Temperature field prediction method and system for grain storage multi-field coupling digital twinning, and storage medium

Abstract

The invention relates to the crossing field of grain storage intelligent monitoring and digital twinning technologies, and provides a grain storage multi-field coupling digital twinning-oriented temperature field prediction method and system and a storage medium, and the method comprises the following steps: S1, modeling and discretization: building a two-dimensional granary section under a unified geometric and regular grid; s2, performing multi-working-condition simulation and unified analysis rearrangement, and performing transient simulation on multi-physical-property and multi-boundary-parameter working conditions by adopting CFD (Computational Fluid Dynamics); and S3, geometric priori construction: generating a fluid domain mask and a grain pile domain mask, and calculating a signed distance field as geometric priori. Scalar conditions such as external temperature, time phase, boundary parameters and the like are mapped into channel-level scaling-offset operators through affine transformation, fine-grained regulation and control of a feature space are achieved, compared with a simple splicing method, the boundary neighborhood prediction precision is improved, the self-adaptive influence radius and assimilation intensity are designed for double domains, and the method has the advantages of being high in adaptability and high in adaptability. The boundary layer temperature correction precision is improved, and the observation information utilization efficiency is improved.

Description

Technical Field

[0001] This invention relates to the intersection of intelligent monitoring and digital twin technology in grain storage, and particularly to a method, system, and storage medium for predicting temperature field in multi-field coupled digital twins for grain storage. Background Technology

[0002] During static grain storage, the spatiotemporal evolution of the internal temperature field of the grain pile is controlled by the strong coupling of multiple physical processes, including heat and mass transfer through porous media, natural convection, and respiration metabolism. Its dynamic characteristics are directly related to the accumulation rate of quality degradation risks such as mold and pests. Existing monitoring methods mainly rely on discretely distributed point sensors, which have inherent limitations due to severe spatial sampling deficiencies, making it difficult to reconstruct a continuous temperature field across the entire area. While high-fidelity numerical simulations based on partial differential equations (PDEs) (such as CFD) can characterize multi-field coupling mechanisms, their computational complexity increases superlinearly at the grid scale, with single predictions taking hours, failing to meet the stringent real-time and interactive requirements of digital twins.

[0003] In recent years, deep learning methods have shown significant potential for proxy modeling in the field of scientific computing, but they still face key challenges in grain storage scenarios: (1) There are drastic jumps in physical properties such as porosity and thermal conductivity in the grain pile-air dual domain, and the weak formal representation of coupling conditions at the interface leads to numerical oscillations in the prediction field; (2) External boundary drives include multi-scale oscillations such as daily and annual cycles and orientation asymmetry caused by solar radiation, and traditional single scalar conditions cannot fully parameterize the spatiotemporal boundary; (3) The distribution drift caused by error accumulation in long-term autoregressive prediction lacks an effective suppression mechanism, and the soft constraints of physical conservation laws are easily marginalized by the optimization process; (4) The assimilation and fusion of sparse field observations lacks adaptive weight design for heterogeneous structures in dual domains, and the boundary layer error correction capability is insufficient. In addition, the complete digital twin workflow from CFD simulation to deep learning model still lacks a systematic methodology that guarantees geometric consistency, data interface standardization, and reproducibility. Therefore, there is an urgent need to develop a hybrid modeling paradigm that integrates physical priors and data-driven approaches, achieving millisecond-level inference speed while maintaining CFD-level physical fidelity, and possessing online observation assimilation and distribution out-generalization capabilities, thereby constructing a self-consistent digital twin closed loop of simulation-prediction-correction. Summary of the Invention

[0004] In view of the problems existing in the prior art, the purpose of this invention is to provide a method, system and storage medium for temperature field prediction of multi-field coupled digital twins for grain storage, so as to solve the problems mentioned in the background art.

[0005] To address the above problems, the present invention adopts the following technical solution: a temperature field prediction method for multi-field coupled digital twins in grain storage, comprising the following steps:

[0006] S1. Modeling and Discretization: A two-dimensional grain silo cross section is established under a unified geometry and regular grid. The grain pile domain and the upper air domain are divided. Boundary conditions are set for bottom insulation, side walls and roof driven by external temperature. The grain pile / air interface satisfies the continuity of temperature and heat flux. A multi-field coupled model is established by discretizing using local thermal equilibrium and finite volume method.

[0007] S2. Multi-condition simulation and unified analytical rearrangement: CFD is used to perform transient simulation of multi-property and multi-boundary parameter conditions, and simulation data containing coordinates and time-stamped temperatures are exported. Time is uniformly converted to seconds and temperature to Kelvin. The temperature fields of each condition are rearranged according to a unified rule grid, and the shared coordinate index is extracted.

[0008] S3. Geometric prior construction: generate fluid domain mask and grain pile domain mask, and calculate the signed distance field (SDF) as geometric prior.

