Deep learning-based rock mass mechanical parameter prediction method and system

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a rock mass mechanical parameter prediction method and system based on deep learning. The method comprises the following steps: acquiring drilling, geophysical prospecting, remote sensing and rock test data; time-space reference unification processing is carried out; respectively coding into modal specific feature vectors; a fusion feature tensor is generated through a cross-modal alignment module; and inputting the deep neural network embedded with Mohr-Coulomb physical constraints, and outputting mechanical parameters such as uniaxial compressive strength and elastic modulus. The system comprises a multi-source data acquisition unit, a space-time unification unit, a feature coding unit, a cross-modal alignment unit and a physical constraint prediction unit. According to the method, multi-source data deep fusion and physical law collaborative optimization are realized, and the prediction precision and the engineering reliability are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method and system for predicting rock mechanics parameters based on deep learning. Background Technology

[0002] With the increasing demand for deep resource development, major geotechnical engineering, and geological disaster prevention, the accurate acquisition of rock mechanics parameters has become a core prerequisite for ensuring engineering safety and design rationality. Traditional acquisition methods mainly rely on field core sampling tests and laboratory physical and mechanical tests, which are time-consuming, costly, and difficult to cover the spatial variability of complex geological bodies.

[0003] To improve efficiency, prediction models based on empirical formulas or statistical regression have been widely introduced. However, such methods usually only use a single type of data and ignore the inherent correlation of multi-source information in the rock mass system, resulting in limited prediction generalization ability.

[0004] Deep learning-based methods for predicting rock mechanics parameters have attracted attention in recent years. These methods attempt to automatically uncover the nonlinear mapping relationship between input features and target parameters through neural networks. They typically use geological exploration data, geophysical well logging curves, remote sensing images, or numerical simulation results as input, aiming to replace some physical experiments.

[0005] However, existing deep learning models generally employ a decision-level fusion strategy when processing heterogeneous data from different acquisition mechanisms, spatiotemporal scales, and representation dimensions—that is, modeling each modality separately and then weighting or concatenating them at the output layer. This fragmented approach fails to establish cross-modal semantic alignment early in feature extraction, resulting in the loss or weakening of key geological structural information before fusion.

[0006] Existing technologies therefore face multiple bottlenecks: Multi-source data is difficult to align due to differences in format, resolution and physical meaning. Direct splicing can easily introduce noise and destroy the original semantics. Decision-level fusion ignores the synergistic constraints between modes in terms of physical mechanisms and cannot capture the deep correlation between the macroscopic mechanical response and microstructural characteristics of rock mass; The lack of a unified feature coding space leads to poor adaptability of the model to unseen geological scenes, and the prediction results often deviate significantly from the actual measured values, especially in complex structural areas such as fault zones and weak interlayers.

[0007] The aforementioned problems severely restrict the reliable deployment of intelligent prediction methods in practical engineering, and there is an urgent need for a rock mechanics parameter prediction framework that can achieve deep fusion of multimodal information at the data level. Summary of the Invention

[0008] This invention provides a method and system for predicting rock mechanics parameters based on deep learning, aiming to solve the technical problems of traditional prediction systems being unable to effectively integrate multi-source heterogeneous data, resulting in the loss of key geological information and large deviations between prediction results and field measurements.

[0009] The method constructs a multi-source geological data fusion framework under a unified spatiotemporal benchmark, and adopts a hierarchical feature extraction and cross-modal alignment mechanism to achieve deep fusion of borehole exploration data, geophysical exploration data, remote sensing image data and indoor rock test data. On this basis, a physically constrained deep neural network model is introduced to generate high-precision and highly robust rock mechanics parameter prediction results.

