An Optimization Method for Steel Control Cooling Process Based on Physical Information Neural Network

Abstract

This invention provides a method for optimizing the steel pipe cooling process based on a physical information neural network, relating to the fields of materials mechanics and engineering. The method includes: Step S1, data augmentation based on physical information: first, process parameters are input, and the temperature field is obtained using the finite difference method. Then, the phase composition of the steel pipe is obtained using the temperature field, and the macroscopic properties of the steel pipe are obtained from the phase composition and temperature field. Finally, flow field analysis and finite element simulation verification are completed. Step S2, prediction of the mechanical properties of the steel pipe: a model based on a physical information neural network is constructed to realize an integrated prediction-optimization framework from process parameters to final mechanical properties. Step S3, prediction of the mechanical properties of the steel pipe: using the physical information neural network model as the objective function, an improved Bayesian optimization algorithm is used to optimize the process parameters. This method can effectively achieve accurate prediction of the mechanical properties of steel pipes and synergistic optimization of their controlled cooling process parameters.

Description

Technical Field

[0001] This invention relates to the fields of materials science and engineering technology, specifically to an optimization method for steel control cooling process based on physical information neural networks. Background Technology

[0002] In the modern steel industry, controlled cooling, a heat treatment process, is a crucial step in producing high-performance steel. By precisely and rapidly cooling steel pipes after hot rolling, the internal microstructure of the material (such as grain size, the proportion and morphology of different phases like ferrite, bainite, and martensite) can be actively adjusted, thereby obtaining excellent mechanical properties such as high strength, high toughness, and good elongation. However, traditional optimization methods for controlled cooling processes mostly rely on trial and error based on experience and limited experiments. These methods depend heavily on the experience of engineers and field operators, making it difficult to efficiently handle the complex multivariate influences and coupling relationships between performance indicators in steel pipes (e.g., oil casing in oil and gas extraction processes). Furthermore, they suffer from limitations such as high cost, long cycle time, and difficulty in achieving global optimization.

[0003] With the rapid development of artificial intelligence, scholars at home and abroad have gradually applied artificial intelligence to optimize controlled cooling processes, replacing traditional trial-and-error and limited experiments. For example, in "Modeling the correlation between heat treatment, chemical composition and bainite fraction of pipeline steels by means of artificial neural networks" (Khalaj G et al. Neural Network World, 2013), a multilayer feedforward artificial neural network was used to map 16 chemical components and cooling rate parameters to bainite volume fraction, enabling prediction of the continuous cooling microstructure of pipeline steel and providing a basis for optimizing the heat treatment window. Similarly, in "Prediction of the deformation of aluminum alloy drill pipes in thermal assembly based on a BPneural network" (Wang X et al. Applied Sciences-Basel, 2013), the network was used to predict the microstructure of pipeline steel under continuous cooling and to provide a basis for optimizing the heat treatment window. In 2022, a BP neural network was used to construct a correlation model between cooling water flow rate, initial heating temperature, and hot deformation of steel joints, which accurately evaluated and optimized the interference of hot fitting of aluminum alloy drill pipes; the influence of ultra-fast cooling process on the microstructure and properties of DP780 hot-rolled strip steel [J] (Sun Lirong et al., Rolling Steel, 2024) used extreme learning machine to establish an ultra-fast cooling heat transfer coefficient model, which significantly improved the control accuracy of the pre-ultra-fast cooling outlet and coiling temperature of DP780 hot-rolled strip steel, and finally obtained low-cost high-performance steel. However, the existing technology still lacks a method for accurately predicting the mechanical properties of steel pipes, especially oil casings, and coordinating the optimization of their controlled cooling process parameters; since the convective heat transfer coefficient of steel pipes during water spray cooling is a nonlinear parameter that changes drastically with temperature and spraying conditions, simple physical models are difficult to capture accurately, leading to biased prediction results; at the same time, traditional machine learning models are constrained by data sparsity and "black box" problems, resulting in "small sample" characteristics of available data, poor generalization ability, and inability to quantify uncertain data, further increasing prediction errors, leading to bias and uncertainty in the optimization of controlled cooling process. Summary of the Invention

[0004] To address the problems existing in the prior art, the present invention aims to provide a method for optimizing the cooling process of steel pipes based on physical information neural networks. This method can utilize physical laws to guide learning and reduce reliance on data, while also quantifying the uncertainty of prediction results, thereby effectively achieving accurate prediction of the mechanical properties of steel pipes and synergistic optimization of their cooling process parameters.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A method for optimizing steel cooling system based on physical information neural network, comprising:

[0007] Step S1: Data augmentation based on physical information: First, input the process parameters, use the finite difference method to obtain the temperature field, then use the temperature field to obtain the phase composition of the steel pipe, and obtain the macroscopic properties of the steel pipe from the phase composition and temperature field. Finally, complete the flow field analysis and finite element simulation verification.

