Converter steelmaking multi-physical field soft measurement method and system for physically constraining teacher-student network

Abstract

The present application provides a kind of physical constraint converter steelmaking multi-physical field soft measurement method and system of teacher-student network, it is related to industrial process control technical field.First, build the mechanism model of converter steelmaking thermal fluid dynamics including parameterized dynamic source term, including thermal fluid dynamics control equation, parameterized dynamic source term;Teacher network is constructed and trained to obtain fluid dynamics benchmark;Student network is then constructed, student network executes selective knowledge distillation mechanism to carry out fluid dynamics calculation and thermochemical reaction dynamics inversion;The student network is parallel with teacher network, and the velocity field and pressure field of knowledge distillation teacher network are distilled, and the final physical output of student network forward propagation is expanded to five-dimensional physical quantity;Finally, unknown thermal fluid reaction source term and student network weight are jointly optimized and parameter inversion.The method greatly improves the self-adaptive ability of model when facing the highly fluctuating reaction kinetics in actual industry.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to a multi-physics soft measurement method and system for converter steelmaking using a physically constrained teacher-student network. Background Technology

[0002] Converter steelmaking is the dominant process in global steel production and plays an irreplaceable role in the modern steel industry. However, due to the extremely harsh environment inside the converter molten pool, characterized by extremely high temperatures, high-pressure oxygen blowing, and intense dust, traditional contact sensors cannot operate stably for extended periods. Therefore, the converter smelting process is typically treated as a "black box."

[0003] Currently, monitoring and reconstructing the multiphysics state inside a converter faces three major technical bottlenecks: First, the chemical reaction kinetics inside the converter are highly volatile, with the actual heat release rate and decarburization rate changing drastically over time, which undermines the validity of the static boundary assumptions made in traditional mechanism-based fluid dynamics models. Second, converter smelting involves strong multiphysics coupling (fluid flow, heat transfer, component transport), and in deep learning inversion, this coupling inevitably leads to severe gradient ill-conditioning, causing the single physical information neural network to antagonize each other during optimization, easily getting trapped in local optima or even failing to converge. Finally, in actual industrial settings, only extremely sparse measurement data can be obtained through secondary guns. Directly applying this sparse data to conventional data-driven models leads to severe local overfitting, making it impossible to reconstruct the true global physical evolution process. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for soft measurement of multi-physics fields in converter steelmaking with physical constraints on teacher-student networks, thereby realizing soft measurement of multi-physics fields in converter steelmaking.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] On one hand, this invention provides a multi-physics soft measurement method for converter steelmaking using a physically constrained teacher-student network, comprising:

[0007] A thermodynamic mechanism model of converter steelmaking, including parameterized dynamic source terms, is constructed. This model includes thermodynamic control equations and parameterized dynamic source terms.

[0008] A teacher network is constructed and trained to obtain a fluid dynamics benchmark. The teacher network adopts a fully connected neural network architecture. The input is a spatiotemporal coordinate composed of continuous two-dimensional spatial coordinates and time coordinates. The macroscopic multiphysics field inside the converter is fitted by the forward propagation algorithm of the fully connected neural network. The fluid velocity field, pressure field and benchmark temperature field are linearly output in the last layer of the teacher network.

[0009] A student network is constructed, which performs fluid dynamics calculations and thermochemical reaction kinetic inversions using a selective knowledge distillation mechanism. The student network operates in parallel with the teacher network, distilling the velocity and pressure fields of the teacher network. The input is also spatiotemporal coordinates. The final physical output of the student network's forward propagation is expanded into a five-dimensional physical quantity. direction and The fluid velocity field, pressure field, temperature field, and carbon concentration field of molten steel in the direction of motion;

[0010] Joint optimization and parameter inversion are performed on the unknown thermal fluid reaction source term and student network weights.

[0011] Furthermore, the governing equations of the thermodynamics include the two-dimensional continuity equation of the converter, the momentum conservation equation, the energy conservation equation, and the carbon component transport equation; the parameterized dynamic source terms include the reaction heat source terms. and local carbon absorption terms As shown in the formula below:

[0012] ;

[0013] ;

[0014] in, Used to define the impact zone of the oxygen lance jet. The coordinates of the points inside the converter. For the concentration of carbon, To distinguish between gas and liquid phases, a smooth phase field indicator, and These are the exothermic reaction coefficient and decarbonization reaction rate parameters to be inverted, respectively.

[0015] Furthermore, the method introduces a fixed, dimensionless prior reference heat source term with known boundaries into the energy conservation equation used to constrain the teacher network; through this known heat source, local high-temperature expansion is induced, thereby deriving the complex convective circulation flow field of the bottom liquid molten pool driven by gravity and density difference based on the Boussinesq approximation.

[0016] Define the total loss function of the teacher network. It consists of four core weighted residual terms coupled together, used to constrain the teacher network parameters. Update direction:

[0017] ;

[0018] in, For the residuals of the governing equations; For boundary condition residuals; The initial condition residual; For measurement data residuals; These are the weights of each residual term;

[0019] By minimizing the total loss function of the teacher network To obtain a high-fidelity hydrodynamic field reference in steady state.