[0009] S4. Input design, construct a static multi-channel vector containing initial temperature, initial moisture content, SDF, domain mask and porosity, and a scalar condition vector containing daily / annual phase, left and right wall / roof boundary temperatures, time step and boundary parameters.

[0010] S5. Data cleaning: Based only on the training subset and statistically analyzing the mean and standard deviation of temperature in the fluid domain, construct the z-space incremental learning objective and perform amplitude soft clipping using hyperbolic tangent.

[0011] S6. Train the proxy network, divide the training / test set into groups according to the run identifier and persist the shared split file, use UNet as the backbone and integrate FiLM conditional modulation, calculate the loss in the domain mask and train using dual-domain weighting or only grain pile domain loss, and save the results.

[0012] S7. Autoregressive rolling prediction: The normalized representation of the current temperature field, static multi-channel and scalar conditions are used as inputs. The output temperature increment is updated to the next time step. Rolling prediction is performed in a multi-step manner.

[0013] S8, Nudging assimilation and posterior generation: Linear incremental fitting of the external temperature function is performed within the most recent D-day observation window, and the observation information is calculated and Nudging assimilated with Gaussian kernel weights. The posterior field is calculated and used as the initial state for the next autoregression step.

[0014] S9. Model Evaluation: Output the temperature field, prior / posterior, and absolute error heatmaps in physical coordinates at a proportional scale, and overlay the inner and outer boundary lines to calculate the target area of ​​the grain pile. The MAE, RMSE, nRMSE, PSNR, and SSIM indices are used to evaluate the final assimilated model and prediction results.

[0015] Preferably, in step S1, modeling and discretization include: establishing continuity, momentum, energy, and moisture migration equations based on local thermal equilibrium and the finite volume method, with the boundary condition being bottom adiabatic and the boundary temperature of the top and sun-facing sidewall being... The shady side wall is ,in, The temperature difference caused by solar radiation is due to the difference, and , Given Kelvin, and the external temperature function is: ,in, The term represents the annual average temperature. and These are the annual cycle amplitude and the daily cycle amplitude, respectively. and These are the angular frequencies of the annual cycle and the daily cycle, respectively. , , In radians per second, For time, the unit is , and These are the phase parameters for annual temperature variation and daily temperature variation, respectively.

[0016] Preferably, in step S2, the data export and unified parsing rearrangement use the minimum interface of coordinate-time-temperature, with the coordinate position being... The time unit is The unit of temperature is The export and parsing formats support at least one of CSV, TXT or HDF5, and the temperature fields of each operating condition are rearranged according to the shared rule grid, and bilinear resampling is performed when necessary to align the grid, while extracting the shared coordinate index.

[0017] Preferably, in step S3, the signed distance field function is: ,in, To describe spatial points The signed distance field to the boundary of the fluid domain. This is a Euclidean distance transformation. For fluid domain mask The complement, when Time indicates the domain, To achieve the closest distance to the fluid domain boundary, the mask boundary is smoothed by 1 pixel.

[0018] Preferably, in step S5, normalization and incremental update satisfy: ,in, This is the updated value of the normalized increment. In time step spatial point The temperature at that location The average temperature. The standard deviation of temperature, To prevent extremely small values ​​where the denominator is zero, the network output increment is pruned to satisfy the following: ,in, This is the incremental output of the pruned network. This is the maximum limit value for the network output increment. It is the hyperbolic tangent function. For the original increment output by the network, the physical temperature is restored to satisfy: ,in, This is the restored physical temperature value.

[0019] Preferably, in step S6, the proxy network adopts a UNet encoder-decoder structure and maps the scalar conditional vector from step S4 to channel-level scaling and offset through a FiLM module at each layer to perform affine modulation based on the FiLM module on the intermediate features. The affine modulation function is: ,in, This refers to the modulated intermediate feature channel. This is the channel scaling factor. This is the intermediate feature channel before modulation. The channel offset factor is used. The loss is calculated within the domain mask and weighted using a dual-domain weighting function, which is: ,in, For the total loss, Here are the loss weighting coefficients, and , Mean square error, The target area for the grain pile. For the air area above the grain pile, training or testing is performed by dividing the entire group by run and persistently sharing the split files.

[0020] Preferably, in step S7, the natural regression update function is: ,in, and The first Step and the first spatial point The normalized increment intermediate quantity at the location, For the first Step, Spatial Point Network output increment at the location, For the first Step, Spatial Point The final temperature value at that location.

