Converter splashing prediction method and device based on normal label and weighted convolution time sequence

Abstract

The application provides a converter spatter prediction method and device based on normal labels and weighted convolution timing, and relates to the technical field of converter steelmaking. The method comprises the following steps: acquiring flame video data of a converter steelmaking process and performing frame segmentation to obtain image data, constructing a normal distribution curve based on the principle of normal distribution, calculating the spatter intensity value at each moment, and then obtaining a continuous spatter intensity data set; performing aggregation and sampling processing on the continuous spatter intensity data set to obtain a processed data set; constructing a deep fusion regression model based on CNN-BiLSTM and an attention mechanism, designing a high-value weighted mean square error loss function, training the model according to the processed data set and the high-value weighted mean square error loss function, and obtaining a spatter intensity prediction value according to the trained deep fusion regression model. The application can realize accurate quantification of spatter intensity and improve the timeliness and accuracy of spatter early warning.

Description

Technical Field

[0001] This invention relates to the field of converter steelmaking technology, and in particular to a method and apparatus for predicting converter splashing based on normal labels and weighted convolution time series. Background Technology

[0002] As a fundamental material in the modern industrial system, steel is widely used in infrastructure construction, equipment manufacturing, energy and transportation, and high-end manufacturing, among other fields. Its output and quality directly impact the operation of the national economy and the stability of the industrial system. Converter steelmaking, with its efficient oxidation reaction mechanism and mature process system, has become one of the main steelmaking methods in current steel production. This process utilizes high-purity oxygen injected at high speed to oxidize and remove impurities in molten iron in a short time, exhibiting significant advantages such as high reaction intensity, fast production pace, and low unit energy consumption, effectively supporting large-scale continuous production.

[0003] Converter steelmaking is a complex multiphase reaction process under high temperature and high pressure, which is highly susceptible to splashing. The splashing of hot molten steel and slag not only causes losses but can also lead to safety accidents such as burns. More importantly, changes in splash intensity are directly related to the intensity of the reaction inside the furnace and can indirectly reflect the stability of the furnace operating conditions. This is an important basis for adjusting process parameters such as oxygen supply flow rate and charging rhythm. Low-intensity splashing can maintain normal operation, while high-intensity splashing requires immediate intervention and control.

[0004] In the early stages of industrialized converter steelmaking, splash identification and prediction relied on the on-site experience of operators, lacking specialized detection equipment. Experienced steelworkers judged the intensity of the internal reaction by visually observing the color, shape, and frequency of the flame at the furnace mouth, as well as the intensity of the flue gas emission, thus predicting the likelihood of splashing. The advantage of this method was that it required no equipment investment, provided immediate response, and could handle sudden changes in furnace conditions. However, its drawbacks were also apparent: differences in experience and judgment among operators, inconsistent judgment standards, and the high labor intensity of operators exposed to high temperatures and noise in the industrial environment led to poor stability in prediction and identification.

[0005] In the early stages of industrial automation, the industry used physical sensors to replace manual vision for predicting and identifying splashes. For example, pressure and gas flow sensors were installed in the flue gas duct of a converter to collect flue gas parameters in real time and compare them with preset thresholds for judgment. Patent CN113564296A discloses a splash warning method, system, and industrial control equipment for steelmaking converters. This includes constructing a safe range for smelting time and carbon monoxide concentration based on historical furnace data, using sensors to monitor the carbon monoxide concentration in the current furnace flue gas in real time; below the lower limit indicates carbon-oxygen reaction inhibition, and above the upper limit indicates a vigorous reaction, generating corresponding warnings and controlling the oxygen lance to raise or lower the flow rate to avoid splashing. This technology achieves objectivity and automation in splash monitoring, reduces manual labor intensity, and significantly improves stability compared to the manual experience stage. However, the physical parameters detected by the sensors are indirect representations of splashes, and interference such as uneven raw material composition and abnormal slag thickness can easily cause parameter changes. This type of technology still cannot quantify splash intensity, only perform qualitative judgments, and the sensors operate in high-temperature and highly corrosive environments for extended periods, resulting in a high failure rate and requiring regular maintenance and calibration.

