Grouting intelligent simulation method and system based on large model

Abstract

The application belongs to the technical field of intelligent grouting simulation, and specifically discloses a grouting intelligent simulation method and system based on a large model, which comprises the following steps: obtaining grouting pressure and grouting flow data monitored on site, obtaining grouting pressure, flow, and grout diffusion form data through grouting experiments, and obtaining multiphase flow field data through numerical simulation; training a teacher model by using preprocessed numerical simulation data, grouting experiment data, and part of the on-site monitoring data; in the calculation domain, for a set of predicted working conditions, adopting a differentiated knowledge distillation strategy, processing non-field data and field data respectively, and migrating the generalization knowledge of the teacher model to a student model; obtaining parameter input under a to-be-tested working condition and inputting the parameter into the trained student model to obtain the prediction results of the grout diffusion field, the pressure field, and the velocity field. The application realizes the requirement of quickly adapting to new working conditions and meets the efficiency requirement of real-time decision-making in grouting engineering.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent grouting simulation, and in particular to a grouting intelligent simulation method and system based on a large model. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] As a key means for foundation reinforcement, seepage prevention treatment and geological disaster prevention, grouting technology has a wide application in the fields of tunnels, mines, buildings and water conservancy projects. In the grouting process, the diffusion behavior of the grout directly affects the construction effect. Due to the concealment and invisibility of the injected medium, accurately predicting the spatial and temporal changes of the grout distribution, flow velocity and pressure is the core requirement for optimizing grouting parameters and improving engineering efficiency. However, due to the involvement of multiple geological media such as pipes, fractures and porous media, as well as complex working conditions such as different grouting pressures, flow rates and medium permeability properties, achieving rapid migration training for new working conditions and accurate prediction combined with physical laws faces major challenges.

[0004] Rapid training and accurate prediction are both important goals of intelligent grouting simulation. In actual engineering, working conditions change frequently, not only including the switching of different geological media, but also involving the adjustment of grouting point positions, the change of grouting parameters, the fluctuation of geological conditions and the diversity of scales such as point dynamics, regional diffusion and global flow field under the same working condition. Existing models cannot be directly used for prediction of new working conditions and need to be retrained, which is time-consuming and inefficient, making it difficult to meet real-time decision-making requirements.

[0005] In addition, traditional models require a large amount of new working condition data for training. In the face of frequent changes in grouting position, parameters, geological conditions and scale size, training is time-consuming and inefficient, making it difficult to quickly adapt to the prediction needs of sparse data on site. SUMMARY

[0006] To solve the above problems, the present application proposes a grouting intelligent simulation method and system based on a large model, which learns deeply on grouting data under multiple scenarios and multiple environmental conditions to build a teacher model with strong generalization ability. For specific prediction working conditions, a knowledge distillation strategy is introduced to extract a lightweight student model, and combined with a transfer learning mechanism, multi-source grouting data, physical constraint models and cross-scale consistency loss functions are fused to significantly improve the accuracy and physical consistency of multi-scale prediction, realizing intelligent and efficient simulation of grouting diffusion behavior under different working conditions.

[0007] In some embodiments, the following technical solutions are adopted:

[0008] A grouting intelligent simulation method based on a large model, comprising:

[0009] The grouting pressure and grouting flow data obtained by grouting site monitoring, the grouting pressure, flow and grout diffusion form data obtained by grouting experiment, and the multiphase flow field data obtained by numerical simulation are acquired;

[0010] The acquired data is divided into field data and non-field data, and preprocessed;

[0011] The preprocessed numerical simulation data, grouting experiment data and part of the site monitoring data are used to train a teacher model;

[0012] In the calculation domain, for the set prediction working condition, a differentiated knowledge distillation strategy is adopted to process the non-field data and the field data respectively, and the generalization knowledge of the teacher model is migrated to the student model;

[0013] The parameters under the to-be-tested working condition are acquired and input to the trained student model to obtain the prediction results of the grout diffusion field, pressure field and velocity field.

[0014] As a further scheme, the grouting pressure and grouting flow data obtained by grouting site monitoring and the grouting pressure and flow data obtained by grouting experiment are taken as non-field data, and the grout diffusion form data obtained by grouting experiment and the multiphase flow field data obtained by numerical simulation are taken as field data.

[0015] As a further scheme, after the acquired data is divided into field data and non-field data and preprocessed, it further includes: extracting time sequence statistical features from the non-field data, extracting spatial distribution features from the field data, forming a feature vector, inputting the feature vector into a trained classification model, and outputting a working condition label probability distribution to determine the working condition type to which the current data belongs.

[0016] As a further scheme, the teacher model is a deep neural network including a time sequence prediction branch, a regional feature extraction branch and a global flow field prediction branch; the time sequence prediction branch processes the non-field data through a full connection layer and a time attention mechanism to predict pressure and flow; the regional feature extraction branch extracts image features in the field data through a two-dimensional convolution layer and a spatial attention mechanism to predict phase fraction distribution; and the global flow field prediction branch predicts grid features in the field data through a three-dimensional convolution layer and a global attention mechanism to predict phase fraction, velocity and pressure.

[0017] As a further scheme, the loss function of the teacher model is:

[0018] ;

[0019] wherein, is the total loss function of the teacher model, is the time sequence mean square error loss for the non-field data, To address the pixel-level mean square error loss in field data, To address the physical mean square error loss of the grid in the field data, This is a cross-scale consistency loss used to bridge point, regional, and global scales; , , and These are the weights for each loss.