[0008] Step S2, Prediction of mechanical properties of steel pipes: Construct a model based on physical information neural network to realize an integrated prediction-optimization framework from process parameters to final mechanical properties;

[0009] Step S3, Steel Control Cooling Parameter Optimization: Using a physical information-based neural network model as the objective function, an improved Bayesian optimization algorithm is employed to optimize the process parameters.

[0010] Based on further optimization of the above scheme, step S1, "inputting process parameters and obtaining the temperature field using the finite difference method," specifically involves: inputting a process parameter vector, including the heating temperature. T heat Cooling temperature T start Final cooling temperature T end Cooling rate V cool With spray pressure P sp ;

[0011] For each set of input process parameters, the finite difference method is used for numerical solution: First, the geometric model of the steel pipe (axial direction z, radial direction r) and the time of the entire cooling process are discretized to form a spatiotemporal network; then, the input process parameters are used as the initial and boundary conditions of this physical scenario to solve the unsteady heat conduction equation:

[0012]

[0013] In the formula: T(r,z,t) express t At any given time, radial position r and axial position z Temperature at that location; This indicates the density of steel as a function of temperature. This indicates the specific heat capacity of steel as a function of temperature. This indicates the thermal conductivity of steel as a function of temperature.

[0014] The final output is a complete four-dimensional spatiotemporal temperature field data matrix. T(r,z,t) .

[0015] Based on further optimization of the above scheme, the step S1 of "obtaining the phase composition of the steel pipe using the temperature field" specifically involves: based on the spatiotemporal temperature field data matrix... T(r,z,t) Select a point at half the pipe wall thickness and extract its complete temperature-time curve. T(t) ;

[0016] Through the curve T(t) The ferrite and pearlite phase transformations during the continuous cooling process of the steel pipe were obtained separately:

[0017] Ferrite phase transformation: along the temperature path T(t) Data integration was performed on the differential equation of ferrite phase transformation kinetics within the temperature range of austenite-to-ferrite transformation:

[0018]

[0019] In the formula: X a This represents the volume fraction of ferrite at the current moment; n a The Avramie index represents the ferrite phase transition and reflects its nucleation and growth mechanism; k a (T) This represents the rate constant of the ferrite phase transformation, which is dependent on temperature. T Highly sensitive;

[0020] After integration, the final ferrite volume fraction is obtained:

[0021]

[0022] Pearlite phase transition:

[0023] The remaining austenite, i.e., the volume fraction is 1- V a The pearlite phase transition is calculated by numerically integrating the differential equation of pearlite phase transition kinetics within the temperature range of the pearlite transition:

[0024]

[0025] In the formula: X p This indicates the volume fraction that has been converted to pearlite, relative to the remaining austenite; n p The Avramie index represents the phase transition of pearlite; k p (T) The rate constant representing the pearlite phase transition;

[0026] After integration, the final transformation fraction of pearlite in the remaining austenite is obtained. X p-final Finally, the pearlite volume fraction was obtained:

[0027]

[0028] The final phase composition includes the volume fraction of ferrite. V a With pearlite volume fraction V p .

[0029] Based on further optimization of the above scheme, the step S1 of "obtaining the macroscopic properties of the steel pipe from the phase composition and temperature field" specifically refers to:

[0030] First, based on the heating temperature T heat Obtain the initial austenite grain size d y And from the temperature change curve over time T (t) Average cooling rate for extracting the austenite-ferrite phase transformation region This yields the final average ferrite grain size. d a :

[0031]

[0032] In the formula: This represents an empirical constant related to the composition of steel;

[0033] Then, the yield strength of the ferrite phase is obtained. (It is determined by the grain size):

[0034]

[0035] In the formula: This represents the frictional stress of the material; k y This represents the Hall-Petch coefficient, which measures the grain boundary strengthening effect;

[0036] The yield strength of the pearlite phase itself (It is determined by the interlamellar spacing):

[0037]

[0038] In the formula: S This represents the interlamellar spacing of pearlite, which is the average distance between the centerlines of two adjacent cementite lamellae in pearlite; This represents the frictional stress of the pearlite phase; k p The coefficients representing the pearlite-related properties are used to measure the strengthening effect of the lamellar interface.

[0039] Finally, the overall macroscopic yield strength of the material is obtained. :

[0040]

[0041] To obtain the overall tensile strength (UTS) of the material:

[0042]

[0043] In the formula: YR represents the yield strength ratio (obtained through a large amount of experimental data);

[0044] Elongation A:

[0045]

[0046] In the formula: C elo This represents the empirical conversion factor from uniform elongation to total elongation. N a The inherent strain hardening index of the ferrite phase; N p This represents the inherent strain hardening index of the pearlite phase.

