A load modeling system and method based on a multi-modal model fusing geological parameters and water pressure time series data

Abstract

This invention discloses a load modeling system and method based on a multimodal model that fuses geological parameters and water pressure time-series data, belonging to the field of computer technology. The method includes: standardizing and spatiotemporally aligning geological parameters and water pressure time-series data to generate multimodal feature data; extracting and encoding prior knowledge of geological structures using a graph neural network; dynamically modeling the modulation effect of geological conditions on water pressure response using a cross-modal attention and gating fusion mechanism to generate a unified feature representation; reconstructing the current load field and predicting its future evolution based on an encoder-decoder architecture; simultaneously employing multiple random forward propagations to quantify uncertainty and generate confidence intervals for predicted values; and dynamically optimizing and updating the model using a sliding window online learning strategy, combined with online monitoring feedback. This invention effectively solves the problem of spatiotemporal semantic mismatch in multi-source data, improving the accuracy, physical consistency, generalization ability, and engineering practical value of load inversion.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a load modeling system and method based on a multimodal model that integrates geological parameters and water pressure time series data. Background Technology

[0002] In the field of geotechnical engineering and intelligent monitoring of geological hazards, load modeling is a core component of structural safety assessment. Its accuracy directly determines the service reliability of critical infrastructure such as tunnel lining, slope protection, and deep foundation pit retaining walls, as well as the actual effectiveness of risk early warning. This technology needs to comprehensively characterize two effects: first, the static spatial heterogeneity determined by complex geological structures; and second, the instantaneous and cumulative load effects caused by dynamic changes in groundwater. Among these, geological parameters (such as rock mass integrity index, joint group number, permeability coefficient, and principal stress value) characterize the constitutive properties of the medium, while water pressure time series data reflects the dynamic process of external hydrological excitation on the structure.

[0003] With the deep integration of IoT sensing networks and artificial intelligence technologies, the engineering community is urgently seeking a high-precision load modeling method capable of collaboratively analyzing multi-source, heterogeneous, and asynchronous spatiotemporal data to provide core support for the safety and intelligent decision-making system of major projects. Load modeling technology based on multimodal large models is widely recognized as a key path to overcome the bottlenecks of traditional methods. Its core lies in building a unified deep learning framework to simultaneously process multidimensional inputs from geological exploration databases, in-situ sensor arrays, and hydrological monitoring systems, ultimately achieving dynamic inversion and evolution prediction of the coupled geological and hydrological load field.

[0004] However, existing technologies still suffer from systemic defects when fusing geological parameters and water pressure time-series data for load modeling: First, geological parameters are mostly derived from sparse boreholes or geophysical profiles, exhibiting spatial discreteness and non-uniform distribution; while water pressure data is collected continuously at fixed monitoring points at high frequency. The two suffer from severe mismatches in spatial coverage density, temporal sampling frequency, and semantic hierarchy, making it difficult for existing models to construct accurate cross-modal spatiotemporal alignment mechanisms. Second, mainstream fusion strategies often employ shallow operations such as feature stitching or early weighted averaging, failing to explicitly model the nonlinear regulatory effects of geological structures on water pressure conduction and load transfer. Third, existing models generally treat geological parameters as ordinary input features, failing to incorporate prior knowledge of regional geological topology or rock mass structure, resulting in a sharp decline in generalization performance in data-scarce areas or when encountering unseen geological combinations. Finally, most methods only output point estimation results, lacking end-to-end quantification capabilities for model prediction uncertainties. These defects collectively lead to the current load modeling techniques generally suffering from insufficient accuracy, weak robustness, and poor interpretability. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a load modeling system and method based on a multimodal model that integrates geological parameters and water pressure time series data.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This application provides a load modeling system based on a multimodal model that integrates geological parameters and water pressure time series data, including:

[0008] The data preprocessing and alignment module is used to standardize the original geological parameters of the target engineering area and the water pressure time series data of the monitoring points, and to establish the spatiotemporal alignment relationship between the geological parameters and the water pressure time series data, generating spatiotemporally aligned multimodal feature data.

