Soil-rock anisotropic interface shear strength prediction model construction method

Abstract

The invention discloses a soil-rock anisotropic interface shear strength prediction model construction method, which comprises the following steps: acquiring lithology, structure and interface bonding parameters of a soil-rock interface, extracting multi-parameter correlation characteristics through a lithology constraint correlation network, learning interface bonding strength characteristics through a heterogeneous interface bonding strength graph neural network, and outputting a characteristic matrix; and then, optimizing the mapping parameters by using a policy gradient algorithm guided by a fabric tensor, inputting the optimized parameters into a multi-modal federated training platform to realize distributed data cooperative training to generate intermediate model parameters, and finally, constructing a shear strength prediction model based on the parameters. According to the method, multiple types of networks and algorithms are integrated step by step, multi-parameter complex association is accurately captured, data privacy and multi-source data utilization are considered through federal training, dependence on a large number of field tests is not needed, and adaptability and prediction accuracy of the model to different engineering scenes are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of soil and rock shear strength prediction technology, and in particular to a method for constructing a soil-rock anisotropic interface shear strength prediction model. Background Technology

[0002] In civil and geological engineering activities, the stability of soil-rock anisotropic interfaces directly affects the safety performance of engineering structures such as slopes, tunnels, and foundations, and shear strength is a key indicator for assessing the stability of this interface. Traditional soil-rock interface shear strength assessments mostly rely on field tests or empirical formulas. However, field tests are limited by factors such as terrain and environment, resulting in high costs, long cycles, and difficulties in data acquisition. Empirical formulas, on the other hand, cannot fully encompass the complex relationships between multiple dimensions of parameters, including lithology, fabric, and interface bonding. With the application of artificial intelligence and big data technologies in engineering, shear strength prediction based on network models has become a research hotspot. However, in the process of building existing models, how to effectively integrate multi-source parameter features, achieve distributed data collaborative training, and improve the model's adaptability to complex interface characteristics remains a key problem to be solved. There is an urgent need for a predictive model construction method that can integrate multi-dimensional parameters, optimize feature extraction and parameter updates, and balance data security and model accuracy.

[0003] Existing technologies have two significant drawbacks in constructing predictive models for the shear strength of soil-rock anisotropic interfaces: Firstly, existing models mostly use a single network structure for feature extraction, failing to fully consider the nonlinear correlation between lithological parameters, fabric parameters, and interface bonding parameters. They also lack targeted constraint mechanisms and feature fusion strategies, resulting in extracted features that cannot accurately reflect the influencing factors of interface shear strength, thus affecting the model's prediction performance. Secondly, during the training process of existing models, they either rely on centralized data training, failing to effectively utilize multi-source data in distributed engineering scenarios, or neglect data privacy protection requirements during collaborative training, failing to construct a reasonable federated training mechanism. This leads to low data utilization, insufficient model generalization ability, and difficulty in adapting to the prediction needs of soil-rock anisotropic interface shear strength in different engineering scenarios. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides a method for constructing a predictive model for the shear strength of soil-rock anisotropic interfaces.

[0005] The technical solution adopted in this invention is a method for constructing a predictive model for the shear strength of a soil-rock heterogeneous interface, comprising the following steps: S1, obtaining lithological parameters, fabric parameters, and interface bonding parameters of the soil-rock heterogeneous interface, wherein the lithological parameters include the proportion of lithological components and the degree of lithological compactness, the fabric parameters include the particle arrangement direction and particle distribution density, and the interface bonding parameters include the magnitude of the interface bonding force and the area of ​​the interface bonding region; S2, inputting the parameters obtained in S1 into a lithologically constrained Transformer network, and using the multi-head attention mechanism of this network to extract features of the correlation between lithological parameters, fabric parameters, and interface bonding parameters, obtaining a lithological-fabric-bonding correlation feature vector; S3, inputting the correlation feature vector obtained in S2 into a heterogeneous interface bonding strength graph neural network, and using the node embedding method of the graph neural network to analyze the interface bonding region. S3: Learn the bonding strength characteristics of each node and output the interface bonding strength feature matrix; S4: Input the feature matrix output in S3 into the structure tensor-guided policy gradient algorithm. Using the structure tensor as the guiding factor, iteratively optimize the policy network parameters of the algorithm through the policy gradient update rule to obtain the optimized structure-bond strength mapping parameters; S5: Input the mapping parameters obtained in S4 into the multimodal federated training platform. This platform performs federated training on the soil-rock interface data of multiple distributed nodes and performs collaborative updates of model parameters without sharing the original data, generating intermediate model parameters after collaborative training; S6: Based on the intermediate model parameters obtained in S5, construct a soil-rock anisotropic interface shear strength prediction model. By adjusting the output layer parameters of the model, the model can output the corresponding shear strength prediction results according to the input soil-rock interface parameters.