[0047] Based on further optimization of the above scheme, the step S1 of "completing flow field analysis and finite element simulation verification" specifically refers to:

[0048] Flow field analysis: Computational fluid dynamics (CFD) was used to simulate the spray cooling process on the surface of the steel pipe in detail; by analyzing the nozzle jet morphology, water flow distribution and boiling heat transfer behavior, the convective heat transfer coefficient distribution was obtained.

[0049] Finite element simulation verification: Several sets of process parameters are randomly selected from the generated large-scale dataset. The finite element software ABAQUS is used to perform a complete multiphysics simulation of several sets of parameters coupled with thermo-mechanical-phase transition. These high-precision finite element simulation results (such as temperature field, microstructure distribution, stress and strain) are used as the calculation benchmark and compared with the output of the data augmentation algorithm in step S1 to evaluate the overall prediction accuracy and perform necessary model parameter calibration.

[0050] Based on further optimization of the above scheme, the model based on the physical information neural network in step S2 is as follows:

[0051]

[0052] In the formula: This represents the optimal set of neural network parameters obtained by the model.

[0053] This represents a single neural network with multiple outputs, whose input is in the spatiotemporal domain:

[0054]

[0055] Where: radial coordinate domain r Limited to the thickness range of the steel pipe r 1 and r Within 2; Axial coordinate domain z Limited to the analysis length of the steel pipe l Within; the time coordinate domain is limited to the total cooldown time. t final Within:

[0056]

[0057] This represents a set of physical constraints, including initial condition constraints, boundary condition constraints, and grain size model constraints. It realizes the transformation of a set of macroscopic process parameters into a specific set of physical constraints.

[0058]

[0059] In the formula: F IC , F BC , F GS These represent initial condition constraints, boundary condition constraints, and grain size model constraints, respectively.

[0060] L total This represents the total loss function, which accounts for the physical losses of heat conduction. L PDE Initial condition loss L IC Boundary condition loss L BC Phase transition kinetic loss L JMAK Grain size model loss L GS Loss in the relationship between material yield strength and grain size L HP Loss due to lack of authenticity of real data L data The weighted fusion method yields:

[0061]

[0062] In the formula: w 1. w 2. w 3. w 4. w 5. w 6. w 7 represents the corresponding loss weight value;

[0063] For any given set of process parameters X proc Its corresponding macroscopic mechanical properties Y for:

[0064]

[0065] That is, the final solution time is determined based on the input process parameters. t final Then, select a point in the spatiotemporal domain at half the pipe wall thickness, and call the function with optimal parameters. The neural network performs a complete forward computation and extracts the result at the final time step. t final The corresponding output, thus obtaining Performance predictions for UTS and A.

[0066] Based on further optimization of the above scheme, the initial condition constraints F IC (The initial temperature state within the entire solution domain at t=0 is defined.) Specifically:

[0067]

[0068] In the formula: This indicates that the physical information-based neural network model is in t Predicted temperature value at time =0;

[0069] Boundary condition constraints F BC (It is defined that the law of energy exchange must be satisfied on the outer surface of the steel pipe.) Specifically:

[0070]

[0071] In the formula: k This indicates the thermal conductivity of the steel pipe material; T NN This represents the temperature prediction value based on a physical information neural network model. T am Indicates ambient temperature; Indicates the emissivity of the steel pipe surface; This represents the Stefan-Boltzmann constant;

[0072] h This represents the convective heat transfer coefficient, which is determined by the spray pressure. P sp get:

[0073]

[0074] In the formula: q w Indicates water flow density; q 1. q 2. q 3 represents the water flow density threshold;

[0075]

[0076] Water flow density is determined by spray pressure. P sp With the surface area of ​​steel pipe A sp get:

[0077]

[0078] In the formula: K Indicates the proportionality factor (including nozzle characteristics, quantity, unit conversion, etc.); N This represents the exponent (usually 0.45 to 0.55);

[0079] Grain size model constraints F GS (A semi-empirical relationship is defined to satisfy the final ferrite grain size.) Specifically:

[0080] .

[0081] Based on further optimization of the above scheme, the physical loss of heat conduction is reduced. L PDE :

[0082]

[0083] Initial condition loss L IC :

[0084]

[0085] Boundary condition loss L BC :

[0086]

[0087] Phase transition kinetic loss LJMAK :

[0088]

[0089] Particle size model loss L GS :

[0090]

[0091] Loss of relationship between material yield strength and grain size L HP :

[0092]

[0093] Loss of Authentic Data L data :

[0094]

[0095] In the formula: Y pre Indicates the model's predictive performance. Y true This indicates the actual performance.

