A digital prediction system for side-blown furnaces based on deep learning

Abstract

The present invention discloses a digital prediction system for side-blown furnaces based on deep learning. The system comprises: a data acquisition and preprocessing module for collecting real-time sensor operating data and preprocessing the real-time sensors; a numerical calculation module for numerically simulating the multi-physics field smelting environment within the side-blown furnace based on the preprocessed operating data; a deep learning module for establishing and training a graph neural network (GNN) model based on the numerical simulation results, and using the GNN model to determine the dynamic behavior patterns of the molten pool smelting process; and a digital twin module for constructing a digital twin model consistent with the actual operating state of the side-blown furnace based on these dynamic behavior patterns. The present invention enables precise monitoring and prediction of side-blown furnaces, thereby optimizing the smelting process and improving operational efficiency.

Description

Technical Field

[0001] The present invention relates to the field of digital twin technology, and in particular to a digital intelligent prediction system for side-blown furnaces based on deep learning. Background Art

[0002] Existing industrial monitoring systems primarily rely on traditional numerical analysis methods, which have limitations when dealing with complex physical fields and dynamic changes. Especially in high-temperature, multi-physics coupled smelting environments such as side-blown furnaces, traditional monitoring systems struggle to accurately capture the dynamic behavior of the molten pool in real time. These systems often lack a deep understanding of the complex flow and heat transfer processes within the furnace, making it impossible to accurately predict and control the internal conditions of the furnace. Furthermore, existing monitoring systems often fail to effectively integrate real-time sensor data and numerical simulation results, resulting in low utilization of monitoring data and an inability to fully realize the data's potential value.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the existing technology: the existing system is unable to realize real-time monitoring and prediction of the multi-physical field status inside the furnace, lacks in-depth analysis of the dynamic behavior of the molten pool, and is unable to provide effective decision-making support. These problems limit the optimization of the smelting process and the improvement of production efficiency. Summary of the Invention

[0004] The embodiments of the present invention aim to provide a digital prediction system for side-blown furnaces based on deep learning to solve the technical problems raised in the prior art.

[0005] The embodiments of the present invention solve the technical problems by adopting the following technical solutions:

[0006] Provides a digital prediction system for side-blown furnaces based on deep learning, including:

[0007] The data acquisition and preprocessing module is used to collect real-time sensor working data through real-time sensors and preprocess the real-time sensors;

[0008] Numerical calculation module, used to perform numerical simulation of the multi-physics field smelting environment in the side-blown furnace based on pre-processed working data;

[0009] A deep learning module is used to establish and train a graph neural network (GNN) model based on numerical simulation results, and to obtain the dynamic behavior of the molten pool smelting process through the GNN model;

[0010] The digital twin module is used to construct a digital twin model consistent with the actual operating state of the side-blown furnace based on the dynamic behavior law.

[0011] Furthermore, in the data acquisition and preprocessing module, preprocessing includes removing random noise, identifying and eliminating outliers, multiple redundant processing, and adaptive sampling rate adjustment.

[0012] Furthermore, removing random noise includes:

[0013] The prediction equation of the Kalman filter is used to estimate the current state. The prediction process is shown in formula (1):

[0014]

[0015] in, is the prior state estimate at time k based on time k-1, A is the state transfer matrix, B is the control input matrix, u k-1 is the control input at time k-1;

[0016] The Kalman filter update equation is used to optimize the predicted state and obtain the updated equation after removing random noise as shown in formula (2):

[0017]

[0018] Among them, K k is the Kalman gain, P k|k-1 is the prior state estimation covariance matrix, H is the observation matrix, and R is the observation noise covariance matrix;

[0019] According to the difference between the observed value and the predicted value, the predicted state is adjusted to obtain the real-time sensor working data after removing random noise.

[0020] Furthermore, identifying and removing outliers includes:

[0021] Calculate the interquartile range IQR,

[0022] IQR=Q3-Q1 (3),

[0023] The outlier range is determined based on the interquartile range (IQR), as shown in formula (4):

[0024] [Q1-1.5×IQR, Q3+1.5×IQR] (4),

[0025] Among them, Q1 is the first quartile and Q3 is the third quartile.

[0026] Furthermore, the numerical calculation module uses computational fluid dynamics (CFD) methods to perform numerical simulations on the multi-physical field smelting environment in the side-blown furnace based on the pre-processed working data.

[0027] Furthermore, computational fluid dynamics (CFD) methods include:

[0028] By solving the mass conservation equation, momentum conservation equation and energy conservation equation to simulate and predict the flow and heat transfer process in the furnace, the numerical simulation results are obtained. The numerical simulation results include mass results, momentum results and energy results;

[0029] The mass conservation equation is shown in formula (5):

[0030]

[0031] The momentum conservation equation is shown in formula (6):

[0032]

[0033] The energy conservation equation is shown in formula (7):

[0034]

[0035] where ρ is the fluid density, is the fluid velocity vector, t is time, p is the fluid pressure, μ is the molecular viscosity, represents the gravitational acceleration vector of the fluid, F ST represents the surface tension between the phases in each control volume.

