Converter steelmaking oxygen supply prediction method based on knowledge and data fusion driving

Abstract

The invention discloses a converter steelmaking oxygen supply prediction method based on knowledge and data fusion driving. The converter steelmaking oxygen supply prediction method comprises the following steps: S1, constructing a method framework comprising a knowledge-based feature engineering and gating constraint module and a data-based dual-attention prediction module; s2, preprocessing and cutting the data; s3, performing feature generation and feature selection through the represented knowledge in a knowledge-based feature engineering and gating constraint module, and establishing a gating constraint mechanism; s4, constructing a multi-scale convolutional neural network and a time perception neural network in a data-based dual-attention prediction module, and training under obtained features and gating constraints; and S5, performing prediction by using the trained model, and performing weighted fusion on prediction results obtained by the multi-scale convolutional neural network and the time sensing neural network to obtain a final oxygen supply amount.

Description

Technical Field

[0001] The present invention relates to the field of prediction of oxygen supply for converter steelmaking and deep learning technology, and in particular to a method for predicting oxygen supply for converter steelmaking driven by knowledge and data fusion. Background Art

[0002] Converter steelmaking is a process that converts raw molten iron into high-quality molten steel. It is one of the most widely used steelmaking technologies in industry. In this process, high-pressure oxygen is blown into the molten iron to quickly remove impurities in the molten iron. The converter rotates at high temperature to ensure that the oxygen reacts evenly with the molten iron to produce gas and slag. The amount of oxygen supply directly affects the smelting effect and is crucial in the entire smelting process. If the oxygen supply is too much, it will not only lead to energy waste and increased costs, but also affect the chemical composition of the steel and lead to a decrease in product quality; while insufficient oxygen supply will lead to slow steelmaking reactions and affect production efficiency. At the same time, since impurities cannot be fully oxidized, the purity and performance of the steel will also be affected. Therefore, accurately predicting the oxygen supply has become a research hotspot.

[0003] Oxygen supply prediction methods are mainly divided into data-driven prediction methods and domain knowledge-driven prediction methods. Data-driven prediction methods solve the problem that it is difficult to establish an accurate mathematical model due to the complexity of the process. For example, the article [Qin Bo, Wu Qingchao, Zhang Juanjuan, et al. Research on oxygen consumption prediction for converter steelmaking based on PSO optimized SVM [J]. Measurement and Control Technology, 2014, 33 (12): 121-124.] uses a particle swarm algorithm to optimize the penalty coefficient, insensitive loss coefficient and Gaussian kernel width coefficient of the support vector machine, thereby improving the prediction accuracy of the converter oxygen consumption; in order to process large-scale, high-dimensional data, the article [Fu Jia. Research and application of converter oxygen supply model based on BP neural network Application Development [D]. Metallurgical Automation Research and Design Institute, 2014.] established a BP neural network converter oxygen supply model, and optimized the BP neural network model by adding momentum terms and using variable step length method to solve the problem of processing complex data; in order to solve the problem of slow convergence of BP neural network, the article [Zhao Hui, Yi Xiaomin, Wang Hongjun, et al. Research on oxygen consumption prediction model for converter steelmaking [J]. Computer Simulation, 2017, 34(01): 380-383+422.] used extreme learning machine to establish oxygen consumption prediction model, which improved the convergence speed of the model. The above methods are based on historical data to build and train models and make predictions, and all achieved good results. However, data-driven models only consider the relationship between input and output in the data, and lack the application of domain knowledge, resulting in not only many outliers in the prediction results, but also poor interpretability, which limits the application of these models in actual production. Therefore, how to realize the application of domain knowledge is very important. In order to realize the application of domain knowledge, the article [Mei Liuliu. Research on intelligent control method of converter steelmaking based on expert system [D]. Liaoning University of Science and Technology, 2023.] established the heat balance and material balance equations of the converter smelting process and obtained the oxygen consumption; the article [Wang Huixian, Xu Anjun, Gu Maoqiang, et al. Prediction model of converter oxygen supply based on segmented oxygen decarburization efficiency [J]. China Metallurgy, 2023, 33(07): 91-99.] proposed a prediction model of converter oxygen supply based on segmented oxygen decarburization efficiency. Through mechanism analysis, the inflection points of the three stages of decarburization in the converter and the oxygen decarburization efficiency at different stages were determined, and the oxygen supply of the furnace to be solved was predicted by case reasoning, realizing the application of knowledge; although the above methods have realized the application of domain knowledge to a certain extent, they have not yet been able to be fully integrated with the data-driven model, and cannot give full play to the advantages of the combination of data and knowledge, which limits the accuracy of model prediction.

[0004] In summary, most traditional oxygen supply prediction models rely solely on knowledge or data for modeling, and domain knowledge fails to be deeply integrated with data models, resulting in the inability to simultaneously bring out the strengths of both. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting the oxygen supply of converter steelmaking driven by knowledge and data fusion. The method uses knowledge extracted from the smelting process and expert experience to perform feature engineering and establish a gated constraint mechanism, and applies complex smelting mechanisms and expert experience knowledge to the prediction model as an auxiliary guide to the data model. At the same time, in order to fully consider the time series information between furnaces caused by factors such as slag during model prediction, a dual-focus prediction model is constructed, which considers both the regression relationship within a single furnace and the time series characteristics between furnaces, thereby achieving full utilization of data and achieving stable and accurate prediction of the oxygen supply of the converter.

[0006] The technical solution adopted by the present invention is:

[0007] A method for predicting oxygen supply for converter steelmaking based on knowledge and data fusion drive includes the following steps:

[0008] S1: Construct a method framework including knowledge-based feature engineering and gating constraint modules and data-based dual attention prediction modules;

[0009] S2: Preprocess the real data of converter steelmaking, including removing null values ​​and discrete values, to obtain cleaned data. Then, cut the data according to the time series to generate a data set that meets the needs of subsequent models;

[0010] S3: In the knowledge-based feature engineering and gated constraint module, the domain knowledge is represented as usable influencing factors or equations. Secondly, the transformed knowledge is combined with the original features for feature generation and feature selection, and a gated constraint mechanism is established based on the represented knowledge to obtain optimized features and gated constraints.