[0096] Based on further optimization of the above scheme, the improved Bayesian optimization algorithm in step S3, building upon the existing Bayesian optimization algorithm which uses a Gaussian process as a surrogate model, introduces a variational autoencoder as the core component of the surrogate model. This injects rich physical prior knowledge into the optimization process. The variational autoencoder includes an encoder. and decoder The encoder consists of two parts: it receives process parameters and encodes them into parameters of a probability distribution; the encoder outputs a mean vector. And the logarithmic variance vector log( Then, a latent vector Z is sampled from the distribution defined at the beginning of the encoding:

[0097]

[0098] In the formula: Indicates standard normal noise; Represents element-wise product;

[0099] Decoder receives latent vector Z And decode and reconstruct it into the corresponding mechanical properties. Y ;

[0100] The prior mean function of the variational autoencoder is:

[0101]

[0102] For a given process parameter x via encoder Obtain its deterministic representation in the latent space, and then pass it through the decoder. Reconstruct it into performance predictions;

[0103] The improved Bayesian optimization algorithm and hybrid surrogate model for new candidate points The posterior mean and variance are:

[0104]

[0105] In the formula: Indicates a new point The covariance vector between X and all previous points; This represents the covariance matrix among all evaluated points; Represents the covariance kernel function; Indicates the noise variance; I Represents the identity matrix.

[0106] The following are the technical effects of the present invention:

[0107] This invention first employs a data augmentation algorithm based on a physical model. Taking a wide range of process parameters as input, it uses the finite difference method to solve thermodynamic equations to obtain a high-precision temperature field. Then, it sequentially derives the material's phase composition, grain size, and final macroscopic mechanical properties, generating a large-scale virtual dataset. The accuracy of this large-scale virtual dataset is cross-validated through flow field analysis and finite element simulation to ensure data precision and avoid error accumulation. Based on this, an innovative physical information neural network is constructed as the core prediction model. This model receives the physical coordinates of the steel pipe as input and transforms specific process parameters into a "physical constraint set" containing initial conditions, boundary conditions, and an intrinsic physical model. By minimizing a composite loss function coupled with multi-physics laws, it achieves end-to-end, high-fidelity prediction from process parameters to macroscopic performance. Finally, in the optimization stage, the trained physical information neural network model is used as the low-cost, high-fidelity objective function of Bayesian optimization, and a variational autoencoder is used as the physical prior of the Gaussian process surrogate model. By learning the residual between the variational autoencoder prediction and the target value, an improved Bayesian optimization algorithm is constructed, which greatly accelerates the convergence speed of finding the optimal process combination in a complex parameter space. Attached Figure Description

[0108] Figure 1 This is an overall flowchart of the steel pipe cooling process optimization method based on physical information neural network in this embodiment of the invention.

[0109] Figure 2 This is a schematic diagram of finite element simulation of steel pipe cooling in an embodiment of the present invention.

[0110] Figure 3 This is the optimization result of the steel pipe cooling process in the embodiments of the present invention; wherein, Figure 3 (a) Optimize the Pareto front for the three objectives of yield strength, tensile strength, and elongation. Figure 3 (b) shows the heating temperature distribution. Figure 3 (c) shows the temperature distribution at the start of cooling. Figure 3 (d) shows the final cooling temperature distribution. Figure 3 (e) shows the spray pressure distribution. Figure 3 (f) represents the cooling rate distribution.

[0111] Figure 4 The stress-strain curves of four sets of trial steel pipes with controlled cooling parameters obtained by the optimization algorithm in the embodiments of the present invention are shown. Detailed Implementation

[0112] The technical solutions in the embodiments of the present invention will be clearly and completely described below. In the following description, specific details such as specific system structures and technologies are presented for illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention.

[0113] Example 1:

[0114] A method for optimizing steel cooling system based on physical information neural network, comprising:

[0115] Step S1: Data augmentation based on physical information: First, input the process parameters and obtain the temperature field using the finite difference method. Specifically, input the process parameter vector, including the heating temperature. T heat Cooling temperature T start Final cooling temperature T end Cooling rate V cool With spray pressure P sp (In this embodiment, T heat 950~1000℃; T start 800~850℃; T end 25~650℃; V cool : 5~98.2℃ / s; P sp (0~400kPa)

[0116] For each set of input process parameters, the finite difference method is used for numerical solution: First, the geometric model of the steel pipe (axial direction z, radial direction r) and the time of the entire cooling process are discretized to form a spatiotemporal network; then, the input process parameters are used as the initial and boundary conditions of this physical scenario (where the convective heat transfer coefficient h of the boundary adjustment is determined by the spray pressure). P sp (Convert the equations) and solve the unsteady heat conduction equations:

[0117]

[0118] In the formula: T(r,z,t) express t At any given time, radial position r and axial position z Temperature at that location; This indicates the density of steel as a function of temperature. This indicates the specific heat capacity of steel as a function of temperature. This indicates the thermal conductivity of steel as a function of temperature.