[0020] Furthermore, the calculation method for the four residual terms in the total loss function of the teacher network is as follows:

[0021] Using automatic differentiation techniques, the fluid velocity field output by the teacher network is calculated. Pressure field and reference temperature field Partial derivatives of various orders with respect to the input spatiotemporal coordinates;

[0022] The residuals of the governing equations are the sum of the mean square errors of the continuity equation, the momentum conservation equation, and the energy conservation equation at all fluid control points. ;

[0023] The boundary condition residual is constituted by the mean square error between the predicted physical quantities of the teacher network at the spatial geometric boundary and the actual values ​​of the set physical boundary conditions. ;

[0024] The initial condition residual is constituted by the mean square error between the predicted physical quantity output by the teacher network at the initial moment of calculation and the actual value of the set initial physical boundary conditions. ;

[0025] The measurement data residual is constituted by the weighted mean square error between the predicted physical quantities at the measurement points by the teacher network and the actual measured values ​​of the secondary gun. .

[0026] Furthermore, the method performs scale normalization on each physical variable when calculating the total loss function of the teacher network.

[0027] Furthermore, the carbon concentration output of the student network is designed with a physical hard constraint based on the Softplus activation function, the mathematical expression of which is as follows:

[0028] ;

[0029] In the formula, The carbon concentration field of molten steel output to students via network. The raw linear output real carbon concentration value of the last layer of the student network before it is processed by the activation function.

[0030] Furthermore, the specific method for the student network to perform the selective knowledge distillation mechanism is as follows:

[0031] The network weight parameters of the teacher network, which has been fully pre-trained to steady-state convergence, are frozen to serve as the baseline generator for the high-fidelity fluid dynamics framework, producing velocity, pressure, and temperature fields. Then, during forward propagation training, the student network is trained with a highly weighted mean square error distillation loss term. By replicating the fluid skeleton through knowledge distillation, the spatial distribution characteristics of the velocity and pressure fields of the teacher network are forcibly copied and fully inherited, while the reference temperature field output of the teacher network is completely shielded. This allows the student network to independently perform spatiotemporal multi-component reconstruction and thermodynamic evolution inversion, relying entirely on real industrial secondary gun sparse physical measurement data and the parameterized dynamic thermochemical source terms to be inverted.

[0032] ;

[0033] in, The total number of pseudo-random uniform control points collected throughout the continuous spatiotemporal fluid domain of the entire converter.

[0034] Furthermore, the specific method for jointly optimizing and inverting the parameters of the unknown thermofluid reaction source term and the student network weights is as follows:

[0035] In the process of supervised and unsupervised mixed training on student networks, the reaction exothermic coefficient, which represents the macroscopic unknown thermochemical reaction kinetics characteristics of the converter, is used. and decarbonization reaction rate parameters As a differentiable physical optimization variable, it is directly embedded into the entire backpropagation computation graph of the student network; in each iteration batch of forward and backpropagation, the connection weights of the student network itself are used as differentiable physical optimization variables. With these two physical inversion parameters By performing fully synchronous and parallel adaptive updates and joint tuning optimization, the problem of solving the macroscopic reaction dynamics in the converter is transformed into the problem of optimizing the weights in the neural network.

[0036] Construct the total loss function of the student network The high-weighted mean square error distillation loss term is deeply integrated with the strong anchoring term of multimodal extended industrial observation data:

[0037] ;

[0038] in, This represents the weight of the mean square error distillation loss term;

[0039] The total loss function of the inherited teacher network The structure contains measurement data residuals. It has been expanded and upgraded from only calculating temperature residuals to a multimodal anchoring term that simultaneously calculates the prediction errors of temperature physical quantities and carbon concentration component physical quantities;

[0040] During the joint optimization of student network weights, an asymmetric physical control weight allocation strategy is adopted: the weight of the mean square error distillation loss term is increased. Reduce the penalty weight priority of the residuals of the energy conservation equation and the carbon component transport equation.

[0041] On the other hand, the present invention also provides a multi-physics soft measurement system for converter steelmaking with physical constraints on teacher-student networks, including a mechanism model construction module, a teacher network benchmark generation module, a student network multi-component reconstruction module, and a joint optimization and parameter inversion module;

[0042] The mechanism model construction module is used to construct a thermodynamic mechanism model of converter steelmaking that includes parameterized dynamic source terms, including thermodynamic control equations and parameterized dynamic source terms;

[0043] The teacher network benchmark generation module is used to construct a fully connected teacher network to obtain the fluid dynamics field benchmark. The teacher network adopts a fully connected neural network architecture, and the input is a spatiotemporal coordinate composed of continuous two-dimensional spatial coordinates and time coordinates. The forward propagation algorithm of the fully connected neural network is used to fit the macroscopic multiphysics field inside the converter. The fluid velocity field, pressure field and benchmark temperature field are linearly output in the last layer of the teacher network.

[0044] The multi-component reconstruction module of the student network is used to construct the student network. The student network performs selective knowledge distillation mechanism to perform fluid dynamics calculations and thermochemical reaction kinetic inversion. The student network runs in parallel with the teacher network, and the velocity and pressure fields of the teacher network are distilled using spatiotemporal coordinates as input. The final physical output of the student network's forward propagation is expanded into a five-dimensional physical quantity. direction and The fluid velocity field, pressure field, temperature field, and carbon concentration field of molten steel in the direction of motion;

[0045] The joint optimization and parameter inversion module is used to perform joint optimization and parameter inversion on the unknown thermofluid reaction source term and the student network weights.

[0046] The beneficial effects of adopting the above technical solution are as follows: The multi-physics soft measurement method and system for converter steelmaking with physical constraint teacher-student network provided by the present invention (1) overcomes the limitations of traditional static assumptions: adopting a parameterized dynamic source term mechanism, it abandons the strong dependence on static boundaries and reaction rates in traditional numerical simulation. By transforming complex unknown chemical reactions into invertible parameterized learning problems, it greatly improves the model's adaptability when facing highly fluctuating reaction dynamics in actual industry.