[0036] Furthermore, the deep learning module combines the real-time sensor working data and numerical simulation results to establish and train the graph neural network GNN model, and obtains the dynamic behavior law of the molten pool smelting process through the GNN model. Among them, the real-time sensor working data includes fluid density ρ, fluid velocity vector Pressure p, molecular viscosity μ, gravitational acceleration vector and surface tension F ST .

[0037] Furthermore, the dynamic behavior of the molten pool smelting process obtained by learning the GNN model includes the following steps:

[0038] The multi-physics smelting environment in the side-blown furnace is discretized into a graph structure, where nodes represent key locations or control volumes in the smelting environment, and edges represent physical connections or interactions between these locations or control volumes.

[0039] Each node contains its corresponding physical quantity, such as fluid density ρ, fluid velocity vector Pressure p, molecular viscosity μ, gravitational acceleration vector and surface tension F ST Etc., these physical quantities are used as node feature vectors;

[0040] The GNN model is used to iteratively update these node features to capture the complex interactions and information transfer between nodes, where the new feature vector of each node is It is calculated from the information of its neighbor nodes and the characteristics of the edges in the tth iteration;

[0041] Through the information propagation mechanism of the GNN model, the update rules of node features are as follows:

[0042]

[0043] Where N(i) is the node v i The neighbor node set of node v, f is the aggregation function used to combine i and its neighbor node v j Information, e ij Represents node v i and v j The characteristics of the edges between them;

[0044] The aggregation function f can be expressed as a weighted sum:

[0045]

[0046] Among them, α ij It is edge e ij The weight of may be related to the distance between nodes or the transmission intensity of physical quantities;

[0047] The update function is:

[0048]

[0049] Where W is the weight matrix and σ is the activation function, such as the ReLU function;

[0050] Through the above iterative update process, the GNN model can learn the dynamic behavior laws of the molten pool smelting process, including the multi-physical field states such as the molten pool temperature distribution, fluid flow pattern, composition distribution and their evolution process over time, thereby realizing the prediction of the multi-physical field states and evolution process in the side-blown furnace.

[0051] Furthermore, the digital twin module constructs a digital twin model consistent with the actual side-blown furnace operating state based on the dynamic behavior rules obtained by the GNN model. The process includes:

[0052] The multi-physics field states and evolution processes predicted by the GNN model are used as the initial conditions and boundary conditions of the digital twin model;

[0053] Combined with the CFD simulation results provided by the numerical calculation module, including the numerical values ​​of the mass conservation equation (Equation 5), momentum conservation equation (Equation 6), and energy conservation equation (Equation 7), to calibrate and refine the physical behavior of the digital twin model;

[0054] Real-time update of the digital twin model, through the real-time sensor data provided by the integrated data acquisition and preprocessing module, such as fluid density ρ, fluid velocity vector Pressure p, molecular viscosity μ, gravitational acceleration vector and surface tension F ST , ensuring that the digital twin model reflects the real-time status of the side-blown kiln;

[0055] The digital twin model provides a global visualization of the process status inside the furnace, enabling real-time simulation and monitoring of the dynamic state inside the furnace, thereby assisting operational decision-making and process optimization.

[0056] Furthermore, the digital twin model includes:

[0057] The data fusion unit is used to integrate the prediction results of the deep learning module, the simulation results of the numerical calculation module, and the real-time data of the data acquisition and preprocessing module;

[0058] A model update unit, used to dynamically adjust the digital twin model based on the latest data and prediction results, ensuring that the model is consistent with the actual operating status of the side-blown kiln;

[0059] A visualization interface unit is used to display the status of the digital twin model, including the multi-physics field distribution and evolution process of the molten pool, as well as possible abnormality warnings, providing intuitive decision support for operators;

[0060] Feedback loop unit allows operator intervention and adjustments to be fed back into the model for self-learning and optimization, improving the accuracy of predictions and simulations.

[0061] The above-mentioned embodiments of the present invention have at least the following beneficial effects: by combining real-time sensor data and deep learning technology, the system can accurately numerically simulate and predict the dynamic behavior of the multi-physical field smelting environment in the side-blown furnace. This integrated monitoring-prediction system can provide a global visualization of the process status in the furnace, and provide intuitive decision-making support for operators, thereby realizing real-time simulation and monitoring of the dynamic state in the furnace. In addition, the system uses digital twin technology to construct a digital twin model that is consistent with the actual operating status of the side-blown furnace. The model can be updated in real time to reflect the real-time status of the furnace, and self-learning and optimization are carried out through the feedback loop unit to improve the accuracy of prediction and simulation. This can not only enhance the safety and reliability of furnace operation, but also improve smelting efficiency and product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] One or more embodiments are exemplarily illustrated by the figures in the corresponding drawings. These exemplifications do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings are represented as similar elements. Unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0063] Figure 1 A schematic structural diagram of a digital prediction system for side-blown furnaces based on deep learning provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0064] For ease of understanding of the present invention, the present invention will be described in more detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that when an element is described as "connected" to another element, it can be directly on another element, or there can be one or more centered elements therebetween. The orientation or positional relationship indicated by the terms "upper", "lower", "left", "right", "upper end", "lower end", "top" and "bottom" used in this specification is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0065] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0066] The following combination Figure 1 , the side-blown furnace digital prediction system 100 based on deep learning provided in the embodiment of the present application is described in detail through specific examples.