[0010] This invention provides a method for predicting rock mechanics parameters based on deep learning, which includes: acquiring multi-source heterogeneous geological data of a target area, wherein the multi-source heterogeneous geological data includes borehole exploration data, geophysical exploration data, remote sensing image data, and indoor rock test data; The multi-source heterogeneous geological data are subjected to spatiotemporal benchmark unification processing to generate a standardized geological dataset with a unified spatial coordinate system and timestamps; Modality-specific feature encoding is performed on various types of data in the standardized geological dataset to obtain borehole feature vectors, geophysical feature vectors, remote sensing feature vectors, and experimental feature vectors. The borehole feature vector, geophysical feature vector, remote sensing feature vector, and experimental feature vector are input into the cross-modal feature alignment module. By sharing the latent semantic space mapping, the geometric and semantic consistency alignment of the multi-source features is achieved, and the aligned multi-source fusion feature tensor is generated. The multi-source fusion feature tensor is input into a physically constrained deep neural network model. The physically constrained deep neural network model embeds the rock mass mechanical constitutive relation as a regularization term in the loss function and outputs the predicted values ​​of rock mass mechanical parameters of the target area. The rock mass mechanical parameters include uniaxial compressive strength, elastic modulus, Poisson's ratio, internal friction angle and cohesion.

[0011] As one embodiment of the present invention, the acquisition of multi-source heterogeneous geological data of the target area specifically includes: The borehole location coordinates, core description records, rock type labels, point load intensity index and sonic velocity profile data within the target area are obtained through the borehole exploration database. Geophysical exploration equipment is used to acquire seismic wave velocity profiles, resistivity imaging data, gravity anomaly fields, and magnetic anomaly field data of the target area. High-resolution optical images, synthetic aperture radar interferograms, and thermal infrared image data of the target area are acquired through satellite or airborne remote sensing platforms. Raw stress-strain curve data of uniaxial compression tests, triaxial shear tests, and Brazilian splitting tests of indoor rock specimens that match the lithology of the target area were obtained from the rock mechanics laboratory database.

[0012] As one embodiment of the present invention, the spatiotemporal benchmark unification processing of multi-source heterogeneous geological data specifically includes: Using the national geodetic coordinate system as a reference, all borehole locations, geophysical survey line nodes, remote sensing image pixel center points, and test sample collection locations are uniformly projected onto the same Cartesian coordinate system; virtual timestamps are assigned to asynchronously acquired data, with regional geological age or engineering exploration stage as the time dimension identifier; data with inconsistent spatial resolution are resampled, and bilinear interpolation or nearest neighbor interpolation methods are used to map all data to a unified spatial grid system based on a preset grid cell size, forming a three-dimensional voxelized standardized geological dataset.

[0013] As one embodiment of the present invention, the modality-specific feature encoding of various types of data in the standardized geological dataset specifically includes: For borehole exploration data, a one-dimensional convolutional neural network is used to extract lithological sequences and physical parameter variation characteristics along the depth direction, and output borehole feature vectors; For geophysical exploration data, a two-dimensional convolutional neural network is used to process the profile images, extract the horizontal and vertical physical property gradients and anomaly boundary features, and output geophysical feature vectors. For remote sensing image data, a pre-trained visual Transformer model is used to extract global semantic features of surface texture, linear structure and topographic relief, and output remote sensing feature vectors. For indoor rock test data, a long short-term memory network is used to perform time-series modeling of the original stress-strain curves, extract the dynamic response features of peak strength, residual strength and plastic deformation stages, and output the test feature vector.

[0014] As one embodiment of the present invention, the cross-modal feature alignment module uses a multi-head attention mechanism to construct a shared latent semantic space. Its input is four types of modal feature vectors. First, each modal feature is mapped to a query vector, key vector and value vector of the same dimension through linear projection. Then, the attention weight matrix between different modalities is calculated, with the borehole feature as the main query source and the other three types of features as key-value pairs, to generate weighted fusion context-aware features; Finally, through residual connections and layer normalization operations, the aligned multi-source fusion feature tensor is output, which is a vector sequence of fixed length, preserving the key discriminative information of each modality and eliminating scale and distribution differences between modalities.

