Loess mechanical parameter space-time interpolation prediction model construction method based on big data

Abstract

The invention discloses a loess mechanical parameter spatio-temporal interpolation prediction model construction method based on big data, and relates to the technical field of geotechnical engineering data analys.The loess mechanical parameter spatio-temporal interpolation prediction model construction method comprises the following steps that loess mechanical parameter data, environmental factor data and geologic structure data are collected through a multi-source sensor, and spatio-temporal unified multi-modal data are constructed; designing a collaborative kernel function fusing spatial heterogeneity and time dynamics on the basis of multi-modal data, and quantifying relevance between mechanical parameters and space-time positions and environmental factors; according to the collaborative kernel function, dynamically optimizing a space weight, a time attenuation coefficient and a multi-source data fusion weight, and generating a parameter-adaptive interpolation model; carrying out distributed training and calculation acceleration on the interpolation model by adopting a block matrix approximation and parallel calculation architecture; through the trained model, a mechanical parameter prediction value of a target space-time position is calculated in real time, and the calculation efficiency is improved through a tensor expansion algorithm.

Description

Technical Field

[0001] This invention belongs to the field of geotechnical engineering data analysis technology, specifically involving a method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data. Background Technology

[0002] In the field of geological data analysis, accurate prediction of loess mechanical parameters is crucial for the stability assessment of infrastructure such as roadbeds and tunnels. Traditional methods for predicting loess mechanical parameters often rely on static models or analysis of single data sources, making it difficult to effectively capture spatiotemporal heterogeneity and dynamic change patterns. Existing spatial interpolation methods, such as Kriging, can handle spatial correlations but ignore the dynamic coupling effect of the time dimension and cannot efficiently integrate multi-source heterogeneous data. Conventional models have high computational complexity when processing large-scale spatiotemporal data, making it difficult to meet real-time prediction requirements. Existing dynamic optimization methods lack the embedding of physical laws, leading to deviations between prediction results and actual conditions. In terms of data preprocessing, missing values ​​are often handled using simple interpolation methods, failing to fully utilize spatiotemporal correlations, resulting in limited data quality. These shortcomings limit the accuracy and practicality of loess mechanical parameter prediction, especially in complex geological environments and variable climate conditions, making it difficult to meet the high-precision requirements of engineering practice. Therefore, we propose a method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data. Summary of the Invention

[0003] The purpose of this invention is to provide a method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data, which solves the problems of poor accuracy and practicality of existing loess mechanical parameter predictions.

[0004] The technical solution adopted in this invention is: a method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data, and the specific operation steps are as follows: Step S1: Collect loess mechanical parameter data, environmental factor data and geological structure data through multi-source sensors, construct spatiotemporally unified multimodal data, and divide the multimodal data into training set and test set; Step S2: Based on the multimodal data, design a collaborative kernel function that integrates spatial heterogeneity and temporal dynamics to quantify the correlation between mechanical parameters and spatiotemporal location and environmental factors; Step S3: Based on the collaborative kernel function, dynamically optimize the spatial weights, temporal decay coefficients, and multi-source data fusion weights to generate an interpolation model with adaptive parameters; Step S4: Using a block matrix approximation and parallel computing architecture, the interpolation model is trained and optimized in a distributed manner using the training set; Step S5: Using the trained interpolation model, calculate the predicted values ​​of the mechanical parameters of the target's spatiotemporal location in real time, and improve computational efficiency through tensor expansion algorithm.