[0020] As a further solution, a dynamic weight adaptive mechanism is adopted for weight updates, specifically as follows:

[0021] ;

[0022] in, To adjust the rate, The output of the classifier is the probability of the working condition label. The verification error corresponding to the loss component, Indicates the current iteration step The weight, The updated weights are represented by j, which is an integer index variable representing the position of the loss component whose weight is being adjusted in the total loss function. k represents the auxiliary index in the summation. The indexes of all loss components are traversed, ranging from 1 to the total number of loss terms. The sum is calculated by traversing all terms.

[0023] As a further solution, for the given prediction conditions, a differentiated knowledge distillation strategy is adopted to process non-field data and field data separately, specifically as follows:

[0024] For time series prediction of non-field data, the teacher model generates the predicted probability distributions of pressure and flow as soft labels, and optimizes the learning of soft labels through KL divergence loss.

[0025] To address the sparsity and uncertainty of time-series data in the grouting process, a time-series uncertainty loss is introduced, and variational inference is used to estimate the uncertainty of the prediction distribution, penalizing regions with high uncertainty. Simultaneously, the mean square error loss is calculated by combining sparse real data of the target working condition.

[0026] For the field data images, the teacher model generates soft labels for the phase fraction distribution, capturing the diffusion boundary and velocity pattern, while the student model optimizes the KL divergence loss; pixel-level mean square error loss is calculated by combining real data from sparse region images; at the same time, a geological heterogeneity sensitivity loss is introduced to penalize the model's prediction bias for heterogeneous regions.

[0027] For the gridded field data, the teacher model generates spatiotemporal soft labels for phase fractions, velocity, and pressure, including eddy current and pressure gradient features, while the student model optimizes the KL divergence loss; at the same time, it calculates the mean square error of the physical field by combining real data from some grid points.

[0028] As a further approach, the generalized knowledge from the teacher model is transferred to the student model, specifically as follows:

[0029] Weights are extracted from the time-series prediction branch, regional feature extraction branch, and global flow field prediction branch of the teacher model and used as initial parameters for the student model. These parameters are then fine-tuned for the target operating condition characteristics to ensure the model adapts to the dynamic and distributional features of the specific operating condition. The specific process is as follows:

[0030] For off-field data, transfer learning uses sparse target working condition data to optimize the time series prediction branch and adjusts the weights through mean squared error loss.

[0031] For images in the field data, transfer learning optimizes the regional feature extraction branch using the target working condition image and refines the diffusion boundary prediction through pixel-level mean square error.

[0032] For the grid in the field data, transfer learning uses partial grid point data to optimize the global flow field prediction branch and adjusts the flow field dynamics through the physical field mean square error.

[0033] As a further step, transfer learning also incorporates cross-scale consistency loss and geological heterogeneity sensitivity loss to improve adaptability across different working conditions; it also introduces temporal uncertainty loss to handle the sparsity of temporal data; and it introduces adversarial training, using a discriminator to optimize feature alignment loss to ensure that student models adapt to the feature distribution of new working conditions.

[0034] In other embodiments, the following technical solutions are adopted:

[0035] A large-model-based intelligent simulation system for grouting includes:

[0036] The data acquisition module is configured to acquire grouting pressure and grouting flow data from on-site grouting monitoring, grouting pressure, flow rate, and grout diffusion morphology data obtained through grouting experiments, and multiphase flow field data obtained through numerical simulation.

[0037] The preprocessing module is configured to divide the acquired data into field data and non-field data, and then perform preprocessing.

[0038] The teacher model training module is configured to train the teacher model using preprocessed numerical simulation data, grouting experiment data, and some field monitoring data.

[0039] The transfer learning module is configured to: within the computational domain, for a given prediction condition, employ a differentiated knowledge distillation strategy to process non-field data and field data separately, transferring the generalized knowledge of the teacher model to the student model.

[0040] The prediction module is configured to: obtain the parameters of the test condition and input them into the trained student model to obtain the prediction results of the slurry diffusion field, pressure field and velocity field.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] (1) The knowledge distillation mechanism designed in this invention significantly improves the model training efficiency. By extracting a lightweight student model from a complex teacher model, and using a small amount of target working condition data for rapid training, the time cost of retraining is reduced, and the need to quickly adapt to new working conditions (such as changes in the location of grouting points and heterogeneity of geological conditions) is realized, thus meeting the efficiency requirements of real-time decision-making in grouting projects.

[0043] (2) The transfer learning technique designed in this invention enhances the model prediction accuracy. By reusing the pre-trained weights of the teacher model and fine-tuning them with a small amount of target working condition data, transfer learning deeply adapts to the dynamic and distribution characteristics of specific working conditions, ensuring that the prediction results can accurately reflect the temporal dynamics (such as pressure change), regional diffusion (such as boundary morphology), and global flow field (such as velocity component distribution) in the grouting process, thereby improving the engineering reliability of the prediction.

[0044] (3) The teacher model of this invention learns general grouting characteristics (such as temporal dynamics, regional diffusion, and global flow field law) across working conditions through pre-training of multi-source data (simulation, laboratory, field). The student model inherits the generalization ability and adapts to diverse working conditions through knowledge distillation and transfer learning for collaborative optimization, thus achieving efficient prediction under different geological media (such as pipelines, fractures, and porous media), effectively enhancing the adaptability of the model.

[0045] (4) This invention innovatively designs loss functions that are consistent across scales, sensitive to geological heterogeneity, and have temporal uncertainty. Combined with a dynamic weight adaptive mechanism, it optimizes the model for the heterogeneity, data sparsity, and multi-scale characteristics of grouting conditions, ensuring that the prediction results conform to the laws of fluid mechanics and significantly improving the prediction accuracy of grouting effect.

[0046] (5) The present invention introduces an adaptive working condition classifier to automatically identify geological media (such as fissures or pores) and grout types (such as high-viscosity cement grout), dynamically adjust model weights and training strategies, reduce manual intervention, and improve the efficiency of grouting deployment in complex geological environments.

