag-forming state; the deep learning model includes: The input layer is used to input the current furnace opening image; The backbone network consists of shallow, medium, and high-level networks with progressively decreasing convolutional kernel sizes, as well as a fusion network. The shallow network is used to extract the morphological features of flames and smoke; the medium network is used to extract the texture features of slag particles; the high-level network is used to extract the motion blur features during the dynamic slag splashing process; and the fusion network is used to fuse morphological features, texture features, and motion blur features to obtain deep features. The output layer is used to output the results of the non-slag-forming state and the slag-forming state based on the deep features; S3: If the slag is not rising, return to step S1; if the slag is rising, determine the current slag amount based on the current furnace opening image. S4: Based on the current slag volume, control the oxygen lance to perform the corresponding processing action.

[0008] Furthermore, the input layer is used to receive and preprocess the current furnace opening image to obtain an image tensor; The backbone network consists of multiple layers of alternating convolutional and pooling layers, forming a hierarchical structure of shallow, mid, and high-level networks; and the high-level networks incorporate a CA coordinate attention module and / or a channel attention module. The CA coordinate attention module is used to redistribute the spatial dimension weights of the input features of the high-level network, reduce the feature weights of large-area flame and smoke regions, and amplify the weights of local particle textures in the slag. The channel attention module is used to perform global channel dimension statistics and weight redistribution on the output features of high-level networks, automatically enhance the effective feature channel weights including slag tails and dynamic blurred trajectories, and suppress the ineffective feature channel weights corresponding to firelight and smoke. The output layer is used to flatten the deep features into a one-dimensional feature vector and output the results of the non-slag-forming state and the slag-forming state.

[0009] Furthermore, the shallow network is configured with the first convolutional kernel, which, with its large receptive field covering the entire image area, extracts large-area spatially continuous, wide- and high-dimensional globally uniform features of the image as flame and smoke morphological features; the corresponding pooling layer completes downsampling and dimensionality reduction. The middle layer network is configured with a second convolutional kernel, which relies on local small receptive fields to finely capture pixel grayscale differences and arrangement patterns, and extracts the surface roughness of the slag block as the texture features of the slag block particles; the matching pooling layer completes the filtering of image redundancy noise; The high-level network learns high-order semantic features based on the feature information obtained by the preceding multi-layer convolution and pooling iterative downsampling, identifies the trailing and blurred trajectory generated by the high-speed splashing of slag blocks, and extracts the motion blur features in the splashing process; the corresponding pooling layer completes the final dimensionality compression.

[0010] Furthermore, the CA coordinate attention module performs global pooling on the input features of the high-level network in the vertical and horizontal independent dimensions to obtain global context information in the height and width directions, generate height attention weight maps and width attention weight maps, and then complete the redistribution of spatial weights. The channel attention module is used to perform global average pooling on the output features of the high-level network, compressing the spatial information into a one-dimensional channel vector; then, two cascaded fully connected layers are used to mine channel dependencies and generate importance weights for each channel; and feature recalibration is completed using the channel importance weights.

[0011] Furthermore, the training process of the deep learning model includes: Acquire furnace mouth image data of at least two furnace bases during the slag splashing process to construct a multi-source dataset; the multi-source dataset includes negative samples of non-slag-splashing images under typical interference conditions and positive samples of slag-splashing images; typical interference conditions include any one or more of the following: smoke surrounding, local overexposure of the image, and molten steel gushing. Based on historical and recent datasets from multiple sources, an incremental training dataset is constructed using an incremental learning approach. The deep learning model is fine-tuned and updated based on the optimization objective function and the incremental training dataset. By jointly training across furnaces, the loss functions of each furnace are summed to minimize the total loss of each furnace and the model parameters are updated to obtain a deep learning model that is adapted to all converters in the plant.

[0012] Furthermore, if the slag formation is in progress, the steps for determining the current slag quantity based on the current furnace opening image include: The current furnace opening image is preprocessed with noise reduction and smoothing to filter out furnace dust noise and preserve the edge features of the furnace opening steel shell; The preprocessed image is decomposed into color channels, and the color channel whose contrast meets the preset conditions is selected as the channel to be processed. Based on the channel to be processed, the region is segmented, and candidate regions that conform to the geometric size characteristics of the furnace opening are extracted. Then, the effective region of the furnace opening is dynamically updated based on the boundary coordinates of the candidate regions. Effective slag blocks are selected within the effective area of ​​the furnace opening, and the current slag quantity is calculated.

[0013] Furthermore, the step of acquiring the current furnace mouth image during the converter slag splashing process in real time includes: When the state is determined to be non-slag-forming, the image acquisition device is controlled by the first imaging strategy to acquire the current furnace mouth image; the first imaging strategy includes reducing the aperture or shortening the exposure time to suppress highlights and highlight the smoke texture features; When the slag-raising state is determined, the image acquisition device is controlled by the second imaging strategy to acquire the current furnace mouth image; the second imaging strategy includes increasing the exposure gain or extending the exposure time to ensure that the clear trajectory of the high-speed moving slag block is captured.

[0014] Furthermore, the steps for controlling the oxygen lance to perform the corresponding processing actions based on the current slag volume include: The processing actions include lowering and raising. The lowering action is the action performed by controlling the oxygen lance through the press-down command, and the raising action is the action performed by controlling the oxygen lance through the lift-up command. The current slag volume is compared with the preset lance trigger threshold. When the current slag volume is less than or equal to the lance trigger threshold, a lance trigger command is generated and output to control the oxygen lance to perform a descent action; otherwise, the output of the lance trigger command is paused. After generating and outputting the lance control command to control the oxygen lance to perform the descent action, the lance position height is monitored in real time. Compare the gun position height with the preset process switching threshold. When the gun position height is greater than or equal to the process switching threshold, return to execute step S1. Otherwise, by judging the process of raising the lance, a lance-raising command is generated and output to control the oxygen lance to perform the upward movement.

[0015] Furthermore, the steps for generating and outputting a lifting command to control the oxygen lance to perform an upward movement, based on the lifting judgment process, include: Get the currently selected furnace type maintenance mode. The furnace type maintenance modes include furnace bottom raising mode, molten pool protection mode, and furnace bottom lowering mode. If the currently selected mode is the furnace bottom raising mode, the lance lifting process will be divided into at least two stages, with different slag amount thresholds set for each stage, and the process will be judged sequentially for each stage, including: When the current slag amount is less than or equal to the first slag amount threshold for lifting the gun, the first gun lifting command is generated and output; when the current slag amount is less than or equal to the second slag amount threshold for lifting the gun, the second gun lifting command is generated and output. The lifting height and holding time corresponding to the first gun lifting command are greater than the lifting height and holding time corresponding to the second gun lifting command. If the currently selected mode is the molten pool protection mode or the furnace bottom lowering mode, then when the current slag amount is less than or equal to the third lance lifting slag amount threshold, a lance lifting command will be generated and output. The relationship between the threshold values ​​for the amount of slag removed from the spear is as follows: the second threshold value for the amount of slag removed from the spear is less than the third threshold value for the amount of slag removed from the spear, which is less than the first threshold value for the amount of slag removed from the spear.