[0011] S4: A multi-scale convolutional neural network and a time-aware neural network are constructed in the data-based dual attention prediction module. The multi-scale convolutional neural network is used to capture the regression information of the current heat, and the time-aware neural network is used to capture the time series information between heats. Both models are optimized using features and constraint gates constructed through knowledge to train the models.

[0012] S5: Use the trained model for prediction, input the data into the multi-scale convolutional neural network and the time-aware neural network for prediction, and perform weighted fusion on the prediction results of the two models (pred1 and pred2) to obtain the final oxygen supply.

[0013] Furthermore, in the described step S3, the domain knowledge is represented as a usable influencing factor or equation form, and then the transformed knowledge is combined with the original features for feature generation and feature selection, and a gating constraint mechanism is established through the represented knowledge to obtain optimized features and gating constraints, including the following steps:

[0014] 1-1) In the knowledge representation, expert experience and metallurgical mechanism knowledge are converted into influencing factors. The specific influencing factors are determined as molten iron weight, temperature, carbon, silicon, manganese, phosphorus, sulfur content, scrap steel weight, raw material addition ratio, C sample carbon content, temperature, end point carbon content, temperature, converter, oxygen lance life, where C sample represents the sampling of a period of time before the end;

[0015] 1-2) In the knowledge representation, expert experience and metallurgical mechanism knowledge are converted into equations. In order to more accurately characterize the oxygen consumption in the smelting process, the calculation of oxygen supply is decomposed into four parts, and the oxygen consumption of each part is represented by an independent equation. The specific equation is:

[0016] Oxygen Elements =(iron_weight×iron_elements-steel_weight×target_elements)×oxides atomic_mass

[0017] Oxygen Slag =slag_weight×slag_fe×oxides atomic_mass

[0018] Oxygen Lining =lining_weight×lining_c×oxides atomic_mass

[0019] Oxygen Smoke =smoke_weight×smoke_fe×oxides atomic_mass

[0020] Among them, Oxygen Elements represents the part of oxygen consumed by the oxidation of elements in the raw materials, iron_weight represents the weight of molten iron, iron_elements represents the element concentration in molten iron, including carbon, silicon, manganese, phosphorus and sulfur, steel_weight represents the weight of molten steel, target_elements represents the smelting target concentration of elements, oxides atomic_mass Represents the atomic mass of the element's oxide, Oxygen Slag , Oxygen Lining , Oxygen SmokeThey represent the oxygen consumption of slag, lining and smoke respectively. slag_weight, lining_weight and smoke_weight represent the weight of slag, lining and smoke respectively. slag_fe, lining_c and smoke_fe represent the oxidation percentage of iron in slag, carbon in lining and iron in smoke respectively. The specific values ​​can be obtained from the data.

[0021] 1-3) In the process of feature generation, the influencing factors that are not directly provided in the original data, such as converter life, oxygen lance life and weight ratio of various added materials, are obtained by calculation. The specific calculation is as follows:

[0022]

[0023]

[0024] Among them, Time represents the time difference between the current furnace and the starting furnace, D represents the start time of the current furnace smelting, D0 represents the time after the last maintenance of the converter and oxygen gun, and per_time represents the average smelting time per furnace. Steel_VS_Iron_ratio represents the ratio of the amount of scrap steel added to the amount of molten iron added, steel_scrap_weight represents the weight of the scrap steel added, and iron_weight represents the weight of the molten iron added. Through calculation, new key features are obtained;

[0025] 1-4) In the feature selection process, after generating new features, the features most relevant to the prediction of oxygen supply are selected according to the key influencing factors determined previously as input variables of the model;

[0026] 1-5) At the same time, dynamic gating constraints are constructed. In the entire smelting process, the most basic requirement is to oxidize each element in the raw material to the target concentration, which constitutes the minimum threshold. On the other hand, considering that the slag, lining and smoke loss of each furnace are uncertain, it is assumed that all three parts have oxidation consumption, so that the maximum threshold of oxygen consumption can be obtained. The specific calculation is as follows:

[0027] Min=Oxygen Elements

[0028] Max=Oxygen Elements +Oxygen Slag +Oxygen Lining +Oxygen Smoke

[0029] Among them, Min is the minimum threshold and MAx is the maximum threshold. The conditions of these four parts in each furnace are different, so the calculated thresholds are dynamically changing;

[0030] 1-6) Based on these dynamic maximum and minimum values, we constructed a dynamic control gate. When the predicted value during model training exceeds the calculated maximum or is lower than the minimum, the dynamic control gate will be activated and a distance-based penalty term will be imposed, which will be integrated into the loss (LOSS). The specific expression is as follows:

[0031]

[0032]

[0033] Among them, L gate represents the penalty term, α is the coefficient that controls the penalty intensity, δ is a positive constant used to scale the distance, V min is the minimum threshold, V max is the maximum threshold, when the predicted value Out of dynamic range [V min , V max ], the loss term increases by the extent of the excess range.