[0021] Preferably, in step S8, the linear increment function is: ,in, The sum is the linear increment of the fitted external temperature function. and The coefficients are those of the linear fit. The slope The intercept is... Given the ambient temperature, the function for calculating the observed information is: ,in, For the first Each sensor at time step Observational information, For the first Each sensor at time step The actual observed value, For the mapping operator from the temperature field to the sensor location, In time step ,sensor The temperature field model prediction at the location, with the Nudging assimilation function being: ,in, For the first A sensor at a spatial point The corresponding Gaussian weight kernel, For spatial points to the first Euclidean distance of each sensor's spatial location Let be the spatial scale parameter of the Gaussian kernel, and the posterior field function be: ,in, and At time step spatial point The posterior and prior temperature fields at the location, For spatial points Assimilation gain coefficient at the location, To sum the products of the weights and observation information of all sensors, To sum the weights of all sensors, To prevent the minimum value where the denominator is zero.

[0022] A temperature field prediction system for multi-field coupled digital twins in grain storage includes:

[0023] The data acquisition and parsing module is used to collect various types of raw data and perform format parsing and preliminary structure processing on them;

[0024] The geometry and prior construction module is used to build geometric models of research objects and establish prior hypotheses by combining domain knowledge or historical experience.

[0025] The conditional encoding module is used to encode geometric, prior, and other conditional information into vectors or features that can be recognized by the deep network, providing conditional constraints for network training and inference.

[0026] The data cleaning module is used to remove outliers, impute missing values, and filter noise from the raw data to improve data quality and ensure the accuracy of subsequent analysis.

[0027] The deep network training module is used to train deep networks with specific functions based on cleaned data and conditional encoding, enabling them to learn patterns and rules in the data.

[0028] The autoregressive inference module is used to derive or predict the dynamic changes of physical quantities step by step in an autoregressive manner using a trained deep network.

[0029] The data assimilation module is used to fuse observation data with model prediction results and correct the model output through algorithms to improve its consistency with reality.

[0030] The evaluation and visualization module is used to quantitatively evaluate model performance and data assimilation effects, and to visually display the results in the form of charts and animations to assist in analysis and decision-making.

[0031] A computer storage medium storing computer-executable instructions, wherein the computer-executable instructions, when executed, implement the temperature field prediction method for multi-field coupled digital twins of grain storage as described in any one of claims 1-8.

[0032] The present invention provides a method, system, and storage medium for predicting temperature fields using multi-field coupled digital twins for grain storage, the advantages of which are:

[0033] 1. Original construction of a two-layer digital twin architecture: By proposing a layered digital twin framework consisting of a high-fidelity simulation layer and a neural network proxy layer, computational efficiency is achieved through offline training while maintaining the accuracy of temperature field prediction;

[0034] 2. Geometric-physical prior tensor field encoding mechanism: By innovatively integrating the signed distance field, porosity distribution, and initial thermal and humid state into a static multi-channel tensor, implicit learning of geometric constraints is achieved through the receptive field propagation of the convolution kernel;

[0035] 3. Multi-scale conditional modulation feature linear injection strategy: By mapping scalar conditions such as external temperature, time phase, and boundary parameters into channel-level scaling-offset operators through affine transformation, fine-grained control of the feature space is achieved. Compared with simple stitching methods, the prediction accuracy of the boundary neighborhood is improved.

[0036] 4. Cooperative stabilization mechanism of incremental learning and soft constraint clipping: By modeling temperature increments rather than absolute values ​​in the standardized latent space, and combining this with hyperbolic tangent soft clipping to limit the single-step change amplitude, the smoothness index of the physics field is improved.

[0037] 5. Innovative fusion of hierarchical Bayesian-variational assimilation: By proposing for the first time a two-level assimilation strategy of maximum a posteriori estimation of boundary parameters and Gaussian weighted update of global state, adaptive influence radius and assimilation intensity are designed for dual domains, improving the accuracy of boundary layer temperature correction and the efficiency of observation information utilization;

[0038] 6. Standardized paradigm for end-to-end engineering processes: By establishing a modular pipeline from geometric topology to evaluation visualization, it provides a path for modular systems and computer-readable storage media implementation, supporting rapid deployment and maintenance at the edge / cloud. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A schematic diagram of the overall process of a temperature field prediction method, system and storage medium for multi-field coupled digital twins in grain storage provided in this application;

[0041] Figure 2 A schematic diagram of the application framework for a temperature field prediction method, system, and storage medium for multi-field coupled digital twins in grain storage, provided in this application;

[0042] Figure 3 A two-dimensional physical schematic diagram of a temperature field prediction method, system, and storage medium for multi-field coupled digital twins in grain storage, provided for this application;

[0043] Figure 4 A schematic diagram of the temperature distribution of a grain pile stored for 150 days, provided in this application, for a temperature field prediction method, system, and storage medium based on multi-field coupled digital twins for grain storage.