[0047] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0048] Figure 1 This is a flowchart of the intelligent simulation method for grouting based on a large model in an embodiment of the present invention. Detailed Implementation

[0049] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0050] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0051] Example 1

[0052] In one or more embodiments, a grouting intelligent simulation method based on a large model is disclosed, combining... Figure 1 Specifically, it includes the following processes:

[0053] S101: Acquire grouting pressure and flow rate data from on-site grouting monitoring, grouting pressure, flow rate, and grout diffusion morphology data obtained through grouting experiments, and multiphase flow field data obtained through numerical simulation.

[0054] In this embodiment, grouting-related multiphase flow field data are collected from three sources, specifically:

[0055] Data acquisition at the site: Data is collected from the site using sensors installed on the grouting equipment, mainly including grouting pressure ( ) and grouting flow rate ( (e.g., pressure sensors and flow meters are used to monitor grouting parameters in real time. The sampling frequency is set according to actual needs to generate sparse time series point data. Each data point contains a timestamp and the corresponding pressure and flow values.)

[0056] Data acquisition for laboratory experiments: Grouting-related parameters and diffusion behavior are recorded using experimental monitoring elements, including grouting pressure and flow rate data (similar to field monitoring, but may vary due to differences in experimental environment), as well as the diffusion morphology of the grout in the medium; for pipe and fractured media, time-series photographs are taken using high-speed cameras to capture the dynamic diffusion of the grout; for three-dimensional porous media, resistivity inversion technology is used to generate time-series image data of grout diffusion, with the image resolution set according to experimental requirements.

[0057] Generation of numerical simulation data: Numerical simulations are performed using computational fluid dynamics software (such as COMSOL Multiphysics) to generate multiphase flow field data containing high-resolution meshes. The simulation results include scalar fields (such as phase fractions). and pressure ) and vector fields such as The grid resolution is set according to computing power to cover the global flow field characteristics during the grouting process.

[0058] S102: Divide the acquired data into field data and non-field data, and perform preprocessing.

[0059] In this embodiment, preliminary classification is performed based on data characteristics and presentation. Grouting pressure and flow rate from field monitoring data, as well as pressure and flow rate from laboratory test data, are classified as non-field data because they are discrete time-series point data lacking spatial distribution information. Time-series photographs and resistivity inversion images from laboratory test data, as well as numerical simulation data, are classified as field data because they provide continuous spatial distribution information in image or grid form. The classified dataset is then output.

[0060] In this embodiment, a lightweight condition classification method is proposed to achieve adaptive modeling and response adjustment for grouting conditions. This method extracts time-series statistical features (such as the mean, variance, and spectrum of pressure and flow rate) from non-field data, extracts spatial distribution features (such as grayscale texture and permeability histogram) from field data, and constructs a unified format feature vector [number of samples, 128] as input to the classification model.

[0061] Specifically, the classification model can employ existing lightweight neural networks, such as MobileNetV2, combined with a single-head self-attention mechanism. The input is a feature vector, and the output is a probability distribution of the working condition label (categories include pipes, cracks, porous media, and grout types such as cement grout, chemical grout, cement-water glass grout, and polyurethane grout). The training process is optimized using cross-entropy loss to ensure cross-validation accuracy ≥ 95%.

[0062] Cross-entropy loss Specifically:

[0063] ;

[0064] in, The total number of samples, For sample indexes (range 1 to N). For working condition category index, For the first Real media type of each sample The label (0 or 1). The model predicts the first Each sample belongs to category The probability (from 0 to 1).

[0065] Generate a probability distribution [number of samples, 7] for the new data, and normalize it using softmax:

[0066] ;

[0067] The model branch weights are adjusted according to the probability distribution. The model branches include: time series prediction branch: processing non-field data (such as grouting pressure and flow time series) and predicting pressure and flow; regional feature extraction branch: processing images in field data and predicting phase fraction distribution; global flow field prediction branch: processing grids in field data and predicting phase fraction, velocity and pressure distribution.

[0068] The specific mapping formula for the model branch weights is as follows:

[0069] ;

[0070] in, Preset the initial weights (using empirical values ​​or uniform distribution, such as 1 / 5). The pre-trained mapping coefficients (determined through validation set optimization) reflect the operating conditions. Weights of loss components The impact of weight normalization. .

[0071] Let represent the probability distribution of the c-th working condition category; The unnormalized score of the output layer of the classification neural network for the c-th media type is generated by the last fully connected layer of the neural network. This represents the unnormalized score of the output layer of the classification neural network for the j-th working condition category, covering all categories to calculate the sum; This represents the weight coefficient of the k-th model branch (time series prediction, regional feature extraction, global flow field prediction); This represents the mapping coefficient of the k-th branch to the c-th working condition medium category.

[0072] The output of the classification model is used to guide the weight adjustment for subsequent knowledge distillation and transfer training.

[0073] In this embodiment, the data preprocessing process is as follows:

[0074] (1) Preprocessing of off-field data: Outliers were removed from off-field data (pressure and flow rate in the field and laboratory). The outlier range was calculated using a box plot method (below the lower quartile minus the weighted interquartile range, or above the upper quartile plus the weighted interquartile range). Data points exceeding the range were checked and removed point by point, and the reasons for removal were recorded. Subsequently, the time series was enriched by cubic spline interpolation. A cubic polynomial function was constructed with the timestamp as the independent variable and grouting pressure and flow rate as the dependent variables. The coefficients were calculated and uniformly sampled to generate the enriched time series. Then, the spatial coordinates (x, y, z) of the grouting point were associated with each time point to form a tensor [number of samples, number of sample time points, 6], where 6 corresponds to the temporal features (time, coordinates, and grouting parameters). Finally, normalization was performed, the mean and standard deviation were calculated, and the output tensor was standardized.