[0015] As one embodiment of the present invention, the physical constraint deep neural network model is a fully connected feedforward neural network, whose input layer receives multi-source fused feature tensors, the hidden layer adopts a modified linear unit activation function, and the output layer generates predicted values ​​of five rock mass mechanical parameters. The loss function of the physical constraint deep neural network model is composed of a weighted sum of a data-driven loss term and a physical constraint regularization term; the data-driven loss term is the mean square error between the predicted value and the known experimental sample label. The physical constraint regularization term is constructed based on the Mohr-Coulomb strength criterion. When the predicted cohesion is less than 0 or the internal friction angle exceeds 0 degrees to the inventiveness range, a secondary penalty term is applied, and the predicted elastic modulus must be greater than 1 / 10 of the uniaxial compressive strength; otherwise, additional constraint loss is introduced. By jointly optimizing the composite loss function, we can ensure that the model output conforms to the basic physical laws of rock mechanics.

[0016] This invention also provides a deep learning-based rock mechanics parameter prediction system, which includes: The multi-source geological data acquisition unit is used to acquire borehole exploration data, geophysical exploration data, remote sensing image data, and indoor rock test data for the target area. The spatiotemporal reference unification processing unit is used to perform coordinate system unification, timestamp assignment, and spatial resampling on the multi-source heterogeneous geological data to generate a standardized geological dataset. Modal-specific feature encoding units are used to extract features from four types of data in a standardized geological dataset, generating borehole feature vectors, geophysical feature vectors, remote sensing feature vectors, and experimental feature vectors. The cross-modal feature alignment unit is used to input four types of feature vectors into the shared latent semantic space mapping module to generate an aligned multi-source fusion feature tensor. The physical constraint prediction unit is used to input the multi-source fused feature tensor into a deep neural network model embedded with the rock mass mechanics constitutive relation, and output the predicted values ​​of rock mass mechanics parameters of the target area.

[0017] In one embodiment of the present invention, the multi-source geological data acquisition unit is connected to a geological database, a geophysical instrument data export module, a remote sensing data service platform, and a rock mechanics test management system through a standardized interface, automatically pulling structured and unstructured geological data and performing format parsing and field mapping.

[0018] As one embodiment of the present invention, the spatiotemporal reference unification processing unit has a built-in coordinate transformation engine that supports seven-parameter Bursa transformation from any local coordinate system to the National 2000 Geodetic Coordinate System, and uses Kriging interpolation to perform spatial continuity processing on sparse borehole data to generate a three-dimensional attribute field covering the entire target area.

[0019] In one embodiment of the present invention, the one-dimensional convolutional neural network in the modality-specific feature encoding unit contains three convolutional layers with kernel sizes of 5, 3, and 3, and channel numbers of 64, 128, and 256, respectively; the two-dimensional convolutional neural network contains four convolutional blocks, each consisting of two convolutional layers and one max-pooling layer; the visual Transformer model contains twelve encoder layers, each containing a multi-head self-attention mechanism and a feedforward network; and the long short-term memory network contains a two-layer stacked structure with 512 hidden units.

[0020] In one embodiment of the present invention, the multi-head attention mechanism in the cross-modal feature alignment unit is set to have 8 heads, each head has a dimension of 64, and the total feature dimension is 512; the residual connection adopts identity mapping, and the layer normalization operation is performed after the output of each sub-layer.

[0021] In one embodiment of the present invention, the fully connected feedforward neural network in the physical constraint prediction unit contains four hidden layers, with 1024, 512, 256, and 128 neurons in each layer, respectively; the weight coefficient of the data-driven loss term in the composite loss function is 0.9, and the weight coefficient of the physical constraint regularization term is 0.1; the model training uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and a maximum iteration round of two hundred rounds.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention solves the problem of information fragmentation caused by differences in data format, coordinate system, sampling density and physical dimensions in traditional methods by constructing a multi-source geological data fusion framework under a unified spatiotemporal benchmark, and achieves seamless integration of borehole, geophysical, remote sensing and experimental data.

[0023] The proposed modality-specific feature encoding and cross-modal alignment mechanism effectively preserves the unique discriminative capabilities of each data source, while eliminating the semantic gap between modalities and improving the integrity and consistency of feature representation.