[0006] Furthermore, the feature extraction process of the lithology-constrained Transformer network satisfies the following expression: ,in, This represents the lithology-texture-cohesion correlation feature vector. This represents the original feature matrix composed of the input lithological parameters, fabric parameters, and interface bonding parameters. This represents the multi-head attention calculation function. Presentation layer normalization operation, Represents the self-attention computation function. This represents the computation function of the feedforward neural network.

[0007] Furthermore, the process by which the heterogeneous interface adhesion strength mapping neural network outputs the interface adhesion strength feature matrix satisfies the following expression: ,in, This represents the characteristic matrix of interfacial bond strength. This represents the activation function. Representing nodes in a graph neural network The set of neighboring nodes, Represents a node With nodes Adjacency matrix elements between them This represents the weight matrix of a graph neural network. Representing nodes respectively ,node The initial eigenvectors.

[0008] Furthermore, the parameter iterative optimization process of the constructed tensor guided strategy gradient algorithm satisfies the following expression: ,in, These represent the policy network parameters before and after optimization, respectively. Indicates the learning rate. Indicates the parameter gradient calculation, Indicating in strategy Expectation calculation below, Representing the trajectory Cumulative rewards Represents the structure tensor.

[0009] Furthermore, the collaborative training process performed by the multimodal federated training platform satisfies the following expression: ,in, This represents the intermediate model parameters after co-training. Indicates the number of distributed nodes. Indicates the first The number of samples per node Indicates the first The model parameters are obtained by training each node locally.

[0010] Furthermore, the process by which the soil-rock anisotropic interface shear strength prediction model outputs the prediction results satisfies the following expression: ,in, This indicates the predicted shear strength. This represents the output layer weight matrix. This represents the hidden layer weight matrix. This represents the input soil-rock interface parameters. This represents the hidden layer bias vector. This represents the output layer bias vector. This represents the ReLU activation function.

[0011] Further, S3 includes the following sub-steps: S31, performing node mapping processing on the lithology-facies-bond correlation feature vector, mapping each feature dimension in the correlation feature vector to a node in the heterogeneous interface bond strength graph neural network, forming an initial node feature set; S32, constructing an adjacency matrix of the graph neural network based on the geometric structure of the interface bond region, where the element values ​​are determined according to the spatial distance between nodes and the correlation of bond strength; S33, inputting the initial node feature set and the adjacency matrix into the convolutional layer of the graph neural network, updating the node features through convolution operations, and capturing the local feature correlation between nodes; S34, performing global pooling on the node features output by the convolutional layer, integrating the feature information of all nodes, and generating an interface bond strength feature matrix.

[0012] Further, step S4 includes the following sub-steps: S41, converting the interface adhesion strength feature matrix into a feature format that can be processed by the policy gradient algorithm, and extracting the calibration feature values ​​in the feature matrix as the input state of the algorithm; S42, introducing a configuration tensor, fusing each component of the configuration tensor with the input state to form an extended state vector with configuration guidance information; S43, based on the extended state vector, generating an action probability distribution for parameter updates through the policy network, and selecting the parameter update direction according to the action probability distribution; S44, calculating the reward value corresponding to the parameter update direction, adjusting the reward weight in combination with the guidance effect of the configuration tensor, and updating the policy network parameters using the policy gradient formula.

[0013] Further, S5 includes the following sub-steps: S51, the multimodal federated training platform receives local model parameter update requests sent by each distributed node and verifies the node identity and model parameter format in the request; S52, after verification, the platform obtains the sample quantity information of each node and calculates the weight coefficient of the model parameters of each node according to the sample quantity; S53, the model parameters of each node are weighted and summed according to the weight coefficient to obtain the preliminary global model parameters; S54, the preliminary global model parameters are checked for consistency. If the parameters meet the preset consistency threshold, they are used as intermediate model parameters after collaborative training; otherwise, the weighted calculation is performed again.

