High-temperature silo digital twin reconstruction and simulation method
By combining UAV data acquisition with a thermal-visible light fusion neural radiation field model and a physically constrained neural network, the problems of the invisible internal state of high-temperature silos and the complexity of modeling were solved, achieving high-precision temperature field reconstruction and real-time simulation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUADIAN ZHENGZHOU MECHANICAL DESIGN INST
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-24
AI Technical Summary
The simulation suffers from low accuracy and difficulty in achieving real-time performance due to the invisibility of the internal state of high-temperature silos, the difficulty in modeling, the complexity of multi-physics coupling, and the sparseness of temperature measurement points.
Data was collected by a drone using a visible light camera and an infrared thermal imager. A physical information neural network was constructed by combining a thermal-visible light fusion neural radiation field model and the Brinkman-Forchheimer-Darcy partial differential equation. The temperature field was inverted through a gradient descent optimization algorithm, and the lattice Boltzmann method was used for real-time simulation and uncertainty quantification.
It achieves accurate reconstruction and real-time simulation of the internal temperature field of high-temperature silos, reduces errors, improves reconstruction accuracy and computational efficiency, and provides reliable confidence indicators.
Smart Images
Figure CN122454040A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computational fluid dynamics, specifically to a method for digital twin reconstruction and simulation of high-temperature silos. Background Technology
[0002] High-temperature sand and gravel thermal storage silos are key equipment in fields such as solar thermal power generation and industrial waste heat utilization. They are typically enormous (tens of meters in height), have complex internal structures (filled with phase change materials or sand and gravel), and operate in high-temperature environments (>500℃). Currently, the operation monitoring and condition assessment of high-temperature silos face the following technical challenges: 1. Invisible internal state: The inside of the silo is in a "black box" state. Traditional sensors can only collect temperature data at discrete points and cannot know the temperature distribution of the whole field, making it difficult to accurately assess the thermal storage state and predict the risk of hot spots. 2. Difficulty in 3D modeling: Traditional laser scanning or manual modeling methods are time-consuming and labor-intensive, and cannot reproduce realistic texture details such as surface rust and insulation layer peeling, affecting the accuracy of digital twin models; 3. Complex multi-physics coupling: The heat conduction between sand and gravel particles, high-temperature radiation and air convection heat transfer are coupled with each other, which is difficult to describe accurately by traditional empirical formulas, resulting in large simulation prediction errors; 4. Sparse temperature measurement points: Due to the high-temperature environment and cost limitations, the number of temperature measurement points inside the silo is limited (usually <1% space coverage), making it difficult to support accurate reconstruction of the entire temperature field.
[0003] In existing technologies, the temperature field simulation method for thermal storage devices based on the finite volume method requires a large amount of mesh generation work, resulting in low computational efficiency and difficulty in achieving real-time simulation; the temperature prediction method based on deep learning has poor generalization ability under sparse temperature measurement points, and the prediction accuracy is difficult to guarantee. Summary of the Invention
[0004] The present invention proposes a digital twin reconstruction and simulation method for high-temperature silos, which can at least solve one of the technical problems in the background art.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for digital twin reconstruction and simulation of high-temperature silos includes the following steps: Step S1, Data Acquisition: Use a drone equipped with a visible light camera and an infrared thermal imager to collect a multi-view image sequence of the silo's exterior, and at the same time collect temperature data of preset temperature measurement points inside the silo. Step S2, 3D Reconstruction: Based on the acquired visible light and infrared images, a thermal-visible light fusion neural radiation field model is constructed. The appearance of the silo is implicitly represented by multi-resolution hash coding to achieve high-fidelity 3D reconstruction of the silo surface. Step S3, Physical Field Inversion: Establish the Brinkman-Forchheimer-Darcy partial differential equation describing the heat transfer of porous media as a physical constraint, construct a physical information neural network, use the permeability K and the inertia coefficient CF as learnable parameters of the neural network, and simultaneously invert the internal temperature field of the silo and the thermal property parameters of the porous media through the gradient descent optimization algorithm to achieve accurate acquisition of the whole field temperature under the condition of sparse temperature measurement points. Step S4, Real-time simulation: The lattice Boltzmann method is used to perform real-time simulation calculations of thermal-fluid coupling on the GPU, and the inversion results are verified and predicted. Step S5, Uncertainty Quantification: The uncertainty of the temperature field prediction result is quantified using a Bayesian physical information neural network, and the confidence interval of the temperature field is output.