[0009] The geological prior coding module is used to extract and encode prior knowledge reflecting the regional geological structure relationship from standardized geological parameters, and generate geological prior features.

[0010] The feature fusion module is used to receive the multimodal feature data and the geological prior features, model the modulation effect of geological conditions on water pressure response through a cross-modal interaction mechanism, and generate a unified feature representation that integrates geological and hydrological information.

[0011] The load modeling module is used to reconstruct the spatial distribution of the load field at the current moment and predict the load field evolution sequence within a preset time window based on the unified feature representation.

[0012] The uncertainty quantification module is used to simultaneously estimate the uncertainty of the model output and generate confidence intervals for the load prediction values ​​during the load modeling process.

[0013] The model optimization and deployment module is used to dynamically adjust model parameters based on the feedback error between online monitoring data and model prediction results, and then deploy the optimized model to the engineering evaluation platform.

[0014] Optionally, in the data preprocessing and alignment module, categorical variables in geological parameters are encoded, and continuous variables are normalized; water pressure time series data are cleaned and feature decomposed; the spatiotemporal alignment relationship is constructed through spatial interpolation and temporal resampling methods.

[0015] Optionally, the geological prior coding module is implemented based on a graph neural network. It generates the geological prior features by constructing a graph structure with spatial locations as nodes and relationships between locations as edges, and by aggregating information from neighboring nodes.

[0016] Optionally, the feature fusion module includes a cross-modal attention unit and a gated fusion unit; the cross-modal attention unit is used to calculate the correlation weight between water pressure features and geological prior features; the gated fusion unit is used to dynamically fuse the two types of features according to the correlation weight.

[0017] Optionally, the load modeling module adopts an encoder and decoder architecture, wherein the encoder is used to extract spatiotemporal features from the unified feature representation, and the decoder is used to reconstruct and predict the load field based on the spatiotemporal features.

[0018] Optionally, the uncertainty quantification module performs multiple random forward propagations during model inference to statistically analyze the distribution of prediction results and then calculates the confidence interval.

[0019] Optionally, the model optimization and deployment module adopts an online learning strategy based on a sliding window, using the latest monitoring data to incrementally fine-tune and update the model.

[0020] Secondly, this application provides a load modeling method based on a multimodal model that integrates geological parameters and water pressure time series data, including:

[0021] The original geological parameters and water pressure time series data are standardized and spatiotemporally aligned to generate multimodal feature data;

[0022] Geological structure prior knowledge is extracted and encoded from the geological parameters to generate geological prior features;

[0023] By fusing the multimodal feature data with the geological prior features, a unified feature representation is generated;

[0024] Based on the unified feature representation, the current load field is reconstructed and the future load field evolution sequence is predicted;

[0025] During the load reconstruction and prediction process, prediction uncertainties are estimated simultaneously and confidence intervals are generated.

[0026] Based on online monitoring feedback, the model parameters are dynamically optimized and the deployed model is updated.

[0027] Optionally, the fusion process calculates the correlation between water pressure features and geological prior features through a cross-modal attention mechanism, and dynamically weights the fusion based on a gating mechanism.

[0028] Optionally, the estimation of prediction uncertainty is achieved by enabling random dropout during model inference and performing multiple forward propagations.

[0029] Compared with the prior art, this application has the following beneficial effects:

[0030] This invention proposes a load modeling system and method based on a multimodal model that integrates geological parameters and water pressure time-series data. By constructing a spatiotemporally aligned multimodal feature tensor and a geological prior knowledge embedding mechanism, the system effectively addresses the mismatch between geological parameters and water pressure time-series data in terms of spatiotemporal scale and semantic hierarchy. A hierarchical cross-modal attention and gating fusion network is employed to dynamically model the nonlinear modulation effect of the geological medium's response to water pressure, significantly improving the accuracy and physical consistency of load inversion. A geological prior knowledge encoding mechanism based on graph neural networks is introduced, enhancing the model's generalization ability and robustness under small sample or unseen geological conditions. An end-to-end uncertainty quantification mechanism is integrated to provide a reliable confidence interval for load prediction results, enhancing the practical value of this method in safety early warning scenarios. A model optimization and deployment module is designed, allowing the system to dynamically adjust based on real-time feedback data, ensuring the long-term effectiveness and accuracy of load modeling. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention.