[0005] The invention is further characterized in that, The multimodal data is a spatiotemporally unified three-dimensional data structure formed by integrating data from different sources, types, and dimensions. Step S1 includes the following steps: Step S11: Analyze the borehole data set. Z-score standardization is performed to eliminate dimensional differences, as shown in the following formula: ; In the formula, For the first Spatial coordinates of a borehole For the first Timestamps of borehole data This is the original borehole data. For standardized borehole data, This represents the mean of the borehole data. The standard deviation of the borehole data. For spatial coordinates, For timestamps, This represents different types of borehole data, including density, specific gravity, shear strength, etc. This represents the total number of borehole samples. Multi-source sensor datasets Normalize to [0,1] using Min-Max. This indicates the data collected by the sensor, including temperature, humidity, pore water pressure, etc. It is the first One sensor 3D feature vectors The number of feature dimensions of the sensor data; the collection of remote sensing data. And perform georegistration and radiometric correction. This is the first in remote sensing data i The feature vector of each sample The number of feature dimensions in the remote sensing data; This indicates that the data source is remote sensing, including vegetation index, surface temperature band data, etc. Step S12: Establish a unified spatiotemporal coordinate system. The space adopts UTM projection, the time is unified to UNIX timestamp, and a three-dimensional spatiotemporal grid is defined. The borehole data, sensor data and remote sensing data are mapped to the unified three-dimensional spatiotemporal grid through bilinear interpolation. Step S13: Use a spatiotemporal conditional generative adversarial network (ST-GAN) to fill in the missing values. The loss function of the generator is defined as a spatiotemporal gradient consistency constraint, as shown in the following formula: ; In the formula, Let the loss function of the generator be... As the expected value, This is a discriminator network used to distinguish between generated data and real data. This is a generator network used to supplement missing mechanical parameters. For spacetime coordinates, The weights of the gradient penalty term are... For the spatiotemporal gradient operator, It consists of multimodal spatiotemporal data, namely borehole data, sensor data, and remote sensing data.

[0006] Step S2 includes the following steps: Step S21: Define an adaptive spatiotemporal collaborative kernel function. The collaborative kernel function integrates spatial proximity, temporal correlation, and multi-source data coupling characteristics. The formula is as follows: ; In the formula, For the spacetime kernel function, , The first The and the first j Spatial coordinates and timestamps of the sample For adaptive weight parameters, Represents the space core, Represents the time core; For the partial derivative terms of embedded physical laws, where, The mathematical expression of the physical equations embedded in the spacetime kernel function: For the first k The input variables are derived from three main categories: borehole data, sensor data, and remote sensing data. For multimodal data mapping functions, For the first Class input variables in space-time points The value on; Step S22: Fuse multi-source heterogeneous data to construct a tensor coupling operator: ; In the formula, For the first One core tensor component Factor vector of data source modality, The factor vector of spatial modes, The factor vector of the time modality; Represents the outer product of vectors; operator Used to define the first k Class data modal and the first k′ The coupling relationship between data modalities; Multimodal projection is achieved using an LSTM network, as shown in the following formula: ; In the formula, For the first The hidden state of class data, Representing the k Trainable weight parameters for class data modalities This is a 3D convolution operation.

[0007] The space core It can capture spatial non-uniformity and employs improved... The function, and its calculation formula, are as follows: ; In the formula, For improvement The spatial kernel of the function, These are the fixed coefficients of the Matern kernel function. For spatial Euclidean distance; The time kernel can characterize the annual periodic variation of loess mechanical parameters. Introducing a periodic component, the calculation formula is as follows: ; In the formula, The weights of the periodic components, This is a yearly cycle parameter, 365 days. This is the absolute value of the time difference.

[0008] Step S3 includes the following steps: Step S31: Construct the feature importance evaluation matrix The impact of quantification weight parameters on the prediction results is expressed by the following formula: ; In the formula, For the first Weight parameters for the first Sensitivity of features It is a linear rectified activation function. For the spatiotemporal kernel function with respect to the weight parameters The partial derivatives, Representing spatial weights, temporal decay coefficients, and multi-source data fusion weight parameters, reflecting... The sensitivity of changes to kernel function values; For the rate of change of predicted values ​​relative to input features The partial derivatives, Representing borehole data characteristics, sensor data characteristics, and remote sensing data characteristics, reflecting features The impact on dynamic prediction of the interpolation model, 3 is a matrix number of rows; For matrix The number of columns represents the total feature dimension after multimodal data fusion; The weight parameters are dynamically updated using a gating mechanism, as shown in the following formula: ; In the formula, For the Softmax function, Update the matrix for weights. For the loss function with respect to the weight parameters The gradient; Step S32: Embed the humidity migration equation as a regularization term to make the predicted mechanical parameters conform to the physical laws of humidity migration. The loss of the regularization term is as follows: ; In the formula, For gradient operators, Let be the humidity diffusivity tensor. The regularization coefficient is . It is the regularization loss, which is part of the loss function and optimizes the interpolation model parameters during training.