[0075] Preprocessing of field data: For time-series photographs (pipeline / fractured media) in the experiment, RGB images are first converted to grayscale images, and the slurry diffusion region is extracted (through threshold segmentation, e.g., grayscale values ​​exceeding a certain preset threshold). Then, the data is resampled to a uniform grid resolution (resolution determined by computational capability) and the pixel values ​​are normalized to [0, 1]. For resistivity inversion images (porous media), the resistivity field is first converted to a phase fraction approximation (through a predefined mapping relationship, such as α = f(resistivity)), then normalized and adjusted to a uniform grid resolution. For simulation data, scalar fields (α, p) and vector fields (u, v, w) are first extracted, retaining the velocity components in the three directions. Then, the data is resampled to a uniform grid and normalized to maintain consistency with non-field data. Finally, all field data is organized into a tensor format [number of samples, number of time steps, height, width, depth, number of channels], where height, width, and depth correspond to the grid dimensions in the x, y, and z directions, and the number of channels corresponds to physical quantities (such as α, u, v, w, p).

[0076] Where α, u, v, w, and p represent the phase fraction, the x-direction of the velocity component, the y-direction of the velocity component, the z-direction of the velocity component, and the pressure, respectively.

[0077] After preprocessing, feature extraction is performed, specifically as follows:

[0078] For time-series photographs and resistivity images, multi-size 2D convolutional kernels (such as 3×3, 5×5, stride 1, padding-preserving size) are applied to generate hierarchical feature maps to capture the spatial patterns of slurry diffusion (such as boundaries and diffusion velocity). The channels are expanded through depthwise separable convolution to reduce computation and output tensors [number of samples, number of time steps, spatial resolution width, spatial resolution height, multiple channels], such as 64 channels.

[0079] For simulated mesh data, multi-size 3D convolutional kernels (such as 3×3×3, 5×5×5, stride 1, fill-and-hold size) are applied to generate global flow field features, capturing the spatial distribution and dynamic changes of phase fraction, velocity, and pressure; the output tensor [number of samples, number of time steps, mesh resolution width, mesh resolution height, mesh resolution depth, multi-channel] is expanded through depthwise separable convolution, with the number of multi-channels being 64.

[0080] S103: Train the teacher model using preprocessed numerical simulation data, grouting experiment data, and some field monitoring data.

[0081] In this embodiment, the teacher model is a deep neural network comprising a temporal prediction branch, a regional feature extraction branch, and a global flow field prediction branch. The temporal prediction branch processes non-field data through fully connected layers and a temporal attention mechanism to predict pressure and flow rate; the regional feature extraction branch extracts image features from the field data through two-dimensional convolutional layers and a spatial attention mechanism to predict phase fraction distribution; and the global flow field prediction branch predicts grid features in the field data through three-dimensional convolutional layers and a global attention mechanism to predict phase fraction, velocity, and pressure.

[0082] The teacher model is a large model with a high number of parameters (e.g., millions of parameters) to capture the characteristics of complex operating conditions. The large model is pre-trained using source domain data, including simulation grid data, laboratory images and off-site time-series data, and some field time-series data. To balance the multi-source data, modules are sampled proportionally (e.g., simulation:laboratory:field = 5:3:2) to ensure the model's adaptability to different data distributions.

[0083] The loss of the teacher model includes:

[0084] (1) For non-field data time series mean square error loss :

[0085] ;

[0086] (2) For image pixel-level mean square error loss in field data :

[0087] ;

[0088] (3) For the grid in the field data Physical field mean square error loss :

[0089] ;

[0090] in, Indicates the first The predicted grouting pressure value for each sample. Indicates the first The actual grouting pressure value of each sample Indicates the first The predicted grouting flow rate for each sample. This represents the actual grouting flow rate value of the nth sample. Indicates the image height. Indicates the image width. Indicates the grid depth. Indicates the first in the image grid Line 1 The phase fraction values ​​predicted by column position. Indicates the first in the image grid Line 1 The actual phase fraction value of the column position. Indicates the first in the grid Altitude Width Physics values ​​for depth location prediction They represent the first Altitude Width The actual physical field values ​​at depth location.

[0091] (4) To bridge point, regional and global scales, this embodiment introduces cross-scale consistency loss. This method enforces consistency in model predictions across different scales (e.g., matching time-series pressure with the global pressure field). It improves the consistency, generalization ability, and accuracy of multi-scale predictions by sampling multi-scale points to calculate biases. This is defined as:

[0092] ;

[0093] in, Indicates the first Each time series sample stress value Indicates the first The stress value of a sample at the global scale (usually referring to a grid). Indicates the first Predicted regional phase scores for each sample Indicates the first Global phase score prediction for each sample.

[0094] In summary, the total loss function of the teacher model is as follows:

[0095] ;

[0096] Among them, weight to The system is dynamically adjusted based on the output of the adaptive operating condition classifier, for example, increasing the output in high-porosity media conditions. To enhance global flow field learning and increase performance under high-viscosity slurry conditions. To enhance cross-scale consistency.

[0097] As an optional implementation, the weight update adopts a dynamic weight adaptive mechanism, which is automatically optimized through gradient analysis, specifically as follows:

[0098] ;

[0099] in, To adjust the rate (value range 0.01 to 0.1), The output of the classifier is the probability of the working condition label. To address the verification error of the loss components, the adjustment process ensures that the weights match the characteristics of the working conditions (such as slurry viscosity or geological heterogeneity), thereby improving model adaptability, optimizing resource allocation, enhancing prediction accuracy, reducing manual intervention, and improving robustness; the formula considers the adjustment rate. Operating condition probability and verification error The dynamic combination of factors, especially considering the diversity of working conditions and error feedback, affects the convergence speed of the calculation results and the targeted allocation of weights, ensuring efficient optimization of the model under complex grouting conditions.