[0119] The final output is a complete four-dimensional spatiotemporal temperature field data matrix. T(r,z,t) .

[0120] The phase composition of the steel pipe is then obtained using the temperature field, specifically: based on the spatiotemporal temperature field data matrix. T(r,z,t) Select a point at half the pipe wall thickness and extract its complete temperature-time curve. T(t) ;

[0121] Through the curve T(t) The ferrite and pearlite phase transformations during the continuous cooling process of the steel pipe were obtained separately:

[0122] Ferrite phase transformation: along the temperature path T(t) Data integration was performed on the differential equation of ferrite phase transformation kinetics within the temperature range of austenite-to-ferrite transformation:

[0123]

[0124] In the formula: X a This represents the volume fraction of ferrite at the current moment; n a The Avramie index represents the ferrite phase transition and reflects its nucleation and growth mechanism; k a (T) This represents the rate constant of the ferrite phase transformation, which is dependent on temperature. T Highly sensitive;

[0125] After integration, the final ferrite volume fraction is obtained:

[0126]

[0127] Pearlite phase transition:

[0128] The remaining austenite, i.e., the volume fraction is 1- V a The pearlite phase transition is calculated by numerically integrating the differential equation of pearlite phase transition kinetics within the temperature range of the pearlite transition:

[0129]

[0130] In the formula: X p This indicates the volume fraction that has been converted to pearlite, relative to the remaining austenite; n p The Avramie index represents the phase transition of pearlite; k p (T) The rate constant representing the pearlite phase transition;

[0131] After integration, the final transformation fraction of pearlite in the remaining austenite is obtained. X p-final Finally, the pearlite volume fraction was obtained:

[0132]

[0133] The final phase composition includes the volume fraction of ferrite. V a With pearlite volume fraction V p .

[0134] The macroscopic properties of the steel pipe are obtained from the phase composition and temperature field, specifically:

[0135] First, based on the heating temperature T heat Obtain the initial austenite grain size d y And from the temperature change curve over time T (t) Average cooling rate for extracting the austenite-ferrite phase transformation region This yields the final average ferrite grain size. d a :

[0136]

[0137] In the formula: Represents an empirical constant related to the composition of steel (where, );

[0138] Then, the yield strength of the ferrite phase is obtained. (It is determined by the grain size):

[0139]

[0140] In the formula: This represents the frictional stress of the material; k y This represents the Hall-Petch coefficient, which measures the grain boundary strengthening effect;

[0141] The yield strength of the pearlite phase itself (It is determined by the interlamellar spacing):

[0142]

[0143] In the formula: S This represents the interlamellar spacing of pearlite, which is the average distance between the centerlines of two adjacent cementite lamellae in pearlite; This represents the frictional stress of the pearlite phase; k p The coefficients representing the pearlite-related properties are used to measure the strengthening effect of the lamellar interface.

[0144] Finally, the overall macroscopic yield strength of the material is obtained. :

[0145]

[0146] To obtain the overall tensile strength (UTS) of the material:

[0147]

[0148] In the formula: YR represents the yield strength ratio (obtained through a large amount of experimental data);

[0149] Elongation A:

[0150]

[0151] In the formula: C elo This represents the empirical conversion factor from uniform elongation to total elongation (in this embodiment, ...). C elo (1.15). N a This indicates the inherent strain hardening index of the ferrite phase (for J55 carbon steel). N a (0.25) N p This represents the inherent strain hardening index of the pearlite phase (for J55 carbon steel).N p (0.15).

[0152] Finally, flow field analysis and finite element simulation verification were completed, specifically as follows:

[0153] Flow field analysis: Computational fluid dynamics (CFD) is used to simulate the spray cooling process on the surface of the steel pipe in detail; by analyzing the nozzle jet morphology, water flow distribution and boiling heat transfer behavior, the convective heat transfer coefficient distribution is obtained; the accuracy of the heat transfer model is verified or corrected through flow field analysis.

[0154] Finite element simulation verification: From the generated large-scale dataset, several sets of process parameters are randomly selected, and the finite element software ABAQUS is used to perform a complete multiphysics simulation of several sets of parameters coupled with thermo-mechanical-phase transition (e.g., Figure 2 (as shown); and these high-precision finite element simulation results (such as temperature field, microstructure distribution, stress and strain, etc.) are used as the calculation benchmark and compared with the output of the data augmentation algorithm in step S1 to evaluate the overall prediction accuracy and perform necessary model parameter calibration.

[0155] Step S2, Prediction of mechanical properties of steel pipes: Construct a model based on physical information neural network to realize an integrated prediction-optimization framework from process parameters to final mechanical properties;

[0156] The model based on the physical information neural network is as follows:

[0157]

[0158] In the formula: This represents the optimal set of neural network parameters obtained by the model.

