Deep learning based hybrid rock dynamics parameter prediction method

Abstract

The application relates to the technical field of geological engineering, and aims to improve the prediction accuracy of rock mechanics parameters, and provides a mixed rock dynamics parameter prediction method based on deep learning, features are extracted based on pretreated multi-source heterogeneous data, a graph neural network is constructed, a double deep learning model is constructed and trained, and the dynamics parameters are predicted based on the trained double deep learning model. The above method improves the prediction accuracy of rock mechanics parameters.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological engineering, and particularly relates to a mixed rock dynamics parameter prediction method based on deep learning. BACKGROUND

[0002] Mixed rocks in tectonic zones are widely distributed in strong tectonic activity regions such as orogenic belts, fault zones and shear zones, and are common and key geological bodies in water conservancy and hydropower engineering (such as dam foundation and tunnel surrounding rock), traffic engineering (such as slope and tunnel), mineral resource development (such as deep ore body surrounding rock) and geological disaster prevention engineering. Such rocks are formed by mechanical mixing, penetration metasomatism or deep melting of multiple lithologies such as gneiss and amphibolite as matrix and granite and pegmatite as vein under the action of strong tectonic stress, and have extremely significant heterogeneity (composition and structure change dramatically in a short distance), strong anisotropy (development of gneissic structure and banded structure), complex internal structure (fracture, mylonitization and fold deformation) and general post-alteration modification. These characteristics make the dynamics behavior (such as deformation, strength and wave velocity propagation) of mixed rocks extremely complex and difficult to accurately characterize.

[0003] Traditional dynamics parameter acquisition methods have great limitations. In-situ testing can reflect the mechanical response in the in-situ state, but the cost is high and the period is long. Large-scale in-situ testing requires special equipment, drilling and a large amount of manpower and material resources, and the cost of single-point testing can reach tens of thousands to hundreds of thousands of yuan. The representation is limited, and the test results only reflect the characteristics of the rock mass in a very small range (usually centimeters to meters) near the probe / sensor, and it is difficult to capture the dramatic heterogeneity and structural changes of mixed rocks in the spatial scale. Indoor rock mechanics tests can accurately control the test conditions and obtain various parameters such as dynamic / static elastic modulus, Poisson's ratio and strength parameters. However, it is difficult and destructive to obtain high-quality, large-size and directional undisturbed samples in strong deformation and easily broken mixed rocks, and the sampling process will cause damage to the rock mass. The scale effect is significant, and small-size samples (usually centimeters) cannot represent the complex structure and heterogeneity of mixed rock mass on the engineering scale (meters to tens of meters), resulting in low reliability of the test results.

[0004] In recent years, artificial intelligence has been applied in rock mechanics parameter prediction. However, the existing intelligent methods have fundamental deficiencies in solving the problem of predicting the dynamics parameters of mixed rocks in tectonic zones: lack of pertinence to the complexity of mixed rocks: general models are not optimized and designed for the unique high heterogeneity, anisotropy and complex structure of mixed rocks, and it is difficult to effectively extract and integrate key geological feature information. Insufficient use of data: mainly relying on single geophysical data (such as wave velocity) or simple lithology coding, failing to deeply mine and effectively integrate multi-source heterogeneous geological information, resulting in inaccurate rock mechanics parameter prediction. SUMMARY

[0005] In order to improve the prediction accuracy of rock mechanics parameters, the application provides a hybrid rock dynamics parameter prediction method based on deep learning.

[0006] The technical scheme adopted by the application to solve the above problems is:

[0007] The hybrid rock dynamics parameter prediction method based on deep learning comprises:

[0008] Step 1: Multi-source heterogeneous data acquisition and preprocessing, the multi-source heterogeneous data comprising geological attribute data, physical mechanics parameters, geophysical data and image data;

[0009] Step 2: Feature extraction based on the preprocessed data to obtain physical features, image features and cross-modal learning features, and construction of a graph neural network to obtain initial feature representation of a graph structure;

[0010] Step 3: Construction and training of a double deep learning model based on the features obtained in Step 2, the double deep learning model comprising:

[0011] GNN branch: directional aggregation of fault zone node messages, outputting construction of enhanced graph feature representation;

[0012] Transformer branch: focusing on the same lithology layer association by constructing an attention mask, outputting a formation-sensitive sequence feature representation;

[0013] Geological gating fusion module: dynamically integrating the features output by the GNN branch and the Transformer branch;

[0014] Hierarchical feature fusion module: fusion of physical features, image features and dynamically integrated features to generate fusion features;

[0015] Prediction module: prediction of mechanics parameters based on the fusion features;

[0016] Step 4: processing of the multi-source heterogeneous data of the region to be predicted in the same way as in Steps 1 and 2, and prediction of the dynamics parameters based on the trained double deep learning model.

[0017] Further, the preprocessing comprises:

[0018] Standardization of data: numerical data is standardized by Z-score, text data is vectorized by TF-IDF or word embedding, and image data is subjected to geometric correction and ROI segmentation;

[0019] Defect processing of data: missing values are filled in by using a geological constraint interpolation method or a geological analogy interpolation method, and abnormal values are identified and removed in combination with physical constraints and statistical detection;

[0020] Quantitative structural characteristics: quantifying the development degree of schistosity and anisotropy index;

[0021] Data augmentation: data is augmented by image rotation, random cropping, noise addition and rock physics model synthesis.

[0022] Further, step 2 specifically includes:

[0023] Based on the preprocessed data, physical features, image features and cross-modal learning features are obtained; the physical features include mineral composition features, structural complexity, physical correlation features and structural damage index; the cross-modal learning feature acquisition step is: based on the preprocessed data, geophysical parameter sequence, lithology word embedding and depth coordinate coding are obtained, and feature fusion is performed through lithology-geophysical cross attention;

[0024] Constructing a graph neural network: taking rock samples as nodes, defining edge weights based on spatial adjacency and lithological similarity, and updating node states through message passing to obtain initial feature representation of the graph structure.

[0025] Further, the image feature acquisition step is: a direction constraint convolution kernel is used to extract features from the preprocessed image, and the extracted features are input into a VGG network; in the VGG network, the feature maps of different convolution layers are fused, and finally an image feature vector is output.

[0026] Further, the attention mask is represented as: , represents the depth difference between nodes i and j, is a threshold value.

[0027] Further, the fusion step of the hierarchical feature fusion module is:

[0028] Feature grouping and linear transformation: the physical features and image features are linearly transformed and spliced to obtain , the linearly transformed and spliced features are linearly transformed to obtain ;

[0029] Element-wise multiplication is performed on and , input into a fully connected layer for dimension transformation, and a fusion feature is generated through a GeLU activation function.

[0030] Further, the fusion step of the hierarchical feature fusion module further includes: dimension reduction is performed on the fusion feature.