[0044] Figure 5 A schematic diagram of Unet autoregressive prediction for day 1 of a temperature field prediction method, system and storage medium for multi-field coupled digital twin of grain storage provided in this application;

[0045] Figure 6A schematic diagram of Unet autoregressive prediction for day 3 of a temperature field prediction method, system and storage medium for multi-field coupled digital twin of grain storage provided in this application;

[0046] Figure 7 A schematic diagram of Unet autoregressive prediction for day 7 of a temperature field prediction method, system and storage medium for multi-field coupled digital twins of grain storage provided in this application;

[0047] Figure 8 A schematic diagram of the UNet prediction prior field for a temperature field prediction method, system, and storage medium for multi-field coupled digital twins of grain storage provided in this application;

[0048] Figure 9 This application provides a temperature field prediction method, system, and a schematic diagram of the nudging assimilation posterior field distribution for a multi-field coupled digital twin for grain storage.

[0049] Figure 10 This application provides a schematic diagram of a temperature field prediction method, system, and assimilation correction amount for a multi-field coupled digital twin for grain storage. Detailed Implementation

[0050] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0051] like Figures 1-10 As shown, this embodiment proposes a temperature field prediction method for multi-field coupled digital twins in grain storage, including the following steps:

[0052] S1. Modeling and Discretization: A two-dimensional grain silo cross section is established under a unified geometry and regular grid. The grain pile domain and the upper air domain are divided. Boundary conditions are set for bottom insulation, side walls and roof driven by external temperature. The grain pile / air interface satisfies the continuity of temperature and heat flux. A multi-field coupled model is established by discretizing using local thermal equilibrium and finite volume method.

[0053] S2. Multi-condition simulation and unified analytical rearrangement: CFD is used to perform transient simulation of multi-property and multi-boundary parameter conditions, and simulation data containing coordinates and time-stamped temperatures are exported. Time is uniformly converted to seconds and temperature to Kelvin. The temperature fields of each condition are rearranged according to a unified rule grid, and the shared coordinate index is extracted.

[0054] S3. Geometric prior construction: generate fluid domain mask and grain pile domain mask, and calculate the signed distance field (SDF) as geometric prior.

[0055] S4. Input design, construct a static multi-channel vector containing initial temperature, initial moisture content, SDF, domain mask and porosity, and a scalar condition vector containing daily / annual phase, left and right wall / roof boundary temperatures, time step and boundary parameters.

[0056] Specifically, the static multichannel includes at least: a feature for identifying the initial temperature distribution of the fluid domain. Channels, in which, in the fluid domain mask Fill the initial temperature position Used to identify the initial moisture content distribution of the grain pile area. The passage, including the masked area in the grain pile region. The initial moisture content of the filling position Used to describe spatial points within a fluid domain. Signed distance from the boundary Channel, in which the signed distance field of the fluid domain Used to distinguish between fluid and non-fluid regions. The passage, among which, This indicates that the point belongs to the fluid domain. This indicates that the point belongs to the non-fluid domain; it is used to describe the pore characteristics of different subdomains. The passageway, including the grain pile target area. The porosity of the grain pile is air domain The air porosity is taken as 1.0;

[0057] The scalar condition vector includes: the first Time step, ambient temperature of the left and right walls and The ambient temperature at the top of the roof Controlling the time discretization accuracy of the simulation Sine / cosine terms of daily temperature variation and The sine / cosine terms of the annual temperature cycle and initial temperature initial moisture content ;

[0058] S5. Data cleaning: Based only on the training subset and statistically analyzing the mean and standard deviation of temperature in the fluid domain, construct the z-space incremental learning objective and perform amplitude soft clipping using hyperbolic tangent.

[0059] S6. Train the proxy network, divide the training / test set into groups according to the run identifier and persist the shared split file, use UNet as the backbone and integrate FiLM conditional modulation, calculate the loss in the domain mask and train using dual-domain weighting or only grain pile domain loss, and save the results.

[0060] S7. Autoregressive rolling prediction: The normalized representation of the current temperature field, static multi-channel and scalar conditions are used as inputs. The output temperature increment is updated to the next time step. Rolling prediction is performed in a multi-step manner.

[0061] S8, Nudging assimilation and posterior generation: Linear incremental fitting of the external temperature function is performed within the most recent D-day observation window, and the observation information is calculated and Nudging assimilated with Gaussian kernel weights. The posterior field is calculated and used as the initial state for the next autoregression step.

[0062] S9. Model Evaluation: Output the temperature field, prior / posterior, and absolute error heatmaps in physical coordinates at a proportional scale, and overlay the inner and outer boundary lines to calculate the target area of ​​the grain pile. The MAE, RMSE, nRMSE, PSNR, and SSIM indices are used to evaluate the final assimilated model and prediction results.

[0063] Specifically, the prior and posterior temperature fields are visualized proportionally in physical coordinates and inner and outer boundary lines are superimposed in the grain pile domain. The system calculates metrics such as MAE, RMSE, nRMSE, PSNR, and SSIM to quantitatively evaluate the one-step prediction and rolling sequence. Representative results of the autoregressive rolling prediction accuracy are shown in Table 1.