[0067] Figure 1 Schematic diagram of the structure of the digital prediction system for side-blown furnaces based on deep learning provided by the present invention. One embodiment of the present invention provides a digital prediction system for side-blown furnaces based on deep learning, including:

[0068] The data acquisition and preprocessing module 101 is used to collect real-time sensor working data through the real-time sensor and preprocess the real-time sensor;

[0069] A numerical calculation module 102 is used to perform numerical simulation of the multi-physical field smelting environment in the side-blown furnace based on the pre-processed working data;

[0070] A deep learning module 103 is used to establish and train a graph neural network (GNN) model based on numerical simulation results, and to obtain the dynamic behavior of the molten pool smelting process through the GNN model;

[0071] The digital twin module 104 is used to construct a digital twin model consistent with the actual operating state of the side-blown furnace based on the dynamic behavior law.

[0072] It should be noted that the system includes a data acquisition and preprocessing module, which is responsible for collecting real-time sensor data and preprocessing this data. Here, the data acquisition and preprocessing module refers to the part of the system that collects data from kiln sensors and performs necessary processing so that subsequent modules can use this data more effectively. Real-time sensors are devices installed on side-blown kilns that continuously monitor and record the kiln's operating status.

[0073] Specifically, the preprocessing steps in the data acquisition and preprocessing module include random noise removal, outlier identification and removal, multiple redundancy processing, and adaptive sampling rate adjustment. These steps ensure data quality and accuracy, providing reliable input for subsequent numerical simulations and deep learning. For example, random noise removal can be achieved using a Kalman filter, a commonly used algorithm for estimating the state of a dynamic system from a series of noisy measurements.

[0074] Preferably, random noise removal during the preprocessing step can be implemented using the prediction and update equations of a Kalman filter. The prediction equation estimates the current state, while the update equation optimizes the predicted state, resulting in data free of random noise. During this process, parameters such as the Kalman gain, the prior state estimate covariance matrix, and the observation noise covariance matrix need to be adjusted and optimized based on the actual operating characteristics of the kiln to ensure optimal filter performance.

[0075] Furthermore, the identification and removal of outliers can be achieved by calculating the interquartile range (IQR) and determining the outlier range based on it. This is a statistical method used to identify outliers in data.

[0076] In some embodiments, in the data acquisition and preprocessing module, preprocessing includes removing random noise, identifying and eliminating outliers, multiple redundant processing, and adaptive sampling rate adjustment.

[0077] It's important to note that within the system's data acquisition and preprocessing module, preprocessing steps include removing random noise, identifying and eliminating outliers, handling multiple redundancies, and adaptively adjusting the sampling rate. Preprocessing, in this context, refers to a series of operations performed on raw data before further analysis to improve data quality and reduce the complexity of subsequent processing. Random noise refers to irregular fluctuations in data due to measurement errors or other random factors.

[0078] Specifically, the removal of random noise in the preprocessing step can be achieved by using the prediction equation of the Kalman filter, which is used to estimate the current state and has the form: in, represents the prior state estimate at time k based on time k-1, Φ represents the state transfer matrix, u k-1 represents the control input at time k-1.

[0079] More specifically, identifying and eliminating outliers can be achieved by calculating the interquartile range (IQR). IQR refers to the difference between the third quartile and the first quartile in statistics, which is used to measure the degree of dispersion of data.

[0080] Preferably, the multiple redundant processing in the preprocessing step can be achieved through data fusion technology, which integrates data from multiple sensors to improve data robustness. Adaptive sampling rate adjustment can dynamically adjust the sampling frequency based on the actual operating status of the kiln to meet different monitoring needs. For example, when the kiln is in a critical operation phase, the sampling rate may need to be increased to capture more detailed changes.

[0081] Furthermore, in the process of removing outliers, in addition to using the IQR method, other statistical methods such as the standard deviation method or anomaly detection algorithm based on machine learning can also be considered to further improve the accuracy of outlier identification.

[0082] In some embodiments, removing random noise includes:

[0083] The prediction equation of the Kalman filter is used to estimate the current state. The prediction process is shown in formula (1):

[0084]

[0085] in, is the prior state estimate at time k based on time k-1, A is the state transfer matrix, B is the control input matrix, u k-1 is the control input at time k-1;

[0086] The Kalman filter update equation is used to optimize the predicted state and obtain the updated equation after removing random noise as shown in formula (2):

[0087]

[0088] Among them, K k is the Kalman gain, P k|k-1 is the prior state estimation covariance matrix, H is the observation matrix, and R is the observation noise covariance matrix;

[0089] According to the difference between the observed value and the predicted value, the predicted state is adjusted to obtain the real-time sensor working data after removing random noise.

[0090] It's important to note that in the system's data acquisition and preprocessing module, the random noise removal step involves estimating the current state using the prediction equation of a Kalman filter. The Kalman filter is an efficient recursive algorithm for estimating the state of a dynamic system from a series of noisy measurements. The filter consists of two parts: a prediction step, which predicts the current state based on previous estimates and a system model, and an update step.

[0091] Specifically, the prediction equation of the Kalman filter is used to predict the state of the system at the next moment, and its formula is: in, represents the prior state estimate at time k based on time k-1, is the state estimate of the previous moment, Φ is the state transfer matrix, u k-1 is the control input at time k-1. This equation takes into account the system dynamics and the impact of the control input on the state, providing a basis for subsequent update steps.