[0031] Further, a geology knowledge distillation loss function is used in the model training process: , wherein, is the value of the feature fused by the hierarchical feature fusion module at the kth sample, is the value of the expert experience feature on the kth sample, is the square of the L2 norm.

[0032] Further, step 3 further comprises: verifying and optimizing the trained model.

[0033] Further, the verification of the model comprises: accuracy verification, lithology specificity verification, physical rationality verification, generalization ability verification and uncertainty quantification, and the optimization strategy comprises: geological knowledge distillation, geological multi-objective Pareto optimization and adversarial sample regularization.

[0034] Compared with the prior art, the present application has the beneficial effects that: multi-source heterogeneous data covers macro to micro scale, providing a more comprehensive data basis for the prediction of mechanical parameters, effectively solving the problems of strong heterogeneity of mixed rock, multiple data missing and strong subjectivity of description, and improving the consistency, comparability and physical rationality of the data; based on the multi-source heterogeneous data, physical features are extracted, a lithology-geophysical cross-attention mechanism is used to associate text description and geophysical parameters, a geological map neural network is constructed, edge weights are defined based on spatial adjacency and lithology similarity, and topological relationship modeling between samples is realized; directional constraint convolution kernels are used to extract structure-sensitive image features, and the perception of key geological structures such as schistosity and anisotropy is enhanced, so that the model can fuse multi-modal information, capture the complex correlation between lithology-geophysical-space, and improve the representation ability of complex structures of mixed rock; the GNN branch focuses on node aggregation within the structure zone, enhances the response to key structures such as faults and shear zones, and the Transformer branch processes geophysical sequences and lithology embedding, focuses on the correlation within the same lithology layer through structure attention mask, so that the model not only has strong nonlinear fitting ability, but also maintains geological rationality; the HFF module fuses physical features, image features, graph features and sequence features in stages, and the multi-task prediction head simultaneously outputs multiple dynamic parameters such as compressive strength, elastic modulus and Poisson's ratio, the geological physics constraint loss ensures that the prediction result conforms to the law of rock mechanics, realizes high-precision, multi-parameter and physically consistent prediction, and is suitable for the simultaneous demand for multiple mechanical parameters in engineering practice; structure-specific analysis of rock group and physical rationality verification are proposed, cross-structure unit migration test, small sample robustness verification and uncertainty quantification are carried out, and optimization strategies such as geological knowledge distillation, adversarial sample regularization and multi-objective Pareto optimization are provided, so as to ensure that the model not only performs well on the training set, but also has strong generalization ability, high robustness and explainability, and is suitable for actual geological engineering scenes; through the knowledge distillation technology, the complex model is compressed into a lightweight model, the inference speed is improved by about 5 times, so that the present method can be deployed on site equipment to support real-time prediction and intelligent decision-making, and has engineering practical value. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1A deep learning-based hybrid rock dynamic parameter prediction method is provided.

[0036] Figure 2 A multi-source heterogeneous data acquisition and preprocessing flowchart is provided.

[0037] Figure 3 A feature extraction and graph neural network construction flowchart is provided.

[0038] Figure 4 A dual deep learning model processing flowchart is provided.

[0039] Figure 5 A model verification and optimization flowchart is provided. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0041] As shown in Figure 1 , the deep learning-based hybrid rock dynamic parameter prediction method comprises:

[0042] Step 1: Multi-source heterogeneous data acquisition and preprocessing, as shown in Figure 2 .

[0043] The multi-source heterogeneous data includes geological attribute data, physical and mechanical parameters, geophysical data and image data, covering macro to micro scales, as shown in Table 1.

[0044] Table 1: Multi-source heterogeneous data table

[0045]

[0046] The preprocessing of multi-source heterogeneous data includes the following steps:

[0047] Standardization of data: numerical data is standardized by Z-score to eliminate dimension: , where μ is the mean, is the standard deviation, and x is the original data; the processing objects include geophysical parameters (such as seismic wave velocity Vp, resistivity , magnetic susceptibility κ); time series / curve data (such as logging curve, seismic reflection profile).

[0048] For text-type geological description, first, a hybrid rock professional dictionary (such as the terms of “eyeball structure” and “mylonitization”) is constructed; then, a TF-IDF or word embedding model (Word2Vec) is used to generate a feature vector: , where d is the embedding dimension.

[0049] For image data, geometric correction is performed to eliminate shooting angle distortion; ROI segmentation is performed to extract the effective area of the core / slice.

[0050] Processing data defects: For the problem of missing field data, use geological constraint interpolation or geological analogy interpolation. The geological constraint interpolation method is: Wherein The difference interpolation function f can select K nearest neighbor mean (KNN): k represents the number of nearest neighbor data points selected most similar to the target data point (under the same tectonic unit, same lithology, etc.); The geological analogy interpolation method maps the missing data based on the known data of similar outcrops / drilling holes.

[0051] Identify and remove outliers by combining physical constraints and statistical detection: Physical constraints can exclude data that does not conform to basic physical laws due to measurement errors or other abnormal conditions, ensuring that the data is reasonable at the physical level: Remove data points that satisfy , where Vp is the P-wave velocity, p is the density, K is the bulk modulus, and G is the shear modulus. Statistical detection uses MAD (Median Absolute Deviation) for geophysical parameter sequences such as resistivity and susceptibility:

[0052] , ; wherein is a single data point in the data set X, and median() represents the median. This method is based on the statistical distribution characteristics of the data, with the median as the center, and identifies outliers by measuring the relative deviation of the data points from the median. Compared with the method based on mean and standard deviation, it is more robust to outliers (not easily affected by extreme values), and can detect data points that deviate from the overall distribution characteristics in the data set from a statistical perspective, helping to find abnormal fluctuations or error data in the data for further processing and improving data quality.

[0053] Quantify structural features: Quantify the development degree of schistosity and anisotropy index. Quantifying structural features converts qualitative geological descriptions into quantitative indicators: Schistosity development degree (D): Wherein is the schistosity development area, is the total area. This index quantifies the development degree of schistosity by calculating the ratio of the schistosity development area to the total area . Schistosity is a planar distribution of tectonic features in rock. This quantification method converts the qualitative schistosity development into a value between 0 and 1. The closer the value is to 1, the higher the proportion of schistosity development area to total area, i.e. the higher the development degree of schistosity; the closer the value is to 0, the lower the development degree of schistosity.

[0054] Anisotropy index ( ): ,in This index represents the P-wave velocity in three directions. It reflects the degree of difference in P-wave velocity across different directions, thus embodying the anisotropic characteristics of the rock. A larger index value indicates a greater difference in P-wave velocity across the three directions, and a stronger anisotropy; a smaller index value indicates a weaker anisotropy.

[0055] Data augmentation: Data is enhanced through image rotation, random cropping, noise addition, and synthesis using rock physics models.