[0034] Furthermore, in the described step S4, a multi-scale convolutional neural network and a time-aware neural network are constructed in the data-based dual attention prediction module, and both models are optimized using features and constraint gates constructed through knowledge, and the models are trained, including the following steps:

[0035] 2-1) Construct the input layer and multi-scale convolutional feature extraction layer in the multi-scale convolutional neural network. The input layer receives the previously processed data, and the multi-scale convolutional feature extraction layer uses multiple convolution kernels of different scales to perform parallel feature extraction, and then fuses the multi-scale features;

[0036] 2-2) Construct an adaptive attention layer (Self-Attention) in a multi-scale convolutional neural network. After fusing the multi-scale features, add the adaptive attention layer. The attention mechanism can dynamically adjust the weight according to the contribution of each feature, allowing the model to focus on those features that are critical to the prediction results while suppressing irrelevant noise information. The specific calculation is as follows:

[0037]

[0038] Among them, Attention is the result of attention calculation, Q, K, and V are query matrices, key matrices, and value matrices respectively, which are obtained by linear transformation of input features, T represents matrix transpose, and d k is the dimension of K, which is 1 here. The softmax activation function makes the values ​​between 0 and 1;

[0039] 2-3) Construct a fully connected layer in a multi-scale convolutional neural network, and follow the adaptive attention layer with a fully connected layer for the final regression output;

[0040] 2-4) At the same time, a time-aware neural network based on the Transformer architecture is constructed. The architecture is a neural network composed of an encoder-decoder structure, and the input layer is constructed. The input layer is used to receive time series data;

[0041] 2-5) Construct two embedding layers (Input Embedding and Output Embedding) after the input layer to embed the position index of the input sequence (Inputs) and the target sequence (Outputs (shifted right)) into high-dimensional vectors. These embeddings are trainable parameters that allow the model to adapt to different sequence lengths during the learning process;

[0042] 2-6) Then construct the positional encoding, which is connected after the embedding layer. Through the positional encoding module, the position information is introduced into the input sequence and the target sequence. The specific calculation is as follows:

[0043]

[0044]

[0045] Among them, PE represents the obtained position information, pos represents the position, i represents the feature dimension, sin() and cos() represent the sine and cosine functions, d model Represents the dimension of the model. The position encoding matrix is ​​finally obtained through the formula calculation, which is used to add position encoding to the input in the forward propagation;

[0046] 2-7) Build an encoder, which is composed of multiple encoder layers. Each encoder layer includes a multi-head self-attention mechanism (Multi-HeadAttention) and a feedforward neural network (Feed Forward). Multiple attention heads are used to learn different patterns in the input sequence. For each head, the input Q, K, and V matrices are linearly transformed to generate the query, key, and value corresponding to each head. The specific calculation is as follows:

[0047]

[0048] Among them, X represents the embedded representation of the input sequence, Q, K, V are the query matrix, key matrix, and value matrix respectively, and i represents the i-th attention head. Represents the projection matrix of each head. Connect the outputs of all heads together, and the specific calculation is as follows:

[0049] MultiHead=Concat(head1,...,head i )×W O

[0050] Among them, MultiHead represents the final result of multi-head attention, W O Represents the final linear transformation matrix, Concat represents connection, head i Represents the attention calculation results of each head;

[0051] 2-8) Construct a feedforward neural network, which is applied after each attention head. It consists of two layers of linear transformation. After each sub-layer (multi-head self-attention mechanism and feedforward network), residual connections and layer normalization (Add&Norm) are added. The specific calculation is as follows:

[0052]

[0053] Among them, LayerNorm is the layer normalization operation, x represents the input, Sublayer(x) is the output of the sublayer, +x represents the residual connection, μ and σ 2 is the mean and variance of Sublayer(x)+x, ε is a small constant, γ and β are learnable parameters;

[0054] 2-9) Construct a decoder, which is composed of multiple decoder layers stacked together. The decoder layer is similar to the encoder layer, but the decoder layer will refer to the previous target sequence (self-attention) and the output of the encoder layer (cross-attention) at the same time. In the calculation of cross-attention, K and V are replaced by the output of the encoder. The rest is the same as the encoder layer. The decoder layer also has a feedforward neural network layer, which is also a two-layer linear transformation. Similarly, after each sub-layer (multi-head self-attention mechanism and feedforward network), residual connections and layer normalization are added;

[0055] 2-10) Finally, the encoder and decoder are connected to build a fully connected layer (Linear) to generate the final output (output). The overall architecture is encoder-decoder;

[0056] 2-11) Use the data to train the established multi-scale convolutional neural network and time-aware neural network. The training process uses knowledge-based features and gating constraints.

[0057] The beneficial effects of adopting the above technical solution are:

[0058] The purpose of the present invention is to provide a method for predicting oxygen supply for converter steelmaking driven by knowledge and data fusion, the method comprising a knowledge-based feature engineering and gating constraint module and a data-based dual-attention prediction module. The knowledge-based feature engineering and gating constraint module first characterizes the domain knowledge, and uses the characterized knowledge information to implement feature engineering, while constructing a gating constraint mechanism. The data-based dual-attention prediction module captures the time series information and non-time series information in the data respectively, and is trained under the guidance of the features obtained from the knowledge and the gating constraint mechanism, and finally the prediction results of the two parts are fused, which not only improves the prediction accuracy of the model, but also, with the assistance of knowledge, significantly reduces the occurrence of outliers and enhances the stability of the prediction.

[0059] The proposed converter steelmaking oxygen supply prediction method is applied to the actual production data of a domestic steel plant. Through experimental analysis, the average absolute error and root mean square error of oxygen supply in different steel grades produced by two different converters are reduced by at least 4% and at most 15% compared with the mainstream method. The optimal hit rate of relative error at an accuracy of 5% can reach 87.16%, which is better than the mainstream method. The oxygen supply prediction ability of the converter steelmaking oxygen supply prediction method driven by knowledge and data fusion is verified, and the probability of abnormal values ​​is reduced while improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a framework diagram of the oxygen supply prediction method for converter steelmaking driven by knowledge and data fusion;

[0061] Figure 2 Module graph for knowledge-based feature engineering and constraint optimization;

[0062] Figure 3 It is a multi-scale convolutional neural network structure diagram;

[0063] Figure 4 This is a diagram of the time-aware neural network structure;

[0064] Figure 5 This is a (partial) diagram of the comparison experiment prediction results;

[0065] Figure 6 This is the cumulative absolute error diagram of the comparative experiment;

[0066] Figure 7 This is a comparison chart of the proportions within each error range of furnace A;

[0067] Figure 8 This is a comparison chart of the proportions within each error range of furnace B;

[0068] Fig. 9 This is a comparison chart of the KACT-NonSeqTran ablation experiment errors;