[0032] Figure 2 This is a schematic diagram of the core principle framework of the feature fusion module in this invention.

[0033] Figure 3 This is a logical flowchart of the data preprocessing and alignment module in this invention.

[0034] Figure 4 This is a logical flow diagram of the load modeling module in this invention.

[0035] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the model optimization and deployment module in this invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Furthermore, in this invention, an element referred to as fixed to or disposed on another element may be directly disposed on the other element, or there may be an intermediate element. When an element is considered to be connected to another element, it may be directly connected to the other element, or there may be an intermediate element present simultaneously. The terms vertical, horizontal, left, right, and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0038] Example 1

[0039] In the safety monitoring and early warning scenarios of deep-buried long tunnel projects or large reservoir dam foundation projects, accurately assessing the load distribution and evolution trend of rock masses under the coupled action of complex geological structures and dynamic groundwater is a core challenge to ensure the long-term stability and safe operation of engineering structures. Traditional methods typically process static spatial parameters obtained from geological exploration and dynamic water pressure time-series data obtained from monitoring separately, using simple empirical formulas or linear superposition for load estimation. This makes it difficult to accurately characterize the nonlinear modulation effect of the geological medium on the dynamic response of water pressure, resulting in insufficient accuracy and poor physical consistency in load inversion and prediction, and failing to provide a risk boundary with statistical confidence for engineering decisions. This embodiment addresses the above problems by providing a load modeling method and system implementation process based on a multimodal large model that integrates geological parameters and water pressure time-series data.

[0040] See Figure 1 This application provides a load modeling system based on a multimodal model that integrates geological parameters and water pressure time series data. The system includes: a data preprocessing and alignment module, a geological prior coding module, a feature fusion module, a load modeling module, an uncertainty quantification module, and a model optimization and deployment module. These modules work together to form an end-to-end intelligent load modeling and analysis system. The functions of each module are described in detail below.

[0041] First, the data preprocessing and alignment module begins operation to standardize the original geological parameters of the target engineering area and the water pressure time series data of the monitoring points, and establish the spatiotemporal alignment relationship between the geological parameters and the water pressure time series data, generating spatiotemporally aligned multimodal feature data. In the data preprocessing and alignment module, categorical variables in the geological parameters are encoded, and continuous variables are normalized; the water pressure time series data is cleaned and feature decomposed; the spatiotemporal alignment relationship is constructed through spatial interpolation and temporal resampling methods.