[0006] As a preferred embodiment of the digital twin reconstruction and simulation method for high-temperature silos described in this invention, the method for constructing the thermal-visible light fused neural radiation field model in step S2 is as follows: Constructing the volume rendering equations for the thermal radiation field:
[0007] in, For along the ray The rendered composite radiation value, For the ray equation, Centered on the camera The direction of the ray. and Let the near and far boundaries of the integral be the boundaries. Cumulative transmittance The absorption coefficient is... The scattering coefficient is... The intensity of spontaneous emission. The intensity of the scattered radiation. The absorption coefficient is... The scattering coefficient is... This represents the intensity of spontaneous emission.
[0008] As a preferred embodiment of the digital twin reconstruction and simulation method for high-temperature silos described in this invention, the loss function construction method for the physical information neural network in step S3 is as follows: Construct the physical consistency loss function:
[0009] in, To monitor the loss based on measured temperature points, For physical residuals, and These are the weighting coefficients; The physical residual term Includes energy equation residuals Continuity equation residuals and momentum equation residuals : .
[0010] As a preferred embodiment of the digital twin reconstruction and simulation method for high-temperature silos described in this invention, the momentum equation residual is constructed based on the Brinkman-Forchheimer-Darcy equation.
[0011] in, For fluid density, Porosity For fluid velocity vector, For pressure, For effective dynamic viscosity, For fluid dynamic viscosity, For penetration rate, Forchheimer inertia coefficient, It is the gravitational acceleration vector. The coefficient of thermal expansion is For temperature, This is a reference temperature.
[0012] As a preferred embodiment of the high-temperature silo digital twin reconstruction and simulation method of the present invention, wherein the calculation method for the effective thermal conductivity of the porous medium in step S3 is as follows:
[0013] in, The thermal conductivity of the solid phase is... The thermal conductivity of the fluid phase. Porosity As a contributor to radiative heat transfer, Contribution to mixed convection; The contribution of radiation heat transfer The calculation formula is:
[0014] in, This is the Stefan-Boltzmann constant. The particle diameter is For temperature, The temperature correction factor is expressed as follows: .
[0015] As a preferred embodiment of the digital twin reconstruction and simulation method for high-temperature silos described in this invention, in step S4, the lattice Boltzmann method adopts the D3Q19 discrete velocity model, and the evolution equation of the particle distribution function is:
[0016] in, For the first Particle distribution function in discrete velocity directions It is a spatial position vector. For the first A velocity vector in a discrete velocity direction For time step, The relaxation time is dimensionless. This is the equilibrium distribution function.
[0017] As a preferred embodiment of the digital twin reconstruction and simulation method for high-temperature silos described in this invention, the uncertainty quantification method in step S5 is as follows: Construct the posterior distribution of the parameters:
[0018] in, For neural network parameters, For observation data, Let be the likelihood function. For the first One physical constraint term; Decompose total uncertainty into cognitive uncertainty and accidental uncertainty:
[0019] in, For predicted values Total variance To understand uncertainty, It is due to chance and uncertainty.
[0020] The beneficial effects of this invention are: This invention embeds the Brinkman-Forchheimer-Darcy equation into the loss function of a neural network and uses physical conservation laws to constrain the spatial distribution of the temperature field. Compared with pure data-driven methods and traditional finite volume methods, the method of this invention reduces errors and improves accuracy. The thermal-visible light fusion neural radiation field of this invention realizes the joint reconstruction of temperature field and geometric field, and fuses the temperature information of infrared thermal image and the texture information of visible light image in implicit space. Compared with processing the two modes separately, the reconstruction accuracy is improved. Attached Figure Description
[0021] Figure 1 This is an overall flowchart of the digital twin reconstruction and simulation method for high-temperature silos of the present invention.
[0022] Figure 2 This is a schematic diagram of the thermal-visible light fusion NeRF model structure of the high-temperature silo digital twin reconstruction and simulation method of the present invention.
[0023] Figure 3 This is a schematic diagram of the physical information neural network structure of the digital twin reconstruction and simulation method for high-temperature silos of the present invention.
[0024] Figure 4 This is a comparison chart of temperature field inversion results and measured values in an embodiment of the digital twin reconstruction and simulation method for high-temperature silos of the present invention. Detailed Implementation
[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.