[0069] Fig.10 This is the comparison diagram of KACT-NonATTCnn ablation experiment error;

[0070] Fig.11 This is a comparison chart of KACT-NonKnow ablation experiment errors. DETAILED DESCRIPTION

[0071] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0072] The present invention takes converter steelmaking oxygen supply prediction as the background, deep learning as the carrier, knowledge-based feature engineering and gated constraint module and data-based dual attention prediction module as the main algorithm framework. Figure 1 As shown, the following steps are included:

[0073] S1: Construct a method framework including knowledge-based feature engineering and gating constraint modules and data-based dual attention prediction modules;

[0074] S2: Preprocess the real data of converter steelmaking, including removing null values ​​and discrete values, to obtain cleaned data. Then, cut the data according to the time series to generate a data set that meets the needs of subsequent models;

[0075] S3: In the knowledge-based feature engineering and gated constraint module, the domain knowledge is represented as usable influencing factors or equations. Secondly, the transformed knowledge is combined with the original features for feature generation and feature selection, and a gated constraint mechanism is established through the represented knowledge to obtain optimized features and gated constraints, including the following steps:

[0076] 1-1) In the knowledge representation, expert experience and metallurgical mechanism knowledge are converted into influencing factors. The specific influencing factors are determined as molten iron weight, temperature, carbon, silicon, manganese, phosphorus, sulfur content, scrap steel weight, raw material addition ratio, C sample carbon content, temperature, end point carbon content, temperature, converter, oxygen lance life, where C sample represents the sampling of a period of time before the end;

[0077] 1-2) In the knowledge representation, expert experience and metallurgical mechanism knowledge are converted into equations. In order to more accurately characterize the oxygen consumption in the smelting process, the calculation of oxygen supply is decomposed into four parts, and the oxygen consumption of each part is represented by an independent equation. The specific equation is:

[0078] Oxygen Elements =(iron_weight×iron_elements-steel_weight×target_elements)×oxides atomic_mass

[0079] Oxygen Slag =slag_weight×slag_fe×oxides atomic_mass

[0080] Oxygen Lining =lining_weight×lining_c×oxides atomic_mass

[0081] Oxygen Smoke =smoke_weight×smoke_fe×oxides atomic_mass

[0082] Among them, Oxygen Elements represents the part of oxygen consumed by the oxidation of elements in the raw materials, iron_weight represents the weight of molten iron, iron_elements represents the element concentration in molten iron, including carbon, silicon, manganese, phosphorus and sulfur, steel_weight represents the weight of molten steel, target_elements represents the smelting target concentration of elements, oxides atomic_mass Represents the atomic mass of the element's oxide, Oxygen Slag , Oxygen Lining , Oxygen Smoke They represent the oxygen consumption of slag, lining and smoke respectively. slag_weight, lining_weight and smoke_weight represent the weight of slag, lining and smoke respectively. slag_fe, lining_c and smoke_fe represent the oxidation percentage of iron in slag, carbon in lining and iron in smoke respectively. The specific values ​​can be obtained from the data.

[0083] 1-3) In the process of feature generation, the influencing factors that are not directly provided in the original data, such as converter life, oxygen lance life and weight ratio of various added materials, are obtained by calculation. The specific calculation is as follows:

[0084]

[0085]

[0086] Among them, Time represents the time difference between the current furnace and the starting furnace, D represents the start time of the current furnace smelting, D0 represents the time after the last maintenance of the converter and oxygen gun, and per_time represents the average smelting time per furnace. Steel_VS_Iron_ratio represents the ratio of the amount of scrap steel added to the amount of molten iron added, steel_scrap_weight represents the weight of the scrap steel added, and iron_weight represents the weight of the molten iron added. Through calculation, new key features are obtained;

[0087] 1-4) In the feature selection process, after generating new features, the features most relevant to the prediction of oxygen supply are selected according to the key influencing factors determined previously as input variables of the model;

[0088] 1-5) At the same time, dynamic gating constraints are constructed. In the entire smelting process, the most basic requirement is to oxidize each element in the raw material to the target concentration, which constitutes the minimum threshold. On the other hand, considering that the slag, lining and smoke loss of each furnace are uncertain, it is assumed that all three parts have oxidation consumption, so that the maximum threshold of oxygen consumption can be obtained. The specific calculation is as follows:

[0089] Min=Oxygen Elements

[0090] Max=Oxygen Elements +Oxygen Slag +Oxygen Lining +Oxygen Smoke

[0091] Among them, Min is the minimum threshold and Max is the maximum threshold. The conditions of these four parts in each furnace are different, so the calculated thresholds are dynamically changing;

[0092] 1-6) Based on these dynamic maximum and minimum values, we constructed a dynamic control gate. When the predicted value during model training exceeds the calculated maximum or is lower than the minimum, the dynamic control gate will be activated and a distance-based penalty term will be imposed, which will be integrated into the loss (LOSS). The specific expression is as follows:

[0093]

[0094]

[0095] Among them, L gate represents the penalty term, α is the coefficient that controls the penalty intensity, δ is a positive constant used to scale the distance, V min is the minimum threshold, V max is the maximum threshold, when the predicted value Out of dynamic range [V min , V max ], the loss term increases by the extent of the excess range.