[0056] Core image rotation (±5°): By rotating the core image within a certain angular range (±5 degrees), the observation of the core under different placement angles is simulated, increasing the diversity of image data. This allows the image-based training model to learn the characteristics of the core at different angles, improving the model's robustness to core image recognition, avoiding the model's adaptation to images at only specific angles, and enhancing its generalization ability.

[0057] Random cropping (preserving mineral boundaries): This involves randomly cropping the core image while preserving mineral boundaries. This is to enrich the data sample by changing the image's field of view and local content without losing key mineral information. Different cropped regions can showcase features from different parts of the core, expanding the variety of samples in the dataset and enabling the model to learn more detailed information about the local core areas.

[0058] Gaussian noise was added: Gaussian noise with a mean of 0 and a standard deviation of 0.01 was added to the core images. Real-world core images may be subject to various noise interferences; adding Gaussian noise simulates this real-world situation. The model is exposed to noisy images during training, which improves its resistance to noise and enhances its stability and accuracy in real-world noisy image recognition tasks.

[0059] Outcrop photos: Rock images of natural outcrops or engineering exposed surfaces (such as trenches and chambers) collected in the field. Due to the influence of lighting (shading in the morning and evening), shooting angle (shooting from below / above), and background interference (vegetation, debris), they need to be pre-processed and adapted.

[0060] 1. Geometric Normalization: Focusing on the Main Lithological Elements

[0061] Cropping ROI (Region of Interest): Use image segmentation tools (such as lithological region recognition based on color thresholds) to select pure lithological regions and remove irrelevant backgrounds such as vegetation and air, so that the model only learns "the rock itself";

[0062] Perspective Correction: For the inclined wall of the exploration trench / chamber, use perspective transformation algorithm to restore the "ladder-shaped rock wall" to "approximately rectangular", repair the true trend of bedding and fracture, and ensure the accuracy of structural feature analysis.

[0063] 2. Illumination Normalization: Eliminate Brightness Bias

[0064] Use Retinex algorithm to simulate the visual mechanism of human eye "adapting to different light", decompose the "light layer" (shadows, highlights) and "reflection layer" (true color of lithology, texture) of the image, remove shadows, balance brightness, and make the photos taken at different times and weather have more consistent lithological features, which is convenient for cross-sample comparison.

[0065] 3. Enhancement Strategy Adjustment: Adapt to Field Randomness

[0066] Relax rotation angle (±15°): There is no fixed angle for field shooting, and the actual rock may also appear "natural tilt" due to landslides and weathering. Rotating the image simulates more angles, allowing the model to learn "lithological features at different angles" and enhance generalization.

[0067] Increase "background blur": Use Gaussian blur to weaken the residual background (such as edge vegetation) and reduce noise interference, so that the model can focus on lithological texture and mineral structure.

[0068] Microscopic thin section images: Microscopic images (mineral crystals, pore structure, etc.) of rock thin sections collected by polarizing microscope (single polarizing, orthogonal polarizing) or scanning electron microscope (SEM) need to be enhanced:

[0069] 1. "Virtual Rotation" Based on Mineral Optical Properties

[0070] Principle: Under the polarizing microscope, mineral rotation will change the "extinction position and interference color order" (e.g., when hornblende is rotated 45°, the interference color is the most vibrant). Use image transformation algorithm (simulate stage rotation 0-360°, step 10°) to generate "series of virtual rotation images".

[0071] Geological value: Let the AI model learn "full-angle optical properties of the same mineral", solve the feature missing caused by "random angle" in field / thin section photos, and especially more accurate for analysis of porphyritic structure (porphyritic rotation characteristics) and directional structure (schistosity mineral orientation).

[0072] Implementation details:

[0073] Under orthogonal polarizing, focus on rotating to "45° extinction position" to enhance interference color features;

[0074] Under single polarizing, preserve the continuity of "mineral cleavage joint direction" to avoid destroying cleavage features.

[0075] 2. Focus on magnified cropping of "mineral assemblage patterns"

[0076] Principle: Mineral assemblages (such as secondary enlarged quartz boundaries and feldspar clayatization boundaries) are direct records of diagenetic events, but traditional random cropping severs the mineral contact relationships. A smart ROI cropping method is used instead: edge detection identifies mineral grain boundaries; the cropping area is defined as the "contact zone + 50 pixels on each side" (e.g., in the quartz-feldspar contact zone, retaining 50μm of minerals on each side).

[0077] For SEM images, focus on diagenetic structures such as "pore-mineral interfaces" and "clay film distribution".

[0078] Geological value: Force AI to learn "mineral assemblage patterns" (such as the distribution pattern of carbonate cementation and the morphology of feldspar dissolution pores), which are core evidence of reservoir properties (porosity, permeability) and diagenetic stages.

[0079] Generate synthetic data based on rock physics models:

[0080] , ,in, For longitudinal wave velocity, For saturated bulk modulus, It is the density of saturated rock. For the bulk modulus of mineral framework, For fluid bulk modulus, Porosity; by changing these parameter values, synthetic P-wave velocity data under different conditions can be generated. This increases the quantity and diversity of numerical data, providing more data support for related research and model training when actual field data is limited, and helps to more comprehensively analyze rock physical properties, etc.

[0081] After preprocessing, a unified feature matrix and label vector are generated:

[0082] Feature matrix Where N is the number of samples, D is the total dimension of the fused features, and the label vector is... The label vector Y is a dynamic parameter matrix, including parameters such as compressive strength. Dynamic parameters such as elastic modulus E and Poisson's ratio ν are used to quantitatively describe the dynamic properties of the sample in terms of mechanics and other aspects. These parameters serve as target variables or reference indicators in subsequent data analysis and modeling.

[0083] Step 2: Based on the preprocessed data, perform feature extraction to obtain physical features, image features, and cross-modal learning features, and construct a graph neural network to obtain the initial feature representation of the graph structure, such as... Figure 3 As shown.

[0084] Four types of interpretable features are constructed based on the feature matrix X output from Step 1 to describe the characteristics of rock samples from different perspectives, as shown in Table 2.

[0085] Table 2: Interpretable features with physical meaning

[0086]

[0087] Since different features have different dimensions and numerical ranges (e.g., The value of (bulk modulus) may be much larger than (fissure density)), to avoid some features dominating in subsequent models, normalization or standardization is needed. For rock sample i, the processed mineral composition features, structural complexity features, physical correlation features, and tectonic damage index are concatenated into a one-dimensional vector in a fixed order to obtain the "traditional physical feature vector" , i.e. , which integrates various features with physical meaning, providing basic data representation for subsequent data analysis and modeling based on rock mechanics and geological constraints, representing the dimension of the feature vector.

[0088] Extract tectonic sensitive image features: use direction-constrained convolution kernels to capture sheeting direction-related features, and combine VGG network multi-level feature fusion (shallow details + deep semantics) to generate image feature vectors.