[0006] With the rise of machine vision technology, the visual characteristics of splashes are being utilized to achieve more intuitive splash prediction and recognition. This stage typically involves installing a high-temperature resistant industrial camera at the furnace opening to acquire color images of the furnace opening area in real time. Then, the acquired raw images are processed using algorithms, retaining the furnace flame and potential splash areas as regions of interest. Geometric and grayscale features of these regions are extracted using traditional image processing algorithms, and these features are input into a pre-set decision tree or SVM (Support Vector Machine) classification model. The model outputs a judgment result. For example, patent CN117496429A discloses a method, device, equipment, and storage medium for monitoring the splash state of a converter flame. This includes first optimizing the brightness of converter flame images to obtain clear candidate images, then generating a flame image dataset based on the labeled candidate images, and training a flame state recognition model. This model focuses on extracting the flame shape features of flame sub-images in the candidate images and establishing a mapping relationship between flame shape and splash state. Further, the system acquires initial images of the target area of ​​the converter in real time, adjusts the brightness, and inputs them into the trained neural network flame state recognition model. The model then identifies the flame shape and matches it with the corresponding splash state. Patent CN110309973A discloses a method and system for predicting converter splashing based on a video intelligent algorithm. The method includes first acquiring video of the furnace mouth flame using an industrial camera, constructing a dataset with the first N frames as input and the next M frames marked as splashes, and preprocessing this dataset by cropping the core region, using Gaussian and median dual filtering for noise reduction, and brightness normalization. Then, a CNN-RNN (Convolutional Neural Networks-Recurrent Neural Network) fusion architecture is used to extract static features such as flame brightness, texture, and morphology. The RNN is used to capture temporal dependencies, achieving feature extraction and fusion. The model is trained using a cross-entropy loss function, an Adam optimizer, and a Dropout layer, iterating until the loss converges. Finally, features are extracted in real-time through a sliding window, and the model outputs the splashing probability. When the output continuously reaches a threshold, an early warning is triggered, thus predicting splashing to support adjustments to process parameters.

[0007] With breakthroughs in deep learning technology in computer vision, splash recognition and prediction technologies have entered a new stage. The core improvement lies in automatically extracting image features through deep neural networks and introducing temporal models to learn the dynamic evolution of splashes. For example, patent CN121170822A discloses a metallurgical multimodal image classification method based on adaptive marginal learning and cross-modal enhancement. This method first expands the splash state category labels to process description text based on the characteristics of the converter steelmaking BOF process, pairing them with collected splash images to construct an image-to-text dataset, while preprocessing the data according to small-sample learning requirements. Visual features of the images are extracted using a pre-trained ResNet50, and semantic features of the text are extracted using a pre-trained Transformer. Then, a dual-adaptor network is used to enhance the representation of the two types of features respectively, and a cross-modal attention mechanism is introduced to achieve text-guided image feature fusion, strengthening the semantic discrimination ability of visual representations. The system constructs an adaptive marginal generator based on the semantic similarity between categories, dynamically adjusting the classification boundary, and using an adaptive marginal loss function to replace the traditional cross-entropy loss, imposing stricter separation constraints on semantically similar categories. Finally, the fused image features and text features are L2 normalized, and the image-text similarity score is calculated by adjusting the temperature parameter. The score is then converted into a probability distribution using the softmax function, and the classification result of the splashing state is output, thereby realizing the prediction and recognition of the converter splashing state.

[0008] Existing converter splash prediction and identification technologies have gone through three development stages: human experience, physical sensing, and machine vision and deep learning. The core technology paths and application characteristics have been iterated step by step, but there are still significant limitations.

[0009] The manual experience stage relies on operators to observe the characteristics of the flame color, shape and flue gas intensity at the furnace opening with their naked eyes, and predict the splash based on experience. Although it does not require equipment investment and has an immediate response, the judgment criteria are greatly affected by individual experience differences and lack uniformity. In addition, operators are in a high-temperature and high-noise environment for a long time, and the high labor intensity leads to poor recognition stability.

[0010] In the early stages of industrial automation, physical sensors such as pressure and gas flow rate were used to collect flue gas parameters inside the furnace. Automated monitoring and early warning were achieved through threshold comparison. For example, Chopin et al. used carbon monoxide concentration to warn of splashing. While this reduced manual labor and improved stability, the physical parameters detected by the sensors were only indirect representations of splashing. They were susceptible to interference from factors such as uneven raw material composition and abnormal slag thickness, and could not quantify splashing intensity, only providing qualitative judgment. Furthermore, the sensors had a high failure rate in high-temperature and highly corrosive environments, requiring frequent maintenance and calibration.

[0011] In the machine vision stage, image features are extracted through neural networks to achieve accurate recognition, such as recognition models based on flame brightness optimization and shape feature matching, and video prediction models with CNN-RNN fusion architecture. Although these have significantly improved recognition accuracy and generalization ability, they still have limitations: 1) Lack of splash intensity quantification capability: Existing methods are limited to binary classification or discrete categories, and cannot output continuous intensity values ​​or represent the gradient change of splash probability.

[0012] 2) Imperfect temporal modeling and feature fusion: Many methods fail to capture the temporal dynamics of flame morphology, are limited to feature extraction from single-frame and multi-frame images, and do not pay attention to the temporal correlation between images per second.

[0013] 3) Poor fit between samples and loss function: The loss function was not optimized for the uneven distribution of splash samples, resulting in poor prediction performance in high splash intensity scenarios. The dataset did not fit the splash intensity changes and time distribution.