[0047] (2) Solved the gradient ill-conditioning problem caused by multi-physics coupling: A teacher-student network architecture with physical field decoupling was designed. The teacher network focuses on building a high-fidelity fluid flow skeleton, while the student network independently performs component transport and thermodynamic parameter inversion on the basis of shielding the temperature field. This asynchronous physical-level decoupling strategy effectively isolates the mutual interference between the momentum equation and the energy and mass conservation equations in the backpropagation of the network, fundamentally alleviates the gradient ill-conditioning trap of individual physical neural network models that are prone to local optima, and significantly improves the stability and global convergence of joint optimization training.

[0048] (3) Breakthrough in soft measurement accuracy under extremely sparse industrial data: Facing the extreme sparsity of secondary gun point source sampling, a hybrid strategy of Gaussian spatiotemporal soft constraints and selective knowledge distillation is introduced into the multimodal data anchoring term to smoothly diffuse the influence weight of sparse data points on the global spatiotemporal evolution manifold. This mechanism effectively avoids the serious local overfitting phenomenon caused by point-to-point mean square error, and in the final prediction mapped to the actual industrial site, it achieves an extremely low relative prediction error for the critical local blowing endpoint temperature below 0.99%, which has strong industrial application value. Attached Figure Description

[0049] Figure 1 This is a flowchart of a multi-physics soft measurement method for converter steelmaking using a physically constrained teacher-student network, provided in Embodiment 1 of the present invention.

[0050] Figure 2 This is a schematic diagram of the modeling process and teacher-student network architecture provided in Embodiment 1 of the present invention;

[0051] Figure 3 This is a teacher network architecture diagram for establishing a fluid dynamics benchmark provided in Embodiment 1 of the present invention;

[0052] Figure 4 This is a student network architecture diagram provided in Embodiment 1 of the present invention, which employs selective knowledge distillation and joint inversion of unknown source terms.

[0053] Figure 5The diagram shows the convergence curves of the loss functions of the teacher network and the student network during the training process, as provided in Embodiment 1 of the present invention. (a) represents the total loss of the teacher network, and (b) represents the total loss of the student network.

[0054] Figure 6 This is a convergence trajectory diagram of the unknown thermochemical parameters provided in Embodiment 1 of the present invention during the joint inversion process, wherein (a) is the reaction exothermic coefficient. The convergence trajectory, (b) is the decarbonization reaction rate parameter. The convergence trajectory;

[0055] Figure 7 This is a comparison and verification diagram of the temperature field reconstructed based on multiphysics field and the actual industrial field measurement data provided in Embodiment 1 of the present invention. Detailed Implementation

[0056] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0057] Example 1:

[0058] Multiphysics soft measurement methods for converter steelmaking in physically constrained teacher-student networks, such as Figures 1-7 As shown, it includes the following steps:

[0059] S1. Construct a thermodynamic mechanism model of converter steelmaking that includes parameterized dynamic source terms, comprising thermodynamic governing equations and parameterized dynamic source terms; the governing equations include the converter two-dimensional continuity equation, momentum conservation equation, energy conservation equation, and carbon component transport equation; the parameterized dynamic source terms include reaction heat source terms. and local carbon absorption terms ;

[0060] S11. Assuming the converter body is geometrically symmetric, the three-dimensional calculation volume of the converter body is simplified to a two-dimensional region. The molten steel flow is considered an incompressible Newtonian fluid, and the temperature gradient is modeled as a thermal buoyancy force causing natural convection based on the Boussinesq approximation. On this basis, the two-dimensional continuity equation and momentum conservation equation for the converter are established:

[0061] (1);

[0062] (2);

[0063] (3);

[0064] Among them, formula (1) is the two-dimensional continuity equation, and formulas (2) and (3) are the momentum conservation equations. and They are respectively inside the converter direction and The fluid velocity component in the direction; This refers to the local static pressure of the fluid (i.e., molten steel) inside the converter; This refers to the fluid density inside the converter. t represents dynamic viscosity; t represents time. The thermal buoyancy term is shown in the following formula:

[0065] (4);

[0066] in, It is the acceleration due to gravity. The coefficient of thermal expansion is The initial temperature inside the converter. This is the current temperature inside the converter. A smooth phase field indicator to distinguish between the gas and liquid phases, used to ensure that thermal buoyancy plays a dominant role only in the liquid molten pool region.

[0067] S12. Construct coupled energy conservation equations to describe the spatiotemporal evolution of heat driven by convection and conduction, and construct carbon component transport equations to describe carbon concentration distribution:

[0068] (5);

[0069] (6);

[0070] Among them, formula (5) is the energy conservation equation, and formula (6) is the carbon component transport equation; and These represent the specific heat capacity and thermal conductivity of the fluid (molten steel) inside the converter, respectively. It is the heat source term for the reaction per unit volume; The concentration of carbon; The effective diffusion coefficient of carbon in molten steel; This is the local carbon absorption term caused by the decarbonization reaction.

[0071] S13. Due to the dramatic fluctuations in the actual heat release rate and decarbonization rate over time, the reaction heat source term in the coupled energy conservation equation will be affected. And the local carbon absorption term in the component transport equation By parameterizing the parameters, the assumptions made in traditional mechanistic models regarding the reaction exothermic rate and decarbonization rate as fixed constants or static boundary conditions are abandoned.

[0072] (7);

[0073] (8);

[0074] in, Used to define the oxygen lance jet impact zone (corresponding to dimensionless coordinates in this embodiment). and (area) and These are the exothermic reaction coefficient and decarbonization reaction rate parameters to be inverted, respectively.