[0096] S4: A multi-scale convolutional neural network and a time-aware neural network are constructed in the data-based dual attention prediction module. The multi-scale convolutional neural network is used to capture the regression information of the current heat, and the time-aware neural network is used to capture the time series information between heats. Both models are optimized using features and constraint gates constructed through knowledge. The training of the models includes the following steps:

[0097] 2-1) Construct the input layer and multi-scale convolutional feature extraction layer in the multi-scale convolutional neural network. The input layer receives the previously processed data, and the multi-scale convolutional feature extraction layer uses multiple convolution kernels of different scales to perform parallel feature extraction, and then fuses the multi-scale features;

[0098] 2-2) Construct an adaptive attention layer (Self-Attention) in a multi-scale convolutional neural network. After fusing the multi-scale features, add the adaptive attention layer. The attention mechanism can dynamically adjust the weight according to the contribution of each feature, allowing the model to focus on those features that are critical to the prediction results while suppressing irrelevant noise information. The specific calculation is as follows:

[0099]

[0100] Among them, Attention is the result of attention calculation, Q, K, and V are query matrices, key matrices, and value matrices respectively, which are obtained by linear transformation of input features, T represents matrix transpose, and d k is the dimension of K, which is 1 here. The softmax activation function makes the values ​​between 0 and 1;

[0101] 2-3) Construct a fully connected layer in a multi-scale convolutional neural network, and follow the adaptive attention layer with a fully connected layer for the final regression output;

[0102] 2-4) At the same time, a time-aware neural network based on the Transformer architecture is constructed. The architecture is a neural network composed of an encoder-decoder structure, and the input layer is constructed. The input layer is used to receive time series data;

[0103] 2-5) Construct two embedding layers (Input Embedding and Output Embedding) after the input layer to embed the position index of the input sequence (Inputs) and the target sequence (Outputs (shifted right)) into high-dimensional vectors. These embeddings are trainable parameters that allow the model to adapt to different sequence lengths during the learning process;

[0104] 2-6) Then construct the positional encoding, which is connected after the embedding layer. Through the positional encoding module, the position information is introduced into the input sequence and the target sequence. The specific calculation is as follows:

[0105]

[0106]

[0107] Among them, PE represents the obtained position information, pos represents the position, i represents the feature dimension, sin() and cos() represent the sine and cosine functions, d model Represents the dimension of the model. The position encoding matrix is ​​finally obtained through the formula calculation, which is used to add position encoding to the input in the forward propagation;

[0108] 2-7) Build an encoder, which is composed of multiple encoder layers. Each encoder layer includes a multi-head self-attention mechanism (Multi-HeadAttention) and a feedforward neural network (Feed Forward). Multiple attention heads are used to learn different patterns in the input sequence. For each head, the input Q, K, and V matrices are linearly transformed to generate the query, key, and value corresponding to each head. The specific calculation is as follows:

[0109]

[0110] Among them, X represents the embedded representation of the input sequence, Q, K, V are the query matrix, key matrix, and value matrix respectively, and i represents the i-th attention head. Represents the projection matrix of each head. Connect the outputs of all heads together, and the specific calculation is as follows:

[0111] MultiHead=Concat(head1,...,head i )×W O

[0112] Among them, MultiHead represents the final result of multi-head attention, W O Represents the final linear transformation matrix, Concat represents connection, head i Represents the attention calculation results of each head;

[0113] 2-8) Construct a feedforward neural network, which is applied after each attention head. It consists of two layers of linear transformation. After each sub-layer (multi-head self-attention mechanism and feedforward network), residual connections and layer normalization (Add&Norm) are added. The specific calculation is as follows:

[0114]

[0115] Among them, LayerNorm is the layer normalization operation, x represents the input, Sublayer(x) is the output of the sublayer, +x represents the residual connection, μ and σ 2 is the mean and variance of Sublayer(x)+x, ε is a small constant, γ and β are learnable parameters;

[0116] 2-9) Construct a decoder, which is composed of multiple decoder layers stacked together. The decoder layer is similar to the encoder layer, but the decoder layer will refer to the previous target sequence (self-attention) and the output of the encoder layer (cross-attention) at the same time. In the calculation of cross-attention, K and V are replaced by the output of the encoder. The rest is the same as the encoder layer. The decoder layer also has a feedforward neural network layer, which is also a two-layer linear transformation. Similarly, after each sub-layer (multi-head self-attention mechanism and feedforward network), residual connections and layer normalization are added;

[0117] 2-10) Finally, the encoder and decoder are connected to build a fully connected layer (Linear) to generate the final output (output). The overall architecture is encoder-decoder;

[0118] 2-11) Use the data to train the established multi-scale convolutional neural network and time-aware neural network. The training process uses knowledge-based features and gating constraints.

[0119] S5: Use the trained model for prediction, input the data into the multi-scale convolutional neural network and the time-aware neural network for prediction, and perform weighted fusion on the prediction results of the two models (pred1 and pred2) to obtain the final oxygen supply.

[0120] Based on the above steps, the present invention effectively solves the problem that most oxygen supply prediction models rely solely on knowledge or data for modeling, and domain knowledge fails to be deeply integrated with the data model, resulting in the inability to simultaneously bring into play the advantages of both. A method for predicting the oxygen supply of converter steelmaking driven by knowledge and data fusion is proposed. This method uses knowledge extracted from the smelting process and expert experience to perform feature engineering and establish a gated constraint mechanism, and applies complex smelting mechanisms and expert experience knowledge to the prediction model as an auxiliary guide to the data model. At the same time, in order to fully consider the time series information between furnaces caused by factors such as slag during model prediction, a dual-focus prediction model is constructed, which not only considers the regression relationship within a single furnace, but also takes into account the time series characteristics between furnaces, so as to fully utilize the data, thereby achieving a stable and accurate prediction of the converter oxygen supply.

[0121] Experimental verification of the oxygen supply prediction method for converter steelmaking driven by knowledge and data fusion in the present invention:

[0122] 1. Test environment

[0123] Windows 11 system, Intel Core i5-1035G7 CPU@1.20GHz and NVIDIA GeForce RTX2080Ti GPU, the development environment is Python 3.9.10, the deep learning framework is PyTorch 2.1.0, and the machine learning method uses the API provided by Scikit-learn1.3.0.

[0124] 2. Test verification

[0125] Experimental results and analysis on actual production data of converter steelmaking in a domestic steel plant

[0126] (1) Dataset description

[0127] The experiment uses the actual production data of two converters in a domestic steel plant, and selects the main steel grades in each converter as the research object, namely SPHC steel grade in furnace A and HRB400E steel grade in furnace B. The production data includes molten iron information, scrap steel information, smelting process information, smelting end point information, oxygen supply, etc. The data selection time is from January 1, 2023 to May 1, 2024. There are 7071 data in furnace A and 15380 data in furnace B. Some of the data are shown in Table 1.