[0014] For example, patent CN117496429A determines the splashing state by extracting features from a single-frame flame image. This method only mines the static visual features of the flame and fails to capture the temporal dynamics of the flame morphology during the blowing process. It cannot quantitatively characterize the splashing intensity and can only complete state recognition after splashing occurs, resulting in significant judgment lag. Patent CN110309973A proposes a CNN-RNN fusion video prediction method. Although it introduces temporal modeling logic to achieve splashing probability prediction and threshold warning, compensating for the lag problem of static recognition, the output of this method is still a binary classification probability judgment result. It does not achieve continuous value regression prediction of splashing intensity and can only determine whether splashing has occurred, but cannot characterize the intensity and gradient changes of splashing. Its feature fusion is a simple vector concatenation form, without focusing on enhancing the key frame features of splashing, resulting in insufficient prediction accuracy for low-intensity, gradual splashing. Meanwhile, the general cross-entropy loss function was used, but it was not optimized for the non-equilibrium distribution of splash samples. The prediction effect was poor in high splash intensity scenarios, and the dataset did not fit the actual intensity variation of splashes.

[0015] As the model is further optimized and integrates visual and physical parameter data, it enters the deep learning stage, utilizing large deep learning models to achieve early warning and automatic control. While the cross-modal enhanced classification method proposed in patent CN121170822A solves the problem of distinguishing semantically similar splash categories through image-text dual-modal fusion, this method is essentially still a discrete classification task, outputting a probability distribution of fixed categories. It cannot output splash intensity values ​​in a continuous range from 0 to 1, making it difficult to match the quantitative requirements for splash intensity in industrial settings. Furthermore, it lacks temporal modeling capabilities, only performing state classification on single-frame images, and the dataset does not specifically construct the temporal distribution and intensity gradient of splashes. Summary of the Invention

[0016] To address the technical problems of existing converter splash prediction methods, such as lack of splash intensity quantification capability, imperfect temporal modeling and feature fusion, and poor sample-loss function compatibility, this invention provides a converter splash prediction method and apparatus based on normal labels and weighted convolutional temporal series. The technical solution is as follows: On the one hand, a converter splash prediction method based on normal labels and weighted convolution time series is provided. This method is implemented by a converter splash prediction device and includes: S1. Obtain flame video data of the converter steelmaking process, segment the flame video data into frames to obtain image data, construct a normal distribution curve for the image data based on the principle of normal distribution, calculate the splash intensity value at each moment according to the normal distribution curve, and then obtain a continuous splash intensity dataset; wherein, the continuous splash intensity dataset includes several frames of image data and the splash intensity label corresponding to each frame of image data.

[0017] S2. Aggregate the splash intensity labels in the continuous splash intensity dataset to obtain a second-level splash intensity dataset. Use an asymmetric sampling strategy to sample the second-level splash intensity dataset to obtain the processed dataset.

[0018] S3. Construct a deep fusion regression model based on CNN-BiLSTM and attention mechanism, design a high-value weighted mean square error loss function, train the deep fusion regression model based on the processed dataset and the high-value weighted mean square error loss function, and obtain the trained deep fusion regression model.

[0019] S4. Obtain the flame video data to be predicted, and obtain the splash intensity prediction value based on the flame video data to be predicted and the trained deep fusion regression model.

[0020] Optionally, in S1, a normal distribution curve is constructed based on the principle of normal distribution for the image data. The splash intensity value at each moment is calculated according to the normal distribution curve, thus obtaining a continuous splash intensity dataset, including: The splashing time, splashing stage, and standard deviation in the image data are determined. The peak splashing time is used as the mean to generate a normal distribution curve. The splashing intensity value at each time is calculated based on the normal distribution curve. The splashing intensity value is normalized. Several splashing probability intensity curves are obtained based on the normalized splashing intensity value and the intensity weight of the splashing stage. Several splashing probability intensity curves are fused to obtain a continuous splashing intensity dataset. The splashing stages include small splashes, medium splashes, and large splashes.

[0021] Optionally, the formula for calculating the splash probability intensity curve is as shown in equation (1) below: (1) In the formula, Represents any moment on the timeline The splash intensity value, Indicates the first The standard deviation of a splash event Indicates the first The peak splash time of a splash event Indicates the first The intensity weight of the splash phase to which each splash event belongs.

[0022] Optionally, the construction in S3 is based on a deep fusion regression model using CNN-BiLSTM and an attention mechanism, including: Static visual features are extracted from each frame of image data in the processed dataset using a pre-trained CNN model to obtain the feature vector of each frame. The feature vectors of all frames corresponding to each second are then concatenated to obtain the concatenated features.

[0023] The concatenated features are input into a bidirectional long short-term memory network temporal feature learner to obtain the temporal dependencies and temporal context information of the features per second.

[0024] The temporal features of all frames are weighted and aggregated according to attention weights to obtain the aggregated global feature vector. The aggregated global feature vector is then input into the regression head to obtain the splash intensity prediction value.

[0025] Optionally, the high-value weighted mean squared error loss function in S3 includes: In the backpropagation process of the deep fusion regression model, weighted mean square error is used to assign differentiated weights to the samples in the processed dataset.