[0064] Step length MAE RMSE MAPE nRMSE PSNR SSIM +1 day 0.02 0.03 0.08 0.30 45.8 0.985 +3 days 0.12 0.16 0.45 1.50 39.6 0.967 +7 days 0.35 0.46 1.30 4.20 34.1 0.945 +14 days 1.10 1.42 3.90 12.50 28.2 0.910

[0065] Table 1 shows representative results of the accuracy of autoregressive rolling prediction.

[0066] In this embodiment, step S1, modeling and discretization, includes: establishing continuity, momentum, energy, and moisture migration equations based on local thermal equilibrium and the finite volume method, with the boundary condition being bottom adiabatic and the boundary temperature of the top and sun-facing sidewall being... The shady side wall is ,in, The temperature difference caused by solar radiation is due to the difference, and , Given Kelvin, and the external temperature function is: ,in, The term represents the annual average temperature. and These are the annual cycle amplitude and the daily cycle amplitude, respectively. and These are the angular frequencies of the annual cycle and the daily cycle, respectively. , , In radians per second, For time, the unit is , and These are the phase parameters for annual temperature variation and daily temperature variation, respectively.

[0067] Specifically, using the local thermal equilibrium theory and the finite volume method, this study considers the hygroscopic and dehygroscopic characteristics of grains during storage, the effects of respiration and temperature gradients, and the natural convection effect. A multi-field coupled mathematical model of thermal and moisture coupling during the natural storage of grain piles is established. Comsol software is then used to create a two-dimensional grain silo cross-section under a unified geometry and regular mesh, dividing the target area of ​​the grain pile. With upper air domain ,like Figure 3 As shown, the height of the grain pile ranges from 0 to 6 meters, the upper air height is 6 to 12 meters, and the bottom is insulated, while the side walls and roof are driven by the external temperature.

[0068] In this embodiment, in step S2, the data export and unified parsing rearrangement adopt the minimum interface of coordinate-time-temperature, with the coordinate position being... The time unit is The unit of temperature is The export and parsing formats support at least one of CSV, TXT or HDF5, and the temperature fields of each operating condition are rearranged according to the shared rule grid, and bilinear resampling is performed when necessary to align the grid, while extracting the shared coordinate index.

[0069] Specifically, transient CFD simulations are performed under the same geometric and mesh conditions as in step S1, focusing on the initial temperature. Initial moisture content The amplitude of annual temperature variation Amplitude of daily temperature variation Sunlight temperature difference Thermal conductivity of grain piles The parameters are used to construct a multi-condition set, where, , , , , , After simulation, the coordinate-time-temperature data is exported and the time is uniformly converted to seconds and the temperature to Kelvin. The temperature field of each working condition is mapped to a shared regular grid, and the coordinates are rearranged according to the y-axis first and the x-axis second. If necessary, bilinear resampling is used to align the grid (interpolation error <0.02K). The shared coordinate indexes X and Y and the working condition metadata are extracted for training.

[0070] In this embodiment, the signed distance field function in step S3 is: ,in, To describe spatial points The signed distance field to the boundary of the fluid domain. This is a Euclidean distance transformation. For fluid domain mask The complement, when Time indicates the domain, To achieve the closest distance to the fluid domain boundary, the mask boundary is smoothed by 1 pixel.

[0071] Specifically, a fluid domain mask is generated on the shared mesh. Grain pile area mask Signed distance field As a geometric prior.

[0072] In this embodiment, in step S5, normalization and incremental update satisfy: ,in, This is the updated value of the normalized increment. In time step spatial point The temperature at that location The average temperature. The standard deviation of temperature, To prevent extremely small values ​​where the denominator is zero, the network output increment is pruned to satisfy the following: ,in, This is the incremental output of the pruned network. This is the maximum limit value for the network output increment. It is the hyperbolic tangent function. For the original increment output by the network, the physical temperature is restored to satisfy: ,in, This is the restored physical temperature value.

[0073] Specifically, only in the training subset and Statistical mean temperature within the fluid domain and the standard deviation of temperature Temperature is represented in z-space, and the network learns incremental values. Hyperbolic tangent clipping is used to suppress single-step amplitude, and obvious erroneous frames and isolated high-gradient pixels are gently culled and filtered out to enhance robustness.

[0074] In this embodiment, in step S6, the proxy network adopts a UNet encoder-decoder structure and maps the scalar conditional vector from step S4 to channel-level scaling and offset through the FiLM module at each layer to perform affine modulation based on the FiLM module on the intermediate features. The affine modulation function is: ,in, This refers to the modulated intermediate feature channel. This is the channel scaling factor. This is the intermediate feature channel before modulation. The channel offset factor is used. The loss is calculated within the domain mask and weighted using a dual-domain weighting function, which is: ,in, For the total loss, Here are the loss weighting coefficients, and , Mean square error, The target area for the grain pile. For the air area above the grain pile, training or testing is performed by dividing the entire group by run and persistently sharing the split files.