[0014] A method for constructing a predictive model for the shear strength of a soil-rock heterogeneous interface is disclosed. This method is implemented through different units, including: a multi-parameter acquisition unit for the soil-rock interface, connected to lithological parameter detection equipment, fabric parameter measurement equipment, and interface bond parameter testing equipment, used to receive and integrate lithological parameters, fabric parameters, and interface bond parameters acquired by each device; a lithological constraint feature extraction unit, connected to the multi-parameter acquisition unit, which incorporates a lithological constraint Transformer network for performing correlation feature extraction on the acquired parameters; and an interface bond strength graph neural network calculation unit, connected to the lithological constraint feature extraction unit, which calculates the correlation feature of the heterogeneous interface bond strength graph. The bonding strength graph neural network processes the associated features and outputs the interface bonding strength feature matrix. The configuration-guided parameter optimization unit, connected to the interface bonding strength graph neural computation unit, optimizes the parameters corresponding to the feature matrix using a configuration tensor-guided gradient algorithm. The federated collaborative training unit, connected to the configuration-guided parameter optimization unit, constructs a multimodal federated training platform for collaborative updating of model parameters across distributed nodes. The shear strength prediction model construction unit, connected to the federated collaborative training unit, adjusts the output layer parameters based on the collaboratively updated model parameters to construct a model capable of outputting prediction results for the shear strength of soil-rock anisotropic interfaces.

[0015] Beneficial Effects: This invention proposes a method for constructing a predictive model for the shear strength of soil-rock anisotropic interfaces. By integrating techniques such as lithological constraint feature extraction, neural computation of the interface bond strength map, fabric-guided parameter optimization, and multimodal federated training in stages, it first utilizes a lithological constraint-related network to fully explore the nonlinear correlations between lithology, fabric, and interface bond parameters. Combined with targeted constraint mechanisms and feature fusion strategies, this solves the problems of inaccurate feature extraction from single-network structures in existing models and their inability to reflect key influencing factors of shear strength, thus improving the model's adaptability to interface characteristics. Then, through a fabric tensor-guided parameter optimization algorithm, key parameters are optimized. Iterative optimization further enhances the model's efficiency in utilizing multi-dimensional parameters. Simultaneously, through a multimodal federated training platform, collaborative training of distributed node data is achieved without sharing the original data. This avoids the limitations of centralized training in utilizing multi-source data while meeting data privacy protection requirements. It solves the problems of low data utilization and insufficient generalization ability of existing models. The final prediction model does not require extensive field trials, reducing costs and time. It can also integrate multi-source parameters, optimize feature extraction and parameter update processes, and balance data security and prediction accuracy, making it effectively applicable to shear strength prediction in different engineering scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method steps of the present invention;

[0017] Figure 2This is a diagram showing the unit composition for implementing the method of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, a method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface includes the following steps:

[0020] S1. Obtain the lithological parameters, fabric parameters and interface bonding parameters of the soil-rock anisotropic interface. The lithological parameters include the proportion of lithological components and the degree of lithological compaction. The fabric parameters include the particle arrangement direction and particle distribution density. The interface bonding parameters include the magnitude of the interface bonding force and the area of ​​the interface bonding region.

[0021] Specifically, step S1 is implemented as follows: The lithological parameters, fabric parameters, and interfacial bonding parameters of the soil-rock anisotropic interface are obtained by combining on-site drilling sampling with precise laboratory testing. The lithological parameters specifically include the proportion of lithological components (sandstone 30%-70%, clay 10%-50%, gravel 5%-20%) and lithological density (measured by porosity, ranging from 5%-35%, measured using mercury intrusion porosimetry to determine the pore volume percentage). The fabric parameters specifically include the particle orientation (expressed as the angle with the interface normal, ranging from 0°-90°, observed and statistically analyzed using a scanning electron microscope to determine the angle of the particle's long axis) and particle density (number of particles per unit volume, ranging from 10³ particles / m³ to 10⁻¹⁰ particles / m³). 5 The particle count is calculated per cubic centimeter and then converted to cubic meters. Specific interfacial bonding parameters include the interfacial bonding force (range 5 kPa-50 kPa, measured in a laboratory using a direct shear apparatus to simulate the interface's stress state) and the interfacial bonding area (range 0.1 m²-10 m², measured based on the actual dimensions of the sampled interface and converted to a standard area). This step collaboratively collects multi-dimensional parameters using multiple methods to ensure that the parameters cover the core influencing factors of the interface's physical and mechanical properties. This provides comprehensive and accurate basic data support for subsequent feature extraction and model construction, avoiding systematic deviations in subsequent model training due to missing parameters, measurement errors, or a single dimension, thus ensuring the reliability and completeness of the model input data.