[0016] On the other hand, the present invention also provides a deep learning-based automatic slag splashing control system for converters, characterized in that it employs any of the above-mentioned deep learning-based automatic slag splashing control methods for converters, including: The image acquisition module, connected to the image processing module, is used to acquire the current furnace mouth image during the converter slag splashing process; The image processing module is connected to the image acquisition module and the decision module respectively. It is used to receive control parameters and control signals sent by the decision module, perform slag identification and slag quantity detection on the current furnace mouth image, and upload the identification results to the decision module. The decision module is connected to the image processing module and the execution module respectively. It is used to determine the smelting status based on the PLC signal uploaded by the execution module, send control parameters and control signals to the image processing module, receive the recognition results, generate the gun pressing command or gun lifting command, and send the command to the execution module. The execution module is connected to both the decision module and the external PLC controller. It is used to collect the nitrogen valve status signal and oxygen lance position height signal uploaded by the external PLC controller, upload the signal to the decision module, receive the instructions fed back by the decision module, and send the instructions to the external PLC controller.

[0017] This application provides a method and system for automatic slag splashing control in converters based on deep learning. Compared with the prior art, the beneficial effects of the technical solution provided by this invention include at least the following: 1. Improve recognition accuracy and anti-interference ability: By establishing a progressive recognition architecture of "classification first and detection later", the deep learning model is first used to identify the non-slag formation period and the slag formation period, which effectively avoids the problem of false detection of slag blocks caused by the interference of fire light and smoke in the non-slag formation state, and improves the accuracy of recognition.

[0018] 2. Achieve refined and adaptive control: Target detection technology is used to detect slag quantity and obtain real-time slag quantity data. The lance pressing rhythm is dynamically adjusted according to the slag quantity changes, achieving precise control based on real-time slag conditions. This avoids the problem of lance pressing timing being too early or too late in the traditional static control model and eliminates the blindness of traditional timing control.

[0019] 3. Improve model generalization ability and maintenance efficiency: Through multi-source data collection, incremental training and cross-furnace joint training, a single deep learning model can be adapted to the complex working conditions of multiple converters in the whole plant, reducing the model maintenance cost.

[0020] 4. Meets diverse furnace maintenance needs: By pre-setting various furnace type maintenance modes such as furnace bottom expansion, molten pool protection, and furnace bottom reduction, and setting differentiated lance lifting trigger conditions, it achieves precise maintenance of furnace types as needed, effectively extending the service life of converters. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment of the deep learning-based automatic slag splash control method of this application; Figure 2 This is a schematic diagram of the overall process of the deep learning-based automatic slag splash control method of this application; Figure 3 This is a schematic diagram of the structure of an embodiment of the deep learning model of the automatic slag splashing control method based on deep learning in this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application 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 this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0023] It should be noted that if the embodiments of this application involve directional indicators, such as up, down, left, right, front, back, etc., these directional indicators are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly. Furthermore, if the embodiments of this application 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 should be included within the scope of protection of this application.

[0024] See Figures 1-2 This application provides an automatic slag splash control method based on deep learning, comprising: S1: Real-time acquisition of the current furnace mouth image during the converter slag splashing process; Specifically, the current furnace mouth image can be continuously acquired by image acquisition equipment at a preset frequency to reflect the slag splashing process in the converter furnace mouth area, and to present the movement trajectory, quantity changes and morphological evolution of slag blocks.

[0025] S2: Input the current furnace opening image into the trained deep learning model to determine whether the current slag splashing type is a non-slag-forming state or a slag-forming state; Specifically, the deep learning model can be a pre-trained convolutional neural network classification model. Its input is a single frame of the current furnace opening image, and its output is the probability distribution of two categories, corresponding to the non-slag-forming state and the slag-forming state, respectively. Specifically, such as... Figure 3 As shown, a deep learning model includes: an input layer, a backbone network, and an output layer; The input layer is used to input the current furnace opening image. Specifically, it can optionally receive and preprocess the current furnace opening image, adjust the image size to the model input size, and normalize the pixel values. Specifically, this includes reading the current furnace opening image from memory, calling an image processing library to complete the size adjustment and normalization, and generating an image tensor with a shape of (number of channels × height × width).

[0026] The backbone network consists of shallow, mid, and high-level networks with progressively decreasing convolutional kernel sizes, as well as a fusion network. The shallow network extracts flame and smoke morphological features; the mid-level network extracts slag particle texture features; the high-level network extracts motion blur features during dynamic slag splashing; and the fusion network fuses morphological, texture, and motion blur features to obtain deep features. Specifically, the backbone network can optionally include multiple alternating convolutional and pooling layers, divided into shallow, mid, and high-level networks according to feature extraction depth. Each layer is configured with progressively decreasing convolutional kernels to extract corresponding morphological, texture, and motion blur features at different levels. These features are then fused to obtain a three-dimensional tensor containing multi-dimensional slag features, which is then passed to the output layer.

[0027] The output layer flattens the deep features into a one-dimensional feature vector and outputs the results of the non-slag-forming state and the slag-forming state. Specifically, it may include a classification head composed of fully connected layers and a Softmax activation function. The deep features output from the backbone network are flattened into a one-dimensional feature vector. Then, this one-dimensional feature vector is input into the fully connected layer of the classification head. The fully connected layer performs a linear weighted mapping on the one-dimensional feature vector, outputting the original scores corresponding to the slag-forming state and the non-slag-forming state. The original scores are input into the Softmax function and converted into a two-dimensional probability distribution through exponential normalization to obtain the first probability that the current furnace mouth image belongs to the non-slag-forming state and the second probability that it belongs to the slag-forming state. The magnitudes of the first probability and the second probability are compared. If the first probability is greater than the second probability, the slag-forming type of the current slag-splashing stage is determined to be the non-slag-forming state; otherwise, it is determined to be the non-slag-forming state.

[0028] S3: If the slag is not rising, return to step S1; if the slag is rising, determine the current slag amount based on the current furnace opening image; S4: Based on the current slag amount, control the oxygen lance to perform the corresponding processing action.