[0075] By using parametric modeling, complex unknown chemical reaction kinetics are transformed into solving key constants. and The optimization problem.

[0076] S2. Construct and train the Teacher Network to obtain a high-fidelity fluid dynamics benchmark;

[0077] S21. Network Architecture and Output Definition of the Teacher Network: This step constructs a fully connected neural network architecture for the teacher network, whose input consists of spatiotemporal coordinates composed of continuous two-dimensional spatial coordinates and time coordinates. The macroscopic multiphysics field inside the converter is fitted using the forward propagation algorithm of the fully connected neural network, and the linear output is obtained in the last layer of the teacher network. direction and Fluid velocity field in the direction Pressure field and reference temperature field By utilizing automatic differentiation techniques, the fluid velocity field output above can be calculated with high precision without relying on traditional grid differencing. Pressure field and reference temperature field The partial derivatives of various orders of the input spatiotemporal coordinates (including the first-order time derivative, the first-order spatial convection term, the second-order spatial diffusion term, etc.) provide a data foundation for the accurate analysis of the residuals of the governing equations.

[0078] S22. Driving Mechanism of Fluid Dynamics Reference Flow Field: In order to enable the fully connected neural network to autonomously learn a real flow field skeleton that conforms to the conservation laws of fluid dynamics (such as natural convection circulation driven by thermal buoyancy) under unsupervised conditions without full-field flow velocity labels, a fixed, dimensionless prior reference heat source term with known boundaries is introduced into the energy conservation equation (i.e., Equation 5) used to constrain the teacher network (in this embodiment, a dimensionless heat source is set). By inducing local high-temperature expansion through the known heat source, the complex convective circulation flow field of the bottom liquid molten pool driven by gravity and density difference is derived based on the Boussinesq approximation, laying a high-fidelity fluid dynamics benchmark for the subsequent multi-component reaction dynamics inversion of the student network.

[0079] S23. Construction of the overall loss function of the teacher network: Define the overall loss function of the teacher network. It consists of four core weighted residual terms coupled together, used to constrain the teacher network parameters. Update direction:

[0080] (9);

[0081] in, The residuals of the governing equations (continuity equation, momentum conservation equation, energy conservation equation); For boundary condition residuals; The initial condition residual; For measurement data residuals; These are the weights for each residual term; to ensure consistency between fundamental physical conservation laws and actual industrial observations and boundary constraints, the priority of the weights is set as follows: By minimizing To obtain a high-fidelity hydrodynamic field reference in steady state.

[0082] (1) Residual of governing equation The equation is composed of the sum of the mean square errors of the continuity equation (Equation 1), the momentum conservation equation (Equations 2 and 3), and the energy conservation equation (Equation 5) at all fluid control points. Since the converter pool flow is modeled as an incompressible fluid and absolute pressure drift is not considered during the solution process, a penalty term for the local static pressure gradient is specifically introduced in the nominal region of non-violent reaction to avoid gradient explosion and stabilize the flow field streamlines. This penalty term is then incorporated into the residual of the governing equation. In order to enhance the numerical optimization stability of the loss function.

[0083] (2) Boundary condition residuals The residual term is composed of the mean square error between the predicted physical quantities of the teacher network at the spatial geometric boundary and the actual values ​​of the set physical boundary conditions. This residual term strictly constrains the physical legality at the spatial geometric boundary, including the forced top-blowing inlet velocity condition at the oxygen lance nozzle, the fluid viscosity no-slip boundary condition of the converter solid sidewall and bottom, the free atmospheric pressure outlet condition at the top of the molten pool, and the axisymmetric boundary condition at the central axis of the converter body.

[0084] (3) Initial condition residual The residual term is composed of the mean square error between the predicted physical quantity output by the teacher network at the initial moment of calculation and the actual value of the set initial physical boundary conditions; this residual term constrains the initial spatiotemporal distribution of temperature and the initial static state of the entire converter liquid region at the initial moment of calculation.

[0085] (4) Measurement data residuals The weighted mean square error is composed of the predicted physical quantity values ​​of the teacher network at the measurement point and the actual measured values ​​of the secondary lance. Due to the technical limitations in actual industrial steelmaking, where the secondary lance can only acquire extremely sparse TSC (thermometry for carbon determination) and TSO (thermometry for oxygen determination) point source sampling data at a very few specific time points, directly using conventional point-to-point mean square error penalties can easily lead to severe local overfitting and non-physical oscillations in the fully connected neural network at sparse points. Therefore, this invention introduces a Gaussian kernel function to apply spatiotemporal soft constraints to the sparse TSC and TSO measurement data samples, specifically using the mask weight coefficients calculated using the Gaussian kernel function. The mask weight coefficient is multiplied into the calculation as a weight for the weighted mean square error, and the formula for this weight coefficient is as follows:

[0086] (10);

[0087] In the formula, The spatiotemporal center coordinates of the physical data actually measured by the secondary gun sensor; and These represent the decay radii of the adaptive control data samples' outward radiation influence in the time and spatial dimensions, respectively. This spatiotemporal soft constraint strategy based on the Gaussian kernel function allows the influence of extremely sparse measurement data to diffuse smoothly and continuously into the adjacent spatiotemporal continuum, achieving regularization of the fluid topology manifold.