[0089] Use direction-constrained convolution kernels to extract features from preprocessed images, and input the extracted features into the VGG network. In the VGG network, the feature maps of different convolution layers are fused, and finally the image feature vector is output.

[0090] Input the corrected image , where H represents the image height, W represents the image width, and 3 represents that the image has RGB three channels; then perform feature extraction based on direction-constrained convolution kernels: , where is the sheeting direction, is the Gaussian kernel, is the control scale, is the second derivative along the direction ; convolution operation: , is the convolution operation, and the output is the direction-sensitive feature map .

[0091] By taking the second derivative of the Gaussian kernel in a certain schistosity direction, the convolution kernel can capture the image feature changes along this direction, highlighting the image structure information related to the schistosity direction, thus achieving sensitive extraction of structural features in rock images. Compared with ordinary convolution kernels, direction-constrained convolution kernels can more specifically extract texture, structure and other features related to the rock schistosity direction, helping to more accurately describe the structural information of the rock and provide more effective feature data for subsequent analysis of the anisotropy and other properties of the rock.

[0092] The output of the direction-constrained convolution is directly taken as the input of the VGG: , , are the feature maps extracted by the 3rd and 5th layers of the VGG network, respectively.

[0093] The two layers of feature maps are fused by the MAXPOOL (max pooling) operation. Max pooling selects the maximum value in the local region of the feature map, which can retain more significant features and suppress noise and unimportant information; fusing feature maps of different levels makes the final image feature vector contains both detailed texture information (from the shallower layer3) and abstract semantic information (from the deeper layer5), which comprehensively improves the expression ability of the features and provides more abundant and representative data for subsequent image feature-based analysis and modeling.

[0094] Cross-modal sequence feature learning is performed. The geophysical parameter sequence, lithology word embedding and deep coordinate encoding are fused, and different modal information is associated through lithology-geophysical cross-attention to enhance the feature correlation.

[0095] The geophysical parameters, geological description words and geological positions from the geophysical feature column of X are taken as input for cross-modal feature learning.

[0096] Geophysical parameter sequence The geophysical parameter sequence comes from the geophysical feature column of X and contains original or derived parameters from geophysical measurements, such as resistivity ( ), magnetic susceptibility (κ), each sample has a scalar value on this column.

[0097] Geological description word embedding ; The geological description word embedding comes from the geological description text in the text vector column of X. The text vector column contains a fixed-dimensional vector representation of the processed geological description text. The geological description word embedding is a vector representation of the text information of the geological description, and the geological description word vector at each time step (or corresponding position) t is d-dimensional. Digitalizing the semantic information of the text facilitates subsequent processing and analysis. ​

[0098] Encode the geological position: where is the depth coordinate, is an adjustable parameter. By encoding the depth coordinate with sine and cosine functions, we assign specific vector representations to different depth positions. When processing sequence data, this allows the model to perceive the depth position information of the data, capture the geological feature rules that vary with depth, enhance the model's understanding of the spatial position relationship of geology, and help analyze the relationship between geophysical parameters and geological descriptions at different depths.

[0099] Lithology-geophysical cross-attention:

[0100] ,

[0101] where is the geophysical parameter sequence, is the geological description word, is the learnable weight matrix, and k is the loop variable used to sum all positions in the geological description sequence for normalization;

[0102] Through matrix multiplication measure the degree of association between the two, and then perform softmax normalization (exponential operation and sum normalization) to obtain the attention weight ; This allows the model to focus on the relevant information between the two when processing cross-modal data (geophysical parameters and geological description text), highlighting the relationship between lithology-related geological descriptions and geophysical parameters, enhancing the model's ability to integrate and understand different modal data, and improving the effectiveness of cross-modal feature learning.

[0103] Construct a geological map neural network: take rock samples as nodes (attributes include physical features and image features), define edge weights based on spatial adjacency and lithology similarity, and update node states through message passing to mine topological relationships between samples.

[0104] First, define the basic graph in the geological map neural network: ; where is the node set, is the edge set, is the feature set related to the node; node : rock sample (attributes: ); Each node represents a rock sample whose attributes combine traditional physical features (such as mineral composition, structural complexity, etc.) and image features (extracted by constructing sensitive image features), providing a comprehensive description of rock sample characteristics. : spatial proximity + lithology similarity, edge connection node, reflecting the relationship between nodes. Here the edge is based on spatial proximity and lithology similarity; , where measures the proximity of the spatial positions of nodes i and j (the closer the spatial distance, the greater the value), , is the scale of the tectonic unit, is the JS divergence of lithology description, used to measure lithology similarity (the smaller the JS divergence, the more similar the lithology). The product of the two determines the weight of the edge, and the greater the weight indicates the closer the connection between nodes.

[0105] Message passing: . In the update from layer l to layer l+1, is the hidden state of node i at layer l. The function aggregates the hidden states of the neighbor nodes of node i at layer l, is a learnable weight matrix that linearly transforms the aggregated information, is an activation function that introduces nonlinearity, and the hidden state of node i after updating at layer l+1 is obtained .

[0106] Through the message passing mechanism, nodes can receive and integrate the information of neighbor nodes, constantly update their own representation, and thus capture the association and topological information between nodes in the graph structure. In the field of geology, it can be used to mine the potential relationship between rock samples based on space and lithology, and better learn and model the features of geological data.

[0107] The geological graph neural network constructs a geological region into a computational graph, and converts the prior knowledge of "spatial relationship" and "lithology similarity" understood by geologists into mathematical structures that machines can learn and calculate. Through the message passing mechanism of the graph, the prediction of each sample fully considers its geological background and tectonic environment, thereby achieving more intelligent, reliable, and geologically regular prediction of the mixed rock dynamics parameters of highly heterogeneous tectonic areas. The initial feature representation of the graph structure is output.

[0108] Step 3: Build a double deep learning model based on the features obtained in step 2 and train it.

[0109] ​​The dual deep learning model includes a GNN branch and a Transformer branch, dynamically integrates the features of the two branches through a geological gating fusion module, and injects geological prior knowledge (such as enhancing tectonic feature weights through a diagonal matrix). The physical features, image features, and preliminary fusion features are integrated through hierarchical feature fusion (HFF) to output the fused deep features. The fused deep features are input to predict the dynamic parameters, such as Figure 4 as shown.

[0110] wherein the GNN branch: directional aggregation of fault zone node messages, output of tectonic-enhanced graph feature representation.

[0111] Node attributes: wherein is the density, is the P-wave velocity, is the quartz index (from physical features), is the schistosity intensity (from image features).