[0159] This represents a single neural network with multiple outputs, whose input is in the spatiotemporal domain:

[0160]

[0161] Where: radial coordinate domain r Limited to the thickness range of the steel pipe r 1 and r Within 2 (in this embodiment, r 1 = 62.13 mm r 2=69.85mm); Axial coordinate domain z Limited to the analysis length of the steel pipe l Within (in this embodiment, l =600mm); the time coordinate domain is limited to the total cooling time. t final Within:

[0162]

[0163] This represents a set of physical constraints, including initial condition constraints, boundary condition constraints, and grain size model constraints. It realizes the transformation of a set of macroscopic process parameters into a specific set of physical constraints.

[0164]

[0165] In the formula: F IC , F BC , F GS These represent initial condition constraints, boundary condition constraints, and grain size model constraints, respectively.

[0166] Initial constraints F IC (The initial temperature state within the entire solution domain at t=0 is defined.) Specifically:

[0167]

[0168] In the formula: This indicates that the physical information-based neural network model is in t Predicted temperature value at time =0;

[0169] Boundary condition constraints F BC (It is defined that the law of energy exchange must be satisfied on the outer surface of the steel pipe.) Specifically:

[0170]

[0171] In the formula: k This indicates the thermal conductivity of the steel pipe material; T NN This represents the temperature prediction value based on a physical information neural network model. T am Indicates ambient temperature; Indicates the emissivity of the steel pipe surface; This represents the Stefan-Boltzmann constant;

[0172] h This represents the convective heat transfer coefficient, which is determined by the spray pressure. P sp get:

[0173]

[0174] In the formula: q w Indicates water flow density;q 1. q 2. q 3 represents the water flow density threshold ( q 1=1361、 q 2=6649、 q 3=8010.5);

[0175]

[0176] Water flow density is determined by spray pressure. P sp With the surface area of ​​steel pipe A sp get:

[0177]

[0178] In the formula: K Indicates the proportionality factor (including nozzle characteristics, quantity, unit conversion, etc.); N This represents the exponent (usually 0.45 to 0.55);

[0179] Grain size model constraints F GS (A semi-empirical relationship is defined to satisfy the final ferrite grain size.) Specifically:

[0180] .

[0181] L total This represents the total loss function, which accounts for the physical losses of heat conduction. L PDE Initial condition loss L IC Boundary condition loss L BC Phase transition kinetic loss L JMAK Grain size model loss L GS Loss in the relationship between material yield strength and grain size L HP Loss due to lack of authenticity of real data L data Weighted fusion is performed to obtain (the total loss function receives the output of the neural network and is based on the set of physical constraints). The provided parameters are used to calculate the total deviation between the network predictions and all physical laws and real data.

[0182]

[0183] In the formula: w 1. w 2.w 3. w 4. w 5. w 6. w 7 represents the corresponding loss weight value;

[0184] Physical loss due to heat conduction L PDE :

[0185]

[0186] Initial condition loss L IC :

[0187]

[0188] Boundary condition loss L BC :

[0189]

[0190] Phase transition kinetic loss L JMAK :

[0191]

[0192] Particle size model loss L GS :

[0193]

[0194] Loss of relationship between material yield strength and grain size L HP :

[0195]

[0196] Loss of Authentic Data L data :

[0197]

[0198] In the formula: Y pre Indicates the model's predictive performance. Y true This indicates the actual performance.

[0199] For any given set of process parameters X proc Its corresponding macroscopic mechanical properties Y for:

[0200]

[0201] That is, the final solution time is determined based on the input process parameters. t final Then, select a point in the spatiotemporal domain at half the pipe wall thickness, and call the function with optimal parameters. The neural network performs a complete forward computation and extracts the result at the final time step. t final The corresponding output, thus obtaining 、 Performance predictions for UTS and A.

[0202] Step S3, Steel Control Cooling Parameter Optimization: Using a physical information-based neural network model as the objective function, an improved Bayesian optimization algorithm is employed to optimize the process parameters;

[0203] Assumption x ={ T heat , T start , T end , V cool , P sp If} represents a set of process parameters, then the optimization objective is:

[0204]

[0205] In the formula: NN PINN (x) This indicates the entire process of calling a physical information-based neural network model, that is, based on process parameters. x Constructing a set of physical constraints Then utilize the optimal parameters trained in step S2. Perform a forward propagation to obtain the final performance prediction.