[0088] S24. Dimensionless and Normalization Specifications for Physical Variables: To eliminate the numerical gap between different physical quantities (such as furnace temperatures of several thousand degrees Celsius and converter geometry of several meters), and to avoid the inundation of small physical quantity gradients by large physical quantity gradients, this invention performs scale normalization on key physical variables before calculating each loss term: the time axis is normalized with an 800-second characteristic scale; the temperature axis is normalized with... The characteristic scale is normalized; the converter characteristic length is normalized based on 2 m. For example, when the secondary lance is at an actual physical time of 633 s and a temperature... Geometric coordinates When TSC data is measured, its dimensionless spatiotemporal and physical field data mapped to the teacher network are automatically converted into… , , , Similarly, when the actual physical time is 768 s and the temperature is... Geometric coordinates When the TSO data was measured, the converted dimensionless training data was: .

[0089] S3. Construct a student network and implement a selective knowledge distillation mechanism;

[0090] S31. Definition of Physical Output and Design of Non-negativity Hard Constraints for the Student Network: Construct an independent student network that runs parallel to the teacher network. The student network also takes spatiotemporal coordinates as input. Its forward propagation ultimately expands to a five-dimensional physical quantity, namely... direction and Fluid velocity field in the direction and Pressure field Temperature field and the carbon concentration field of molten steel In multi-component transport evolution, since carbon concentration is absolutely not allowed to be negative in the real physical world, in order to avoid the neural network generating non-physical negative solutions in complex unsupervised multi-objective optimization, this invention designs a physical hard constraint based on the Softplus activation function at the carbon concentration output end of the student network, the mathematical expression of which is as follows:

[0091] (11);

[0092] In the formula, The raw linear output real carbon concentration value of the last layer of the student network before it is processed by the activation function.

[0093] S32. Decoupling Principle of Selective Knowledge Distillation Mechanism: Since the calculation of fluid momentum transport (Navier-Stokes equations) within the converter is extremely time-consuming and accompanied by strong nonlinearity, directly constructing a single pure physical information neural network to simultaneously fit the flow field and invert unknown chemical reaction source terms would result in severe gradient ill-conditioning between multiple physical fields due to differences in physical magnitudes. This would cause the residual gradients of different equations to antagonize each other during backpropagation, leading to the technical disaster of local minima or model non-convergence. Therefore, this invention employs a physical-level asynchronous decoupling strategy to separate fluid dynamics calculations from thermochemical reaction dynamics inversion: First, the network weight parameters of the teacher network, which has been fully pre-trained to steady-state convergence, are frozen, serving as the benchmark source for generating velocity, pressure, and temperature fields within a high-fidelity fluid dynamics framework. Then, during forward propagation training, a high-weight mean square error distillation loss term is introduced into the student network. The fluid skeleton is replicated through knowledge distillation, that is, the spatial distribution characteristics of the velocity and pressure fields of the teacher network are forcibly replicated and fully inherited.

[0094] (12);

[0095] in, The total number of pseudo-random uniform control points collected in the continuous spatiotemporal fluid domain of the entire converter is defined as the spatiotemporal coordinate sampling points used to calculate the residuals of the control equations.

[0096] S33. Selective Shielding of Temperature Field and Independent Thermodynamic Evolution Reconstruction: In the process of replicating the fluid skeleton through knowledge distillation, this invention specifically designs the student network to completely shield the reference temperature field output of the teacher network. That is, in the above distillation loss item The mean square error of the predicted temperature field is not calculated. The core design reason is that the temperature field of the teacher network is generated by an artificially imagined fixed reference heat source, which cannot reflect the real transient exothermic laws driven by decarburization and violent redox reactions inside the converter. Through this selective shielding mechanism, the student network perfectly and inexpensively inherits the fluid flow and convection characteristics of the underlying layer, while completely releasing its own optimization degrees of freedom in the temperature field and carbon component transport field. This allows it to independently perform high-quality spatiotemporal multi-component reconstruction and thermodynamic evolution inversion, relying entirely on real industrial secondary gun sparse physical measurement data and the parameterized dynamic thermochemical source terms to be inverted.

[0097] S4. Perform joint optimization and parameter inversion of unknown thermal fluid reaction source terms and student network weights;

[0098] S41. Synchronous tuning of back propagation and unknown reaction kinetic parameters: During the supervised and unsupervised mixed training of the student network, the reaction exothermic coefficients representing the macroscopic unknown thermochemical reaction kinetics of the converter in equations (7) and (8) are used. and decarbonization reaction rate parameters As a differentiable physical optimization variable, it is directly embedded into the entire backpropagation computation graph of the student network. In each iteration batch of forward and backpropagation, the connection weights of the student network itself are used as differentiable physical optimization variables. With these two physical inversion parameters By performing fully synchronous and parallel adaptive updates and joint tuning optimization, the complex macroscopic reaction dynamics problem in the converter is transformed into the weight optimization problem within the neural network.

[0099] S42. Comprehensive Coupling Design of the Total Loss Function of the Student Network: Constructing the Total Loss Function of the Student Network It deeply integrates the heavily weighted mean square error distillation loss term with the strong anchoring term of multimodal extended industrial observation data:

[0100] (13);

[0101] in, This represents the weight of the mean square error distillation loss term;

[0102] The total loss function of the inherited teacher network The structure contains measurement data residuals. The method has been expanded and upgraded from calculating only the temperature residual to simultaneously calculating the prediction errors of temperature physical quantities and carbon concentration component physical quantities in a multimodal anchoring term. To ensure that the numerical sensitivity of the residuals in the carbon component transport equation (Equation 6) remains on the same order of magnitude as that in the energy conservation equation, this invention performs a characteristic dimensionless normalization on the overall carbon concentration: uniformly using the standard maximum mass fraction of converter steel, 4.5%, as the characteristic scaling value for the component. Through this specification, the initial carbon concentration of molten iron at the blowing stage, 4.26%, is within the total loss function... The TSC carbon concentration data of 0.169% measured by the industrial auxiliary gun is adaptively mapped to a dimensionless real value of 0.0376; the TSO carbon concentration data of 0.042% at the blowing endpoint is mapped to a dimensionless eigenvalue of 0.0093. The Gaussian mask weighting coefficients defined by the aforementioned equation (10) are used. (i.e., specifically used for calculating measurement data residuals) (Gaussian spatiotemporal local weights), multimodal extended industrial observation data are forcibly anchored to a continuous spatiotemporal distribution network inside the converter.