[0009] Step S4 includes the following steps: Step S41: Divide the spatial domain into For grid blocks, a block diagonal approximation method is used to accelerate matrix inversion, as shown in the following formula: ; In the formula, For the first Local kernel matrix of each spatial block, For noise variance, This is the identity matrix for the corresponding dimension; B It represents the number of blocks in the grid. Among them, the convolution kernel weights and bias terms Distributed training is achieved through a data parallelism strategy; Step S42: Use CUDA streams to process each grid block in parallel, allocate an independent CUDA stream to each block, and compute the local inverse matrix of all blocks in parallel; use a mixed precision storage strategy to train the interpolation model.

[0010] The mixed-precision storage strategy in step S42 is as follows: The space kernel parameters are stored as FP16 to reduce memory usage; Time kernel parameters are stored in FP32 to ensure calculation accuracy; The coupling term uses INT8 quantization to reduce data transmission latency.

[0011] In step S5, based on the trained interpolation model, the new input is processed. The predicted value can be quickly calculated using tensor expansion: ; In the formula, For new input point The predicted value, Let covariance vector be the sum of the variances between the new input point and the training set data. The covariance matrix of the training set data, For observation data vectors; in, The calculation formula is as follows: ; In the formula, Indicates vectorization operation, These correspond to data mode, spatial mode, and temporal mode parameters, respectively. It is a multimodal coupled core tensor.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention proposes an adaptive spatiotemporal collaborative kernel function, which integrates spatial proximity, temporal dynamics, and multi-source coupling characteristics. By dynamically optimizing weight parameters and introducing a periodic time kernel and improving the Matern spatial kernel, it effectively characterizes the spatiotemporal heterogeneity and annual periodic variation of loess mechanical parameters. Combined with an LSTM network to achieve multimodal projection, and embedding a moisture migration equation as a physical regularization term, the model's prediction results conform to actual physical laws, enhancing the model's physical interpretability and predictive rationality, and providing more scientific theoretical support for loess engineering stability assessment and disaster early warning.

[0013] 2. This invention effectively eliminates the dimensional differences and distribution inconsistencies among borehole data, sensor data, and remote sensing data by integrating multi-source heterogeneous data and employing standardization and normalization processing. This significantly improves data consistency and comparability, establishes a unified spatiotemporal coordinate system and a three-dimensional grid, and combines bilinear interpolation with spatiotemporal conditional generative adversarial networks to fill in missing values, thereby enhancing data integrity and quality. This provides high-precision input for model training and improves the reliability and accuracy of loess mechanical parameter prediction.

[0014] 3. This invention adopts a block matrix approximation and parallel computing architecture, combined with CUDA streaming parallel processing and mixed precision storage strategy, which significantly reduces the time and resource consumption of massive spatiotemporal data computation. The block diagonal approximation method accelerates matrix inversion and optimizes model training and inference efficiency. The tensor expansion algorithm and fast covariance vector calculation realize efficient prediction of the mechanical parameters of the target spatiotemporal point.

[0015] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data, as described in this invention. Detailed Implementation

[0018] 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.

[0019] Example 1 Please see Figure 1 As shown, the present invention provides a method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data, comprising the following steps: Step S1: Collect loess mechanical parameter data, environmental factor data and geological structure data through multi-source sensors, construct spatiotemporally unified multimodal data, and divide the multimodal data into training set and test set; Step S2: Based on the multimodal data, design a collaborative kernel function that integrates spatial heterogeneity and temporal dynamics to quantify the correlation between mechanical parameters and spatiotemporal location and environmental factors; Step S3: Based on the collaborative kernel function, dynamically optimize the spatial weights, temporal decay coefficients, and multi-source data fusion weights to generate an interpolation model with adaptive parameters; Step S4: Using a block matrix approximation and parallel computing architecture, the interpolation model is trained and optimized in a distributed manner using the training set; Step S5: Using the trained interpolation model, calculate the predicted values ​​of the mechanical parameters of the target's spatiotemporal location in real time, and improve computational efficiency through tensor expansion algorithm.