[0112] Geological edge weight: The embodiment continues the logic of step 2 "define edges based on spatial proximity + lithology similarity", and the graph structure is directly reused and optimized without the need to rebuild. Wherein spatial distance decay, wherein and are the spatial positions of nodes i and j, is the tectonic unit scale. This formula measures the influence of spatial distance on the connection between nodes. The closer the distance, the greater the weight, reflecting the constraint of spatial proximity on rock sample relationships. is the damage index similarity constraint, and are the tectonic damage indices corresponding to nodes i and j. The closer the damage index, the greater the weight, emphasizing the role of tectonic damage similarity in determining node connection relationships.

[0113] Directional message aggregation (for tectonic zones): , is the set of neighbor nodes related to node i in the tectonic zone, and the attributes , of node i and its neighbor node j are fused (input to MLP after splicing). This way can capture the correlation information between nodes in the tectonic zone based on tectonic features (such as fault strike), and aggregate relevant messages.

[0114] Output: , first extract all node hidden states at layer L through the max-pooling operation max pooling to highlight the most significant feature information. In addition to the directional message aggregation , the feature representation related to the structural belt is enhanced, and finally the structural enhanced graph representation is obtained for subsequent geological data analysis and modeling.

[0115] Transformer branch: input geophysical parameter, lithology and structural feature embedding vector, focus on the association with the same lithology layer through structural attention mask, output stratum sensitive sequence feature representation.

[0116] Input embedding: , geophysical parameters (P-wave velocity , resistivity ), lithology embedding (the basic data of lithology embedding comes from geological attribute data in multi-source heterogeneous data acquisition), structural feature vector (comes from geological attribute data, image data and physical parameters in multi-source heterogeneous data acquisition) represents the splicing operation), among them “geophysical parameters , , lithology embedding , structural features ” come from the output of cross-modal sequence learning, which integrates information from multiple sources to provide a more comprehensive input representation for the model.

[0117] Structural attention mask: , according to the depth difference and whether the lithology is the same, when the depth difference is less than the threshold and the lithology is the same, =0, allowing the model to focus on the sequence association at the corresponding position; otherwise, , in the attention calculation, the model is inhibited from paying attention to the association of these positions. The model is forced to focus on the sequence association within the same lithology layer and with similar depth, avoiding the model from learning irrelevant or cross-lithology layer noise association information, so as to more accurately capture the sequence features and change rules within the stratum, and improve the model's understanding of the stratum structure and lithology change.

[0118] Output stratum sensitive sequence representation; , where . First, calculate the attention weight by linear transformation and depth coordinate , then weighted sum the hidden state at each time step, realize depth weighted aggregation, highlight the more representative information of the stratum features at different depth positions, and obtain the stratum sensitive sequence representation This weighting method enables the model to adaptively adjust the attention to features in different positions according to the depth difference, enhancing the model's ability to capture the changes in the stratigraphic sequence in the depth dimension, and providing more effective feature representation for subsequent geological parameter prediction tasks.

[0119] The geology gating fusion module uses a dynamic feature selection mechanism: wherein is the graph embedding output by the geology graph neural network (GNN), representing topological structure and other related features between rock samples; is the sequence embedding output by the geology sequence Transformer, containing features of the stratigraphic sequence in the depth dimension. After linear transformation of , the Sigmoid gating function is processed to obtain a value between 0 and 1, controlling the contribution of sequence features, , on the basis of the original weight matrix , a diagonal matrix is added to enhance the weight of the part of the weight matrix related to the structural features. Then is multiplied element by element , and finally is linearly transformed and a bias is added to obtain the preliminary fusion features . Through the gating mechanism, the model can dynamically adjust the fusion ratio of graph features and sequence features according to task requirements, adaptively select features more important to the current task, avoid irrelevant or redundant features interference, and enhance the effectiveness and relevance of feature representation.

[0120] Subsequently, the hierarchical feature fusion (HFF) architecture further integrates physical features, image features, and graph features and sequence features output by the GNN and Transformer branches to generate the final fusion features for prediction.

[0121] Hierarchical feature fusion (HFF) is an innovative architecture designed to fuse different types of features. The types of features involved include (physical features), image features , and preliminary fusion features , which are fused through multiple stages of processing:

[0122] Feature grouping and linear transformation:

[0123] HFF first divides the input features of different types into two groups and performs linear transformation (weighted summation) to map them to a common feature space;

[0124] The first group (physical-image feature group): the intrinsic properties of the rock.

[0125] Physical feature vector : numerical values describing the "intrinsic physical nature" of the rock, such as density, resistivity, mineral composition proportion (such as quartz content 35%), mechanical parameters (such as elastic modulus), etc.

[0126] Image feature vector : "morphology / texture features" extracted from rock images (such as core photos, microscopic sections, etc.);

[0127] Linear combination and splicing of physical features ( ) and image features ( ) to form the first group of fusion features g1;

[0128] Let the physical feature vector be , and the image feature vector be , first do linear transformation on both, introduce learnable weight matrix , ( is the intermediate dimension, which can be set as needed), calculate: ; then splice and superimpose the bias , get the first group of fusion features .

[0129] The second group: the output features of the geology gating fusion module;

[0130] Input: output features of the geology gating fusion module ;

[0131] Linear transformation of : , where is a learnable weight matrix; splice and add bias , get the second group of fusion features ;

[0132] Perform element-wise multiplication operation on and to realize the deep interaction of the two groups of features: the multiplication operation strengthens the nonlinear correlation between features, automatically identifies the complementary relationship between physical-image features and sequence-image features; input the interaction result into the fully connected layer for dimension transformation, and generate the final fusion features through GeLU activation function.

[0133] Further, it also includes adaptive dimension reduction: , where These are the first d eigenvectors obtained from Principal Component Analysis (PCA). PCA is a commonly used dimensionality reduction technique that uses the fused feature matrix... Perform PCA analysis to identify the d most prominent feature directions in the data (by... (Representation), and then project the original fused features onto these main feature directions to obtain the final dimensionality-reduced features. Reducing feature dimensionality decreases redundant information in the data, simplifies model complexity, and retains the most critical information. Dimensionally reduced features not only accelerate model training and inference but also avoid the "curse of dimensionality" problem caused by high-dimensional data, improving the model's generalization ability and stability, and enabling the model to run more efficiently in subsequent applications.

[0134] The hierarchical feature fusion architecture can fully explore the intrinsic relationships between various features by fusing different types of features in stages and hierarchically. It integrates information from multiple aspects such as physics, images, sequences, and graph structures, providing more comprehensive and representative feature representations for subsequent geological data analysis, modeling, and prediction tasks, and improving the performance of the model in geological applications.

[0135] Parameter prediction head: Regarding the compressive strength of rock Elastic modulus E, Poisson's ratio Prediction is made using parameters such as... It is the fusion feature of the final output of the HFF module, first processed... Linear transformation, followed by activation function (e.g., ReLU) introduces nonlinearity, which is then combined with the weight vector. Perform the dot product and add a bias. , obtain the predicted values ​​of the parameters .