[0103] S43. Adaptive Weight Relaxation and Non-convex Loss Topography Locking Control Strategy: During the joint optimization of student network weights, this invention implements an asymmetric physical control weight configuration strategy: the weights of the mean square error distillation loss term are adjusted. By assigning extremely high orders of magnitude (e.g., more than two orders of magnitude higher than the equation residuals), strong physical regularization is formed, firmly locking the velocity and pressure fields of the student network during forward propagation, ensuring it never deviates from the high-fidelity fluid dynamics framework. Conversely, the penalty weights of the residuals of the energy conservation equation (Equation 5) and the carbon component transport equation (Equation 6) are appropriately relaxed and their priority is significantly increased. This asymmetric relaxation mechanism endows the student network with the flexibility to adaptively deform the temperature and carbon concentration fields in local space and reasonably deviate from the nominal fluid manifold. This allows the student network to spontaneously utilize the convection entrainment effect of the flow field when facing highly complex non-convex loss function terrain, bypassing various deep local minima optimization traps caused by the highly nonlinear coupling of multiphysics fields, and ultimately achieving the global optimal locking and accurate reconstruction of unknown macroscopic thermochemical reaction kinetic source term parameters.

[0104] To demonstrate the superior effectiveness of the proposed teacher-student network soft measurement method based on selective knowledge distillation of physical information, this embodiment presents a complete numerical simulation and parameter inversion reproduction of the dual-network architecture based on actual industrial field operation data. The specific algorithm configuration and technical disclosure are as follows:

[0105] Geometric and computational spatiotemporal domain partitioning specifications: In this embodiment, the converter body adopts a two-dimensional axisymmetric geometric space model for dimensionless simplification, and its continuous dimensionless computational spatiotemporal domain is defined as follows: The dimensionless time domain is set as follows. (Corresponding to an actual physical blowing time of 800 s in steel production). The computational domain of the internal multiphase flow of molten steel is defined as an inclined geometric region: The four physical cross-sectional boundaries of this geometric space. Defined respectively as: Top pressure free outlet boundary Side fireplace bottom boundary ; axisymmetric boundary of converter center axis ; and the boundary of the refractory brick wall of the inclined furnace lining. .

[0106] Meshless Random Point Cloud Uniform Sampling Strategy: Since the physical information neural network uses a meshless random point cloud mechanism instead of the traditional CFD differential grid, this embodiment employs a pseudo-random consistent uniform sampling strategy to generate control points in the entire spatiotemporal domain. Randomly scattered within the continuous flow domain of molten steel inside the converter... The residuals of the governing equations are used to calculate the collocation points, and the results are based on the mask conditions. Perform physical boundary clipping and filtering; at the geometric space boundary Upper pseudo-random uniform sampling Boundary constraint points are used to apply Dirichlet and Neumann boundary constraints, with 1,333 boundary points allocated to the top free exit, and 1,000 boundary points evenly allocated to the furnace bottom, central axis, and inclined sidewalls; initial spatiotemporal condition points are allocated in the initial time period. Independently generated internally. Extremely sparse secondary weapon industrial-scale anchor points are set at dimensionless time. and The location is centered at the dimensionless spatial physical coordinates within the two-dimensional axisymmetric geometric space model. .

[0107] Hardware and software environment and adaptive network training strategy: This algorithm is written based on the PyTorch deep learning engine and is deployed on an industrial computing hardware platform equipped with an Intel i7-11700K CPU, 32GB of RAM, and an NVIDIA RTX 3070 graphics card. Both network structures use the Adam gradient optimizer for parameter optimization, and the initial learning rates for the teacher and student networks are set to... In the forward propagation, a StepLR learning rate decay mechanism is adopted, which automatically decays the current learning rate by 50% every 5,000 training steps. The total number of training iterations is set to 50,000 to ensure that the multiphysics loss residuals converge completely and smoothly to the numerical baseline.

[0108] Inversion trajectory and steady-state locking behavior of unknown chemical reaction kinetic parameters: In an unsupervised joint inversion experiment, the reaction exothermic coefficient was... and decarbonization reaction rate parameters The initial guesses were intentionally incorrectly and deviated from the initial values, being initialized to 5.0 and 10.0, respectively. After 50,000 rounds of alternating backpropagation optimization, the two unknown physical parameters to be inverted perfectly adaptively converged to their true macroscopic physical values, i.e., the reaction heat release coefficient stably converged to... The decarbonization reaction rate parameters were precisely locked at The reaction exothermic parameter The convergence exhibits an extremely smooth, monotonically decreasing evolution trajectory; while the decarbonization reaction rate parameter The convergence trajectory exhibits a more complex, nonlinear, and nonconvex topographical feature with a more realistic industrial feel: in the first 20,000 training rounds, its value rapidly and adaptively decreases from an initial 10.0 to around 1.9. Subsequently, due to the momentum inertia of the Adam optimizer, a slight rebound and fine-tuning phase is triggered on the surface of the nonconvex loss function. Finally, it is precisely locked at 2.0144 with an adaptive nonlinear evolution trajectory of "rapid decline - slight pullback - robust locking".