[0020] Example 2 Based on Example 1, step S1 includes the following steps: Step S11: Drilling data Z-score standardization is performed to eliminate dimensional differences, as follows: ; In the formula, For the first Spatial coordinates of a borehole For the first Timestamps of borehole data This is the original borehole data. For standardized borehole data, This represents the mean of the borehole data. The standard deviation of the borehole data. For spatial coordinates, For timestamps, This represents different types of borehole data, including density, specific gravity, shear strength, etc. This represents the total number of borehole samples. Multi-source sensor data Normalize to [0,1] using Min-Max. This indicates the data collected by the sensor, including temperature, humidity, pore water pressure, etc. It is the first One sensor 3D feature vectors The number of feature dimensions of the sensor data; the collection of remote sensing data. And perform georegistration and radiometric correction. This is the first in remote sensing data i The feature vector of each sample The number of feature dimensions in the remote sensing data; This indicates that the data source is remote sensing, including vegetation index, surface temperature band data, etc. Step S12: Establish a unified spatiotemporal coordinate system. The spatial coordinates are UTM projection, and the time is unified as UNIX timestamp. Define a three-dimensional spatiotemporal grid. Map borehole data, sensor data, and remote sensing data to the unified three-dimensional spatiotemporal grid through bilinear interpolation. The spatial resolution is Δs=50m, and the temporal resolution is Δt=6h. Map each data source to the unified grid through bilinear interpolation. Step S13: Based on the data features and relationships of adjacent spatiotemporal points in the unified grid, missing values ​​are generated. A spatiotemporal conditional generative adversarial network (ST-GAN) is used to fill in the missing values. Spatiotemporal conditions refer to the dynamic correlation in the spatial and temporal dimensions. Establishing a unified spatiotemporal coordinate system and grid mapping is the foundation for spatiotemporal conditional modeling. Missing values ​​are data that still exist after being mapped to the unified spatiotemporal grid through bilinear interpolation. The ST-GAN is used to fill in the missing values. The generator's loss function is defined as a spatiotemporal gradient consistency constraint, as shown in the following formula: ; In the formula, Let the loss function of the generator be... As the expected value, This is a discriminator network used to distinguish between generated data and real data. This is a generator network used to supplement missing mechanical parameters. For spacetime coordinates, The weights of the gradient penalty term are... For the spatiotemporal gradient operator, It consists of multimodal spatiotemporal data, namely borehole data, sensor data, and remote sensing data.

[0021] Example 3 Based on Example 1, step S2 includes the following steps: Step S21: Based on the constructed multimodal data, define an adaptive spatiotemporal collaborative kernel function to fuse spatial proximity, temporal correlation, and multi-source data coupling features, as shown in the following formula: ; In the formula, For the spacetime kernel function, , The first The and the first j Spatial coordinates and timestamps of the sample For adaptive weight parameters, Represents the space core, Represents the time core; For the partial derivative terms of embedded physical laws, where, The mathematical expression of the physical equations embedded in the spacetime kernel function: For the first k The input variables are derived from three main categories: borehole data, sensor data, and remote sensing data. For multimodal data mapping functions, For the first Class input variables in space-time points The value on; physical equations include: (1) Cohesive field function:

[0022] (2) Internal friction angle field function:

[0023] (3) Compression modulus field function:

[0024] in, It is the initial cohesion. It is the initial internal friction angle. The initial compressive modulus, The spatiotemporal distribution of loess layer thickness. For the spatial and temporal distribution of water content, The spatial and temporal distribution of dry density For the spatiotemporal distribution of fracture density, For reference thickness, The above data are all from three main categories: borehole data, sensor data, and remote sensing data. These are the weight parameters.

[0025] Step S22: Fuse multi-source heterogeneous data to construct a tensor coupling operator: ; In the formula, For the first One core tensor component Factor vector of data source modality, The factor vector of spatial modes, The factor vector of the time modality; Represents the outer product of vectors; operator Used to define the first k Class data modal and the first k′ The coupling relationship between data modalities; Multimodal projection is achieved using an LSTM network, as shown in the following formula: ; In the formula, For the first The hidden state of class data, Representing the k Trainable weight parameters for class data modalities This is a 3D convolution operation.

[0026] The space core It can capture spatial non-uniformity and employs improved... The function, and its calculation formula, are as follows: ; In the formula, For improvement The spatial kernel of the function, These are the fixed coefficients of the Matern kernel function. For spatial Euclidean distance; The time kernel can characterize the annual periodic variation of loess mechanical parameters. Introducing a periodic component, the calculation formula is as follows: ; In the formula, The weights of the periodic components, This is a yearly cycle parameter, 365 days. This is the absolute value of the time difference.