[0042] Specifically, this module receives raw, multi-source, heterogeneous data from the target engineering area. Geological parameter data originates from pre-project geological survey reports, borehole core records, 3D seismic wave velocity inversion results, and ground-penetrating radar data. This includes lithology classification, joint and fracture density, in-situ stress components, elastic modulus, Poisson's ratio, permeability coefficient, and rock mass integrity index at different spatial locations. These data are spatially discrete and have uneven sampling density. Water pressure time-series data comes from a network of pore water pressure gauges pre-deployed within boreholes or structures. These sensors continuously monitor pore water pressure changes at different spatial locations at a fixed sampling frequency, resulting in a continuous time series. The first step in data preprocessing is data cleaning and standardization. For categorical variables in the geological parameters, such as lithology, one-hot encoding is used to convert them into binary vectors, with the vector dimension equal to the total number of lithology categories. For continuous variables, such as joint density and in-situ stress components, a maximum-minimum normalization method is used to linearly transform all values ​​to the zero-to-one interval. The transformation formula is as follows: ,in The original value, and These represent the minimum and maximum values ​​of the parameter across the entire engineering area. For water pressure time-series data, missing value detection is first performed. For missing values ​​caused by brief communication interruptions of the sensor, a Kalman filter method based on a time-series autoregressive model is used for imputation. For obvious outliers, such as spikes caused by sensor impacts or power fluctuations, the three-sigma criterion based on statistical distribution is used for identification and removal, i.e., data points deviating from the series mean by more than three standard deviations are removed. Subsequently, the cleaned water pressure time-series data undergoes detrending and seasonal decomposition processing. The STL decomposition method is used to separate the long-term trend term, the periodic seasonal term, and the residual random term, and the residual sequence is used as a feature reflecting the dynamic fluctuations of water pressure. See [link to relevant documentation] Figure 3After preprocessing, the module performs a crucial spatiotemporal alignment operation. This operation aims to map discrete spatially distributed geological parameters to continuous time-series water pressure data into a unified spatiotemporal reference frame. Specifically, firstly, based on the scope and accuracy requirements of the engineering area, a three-dimensional Cartesian spatial grid is defined, with grid nodes corresponding to the spatial locations where loads need to be calculated. The grid resolution is set according to the engineering safety level. For geological parameters, ordinary kriging spatial interpolation based on variogram theory is used to interpolate the geological parameter values ​​of discrete exploration points to each spatial grid node. The interpolation process first calculates the semi-variogram between each exploration point, fits the theoretical variogram model, and then calculates the optimal weights based on the model and the spatial positional relationship between the interpolation point and known points. Finally, a weighted sum is obtained to obtain the parameter estimate value of each grid node, forming a continuous spatial distribution field for each geological parameter. For water pressure data, since the monitoring point locations and spatial grid nodes usually do not completely coincide, the kriging interpolation method is also used to interpolate the water pressure values ​​of all monitoring points at each sampling time to all spatial grid nodes, thereby obtaining the water pressure value of each grid node at each sampling time. In the time dimension, the water pressure data is resampled to a unified timestamp sequence, with the time step set according to the engineering early warning response time requirements. Finally, a three-dimensional feature tensor is generated, with dimensions equal to the time step multiplied by the number of spatial grid nodes and the number of feature channels. The feature channels include all processed geological parameter features and detrended water pressure dynamic features. This three-dimensional tensor is the spatiotemporally aligned multimodal feature tensor, providing a rigorously aligned data foundation for subsequent deep fusion.

[0043] Next, the geological prior coding module starts working, which is used to extract and encode prior knowledge reflecting the regional geological structure relationship from the standardized geological parameters, and generate geological prior features. The geological prior coding module is based on graph neural network, which generates the geological prior features by constructing a graph structure with spatial location as nodes and the relationship between locations as edges, and aggregating the information of neighboring nodes.

[0044] Specifically, the core task of this module is to extract and encode prior knowledge reflecting the macroscopic laws and constraints of the regional geological structure from standardized geological parameters, rather than simply stacking parameters. To achieve this goal, this embodiment uses a graph neural network as the encoder. First, a geological attribute map is constructed. Each spatial grid node is taken as a node in the graph, and the feature vector of this node is the vector composed of all its corresponding geological parameter values. Then, edges between nodes are constructed based on the spatial proximity relationship and the similarity of geological attributes. Specifically, a dual connection strategy is adopted: one is based on spatial distance, setting a spatial neighborhood radius, and establishing an undirected edge between two nodes whose distance is less than this radius; the other is based on the similarity of geological attributes, calculating the cosine similarity of the geological feature vectors of two nodes, and establishing a connection between nodes whose similarity is higher than a set threshold. In this way, the geological space of the entire engineering area is abstracted into a graph structure, where nodes carry local geological information, and edges not only represent spatial proximity but also represent the similarity and correlation of geological attributes. Subsequently, this geological attribute map is input into a network containing two layers of graph convolutional layers. Each graph convolutional operation aggregates the feature information of each node's first-order neighbors and updates the node's feature representation using a non-linear activation function. After two layers of graph convolutional operations, the output feature vector of each node is no longer just its own raw parameters, but rather a "context-aware" feature that incorporates geological structural information within a certain surrounding range. This feature vector contains the local structural pattern of the geological conditions at that location; for example, it can effectively distinguish nodes located inside intact rock blocks, in fault zones, at contact zones between different lithologies, or in areas with dense joints. The output feature vectors of all nodes together constitute a geological prior feature matrix, which serves as an explicit regional geological structural constraint and is fed into the subsequent fusion network.