[0029] Specifically, if the slag is not being formed, return to step S1 and re-acquire the current furnace opening image until step S2 determines that the slag has been formed, then begin the "slag quantity counting" step to determine the current slag quantity; then proceed to step S4 to determine the subsequent actions of the oxygen lance based on the current slag quantity.

[0030] This embodiment presents an automatic slag splashing control method based on deep learning according to the present invention. Its core inventive point lies in: after step S2, determining whether the current slag splashing type is a non-slag-forming state or a slag-forming state; only in the slag-forming state is the "slag counting" step of "determining the current slag quantity" initiated. This is because traditional solutions either employ a neural network model throughout the entire process to identify the slag splashing state stage by stage, or directly use a "slag counting" step based on slag quantity identification; however, the "critical period before slag formation" and the "effective slag splashing period" are easily misjudged by the neural network model, and interference from firelight and smoke also hinders accurate slag quantity identification and counting. Step S2 of this invention addresses the technical challenge of traditional visual technologies failing easily in the early stages of slag splashing. It introduces a deep learning model specifically designed to differentiate between the "critical period before slag formation" and the "effective slag splashing period" as an anti-interference classification model. Its core purpose is to distinguish the initial phenomena: In the initial stage of the converter slag splashing process (i.e., before slag formation), the physicochemical reactions within the furnace are intense. Nitrogen gas streams impact the liquid surface, causing turbulence and resulting in a large amount of smoke enveloping the furnace mouth. This is accompanied by violent shaking of the molten steel surface and even instantaneous spurting of molten steel. During this stage, traditional detection algorithms based on brightness thresholds or simple textures will generate numerous artifacts due to the high-brightness reflection of the smoke and the specular reflection of the molten steel, misjudging them as "slag lumps." This means that identifying a certain feature in the "early stage of slag splashing" and directly providing a lance position adjustment can lead to the system erroneously triggering the oxygen lance action when slag splashing conditions are not actually present. For example, triggering the lance-pressing logic could pose a significant safety hazard. This invention introduces a deep learning model to accurately capture the "slag-raising moment" under complex interference such as flames and dense smoke, providing a clean trigger signal for subsequent slag quantity statistics. In other words, the deep learning model in step S2, as an anti-interference classification model, is a "pre-judgment and then control trigger mechanism," serving as the "switch" or "precondition" for step S3 to initiate the "determine the current slag quantity, i.e., slag counting" step. Only when "slag raising" is determined is the "slag quantity counter" activated; if "slag raising has not occurred," it continues to wait, without counting slag or performing subsequent control, thus eliminating complex interference such as flames and dense smoke and distinguishing it from actual slag.

[0031] Based on this, in order to eliminate the complex interference of the specific environment of "flame and dense smoke" and to distinguish the characteristic differences between the "critical period before slag splashing" and the "effective slag splashing period," the specific structure of the deep learning model of this invention is a proprietary inventive design, not a simple technical choice. Specifically: By configuring shallow, mid, and high-level networks with progressively decreasing convolutional kernel sizes and performing layer-by-layer downsampling, it is possible to substantially distinguish between interference such as "flames and thick smoke" and "real slag". Because: shallow networks, with the largest convolutional kernels, extract large-area, spatially continuous, and globally uniform features, perfectly reflecting the overall morphology of flames and smoke; mid-layer networks, with centered convolutional kernels, extract locally discrete, spatially sparse features with abrupt changes at particle edges, perfectly reflecting the texture features of slag particles; high-layer networks, with the smallest convolutional kernels, extract detailed features of trajectories and trailing effects, perfectly reflecting the motion blur features during dynamic slag splashing; based on this, a fusion module integrates morphological features, texture features, and motion blur features to obtain deep features, which can distinguish the subtle differences in morphology, texture gradient, and motion blur between "static / semi-static light curtains and smoke" and "high-speed moving solid slag blocks," identifying that flames in a non-slag-splashing state typically exhibit continuous, blurred-edge fluid features, while slag blocks in an effective slag-splashing state exhibit discrete, bright particle features with clear motion trailing effects; effectively eliminating complex interference from flames, dense smoke, etc., and distinguishing them from real slag.

[0032] Preferably, the backbone network can be a lightweight and efficient architecture such as ResNet or MobileNet, and may include: (1) Multiple layers of alternating convolutional and pooling layers are used to hierarchically form shallow, intermediate, and high-level networks; in a preferred embodiment: The shallow network is configured with the first convolutional kernel, which, with its large receptive field covering the entire image area, extracts large-area spatially continuous, wide and high-dimensional globally uniform features of the image as flame and smoke morphological features; the corresponding pooling layer completes downsampling and dimensionality reduction. The middle layer network is configured with a second convolutional kernel, which relies on local small receptive fields to finely capture pixel grayscale differences and arrangement patterns, and extracts the surface roughness of the slag block as the texture features of the slag block particles; the matching pooling layer completes the filtering of image redundancy noise; The high-level network learns high-order semantic features based on the feature information obtained by the preceding multi-layer convolution and pooling iterative downsampling, identifies the trailing and blurred trajectory generated by the high-speed splashing of slag blocks, and extracts the motion blur features in the splashing process; the corresponding pooling layer completes the final dimensionality compression.

[0033] It is worth noting that the specific structures of the shallow, middle and high-level networks are not limited to this. As long as downsampling is performed layer by layer to extract the morphological features of flames and smoke, the texture features of slag particles and the motion blur features respectively, the specific number of convolutional layers and pooling layers, the size of the convolutional kernels, etc., can be arbitrarily set by those skilled in the art.

[0034] (2) In high-level networks, a CA coordinate attention module and / or a channel attention module are built in; The CA coordinate attention module is used to redistribute the spatial dimension weights of the input features of the high-level network, reducing the feature weights of large-area flame and smoke regions and amplifying the weights of local particle textures of the slag. The channel attention module is used to perform global channel dimension statistics and weight redistribution of the output features of the high-level network, automatically enhancing the effective feature channel weights containing slag tails and dynamic blurred trajectories, and suppressing the ineffective feature channel weights corresponding to fire and smoke.

[0035] In a preferred embodiment, the CA coordinate attention module specifically performs global pooling on the input features of the high-level network in independent vertical and horizontal dimensions to obtain global context information in the height and width directions, generating a height attention weight map and a width attention weight map, thereby completing the redistribution of spatial weights. This method can reduce the feature weights of large-area, continuous, and uniformly distributed flame and smoke regions, while amplifying the weights of local regions of the furnace slag body with abrupt grayscale changes and discrete particle textures.