[0092] Preferably, the parameters in the prediction equation of the Kalman filter can be adjusted according to the actual system characteristics. For example, the state transfer matrix Φ can be determined according to the physical model and dynamic characteristics of the system, and the control input u k-1 It can be an external control signal or feedback control within the system.

[0093] Furthermore, to improve filter performance, one can consider using an extended Kalman filter (EKF) or an unscented Kalman filter (UKF). These are variants of the Kalman filter that are better able to handle nonlinear systems. In practical applications, filtering performance can also be optimized by adjusting the filter's noise covariance parameters, which describe the characteristics of process noise and measurement noise.

[0094] In some embodiments, identifying and removing outliers includes:

[0095] Calculate the interquartile range IQR,

[0096] IQR=Q3-Q1 (3),

[0097] The outlier range is determined based on the interquartile range (IQR), as shown in formula (4):

[0098] [Q1-1.5×IQR, Q3+1.5×IQR] (4),

[0099] Among them, Q1 is the first quartile and Q3 is the third quartile.

[0100] It's important to note that the system's steps for identifying and removing outliers include calculating the interquartile range (IQR), which is used to determine the range of outliers. The IQR is a statistical term that refers to the difference between the third quartile (Q3) and the first quartile (Q1) in a dataset and is used to measure the degree of dispersion in the data. The IQR provides a robust measure of data distribution because it's not affected by extreme values.

[0101] Specifically, the formula for calculating the interquartile range (IQR) is: IQR = Q3 - Q1, where Q3 is the value at the 75th percentile in the data set, and Q1 is the value at the 25th percentile. To determine the range of outliers based on the interquartile range (IQR), multiply the data value by 1.5 times the IQR, then subtract and add this value from Q1 and Q3 to obtain the range of outliers: [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR]. Data within this range is considered normal, while data outside this range is considered an outlier and needs to be removed.

[0102] Preferably, the specific steps for identifying outliers can be further refined. For example, when calculating the IQR, data analysis tools such as R or Python can be used to automate the process. When removing outliers, a threshold can be set, such as 1.5 times the IQR. This threshold can be adjusted based on the actual application scenario and data characteristics.

[0103] Furthermore, in addition to IQR-based methods, other outlier detection techniques such as standard deviation, Z-score, or machine learning-based anomaly detection algorithms can be considered to improve the accuracy and robustness of outlier identification. These methods can be used as alternatives or in conjunction with IQR methods to enhance the system's outlier handling capabilities.

[0104] In some embodiments, the numerical calculation module uses a computational fluid dynamics (CFD) method to perform numerical simulation on the multi-physical field smelting environment in the side-blown furnace based on the preprocessed working data.

[0105] It should be noted that the numerical calculation module in this system uses computational fluid dynamics (CFD) to numerically simulate the multi-physics smelting environment within the side-blown furnace based on preprocessed operating data. Computational fluid dynamics (CFD) is a numerical analysis and data structuring technique used to solve and analyze problems involving fluid flow. CFD simulates fluid flow and heat transfer by solving fluid dynamics equations.

[0106] Specifically, the CFD method simulates and predicts the flow and heat transfer process within the furnace by solving the mass conservation equation, momentum conservation equation, and energy conservation equation. These equations describe the conservation laws of fluid mass, momentum, and energy, respectively, and are fundamental equations in fluid mechanics. The mass conservation equation is expressed as:

[0107]

[0108] The momentum conservation equation is expressed as:

[0109]

[0110] The energy conservation equation is expressed as:

[0111]

[0112] Where ρ is the fluid density, u is the fluid velocity vector, p is the fluid pressure, μ is the molecular viscosity, g is the gravitational acceleration vector, k is the thermal conductivity, T is the temperature, is the heat source term per unit mass.

[0113] Preferably, the specific parameter settings and concepts used in the CFD simulation process include the physical properties of the fluid, such as density, viscosity, and thermal conductivity. These parameters can be set based on the characteristics of the materials actually used in the kiln. During the simulation process, various numerical methods, such as finite difference method, finite volume method, or finite element method, can be used to discretize and solve the above equations.

[0114] Furthermore, the internal geometry and boundary conditions of the furnace can be considered, such as the heat exchange characteristics of the furnace walls and the material input and output conditions. During the solution process, commercial software such as ANSYS Fluent and OpenFOAM, or independently developed CFD software, can be used for simulation. To improve simulation accuracy, experimental data can be used to calibrate and validate the CFD model.

[0115] In some embodiments, the computational fluid dynamics (CFD) method includes:

[0116] By solving the mass conservation equation, momentum conservation equation and energy conservation equation to simulate and predict the flow and heat transfer process in the furnace, the numerical simulation results are obtained. The numerical simulation results include mass results, momentum results and energy results;

[0117] The mass conservation equation is shown in formula (5):

[0118]

[0119] The momentum conservation equation is shown in formula (6):

[0120]

[0121] The energy conservation equation is shown in formula (7):

[0122]

[0123] where ρ is the fluid density, is the fluid velocity vector, t is time, p is the fluid pressure, μ is the molecular viscosity, represents the gravitational acceleration vector of the fluid, F ST represents the surface tension between the phases in each control volume.