[0045] Next, the feature fusion module begins operation, receiving the multimodal feature data and the geological prior features. It models the modulation effect of geological conditions on the water pressure response through a cross-modal interaction mechanism, generating a unified feature representation that integrates geological and hydrological information. The feature fusion module includes a cross-modal attention unit and a gated fusion unit. The cross-modal attention unit calculates the correlation weight between the water pressure features and the geological prior features. The gated fusion unit dynamically fuses the two types of features based on the correlation weight.

[0046] Specifically, the feature fusion module receives the aforementioned multimodal feature tensor and geological prior feature matrix. Its core objective is to dynamically model the modulation effect of geological conditions on the dynamic response of water pressure and generate a unified feature representation that couples geological and hydrological information. See also Figure 2The network employs a hierarchical design, primarily comprising a cross-modal attention sub-network and a gated fusion sub-network. The cross-modal attention sub-network operates first. It uses water pressure temporal features as the query source and geological prior features as the key and value sources. Specifically, water pressure features are processed through a trainable linear transformation layer to generate the query matrix, while geological prior features are processed through two additional independent trainable linear transformation layers to generate the key and value matrices, respectively. Subsequently, the dot product of the query matrix and the key matrix is ​​calculated and scaled by dividing by the square root of the key vector dimension to stabilize the gradient. A soft maximization function is applied to the scaled result to obtain the attention weight matrix. The physical meaning of this weight matrix is ​​very clear: it quantifies the correlation strength between each water pressure dynamic state and various surrounding geological conditions at each spatiotemporal location. For example, in highly permeable sandstone areas, water pressure changes may be rapidly transmitted, causing significant load changes; therefore, geological features in this area will receive higher attention weights. Conversely, in extremely low permeability mudstone areas, water pressure changes have limited impact, and the corresponding geological features will have lower attention weights. Finally, the attention weight matrix is ​​multiplied by the value matrix to obtain the water pressure feature representation after weighted modulation by geological information. This process can be expressed by the following formula: Where Q is the query matrix after water pressure feature transformation, and K and V are the key matrix and value matrix after geological prior feature transformation. Let G' be the dimension of the key vector. This mechanism allows the model to automatically focus on the geological factors most relevant to the current water pressure change. Next, the gated fusion subnetwork begins operation. It receives two inputs: one is the geologically modulated water pressure feature after cross-modal attention weighting, and the other is the original water pressure temporal feature. This subnetwork contains a learnable gating unit, consisting of a fully connected layer and a sigmoid activation function. The gating unit calculates a fusion weight vector between zero and the two inputs based on the current feature state. Finally, the fusion feature is obtained by weighting the two input features with the gating weights. Here, G' represents the attention-modulated geological feature, H represents the original water pressure feature, and g represents the gating weight. The gating mechanism gives the model the ability to dynamically adjust the information fusion strength. For example, in areas with homogeneous geological conditions and little influence on water pressure response, the gating weight may approach zero, and the model mainly relies on water pressure features; while in areas with complex geological conditions and a controlling effect on the load, the gating weight increases, and geologically modulated information is significantly injected. After processing by this network, the output is a unified feature representation that deeply couples geological structural constraints with dynamic changes in water pressure.

[0047] Subsequently, the load modeling module begins operation, reconstructing the spatial distribution of the load field at the current moment and predicting the load field evolution sequence within a preset time window based on the unified feature representation. The load modeling module employs an encoder and decoder architecture. The encoder extracts spatiotemporal features from the unified feature representation. The decoder reconstructs and predicts the load field based on these spatiotemporal features.