[0100] Finally, the trained teacher model is tested across different working conditions to verify its generalization ability at various grouting point locations and geological conditions (such as fractures and porous media), ensuring the reliability of the prediction tensor. The module generates a prediction tensor containing the number of samples, time steps, and multiple physical quantity channels, learns general features across working conditions (such as temporal dynamics, regional diffusion, and global flow field patterns), and extracts key working condition features (such as high-concentration areas at grouting points and diffusion boundaries) for subsequent distillation, outputting pre-trained weights and prediction tensors.

[0101] S104: Within the computational domain, for the given prediction conditions, a differentiated knowledge distillation strategy is adopted to process non-field data and field data separately, transferring the generalized knowledge of the teacher model to the student model.

[0102] In this embodiment, within the computational domain, for a specific prediction condition, a lightweight student model is extracted through knowledge distillation using the weights and prediction tensors of the aforementioned pre-trained teacher model, and rapid training is achieved using a small number of new conditions.

[0103] Specifically, the student model is designed with a lightweight structure to efficiently handle non-field data (such as grouting pressure and flow time series) and field data (such as phase fraction, velocity, and pressure distribution), ensuring rapid training and accurate prediction for specific working conditions (such as new grouting point locations, parameter configurations, or changes in geological conditions). Compared to the high number of parameters in the teacher model, the student model significantly reduces computational requirements by simplifying the network structure while retaining core predictive capabilities.

[0104] The student model also includes three functional branches: a time-series prediction branch, a region feature extraction branch, and a global flow field prediction branch.

[0105] The time series prediction branch processes non-field data, reducing it to... A fully connected network of layers (compared to the teacher model) (The layers are significantly simplified), the number of neurons per layer is reduced from Down to A lightweight single-head self-attention mechanism is introduced (the number of channels is reduced from...). Down to This allows for the efficient capture of dynamic characteristics of pressure and flow, such as abrupt changes in the location of the grouting point.

[0106] The region feature extraction branch processes images in the field data, simplifying them to... Two-dimensional convolution (teacher model is) (layers), number of convolutional kernels from Down to It combines depthwise separable convolution to optimize computational cost, preserves spatial attention mechanism to extract phase fractional diffusion boundaries, and adapts to changes in geological conditions.

[0107] The global flow field prediction branch processes the grid in the field data and compresses it to a P-layer 3D convolution (the teacher model is...). (layers), number of channels from Down to Grouped convolution is used to reduce the number of parameters, while global attention mechanism is retained to predict phase fraction, velocity and pressure distribution, ensuring the integrity of flow field characteristics.

[0108] In this embodiment, for specific prediction conditions (such as new grouting point locations, pressure and flow parameters, or changes in geological conditions), a differentiated knowledge distillation strategy is adopted in the computational domain to process non-field data and field data separately, efficiently transferring the generalized knowledge of the teacher model to the student model, ensuring a balance between rapid training and accurate prediction.

[0109] Specifically, the process of the knowledge distillation strategy is as follows:

[0110] (1) For time series prediction of off-field data, the teacher model generation pressure and traffic The predicted probability distribution (the confidence level of the teacher model in predicting pressure or flow at each time point) is used as a soft label. Soft label learning is optimized using KL divergence loss, and the loss function is... Defined as:

[0111]

[0112] in, The temperature parameter is dynamically adjusted to balance the smoothness of the distribution. The teacher model represents the first The probability distribution vector of the predicted pressure or flow rate for each sample (time point). Represents the student model for the first The probability distribution vector of the predicted pressure or flow rate for each sample (time point).

[0113] (2) To address the sparsity and uncertainty of time-series data in the grouting process, a time-series uncertainty loss is introduced. Using variational inference to estimate the uncertainty of the predicted distribution and penalize high uncertainty regions (such as stress mutation points) can improve the robustness of the prediction, enhance the model's perception of uncertainty, optimize the utilization of sparse data, balance knowledge distillation with fitting to real data, and improve engineering practicality, especially in terms of safety during critical operating condition changes (such as stress mutations).

[0114] Specifically defined as:

[0115]

[0116] in, This indicates the predicted pressure or flow rate. Estimated by the additional output head of the student model, This is the actual value.

[0117] (3) Calculate the mean square error loss by combining the sparse real data of the target working condition. To ensure accurate fitting; for working conditions such as changes in the location of grouting points, the distillation weight is dynamically increased to prioritize the learning of key temporal features, such as pressure mutation points, and the perceptual ability of the student model is enhanced through a time attention mechanism.

[0118] Mean square error loss Specifically:

[0119] .

[0120] (4) For the field data image, the teacher model generates soft labels for the phase fraction distribution, captures the diffusion boundary and velocity pattern, and the student model optimizes the KL divergence loss:

[0121] ;

[0122] in, The teacher model is represented in the image grid. Line 1 The probability distribution of the phase fraction values ​​(range 0 to 1) predicted for column positions. This indicates that the student model is in the image grid at the 1st position. Line 1 The probability distribution of the phase values ​​(range 0 to 1) predicted by column position.

[0123] (5) Calculate the pixel-level mean square error loss by combining the real data of the sparse region image:

[0124] ;

[0125] in, This indicates that the student model is in the image grid at the 1st position. Line 1 The phase fraction values ​​predicted by column position. This indicates that the student model is in the image grid at the 1st position. Line 1 The actual phase fraction values ​​of the column positions.