[0109] Multiphysics Soft Measurement Accuracy and Practical Industrial Verification: To verify the effectiveness of the aforementioned inversion parameter locking technique, the converged student network was mapped back to the real physical space. In physical time... At the end point of the refining process (TSO node), the student network uses a virtual probe to locate the actual coordinates of the secondary weapon. The continuous spatial predicted temperature extracted at the location is The actual temperature of the molten steel in the industrial field, measured by sensors at the same location and time, was... Experimental results show that, under extremely high-temperature, opaque, and harsh industrial conditions, the absolute error of soft measurement of key local temperatures using this invention is only [missing information]. The relative prediction error is controlled within The following thoroughly demonstrates the extremely high accuracy of soft measurement, excellent industrial applicability, and strong technical superiority of the present invention from two aspects: continuous physical image of the entire field and the accuracy of key points.

[0110] Example 2:

[0111] This embodiment provides a multi-physics soft measurement system for converter steelmaking with a physically constrained teacher-student network, including a mechanism model construction module, a teacher network benchmark generation module, a student network multi-component reconstruction module, and a joint optimization and parameter inversion module.

[0112] The mechanism model construction module is used to construct a thermodynamic mechanism model of converter steelmaking that includes parameterized dynamic source terms, including thermodynamic control equations and parameterized dynamic source terms;

[0113] The teacher network benchmark generation module is used to construct a fully connected teacher network to obtain the fluid dynamics field benchmark. The teacher network adopts a fully connected neural network architecture, and the input is a spatiotemporal coordinate composed of continuous two-dimensional spatial coordinates and time coordinates. The forward propagation algorithm of the fully connected neural network is used to fit the macroscopic multiphysics field inside the converter. The fluid velocity field, pressure field and benchmark temperature field are linearly output in the last layer of the teacher network.

[0114] The multi-component reconstruction module of the student network is used to construct the student network. The student network performs selective knowledge distillation mechanism to perform fluid dynamics calculations and thermochemical reaction kinetic inversion. The student network runs in parallel with the teacher network, and the velocity and pressure fields of the teacher network are distilled using spatiotemporal coordinates as input. The final physical output of the student network's forward propagation is expanded into a five-dimensional physical quantity. direction and The fluid velocity field, pressure field, temperature field, and carbon concentration field of molten steel in the direction of motion;

[0115] The joint optimization and parameter inversion module is used to perform joint optimization and parameter inversion on the unknown thermofluid reaction source term and the student network weights.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. A multi-physics soft measurement method for converter steelmaking using a physically constrained teacher-student network, characterized in that, include: A thermodynamic mechanism model of converter steelmaking, including parameterized dynamic source terms, is constructed. This model includes thermodynamic control equations and parameterized dynamic source terms. A teacher network is constructed and trained to obtain a fluid dynamics benchmark. The teacher network adopts a fully connected neural network architecture. The input is a spatiotemporal coordinate composed of continuous two-dimensional spatial coordinates and time coordinates. The macroscopic multiphysics field inside the converter is fitted by the forward propagation algorithm of the fully connected neural network. The fluid velocity field, pressure field and benchmark temperature field are linearly output in the last layer of the teacher network. A student network is constructed, which performs fluid dynamics calculations and thermochemical reaction kinetic inversions using a selective knowledge distillation mechanism. The student network operates in parallel with the teacher network, distilling the velocity and pressure fields of the teacher network. The input is also spatiotemporal coordinates. The final physical output of the student network's forward propagation is expanded into a five-dimensional physical quantity. direction and The fluid velocity field, pressure field, temperature field, and carbon concentration field of molten steel in the direction of motion; Joint optimization and parameter inversion are performed on the unknown thermal fluid reaction source term and student network weights.

2. The multi-physics soft measurement method for converter steelmaking using a physically constrained teacher-student network according to claim 1, characterized in that, The governing equations of the thermodynamics include the two-dimensional continuity equation of the converter, the momentum conservation equation, the energy conservation equation, and the carbon component transport equation; the parameterized dynamic source terms include the reaction heat source terms. and local carbon absorption terms As shown in the formula below: ； ； in, Used to define the impact zone of the oxygen lance jet. The coordinates of the points inside the converter. For the concentration of carbon, To distinguish between gas and liquid phases, a smooth phase field indicator, and These are the exothermic reaction coefficient and decarbonization reaction rate parameters to be inverted, respectively.

3. The multi-physics soft measurement method for converter steelmaking using a physically constrained teacher-student network according to claim 2, characterized in that, The method introduces a fixed, dimensionless prior reference heat source term with known boundaries into the energy conservation equation used to constrain the teacher network; through this known heat source, local high-temperature expansion is induced, thereby deriving the complex convective circulation flow field of the bottom liquid molten pool driven by gravity and density difference based on the Boussinesq approximation. Define the total loss function of the teacher network. It consists of four core weighted residual terms coupled together, used to constrain the teacher network parameters. Update direction: ； in, For the residuals of the governing equations; For boundary condition residuals; The initial condition residual; For measurement data residuals; These are the weights of each residual term; By minimizing the total loss function of the teacher network To obtain a high-fidelity hydrodynamic field reference in steady state.