[0048] Specifically: the load modeling module, based on the aforementioned unified feature representation, performs spatial reconstruction and temporal prediction tasks of the load field; see [link to related documentation]. Figure 4 This module employs an encoder-decoder architecture. The encoder is responsible for extracting and compressing key spatiotemporal patterns from high-dimensional fused features. It consists of two 3D convolutional layers and one bidirectional gated recurrent unit (GRU) layer. The first 3D convolutional layer uses a small kernel to perform convolution operations in two spatial dimensions and one temporal dimension, extracting subtle local spatiotemporal features. The second 3D convolutional layer uses a larger receptive field to capture a more macroscopic spatiotemporal structure. Following the convolutional layers is a bidirectional gated recurrent unit network, which further aggregates historical information and short-term future context along the temporal dimension, thereby enhancing the ability to model temporal dependencies. The encoder output is a highly abstract and condensed spatiotemporal feature encoding. The decoder is responsible for progressively mapping the encoded features back to the original load field space and generating future sequences. The decoder consists of one transposed convolutional layer and one forward-gated recurrent unit in the temporal dimension. The transposed convolutional layer progressively upsamples the spatial size of the feature map to the spatial resolution of the target load field. Simultaneously, in the temporal dimension, the forward-gated recurrent unit is used to autoregressively generate the load field for future time steps. The decoding process is as follows: given the fused feature encodings of the current moment and history, the decoder first reconstructs the spatial distribution of the load field at the current moment; then, using this reconstruction result as a starting point, and combining the model's latent state prediction of the system state at the next moment, iteratively generates the load field evolution sequence within a preset time window. This module is optimized using a composite loss function during training. The loss function is defined as the weighted sum of the mean squared error and the spatial gradient smoothing constraint term, i.e. ,in For the predicted load field, For spatial gradient operators, and where is the weighting coefficient. This loss function ensures that the reconstructed load field is spatially continuous and smooth, conforming to physical principles, while guaranteeing prediction accuracy.

[0049] Next, the uncertainty quantification module starts working to simultaneously estimate the uncertainty of the model output and generate confidence intervals for the load prediction values ​​during the load modeling process. The uncertainty quantification module calculates the confidence intervals by performing multiple random forward propagations during model inference, statistically analyzing the distribution of the prediction results.

[0050] Specifically, while the load modeling module performs calculations, the uncertainty quantification module is simultaneously activated to provide a reliable probabilistic interpretation for each prediction result. This embodiment employs the Monte Carlo Dropout method to achieve uncertainty estimation. During model training, Dropout layers are randomly inserted after the key fully connected layers and convolutional layers in the feature fusion module and load modeling module, with a dropout rate set to 20%. During the inference phase after model training, these Dropout layers are not turned off as in traditional methods but remain active. When a prediction is needed, the model runs fifty times via forward propagation. Due to the randomness of Dropout, each forward propagation is equivalent to sampling a slightly different subnetwork from the model's posterior distribution for prediction, resulting in fifty slightly different prediction results. For the predicted load value of a spatial grid node at a specific future time step, these fifty results constitute an empirical distribution. The sample mean of this distribution is calculated as the best estimate of the load at that point and time, and the sample standard deviation is calculated as a measure of prediction uncertainty. Based on the normal distribution assumption, the interval formed by adding or subtracting 1.96 times the standard deviation from the mean is defined as the 95% confidence interval for the predicted value. This module ultimately outputs the mean load, upper confidence bound, and lower confidence bound for each spatial location at each future prediction time point, forming a complete set of probabilistic prediction results. This mechanism provides crucial quantitative uncertainty information for engineering risk assessment; for example, when the upper bound of the predicted load confidence interval exceeds the structural design safety threshold, the system can issue a high-risk warning.

[0051] Finally, the model optimization and deployment module begins its work, dynamically adjusting model parameters based on feedback errors between online monitoring data and model prediction results, and deploying the optimized model to the engineering evaluation platform. This module employs a sliding window-based online learning strategy, using the latest monitoring data to incrementally fine-tune and update the model.

[0052] Specifically, the model optimization and deployment module ensures the entire system can adapt to long-term changes in the engineering environment, maintaining the model's timeliness and accuracy. This module implements an online learning strategy based on a sliding window. During online operation, the system continuously receives real, newly generated water pressure data from monitoring points and corresponding load inversion reference values. The module maintains a data buffer with a fixed capacity of the most recent 1,000 time steps to store monitoring data, model prediction results at the corresponding time points, and reference values. When the buffer is filled with new data, the module automatically calculates the average prediction error of the current model on this latest batch of data. If this error exceeds a threshold pre-set according to engineering safety standards, the model fine-tuning process is triggered. During fine-tuning, the data in the buffer is used as training samples, and the trained model is optimized with a small learning rate for one round of gradient descent optimization, updating the model's parameters. This fine-tuning is incremental, aiming to allow the model to quickly adapt to the latest drift in data distribution, such as changes in load response patterns caused by periodic reservoir water level adjustments, continuous tunnel excavation unloading, or long-term rock creep, without forgetting previously learned fundamental knowledge. After fine-tuning, the module automatically deploys the updated model parameter file to the online service interface of the engineering safety assessment platform, replacing the old model version. The entire update process strives for automation and seamless integration, ensuring uninterrupted load assessment services and that assessment results are always based on the current optimal model, thereby achieving continuous self-evolution and improvement of load modeling capabilities. See also Figure 5 This module forms a closed-loop interaction relationship with the data flow, model flow, and engineering application platform.