[0024] By embedding fundamental physical laws of rock mechanics, such as the Mohr-Coulomb strength criterion, into the loss function of the deep neural network as hard constraints, the prediction results are not only statistically optimal but also physically reasonable, avoiding the non-physical interpretations that may occur in traditional pure data-driven models. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical architecture of the rock mass mechanics parameter prediction method and system based on deep learning proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the cross-modal feature alignment module in this invention; Figure 3 This is a logical flowchart of the unified spatiotemporal reference processing of multi-source heterogeneous geological data in this invention; Figure 4 This is a flowchart of the multi-branch feature extraction logic of the modality-specific feature encoding unit in this invention; Figure 5 This is a schematic diagram illustrating the loss function composition and prediction mechanism of the physically constrained deep neural network model in this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the multi-source geological data acquisition unit and various external data sources in this invention. Detailed Implementation

[0026] Please refer to Figures 1 to 6 This invention provides a method and system for predicting rock mechanics parameters based on deep learning, aiming to solve the technical problems in traditional rock mechanics parameter prediction technology, such as the loss of key geological information and large deviation between prediction results and field measurements due to the inability to effectively integrate multi-source heterogeneous geological data.

[0027] The method constructs a multi-source geological data fusion framework under a unified spatiotemporal benchmark, adopts modality-specific feature encoding and cross-modal alignment mechanism to achieve deep fusion of borehole exploration data, geophysical exploration data, remote sensing image data and indoor rock test data, and introduces a physically constrained deep neural network model embedded with rock mass mechanical constitutive relations to output rock mass mechanical parameter prediction values ​​that conform to physical laws and have high accuracy.

[0028] As one embodiment of the present invention, the deep learning-based rock mechanics parameter prediction method includes the following steps: S1, acquire multi-source heterogeneous geological data of the target area; S2, perform spatiotemporal benchmark unification processing on the multi-source heterogeneous geological data to generate a standardized geological dataset; S3, perform mode-specific feature encoding on each type of data in the standardized geological dataset to obtain borehole feature vectors, geophysical feature vectors, remote sensing feature vectors and experimental feature vectors; S4, input the four types of feature vectors into the cross-modal feature alignment module to generate an aligned multi-source fusion feature tensor; S5, input the multi-source fusion feature tensor into the physical constraint deep neural network model, and output the predicted values ​​of rock mechanics parameters of the target area.

[0029] In step S1, multi-source heterogeneous geological data of the target area are acquired. The multi-source heterogeneous geological data includes borehole exploration data, geophysical exploration data, remote sensing image data, and indoor rock test data.

[0030] Specifically, the borehole location coordinates, core description records, rock type labels, point load intensity index and sonic velocity profile data within the target area are obtained through the borehole exploration database. Geophysical exploration equipment is used to acquire seismic wave velocity profiles, resistivity imaging data, gravity anomaly fields, and magnetic anomaly field data of the target area. High-resolution optical images, synthetic aperture radar interferograms, and thermal infrared image data of the target area are acquired through satellite or airborne remote sensing platforms. Raw stress-strain curve data of uniaxial compression tests, triaxial shear tests, and Brazilian splitting tests of indoor rock specimens that match the lithology of the target area were obtained from the rock mechanics laboratory database.

[0031] All data is stored in local or cloud databases in structured or unstructured form, and field content is automatically retrieved and parsed through standardized application programming interfaces to ensure data integrity and consistency.

[0032] In step S2, the multi-source heterogeneous geological data is subjected to spatiotemporal benchmark unification processing to generate a standardized geological dataset with a unified spatial coordinate system and timestamp.

[0033] The process first uses the National Geodetic Coordinate System 2000 as a reference to project all borehole locations, geophysical survey line nodes, remote sensing image pixel center points, and test sample collection locations onto the same Cartesian coordinate system.

[0034] For local coordinate system data, a seven-parameter Bursa transformation model is used to complete the coordinate system transformation, ensuring accurate spatial alignment.