[0036] More specifically: the CA coordinate attention module includes: The global pooling unit is used to perform global pooling along the vertical and horizontal independent dimensions of the input features of the high-level network (the output features of the middle-level network, such as the output features of the third downsampling module), to separate and obtain global context information in the height and width directions, and output two sets of one-dimensional feature vectors with coordinate position information. The feature fusion and dimensionality reduction unit is used to concatenate two sets of feature vectors along the channel dimension to reduce the dimensionality, thereby achieving interactive fusion of information in two spatial directions. Specifically, channel dimensionality reduction can be achieved by 1×1 convolution, batch normalization, or nonlinear activation. The bidirectional branch convolutional weight generation unit is used to split the fused features and feed them into two independent convolutional branches to generate a height attention weight map and a width attention weight map, respectively. Specifically, the convolutional branches can be selected as two 1×1 convolutional branches. Among them, the height convolutional branch generates a height attention weight map, and the width convolutional branch generates a width attention weight map. The ends of the two branches are configured with a Sigmoid activation function to normalize the weights to the 0~1 range. The channel-wise feature weighting unit is used to multiply the height attention weight map, width attention weight map, and high-level network input features (original input feature map) channel by channel to redistribute spatial weights. Through the above operation, the feature weights of large-area continuous and uniformly distributed flame and smoke regions can be reduced, while the weights of local regions of furnace slag with gray-level abrupt changes and discrete particle textures can be amplified. The spatially enhanced features are then fed into the high-level backbone to extract motion blur features.

[0037] More preferably, a channel attention module is used to perform global average pooling on the output features of the high-level network, compressing spatial information into a one-dimensional channel vector; then, two cascaded fully connected layers are used to mine channel dependencies and generate importance weights for each channel; feature recalibration is completed using these channel importance weights. This approach enhances effective channels containing blurred trajectories of slag block motion and suppresses ineffective channels containing firelight and smoke.

[0038] More specifically, the channel attention module includes: The global spatial compression unit is used to perform two-dimensional global average pooling on the output features (high-order semantic feature maps) of the high-level network, compressing each channel feature (H×W size) into a single scalar, and summarizing the spatial information of the entire feature map into a one-dimensional channel vector (C×1×1). The channel correlation activation unit consists of two fully connected layers connected in series to form a bottleneck structure. The first fully connected layer is used to reduce the dimensionality of the channel vector to reduce the amount of computation, and is combined with ReLU nonlinear activation to mine the dependencies between channels. The second fully connected layer is used to increase the dimensionality to restore the original number of channels, and the end is activated by Sigmoid to generate one-dimensional channel weights for each channel. The channel feature recalibration unit is used to multiply the one-dimensional channel weights with the original high-level network output features channel by channel, automatically increasing the effective feature channel weights of storage slag block trails and dynamic blurred trajectories, suppressing the ineffective feature channel weights that only reflect firelight and smoke, and outputting high-level features optimized by the channel dimension.

[0039] Furthermore, in this embodiment, the training process of the deep learning model includes: S21: Obtain furnace mouth image data of at least two furnace bases during the slag splashing process and construct a multi-source dataset; the multi-source dataset includes negative samples of non-slag-splashing images under typical interference conditions and positive samples of slag-splashing images; typical interference conditions include, but are not limited to, any one or more of the following: smoke surrounding, local overexposure of the image, and molten steel gushing. Specifically, during the sample construction phase, a large number of images without slag formation under typical interference conditions such as "smoke surrounding", "local overexposure of the image", and "molten steel gushing" can be collected as negative samples, and images with clear slag block trajectories can be collected as positive samples for subsequent comparative training. Preferably, since the conditions of "smoke surrounding" and "local overexposure of the image" are more common and typical, a higher weight coefficient can be configured, while a lower weight coefficient can be configured for the condition of "molten steel gushing".

[0040] More specifically, the multi-source dataset can optionally include furnace mouth images from different shifts, furnace batches, and furnace seats, and label them with stage category labels (0 for no slag removal and 1 for slag removal). Different furnace batches can also be numbered to distinguish each sample data. More specifically, furnace mouth image data of corresponding shifts and furnace batches in at least two furnace seats during the slag splashing process can be obtained, and the images can be standardized, labeled with stage category labels, furnace seat numbers, and shift numbers to construct a multi-source dataset.

[0041] S22: Construct an incremental training dataset based on historical and recent datasets from multiple sources using an incremental learning approach; specifically, the incremental training dataset can be constructed using an incremental learning approach based on historical and recent datasets from multiple sources, and the sample distribution of different data sources can be balanced; the methods for balancing the sample distribution of different data sources can include sliding window, sample weighting, etc.

[0042] S23: Fine-tune and update the deep learning model based on the optimization objective function and the incremental training dataset; through joint training across furnaces, sum the loss functions of each furnace to minimize the total loss of each furnace and update the model parameters to obtain a deep learning model that is adapted to all converters in the plant.

[0043] Specifically, image data of different furnace bases, shifts, and furnace cycles during the slag splashing process were collected to construct the following multi-source dataset: ;

[0044] in, Image of the furnace opening. For stage category labels, Number the furnace base. Number the work groups. Then obtain the mean of the training set. Based on the standard deviation σ, the image is standardized as follows: ;

[0045] Secondly, based on historical data sets from multiple sources. and recent datasets An incremental training dataset is constructed using an incremental learning approach: ;

[0046] Next, the deep learning model is fine-tuned and updated using the incremental training dataset, and the objective function is optimized as follows: ;

[0047] in, For the sample size, The probability value predicted by the model. This represents the regularization coefficient. Through joint training across furnace stands, the model parameters are optimized. Minimize the loss function for all furnace bases: ;

[0048] Where K is the total number of furnace bases, Let be the loss function for the k-th furnace base.

[0049] After training, a single model is adapted to meet the slag splashing stage recognition requirements of multiple converters in the entire plant, and then an updated deep learning model is output.

[0050] Furthermore, the process of calling and recognizing the model includes: (1) Take a frame of the current furnace opening image from the sequence image queue in chronological order, and perform normalization preprocessing on the image to conform to the model input format; (2) The preprocessed image tensor is passed to the pre-trained deep learning model. The model outputs a two-dimensional probability distribution, which represents the probability that the current image belongs to the non-slag-forming state and the slag-forming state, respectively.

[0051] (3) Compare the probability values ​​of the two categories and take the category with the higher probability as the final judgment result.

[0052] (4) Repeat the above process for each subsequent frame to achieve continuous real-time recognition.

[0053] Specifically, the output of a deep learning model can be represented as: ;

[0054] in, The final result (value is 0 or 1). The category label indicates whether or not slagging occurs. For the preprocessed image tensor, These are the model parameters.