[0053] Example 2

[0054] Based on the above embodiment 1, this embodiment provides a load modeling method based on a multimodal model that integrates geological parameters and water pressure time series data, including: steps S110-S160:

[0055] Step S110. Standardize and align the original geological parameters and water pressure time series data to generate multimodal feature data.

[0056] Specifically, the original geological parameters of the target engineering area and the water pressure time series data of monitoring points are collected and preprocessed. The lithology classification, joint density, and geostress components of the geological parameters are then subjected to unique thermal coding or numerical normalization. The numerical normalization process employs a formula... The continuous variables are linearly transformed to the [0,1] interval. Missing values ​​are filled, outliers are removed, and the water pressure time series data are detrended and seasonally decomposed to generate a standardized geological parameter matrix and water pressure time series.

[0057] A spatiotemporal alignment mechanism is constructed. A Kriging spatial interpolation method is used to interpolate the discrete spatially distributed geological parameters to the spatial grid corresponding to the monitoring point location. A linear interpolation method is used to resample the water pressure data to a unified timestamp sequence, generating a spatiotemporally aligned multimodal feature tensor.

[0058] The unified feature representation is input into the load modeling module, which uses an encoder consisting of two 3D convolutional layers and one bidirectional gated recurrent unit to extract spatiotemporal features. Then, a decoder consisting of one transposed convolutional layer and one feedforward recurrent network is used to reconstruct the current load field and predict the future load field evolution sequence.

[0059] Geological structure prior knowledge is extracted and encoded from the geological parameters to generate geological prior features;

[0060] Specifically, prior knowledge of geological structure is extracted from standardized geological parameters, a geological attribute map is constructed with each spatial grid point as a graph node, and the topological relationships between geological units are encoded through 2-3 layers of graph convolution operations to generate a geological prior feature vector representing the constraints of regional geological conditions.

[0061] The multimodal feature data and the geological prior features are fused to generate a unified feature representation. The fusion process calculates the correlation between water pressure features and geological prior features through a cross-modal attention mechanism, and dynamically weights the fusion based on a gating mechanism.

[0062] Specifically: the multimodal feature tensor and the geological prior feature vector are input into the feature fusion module, and the mutual attention score between the hydraulic pressure time series features and the geological prior features is calculated through a hierarchical cross-modal attention mechanism. The attention calculation adopts the formula... Then, the fusion weight $g_t$ is calculated through the gated fusion unit to generate a unified feature representation that integrates geological and hydrological coupled information. The output of the gated fusion unit. Activated by the sigmoid function.

[0063] Based on the unified feature representation, the current load field is reconstructed and the future load field evolution sequence is predicted.

[0064] The unified feature representation is input into the load modeling module, and the spatiotemporal features are extracted using an encoder consisting of two 3D convolutional layers and one bidirectional gated recurrent unit. Then, the current load field is reconstructed and the future load field evolution sequence is predicted using a decoder consisting of one transposed convolutional layer and one feedforward recurrent network.

[0065] During load reconstruction and prediction, prediction uncertainty is estimated simultaneously and confidence intervals are generated. The estimation of prediction uncertainty is achieved by enabling random dropout techniques and performing multiple forward propagations during model inference.

[0066] During the dynamic reconstruction and prediction of the load field, uncertainty quantification is performed simultaneously by using the Monte Carlo Dropout method to perform T random forward propagations and estimate the mean of the model predictions. and standard deviation Generate 95% confidence intervals for predicted load values. The number of random forward propagations, T ≥ 50.