[0136] Geological and physical constraint loss: , Forced elastic modulus With strength It satisfies the empirical proportional relationship (the ratio of elastic modulus to strength is close to 200). Constrained Poisson's ratio With strength The ReLU function ensures that only differences that do not conform to empirical relationships are penalized.

[0137] Joint training objectives: , It is a predicted value. Compared with the true value The mean squared error measures the accuracy of the model's predictions; It is the introduced geological and physical constraint loss. is a weight coefficient, controlling the degree of influence of the geophysical constraint; is a weight decay term representing model parameters, is a weight decay coefficient, preventing model overfitting.

[0138] In the model training process, a geology knowledge distillation loss function is adopted: Here is the value of the fused feature of the hierarchical feature fusion module at the kth sample, is the value of the expert experience feature at the kth sample. By calculating the mean square error (square of L2 norm) of the two and summing them up, the difference between the fused feature and the expert experience feature is measured.

[0139] The prior knowledge accumulated by experts in the field of geology is integrated into the model training. By minimizing this loss function, the fused feature learned by the model is brought closer to the expert experience feature, guiding the model to learn a feature representation that conforms to the geological professional knowledge, improving the accuracy and reliability of the model in geological data analysis, and enhancing the model's understanding and modeling ability of geological phenomena using expert knowledge.

[0140] Further, after training is completed, model verification and optimization are also included. Model verification includes: precision verification, rock class specificity verification, physical reasonableness verification, generalization ability verification, and uncertainty quantification, as shown in Figure 4 .

[0141] Precision verification: benchmark model comparison experiment is conducted, and the comparison objects are as follows:

[0142] Hoek-Brown criterion: an empirical criterion widely used in rock mechanics, used to estimate the strength parameters of rock. It is based on rock integrity, geological strength index (GSI), and other parameters, and calculates rock compressive strength and shear strength through specific formulas, and has important applications in rock engineering design, stability analysis, etc.

[0143] Q-system correction formula: a rock mass quality evaluation and support design method proposed by the Norwegian Geotechnical Institute (NGI). By scoring multiple factors such as joint set number, joint roughness, and joint alteration degree of rock mass, the rock mass quality index Q is obtained, which is then used for design and analysis of tunnels, slopes, and other engineering projects.

[0144] Machine learning method:

[0145] Random Forest (RF) is an ensemble learning algorithm based on decision trees. It trains each tree by constructing multiple decision trees, randomly sampling samples and randomly selecting features, and finally combines the prediction results of multiple trees through voting or averaging. It has good noise resistance and generalization performance and is often used for classification and regression tasks.

[0146] Gradient Boosting Tree (GBDT): This is also an ensemble learning algorithm that trains a series of decision trees iteratively. Each tree is trained based on the prediction residuals of the previous trees, and the prediction accuracy of the model is improved by continuously fitting the residuals.

[0147] Single model:

[0148] Independent GNN: This refers to using graph neural networks alone for geological parameter prediction without fusing with other models. It primarily extracts and learns features based on rock sample relationship maps, mining topological information between samples.

[0149] Independent Transformer: This model processes geological sequence data independently, learning the characteristics of borehole data in the depth direction through mechanisms such as depth-lithology coding, without being combined with other models.

[0150] Evaluation indicators: , RMSE It is the actual value. is the predicted value, N is the sample size, and RMSE measures the average magnitude of the error between the predicted value and the true value. It is more sensitive to larger errors, and the smaller the value, the higher the prediction accuracy. middle It is the mean of the true values. It reflects the goodness of fit of the model to the data, and the value ranges from 0 to 1. The closer it is to 1, the better the model fits the data, that is, the closer the predicted value is to the true value.

[0151] Statistical significance test: ,Require ,in It is the root mean square error of the benchmark model (the object of comparison). It is the root mean square error of the model under study. It is the pooled variance, a statistic used to combine the variance information of two sample groups (the baseline model sample group and the model sample group under study). Its function is to more accurately estimate the overall variance of the prediction errors of the two models. and These represent the sample sizes of the baseline model and the model under study, respectively. The formula calculates the t-statistic, used to determine whether there is a significant difference in predictive accuracy between the studied model and the baseline model. Requirements: (corresponding (p<0.01), that is, when the absolute value of t-statistic is greater than 2.58, it is considered that the prediction accuracy of the studied model and the benchmark model is significantly different (at the significance level of 0.01), otherwise the model needs to be optimized.

[0152] Tectonite group-specific analysis;

[0153] Tectonite group grouping: By grouping, rocks with different mylonitization degrees can be distinguished, facilitating the study of the characteristics of rocks at different mylonitization stages, and providing a basis for the performance of subsequent analysis models on different types of rocks;

[0154] In-group accuracy evaluation: Among them represents that sample i belongs to the k group (k corresponds , , ), is the true value, is the predicted value, is the mean value of the true value of the k group. The formula calculates the determination coefficient within each group, which measures the goodness of fit of the model on the rock data in the group. It is required that , that is, the fitting effect of the model on each group of rocks reaches a certain standard, indicating that the fitting ability of the model on rock data with different mylonitization degrees is strong.

[0155] Physical rationality verification;

[0156] (1) Geomechanical law test

[0157] Strength-modulus relationship: , where N is the number of samples, is the predicted elastic modulus, is the predicted compressive strength, is an indicator function, which takes the value of 1 when the condition in the parentheses is true, and 0 when it is not. The formula calculates the proportion of samples that satisfy the elastic modulus to compressive strength ratio in the range of 50-200 to the total number of samples.

[0158] Because different migmatites have large differences in mineral composition and structure, the internal law of strength-modulus relationship may be different. Therefore, for different rock types, the proportion of samples that satisfy (i.e., the proportion of samples in each rock type that satisfy ) can be calculated. It is required that the of each rock type is greater than 85% (the threshold value can be adjusted according to the actual geological law), rather than only calculating the overall >90% for all samples.

[0159] Wave velocity-density correlation: where is the rock density, is the density mean, is the predicted P-wave velocity, is the P-wave velocity mean. This formula calculates the Pearson correlation coefficient between rock density and P-wave velocity, measuring the degree of linear correlation between the two.

[0160] In petrophysical properties, density and P-wave velocity usually have a certain correlation. By judging whether the correlation coefficient is greater than 0.7, it is tested whether the wave velocity-density correlation predicted by the model conforms to the actual physical law. If the correlation coefficient meets the condition, it means that the model can reasonably reflect the correlation between density and P-wave velocity, and the prediction result is physically reasonable; if it does not meet the condition, it means that the model has deviation in describing the relationship between the two physical parameters, and needs to be further adjusted and optimized.