[0035] Secondly, virtual timestamps are assigned to asynchronously acquired data, using regional geological age or engineering exploration stage as the time dimension identifier. For example, all borehole data are marked as "detailed exploration stage", remote sensing images are marked as "summer of 2023", and geophysical data are marked as "preliminary exploration stage".

[0036] Next, data with inconsistent spatial resolution are resampled. Based on a preset grid cell size (e.g., 10m×10m×5m), bilinear interpolation is used to smooth the continuous data (e.g., seismic wave velocity, resistivity), and nearest neighbor interpolation is used to fill the discrete data (e.g., rock type labels). Finally, all data are mapped to a unified three-dimensional voxel grid system to form a standardized geological dataset covering the entire target area.

[0037] For sparsely distributed borehole data, Kriging interpolation is used to perform spatial continuity processing to generate a complete three-dimensional attribute field, avoiding prediction bias caused by missing data.

[0038] In step S3, modality-specific feature encoding is performed on various types of data in the standardized geological dataset to obtain borehole feature vectors, geophysical feature vectors, remote sensing feature vectors, and experimental feature vectors. For borehole exploration data, a one-dimensional convolutional neural network is used to extract lithological sequences and physical parameter variation features along the depth direction.

[0039] This one-dimensional convolutional neural network contains three convolutional layers. The first layer has a kernel size of 5 and 64 channels. The second convolutional kernel has a size of 3 and 128 channels; the third convolutional kernel has a size of 3 and 256 channels; each layer is followed by batch normalization and modified linear unit activation functions, and finally outputs a fixed-length borehole feature vector through a global average pooling layer.

[0040] For geophysical exploration data, a two-dimensional convolutional neural network is used to process profile images and extract lateral and longitudinal physical property gradients and anomaly boundary features.

[0041] This two-dimensional convolutional neural network contains four convolutional blocks. Each block consists of two convolutional layers (with a kernel size of 3 and channel numbers of 64, 128, 256, and 512 respectively) and a max pooling layer (with a pooling window of 2×2). Finally, it outputs the geophysical feature vector through a flattening operation.

[0042] For remote sensing image data, a pre-trained visual Transformer model is used to extract global semantic features of surface texture, linear structures, and topographic relief. This visual Transformer model consists of twelve encoder layers, each containing a multi-head self-attention mechanism (12 heads, 64 dimensions per head) and a feedforward network (3072 hidden layer dimensions). The input image is divided into 16×16 pixel image patches, linearly embedded, and then positional encoding is added. Finally, the output of the category token is taken as the remote sensing feature vector.

[0043] For indoor rock test data, a long short-term memory network was used to perform time-series modeling of the original stress-strain curves, and dynamic response characteristics of peak strength, residual strength and plastic deformation stage were extracted.

[0044] This Long Short-Term Memory (LSTM) network consists of two stacked layers, with 512 hidden units in each layer. The input is a normalized stress-strain time series, and the output is the hidden state of the last time step as the experimental feature vector.

[0045] In step S4, the borehole feature vector, geophysical feature vector, remote sensing feature vector, and experimental feature vector are input to the cross-modal feature alignment module. By sharing the latent semantic space mapping, the geometric and semantic consistency alignment of the multi-source features is achieved, and the aligned multi-source fusion feature tensor is generated.

[0046] This cross-modal feature alignment module employs a multi-head attention mechanism to construct a shared latent semantic space.

[0047] First, the four modal feature vectors are mapped to query vectors, key vectors and value vectors of the same dimension through independent linear projection matrices. The total feature dimension is 512, the number of heads is 8, and the dimension of each head is 64.

[0048] Subsequently, using the borehole feature vector as the main query source and the other three types of feature vectors as key-value pairs, the attention weight matrix is ​​calculated.

[0049] The formula for calculating attention weights is: ; in This is the query matrix after projecting the borehole features. For the joint bond matrix, For the joint value matrix, For each head dimension, This represents the transpose of a matrix.