[0206] The improved Bayesian optimization algorithm, based on the existing Bayesian optimization algorithm which uses a Gaussian process as a surrogate model, introduces a variational autoencoder as the core component of the surrogate model. This injects rich physical prior knowledge into the optimization process. The variational autoencoder includes an encoder... and decoder The encoder consists of two parts: it receives process parameters and encodes them into parameters of a probability distribution; the encoder outputs a mean vector. And the logarithmic variance vector log( Then, a latent vector Z is sampled from the distribution defined at the beginning of the encoding:

[0207]

[0208] In the formula: Indicates standard normal noise; Represents element-wise product;

[0209] Decoder receives latent vector Z And decode and reconstruct it into the corresponding mechanical properties. Y ;

[0210] The prior mean function of the variational autoencoder is:

[0211]

[0212] For a given process parameter x via encoder Obtain its deterministic representation in the latent space, and then pass it through the decoder. Reconstruct it into performance predictions;

[0213] The improved Bayesian optimization algorithm and hybrid surrogate model for new candidate points The posterior mean and variance are:

[0214]

[0215] In the formula: Indicates a new point The covariance vector between X and all previous points; This represents the covariance matrix among all evaluated points; Represents the covariance kernel function; Indicates the noise variance; I Represents the identity matrix.

[0216] Finally, the performance of the prototype steel pipe was verified by uniaxial tensile testing.

[0217] Example 2:

[0218] As another preferred embodiment of the technical solution of the present invention, based on the above embodiment 1, in the improved Bayesian optimization algorithm in step S3, for multi-objective optimization scenarios, the posterior distribution provided by the above hybrid model is used, combined with the "Expected Hypervolume Boost (EHVI)" acquisition function, so as to intelligently select the next most valuable process parameter point for evaluation, and finally find a set of Pareto optimal solutions.

Claims

1. A method for optimizing a controlled cooling process of a steel pipe based on a physical information neural network, characterized in that: Comprise: Step S1, data enhancement based on physical information: first input process parameters, obtain temperature field by finite difference method, then obtain phase composition of steel pipe by temperature field, and obtain macro performance of steel pipe by phase composition and temperature field, finally complete flow field analysis and finite element simulation verification; "Obtaining the phase composition of steel pipes using temperature fields" specifically refers to: based on the spatiotemporal temperature field data matrix. T(r,z,t) Select a point at half the pipe wall thickness and extract its complete temperature-time curve. T(t) ; By curve T(t) The ferrite transformation and the pearlite transformation in the continuous cooling process of the steel pipe were obtained respectively: Ferrite phase transformation: along the temperature path T(t) Within the temperature interval of the austenite to ferrite transformation, the data are integrated for the ferrite phase transformation kinetics differential equation: wherein: X a represents the volume fraction of ferrite at the current time instant; n a represents the Avrami exponent of the ferrite phase transformation, reflecting its nucleation and growth mechanism; k a (T) represents the rate constant of the ferrite phase transformation, which is highly sensitive to temperature T . After integration, the final ferrite volume fraction is obtained: Pearlite phase transformation: in the remaining austenite, i.e. with a volume fraction of 1- V a The pearlite phase transformation was calculated in the remaining austenite, i.e. with a volume fraction of 1- wherein: X p represents the volume fraction of the transformed to pearlite, relative to the remaining austenite; n p represents the Avrami exponent of the pearlite phase transformation; k p (T) represents the rate constant of the pearlite phase transformation; After the integration, the final fraction of pearlite transformation in the remaining austenite is obtained X p-final and finally the volume fraction of pearlite is obtained Final phase composition includes ferrite volume fraction V a Pearlite volume fraction V p ; "Obtaining macro performance of steel pipe by phase composition and temperature field" is specifically: First, the heating temperature T heat The initial austenite grain size d y is obtained from the temperature-time curve T(t) and the average cooling rate in the austenite to ferrite phase transformation interval is extracted , thus obtaining the final ferrite average grain size d a : In the formula: represents an empirical constant related to the composition of the steel material; Then, the yield strength of the ferrite phase is acquired : In the formula: represents the friction stress of the material; k y represents the Hall-Petch coefficient, which measures the grain boundary strengthening effect; the yield strength of the pearlite phase itself : In the formulae: S represents the lamellar spacing of pearlite, i.e. the average distance between the center lines of two adjacent cementite lamellae in the pearlite; represents the friction stress of the pearlite phase; k p represents a coefficient related to the pearlite, for measuring the lamellar interface strengthening effect; Finally, the macroscopic yield strength of the material as a whole is obtained : Obtain the overall tensile strength UTS of the material: Wherein: YR represents the yield ratio; Elongation A: In the formulae: C elo represents an empirical conversion factor from uniform elongation to total elongation; N a represents the intrinsic strain hardening exponent of the ferrite phase; N p represents the intrinsic strain hardening exponent of the pearlite phase; Step S2, steel pipe mechanical property prediction: build a model based on physical information neural network to realize the integrated prediction-optimization framework from process parameters to final mechanical properties; Step S3, steel pipe controlled cooling parameter optimization: use the physical information neural network model as the objective function, and use the improved Bayesian optimization algorithm to realize process parameter optimization.