4. The multi-physics soft measurement method for converter steelmaking using a physically constrained teacher-student network according to claim 3, characterized in that, The calculation method for the four residual terms in the total loss function of the teacher network is as follows: Using automatic differentiation techniques, the fluid velocity field output by the teacher network is calculated. Pressure field and reference temperature field Partial derivatives of various orders with respect to the input spatiotemporal coordinates; The residuals of the governing equations are the sum of the mean square errors of the continuity equation, the momentum conservation equation, and the energy conservation equation at all fluid control points. ; The boundary condition residual is constituted by the mean square error between the predicted physical quantities of the teacher network at the spatial geometric boundary and the actual values ​​of the set physical boundary conditions. ; The initial condition residual is constituted by the mean square error between the predicted physical quantity output by the teacher network at the initial moment of calculation and the actual value of the set initial physical boundary conditions. ; The measurement data residual is constituted by the weighted mean square error between the predicted physical quantities at the measurement points by the teacher network and the actual measured values ​​of the secondary gun. .

5. The multi-physics soft measurement method for converter steelmaking using a physically constrained teacher-student network according to claim 4, characterized in that, The method performs scale normalization on each physical variable when calculating the total loss function of the teacher network.

6. The multi-physics soft measurement method for converter steelmaking using a physically constrained teacher-student network according to claim 5, characterized in that, The carbon concentration output of the student network is designed with a physical hard constraint based on the Softplus activation function, and the mathematical expression is as follows: ； In the formula, The carbon concentration field of molten steel output to students via network. The raw linear output real carbon concentration value of the last layer of the student network before it is processed by the activation function.

7. The method for multi-physics soft measurement of converter steelmaking in a physically constrained teacher-student network according to claim 6, characterized in that, The specific method by which the student network performs the selective knowledge distillation mechanism is as follows: The network weight parameters of the teacher network, which has been fully pre-trained to steady-state convergence, are frozen to serve as the baseline generator for the high-fidelity fluid dynamics framework, producing velocity, pressure, and temperature fields. Then, during forward propagation training, the student network is trained with a highly weighted mean square error distillation loss term. By replicating the fluid skeleton through knowledge distillation, the spatial distribution characteristics of the velocity and pressure fields of the teacher network are forcibly copied and fully inherited, while the reference temperature field output of the teacher network is completely shielded. This allows the student network to independently perform spatiotemporal multi-component reconstruction and thermodynamic evolution inversion, relying entirely on real industrial secondary gun sparse physical measurement data and the parameterized dynamic thermochemical source terms to be inverted. ； in, The total number of pseudo-random uniform control points collected throughout the continuous spatiotemporal fluid domain of the entire converter.

8. The method for multi-physics soft measurement of converter steelmaking using a physically constrained teacher-student network according to claim 7, characterized in that, The specific method for jointly optimizing and inverting parameters of the unknown thermal fluid reaction source term and student network weights is as follows: In the process of supervised and unsupervised mixed training on student networks, the reaction exothermic coefficient, which represents the macroscopic unknown thermochemical reaction kinetics characteristics of the converter, is used. and decarbonization reaction rate parameters As a differentiable physical optimization variable, it is directly embedded into the entire backpropagation computation graph of the student network; in each iteration batch of forward and backpropagation, the connection weights of the student network itself are used as differentiable physical optimization variables. With these two physical inversion parameters By performing fully synchronous and parallel adaptive updates and joint tuning optimization, the problem of solving the macroscopic reaction dynamics in the converter is transformed into the problem of optimizing the weights in the neural network. Construct the total loss function of the student network The high-weighted mean square error distillation loss term is deeply integrated with the strong anchoring term of multimodal extended industrial observation data: ； in, This represents the weight of the mean square error distillation loss term; The total loss function of the inherited teacher network The structure contains measurement data residuals. It has been expanded and upgraded from only calculating temperature residuals to a multimodal anchoring term that simultaneously calculates the prediction errors of temperature physical quantities and carbon concentration component physical quantities; During the joint optimization of student network weights, an asymmetric physical control weight allocation strategy is adopted: the weight of the mean square error distillation loss term is increased. Reduce the penalty weight priority of the residuals of the energy conservation equation and the carbon component transport equation.

9. A multiphysics soft measurement system for converter steelmaking using a physically constrained teacher-student network, implemented based on the multiphysics soft measurement method for converter steelmaking using a physically constrained teacher-student network as described in claim 1, characterized in that... It includes a mechanism model construction module, a teacher network benchmark generation module, a student network multi-component reconstruction module, and a joint optimization and parameter inversion module; The mechanism model construction module is used to construct a thermodynamic mechanism model of converter steelmaking that includes parameterized dynamic source terms, including thermodynamic control equations and parameterized dynamic source terms; The teacher network benchmark generation module is used to construct a fully connected teacher network to obtain the fluid dynamics field benchmark. The teacher network adopts a fully connected neural network architecture, and the input is a spatiotemporal coordinate composed of continuous two-dimensional spatial coordinates and time coordinates. The forward propagation algorithm of the fully connected neural network is used to fit the macroscopic multiphysics field inside the converter. The fluid velocity field, pressure field and benchmark temperature field are linearly output in the last layer of the teacher network. The multi-component reconstruction module of the student network is used to construct the student network. The student network performs selective knowledge distillation mechanism to perform fluid dynamics calculations and thermochemical reaction kinetic inversion. The student network runs in parallel with the teacher network, and the velocity and pressure fields of the teacher network are distilled using spatiotemporal coordinates as input. The final physical output of the student network's forward propagation is expanded into a five-dimensional physical quantity. direction and The fluid velocity field, pressure field, temperature field, and carbon concentration field of molten steel in the direction of motion; The joint optimization and parameter inversion module is used to perform joint optimization and parameter inversion on the unknown thermofluid reaction source term and the student network weights.

Images