[0050] Through this mechanism, the model dynamically assigns contribution weights of different modal features to the current borehole location. For example, in areas with obvious faults, the weight of linear structural features in remote sensing images is increased. In regions of high resistivity anomalies, the weight of geophysical features is enhanced.

[0051] Finally, the weighted fused context-aware features are residually connected with the original borehole features, and layer normalization is performed to output the aligned multi-source fused feature tensor.

[0052] This tensor is a fixed-length vector sequence that retains key discriminative information for each mode and eliminates scale and distribution differences between modes, providing a consistent and rich input representation for subsequent predictions.

[0053] In step S5, the multi-source fusion feature tensor is input into the physical constraint deep neural network model. The physical constraint deep neural network model embeds the rock mass mechanics constitutive relation as a regularization term in the loss function and outputs the predicted value of the rock mass mechanics parameters of the target area.

[0054] The physical constraint deep neural network model is a fully connected feedforward neural network containing four hidden layers with 1024, 512, 256, and 128 neurons per layer, respectively. The activation function for each layer is a modified linear unit (MRU). The output layer contains five neurons, corresponding to uniaxial compressive strength, elastic modulus, Poisson's ratio, internal friction angle, and cohesion. The model's loss function is a weighted sum of a data-driven loss term and a physical constraint regularization term, expressed as: ; in, , Data-driven loss term This represents the mean squared error between the predicted value and the known test sample labels. Physical constraint regularization term. Constructed based on the Mohr-Coulomb strength criterion, its specific form is as follows: ; in, For the predicted cohesion, For the predicted internal friction angle, For the predicted uniaxial compressive strength, This is the predicted elastic modulus.

[0055] This regularization term ensures that the cohesion is non-negative, the internal friction angle is between zero and zero degrees, and the elastic modulus is greater than 1 / 10 of the uniaxial compressive strength.

[0056] If the predicted value violates the above physical laws, a secondary penalty is applied to guide the model parameter update direction toward physical plausibility.

[0057] The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and a maximum of 200 iterations.

[0058] An early stopping strategy is adopted during training. Training is terminated when the validation set loss does not decrease for ten consecutive rounds to prevent overfitting.

[0059] The deep learning-based rock mechanics parameter prediction system includes a multi-source geological data acquisition unit, a spatiotemporal benchmark unification processing unit, a modal-specific feature encoding unit, a cross-modal feature alignment unit, and a physical constraint prediction unit.

[0060] The multi-source geological data acquisition unit connects to the geological database, geophysical instrument data export module, remote sensing data service platform, and rock mechanics test management system through standardized interfaces, automatically pulling structured and unstructured geological data and performing format parsing and field mapping.

[0061] The spatiotemporal reference unification processing unit has a built-in coordinate transformation engine that supports seven-parameter Bursa transformation from any local coordinate system to the national 2000 geodetic coordinate system. It also uses Kriging interpolation to perform spatial continuity processing on sparse borehole data, generating a three-dimensional attribute field covering the entire target area.

[0062] The one-dimensional convolutional neural network, two-dimensional convolutional neural network, visual Transformer model, and long short-term memory network in the modality-specific feature encoding unit are configured according to the aforementioned structure and independently process their respective modal data.

[0063] The multi-head attention mechanism in the cross-modal feature alignment unit is set to 8 heads, each head has a dimension of 64, and the total feature dimension is 512; the residual connection adopts identity mapping, and the layer normalization operation is performed after the output of each sub-layer.

[0064] The fully connected feedforward neural network in the physical constraint prediction unit contains four hidden layers. The weight coefficient of the data-driven loss term in the composite loss function is 0.9, and the weight coefficient of the physical constraint regularization term is 0.1. The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and a maximum number of iterations of 200.