[0027] Example 4 Based on Example 1, step S3 includes the following steps: Step S31: Construct the feature importance evaluation matrix The influence of quantification parameters on the prediction results is shown in the following formula: ; In the formula, For the first Weight parameters for the first Sensitivity of features It is a linear rectified activation function. For the spatiotemporal kernel function with respect to the weight parameters The partial derivatives reflect The sensitivity of changes to kernel function values, For the rate of change of predicted values ​​relative to input features The partial derivatives, Representing borehole data characteristics, sensor data characteristics, and remote sensing data characteristics, reflecting features Impact on model dynamic prediction; The weight parameters are dynamically updated using a gating mechanism, as shown in the following formula: ; In the formula, For the Softmax function, Update the matrix for weights. For loss function pairs The gradient; Step S32: During dynamic weight optimization, the regularization term acts as a constraint, working in conjunction with the loss function to ensure that the model parameters not only fit the data but also satisfy the humidity migration equation. Embedding the humidity migration equation as a regularization term ensures that the prediction results conform to the physical laws of humidity migration. The formula is as follows: ; In the formula, For gradient operators, Let be the humidity diffusivity tensor. The regularization coefficient is used. It is the regularization loss, which is part of the loss function and optimizes the interpolation model parameters during training.

[0028] Example 5 Based on Example 4, the interpolation model generated in step S4 involves a large-scale covariance matrix. After block processing, the local kernel matrix of each block only contains data from adjacent grid blocks. The inverse matrix is ​​approximated using the block diagonal structure, providing a data partitioning basis for parallel computing, including the following steps: Step S41: Divide the spatial domain into For grid blocks, a block diagonal approximation method is used to accelerate matrix inversion, as shown in the following formula: ; In the formula, For the first Local kernel matrix of each spatial block, For noise variance, This is the identity matrix for the corresponding dimension; Step S42: The grid block data after being divided into blocks is processed in parallel using CUDA streaming. Each stream independently computes matrix operations within the block. The mixed-precision storage strategy is as follows: The space kernel parameters are stored as FP16 to reduce memory usage; Time kernel parameters are stored in FP32 to ensure calculation accuracy; The coupling term uses INT8 quantization to reduce data transmission latency.

[0029] In step S5, based on the trained model, the new input is processed. The predicted value can be quickly calculated using tensor expansion: ; In the formula, For new input point The predicted value, Let the covariance vector of the new point and the training data be denoted as . The covariance matrix of the training data, For observation data vectors; in The calculation formula is as follows: ; In the formula, For tensor modes The product of This is a vectorization operation.

[0030] Example 6 Implementation scenarios and preparations: Target area: A roadbed engineering area in the Loess Plateau (area 10 km²), with complex geological conditions and significant spatiotemporal heterogeneity; Data source: Drilling data: The shear strength parameter cohesion of 15 boreholes was collected over a period of 2 years (sampling interval of 1 month). Sensor data: 50 sensor nodes are deployed to monitor soil moisture and temperature in real time (sampling frequency 1 time / hour). Remote sensing data: acquire multispectral satellite imagery (10m resolution), extract vegetation index (NDVI) and land surface temperature (LST) over a time span of 1 year (sampling interval of 15 days). Implementation steps: Step 1 Standardization and Normalization: Borehole data were standardized using Z-score and the mean was calculated. Standard deviation The standardized data distribution is as follows ; Sensor data (humidity, temperature) are normalized to [0,1]. Spatiotemporal mesh mapping: Define spatial resolution Δs=50m, temporal resolution Δt=6h, and generate a total of 200×200×2920 three-dimensional mesh nodes; The borehole data was mapped to grid nodes using bilinear interpolation, with a fill rate of 92%. Missing value imputation: A spatiotemporal conditional generative adversarial network (ST-GAN) was used, with a gradient penalty weight λ=0.1. The generator was trained iteratively for 500 epochs to impute the remaining 8% of missing data. The mean squared error between the imputed data and the measured values ​​is... ; Step Two Kernel function parameter settings: Space Core: Improvement Function, scale .