[0161] (2) Tectonic response sensitivity verification

[0162] Fault zone strength attenuation detection: where is the fault zone sample set, denotes the number of samples in the fault zone sample set. is the predicted compressive strength of the jth sample in the fault zone, is the predicted compressive strength of the jth sample in the normal area (non-fault zone). This formula first calculates the absolute value of the difference between the predicted compressive strength of each sample in the fault zone and the corresponding sample in the normal area, then sums and averages these absolute values to obtain the average strength attenuation value .

[0163] By detecting whether the average strength attenuation value is greater than 30 MPa, the sensitivity of the model to tectonic response is verified. If the condition is met, it means that the model can capture the strength attenuation of rocks in the fault zone, i.e. the model is sensitive to tectonic changes, and the prediction result conforms to the influence of tectonics on rock strength in the actual geological situation; if it does not meet the condition, it means that the model may not accurately reflect the changes in rock strength in tectonic areas such as fault zones, and needs to be further optimized and adjusted to improve the ability of the model to describe tectonic characteristics.

[0164] Generalization ability verification;

[0165] (1) Cross-tectonic unit migration test, as shown in Table 3.

[0166] Table 3 Test set

[0167]

[0168] Generalization decay rate: , is the root mean square error of the model in the test region, is the root mean square error of the model in the training region. This formula measures the degree of attenuation of the prediction accuracy of the model when migrating from the training region to the test region.

[0169] By comparing the training region and the test region under different geological conditions, the generalization ability of the model when crossing tectonic units is tested. The allowable attenuation threshold (such as δ<15% under stress mechanism difference, δ<20% under temperature and pressure difference) is specified. If the generalization attenuation rate is less than the threshold, it means that the model can better adapt to the geological differences of different tectonic units and has strong generalization ability; otherwise, it means that the prediction accuracy of the model decreases significantly when facing different geological conditions, and the generalization ability needs to be improved.

[0170] Small sample robustness verification:

[0171] Convergence requirement: , where n is the sample size, a, b, c are model parameters. By adjusting the parameters to fit the curve of the root mean square error RMSE with the sample size n, it is required that b>0.3. The larger the value of b, the faster the RMSE decreases with the increase of the sample size; it is used to evaluate the robustness of the model under small sample conditions. When b>0.3 is satisfied, it means that the prediction accuracy can be quickly improved and stabilized with the increase of the sample size when the sample size is small, that is, the model has good adaptability and robustness to small sample data; if this condition is not met, it means that the performance of the model improves slowly in the small sample scenario, which may have problems of overfitting or insufficient use of small sample data, and needs to be improved and optimized to improve the small sample robustness.

[0172] Uncertainty quantification verification:

[0173] (1) Prediction interval effectiveness

[0174] Coverage probability test: , (k=1.96 corresponds to 95% confidence level), where N is the sample size, is the true value, is the predicted value, is the prediction uncertainty (such as standard deviation), (·) is an indicator function that takes the value of 1 when the condition in the parentheses is true, and is the prediction uncertainty (such as standard deviation), (·) is an indicator function that takes the value of 1 when the condition in the parentheses is true, and 0 when it is not. k=1.96 corresponds to 95% confidence level. This formula calculates the proportion of the total sample size that the true value falls within the prediction interval . It is used to test the effectiveness of the model's prediction interval. If Greater than 90% means that at a 95% confidence level, the prediction interval given by the model can cover at least 90% of the true values, i.e. the prediction interval has high reliability and can reasonably reflect the uncertainty of prediction; if this condition is not met, it means that the model prediction interval may be too narrow or too wide and cannot effectively cover the true value, and the model's uncertainty quantification method needs to be adjusted;

[0175] (2) Interval width rationality

[0176] Average width constraint: where 2 is the width of each sample prediction interval, and the average of the sum of the prediction interval widths of all samples is , is the maximum value in the true value. This formula measures the rationality of the prediction interval width by comparing the average prediction interval width with a certain proportion (here is 0.2 times) of the maximum value of the true value.

[0177] Model optimization strategy;

[0178] (1) Geological knowledge distillation

[0179] Expert feature alignment: ,

[0180] where N is the number of samples, is the feature representation after model fusion, is the expert feature (such as foliation density, anisotropy index, mineral orientation, etc.). This formula calculates the mean square error (L2 norm) between the model fusion feature and the expert feature, measuring the difference between the two;

[0181] (2) Geological multi-objective Pareto optimization

[0182] Optimization objective function: , , is the model parameter, is the mean square error loss, measuring the difference between the model prediction value and the true value; is the geophysical constraint loss, ensuring that the model prediction conforms to the geophysical law; is the average prediction interval width, used to measure the rationality of the prediction interval width. This method optimizes multiple objectives simultaneously to find the Pareto optimal solution.

[0183] (3) Adversarial sample regularization

[0184] Construct perturbed samples: where is the original sample, is the perturbation intensity hyperparameter, is the loss function with respect to the sampleGradient of , (·) is the sign function. By adding gradient-based perturbations on the original samples, the adversarial samples are constructed .

[0185] Robustness loss: , where M is the number of adversarial samples, (·) is the model prediction function, is the true value. This formula calculates the mean square error of the model's prediction value on the adversarial sample and the true value, measuring the prediction error of the model on the adversarial sample.

[0186] By introducing adversarial samples, the diversity and complexity of the model training data are increased, enabling the model to learn more robust feature representations, improving the model's resistance to input data perturbations, preventing model overfitting, enhancing the model's stability and reliability in practical applications, and improving the model's generalization ability.

[0187] Optimization strategies can be selected according to the problems exposed during verification, and can be used alone or in combination.

[0188] Table 4 Optimization strategies

[0189]

[0190] Finally, the prediction model is deployed to realize intelligent decision support for geological engineering scenarios. This includes model distillation compression:

[0191] Knowledge transfer formula: , This part is the mean square error (MSE) loss term. is the prediction value of the original complex model (teacher model), is the true label value. is the weight of this term. Mean square error is used to measure the gap between the teacher model's prediction and the true value, prompting the student model to learn an output close to the true value; This part is the loss term based on Kullback-Leibler (KL) divergence. is the prediction value of the lightweight student model, T is the temperature parameter, is the balance factor. KL divergence is used to measure the difference between the probability distributions of the student model and the teacher model after temperature scaling, allowing the student model to learn the generalization ability and internal representation of the teacher model. The temperature parameter T can adjust the smoothness of the probability distribution, and the balance factor is used to weigh the relative importance of the mean square error loss and the KL divergence loss;

[0192] Through model distillation compression and other technical means, redundant parameters and structures in the model are removed, the space required for model storage is greatly reduced, making the model more efficient in storage and transmission; the calculation amount is reduced, the required calculation resources and time are reduced when performing prediction and reasoning, thereby significantly accelerating the reasoning speed, which can be increased by about 5 times, enabling faster prediction of new data to meet the application scenarios with higher real-time requirements.