[0026] The high-value weighted mean square error loss function is shown in equation (2) below: (2) In the formula, This represents the high-value weighted mean square error loss function. Indicates the number of samples in the batch. Indicates the dynamic weighting coefficient. Indicates the first The true label of each sample Indicates the first Predicted splash intensity values ​​for each sample.

[0027] On the other hand, a converter splash prediction device based on normal labels and weighted convolution time series is provided. This device is applied to a converter splash prediction method based on normal labels and weighted convolution time series. The device includes: The data acquisition module is used to acquire flame video data of the converter steelmaking process, segment the flame video data into frames to obtain image data, construct a normal distribution curve for the image data based on the normal distribution principle, calculate the splash intensity value at each moment based on the normal distribution curve, and thus obtain a continuous splash intensity dataset; wherein, the continuous splash intensity dataset includes several frames of image data and the splash intensity label corresponding to each frame of image data.

[0028] The data processing module is used to aggregate the splash intensity labels in the continuous splash intensity dataset to obtain a second-level splash intensity dataset. An asymmetric sampling strategy is used to sample the second-level splash intensity dataset to obtain the processed dataset.

[0029] The training module is used to construct a deep fusion regression model based on CNN-BiLSTM and attention mechanism. A high-value weighted mean square error loss function is designed. The deep fusion regression model is trained based on the processed dataset and the high-value weighted mean square error loss function to obtain the trained deep fusion regression model.

[0030] The output module is used to acquire the flame video data to be predicted and to obtain the splash intensity prediction value based on the flame video data to be predicted and the trained deep fusion regression model.

[0031] On the other hand, a converter splash prediction device is provided, the converter splash prediction device comprising: a processor; a memory storing computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, any one of the above-described converter splash prediction methods based on normal labels and weighted convolution time series is implemented.

[0032] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described methods for predicting converter splashing based on normal labels and weighted convolutional timing.

[0033] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention overcomes the limitations of existing splash dataset construction. Based on the physical characteristics of converter splash, it generates a normal distribution curve with the splash peak time as the mean by pre-calibrating the splash occurrence time, intensity weight, and attenuation coefficient. This yields continuous normalized splash intensity values, constructing a continuous splash intensity dataset that fits the actual smelting scenario. Addressing the shortcomings of previous models that could only qualitatively identify splash intensity and failed to explore the correlation between flame images, this invention constructs a deep fusion regression model based on CNN-BiLSTM and an attention mechanism. It aggregates images into second-level features, divides video frames by second, and constructs samples using the continuous intensity values ​​generated from the dataset as second-level labels. Simultaneously, it samples zero-splash samples to solve the sample imbalance problem and introduces weighted temporal features to assign higher weights to moments with high splash probability. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of a converter splash prediction method based on normal labels and weighted convolution time series provided by an embodiment of the present invention; Figure 2 This is a structural diagram of a CNN-BiLSTM attention model provided in an embodiment of the present invention; Figure 3 This is an overall structural diagram of the model provided in the embodiment of the present invention; Figure 4 This is a block diagram of a converter splash prediction device based on normal labels and weighted convolution time series provided by an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a converter splash prediction device provided in an embodiment of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0037] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0038] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent.

[0039] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0040] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0041] This invention provides a converter splash prediction method based on normal labels and weighted convolution time series. This method can be implemented by a converter splash prediction device, which can be a terminal or a server. Figure 1 The flowchart shown is for a converter splash prediction method based on normal labels and weighted convolution time series. The processing flow of this method may include the following steps: S1. Obtain flame video data of the converter steelmaking process, segment the flame video data into frames to obtain image data, construct a normal distribution curve for the image data based on the principle of normal distribution, calculate the splash intensity value at each moment according to the normal distribution curve, and then obtain a continuous splash intensity dataset; wherein, the continuous splash intensity dataset includes several frames of image data and the splash intensity label corresponding to each frame of image data.

[0042] Optionally, in S1, a normal distribution curve is constructed based on the principle of normal distribution for the image data. The splash intensity value at each moment is calculated according to the normal distribution curve, thereby obtaining a continuous splash intensity dataset, specifically including: The splashing time, splashing stage, and standard deviation in the image data are determined. The peak splashing time is used as the mean to generate a normal distribution curve. The splashing intensity value at each time is calculated based on the normal distribution curve. The splashing intensity value is normalized. Several splashing probability intensity curves are obtained based on the normalized splashing intensity value and the intensity weight of the splashing stage. Several splashing probability intensity curves are fused to obtain a continuous splashing intensity dataset. The splashing stages include small splashes, medium splashes, and large splashes.

[0043] In one feasible implementation, the present invention first constructs a splash intensity curve. Specifically, in order to achieve splash prediction rather than simply splash identification, since classification models can only determine whether splash exists but cannot quantify the probability of splash, a regression model is chosen to achieve splash prediction. In this case, it is necessary to label the probability of splash, rather than simply classifying the splash.