[0026] like Figures 1-3 As shown, a digital twin reconstruction and simulation method for high-temperature silos is provided, including the following steps: Step S1, Data Acquisition: Use a drone equipped with a visible light camera and an infrared thermal imager to collect a multi-view image sequence of the silo's exterior, and at the same time collect temperature data of preset temperature measurement points inside the silo. Step S2, 3D Reconstruction: Based on the acquired visible light and infrared images, a thermal-visible light fusion neural radiation field model is constructed. The appearance of the silo is implicitly represented by multi-resolution hash coding to achieve high-fidelity 3D reconstruction of the silo surface. Step S3, Physical Field Inversion: Establish the Brinkman-Forchheimer-Darcy partial differential equation describing the heat transfer of porous media as a physical constraint, construct a physical information neural network, use the permeability K and the inertia coefficient CF as learnable parameters of the neural network, and simultaneously invert the internal temperature field of the silo and the thermal property parameters of the porous media through the gradient descent optimization algorithm to achieve accurate acquisition of the whole field temperature under the condition of sparse temperature measurement points. Step S4, Real-time simulation: The lattice Boltzmann method is used to perform real-time simulation calculations of thermal-fluid coupling on the GPU, and the inversion results are verified and predicted. Step S5, Uncertainty Quantification: The uncertainty of the temperature field prediction result is quantified using a Bayesian physical information neural network, and the confidence interval of the temperature field is output.
[0027] Specifically, the method for constructing the thermal-visible light fused neural radiation field model in step S2 is as follows: Constructing the volume rendering equations for the thermal radiation field:
[0028] in, For along the ray The rendered composite radiation value, For the ray equation, Centered on the camera The direction of the ray. and Let the near and far boundaries of the integral be the boundaries. Cumulative transmittance The absorption coefficient is... The scattering coefficient is... The intensity of spontaneous emission. The intensity of the scattered radiation. The absorption coefficient is... The scattering coefficient is... This represents the intensity of spontaneous emission.
[0029] Furthermore, the method for constructing the loss function of the physical information neural network in step S3 is as follows: Construct the physical consistency loss function:
[0030] in, To monitor the loss based on measured temperature points, For physical residuals, and These are the weighting coefficients; The physical residual term Includes energy equation residuals Continuity equation residuals and momentum equation residuals : .
[0031] Specifically, the momentum equation residuals are constructed based on the Brinkman-Forchheimer-Darcy equations:
[0032] in, For fluid density, Porosity For fluid velocity vector, For pressure, For effective dynamic viscosity, For fluid dynamic viscosity, For penetration rate, Forchheimer inertia coefficient, It is the gravitational acceleration vector. The coefficient of thermal expansion is For temperature, This is a reference temperature.
[0033] Furthermore, the method for calculating the effective thermal conductivity of the porous medium in step S3 is as follows:
[0034] in, The thermal conductivity of the solid phase is... The thermal conductivity of the fluid phase. Porosity As a contributor to radiative heat transfer, Contribution to mixed convection; The contribution of radiation heat transfer The calculation formula is:
[0035] in, This is the Stefan-Boltzmann constant. The particle diameter is For temperature, The temperature correction factor is expressed as follows: .
[0036] In step S4, the lattice Boltzmann method employs the D3Q19 discrete velocity model, and the evolution equation of the particle distribution function is:
[0037] in, For the first Particle distribution function in discrete velocity directions It is a spatial position vector. For the first A velocity vector in a discrete velocity direction For time step, The relaxation time is dimensionless. This is the equilibrium distribution function.
[0038] Furthermore, the uncertainty quantification method in step S5 is as follows: Construct the posterior distribution of the parameters:
[0039] in, For neural network parameters, For observation data, Let be the likelihood function. For the first One physical constraint term; Decompose total uncertainty into cognitive uncertainty and accidental uncertainty:
[0040] in, For predicted values Total variance To understand uncertainty, It is due to chance and uncertainty.
[0041] The specific implementation examples are as follows: Example 1 Taking a high-temperature sand and gravel thermal storage silo in a solar thermal power plant as an example, the silo is 25m high and 12m in diameter, and is filled with sand and gravel particles (average particle size...). =5mm), operating temperature range 200-600℃.
[0042] Step S1: Data Acquisition A DJI M300RTK drone, equipped with a Zenmuse P1 visible light camera and an H20T infrared thermal imager, was used to circle the silo at an altitude of 5-30 meters, capturing multi-view images at 8° intervals. The visible light images have a resolution of 8192×6144 pixels, and the infrared images have a resolution of 640×512 pixels. A total of 360 visible light images and 360 infrared images were captured.
[0043] Inside the silo, temperature measuring points are arranged every 2.5m along the height, for a total of 10 layers, with 6 temperature measuring points on each layer, for a total of 60 K-type thermocouple temperature measuring points, and the temperature acquisition frequency is 1Hz.