[0128] Table 1 Converter production data variable table

[0129]

[0130]

[0131] (2) Implementation details and evaluation indicators

[0132] The multi-scale convolutional neural network used in this section has three convolution kernels: 1, 3, and 5. Each convolution scale contains three convolution layers. Each round of training uses the adaptive moment estimation optimizer (Adam), with a learning rate of 0.0005, 300 training rounds, and a batch size of 2. The time-aware neural network used has an encoder with 4 encoder layers, a decoder with 6 decoder layers, 8 multi-head attention heads, and d model The model dimension is set to 512, and the adaptive moment estimation optimizer is used in each round of training, with a learning rate of 0.0001, 200 training rounds, and a batch size of 2. In the final weighted fusion, the weight coefficient of the two models is set to 7:3 for the multi-scale convolutional neural network and the time-aware neural network.

[0133] The data set is shuffled and divided into training set and test set in a ratio of 8:2. For the time-aware network, the data is divided into 4 pieces as time series data. The input variable is determined according to the knowledge-based feature engineering module, and the output variable is the oxygen supply.

[0134] In order to evaluate the prediction performance of the model, the mean absolute error (MAE), root mean square error (RMSE), mean absolute relative error (MAPE), accuracy within the error range (Acc ε ) as the evaluation criteria, defined as follows:

[0135]

[0136]

[0137]

[0138]

[0139] Among them, i represents the i-th sample, y i and Represent the true value and predicted value of the target variable, n represents the number of samples in the test set, 1(·) is the indicator function. If the expression in the brackets is true, the value of the indicator function is 1, otherwise it is 0. ε is the absolute value of the given error range, expressed as a decimal. MAE describes the average absolute difference between the predicted value and the actual value; RMSE describes the average error between the predicted value and the true value, and is more sensitive to outliers; MAPE describes the average percentage difference between the predicted value and the actual value. The smaller these three indicators are, the higher the prediction accuracy. Acc ε It is an intuitive indicator that reflects the prediction performance. The larger the index is, the better the prediction performance is. Here Acc ε The error ranges are [-3%, 3%] and [-5%, 5%] respectively.

[0140] (3) Analysis of comparative experimental results

[0141] In order to verify the effectiveness of integrating time series features and non-time series features into modeling and introducing mechanism knowledge to assist model training and prediction, a prediction method for oxygen supply in converter steelmaking based on knowledge and data fusion driven is proposed in this paper as KACT. A comparative experiment is conducted on the model. The experimental results are shown in Tables 2 and 3. A brief description of each comparative model is as follows:

[0142] Improved Deep Belief Network (GBRBM-DBN): This model is based on deep learning theory and introduces Gaussian Bernoulli Restricted Boltzmann Machine into the deep belief network. Hybrid Model (HyGPR): This model combines the two methods of multivariate linear regression and Gaussian process regression. Integrated Learning Model (Stacking): This method first uses the random forest and extreme gradient boosting models to independently make the first prediction and generate their own prediction results. Then, these results are used as the input of the second-layer model. The linear regression model is selected to integrate the predictions of the first-layer model to make the final oxygen supply prediction. Bayesian Optimized Gradient Boosting Machine (BOA-LGBM): This method is a data-driven model that combines the Bayesian optimization algorithm and the lightweight gradient boosting machine. The Bayesian optimization algorithm is used to automatically optimize the parameters of the gradient boosting machine model.

[0143] Table 2A furnace comparison test results

[0144]

[0145]

[0146]

[0147] Table 3 B furnace comparison test results

[0148] According to the experimental results, the performance of the KACT model in the A furnace and B furnace data sets is the best compared with other models. For the GBRBM-DBN model, KACT achieved a reduction of 7.22% (A furnace) and 15.76% (B furnace) in MAE, 8.20% (A furnace) and 16.74% (B furnace) in RMSE, and 0.19 (A furnace) and 0.44 (B furnace) in MAPE, and both accuracy rates were improved. For the HyGPR model, KACT has made great progress in all five evaluation indicators. For the Stracing and BOA-LGBM models, there is a significant reduction in three error indicators, but no significant improvement in two accuracy indicators. The significant reduction in these error indicators is mainly due to the integration of the knowledge guidance part into the model, which reduces the sensitivity of the model to specific outliers, thereby reducing the error in the prediction. The improvement in accuracy can be attributed to the dual attention module proposed in this paper, which can capture more key information, thereby improving the prediction accuracy. Overall, compared with classic prediction methods and mainstream methods, KACT has made certain progress, indicating that KACT can complete the prediction task well.

[0149] Furthermore, in order to explore the contribution of each module in the KACT model to the performance, an ablation experiment was conducted, that is, each module was gradually removed from the complete model to observe the changes in model performance to prove the effectiveness of each module of KACT. The experiment was also conducted on the data of furnace A and furnace B, and the experimental results are shown in Tables 4 and 5. A brief description of each model:

[0150] KACT-NonSeqTran represents the module that removes the prediction module using time series information on the full model.

[0151] KACT-NonATTCnn represents the complete model with the module that uses the static regression information in the furnace for prediction removed.

[0152] KACT-NonKnow represents a module that removes the knowledge part on the full model and only uses data selection features and the original loss function.

[0153] KACT represents the complete model of this method.