[0055] For example, if the model output probability distribution is [0, 0.1], then The current image is determined to be in a non-slag-forming state; if the output is [0.1, 0.9], then... =1, indicating a slagging state.

[0056] Furthermore, if the determination result of step S2 is a non-slag-forming state, then the non-slag-forming state loop mechanism is executed; if the determination result of step S2 is a slag-forming state, then the subsequent slag quantity detection process is executed.

[0057] The circulation mechanism in the non-slag-forming state includes: (1) Clear the temporary marker of the current frame image in the cache, release the memory, return to step S1, read the latest frame image from the image acquisition buffer queue, and overwrite the old image in the current work area. For real-time streaming mode, simply wait for the next frame to arrive.

[0058] (2) After acquiring a new frame image, repeat step S2. Perform the same preprocessing and forward inference of the deep learning model on the new image, recalculate the probability that it belongs to the slag-raising state and the non-slag-raising state, and determine the slag splash type again. Form this "acquisition-identification-judgment" process into a closed loop, with each frame processed independently and without interference.

[0059] (3) The loop continues until the judgment result of a certain frame is "slag raising" state, then the loop is exited and the slag quantity detection process is entered.

[0060] More preferably, to adapt to the drastic changes in the lighting environment throughout the slag splashing process, the step of acquiring the current furnace mouth image during the converter slag splashing process in real time further includes: When the state is determined to be non-slag-forming, the image acquisition device is controlled by the first imaging strategy to acquire the current furnace mouth image; the first imaging strategy includes reducing the aperture or shortening the exposure time to suppress highlights and highlight the smoke texture features; When the slag-raising state is determined, the image acquisition device is controlled by the second imaging strategy to acquire the current furnace mouth image; the second imaging strategy includes increasing the exposure gain or extending the exposure time to ensure that the clear trajectory of the high-speed moving slag block is captured.

[0061] In this embodiment, the present application also introduces an adaptive adjustment mechanism for image acquisition parameters. In the non-slag-forming state, considering the high brightness of the flame and the tendency for overexposure, the system automatically adjusts the imaging parameters of the industrial camera (e.g., reducing the aperture and shortening the exposure time) to suppress highlights, highlight smoke texture features, and assist the model in accurate anti-interference classification. Once the model determines that slag formation has occurred, the system immediately switches the imaging parameters (e.g., appropriately extending the exposure time and adjusting the gain) to ensure the capture of a clear trajectory of high-speed moving slag blocks. Through this closed-loop control of "condition perception - parameter adaptation - accurate feature classification," accurate state determination under complex interference is achieved.

[0062] Further, in this embodiment, step S3 includes: S31: Perform noise reduction and smoothing preprocessing on the current furnace opening image to filter out furnace dust noise and retain the edge features of the furnace opening steel shell; Specifically, an image denoising and smoothing filtering algorithm with edge preservation characteristics can be used to preprocess the image, aiming to filter out high-frequency interference such as salt and pepper noise and sparks caused by furnace dust on site, while preserving the edge details between the furnace opening steel shell and the background to the greatest extent, and preventing edge blurring from causing positioning deviation.

[0063] S32: Decompose the preprocessed image into color channels, and select the color channel whose contrast meets the preset conditions as the channel to be processed. Specifically, the preprocessed image can be decomposed into RGB color channels, and the channel with the strongest contrast and least interference from flames (such as the green or red channel) can be selected as the analysis benchmark to obtain the channel to be processed. S33: Perform region segmentation based on the channel to be processed, extract candidate regions that conform to the geometric size characteristics of the furnace opening, and generate dynamically updated effective furnace opening regions based on the boundary coordinates of the candidate regions; Specifically, threshold segmentation technology can be used to extract highlighted areas, and connected component analysis can be used to select large areas that meet the geometric size characteristics of the furnace opening as candidate furnace opening areas. Based on this, the actual boundary of the furnace opening is calculated by combining the regional coordinate parameters, and a dynamically updated region of interest (ROI) is generated, which is the effective area of ​​the furnace opening. S34: Select effective slag blocks within the effective area of ​​the furnace mouth and calculate the current slag quantity. Specifically, extract candidate slag block areas within the effective area of ​​the furnace mouth to shield the fixed slag-forming areas at the edge of the furnace mouth, and select effective slag blocks; then count the number of all effective slag blocks to obtain the current slag quantity.

[0064] More specifically, the options include: S341: Within the effective area of ​​the furnace opening, the highlighted area is extracted as a candidate slag block area by threshold segmentation; S342: Perform connected component analysis on the extracted candidate slag block regions to obtain the area parameter and at least one shape parameter of each candidate slag block region; S343: Based on the area parameter and the shape parameter, as well as the preset area threshold range and shape threshold range, select the target area that simultaneously meets the area threshold range and shape threshold range as the effective slag block, and calculate the current slag quantity.

[0065] Specifically, based on preset area and shape threshold ranges, only connected regions with areas within the preset area threshold range and shape parameters satisfying the preset shape threshold range are retained as valid slag blocks. The area constraint refers to the upper and lower limits set for the number of pixels contained in the candidate slag block region. Candidate regions with too small an area are typically sparks, noise, or small smoke spots, while candidate regions with too large an area may be furnace edges, large pieces of furnace lining debris, or reflective areas in the image. The shape constraint refers to the parameter requirements set for the geometric shape of the candidate slag block region, including but not limited to roundness, rectangularity, aspect ratio, or contour complexity.

[0066] In this embodiment, a preferred embodiment of step S3 is given. Instead of directly performing slag quantity statistics on the current furnace mouth image, a dynamic furnace mouth positioning step S31-S33 is added before performing slag quantity statistics to solve the positioning interference problem of "where to count slag". This is because, considering the unstructured characteristics of the converter site environment, the effective imaging area of ​​the furnace mouth in the image is not constant. Specifically, due to the mechanical deviation of the fume hood lifting mechanism, the long-term adhesion of slag at the furnace mouth (commonly known as "furnace mouth nodules"), and the change in the field of view caused by the tilting of the furnace body, the actual physical boundary of the furnace mouth will drift or deform. If only a preset fixed rectangular area is used as the slag quantity statistics window, it is very easy to misidentify the "stacking slag" fixed at the edge of the furnace mouth as "flowing slag blocks", or some slag blocks will not be counted due to area offset, thus causing distortion of slag quantity data. This application adds a dynamic furnace mouth positioning process. Only when the system successfully detects the effective area of ​​the furnace mouth will this real-time generated area be used as the sole spatial range for subsequent slag extraction. This automatically shields against interference from fixed slag buildup at the furnace mouth edge and captures all slag within the effective area of ​​the furnace mouth. Through this dynamic positioning mechanism, the system can accurately distinguish between "static slag accumulation" and "dynamic splashing," further improving the accuracy of subsequent slag quantity statistics and ensuring that the slag quantity statistics reflect the true splashing intensity inside the furnace, significantly improving the accuracy of subsequent lance position control decisions. It is worth noting that if, due to extreme operating conditions, such as complete obscuring by dense smoke, the effective area of ​​the furnace mouth is not detected, the system automatically calls up the effective area of ​​the previous frame or a preset default area as a substitute.