[0031] Time kernel: weights of periodic components The annual cycle is T=365 days, and the time scale is... sky; Multi-source data coupling weights The latent features of sensor and remote sensing data are extracted using an LSTM network (hidden layer dimension 64). Kernel function fusion: Adaptive weights Dynamically updated via a gating mechanism, with initial values ​​as follows: ; Step 3 Distributed training: The spatial domain is divided into 10×10×10 grid blocks, and CUDA streaming parallel computation is used. The mixed precision strategy is as follows: Spatial kernel parameters are stored as FP16 (reducing memory usage by 50%), and temporal kernel parameters are stored as FP32 (ensuring computational accuracy). The block diagonal approximation method accelerates the inversion of the covariance matrix, reducing the computation time from 8.2 hours to 1.5 hours using the traditional method; Physical constraint embedding: regularization term of humidity migration equation, diffusion coefficient Regularization coefficient ; Step Four Prediction target: A certain roadbed area (coordinates) within the next 24 hours. The shear strength parameter of cohesion; Prediction results: Model predicted values The measured value is absolute error (Relative error 3.8%) Traditional kriging prediction value ,error (Relative error 7.7%).

[0032] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0033] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data, characterized in that, The specific operating steps are as follows: Step S1: Collect loess mechanical parameter data, environmental factor data and geological structure data through multi-source sensors, construct spatiotemporally unified multimodal data, and divide the multimodal data into training set and test set; Step S2: Based on the multimodal data, design a collaborative kernel function that integrates spatial heterogeneity and temporal dynamics to quantify the correlation between mechanical parameters and spatiotemporal location and environmental factors; Step S3: Based on the collaborative kernel function, dynamically optimize the spatial weights, temporal decay coefficients, and multi-source data fusion weights to generate an interpolation model with adaptive parameters; Step S4: Using a block matrix approximation and parallel computing architecture, the interpolation model is trained and optimized in a distributed manner using the training set; Step S5: Using the trained interpolation model, calculate the predicted values ​​of the mechanical parameters of the target's spatiotemporal location in real time, and improve computational efficiency through tensor expansion algorithm.

2. The method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data according to claim 1, characterized in that: The multimodal data is a spatiotemporally unified three-dimensional data structure formed by integrating data from different sources, types, and dimensions. Step S1 includes the following steps: Step S11: Analyze the borehole data set. Z-score standardization is performed to eliminate dimensional differences, as shown in the following formula: ; In the formula, For the first Spatial coordinates of a borehole For the first Timestamps of borehole data This is the original borehole data. For standardized borehole data, This represents the mean of the borehole data. The standard deviation of the borehole data. For spatial coordinates, For timestamps, This represents different types of borehole data, including density, specific gravity, and shear strength. This represents the total number of borehole samples. Multi-source sensor datasets Normalize to [0,1] using Min-Max. This indicates the data collected by the sensor, including temperature, humidity, and pore water pressure. It is the first One sensor 3D feature vectors The number of feature dimensions of the sensor data; the collection of remote sensing data. And perform georegistration and radiometric correction. This is the first in remote sensing data i The feature vector of each sample The number of feature dimensions in the remote sensing data; This indicates that the data source is remote sensing, including spectral data of vegetation index and surface temperature; Step S12: Establish a unified spatiotemporal coordinate system. The space adopts UTM projection, the time is unified to UNIX timestamp, and a three-dimensional spatiotemporal grid is defined. The borehole data, sensor data and remote sensing data are mapped to the unified three-dimensional spatiotemporal grid through bilinear interpolation. Step S13: Use a spatiotemporal conditional generative adversarial network (ST-GAN) to fill in the missing values. The loss function of the generator is defined as a spatiotemporal gradient consistency constraint, as shown in the following formula: ； In the formula, Let the loss function of the generator be... For the expected value, This is a discriminator network used to distinguish between generated data and real data. This is a generator network used to supplement missing mechanical parameters. For spacetime coordinates, The weights of the gradient penalty term are... For the spatiotemporal gradient operator, It consists of multimodal spatiotemporal data, namely borehole data, sensor data, and remote sensing data.