[0067] Based on online monitoring feedback, the model parameters are dynamically optimized and the deployed model is updated.

[0068] Based on real-time monitoring data acquired online, the model prediction error is calculated. An online learning strategy based on a sliding window is adopted to dynamically fine-tune the model using the latest data in the data buffer, and the updated model is deployed to the engineering safety assessment platform.

[0069] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0070] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A load modeling system based on a multimodal model that integrates geological parameters and water pressure time series data, characterized in that, include: The data preprocessing and alignment module is used to standardize the original geological parameters of the target engineering area and the water pressure time series data of the monitoring points, and to establish the spatiotemporal alignment relationship between the geological parameters and the water pressure time series data, generating spatiotemporally aligned multimodal feature data. The geological prior coding module is used to extract and encode prior knowledge reflecting the regional geological structure relationship from standardized geological parameters, and generate geological prior features. The feature fusion module is used to receive the multimodal feature data and the geological prior features, model the modulation effect of geological conditions on water pressure response through a cross-modal interaction mechanism, and generate a unified feature representation that integrates geological and hydrological information. The load modeling module is used to reconstruct the spatial distribution of the load field at the current moment and predict the load field evolution sequence within a preset time window based on the unified feature representation. The uncertainty quantification module is used to simultaneously estimate the uncertainty of the model output and generate confidence intervals for the load prediction values ​​during the load modeling process. The model optimization and deployment module is used to dynamically adjust model parameters based on the feedback error between online monitoring data and model prediction results, and then deploy the optimized model to the engineering evaluation platform.

2. The load modeling system according to claim 1, characterized in that, In the data preprocessing and alignment module, categorical variables in geological parameters are encoded, and continuous variables are normalized; water pressure time series data are cleaned and feature decomposed; the spatiotemporal alignment relationship is constructed through spatial interpolation and temporal resampling methods.

3. The load modeling system according to claim 1, characterized in that, The geological prior coding module is implemented based on a graph neural network. It generates the geological prior features by constructing a graph structure with spatial locations as nodes and relationships between locations as edges, and by aggregating information from neighboring nodes.

4. The load modeling system according to claim 1, characterized in that, The feature fusion module includes a cross-modal attention unit and a gated fusion unit; the cross-modal attention unit is used to calculate the correlation weight between water pressure features and geological prior features; the gated fusion unit is used to dynamically fuse the two types of features according to the correlation weight.

5. The load modeling system according to claim 1, characterized in that, The load modeling module adopts an encoder and decoder architecture. The encoder is used to extract spatiotemporal features from the unified feature representation, and the decoder is used to reconstruct and predict the load field based on the spatiotemporal features.

6. The load modeling system according to claim 1, characterized in that, The uncertainty quantification module performs multiple random forward propagations during model inference to statistically analyze the distribution of prediction results and then calculates the confidence interval.

7. The load modeling system according to claim 1, characterized in that, The model optimization and deployment module adopts an online learning strategy based on a sliding window, using the latest monitoring data to incrementally fine-tune and update the model.

8. A load modeling method based on a multimodal model that integrates geological parameters and water pressure time series data, characterized in that, include: The original geological parameters and water pressure time series data are standardized and spatiotemporally aligned to generate multimodal feature data; Geological structure prior knowledge is extracted and encoded from the geological parameters to generate geological prior features; By fusing the multimodal feature data with the geological prior features, a unified feature representation is generated; Based on the unified feature representation, the current load field is reconstructed and the future load field evolution sequence is predicted; During the load reconstruction and prediction process, prediction uncertainties are estimated simultaneously and confidence intervals are generated. Based on online monitoring feedback, the model parameters are dynamically optimized and the deployed model is updated.

9. The load modeling method according to claim 8, characterized in that, The fusion process calculates the correlation between water pressure features and geological prior features through a cross-modal attention mechanism, and dynamically weights the fusion based on a gating mechanism.

10. The load modeling method according to claim 8, characterized in that, The estimation of prediction uncertainty is achieved by enabling random dropout during model inference and performing multiple forward propagations.

Images