[0044] Step S2: 3D Reconstruction A thermal-visible light fusion neural radiation field model is constructed using the Instant-NGP framework. The network structure is as follows: Multi-resolution hashing: 16-layer resolution, from arrive The feature dimension is 2; MLP network: 2 hidden layers, 64 neurons in each layer; Output: Volume density and color and temperature ; Training parameters: Learning rate: (exponential decay to) ); Batch size: ray; Number of training rounds: 20,000; Optimizer: Adam; The training time is about 15 minutes, and the reconstruction accuracy reaches 8mm.
[0045] Step S3: Physics Field Inversion Physical information neural network structure: Input layer: 4 neurons (spatial coordinates) and time ); Hidden layers: 8 fully connected layers, each with 256 neurons, using the tanh activation function; Output layer: 5 neurons (temperature) velocity components ,pressure ); Training parameters: Learning rate: (exponential decay to) ); Batch size: 4096 points; Training rounds: 50,000 rounds; Optimizer: Adam; Loss function weights: Data loss weights ; Physical residual weights ; Energy equation weights ; Weights of the continuity equation ; Momentum equation weights ; Physical parameter settings: Porosity ; thermal conductivity of solid phase W / (m·K); Fluid phase thermal conductivity W / (m·K); Penetration Initialize to m²; Forchheimer inertia coefficient Initialize to 0.5; The training time is approximately 2 hours, and the penetration rate obtained through inversion is... m², Forchheimer coefficient of inertia .
[0046] Step S4: Real-time simulation The D3Q19 lattice Boltzmann model is used, with a mesh resolution of [missing information]. Time step s, relaxation time .
[0047] GPU configuration: NVIDIA RTX 4090, 24GB VRAM.
[0048] The single-step calculation time is approximately 0.5ms, achieving a simulation speed of 2000 steps / second, which meets the real-time requirements.
[0049] Step S5: Uncertainty Quantification The Dropout variational inference method was used, with the Dropout rate set to 0.1. 100 Monte Carlo samplings were performed to calculate the mean and variance of the temperature prediction.
[0050] Result Validation The temperature field obtained by inversion was compared with the 10 reserved verification temperature measurement points, and the results showed: Mean absolute error: 12.3°C; Average relative error: 3.2%; 95% confidence interval coverage: 96%.
[0051] Example 2: Taking a phase change material thermal storage silo for an industrial waste heat utilization project as an example, the silo is 15m high and 8m in diameter, filled with paraffin / expanded graphite composite phase change material, and operates in a temperature range of 100-300℃.
[0052] Digital twin reconstruction was performed using the same method as in Example 1, with the main parameters adjusted as follows: Porosity ; thermal conductivity of solid phase W / (m·K); Fluid phase thermal conductivity W / (m·K); The temperature field inversion results show that: Mean absolute error: 8.7°C; Average relative error: 4.1%; 95% confidence interval coverage: 94%; The comparison and analysis with Example 1 is shown in Table 1: Table 1 Comparative Analysis
[0053] Example 3: Method Comparison Experiment Reference Figure 4To verify the technical effect of the present invention, a comparative experiment was conducted between the method of the present invention and the prior art.
[0054] Comparison method: 1. Pure data-driven method (DL): The temperature field is directly fitted using a multilayer perceptron without physical constraints; 2. Finite Volume Method (FVM): Traditional CFD simulation is performed using ANSYS Fluent; 3. The method of this invention (PINNs): Physical information neural network + parameter self-learning; Experimental conditions Temperature measurement point coverage: 0.8% (60 temperature measurement points / total space); Number of verification points: 20 reserved verification points; Evaluation metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Maximum Error (MAX); The comparison results are shown in Table 2 below: Table 2 Comparison Results
[0055] Technical effect analysis: 1. Advantages under sparse temperature measurement point conditions: The method of this invention can still maintain an inversion error of 3.2% even when the temperature measurement point coverage is only 0.8%, while the pure data-driven method has an error as high as 45.2℃ under the same conditions, which proves the effectiveness of physical constraints.
[0056] 2. Self-learning effect of porous media parameters: The permeability obtained by the method of this invention is automatically inverted. m² and laboratory measurements Compared to m², the error is only 8.6%, proving the accuracy of the parameter self-learning mechanism.
[0057] 3. Improved computational efficiency: Compared with the traditional finite volume method, the computation time of the method of this invention is reduced from 4 hours to 2 hours, with an efficiency improvement of 50% and an accuracy improvement of 57%.