3. The method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data according to claim 2, characterized in that: Step S2 includes the following steps: Step S21: Define an adaptive spatiotemporal collaborative kernel function. The collaborative kernel function integrates spatial proximity, temporal correlation, and multi-source data coupling characteristics. The formula is as follows: ； In the formula, For the spacetime kernel function, , The first The and the first j Spatial coordinates and timestamps of the sample For adaptive weight parameters, Represents the space core, Represents the time core; For the partial derivative terms of embedded physical laws, where, The mathematical expression of the physical equations embedded in the spacetime kernel function: For the first k The input variables are derived from three main categories: borehole data, sensor data, and remote sensing data. For multimodal data mapping functions, For the first Class input variables in space-time points The value on; Step S22: Fuse multi-source heterogeneous data to construct a tensor coupling operator: ； In the formula, For the first One core tensor component Factor vector of data source modality, The factor vector of spatial modes, The factor vector of the time modality; Represents the outer product of vectors; operator Used to define the first k Class data modal and the first k′ The coupling relationship between data modalities; Multimodal projection is achieved using an LSTM network, as shown in the following formula: ； In the formula, For the first The hidden state of class data, Representing the k Trainable weight parameters for class data modalities This is a 3D convolution operation.

4. The method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data according to claim 3, characterized in that: The space core It can capture spatial non-uniformity and employs improved... The function, and its calculation formula, are as follows: ； In the formula, For improvement The spatial kernel of the function, These are the fixed coefficients of the Matern kernel function. For spatial Euclidean distance; The time kernel can characterize the annual periodic variation of loess mechanical parameters. Introducing a periodic component, the calculation formula is as follows: ； In the formula, The weights of the periodic components, This is a yearly cycle parameter, 365 days. This is the absolute value of the time difference.

5. The method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data as described in claim 3, characterized in that: Step S3 includes the following steps: Step S31: Construct the feature importance evaluation matrix The impact of quantification weight parameters on the prediction results is expressed by the following formula: ； In the formula, For the first Weight parameters for the first Sensitivity of features It is a linear rectified activation function. For the spatiotemporal kernel function with respect to the weight parameters The partial derivatives, Representing spatial weights, temporal decay coefficients, and multi-source data fusion weight parameters, reflecting... The sensitivity of changes to kernel function values; For the rate of change of predicted values ​​relative to input features The partial derivatives, Representing borehole data characteristics, sensor data characteristics, and remote sensing data characteristics, reflecting features The impact on dynamic prediction of the interpolation model, 3 is a matrix number of rows; For matrix The number of columns represents the total feature dimension after multimodal data fusion; The weight parameters are dynamically updated using a gating mechanism, as shown in the following formula: ； In the formula, For the Softmax function, Update the matrix for weights. For the loss function with respect to the weight parameters The gradient; Step S32: Embed the humidity migration equation as a regularization term to make the predicted mechanical parameters conform to the physical laws of humidity migration. The loss of the regularization term is as follows: ； In the formula, For gradient operators, Let be the humidity diffusivity tensor. The regularization coefficient is . It is the regularization loss, which is part of the loss function and optimizes the interpolation model parameters during training.

6. The method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data as described in claim 5, characterized in that: Step S4 includes the following steps: Step S41: Divide the spatial domain into For grid blocks, a block diagonal approximation method is used to accelerate matrix inversion, as shown in the following formula: ； In the formula, For the first Local kernel matrix of each spatial block, For noise variance, This is the identity matrix for the corresponding dimension; B It represents the number of blocks in the grid. Among them, the convolution kernel weights and bias terms Distributed training is achieved through a data parallelism strategy; Step S42: Use CUDA streams to process each grid block in parallel, allocate an independent CUDA stream to each block, and compute the local inverse matrix of all blocks in parallel; use a mixed precision storage strategy to train the interpolation model.

7. The method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data as described in claim 6, characterized in that: The mixed-precision storage strategy in step S42 is as follows: The space kernel parameters are stored as FP16 to reduce memory usage; Time kernel parameters are stored in FP32 to ensure calculation accuracy; The coupling term uses INT8 quantization to reduce data transmission latency.

8. The method for constructing a spatiotemporal interpolation prediction model for loess mechanical parameters based on big data according to claim 6, characterized in that, In step S5, based on the trained interpolation model, the new input is processed. The predicted value can be quickly calculated using tensor expansion: ； In the formula, For new input point The predicted value, Let covariance vector be the sum of the variances between the new input point and the training set data. The covariance matrix of the training set data, For observation data vectors; in, The calculation formula is as follows: ； In the formula, Indicates vectorization operation, These correspond to data mode, spatial mode, and temporal mode parameters, respectively. It is a multimodal coupled core tensor.

Images