[0075] Specifically, the training and testing groups are divided into 8 training groups and 2 testing groups based on the run identifier, and the split files are persisted and shared to ensure reproducibility. The UNet encoder-decoder backbone is used, with a 4-layer encoder (32→64→128→256 channels) and a 4-layer decoder (256→128→64→32 channels). Skip connections are set between the encoder and decoder. The output layer uses a 1×1 convolution to generate a single channel. And integrate FiLM conditional modulation to map the scalar conditions of step S4 to channel-level scaling. and offset Then, affine modulation is performed on the intermediate features, the temperature loss is calculated within the domain mask, and the algorithm is trained using either a dual-domain weighted approach or a grain pile domain-only approach. Optimization is achieved using Adam (initial learning rate). Cosine annealing to Batch size 8, training for 80 epochs, dropout=0.1 at the bottleneck.

[0076] In this embodiment, the natural regression update function in step S7 is: ,in, and The first Step and the first spatial point The normalized increment intermediate quantity at the location, For the first Step, Spatial Point Network output increment at the location, For the first Step, Spatial Point The final temperature value at that location.

[0077] Specifically, the normalized representation of the current temperature field and static multi-channel and scalar conditions are used as inputs to generate and update a multi-step rolling prediction sequence. This process is repeated cyclically to form a sequence. When necessary, out-of-domain safety filling and boundary neighborhood smoothing are combined to maintain numerical and physical consistency. The autoregression starting point is set to the mid-to-late stage of storage. Time step Rolling forecast When 3 consecutive steps Approaching saturation value At that time, temporarily halve the step size. Two steps to suppress long-term extrapolation drift, unifying the evaluation grid as the length direction Altitude Only in the grain storage area Statistics, such as Figures 4-7 As shown, with day 150 as the starting day, the MAE on days 1, 3, and 7 of the autoregression were 0.028℃, 0.115℃, and 0.370℃, respectively.

[0078] In this embodiment, the linear increment function in step S8 is: ,in, The sum is the linear increment of the fitted external temperature function. and The coefficients are those of the linear fit. The slope The intercept is... Given the ambient temperature, the function for calculating the observed information is: ,in, For the first Each sensor at time step Observational information, For the first Each sensor at time step The actual observed value, For the mapping operator from the temperature field to the sensor location, In time step ,sensor The temperature field model prediction at the location, with the Nudging assimilation function being: ,in, For the first A sensor at a spatial point The corresponding Gaussian weight kernel, For spatial points to the first Euclidean distance of each sensor's spatial location Let be the spatial scale parameter of the Gaussian kernel, and the posterior field function be: ,in, and At time step spatial point The posterior and prior temperature fields at the location, For spatial points Assimilation gain coefficient at the location, To sum the products of the weights and observation information of all sensors, To sum the weights of all sensors, To prevent the minimum value where the denominator is zero.

[0079] Specifically, firstly, within the most recent 10-day observation window, a linear incremental fit is performed on the external temperature function using least squares to correct for boundary driving and constraints. , The external ambient temperature acting on the left and right walls and The ambient temperature at the top of the roof Then, nudging assimilation is performed at each observation time, the observation information of each sensor is calculated, and the posterior field is obtained by weighted diffusion within the mask domain using Gaussian kernel weights. Next, the assimilation intensity and influence radius are set according to the subdomain, and... As the initial closed loop for the next autoregression step, the sensors are arranged in a 6×4=24-point configuration, horizontally covering 65–90% of the warehouse width (avoiding the 0.5m boundary zone) and vertically covering 55–85% of the grain pile height. The assimilation frequency is twice daily (10:00 and 22:00). The data assimilation results are as follows: Figures 8-10 As shown.

[0080] A temperature field prediction system for multi-field coupled digital twins in grain storage includes:

[0081] The data acquisition and parsing module is used to collect various types of raw data and perform format parsing and preliminary structure processing on them;

[0082] The geometry and prior construction module is used to build geometric models of research objects and establish prior hypotheses by combining domain knowledge or historical experience.

[0083] The conditional encoding module is used to encode geometric, prior, and other conditional information into vectors or features that can be recognized by the deep network, providing conditional constraints for network training and inference.

[0084] The data cleaning module is used to remove outliers, impute missing values, and filter noise from the raw data to improve data quality and ensure the accuracy of subsequent analysis.

[0085] The deep network training module is used to train deep networks with specific functions based on cleaned data and conditional encoding, enabling them to learn patterns and rules in the data.

[0086] The autoregressive inference module is used to derive or predict the dynamic changes of physical quantities step by step in an autoregressive manner using a trained deep network.

[0087] The data assimilation module is used to fuse observation data with model prediction results and correct the model output through algorithms to improve its consistency with reality.

[0088] The evaluation and visualization module is used to quantitatively evaluate model performance and data assimilation effects, and to visually display the results in the form of charts and animations to assist in analysis and decision-making.