[0065] This embodiment achieves deep fusion of multi-source heterogeneous geological data and high-precision prediction under physical constraints through the aforementioned methods and systems. Practical applications show that, in verification at multiple typical engineering sites, the average absolute error of the uniaxial compressive strength prediction is low, and the relative error of the elastic modulus prediction is small, outperforming existing single-source data source models or simple multi-source splicing models. This reduces the uncertainty in the geotechnical engineering investigation stage and provides highly reliable parameter inputs for slope stability analysis, underground cavern support design, and foundation bearing capacity assessment.

[0066] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting rock mass mechanical parameters based on deep learning, characterized in that, include: Acquire multi-source heterogeneous geological data of the target area, including borehole exploration data, geophysical exploration data, remote sensing image data, and indoor rock test data; The multi-source heterogeneous geological data are subjected to spatiotemporal benchmark unification processing to generate a standardized geological dataset with a unified spatial coordinate system and timestamps; Modality-specific feature encoding is performed on various types of data in the standardized geological dataset to obtain borehole feature vectors, geophysical feature vectors, remote sensing feature vectors, and experimental feature vectors. The borehole feature vector, geophysical feature vector, remote sensing feature vector, and experimental feature vector are input into the cross-modal feature alignment module. By sharing the latent semantic space mapping, the geometric and semantic consistency alignment of the multi-source features is achieved, and the aligned multi-source fusion feature tensor is generated. The multi-source fusion feature tensor is input into a physically constrained deep neural network model. The physically constrained deep neural network model embeds the rock mass mechanical constitutive relation as a regularization term in the loss function and outputs the predicted values ​​of rock mass mechanical parameters of the target area. The rock mass mechanical parameters include uniaxial compressive strength, elastic modulus, Poisson's ratio, internal friction angle and cohesion.

2. The rock mass mechanics parameter prediction method based on deep learning according to claim 1, characterized in that, The process of unifying the spatiotemporal reference of the multi-source heterogeneous geological data to generate a standardized geological dataset with a unified spatial coordinate system and timestamps includes: Using the national geodetic coordinate system as a reference, all borehole locations, geophysical survey line nodes, remote sensing image pixel center points, and test sample collection locations are uniformly projected onto the same plane rectangular coordinate system. A virtual timestamp is assigned to asynchronously acquired data, using regional geological age or engineering exploration stage as the time dimension identifier; Data with inconsistent spatial resolution are resampled. Based on the preset grid cell size, bilinear interpolation or nearest neighbor interpolation methods are used to map all data to a unified spatial grid system, forming a standardized three-dimensional voxelized geological dataset.

3. The rock mass mechanics parameter prediction method based on deep learning according to claim 2, characterized in that, The process involves performing mode-specific feature encoding on various types of data in the standardized geological dataset to obtain borehole feature vectors, geophysical feature vectors, remote sensing feature vectors, and experimental feature vectors, including: For borehole exploration data, a one-dimensional convolutional neural network is used to extract lithological sequences and physical parameter variation characteristics along the depth direction, and output borehole feature vectors; For geophysical exploration data, a two-dimensional convolutional neural network is used to process the profile images, extract the horizontal and vertical physical property gradients and anomaly boundary features, and output geophysical feature vectors. For remote sensing image data, a pre-trained visual Transformer model is used to extract global semantic features of surface texture, linear structure and topographic relief, and output remote sensing feature vectors. For indoor rock test data, a long short-term memory network is used to perform time-series modeling of the original stress-strain curves, extract the dynamic response features of peak strength, residual strength and plastic deformation stages, and output the test feature vector.

4. The rock mass mechanics parameter prediction method based on deep learning according to claim 3, characterized in that, The step of inputting the borehole feature vector, geophysical feature vector, remote sensing feature vector, and experimental feature vector into the cross-modal feature alignment module, and achieving geometric and semantic consistency alignment of multi-source features through shared latent semantic space mapping, generates an aligned multi-source fusion feature tensor, including: The four types of modal feature vectors are mapped to query vectors, key vectors, and value vectors of a unified dimension through linear projection. Using borehole feature vectors as the main query source and the other three types of feature vectors as key-value pairs, we calculate the attention weight matrix between different modalities. Based on the attention weight matrix, a weighted fusion context-aware feature is generated, and an aligned multi-source fusion feature tensor is output through residual connections and layer normalization operations.