[0044] The splashing during converter steelmaking is essentially a dynamic process of molten iron reacting violently with oxygen and slag being ejected from the furnace mouth. The intensity of this splashing, changing over time, conforms to the mathematical characteristics of a normal distribution. There is a clearly defined core splashing moment; the increase and decrease in intensity are processes, not instantaneous events, and are continuously changing. Therefore, a normal distribution curve can be used to simulate the change in the intensity of steelmaking splashing over time.

[0045] Based on on-site investigations and relevant theoretical knowledge, the degree of splashing can be roughly divided into three stages: Small splash: mainly consists of a small amount of steel slag particles or sparks, not forming a continuous slag flow, with a short duration and little impact on production. Medium splash: the amount of splashed material increases significantly, resulting in a mixed form of small slag streams, with noticeable flames shooting outwards from the furnace mouth, and a longer duration. Large splash: a large amount of splashed material is ejected from the furnace mouth, with large size, high speed, and long duration.

[0046] Therefore, constructing a single splash probability intensity curve requires three steps. First, after determining the peak point and approximate range of influence of the splash, the formula for the normal distribution is used to generate the curve for any time point on the calculation time axis. Splash intensity value Then normalize, and finally multiply by the weights to get the complete curve:

[0047] In the formula, Indicates the first The standard deviation of a splash event The larger the value, the larger the denominator, the flatter the curve, and the longer the spray duration; Indicates the first The peak time of each splash event. The intensity setting varies depending on the splash level; high, medium, and low splashes have different weights, with higher splash levels receiving higher weights. After weighting, the peak value of the curve is affected by the weights. This influences and allows for differentiation of the intensity of different splash events.

[0048] Furthermore, in the actual converter steelmaking process, there is more than one peak of the large spray, and the splashing also occurs more than once. Therefore, it is necessary to fuse multiple splashing curves. After the curves are fused, data cleaning is also required to filter out invalid noise and standardize accuracy.

[0049] S2. Aggregate the splash intensity labels in the continuous splash intensity dataset to obtain a second-level splash intensity dataset. Use an asymmetric sampling strategy to sample the second-level splash intensity dataset to obtain the processed dataset.

[0050] In one feasible implementation, the present invention further involves data processing. Specifically, unlike most models that are trained directly using images, the model of the present invention learns through temporal image features, which utilizes the correlation between images per second, thus requiring data preprocessing first.

[0051] The model's prediction scale is on the order of seconds, but the original labels are on the order of frames, so the labels need to be aggregated:

[0052] in It is the first Second-level regression labels It is the first The original frame-level label of the frame.

[0053] Because in the converter steelmaking scenario, there is generally no splashing, resulting in a very high proportion of splash-free data and a low proportion of splashed data, making it a typical imbalanced dataset. The model employs an asymmetric sampling strategy to ensure training effectiveness. All seconds with splashing are retained without sampling, ensuring sufficient learning of splashing features. Seconds without splashing are randomly sampled, retaining only a subset of splash-free samples. Furthermore, after sampling, all valid seconds are sorted by time to ensure temporal continuity.

[0054] S3. Construct a deep fusion regression model based on CNN-BiLSTM and attention mechanism, design a high-value weighted mean square error loss function, train the deep fusion regression model based on the processed dataset and the high-value weighted mean square error loss function, and obtain the trained deep fusion regression model.

[0055] Optionally, constructing a deep fusion regression model based on CNN-BiLSTM and attention mechanism in S3 may include the following steps S31-S33: S31. Static visual features are extracted from each frame of image data in the processed dataset using a pre-trained CNN model to obtain the feature vector of each frame. The feature vectors of all frames corresponding to each second are then concatenated to obtain the concatenated features.

[0056] One feasible implementation method is, for example Figure 2The diagram shows the overall framework of the CNN-BiLSTM attention model. The CNN model acts as a static frame feature extractor, independently extracting static visual features from each frame of the image sequence. It maps a single RGB image into a fixed-dimensional feature vector, stripping away spatial redundancy and retaining the core visual features related to the splashes. A pre-trained CNN is used as the backbone network. The single-frame feature extraction formula is as follows:

[0057] in, It is the first The input image of the frame, It is the feature extraction function of a pre-trained CNN. , These are the weights and biases of the feature adjustment layer. Then, the feature vectors of all frames per second are concatenated temporally to obtain the frame feature sequence, which serves as the input to the BiLSTM.

[0058] S32. Input the concatenated features into the bidirectional long short-term memory network temporal feature learner to obtain the temporal dependency and temporal context information of the features per second.

[0059] In one feasible implementation, the concatenated features are input into a bidirectional LSTM (Long Short-Term Memory) temporal feature learner to learn the temporal dependencies and temporal context information of the feature sequence per second. The bidirectional LSTM can simultaneously capture the forward dependencies from past frames to the current frame and the backward dependencies from future frames to the current frame, thus fully learning the temporal evolution of the splash. Finally, the output dimension of the bidirectional LSTM is formed by concatenating the forward and backward hidden states.

[0060] in Is it a forward LSTM in The hidden state at all times Is it a backward LSTM in The hidden state at any given moment. It is a concatenation operation that joins two outputs together.