[0089] A computer storage medium storing computer-executable instructions, wherein the computer-executable instructions, when executed, implement the temperature field prediction method for multi-field coupled digital twins of grain storage as described in any one of claims 1-8.

[0090] Specifically, this invention employs a dual-layer digital twin architecture that integrates physical simulation and deep learning. It performs multi-condition CFD simulations on a unified geometry and regular mesh, deriving the minimum interface between coordinates, time, and temperature. A fluid domain / grain pile domain mask and a signed distance field (SDF) are constructed as geometric priors. Initial thermal and humidity states, porosity, etc., are integrated into a static multi-channel tensor, and the boundary is parameterized with explicit daily / annual phase and orientation differences. A UNet backbone is used, integrating FiLM feature linear modulation to learn temperature increments in z-space. Hyperbolic tangent soft pruning, domain-weighted loss, and domain projection are combined for stable training. Autoregressive rolling prediction is performed during the inference phase. In the online phase, the external temperature function is first linearly incrementally fitted, then assimilated using Gaussian kernel weights through nudging to generate the posterior field. The closed loop is used for the next autoregression step. A modular system and computer-readable storage medium are provided for implementation, supporting rapid deployment and maintenance at the edge / cloud.

[0091] The above embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Although the invention has been described in detail with reference to the embodiments, those skilled in the art should understand that various combinations, modifications, or equivalent substitutions of the technical solutions of the invention do not depart from the spirit and scope of the invention and should be covered within the scope of the claims of the invention.

Claims

1. A method for predicting temperature field using multi-field coupled digital twins for grain storage, characterized in that, Includes the following steps: S1. Modeling and Discretization: A two-dimensional grain silo cross section is established under a unified geometry and regular grid. The grain pile domain and the upper air domain are divided. Boundary conditions are set for bottom insulation, side walls and roof driven by external temperature. The grain pile / air interface satisfies the continuity of temperature and heat flux. A multi-field coupled model is established by discretizing using local thermal equilibrium and finite volume method. S2. Multi-condition simulation and unified analytical rearrangement: CFD is used to perform transient simulation of multi-property and multi-boundary parameter conditions, and simulation data containing coordinates and time-stamped temperatures are exported. Time is uniformly converted to seconds and temperature to Kelvin. The temperature fields of each condition are rearranged according to a unified rule grid, and the shared coordinate index is extracted. S3. Geometric prior construction: generate fluid domain mask and grain pile domain mask, and calculate the signed distance field (SDF) as geometric prior. S4. Input design, construct a static multi-channel vector containing initial temperature, initial moisture content, SDF, domain mask and porosity, and a scalar condition vector containing daily / annual phase, left and right wall / roof boundary temperatures, time step and boundary parameters. S5. Data cleaning: Based only on the training subset and statistically analyzing the mean and standard deviation of temperature in the fluid domain, construct the z-space incremental learning objective and perform amplitude soft clipping using hyperbolic tangent. S6. Train the proxy network by dividing the training / test set into groups according to the run identifier and persistently sharing the split files. Use UNet as the backbone network and integrate Feature-wise Linear Modulation (FiLM). Calculate the loss within the domain mask and train using dual-domain weighting or only grain pile domain loss. Save the results. S7. Autoregressive rolling prediction: The normalized representation of the current temperature field, static multi-channel and scalar conditions are used as inputs. The output temperature increment is updated to the next time step. Rolling prediction is performed in a multi-step manner. S8, Nudging assimilation and posterior generation: Linear incremental fitting of the external temperature function is performed within the most recent D-day observation window, and the observation information is calculated and Nudging assimilated with Gaussian kernel weights. The posterior field is calculated and used as the initial state for the next autoregression step. S9. Model Evaluation: Output the temperature field, prior / posterior, and absolute error heatmaps in physical coordinates at a proportional scale, and overlay the inner and outer boundary lines to calculate the target area of ​​the grain pile. The mean absolute error (MAE), root mean square error (RMSE), normalized root mean square error (nRMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) are used to evaluate the final assimilated model and prediction results.

2. The temperature field prediction method for multi-field coupled digital twins in grain storage according to claim 1, characterized in that, In step S1, modeling and discretization include: establishing continuity, momentum, energy, and moisture migration equations based on local thermal equilibrium and the finite volume method, with the boundary condition being bottom adiabatic and the boundary temperature of the top and sun-facing sidewalls being... The shady side wall is ,in, The temperature difference caused by solar radiation is due to the difference, and , Given Kelvin, and the external temperature function is: ,in, The term represents the annual average temperature. and These are the annual cycle amplitude and the daily cycle amplitude, respectively. and These are the angular frequencies of the annual cycle and the daily cycle, respectively. , , In radians per second, For time, the unit is , and These are the phase parameters for annual temperature variation and daily temperature variation, respectively.