2. The method of claim 1, wherein the method is based on a physical information neural network. The step S1 of "inputting process parameters and obtaining temperature field by finite difference method" is specifically: inputting a process parameter vector, including heating temperature T heat , open cooling temperature T start , final cooling temperature T end , cooling rate V cool , and spraying pressure P sp ; For each set of input process parameters, numerical solution is carried out by finite difference method: first, discretize the geometric model of the steel pipe and the time of the entire cooling process to form a space-time network; Then, input the process parameters as the initial conditions and boundary conditions of the physical scene, solve the unsteady heat conduction equation: In the formulae: T(r,z,t) denotes t the temperature at the radial position r and the axial position z at the time instant t; denotes the density of the steel material as a function of temperature; denotes the specific heat capacity of the steel material as a function of temperature; denotes the thermal conductivity of the steel material as a function of temperature; The final output is a complete, four-dimensional spatio-temporal temperature field data matrix T(r,z,t) .

3. The method for optimizing steel pipe cooling process based on physical information neural network according to claim 2, characterized in that: In the step S1, "complete flow field analysis and finite element simulation verification" is specifically: Flow field analysis: the spray cooling process on the surface of the steel pipe is simulated in detail by computational fluid dynamics; By analyzing the jet shape, water distribution and boiling heat transfer behavior of the nozzle, the distribution of convective heat transfer coefficient is obtained; Finite element simulation verification: from the generated large-scale data set, a number of process parameters are randomly selected, and several groups of parameters are simulated by using finite element software ABAQUS to complete the multi-physical field simulation coupled with heat-force-phase change; And these high-precision finite element simulation results are used as the calculation benchmark, compared with the output of the data enhancement algorithm in step S1, to evaluate the overall prediction accuracy, and calibrate the model parameters.

4. The method of claim 3, wherein the method is characterized by: In the step S2, the model based on physical information neural network is: In the formula: represents a set of optimal neural network parameters obtained by the model; represents a single, multi-output neural network constructed with inputs in the spatio-temporal domain: wherein: radial coordinate domain r is limited to the thickness of the steel pipe r 1 and r 2; axial coordinate domain z is limited to the analysis length of the steel pipe l ; time coordinate domain is limited to the total cooling time t final ; denotes a set of physical constraints, which includes initial condition constraints, boundary condition constraints, and grain size model constraints, which enables a transformation of a set of macroscopic process parameters to a specific set of physical constraints: In the formula: F IC , F BC , F GS respectively represent initial condition constraints, boundary condition constraints, and grain size model constraints. L total represents the total loss function, which is obtained by weighting and fusing the heat conduction physics loss L PDE , the initial condition loss L IC , the boundary condition loss L BC , the phase change kinetics loss L JMAK , the grain size model loss L GS , the material yield strength and grain size relationship loss L HP , the real data fidelity loss L data ​ In the formula: w 1、 w 2、 w 3、 w 4、 w 5、 w 6、 w 7 respectively represent the corresponding loss weight values; For any given set of process parameters X proc its corresponding macroscopic mechanical properties Y are: i.e. the final time of solution is determined according to the input process parameters t final Then, a point at 1 / 2 of the tube wall thickness is selected in the time-space domain, a complete forward calculation is performed by calling the neural network with the optimal parameters , and the corresponding output at the final time t final is extracted, thereby obtaining the performance prediction values of , UTS and A.

5. The method of claim 4, wherein: The improved Bayesian optimization algorithm in the step S3 is based on the existing Bayesian optimization algorithm using a Gaussian process as a proxy model, introduces a variational autoencoder as a core component of the proxy model, thereby injecting rich physical prior knowledge into the optimization process, the variational autoencoder includes an encoder and a decoder Two parts, the encoder receives the process parameters, and encodes them into parameters of a probability distribution, the encoder outputs a mean vector and a log variance vector log( );Then, a latent vector Z is sampled from the distribution defined from the encoding: In the formulae: denotes a standard normal noise; denotes an element-wise multiplication; The decoder receives the latent vector Z and decodes it to reconstruct the corresponding mechanical properties Y ; The variational autoencoder prior mean function is: For a given process parameter x its deterministic representation in latent space is acquired by an encoder and then reconstructed by a decoder into a performance prediction value; Improved bayesian optimization algorithm hybrid proxy model for new candidate points The posterior mean and variance for the new candidate point x* are: where: denotes the new point denotes the covariance vector between the new point and all old points X; denotes the covariance matrix between all evaluated points; denotes the covariance kernel function; denotes the noise variance; I denotes the identity matrix.

Images