[0058] 4. Uncertainty Quantification Effect: The 95% confidence interval coverage rate output by the method of this invention reaches 96%, providing a reliable confidence index for operational decision-making.
[0059] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0060] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0061] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for digital twin reconstruction and simulation of high-temperature silos, characterized in that, Includes the following steps: Step S1, Data Acquisition: Use a drone equipped with a visible light camera and an infrared thermal imager to collect a multi-view image sequence of the silo's exterior, and at the same time collect temperature data of preset temperature measurement points inside the silo. Step S2, 3D Reconstruction: Based on the acquired visible light and infrared images, a thermal-visible light fusion neural radiation field model is constructed. The appearance of the silo is implicitly represented by multi-resolution hash coding to achieve high-fidelity 3D reconstruction of the silo surface. Step S3, Physical Field Inversion: Establish the Brinkman-Forchheimer-Darcy partial differential equation describing the heat transfer of porous media as a physical constraint, construct a physical information neural network, use the permeability K and the inertia coefficient CF as learnable parameters of the neural network, and simultaneously invert the internal temperature field of the silo and the thermal property parameters of the porous media through the gradient descent optimization algorithm to achieve accurate acquisition of the whole field temperature under the condition of sparse temperature measurement points. Step S4, Real-time simulation: The lattice Boltzmann method is used to perform real-time simulation calculations of thermal-fluid coupling on the GPU, and the inversion results are verified and predicted. Step S5, Uncertainty Quantification: The uncertainty of the temperature field prediction result is quantified using a Bayesian physical information neural network, and the confidence interval of the temperature field is output.
2. The method for digital twin reconstruction and simulation of high-temperature silos according to claim 1, characterized in that: The method for constructing the thermal-visible light fused neural radiation field model in step S2 is as follows: Constructing the volume rendering equations for the thermal radiation field: in, For along the ray The rendered composite radiation value, For the ray equation, Centered on the camera The direction of the ray. and Let the near and far boundaries of the integral be the boundaries. Cumulative transmittance The absorption coefficient is... The scattering coefficient is... The intensity of spontaneous emission. The intensity of the scattered radiation. The absorption coefficient is... The scattering coefficient is... This represents the intensity of spontaneous emission.
3. The method for digital twin reconstruction and simulation of high-temperature silos according to claim 1, characterized in that: The method for constructing the loss function of the physical information neural network in step S3 is as follows: Construct the physical consistency loss function: in, To monitor the loss based on measured temperature points, For physical residuals, and These are the weighting coefficients; The physical residual term Includes energy equation residuals Continuity equation residuals and momentum equation residuals : 。 4. The method for digital twin reconstruction and simulation of high-temperature silos according to claim 3, characterized in that: The momentum equation residuals are constructed based on the Brinkman-Forchheimer-Darcy equations: in, For fluid density, Porosity For fluid velocity vector, For pressure, For effective dynamic viscosity, For fluid dynamic viscosity, For penetration rate, Forchheimer inertia coefficient, It is the gravitational acceleration vector. The coefficient of thermal expansion is... For temperature, This is a reference temperature.
5. The method for digital twin reconstruction and simulation of high-temperature silos according to claim 1, characterized in that: The method for calculating the effective thermal conductivity of the porous medium in step S3 is as follows: in, The thermal conductivity of the solid phase is... The thermal conductivity of the fluid phase, Porosity As a contributor to radiative heat transfer, Contribution to mixed convection; The contribution of radiation heat transfer The calculation formula is: in, This is the Stefan-Boltzmann constant. The particle diameter is For temperature, The temperature correction factor is expressed as follows: 。 6. The method for digital twin reconstruction and simulation of high-temperature silos according to claim 1, characterized in that: In step S4, the lattice Boltzmann method employs the D3Q19 discrete velocity model, and the evolution equation of the particle distribution function is: in, For the first Particle distribution function in discrete velocity directions It is a spatial position vector. For the first A velocity vector in a discrete velocity direction For time step, The relaxation time is dimensionless. This is the equilibrium distribution function.
7. The method for digital twin reconstruction and simulation of high-temperature silos according to claim 1, characterized in that: The uncertainty quantification method in step S5 is as follows: Construct the posterior distribution of the parameters: in, For neural network parameters, For observation data, Let be the likelihood function. For the first One physical constraint term; Decompose total uncertainty into cognitive uncertainty and accidental uncertainty: in, For predicted values Total variance To understand uncertainty, It is due to chance and uncertainty.