[0124] It should be noted that the numerical calculation module in this system simulates and predicts the flow and heat transfer processes within the furnace by solving the mass conservation equations, momentum conservation equations, and energy conservation equations. These are fundamental equations in fluid mechanics that describe fluid flow and heat transfer. They represent the conservation laws of fluid mass, momentum, and energy within a control volume, respectively.

[0125] Specifically, the mass conservation equation is expressed as:

[0126]

[0127] Where ρ represents the fluid density, u represents the fluid velocity vector, and t represents time. The momentum conservation equation is expressed as:

[0128]

[0129] Where p represents the pressure of the fluid, μ represents the molecular viscosity, g represents the gravity acceleration vector, and f represents other external forces. The energy conservation equation is expressed as:

[0130]

[0131] Where E represents the total energy per unit mass, k represents thermal conductivity, and T represents temperature. Represents the heat source term per unit mass.

[0132] Preferably, these equations can be solved using numerical methods such as the finite volume method (FVM) or the finite element method (FEM). In practical implementation, appropriate boundary and initial conditions need to be set, such as adiabatic or convective boundary conditions on the furnace walls, and the initial temperature and fluid velocity distribution at furnace startup.

[0133] Furthermore, the physical properties of the materials in the furnace, such as density, viscosity, and thermal conductivity, can be considered. These parameters can be set based on the actual material characteristics. During the solution process, commercial CFD software such as ANSYS Fluent and OpenFOAM, or independently developed CFD software, can be used for simulation. To improve simulation accuracy, experimental data can be used to calibrate and validate the CFD model.

[0134] Furthermore, multiphase flow characteristics inside the furnace can also be considered, such as gas-liquid or gas-solid flow, which may require the introduction of additional phase interface tracking and interphase force models.

[0135] In some embodiments, the deep learning module combines real-time sensor working data and numerical simulation results to establish and train a graph neural network GNN model, and obtains the dynamic behavior law of the molten pool smelting process through the GNN model, wherein the real-time sensor working data includes fluid density ρ, fluid velocity vector Pressure p, molecular viscosity μ, gravitational acceleration vector and surface tension F ST .

[0136] It should be noted that the system's deep learning module combines real-time sensor data with numerical simulation results to build and train a graph neural network (GNN) model, which then captures the dynamic behavior of the molten pool smelting process. A GNN is a deep learning model that processes graph-structured data and learns node representations through the relationships between nodes and edges, making it suitable for capturing the interactions between components in complex systems.

[0137] Specifically, the establishment and training of the GNN model involves real-time sensor data, including fluid density ρ, fluid velocity vector u, pressure p, molecular viscosity μ, gravity acceleration vector g, and surface tension σ. These physical quantities are input into the GNN model as node feature vectors. The model learns the dynamic behavior of the molten pool smelting process by iteratively updating the node features. During the iterative update process, the new feature vector h of each node is (new) The update rule is calculated based on the information of its neighbor nodes and the characteristics of the edges and can be expressed as:

[0138]

[0139] Among them, N(v) is the set of neighbor nodes of node v, AGGREGATE is an aggregation function used to combine the information of node v and its neighbor node n, e vn Features representing the edge between nodes v and n.

[0140] Preferably, the iterative update process of the GNN model can be further refined. For example, the aggregation function AGGREGATE can be a weighted sum, average pooling, or attention mechanism to combine information about a node and its neighbors. The update function UPDATE can include linear transformations and nonlinear activation functions, such as the ReLU function, to enhance the model's nonlinear expression capabilities.

[0141] Furthermore, during model training, the cross-entropy loss function can be used to optimize classification tasks, or the mean squared error loss function can be used to optimize regression tasks. In practical applications, different GNN variants can also be considered, such as graph convolutional networks (GCNs), graph attention networks (GATs), or message passing neural networks (MPNNs). These models have different advantages and characteristics when processing graph-structured data. To improve the generalization and robustness of the model, techniques such as data augmentation, dropout regularization, or early stopping can also be used.

[0142] In some embodiments, learning the dynamic behavior of the molten pool smelting process through the GNN model includes the following steps:

[0143] The multi-physics smelting environment in the side-blown furnace is discretized into a graph structure, where nodes represent key locations or control volumes in the smelting environment, and edges represent physical connections or interactions between these locations or control volumes.

[0144] Each node contains its corresponding physical quantity, such as fluid density ρ, fluid velocity vector Pressure p, molecular viscosity μ, gravitational acceleration vector and surface tension F ST Etc., these physical quantities are used as node feature vectors;

[0145] The GNN model is used to iteratively update these node features to capture the complex interactions and information transfer between nodes, where the new feature vector of each node is It is calculated from the information of its neighbor nodes and the characteristics of the edges in the tth iteration;

[0146] Through the information propagation mechanism of the GNN model, the update rules of node features are as follows:

[0147]

[0148] Where N(i) is the node v iThe neighbor node set of node v, f is the aggregation function used to combine i and its neighbor node v j Information, e ij Represents node v i and v j The characteristics of the edges between them;

[0149] The aggregation function f can be expressed as a weighted sum:

[0150]

[0151] Among them, α ij It is edge e ij The weight of may be related to the distance between nodes or the transmission intensity of physical quantities;

[0152] Update function i is a linear transformation and can be expressed as:

[0153]

[0154] Where W is the weight matrix and σ is the activation function, such as the ReLU function;

[0155] Through the above iterative update process, the GNN model can learn the dynamic behavior laws of the molten pool smelting process, including the multi-physical field states such as the molten pool temperature distribution, fluid flow pattern, composition distribution and their evolution process over time, thereby realizing the prediction of the multi-physical field states and evolution process in the side-blown furnace.