[0061] S33. The temporal features of all frames are weighted and aggregated according to the attention weight to obtain the aggregated global feature vector. The aggregated global feature vector is then input into the regression head to obtain the splash intensity prediction value.

[0062] In one feasible implementation, after passing through the BiLSTM, the temporal features of all frames need to be weighted and aggregated to address the deficiency of equal weighting of features across all frames in the BiLSTM. Key frames are assigned high weights, while irrelevant frames are assigned low weights, improving prediction accuracy. The temporal output of the BiLSTM is then input to perform attention weight calculations.

[0063] Then, the features of all frames are aggregated based on the attention weights:

[0064] in, It is the attention-weighted global temporal feature, and it is the final feature representation of the splash intensity for that second. It is attention weight.

[0065] Finally, a regression head is used to predict splash intensity. The global feature vector after attention aggregation is mapped to a continuous splash probability value of 0~1, completing the final step of the regression task.

[0066] in, This is the attention context vector. , These are intermediate layer parameters that compress and map the dimensions of the input. , As an output layer parameter, it maps dimensions to output dimensions. It is the model's final prediction, meaning the probability of the splash intensity being continuously between 0 and 1 for that second.

[0067] Optionally, the high-value weighted mean squared error loss function in S3 includes: In the backpropagation process of the deep fusion regression model, weighted mean square error is used to assign differentiated weights to the samples in the processed dataset.

[0068] In one feasible implementation, the loss function was improved during the model's backpropagation. High splash intensity samples constitute a very small percentage, resulting in an imbalanced data label ratio. A weighted MSE (Mean-Square Error) is used to assign differentiated weights to samples in different splash intensity ranges. This ensures that the model training focuses more on the fitting accuracy of high splash samples, penalizing prediction errors related to these samples. The complete loss formula is as follows:

[0069] In the formula, This represents the high-value weighted mean square error loss function. Indicates the number of samples in the batch. This represents the dynamic weighting coefficient, which is determined by the range of the actual labels. Indicates the first The true label of each sample Indicates the first Predicted splash intensity values ​​for each sample.

[0070] This invention designs a method for generating splash intensity labels based on the physical laws of converter splashing according to normal distribution, constructs a CNN-BiLSTM attention temporal prediction model, and uses the CNN-BiLSTM attention model to learn the mapping relationship between labels and visual temporal features, accurately outputting splash intensity prediction values, thus providing a real-time prediction scheme for converter splash intensity.

[0071] S4. Obtain the flame video data to be predicted, and obtain the splash intensity prediction value based on the flame video data to be predicted and the trained deep fusion regression model.

[0072] The overall concept of this invention is as follows: Figure 3 As shown, the system comprises two parts: the construction of a continuous splash intensity dataset based on a normal distribution and a self-designed CNN-BiLSTM attention model. The system operation flow is as follows: 1) Based on the principle of normal distribution, a splash intensity curve is constructed to obtain the image-label pair dataset required for training.

[0073] 2) Data preprocessing: Frame-level labels are aggregated into second-level labels. An asymmetric sampling strategy is used to handle the sample imbalance problem while maintaining temporal continuity.

[0074] 3) Construct a CNN-BiLSTM attention model, using the constructed dataset as the training basis, and train the model by combining the improved weighted MSE loss function to achieve continuous regression prediction of splash intensity.

[0075] 4) The trained model is then applied to the actual scenario of predicting the splash intensity in converter steelmaking.

[0076] This invention proposes a method for continuous quantitative prediction of splash intensity in converter steelmaking. It integrates physically driven label generation with deep learning temporal regression technology to address the limitation of traditional binary classification prediction in quantifying splash intensity. First, based on manually labeled splash times, intensity weights, and standard deviations, a second-level continuous splash intensity label in the 0-1 range is generated using a weighted normal distribution curve overlay and maximum value fusion strategy. Then, a CNN-BiLSTM attention hybrid model is constructed. Taking steelmaking image sequences as input, visual and temporal features are extracted, and the model is trained using a high-value weighted MSE loss function. The trained model is then used to infer real-time splash intensity values ​​from new steelmaking video frame sequences. This method achieves accurate quantification of splash intensity, enabling splash prediction and identification, and improving the timeliness and accuracy of splash warnings. It lays the foundation for subsequent research on splash prediction using multi-source data fusion and the linkage between splash warnings and steelmaking process parameter adjustments.

[0077] Figure 4 This is a block diagram illustrating a converter splash prediction device based on normal labels and weighted convolution time series, according to an exemplary embodiment. The device is used for a converter splash prediction method based on normal labels and weighted convolution time series. (Refer to...) Figure 4 The device includes a data acquisition module 310, a data processing module 320, a training module 330, and an output module 340. Wherein: The data acquisition module 310 is used to acquire flame video data of the converter steelmaking process, segment the flame video data into frames to obtain image data, construct a normal distribution curve for the image data based on the principle of normal distribution, calculate the splash intensity value at each moment according to the normal distribution curve, and then obtain a continuous splash intensity dataset; wherein, the continuous splash intensity dataset includes several frames of image data and the splash intensity label corresponding to each frame of image data.