5. The rock mass mechanics parameter prediction method based on deep learning according to claim 4, characterized in that, The step of mapping the four types of modal feature vectors to query vectors, key vectors, and value vectors of a unified dimension through linear projection includes: Independent linear transformations are applied to the borehole feature vector, geophysical feature vector, remote sensing feature vector, and experimental feature vector respectively, so that their dimensions match the unified dimension.

6. The rock mass mechanics parameter prediction method based on deep learning according to claim 5, characterized in that, The process of using borehole feature vectors as the primary query source and the other three types of feature vectors as key-value pairs to calculate the attention weight matrix between different modalities includes: The geophysical feature vectors, remote sensing feature vectors, and experimental feature vectors are concatenated into a joint bond matrix and a joint value matrix; Attention weights are calculated using a scaled dot product attention mechanism, and the formula is as follows: ； in This is the query matrix after projecting the borehole features. For the joint bond matrix, For the joint value matrix, For each head dimension, This represents the transpose of a matrix.

7. The rock mass mechanics parameter prediction method based on deep learning according to claim 6, characterized in that, The process of inputting the multi-source fused feature tensor into a physically constrained deep neural network model, wherein the physically constrained deep neural network model embeds rock mechanics constitutive relations as regularization terms in the loss function, and outputs predicted values ​​of rock mechanics parameters for the target region, includes: The multi-source fusion feature tensor is input into a fully connected feedforward neural network, which contains four hidden layers and the output layer generates predicted values ​​for five rock mass mechanics parameters. Construct a composite loss function, which is composed of a weighted sum of a data-driven loss term and a physical constraint regularization term; The data-driven loss term is the mean square error between the predicted value and the known experimental sample label; The physical constraint regularization terms are constructed based on the Mohr-Coulomb strength criterion.

8. The rock mass mechanics parameter prediction method based on deep learning according to claim 7, characterized in that, The expression for the physical constraint regularization term is: ； in, For the predicted cohesion, For the predicted internal friction angle, For the predicted uniaxial compressive strength, This is the predicted elastic modulus.

9. A rock mechanics parameter prediction system based on deep learning, characterized in that, include: The multi-source geological data acquisition unit is used to acquire borehole exploration data, geophysical exploration data, remote sensing image data, and indoor rock test data for the target area. The spatiotemporal reference unification processing unit is used to perform coordinate system unification, timestamp assignment, and spatial resampling on the multi-source heterogeneous geological data to generate a standardized geological dataset. Modal-specific feature encoding units are used to extract features from four types of data in a standardized geological dataset, generating borehole feature vectors, geophysical feature vectors, remote sensing feature vectors, and experimental feature vectors. The cross-modal feature alignment unit is used to input four types of feature vectors into the shared latent semantic space mapping module to generate an aligned multi-source fusion feature tensor. The physical constraint prediction unit is used to input the multi-source fused feature tensor into a deep neural network model embedded with the rock mass mechanics constitutive relation, and output the predicted values ​​of rock mass mechanics parameters of the target area.

10. The rock mass mechanics parameter prediction system based on deep learning according to claim 9, characterized in that, The modality-specific feature encoding unit is used for: For borehole exploration data, a one-dimensional convolutional neural network is used to extract lithological sequences and physical parameter variation characteristics along the depth direction, and output borehole feature vectors; For geophysical exploration data, a two-dimensional convolutional neural network is used to process the profile images, extract the horizontal and vertical physical property gradients and anomaly boundary features, and output geophysical feature vectors. For remote sensing image data, a pre-trained visual Transformer model is used to extract global semantic features of surface texture, linear structure and topographic relief, and output remote sensing feature vectors. For indoor rock test data, a long short-term memory network is used to perform time-series modeling of the original stress-strain curves, extract the dynamic response features of peak strength, residual strength and plastic deformation stages, and output the test feature vector.

Images