[0126] (6) To address the heterogeneity of the grouting medium, a geological heterogeneity-sensitive loss is introduced to penalize the model's prediction bias for heterogeneous regions. This is achieved by calculating the spatial gradient weighted error of the phase fraction α, thereby improving the prediction accuracy of heterogeneous regions, enhancing the model's robustness, and increasing engineering reliability. This is defined as:

[0127] ;

[0128] in, Spatial gradients are calculated using automatic differentiation. Represents the student model for the first The phase fraction value predicted for each sample (usually referring to a point in an image or grid). Indicates the first The true phase fraction values ​​of each sample.

[0129] (7) In response to changes in geological conditions such as the heterogeneity of porous media, priority should be given to distillation diffusion boundary characteristics, and key areas should be focused on through spatial attention mechanisms.

[0130] (8) For the grid of field data, the teacher model generates spatiotemporal soft labels for phase fractions, velocity, and pressure, including eddy current and pressure gradient features, while the student model optimizes the KL divergence loss:

[0131] ;

[0132] in, , Indicates the teacher model in the grid. Altitude Width The probability distribution of the physics field values ​​predicted by depth location. Indicates the student model in the grid. Altitude Width The probability distribution of physical field values ​​predicted by depth location.

[0133] (9) Calculate the mean square error loss of the physical field by combining the real data of some grid points:

[0134] ;

[0135] in, Indicates the student model in the grid. Altitude Width Physics values ​​for depth location prediction Indicates the grid number Altitude Width The actual physical field values ​​at depth location.

[0136] For high-velocity operating conditions, the weights are dynamically adjusted to enhance velocity component learning, and the dynamic capture of the flow field is improved through a global attention mechanism.

[0137] Taking into account the above losses, the optimized total loss function is as follows:

[0138]

[0139] Among them, weight to The output is dynamically adjusted based on the adaptive operating condition classification results, for example, increasing the output under high-viscosity slurry conditions. To enhance heterogeneous learning and increase performance in sparse data scenarios. To improve timing robustness.

[0140] As an optional implementation, the weight update also adopts a dynamic weight adaptive mechanism, which is automatically optimized through gradient analysis. The specific method is the same as described above and will not be detailed here.

[0141] Finally, the trained teacher model is tested across different working conditions to verify its generalization ability at various grouting point locations and geological conditions (such as fractures and porous media), ensuring the reliability of the prediction tensor. The module generates a prediction tensor containing the number of samples, time steps, and multiple physical quantity channels, learns general features across working conditions (such as temporal dynamics, regional diffusion, and global flow field patterns), and extracts key working condition features (such as high-concentration areas at grouting points and diffusion boundaries) for subsequent distillation, outputting pre-trained weights and prediction tensors.

[0142] Specifically, the process of transfer learning training is as follows:

[0143] Transfer learning extracts weights from the time-series prediction branch, regional feature extraction branch, and global flow field prediction branch of the teacher model, which are used as the initial parameters of the student model. The model is then fine-tuned for the characteristics of the target working condition (such as new grouting point locations, pressure and flow parameter adjustments, or changes in geological conditions) to ensure that the model adapts to the dynamic and distribution characteristics of the specific working condition.

[0144] (1) For non-field data, such as grouting pressure and traffic For time series analysis, transfer learning utilizes sparse target working condition data (such as finite time series points at the location of new grouting points) to optimize the time series prediction module, adjusting the weights through mean squared error loss.

[0145] ;

[0146] First, fit key dynamic features, such as pressure changes or flow fluctuations caused by changes in the grouting point location, and then enhance the model's ability to capture these features through a time attention mechanism.

[0147] (2) For images in field data, such as phase fractions The regional distribution is analyzed, and transfer learning utilizes a small number of target condition images (such as sparse local diffusion regions) to optimize the regional feature extraction module. The diffusion boundary prediction is refined through pixel-level mean square error.

[0148] ;

[0149] Specifically targeting changes in geological conditions (such as the heterogeneity of porous media), we enhance boundary feature learning and improve our focus on key diffusion areas through spatial attention mechanisms.

[0150] (3) For the grid in the field data, such as the phase fraction, velocity component, and pressure global flow field, transfer learning utilizes partial grid point data to optimize the global flow field prediction module, through the physical field mean square error:

[0151] ;

[0152] Adjust the flow field dynamics, prioritize the optimization of velocity components for high-velocity conditions, and enhance the capture of flow field features through a global attention mechanism.

[0153] (4) To further improve cross-condition adaptability, cross-scale consistency loss and geological heterogeneity sensitivity loss are introduced to enhance the coordination of multi-scale predictions, improve model generalization ability, and increase prediction accuracy. These losses are defined as follows:

[0154] ;

[0155] in, Indicates the first Predicted point pressure values ​​for each sample. Indicates the first Global stress prediction for each sample. Indicates the first Predicted regional phase scores for each sample Indicates the first Global phase score prediction for each sample. Indicates the first Spatial gradient of each sample phase fraction Indicates the first The predicted phase fraction values ​​for each sample. Indicates the first The true phase fraction values ​​of each sample.

[0156] (5) To address the sparsity of time-series data, a time-series uncertainty loss is introduced, defined as:

[0157] ;

[0158] in, This indicates the predicted pressure or flow rate. Estimated by the additional output head of the student model, This is the actual value.

[0159] (6) To bridge the distribution differences between the source domain (simulation and laboratory data) and the target domain (sparse field data or new working conditions), the transfer learning method introduces adversarial training and uses a discriminator to optimize the feature alignment loss, ensuring that the student model adapts to the feature distribution of the new working conditions:

[0160] ;

[0161] in, A value close to 1 indicates that the discriminator considers the feature to originate from the target domain. This indicates that the discriminator D considers the features of the source domain. The output probability, It represents the mathematical expectation.