[0078] The data processing module 320 is used to aggregate the splash intensity labels in the continuous splash intensity dataset to obtain a second-level splash intensity dataset. An asymmetric sampling strategy is used to sample the second-level splash intensity dataset to obtain the processed dataset.

[0079] Training module 330 is used to construct a deep fusion regression model based on CNN-BiLSTM and attention mechanism. A high-value weighted mean square error loss function is designed. The deep fusion regression model is trained based on the processed dataset and the high-value weighted mean square error loss function to obtain the trained deep fusion regression model.

[0080] The output module 340 is used to acquire the flame video data to be predicted and obtain the splash intensity prediction value based on the flame video data to be predicted and the trained deep fusion regression model.

[0081] This invention proposes a method for continuous quantitative prediction of splash intensity in converter steelmaking. It integrates physically driven label generation with deep learning temporal regression technology to address the limitation of traditional binary classification prediction in quantifying splash intensity. First, based on manually labeled splash times, intensity weights, and standard deviations, a second-level continuous splash intensity label in the 0-1 range is generated using a weighted normal distribution curve overlay and maximum value fusion strategy. Then, a CNN-BiLSTM attention hybrid model is constructed. Taking steelmaking image sequences as input, visual and temporal features are extracted, and the model is trained using a high-value weighted MSE loss function. The trained model is then used to infer real-time splash intensity values ​​from new steelmaking video frame sequences. This method achieves accurate quantification of splash intensity, enabling splash prediction and identification, and improving the timeliness and accuracy of splash warnings. It lays the foundation for subsequent research on splash prediction using multi-source data fusion and the linkage between splash warnings and steelmaking process parameter adjustments.

[0082] Figure 5 This is a schematic diagram of the structure of a converter splash prediction device provided in an embodiment of the present invention, as shown below. Figure 5 As shown, the converter splash prediction device may include the above-mentioned Figure 4 The illustrated converter splash prediction device is based on normal labels and weighted convolution time series. Optionally, the converter splash prediction device 410 may include a first processor 2001.

[0083] Optionally, the converter splash prediction device 410 may also include a memory 2002 and a transceiver 2003.

[0084] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0085] The following is combined with Figure 5 A detailed description of each component of the converter splash prediction device 410 is provided below: The first processor 2001 is the control center of the converter splash prediction device 410. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0086] Optionally, the first processor 2001 can perform various functions of the converter splash prediction device 410 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0087] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 5 CPU0 and CPU1 are shown in the diagram.

[0088] In a specific implementation, as one example, the converter splash prediction device 410 may also include multiple processors, for example... Figure 5 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0089] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0090] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently, and may be connected via the interface circuit of the converter splash prediction device 410. Figure 5 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0091] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0092] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 5 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0093] Optionally, the transceiver 2003 can be integrated with the first processor 2001, or it can exist independently and be connected to the interface circuit of the converter splash prediction device 410. Figure 5 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0094] It should be noted that, Figure 5 The structure of the converter splash prediction device 410 shown does not constitute a limitation on the router. Actual knowledge structure identification devices may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0095] Furthermore, the technical effect of the converter splash prediction device 410 can be referred to the technical effect of the converter splash prediction method based on normal labels and weighted convolution time series described in the above method embodiments, and will not be repeated here.

[0096] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be 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. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0097] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0098] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0099] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0100] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0101] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0102] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0103] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0104] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0105] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0106] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0107] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A converter splash prediction method based on normal labels and weighted convolutional temporal series, characterized in that, The method includes: S1. Obtain flame video data of the converter steelmaking process, segment the flame video data into frames to obtain image data, construct a normal distribution curve for the image data based on the principle of normal distribution, calculate the splash intensity value at each moment according to the normal distribution curve, and then obtain a continuous splash intensity dataset; wherein, the continuous splash intensity dataset includes several frames of image data and the splash intensity label corresponding to each frame of image data; S2. Aggregate the splash intensity labels in the continuous splash intensity dataset to obtain a second-level splash intensity dataset. Use an asymmetric sampling strategy to sample the second-level splash intensity dataset to obtain the processed dataset. S3. Construct a deep fusion regression model based on CNN-BiLSTM and attention mechanism, design a high-value weighted mean square error loss function, train the deep fusion regression model according to the processed dataset and the high-value weighted mean square error loss function, and obtain the trained deep fusion regression model. S4. Obtain the flame video data to be predicted, and obtain the splash intensity prediction value based on the flame video data to be predicted and the trained deep fusion regression model.