[0067] More preferably, step S4 may include: generating and outputting control commands to control the oxygen lance to perform a pressing or lifting action. The pressing command is used to control the oxygen lance to perform a lowering action, and the lifting command is used to control the oxygen lance to perform a raising action. Pressing and lifting are sequential actions. It is necessary to determine whether the oxygen lance has lowered to a suitable position for lifting based on the process switching threshold. If the oxygen lance position height has not reached the preset process switching threshold, the pressing judgment process is initiated; if the oxygen lance position height has reached the preset process switching threshold, the lifting judgment process is initiated.

[0068] Furthermore, the recoil control judgment process includes: (1) Compare the current slag quantity with the preset gun-pressing trigger threshold. If the current slag quantity is less than or equal to the gun-pressing trigger threshold, generate and output a gun-pressing command; otherwise, pause the output of the gun-pressing command and return to step S1; (2) After generating and outputting the lance pressing command, monitor the oxygen lance height in real time; (3) The lance height is compared with the preset process switching threshold. When the oxygen lance height is not lower than the process switching threshold, return to step S1 to continue collecting images for slag quantity detection and lance pressure judgment; otherwise, end the lance pressure judgment process and switch to the lance lifting judgment process.

[0069] In addition, slag volume data and corresponding slag splashing effect evaluation indicators from historical furnace cycles can be collected to construct a mapping relationship model between slag volume threshold and slag splashing effect; and the lance triggering threshold can be dynamically optimized based on the furnace type characteristics, steel grade requirements and furnace age of the current furnace cycle.

[0070] Furthermore, the gun-handling judgment process includes: (1) Obtain the current furnace type maintenance mode and determine the corresponding lance lifting trigger condition; (2) Keep the gun position unchanged, continue to collect the current furnace mouth image, and repeat the slag quantity detection to obtain real-time slag quantity data; (3) Determine whether the real-time slag volume data meets the lance-lifting trigger condition. If the current slag volume data is less than or equal to the lance-lifting trigger threshold, it indicates that the number of slag blocks at the furnace mouth has decreased to the target value and the slag splashing effect has reached the expected level. Then generate and output the lance-lifting command.

[0071] Furthermore, the furnace maintenance mode is a pre-selected control strategy by the operator based on the actual erosion condition of the converter lining, including but not limited to a thin bottom, severe wear in the molten pool area, or an excessively thick slag layer at the bottom. This strategy includes three modes: bottom expansion mode, molten pool protection mode, and bottom reduction mode. For example, if the converter bottom erosion is severe and needs to be increased in thickness, the "bottom expansion mode" is selected; if the molten pool area wear needs to be balanced and protected, the "molten pool protection mode" is selected; and if the slag layer at the bottom is too thick and its growth needs to be controlled, the "bottom reduction mode" is selected. Based on the selected mode, the corresponding lance lifting trigger condition is extracted from the mode-parameter mapping table.

[0072] Furthermore, the lance-lifting trigger conditions corresponding to the furnace maintenance mode include: The furnace bottom raising mode: The lance raising process can be divided into multiple stages, each with different trigger conditions to achieve step-by-step splashing and reinforcement of the furnace bottom. As a preferred implementation, the lance raising process in this mode can be divided into two stages: the first lance raising sets a higher slag amount threshold, and the second lance raising sets a lower slag amount threshold. If the current slag amount is less than or equal to the first lance raising slag amount threshold, a first lance raising command is generated and output to induce initial splashing of slag towards the furnace bottom area; if the current slag amount is less than or equal to the second lance raising slag amount threshold, a second lance raising command is generated and output, thereby achieving sufficient splashing and reinforcement of the furnace bottom and ending automatic slag splashing. The lance raising height and holding time corresponding to the first lance raising command are greater than those corresponding to the second lance raising command. Molten pool protection or furnace bottom lowering mode: Although these two modes have different application scenarios, the lance lifting logic is the same. As a preferred implementation method, both modes adopt a single lance lifting method and set a moderate slag amount threshold. When the current slag amount is less than or equal to the third lance lifting slag amount threshold, a lance lifting command is generated and output, ending slag splashing.

[0073] The relationship between the threshold values ​​for the amount of slag removed from the spear is as follows: the second threshold value for the amount of slag removed from the spear < the third threshold value for the amount of slag removed from the spear < the first threshold value for the amount of slag removed from the spear.

[0074] By using the differentiated control of the above three modes, the refined maintenance needs under different furnace conditions can be met, effectively extending the service life of the converter.

[0075] For example, refer to Figure 2 The overall process of the solution involved in the embodiments of this application is as follows: First, sequential image data of the furnace mouth during the converter slag splashing process is acquired, and slag formation is identified in the furnace mouth images based on a deep learning model. The model outputs the current slag splashing stage category (no slag formation or slag formation). Next, branching processing is performed based on the identification result. If the identification result is "no slag formation," the process returns to acquire the next frame image and repeats the identification, without performing slag quantity detection. If the identification result is "slag formation," the slag quantity detection process is initiated to acquire real-time slag quantity data. Then, the slag quantity data is compared with the lance trigger threshold: if the real-time slag quantity is less than or equal to the lance trigger threshold, a lance trigger command is generated and output; otherwise, the output of the lance trigger command is paused. The oxygen lance position height is then monitored in real-time: if the lance position height has not decreased to the preset process switching threshold, the process returns to acquire images for slag quantity detection and lance trigger judgment; otherwise, the process switches to the lance lifting judgment process. Finally, according to the lance lifting trigger condition corresponding to the furnace type maintenance mode, a lance lifting command is output.

[0076] Unlike traditional methods that continuously identify slag blocks and calculate parameters in images regardless of whether slag is forming at the furnace opening, this invention provides an automatic slag splashing control method for converters based on deep learning. By using a deep learning model to perform binary classification judgment on the slag splashing stage, slag quantity detection is only performed after it is confirmed that the slag forming state has been entered. This effectively avoids the false detection problem caused by interference from fire and smoke in the non-slag forming state, and overcomes the core pain points of traditional slag splashing detection technology, such as detection redundancy, poor accuracy, and control lag.