[0162] To address data sparsity, the method incorporates a pseudo-label generation strategy, utilizing a teacher model to predict and fill in missing data. For example, time series points in non-field data are filled in using predicted values ​​through interpolation, while grid points in field data are filled in using the teacher model to generate auxiliary training samples. Simultaneously, data augmentation is introduced, such as temporal random perturbations, image flipping, or slight grid point shifts, to expand the sample size and improve model robustness. The optimization process employs mini-batch gradient descent, with the batch size adjusted according to the data volume to ensure rapid convergence and generate a prediction tensor containing temporal, regional, and global features.

[0163] The total loss is calculated by weighting all losses together.

[0164] ;

[0165] in, to The output of the adaptive operating condition classifier is dynamically adjusted, for example, by increasing the output under highly heterogeneous operating conditions. Enhance geological heterogeneity learning to increase performance in sparse data scenarios. To enhance competitive training.

[0166] The weight update also adopts a dynamic weight adaptive mechanism, and the specific method is the same as described above.

[0167] Transfer learning is automatically triggered based on the classifier output P(c). For example, the initial weights that match the working conditions are selected through P(c) (such as prioritizing the global flow field branch weights for high-viscosity slurry) and the learning rate is adjusted (such as reducing it to 0.001 for stable optimization under porous media conditions).

[0168] S105: Input the parameters under the test condition into the trained student model to obtain the prediction results of the slurry diffusion field, pressure field and velocity field.

[0169] The parameters to be obtained under the test conditions include: the three-dimensional location of the grouting point, the type of geological medium, the type of grout, the sand content, the water inflow, the hydrodynamic data (water level, flow velocity, water pressure) and the hydrochemical data (ion concentration, pH value, redox potential), etc.

[0170] Within the computational domain, the target operating condition parameters are input into the optimized student model to predict the temporal variations of non-field data and the spatial distribution of field data, generating a prediction tensor that includes pressure, flow rate, phase fraction, velocity, and pressure distribution.

[0171] Furthermore, time-series curves of non-field data, regional distribution maps and global distribution maps of field data are plotted to demonstrate the predicted results of the slurry diffusion field, pressure field, and velocity field. The visualized images and prediction results are saved as files for easy verification and analysis.

[0172] Example 2

[0173] In one or more embodiments, a grouting intelligent simulation system based on a large model is disclosed, comprising:

[0174] The data acquisition module is configured to acquire grouting pressure and grouting flow data from on-site grouting monitoring, grouting pressure, flow rate, and grout diffusion morphology data obtained through grouting experiments, and multiphase flow field data obtained through numerical simulation.

[0175] The preprocessing module is configured to divide the acquired data into field data and non-field data, and then perform preprocessing.

[0176] The teacher model training module is configured to train the teacher model using preprocessed numerical simulation data, grouting experiment data, and some field monitoring data.

[0177] The transfer learning module is configured to: within the computational domain, for a given prediction condition, employ a differentiated knowledge distillation strategy to process non-field data and field data separately, transferring the generalized knowledge of the teacher model to the student model.

[0178] The prediction module is configured to: obtain the parameters of the test condition and input them into the trained student model to obtain the prediction results of the slurry diffusion field, pressure field and velocity field.

[0179] It should be noted that the specific implementation methods of the above modules are exactly the same as those in Example 1, and will not be described in detail again.

[0180] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A grouting intelligent simulation method based on a large model, characterized in that, include: Data on grouting pressure and flow rate from on-site grouting monitoring, data on grouting pressure, flow rate, and grout diffusion morphology obtained through grouting experiments, and multiphase flow field data obtained through numerical simulation were acquired. The acquired data is divided into field data and non-field data, and then preprocessed. The teacher model was trained using preprocessed numerical simulation data, grouting experiment data, and some field monitoring data. Within the computational domain, for the given prediction conditions, a differentiated knowledge distillation strategy is adopted to process non-field data and field data separately, transferring the generalized knowledge of the teacher model to the student model. The parameters under the test condition are obtained and input into the trained student model to obtain the prediction results of the slurry diffusion field, pressure field and velocity field; Specifically, for the defined prediction conditions, a differentiated knowledge distillation strategy is adopted to process non-field data and field data separately, as follows: For time series prediction of non-field data, the teacher model generates the predicted probability distributions of pressure and flow as soft labels, and optimizes the learning of soft labels through KL divergence loss. To address the sparsity and uncertainty of time-series data in the grouting process, a time-series uncertainty loss is introduced, and variational inference is used to estimate the uncertainty of the prediction distribution, penalizing regions with high uncertainty. Simultaneously, the mean square error loss is calculated by combining sparse real data of the target working condition. For the field data images, the teacher model generates soft labels for the phase fraction distribution, capturing the diffusion boundary and velocity pattern, while the student model optimizes the KL divergence loss; pixel-level mean square error loss is calculated by combining real data from sparse region images; at the same time, a geological heterogeneity sensitivity loss is introduced to penalize the model's prediction bias for heterogeneous regions. For the gridded field data, the teacher model generates spatiotemporal soft labels for phase fractions, velocity, and pressure, including eddy current and pressure gradient features, while the student model optimizes the KL divergence loss; at the same time, it calculates the mean square error of the physical field by combining real data from some grid points. The generalized knowledge from the teacher model is transferred to the student model, specifically as follows: Weights are extracted from the time-series prediction branch, regional feature extraction branch, and global flow field prediction branch of the teacher model and used as initial parameters for the student model. These parameters are then fine-tuned for the target operating condition characteristics to ensure the model adapts to the dynamic and distributional features of the specific operating condition. The specific process is as follows: For off-field data, transfer learning uses sparse target working condition data to optimize the time series prediction branch and adjusts the weights through mean squared error loss. For images in the field data, transfer learning optimizes the regional feature extraction branch using the target working condition image and refines the diffusion boundary prediction through pixel-level mean square error. For the grid in the field data, transfer learning uses partial grid point data to optimize the global flow field prediction branch and adjusts the flow field dynamics through the physical field mean square error.