[0156] It should be noted that the dynamic behavior of the melt pool smelting process learned through the GNN model in this system involves discretizing the multi-physics smelting environment within the side-blown furnace into a graph structure. Here, a graph structure is a mathematical model used to represent objects (nodes) and their relationships (edges). In this system, nodes represent key locations or control volumes in the smelting environment, while edges represent the physical connections or interactions between these locations or control volumes.

[0157] Specifically, each node contains its corresponding physical quantities, such as fluid density ρ, fluid velocity vector u, pressure p, molecular viscosity μ, gravity acceleration vector g, and surface tension σ, etc. These physical quantities serve as node feature vectors. The GNN model iteratively updates these node features to capture the complex interactions and information transfer between nodes. During the iterative update process, the new feature vector of each node is The update rule is calculated based on the information of its neighbor nodes and the characteristics of the edges and can be expressed as:

[0158]

[0159] Among them, N(v) is the set of neighbor nodes of node v, AGGREGATE is an aggregation function used to combine the information of node v and its neighbor node n, e vn Features representing the edge between nodes v and n.

[0160] Preferably, the information propagation mechanism of the GNN model can be further refined. For example, the aggregation function AGGREGATE can adopt an attention mechanism, allowing the model to dynamically assign different weights to different neighboring nodes, thereby better capturing the complex relationships between nodes. The update function UPDATE can include nonlinear activation functions, such as ReLU, to increase the expressive power of the model.

[0161] Furthermore, a variety of loss functions can be used during model training, such as mean squared error loss or cross entropy loss, to adapt to different prediction tasks. In practical applications, different GNN architectures can also be considered, such as graph convolutional networks (GCNs) or graph attention networks (GATs), to adapt to different data characteristics and task requirements. To improve the generalization ability of the model, regularization techniques such as L2 regularization or dropout, as well as data augmentation strategies such as random perturbations of node features or perturbations of the graph structure, can be used.

[0162] In some embodiments, the process of constructing a digital twin model consistent with the actual side-blown furnace operating state based on the dynamic behavior law obtained by the GNN model by the digital twin module includes:

[0163] The multi-physics field states and evolution processes predicted by the GNN model are used as the initial conditions and boundary conditions of the digital twin model;

[0164] Combined with the CFD simulation results provided by the numerical calculation module, including the numerical values ​​of the mass conservation equation (Equation 5), momentum conservation equation (Equation 6), and energy conservation equation (Equation 7), to calibrate and refine the physical behavior of the digital twin model;

[0165] Real-time update of the digital twin model, through the real-time sensor data provided by the integrated data acquisition and preprocessing module, such as fluid density ρ, fluid velocity vector Pressure p, molecular viscosity μ, gravitational acceleration vector and surface tension F ST , ensuring that the digital twin model reflects the real-time status of the side-blown kiln;

[0166] The digital twin model provides a global visualization of the process status inside the furnace, enabling real-time simulation and monitoring of the dynamic state inside the furnace, thereby assisting operational decision-making and process optimization.

[0167] It should be noted that the digital twin module in this system constructs a digital twin model that is consistent with the actual operating state of the side-blown kiln based on the dynamic behavior patterns derived from the GNN model. A digital twin model, in this context, refers to a virtual, digital representation that accurately maps and updates in real time to reflect the state and behavior of the physical entity. This model uses the multi-physics state and evolution predicted by the GNN model as initial and boundary conditions.

[0168] Specifically, the digital twin module combines the CFD simulation results provided by the numerical calculation module, including the numerical values ​​of the mass conservation equations, momentum conservation equations, and energy conservation equations, to calibrate and refine the physical behavior of the digital twin model. This involves using the distribution and changes in physical quantities such as fluid density, velocity, and pressure obtained from the CFD simulation as input, and adjusting the digital twin model to ensure that its predictions match the actual operating conditions of the kiln.

[0169] More specifically, real-time updating of the digital twin model involves integrating real-time sensor data provided by the data acquisition and preprocessing module, such as fluid density ρ, fluid velocity vector Pressure p, molecular viscosity μ, gravitational acceleration vector and surface tension F ST , ensuring that the digital twin model reflects the real-time status of the side-blown kiln.

[0170] Preferably, the implementation of the digital twin module can be further refined. For example, the data fusion unit can be designed as an advanced data integration system that not only integrates the prediction results of the deep learning module and the simulation results of the numerical calculation module, but also integrates data streams from different sensors, such as temperature, pressure, and flow sensors. The model update unit can use machine learning algorithms, such as online learning or incremental learning, to adapt to new data and dynamically adjust model parameters. The visualization interface unit can be designed as an interactive dashboard that displays the status of the digital twin model, including the multi-physics field distribution and evolution process of the melt pool, as well as possible abnormality warnings, providing intuitive decision support for operators. The feedback loop unit can be designed as a closed-loop control system that allows operator interventions and adjustments to be fed back into the model for self-learning and optimization, thereby improving the accuracy of predictions and simulations. In addition, the use of cloud computing platforms to support large-scale data storage and computing needs, as well as the adoption of edge computing technologies to reduce data transmission delays and improve the system's response speed, can be considered.