[0077] This invention innovatively introduces a hierarchical processing logic of "first judging the stage, then measuring the slag quantity," using state recognition as a prerequisite for slag quantity detection. This algorithmically shields against interference from various scenarios in the non-slag-forming state, fundamentally eliminating invalid and false detections and significantly improving the accuracy and validity of slag quantity data. Simultaneously, this pre-screening mode greatly reduces invalid computational processes, lowers model computational power consumption and equipment operating load, and ensures accurate matching of subsequent slag splashing control parameters. This enables refined, stable, and intelligent control of the steelmaking slag splashing process, effectively improving steelmaking production quality and operational stability. Subsequently, through explicit image processing procedures such as filtering, segmentation, connected component analysis, and geometric constraint screening, accurate and quantifiable real-time slag quantity data is obtained. This data is then compared with a threshold to trigger the lance pressing command, achieving closed-loop precise control based on real-time slag conditions and significantly improving the accuracy of slag splashing control.

[0078] On the other hand, this application also provides a deep learning-based automatic slag splashing control system for executing any of the above-mentioned deep learning-based converter automatic slag splashing control methods, the system comprising: An image acquisition module, connected to an image processing module, is used to acquire images of the furnace mouth during the converter slag splashing process. This module preferably includes an industrial camera and a cooling and protection device, which is used to purge and cool the camera to prevent damage from high temperatures.

[0079] The image processing module is connected to both the image acquisition module and the decision module. It controls the start and stop of the image acquisition module according to the control signals issued by the decision module; acquires furnace mouth images at a preset frequency according to the control parameters issued by the decision module; performs slag identification and slag quantity detection on the furnace mouth images; and uploads the identification results to the decision module.

[0080] The decision-making module, connected to both the image processing module and the execution module, determines the smelting status of the converter based on PLC signals uploaded by the execution module. These PLC signals include, but are not limited to, nitrogen valve status and oxygen lance height signals. Based on the determination result, it sends control signals (start or stop acquisition) and control parameters to the image processing module. It also receives the recognition results uploaded by the image processing module, generates lance-pressing or lance-lifting commands, and sends these commands to the execution module.

[0081] The decision-making module also integrates human-computer interaction functions, allowing operators to set furnace maintenance modes (raising the furnace bottom / protecting the molten pool / lowering the furnace bottom) and adjust control parameters during the slag splashing process. These control parameters include, but are not limited to, slag-raising identification threshold, lance-pressing trigger threshold, and lance-lifting trigger threshold. Simultaneously, it displays the slag volume change curve and slag splashing status in real time, enabling visualized monitoring of slag-raising identification and lance-pressing effects. The decision-making module also integrates storage management functions to retain historical furnace image data, model parameters, control logs, and process parameters, providing data support for model optimization and process analysis.

[0082] The execution module is connected to both the decision module and the external PLC. It is used to collect external nitrogen valve status signals and oxygen lance position height signals, and upload the signals to the decision module. It also receives lance pressing or lifting commands from the decision module and sends the commands to the external PLC controller to drive the oxygen lance actuator to perform the corresponding lifting and lowering actions.

[0083] Furthermore, the execution module sends instructions to the external PLC to drive the oxygen lance actuator to perform corresponding actions, including: Before receiving the slag removal instruction, the oxygen lance performs a preset reciprocating motion to uniformly cool the furnace opening; Upon receiving the slag removal command, the oxygen lance is raised to the preset slag removal lance height and awaits the lance pressing command. Upon receiving the lance lowering command, the oxygen lance performs a descent action according to the preset descent height; during the descent action, receiving new control commands is paused. Upon receiving the lance lifting command, the oxygen lance executes the corresponding lifting action based on the currently selected furnace type maintenance mode: If the currently selected mode is the furnace bottom raising mode, after receiving the first lance raising command, the oxygen lance will be raised to the first preset height; after receiving the second lance raising command, the oxygen lance will be raised to the second preset height, and slag splashing will end. If the currently selected mode is the molten pool protection mode or the furnace bottom lowering mode, then control the oxygen lance to rise to the second preset height and end the slag splashing.

[0084] The above system is created based on the above slag control method. Its technical function and beneficial effects will not be repeated 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.

[0085] The embodiments described above are merely illustrative of several implementation methods of this application, 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 this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for automatic slag splashing control in a converter based on deep learning, characterized in that, include: S1: Real-time acquisition of the current furnace mouth image during the converter slag splashing process; S2: Input the current furnace opening image into the trained deep learning model to determine whether the current slag splashing type is a non-slag-forming state or a slag-forming state; Deep learning models include: The input layer is used to input the current furnace opening image; The backbone network consists of shallow, medium, and high-level networks with progressively decreasing convolutional kernel sizes, as well as a fusion network. The shallow network is used to extract the morphological features of flames and smoke; the medium network is used to extract the texture features of slag particles; the high-level network is used to extract the motion blur features during the dynamic slag splashing process; and the fusion network is used to fuse morphological features, texture features, and motion blur features to obtain deep features. The output layer is used to output the results of the non-slag-forming state and the slag-forming state based on the deep features; S3: If the slag is not rising, return to step S1; if the slag is rising, determine the current slag amount based on the current furnace opening image. S4: Based on the current slag volume, control the oxygen lance to perform the corresponding processing action.

2. The deep learning-based automatic slag splashing control method for converters according to claim 1, characterized in that, The input layer is used to receive and preprocess the current furnace opening image to obtain an image tensor. The backbone network consists of multiple layers of alternating convolutional and pooling layers, forming a hierarchical structure of shallow, mid, and high-level networks; and the high-level networks incorporate a CA coordinate attention module and / or a channel attention module. The CA coordinate attention module is used to redistribute the spatial dimension weights of the input features of the high-level network, reduce the feature weights of large-area flame and smoke regions, and amplify the weights of local particle textures in the slag. The channel attention module is used to perform global channel dimension statistics and weight redistribution on the output features of high-level networks, automatically enhance the effective feature channel weights including slag tails and dynamic blurred trajectories, and suppress the ineffective feature channel weights corresponding to firelight and smoke. The output layer is used to flatten the deep features into a one-dimensional feature vector and output the results of the non-slag-forming state and the slag-forming state.