2. The intelligent simulation method for grouting based on a large model as described in claim 1, characterized in that, The grouting pressure and flow rate data monitored at the grouting site, as well as the grouting pressure and flow rate data obtained from the grouting experiment, are used as non-field data. The slurry diffusion morphology data obtained from the grouting experiment and the multiphase flow field data obtained from numerical simulation are used as field data.

3. The intelligent simulation method for grouting based on a large model as described in claim 1, characterized in that, After dividing the acquired data into field data and non-field data and performing preprocessing, the process also includes: extracting time-series statistical features from the non-field data and extracting spatial distribution features from the field data to form a feature vector. The feature vector is then input into a trained classification model to output a probability distribution of the working condition label, thereby determining the working condition type to which the current data belongs.

4. The intelligent simulation method for grouting based on a large model as described in claim 1, characterized in that, The teacher model is a deep neural network that includes a temporal prediction branch, a regional feature extraction branch, and a global flow field prediction branch; the temporal prediction branch processes non-field data through fully connected layers and a temporal attention mechanism to predict pressure and flow. The regional feature extraction branch extracts image features from the field data through two-dimensional convolutional layers and spatial attention mechanisms to predict the phase fraction distribution; the global flow field prediction branch predicts grid features in the field data through three-dimensional convolutional layers and global attention mechanisms to predict the phase fraction, velocity, and pressure.

5. The intelligent simulation method for grouting based on a large model as described in claim 1, characterized in that, The loss function of the teacher model is: ； in, The total loss function for the teacher model. To address the time-series mean square error loss for non-field data, To address the pixel-level mean square error loss in field data, To address the physical mean square error loss of the grid in the field data, This is a cross-scale consistency loss used to bridge point, regional, and global scales; , , and These are the weights for each loss.

6. The intelligent simulation method for grouting based on a large model as described in claim 5, characterized in that, A dynamic weight adaptive mechanism is used for weight updates, specifically: ； in, To adjust the rate, The output of the classifier is the probability of the working condition label. The verification error corresponding to the loss component, Indicates the current iteration step The weight, The updated weights are represented by j, which is an integer index variable representing the position of the loss component whose weight is being adjusted in the total loss function. k represents the auxiliary index in the summation. The indexes of all loss components are traversed, ranging from 1 to the total number of loss terms. The sum is calculated by traversing all terms.

7. The intelligent simulation method for grouting based on a large model as described in claim 1, characterized in that, In transfer learning, cross-scale consistency loss and geological heterogeneity sensitivity loss are introduced to improve adaptability across working conditions; temporal uncertainty loss is introduced to handle the sparsity of time-series data; and adversarial training is introduced, using a discriminator to optimize feature alignment loss to ensure that student models adapt to the feature distribution of new working conditions.

8. A grouting intelligent simulation system based on a large model, characterized in that, include: The data acquisition module is configured to acquire grouting pressure and grouting flow data from on-site grouting monitoring, grouting pressure, flow rate, and grout diffusion morphology data obtained through grouting experiments, and multiphase flow field data obtained through numerical simulation. The preprocessing module is configured to divide the acquired data into field data and non-field data, and then perform preprocessing. The teacher model training module is configured to train the teacher model using preprocessed numerical simulation data, grouting experiment data, and some field monitoring data. The transfer learning module is configured to: within the computational domain, for a given prediction condition, employ a differentiated knowledge distillation strategy to process non-field data and field data separately, transferring the generalized knowledge of the teacher model to the student model. The prediction module is configured to: obtain the parameters under the test condition and input them into the trained student model to obtain the prediction results of the slurry diffusion field, pressure field and velocity field; Specifically, for the defined prediction conditions, a differentiated knowledge distillation strategy is adopted to process non-field data and field data separately, as follows: For time series prediction of non-field data, the teacher model generates the predicted probability distributions of pressure and flow as soft labels, and optimizes the learning of soft labels through KL divergence loss. To address the sparsity and uncertainty of time-series data in the grouting process, a time-series uncertainty loss is introduced, and variational inference is used to estimate the uncertainty of the prediction distribution, penalizing regions with high uncertainty. Simultaneously, the mean square error loss is calculated by combining sparse real data of the target working condition. For the field data images, the teacher model generates soft labels for the phase fraction distribution, capturing the diffusion boundary and velocity pattern, while the student model optimizes the KL divergence loss; pixel-level mean square error loss is calculated by combining real data from sparse region images; at the same time, a geological heterogeneity sensitivity loss is introduced to penalize the model's prediction bias for heterogeneous regions. For the gridded field data, the teacher model generates spatiotemporal soft labels for phase fractions, velocity, and pressure, including eddy current and pressure gradient features, while the student model optimizes the KL divergence loss; at the same time, it calculates the mean square error of the physical field by combining real data from some grid points. The generalized knowledge from the teacher model is transferred to the student model, specifically as follows: Weights are extracted from the time-series prediction branch, regional feature extraction branch, and global flow field prediction branch of the teacher model and used as initial parameters for the student model. These parameters are then fine-tuned for the target operating condition characteristics to ensure the model adapts to the dynamic and distributional features of the specific operating condition. The specific process is as follows: For off-field data, transfer learning uses sparse target working condition data to optimize the time series prediction branch and adjusts the weights through mean squared error loss. For images in the field data, transfer learning optimizes the regional feature extraction branch using the target working condition image and refines the diffusion boundary prediction through pixel-level mean square error. For the grid in the field data, transfer learning uses partial grid point data to optimize the global flow field prediction branch and adjusts the flow field dynamics through the physical field mean square error.

Images