[0171] In some embodiments, the digital twin model includes:

[0172] The data fusion unit is used to integrate the prediction results of the deep learning module, the simulation results of the numerical calculation module, and the real-time data of the data acquisition and preprocessing module;

[0173] A model update unit, used to dynamically adjust the digital twin model based on the latest data and prediction results, ensuring that the model is consistent with the actual operating status of the side-blown kiln;

[0174] A visualization interface unit is used to display the status of the digital twin model, including the multi-physics field distribution and evolution process of the molten pool, as well as possible abnormality warnings, providing intuitive decision support for operators;

[0175] Feedback loop unit allows operator intervention and adjustments to be fed back into the model for self-learning and optimization, improving the accuracy of predictions and simulations.

[0176] It should be noted that the digital twin model in this system includes a data fusion unit, a model update unit, a visualization interface unit, and a feedback loop unit. The data fusion unit is a component that integrates data from different sources. It combines the prediction results of the deep learning module, the simulation results of the numerical calculation module, and the real-time data of the data acquisition and preprocessing module.

[0177] Specifically, the data fusion unit is responsible for receiving and processing data from multiple modules. This data includes but is not limited to physical quantities such as fluid density, fluid velocity vector, pressure, molecular viscosity, gravitational acceleration vector, and surface tension. The model update unit dynamically adjusts the digital twin model based on this fused data and prediction results to ensure that the model is consistent with the actual operating status of the side-blown furnace. The visual interface unit is used to display the status of the digital twin model and provide intuitive decision support, while the feedback loop unit allows operator interventions and adjustments to be fed back into the model for self-learning and optimization.

[0178] Preferably, the operating steps of the digital twin model can be further refined. For example, the data fusion unit can adopt a specific data fusion algorithm, such as Kalman filtering or particle filtering, to optimize the data integration process. The model update unit can set specific parameters, such as learning rate and model weight update strategy, to adapt to new data and dynamically adjust model parameters. The visualization interface unit can be designed as an interactive dashboard to display the multi-physical field distribution and evolution process of the molten pool, as well as possible abnormal warnings, to provide intuitive decision support for operators. The feedback loop unit can be designed as a closed-loop control system, allowing the operator's intervention and adjustment to be fed back to the model in real time, so as to carry out self-learning and optimization of the model and improve the accuracy of prediction and simulation. In addition, it is possible to consider using artificial intelligence algorithms to enhance the intelligent decision-making ability of the feedback loop unit, such as using reinforcement learning to automatically adjust operating parameters to optimize furnace performance.

[0179] The above-mentioned embodiments of the present invention have the following beneficial effects: the system described in the present invention can provide an innovative monitoring and prediction method, which can realize comprehensive monitoring and dynamic prediction of the multi-physical field smelting environment of the side-blown furnace through the coupling of real-time data acquisition and preprocessing, numerical calculation, deep learning and digital twin technology. This integrated system can effectively remove random noise, eliminate outliers, and perform multiple redundant processing and adaptive sampling rate adjustment, thereby improving the accuracy and reliability of the data. In addition, the system learns the dynamic behavior laws of the molten pool smelting process through the graph neural network model, which can capture the complex interactions and information transmission between nodes, and realize accurate prediction of the multi-physical field state and evolution process in the side-blown furnace.

[0180] The system also enables real-time simulation and monitoring of the furnace's internal state, providing a global visualization through a digital twin model to assist in operational decision-making and process optimization. The system's data fusion unit integrates the predictions of the deep learning module, the simulation results of the numerical calculation module, and real-time data. The model update unit dynamically adjusts the digital twin model based on the latest data and predictions, ensuring that the model remains consistent with the actual operating state of the side-blown furnace. This real-time update and feedback loop mechanism improves the system's prediction accuracy and response speed, providing strong technical support for the intelligent control and optimization of industrial furnaces.