[0154] Table 4 A furnace ablation test results

[0155]

[0156] Table 5 B furnace ablation test results

[0157]

[0158] According to Tables 4 and 5, for KACT-NonSeqTran, after removing the time-aware neural network model that captures time series information, the effects of various indicators have decreased, with an increase of 7.76% (furnace A) and 11.57% (furnace B) in MAE, and an increase of 5.65% (furnace A) and 9.18% (furnace B) in RMSE. The performance on MAPE also increased, and at the same time, the accuracy rates of various indicators decreased. This shows that the time-aware neural network model that captures time series information is significantly effective. For KACT-NonATTCnn, after removing the multi-scale convolutional neural network model that captures the regression information in the furnace, in terms of error, MAE increased by 33.01% (furnace A) and 47.22% (furnace B), RMSE increased by 29.31% (furnace A) and 44.09% (furnace B), and MAPE increased by 0.97 (furnace A) and 1.25 (furnace B). There is a significant decline in accuracy. The accuracy within the error range of [-3%, 3%] is reduced by 15.47 (furnace A) and 20.82 (furnace B), and the accuracy within the error range of [-5%, 5%] is reduced by 13.27 (furnace A) and 16.18 (furnace B). This shows that the multi-scale convolutional neural network model that captures the regression information in the furnace is significantly effective. For KACT-NonKnow, after removing the knowledge module, the effect of furnace B is significantly reduced in terms of error, MAE increases by 20.77%, RMSE increases by 18.68%, and MAPE increases by 0.53. In terms of accuracy, the effects of both data sets have declined significantly. This shows that the knowledge-based guidance module is significantly effective.

[0159] In summary, the prediction results of the KACT prediction method proposed in this paper on different converters are better than other prediction models. KACT applies complex smelting mechanisms and expert experience knowledge to the prediction model as an auxiliary guide to the data model by performing feature engineering and establishing a gated constraint mechanism based on knowledge extracted from the smelting process and expert experience. At the same time, a dual-focus prediction model is constructed, which not only considers the regression relationship within a single furnace, but also takes into account the time series characteristics between furnaces, making full use of the data, thereby achieving a stable and accurate prediction of the converter oxygen supply.

[0160] 3. Conclusion

[0161] In order to solve the problem that traditional prediction methods cannot fully capture the complexity of the steelmaking process by relying solely on knowledge or data, this paper proposes a prediction method for oxygen supply in converter steelmaking driven by knowledge and data fusion. This method consists of two parts. First, the knowledge of smelting mechanism and expert experience are systematically characterized, and these characterization results are applied to the construction of feature engineering and gate constraint mechanism; second, two models are established to capture the time series information and non-time series information in the data, and they are trained under knowledge guidance, and the prediction results of the two models are weighted fused; the fusion drive of the two parts finally realizes a more comprehensive use of data with the assistance of knowledge.

[0162] The above examples of the present invention are described in detail, but the contents are only preferred embodiments of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for predicting oxygen supply for converter steelmaking based on knowledge and data fusion, characterized in that: The steps include: S1: Construct a method framework including knowledge-based feature engineering and gating constraint modules and data-based dual attention prediction modules; S2: Preprocess the real data of converter steelmaking, including removing null values ​​and outliers, to obtain cleaned data. Then, cut the data according to time series to generate a data set that meets the requirements of subsequent models. S3: In the knowledge-based feature engineering and gated constraint module, the domain knowledge is represented as usable influencing factors or equations. Secondly, the transformed knowledge is combined with the original features for feature generation and feature selection, and a gated constraint mechanism is established based on the represented knowledge to obtain optimized features and gated constraints. S4: A multi-scale convolutional neural network and a time-aware neural network are constructed in the data-based dual attention prediction module. The multi-scale convolutional neural network is used to capture the regression information of the current heat, and the time-aware neural network is used to capture the time series information between heats. Both models are optimized using features and constraint gates constructed through knowledge to train the models. S5: Use the trained model to make predictions, input the data into the multi-scale convolutional neural network and the time-aware neural network for prediction, and perform weighted fusion on the prediction results of the two models to obtain the final oxygen supply.

2. The method for predicting oxygen supply for converter steelmaking based on knowledge and data fusion drive according to claim 1, characterized in that: In the described step S3, the domain knowledge is represented as a usable influencing factor or equation form. Secondly, the transformed knowledge is combined with the original features for feature generation and feature selection, and a gating constraint mechanism is established through the represented knowledge to obtain optimized features and gating constraints, including the following steps: 1-1) In the knowledge representation, expert experience and metallurgical mechanism knowledge are converted into influencing factors. The specific influencing factors are determined as molten iron weight, temperature, carbon, silicon, manganese, phosphorus, sulfur content, scrap steel weight, raw material addition ratio, C sample carbon content, temperature, endpoint carbon content, temperature, converter, oxygen lance life, where C sample represents sampling a period of time before the end; 1-2) In the knowledge representation, expert experience and metallurgical mechanism knowledge are converted into equations. In order to more accurately characterize the oxygen consumption in the smelting process, the calculation of oxygen supply is decomposed into four parts. The oxygen consumption of each part is represented by an independent equation. The specific equation is: Oxygen Elements =(iron_weight×iron_elements-steel_weight×target_elements×oxides atomic_mass Oxygen Slag =slag_weight×slag_fe×oxides atomic_mass Oxygen Lining =lining_weight×lining_c×oxides atomic_mass Oxygen Smoke =smoke_weight×smoke_fe×oxides atomic_mass Among them, Oxygen Elements represents the part of oxygen consumed by the oxidation of elements in the raw materials, iron_weight represents the weight of molten iron, iron_elements represents the element concentration in molten iron, including carbon, silicon, manganese, phosphorus and sulfur, steel_weight represents the weight of molten steel, target_elements represents the smelting target concentration of elements, oxides atomic_mass Represents the atomic mass of the element's oxide, Oxygen Slag , Oxygen Lining , Oxygen Smoke They represent the oxygen consumption of slag, lining and smoke respectively. slag_weight, lining_weight and smoke_weight represent the weight of slag, lining and smoke respectively. slag_fe, lining_c and smoke_fe represent the oxidation percentage of iron in slag, carbon in lining and iron in smoke respectively. The specific values ​​can be obtained from the data. 1-3) In the process of feature generation, the influencing factors that are not directly provided in the original data, such as converter life, oxygen lance life and weight ratio of various added materials, are obtained by calculation. The specific calculation is as follows: Among them, Time represents the time difference between the current heat and the starting heat, D represents the start time of the current heat, D0 represents the time after the last maintenance of the converter and oxygen lance, per_time represents the average smelting time of each heat, Steel_VS_Iron_ratio represents the ratio of the amount of scrap steel added to the amount of molten iron added, steel_scrap_weight represents the weight of the scrap steel added, and iron_weight represents the weight of the molten iron added. Through calculation, new key features are obtained; 1-4) In the feature selection process, after generating new features, the features most relevant to the prediction of oxygen supply are selected according to the key influencing factors determined previously as input variables of the model; 1-5) At the same time, a dynamic gating constraint is constructed. In the entire smelting process, the most basic requirement is to oxidize each element in the raw material to the target concentration, which constitutes the minimum threshold. On the other hand, considering that the slag, lining and smoke loss of each furnace are uncertain, it is assumed that all three parts have oxidation consumption, and the maximum threshold of oxygen consumption is obtained. The specific calculation is as follows: Min=Oxygen Elements Max=Oxygen Elements +Oxygen Slag +Oxygen Lining +Oxygen Smoke Among them, Min is the minimum threshold, Max is the maximum threshold, and the conditions of these four parts in each furnace are different, so the calculated thresholds are dynamically changing; 1-6) Based on these dynamic maximum and minimum values, we constructed a dynamic control gate. When the predicted value during model training exceeds the calculated maximum value or is lower than the minimum value, the dynamic control gate will be activated and a distance-based penalty term will be imposed. This penalty term will be integrated into the loss, as shown below: like like Among them, L gate represents the penalty term, α is the coefficient that controls the penalty intensity, δ is a positive constant used to scale the distance, V min is the minimum threshold, V max is the maximum threshold, when the predicted value Out of dynamic range [V min , V max ], the loss term increases by the extent of the excess range.