3. The deep learning-based automatic slag splashing control method for converters according to claim 2, characterized in that, The shallow network is configured with the first convolutional kernel, which, with its large receptive field covering the entire image area, extracts large-area spatially continuous, wide and high-dimensional globally uniform features of the image as flame and smoke morphological features; the corresponding pooling layer completes downsampling and dimensionality reduction. The middle layer network is configured with a second convolutional kernel, which relies on local small receptive fields to capture pixel grayscale differences and arrangement patterns, and extracts the surface roughness of the slag block as the slag block particle texture features. The accompanying pooling layer completes the filtering of redundant noise in the image; The high-level network learns high-order semantic features based on the feature information obtained by the preceding multi-layer convolution and pooling iterative downsampling, identifies the trailing and blurred trajectory generated by the high-speed splashing of slag blocks, and extracts motion blur features in the splashing process. The accompanying pooling layer completes the final dimensionality compression.

4. The deep learning-based automatic slag splashing control method for converters according to claim 2, characterized in that, The CA coordinate attention module performs global pooling on the input features of the high-level network in the vertical and horizontal independent dimensions to obtain global context information in the height and width directions, generate height attention weight maps and width attention weight maps, and then complete the redistribution of spatial weights. The channel attention module is used to perform global average pooling on the output features of the high-level network, compressing the spatial information into a one-dimensional channel vector; then, two cascaded fully connected layers are used to mine channel dependencies and generate importance weights for each channel; and feature recalibration is completed using the channel importance weights.

5. The deep learning-based automatic slag splashing control method for converters according to claim 1, characterized in that, The training process of the deep learning model includes: Acquire furnace mouth image data of at least two furnace bases during the slag splashing process to construct a multi-source dataset; the multi-source dataset includes negative samples of non-slag-splashing images under typical interference conditions and positive samples of slag-splashing images; typical interference conditions include any one or more of the following: smoke surrounding, local overexposure of the image, and molten steel gushing. Based on historical and recent datasets from multiple sources, an incremental training dataset is constructed using an incremental learning approach. The deep learning model is fine-tuned and updated based on the optimization objective function and the incremental training dataset. By jointly training across furnaces, the loss functions of each furnace are summed to minimize the total loss of each furnace and the model parameters are updated to obtain a deep learning model that is adapted to all converters in the plant.

6. The deep learning-based automatic slag splashing control method for converters according to claim 1, characterized in that, If the furnace is in a slagging state, the steps to determine the current slagging amount based on the current furnace opening image include: The current furnace opening image is preprocessed with noise reduction and smoothing to filter out furnace dust noise and preserve the edge features of the furnace opening steel shell; The preprocessed image is decomposed into color channels, and the color channel whose contrast meets the preset conditions is selected as the channel to be processed. Based on the channel to be processed, the region is segmented, and candidate regions that conform to the geometric size characteristics of the furnace opening are extracted. Then, the effective region of the furnace opening is dynamically updated based on the boundary coordinates of the candidate regions. Effective slag blocks are selected within the effective area of ​​the furnace opening, and the current slag quantity is calculated.

7. The deep learning-based automatic slag splashing control method for converters according to claim 1, characterized in that, The steps for real-time acquisition of the current furnace mouth image during the converter slag splashing process include: When the state is determined to be non-slag-forming, the image acquisition device is controlled by the first imaging strategy to acquire the current furnace mouth image; the first imaging strategy includes reducing the aperture or shortening the exposure time to suppress highlights and highlight the smoke texture features; When the slag-raising state is determined, the image acquisition device is controlled by the second imaging strategy to acquire the current furnace mouth image; the second imaging strategy includes increasing the exposure gain or extending the exposure time to ensure that the clear trajectory of the high-speed moving slag block is captured.

8. The deep learning-based automatic slag splashing control method for converters according to any one of claims 1 to 7, characterized in that, Based on the current slag volume, the steps for controlling the oxygen lance to perform the corresponding processing actions include: The processing actions include lowering and raising. The lowering action is the action performed by controlling the oxygen lance through the press-down command, and the raising action is the action performed by controlling the oxygen lance through the lift-up command. The current slag volume is compared with the preset lance trigger threshold. When the current slag volume is less than or equal to the lance trigger threshold, a lance trigger command is generated and output to control the oxygen lance to perform a descent action; otherwise, the output of the lance trigger command is paused. After generating and outputting the lance control command to control the oxygen lance to perform the descent action, the lance position height is monitored in real time. Compare the gun position height with the preset process switching threshold. When the gun position height is greater than or equal to the process switching threshold, return to execute step S1. Otherwise, by judging the process of raising the lance, a lance-raising command is generated and output to control the oxygen lance to perform the upward movement.

9. The deep learning-based automatic slag splashing control method for converters according to claim 8, characterized in that, The process of generating and outputting a lifting command to control the oxygen lance to perform an upward movement, based on the lifting judgment process, includes the following steps: Get the currently selected furnace type maintenance mode. The furnace type maintenance modes include furnace bottom raising mode, molten pool protection mode, and furnace bottom lowering mode. If the currently selected mode is the furnace bottom raising mode, the lance lifting process will be divided into at least two stages, with different slag amount thresholds set for each stage, and the process will be judged sequentially for each stage, including: When the current slag amount is less than or equal to the first slag amount threshold for lifting the gun, the first gun lifting command is generated and output; when the current slag amount is less than or equal to the second slag amount threshold for lifting the gun, the second gun lifting command is generated and output. The lifting height and holding time corresponding to the first gun lifting command are greater than the lifting height and holding time corresponding to the second gun lifting command. If the currently selected mode is the molten pool protection mode or the furnace bottom lowering mode, then when the current slag amount is less than or equal to the third lance lifting slag amount threshold, a lance lifting command will be generated and output. The relationship between the threshold values ​​for the amount of slag removed from the spear is as follows: the second threshold value for the amount of slag removed from the spear is less than the third threshold value for the amount of slag removed from the spear, which is less than the first threshold value for the amount of slag removed from the spear.

10. A deep learning-based automatic slag splashing control system for converters, characterized in that, The automatic slag splashing control method for converters based on deep learning, as described in any one of claims 1 to 9, includes: The image acquisition module, connected to the image processing module, is used to acquire the current furnace mouth image during the converter slag splashing process; The image processing module is connected to the image acquisition module and the decision module respectively. It is used to receive control parameters and control signals sent by the decision module, perform slag identification and slag quantity detection on the current furnace mouth image, and upload the identification results to the decision module. The decision module is connected to the image processing module and the execution module respectively. It is used to determine the smelting status based on the PLC signal uploaded by the execution module, send control parameters and control signals to the image processing module, receive the recognition results, generate the gun pressing command or gun lifting command, and send the command to the execution module. The execution module is connected to both the decision module and the external PLC controller. It is used to collect the nitrogen valve status signal and oxygen lance position height signal uploaded by the external PLC controller, upload the signal to the decision module, receive the instructions fed back by the decision module, and send the instructions to the external PLC controller.

Images