[0181] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, a server, a mobile phone, and a tablet.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Under the idea of ​​the present invention, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as above, which are not provided in detail for the sake of simplicity. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A digital prediction system for side-blown furnaces based on deep learning, characterized in that: Includes the following modules: The data acquisition and preprocessing module is used to collect real-time sensor working data through real-time sensors and preprocess the real-time sensors; Numerical calculation module, used to perform numerical simulation of the multi-physics field smelting environment in the side-blown furnace based on pre-processed working data; A deep learning module is used to establish and train a graph neural network (GNN) model based on numerical simulation results, and to obtain the dynamic behavior of the molten pool smelting process through the GNN model; A digital twin module is used to construct a digital twin model consistent with the actual side-blown furnace operating state based on the dynamic behavior law; In the data acquisition and preprocessing module, preprocessing includes removing random noise, identifying and eliminating outliers, multiple redundancy processing, and adaptive sampling rate adjustment; Identifying and removing outliers includes: Calculate the interquartile range IQR, IQR=Q3-Q1 (3), The outlier range is determined based on the interquartile range (IQR), as shown in formula (4): [Q1-1.5×IQR, Q3+1.5×IQR] (4), Among them, Q1 is the first quartile and Q3 is the third quartile; The numerical calculation module uses computational fluid dynamics (CFD) methods to perform numerical simulations on the multi-physical field smelting environment in the side-blown furnace based on pre-processed working data. Computational fluid dynamics (CFD) methods include: By solving the mass conservation equation, momentum conservation equation and energy conservation equation to simulate and predict the flow and heat transfer process in the furnace, the numerical simulation results are obtained. The numerical simulation results include mass results, momentum results and energy results; The mass conservation equation is shown in formula (5): The momentum conservation equation is shown in formula (6): The energy conservation equation is shown in formula (7): where ρ is the fluid density, is the fluid velocity vector, t is time, p is the fluid pressure, μ is the molecular viscosity, represents the gravitational acceleration vector of the fluid, F ST represents the surface tension between the phases in each control volume.

2. The system according to claim 1, wherein: Removing random noise includes: The prediction equation of the Kalman filter is used to estimate the current state. The prediction process is shown in formula (1): in, is the prior state estimate at time k based on time k-1, A is the state transfer matrix, B is the control input matrix, u k-1 is the control input at time k-1; The Kalman filter update equation is used to optimize the predicted state and obtain the updated equation after removing random noise as shown in formula (2): Among them, K k is the Kalman gain, P k|k-1 is the prior state estimation covariance matrix, H is the observation matrix, and R is the observation noise covariance matrix; According to the difference between the observed value and the predicted value, the predicted state is adjusted to obtain the real-time sensor working data after removing random noise.

3. The system according to claim 2, characterized in that The deep learning module combines real-time sensor working data and numerical simulation results to establish and train the graph neural network GNN model. The dynamic behavior law of the molten pool smelting process is obtained through the GNN model. Among them, the real-time sensor working data includes fluid density ρ, fluid velocity vector u → , pressure p, molecular viscosity μ, gravitational acceleration vector g → and surface tension F ST .

4. The system according to claim 3, characterized in that Learning the dynamic behavior of the molten pool smelting process through the GNN model includes the following steps: The multi-physics smelting environment in the side-blown furnace is discretized into a graph structure, where nodes represent key locations or control volumes in the smelting environment, and edges represent physical connections or interactions between these locations or control volumes. Each node contains its corresponding physical quantity, such as fluid density ρ, fluid velocity vector Pressure p, molecular viscosity μ, gravitational acceleration vector and surface tension F ST , these physical quantities serve as node eigenvectors; The GNN model is used to iteratively update these node features to capture the complex interactions and information transfer between nodes, where the new feature vector of each node is It is calculated from the information of its neighbor nodes and the characteristics of the edges in the tth iteration; Through the information propagation mechanism of the GNN model, the update rule of node features is: Where N(i) is the node v i The neighbor node set of node v, f is the aggregation function used to combine i and its neighbor node v j Information, e ij Represents node v i and v j Features of the edges between them; The aggregate function f is: Among them, α ij It is edge e ij The weight of may be related to the distance between nodes or the transmission intensity of physical quantities; The update function is: Where W is the weight matrix and σ is the activation function; Through the above iterative update process, the GNN model can learn the dynamic behavior laws of the molten pool smelting process, including the temperature distribution, fluid flow pattern, composition distribution multi-physical field state and its evolution process over time, thereby realizing the prediction of the multi-physical field state and evolution process in the side-blown furnace.

5. The system according to claim 4, characterized in that The digital twin module constructs a digital twin model consistent with the actual side-blown furnace operating status based on the dynamic behavior laws obtained by the GNN model. The process includes: The multi-physics field states and evolution processes predicted by the GNN model are used as the initial conditions and boundary conditions of the digital twin model; Combined with the CFD simulation results provided by the numerical calculation module, including the numerical values ​​of the mass conservation equation, momentum conservation equation, and energy conservation equation, to calibrate and refine the physical behavior of the digital twin model; Real-time update of the digital twin model, through the real-time sensor data provided by the integrated data acquisition and preprocessing module, such as fluid density ρ, fluid velocity vector Pressure p, molecular viscosity μ, gravitational acceleration vector and surface tension F ST , ensuring that the digital twin model reflects the real-time status of the side-blown kiln; The digital twin model provides a global visualization of the process status inside the furnace, enabling real-time simulation and monitoring of the dynamic state inside the furnace, thereby assisting operational decision-making and process optimization.

6. The system according to claim 5, characterized in that The digital twin model includes: The data fusion unit is used to integrate the prediction results of the deep learning module, the simulation results of the numerical calculation module, and the real-time data of the data acquisition and preprocessing module; A model update unit, used to dynamically adjust the digital twin model based on the latest data and prediction results, ensuring that the model is consistent with the actual operating status of the side-blown kiln; A visualization interface unit is used to display the status of the digital twin model, including the multi-physics field distribution and evolution process of the molten pool, as well as possible abnormality warnings, providing intuitive decision support for operators; Feedback loop unit allows operator intervention and adjustments to be fed back into the model for self-learning and optimization, improving the accuracy of predictions and simulations.

Images