3. The method for predicting oxygen supply for converter steelmaking based on knowledge and data fusion drive according to claim 1, characterized in that: In the described step S4, a multi-scale convolutional neural network and a time-aware neural network are constructed in the data-based dual attention prediction module. Both models are optimized using features and constraint gates constructed through knowledge, and the models are trained, including the following steps: 2-1) Construct the input layer and multi-scale convolutional feature extraction layer in the multi-scale convolutional neural network. The input layer receives the previously processed data. The multi-scale convolutional feature extraction layer uses multiple convolution kernels of different scales to perform parallel feature extraction. Subsequently, the multi-scale features are fused. 2-2) Construct an adaptive attention layer in a multi-scale convolutional neural network. After fusing the multi-scale features, add the adaptive attention layer. The attention mechanism can dynamically adjust the weight according to the contribution of each feature, allowing the model to focus on those features that are critical to the prediction results while suppressing irrelevant noise information. The specific calculation is as follows: Among them, Attention is the result of attention calculation, Q, K, and V are query matrices, key matrices, and value matrices respectively, which are obtained by linear transformation of input features, T represents matrix transpose, and d k is the dimension of K, which is 1 here. The softmax activation function makes the values ​​between 0 and 1; 2-3) Construct a fully connected layer in a multi-scale convolutional neural network, and follow the adaptive attention layer with a fully connected layer for the final regression output; 2-4) At the same time, a time-aware neural network based on the Transformer architecture is constructed. The architecture is a neural network composed of an encoder-decoder structure, and the input layer is constructed. The input layer is used to receive time series data; 2-5) Construct an embedding layer, which is connected after the input layer, to embed the position index of the input sequence and the target sequence into high-dimensional vectors. These embeddings are trainable parameters that allow the model to adapt to different sequence lengths during the learning process; 2-6) Then, the position encoding is constructed, which is connected to the embedding layer. Through the position encoding module, the position information is introduced into the input sequence and the target sequence. The specific calculation is as follows: Among them, PE represents the obtained position information, pos represents the position, i represents the feature dimension, sin() and cos() represent the sine and cosine functions, d model Represents the dimension of the model. The position encoding matrix is ​​finally obtained through the formula calculation, which is used to add position encoding to the input in the forward propagation; 2-7) Build an encoder, which is composed of multiple encoder layers. Each encoder layer includes a multi-head self-attention mechanism and a feedforward neural network. Multiple attention heads are used to learn different patterns in the input sequence. For each head, the input Q, K, and V matrices are linearly transformed to generate the query, key, and value corresponding to each head. The specific calculation is as follows: Among them, X represents the embedded representation of the input sequence, Q, K, V are the query matrix, key matrix, and value matrix respectively, and i represents the i-th attention head. Represents the projection matrix for each head, concatenating the outputs of all heads together, which is calculated as follows: MultiHead=Concat(head1,...,head i )×W O Among them, MultiHead represents the final result of multi-head attention, W O Represents the final linear transformation matrix, Concat represents connection, head i Represents the attention calculation results of each head; 2-8) Construct a feedforward neural network, which is applied after each attention head. It consists of two layers of linear transformations. After each sub-layer, that is, after the multi-head self-attention mechanism and the feedforward network, residual connections and layer normalization are added. The specific calculation is as follows: Among them, LayerNorm is the layer normalization operation, x represents the input, Sublayer(x) is the output of the sublayer, +x represents the residual connection, μ and σ 2 is the mean and variance of Sublayer(x)+x, ε is a small constant, γ and β are learnable parameters; 2-9) Construct a decoder, which is composed of multiple decoder layers stacked together. The decoder layer is similar to the encoder layer, but the decoder layer will refer to the previous target sequence and the output of the encoder layer at the same time. In the calculation of cross-attention, K and V are replaced by the output of the encoder. The rest is the same as the encoder layer. The decoder layer also has a feedforward neural network layer and two layers of linear transformation. Similarly, residual connections and layer normalization are added after each sub-layer. 2-10) Finally, the encoder and decoder are connected to build a fully connected layer to generate the final output. The overall architecture is encoder-decoder; 2-11) Use the data to train the established multi-scale convolutional neural network and time-aware neural network. The training process uses knowledge-based features and gating constraints.

Images