3. The temperature field prediction method for multi-field coupled digital twins in grain storage according to claim 1, characterized in that, In step S2, the data export and unified parsing rearrangement use the minimum interface of coordinate-time-temperature, with the coordinate position being... The time unit is The unit of temperature is The export and parsing formats support at least one of CSV, TXT or HDF5, and the temperature fields of each operating condition are rearranged according to the shared rule grid, and bilinear resampling is performed when necessary to align the grid, while extracting the shared coordinate index.

4. The temperature field prediction method for multi-field coupled digital twins in grain storage according to claim 1, characterized in that, In step S3, the signed distance field function is: ,in, To describe spatial points The signed distance field to the boundary of the fluid domain. This is a Euclidean distance transformation. For fluid domain mask The complement, when Time indicates the domain, To achieve the closest distance to the fluid domain boundary, the mask boundary is smoothed by 1 pixel.

5. A temperature field prediction method for multi-field coupled digital twins in grain storage according to claim 2, characterized in that, In step S5, normalization and incremental update satisfy the following: ,in, This is the updated value of the normalized increment. In time step spatial point The temperature at that location The average temperature. The standard deviation of temperature, To prevent extremely small values ​​where the denominator is zero, the network output increment is pruned to satisfy the following: ,in, This is the incremental output of the pruned network. This is the maximum limit value for the network output increment. It is the hyperbolic tangent function. For the original increment output by the network, the physical temperature is restored to satisfy: ,in, This is the restored physical temperature value.

6. The temperature field prediction method for multi-field coupled digital twins in grain storage according to claim 5, characterized in that, In step S6, the proxy network adopts a UNet encoder-decoder structure and maps the scalar conditional vector from step S4 to channel-level scaling and offset through the FiLM module at each layer to perform affine modulation based on the FiLM module on the intermediate features. The affine modulation function is: ,in, This refers to the modulated intermediate feature channel. This is the channel scaling factor. This is the intermediate feature channel before modulation. The channel offset factor is used. The loss is calculated within the domain mask and weighted using a dual-domain weighting function, which is: ,in, For the total loss, Here are the loss weighting coefficients, and , Mean square error, The target area for the grain pile. For the air area above the grain pile, training or testing is performed by dividing the entire group by run and persistently sharing the split files.

7. A method for predicting temperature field using multi-field coupled digital twins for grain storage, as described in claim 6, is characterized in that... In step S7, the natural regression update function is: ,in, and The first Step and the first spatial point The normalized increment intermediate quantity at the location, For the first Step, Spatial Point Network output increment at the location, For the first Step, Spatial Point The final temperature value at that location.

8. The temperature field prediction method for multi-field coupled digital twins in grain storage according to claim 1, characterized in that, In step S8, the linear increment function is: ,in, The sum is the linear increment of the fitted external temperature function. and The coefficients are those of the linear fit. The slope The intercept is... Given the ambient temperature, the function for calculating the observed information is: ,in, For the first Each sensor at time step Observational information, For the first Each sensor at time step The actual observed value, For the mapping operator from the temperature field to the sensor location, In time step ,sensor The temperature field model prediction at the location, with the Nudging assimilation function being: ,in, For the first A sensor at a spatial point The corresponding Gaussian weight kernel, For spatial points to the first Euclidean distance of each sensor's spatial location Let be the spatial scale parameter of the Gaussian kernel, and the posterior field function be: ,in, and At time step spatial point The posterior and prior temperature fields at the location, For spatial points Assimilation gain coefficient at the location, To sum the products of the weights and observation information of all sensors, To sum the weights of all sensors, To prevent the minimum value where the denominator is zero.

9. A temperature field prediction system for multi-field coupled digital twins in grain storage, characterized in that, include: The data acquisition and parsing module is used to collect various types of raw data and perform format parsing and preliminary structure processing on them; The geometry and prior construction module is used to build geometric models of research objects and establish prior hypotheses by combining domain knowledge or historical experience. The conditional encoding module is used to encode geometric, prior, and other conditional information into vectors or features that can be recognized by the deep network, providing conditional constraints for network training and inference. The data cleaning module is used to remove outliers, impute missing values, and filter noise from the raw data to improve data quality and ensure the accuracy of subsequent analysis. The deep network training module is used to train deep networks with specific functions based on cleaned data and conditional encoding, enabling them to learn patterns and rules in the data. The autoregressive inference module is used to derive or predict the dynamic changes of physical quantities step by step in an autoregressive manner using a trained deep network. The data assimilation module is used to fuse observation data with model predictions and correct model outputs through algorithms to improve their consistency with reality. The evaluation and visualization module is used to quantitatively evaluate model performance and data assimilation effects, and to visually display the results in the form of charts and animations to assist in analysis and decision-making.

10. A computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which, when executed, implement the temperature field prediction method for multi-field coupled digital twins of grain storage as described in any one of claims 1-8.

Images