2. The converter splash prediction method based on normal labels and weighted convolution time series according to claim 1, characterized in that, In step S1, a normal distribution curve is constructed from the image data based on the principle of normal distribution. The splash intensity value at each moment is calculated based on the normal distribution curve, thus obtaining a continuous splash intensity dataset, including: The splashing time, splashing stage, and standard deviation in the image data are determined. The peak splashing time is used as the mean to generate a normal distribution curve. The splashing intensity value at each time is calculated based on the normal distribution curve. The splashing intensity value is normalized. Several splashing probability intensity curves are obtained based on the normalized splashing intensity value and the intensity weight of the splashing stage. Several splashing probability intensity curves are fused to obtain a continuous splashing intensity dataset. The splashing stages include small splashes, medium splashes, and large splashes.

3. The converter splash prediction method based on normal labels and weighted convolution time series according to claim 2, characterized in that, The formula for calculating the splash probability intensity curve is shown in equation (1) below: （1） In the formula, Represents any moment on the timeline The splash intensity value, Indicates the first The standard deviation of a splash event Indicates the first The peak splash time of a splash event Indicates the first The intensity weight of the splash phase to which each splash event belongs.

4. The converter splash prediction method based on normal labels and weighted convolution time series according to claim 1, characterized in that, The construction of the deep fusion regression model based on CNN-BiLSTM and attention mechanism in S3 includes: Static visual features are extracted from each frame of image data in the processed dataset using a pre-trained CNN model to obtain the feature vector of each frame. The feature vectors of all frames corresponding to each second are then concatenated to obtain the concatenated features. The concatenated features are input into the bidirectional long short-term memory network temporal feature learner to obtain the temporal dependencies and temporal context information of the features per second; The temporal features of all frames are weighted and aggregated according to attention weights to obtain the aggregated global feature vector. The aggregated global feature vector is then input into the regression head to obtain the splash intensity prediction value.

5. The converter splash prediction method based on normal labels and weighted convolution time series according to claim 1, characterized in that, The high-value weighted mean square error loss function in S3 includes: In the backpropagation process of the deep fusion regression model, weighted mean square error is used to assign differentiated weights to the samples in the processed dataset; The high-value weighted mean square error loss function is shown in equation (2) below: （2） In the formula, This represents the high-value weighted mean square error loss function. Indicates the number of samples in the batch. Indicates the dynamic weighting coefficient. Indicates the first The true label of each sample Indicates the first Predicted splash intensity values ​​for each sample.

6. A converter splash prediction device based on normal labels and weighted convolution time series, wherein the converter splash prediction device based on normal labels and weighted convolution time series is used to implement the converter splash prediction method based on normal labels and weighted convolution time series as described in any one of claims 1-5, characterized in that, The device includes: The data acquisition module is used to acquire flame video data of the converter steelmaking process, segment the flame video data into frames to obtain image data, construct a normal distribution curve for the image data based on the normal distribution principle, calculate the splash intensity value at each moment based on the normal distribution curve, and thus obtain a continuous splash intensity dataset; wherein, the continuous splash intensity dataset includes several frames of image data and the splash intensity label corresponding to each frame of image data; The data processing module is used to aggregate the splash intensity labels in the continuous splash intensity dataset to obtain a second-level splash intensity dataset. An asymmetric sampling strategy is used to sample the second-level splash intensity dataset to obtain the processed dataset. The training module is used to build a deep fusion regression model based on CNN-BiLSTM and attention mechanism. A high-value weighted mean square error loss function is designed. The deep fusion regression model is trained based on the processed dataset and the high-value weighted mean square error loss function to obtain the trained deep fusion regression model. The output module is used to acquire the flame video data to be predicted and to obtain the splash intensity prediction value based on the flame video data to be predicted and the trained deep fusion regression model.

7. The converter splash prediction device based on normal labels and weighted convolution time series according to claim 6, characterized in that, The image data is used to construct a normal distribution curve based on the principle of normal distribution. The splash intensity value at each moment is calculated based on the normal distribution curve, thus obtaining a continuous splash intensity dataset, including: The splashing time, splashing stage, and standard deviation in the image data are determined. The peak splashing time is used as the mean to generate a normal distribution curve. The splashing intensity value at each time is calculated based on the normal distribution curve. The splashing intensity value is normalized. Several splashing probability intensity curves are obtained based on the normalized splashing intensity value and the intensity weight of the splashing stage. Several splashing probability intensity curves are fused to obtain a continuous splashing intensity dataset. The splashing stages include small splashes, medium splashes, and large splashes.

8. The converter splash prediction device based on normal labels and weighted convolution time series according to claim 6, characterized in that, The construction of the deep fusion regression model based on CNN-BiLSTM and attention mechanism includes: Static visual features are extracted from each frame of image data in the processed dataset using a pre-trained CNN model to obtain the feature vector of each frame. The feature vectors of all frames corresponding to each second are then concatenated to obtain the concatenated features. The concatenated features are input into the bidirectional long short-term memory network temporal feature learner to obtain the temporal dependencies and temporal context information of the features per second; The temporal features of all frames are weighted and aggregated according to attention weights to obtain the aggregated global feature vector. The aggregated global feature vector is then input into the regression head to obtain the splash intensity prediction value.

9. A converter splash prediction device, characterized in that, The converter splash prediction device includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 5.

Images