[0193] Step 4: Process the multi-source heterogeneous data of the to-be-predicted region in the same way as steps 1 and 2, and predict the kinetic parameters based on the trained double deep learning model.

[0194] 1. Multi-source heterogeneous data acquisition and preprocessing

[0195] Collect multi-source heterogeneous data of the target region with the same type and format as in step 1, and use the same preprocessing method as in step 1 for preprocessing.

[0196] 2. Feature extraction and construction of prediction graph structure

[0197] Calculate four types of interpretable features (mineral composition features, structural complexity, physical correlation features, and tectonic damage index) consistent with Table 2 to form the physical feature vector of each sample ;

[0198] Use the trained directional constraint convolution kernel and VGG feature extraction network to process new images to generate image feature vectors ;

[0199] Splice and as the initial attributes of the nodes of the graph neural network;

[0200] The sequence data of the to-be-predicted region is subjected to the same cross-modal sequence feature learning (deep encoding, lithology-geophysical cross-attention) as in step 2 to generate sequence feature representations;

[0201] Each rock sample in the to-be-predicted region is regarded as a node in the graph, and according to the spatial coordinates and lithology description of the sample, the edge weight formula (based on spatial distance and lithology similarity) defined in step 2 is used to construct the edges between nodes to form the graph structure of the to-be-predicted region.

[0202] This graph can be connected with the graph in the training phase, and the model will use the learned rules to infer the attributes of unknown nodes through the message passing mechanism using known node information.

[0203] 3. Model inference and parameter prediction

[0204] The processed features and data are input into the trained and optimized dual deep learning model for forward propagation (inference):

[0205] GNN branch: input the graph structure into the GNN branch, and obtain the constructed enhanced graph representation of the to-be-predicted sample through message passing and directional aggregation;

[0206] Transformer branch: input the processed sequence features into the Transformer branch to obtain a formation-sensitive sequence representation;

[0207] Feature fusion and prediction: dynamically integrate the output features of the dual branches through the geological gating fusion module, and then obtain the final fused features through hierarchical feature fusion (HFF) and adaptive dimension reduction ;

[0208] Parameter prediction head: input the fused features into the multi-task prediction head (·) to directly output the predicted values of the dynamic parameters of each to-be-predicted sample (node) .

[0209] 4. Result output

[0210] The model outputs the predicted values of the dynamic parameters of all to-be-predicted samples; associates the predicted results with the spatial position information of the samples to form a structured prediction result table; and automatically generates a prediction report including overall statistical information, predicted values of key areas, uncertainty analysis conclusions, etc.

Claims

1. A method for predicting rock dynamic parameters based on deep learning, characterized in that, The method comprises the following steps: Step 1: multi-source heterogeneous data acquisition and preprocessing, the multi-source heterogeneous data comprising geological attribute data, physical and mechanical parameters, geophysical data and image data; Step 2: feature extraction based on the preprocessed data to obtain physical features, image features and cross-modal learning features, and construction of a graph neural network to obtain initial feature representation of a graph structure; Step 3: construction and training of a double deep learning model based on the features obtained in step 2, the double deep learning model comprising: a GNN branch: directional aggregation of fault zone node messages, outputting enhanced graph feature representation; a Transformer branch: focusing on the same lithology layer association by constructing an attention mask, outputting a sequence feature representation sensitive to strata; a geological gating fusion module: dynamically integrating the features output by the GNN branch and the Transformer branch; a hierarchical feature fusion module: fusion of physical features, image features and dynamically integrated features to generate fused features; a prediction module: prediction of mechanical parameters based on the fused features; Step 4: processing of multi-source heterogeneous data of a region to be predicted in the same manner as in steps 1 and 2, and prediction of mechanical parameters based on the trained double deep learning model; Step 2 specifically comprises: obtaining physical features, image features and cross-modal learning features based on the preprocessed data; the physical features include mineral composition features, structural complexity, physical correlation features and tectonic damage index; the cross-modal learning features are obtained by obtaining geophysical parameter sequences, lithology word embedding and deep coordinate encoding based on the preprocessed data, and feature fusion is performed through lithology-geophysical cross-attention; construction of a graph neural network: taking rock samples as nodes, defining edge weights based on spatial adjacency and lithological similarity, and updating node states through message passing to obtain initial feature representation of a graph structure. 2.The hybrid rock-dynamics parameter prediction method based on deep learning according to claim 1, wherein, The preprocessing comprises: standardization of data: numerical data is standardized by Z-score, textual data is vectorized by TF-IDF or word embedding, and image data is subjected to geometric correction and ROI segmentation; processing of data defects: missing values are filled by geological constraint interpolation or geological analogy interpolation, and abnormal values are identified and removed in combination with physical constraints and statistical detection; quantification of tectonic features: quantification of schistosity development degree and anisotropy index; data enhancement: data is enhanced by image rotation, random cropping, noise addition and rock physics model synthesis. 3.The hybrid rock-dynamics parameter prediction method based on deep learning according to claim 1, wherein, The image feature acquisition step comprises: feature extraction of the preprocessed image by using a directionally constrained convolution kernel, input of the extracted features into a VGG network, fusion of feature maps of different convolution layers in the VGG network, and finally output of an image feature vector. 4.The hybrid rock-dynamics parameter prediction method based on deep learning according to claim 1, wherein, The attention mask is represented as: , represents the depth difference between nodes i, j, is a threshold value. 5.The hybrid rock-dynamics parameter prediction method based on deep learning according to claim 1, wherein, The fusion step of the hierarchical feature fusion module comprises: Feature grouping and linear transformation: linearly transform physical features and image features and concatenate to get Linearly transform the dynamically integrated features to get ; To and Element-wise multiplication is performed, the full connection layer is input for dimension transformation, and the fusion features are generated through the GeLU activation function. 6.The hybrid rock-dynamics parameter prediction method based on deep learning according to claim 5, wherein, The fusion step of the hierarchical feature fusion module further comprises dimensionality reduction of the fused features. 7.The hybrid rock-dynamics parameter prediction method based on deep learning according to claim 1, wherein, A geological knowledge distillation loss function is used in the model training process: wherein, is a value of the fused feature of the hierarchical feature fusion module at the kth sample, is a value of the expert experience feature at the kth sample, is a square of an L2 norm. 8.The hybrid rock-dynamics parameter prediction method based on deep learning according to claim 1, wherein, Step 3 further comprises verification and optimization of the trained model. 9.The hybrid rock-dynamics parameter prediction method based on deep learning according to claim 8, wherein, Verification of the model comprises accuracy verification, rock class specificity verification, physical reasonableness verification, generalization ability verification and uncertainty quantification, and optimization strategies comprise geological knowledge distillation, geological multi-objective Pareto optimization and adversarial sample regularization.

Images