[0022] S2. Input the parameters obtained in S1 into the lithology-constrained Transformer network, and use the multi-head attention mechanism of the network to extract the features of the relationship between lithology parameters, fabric parameters, and interface bonding parameters to obtain the lithology-fabrication-bond correlation feature vector.

[0023] Specifically, the implementation process of step S2 is as follows: the lithological parameters, fabric parameters, and interface bonding parameters obtained in S1 are converted into a standardized data format according to the preset data format specifications (numerical values ​​are uniformly retained to two decimal places, and parameter dimensions are classified and labeled according to lithology, fabric, and bonding), and then input into the lithological constrained Transformer network; this network uses a multi-head attention mechanism composed of 8-12 attention heads to process multiple sets of data, such as the proportion of lithological components and the direction of grain arrangement, the degree of lithological compactness and the grain distribution density, the magnitude of interface bonding force and the area of ​​the bonding region, and the proportion of lithological components and the magnitude of interface bonding force. The relationships between parameters are explored layer by layer, with each attention head focusing on capturing a set of parameter relationships. The encoder layer of the network is set to 6-10 layers. Each layer first captures the local relationships between parameters through self-attention calculation (the attention weights are calculated using a scaled dot product). Then, a feedforward neural network (with 512-1024 hidden layer neurons) performs a nonlinear transformation on the relationship information. After layer normalization, the final output is a lithology-facies-cohesion relationship feature vector with a dimension of 256-512 (each dimension of the vector corresponds to a set of feature values ​​for parameter relationships). This step, through targeted network structure design, strengthens the constraint effect of lithological parameters on fabric parameters and interface cohesion parameters, enabling the extracted feature vector to accurately reflect the nonlinear intrinsic relationship between multiple parameters. This avoids the problem of isolated single-parameter features, laying a high-quality, highly correlated feature foundation for subsequent interface cohesion strength feature learning and improving the ability of features to represent the influence of interface shear strength.

[0024] S3. Input the associated feature vector obtained in S2 into the heterogeneous interface bonding strength graph neural network. Based on the node embedding method of the graph neural network, learn the bonding strength features of each node in the interface bonding region and output the interface bonding strength feature matrix.

[0025] Specifically, step S3 is implemented as follows: the lithology-facies-cohesion correlation feature vector obtained in S2 is uniformly split according to the node dimension, with each feature dimension corresponding to a node in the heterogeneous interface bonding strength graph neural network, forming an initial node feature set including 128-256 nodes (each node has a feature dimension of 1, and the feature value directly uses the value of the corresponding dimension of the correlation feature vector); based on the three-dimensional geometric coordinates of the interface bonding region (obtained by a laser scanner to acquire interface surface coordinate data), the Euclidean distance between any two nodes is calculated, and nodes with a distance less than 0.5m are determined to be adjacent nodes, while nodes with a distance greater than or equal to 0.5m are determined to be non-adjacent nodes, thereby constructing a 0-1 adjacency matrix (adjacent nodes correspond to...). (Position elements are set to 1, non-adjacent nodes to 0); the graph neural network is configured with 3-5 convolutional layers. Each layer uses a mean aggregation function to aggregate the features of adjacent nodes (the arithmetic mean of the feature values ​​of all adjacent nodes for each node is taken during aggregation). Then, a ReLU activation function is used to perform a non-linear transformation on the aggregated features (the activation function threshold is set to 0, feature values ​​less than 0 are mapped to 0, and feature values ​​greater than 0 remain unchanged). Finally, a global average pooling operation is used (calculating the average of the feature values ​​of all nodes after convolution processing) to integrate the features of all nodes into an interface adhesion strength feature matrix with dimensions of 64-128 (the number of rows in the matrix equals the feature dimension, the number of columns is 1, and each element corresponds to the mean of a type of adhesion strength feature). This step, through the node embedding and feature aggregation capabilities of the graph neural network, accurately learns the adhesion strength features of each local region within the interface adhesion area, while also taking into account the feature correlation between regions. This ensures that the output feature matrix can comprehensively reflect the spatial distribution characteristics and overall variation law of interface adhesion strength, avoiding the problems of missing local features or blurring overall features, and providing accurate feature basis for subsequent parameter optimization.

[0026] S4. Input the feature matrix output from S3 into the configuration tensor to guide the policy gradient algorithm. Using the configuration tensor as the guiding factor, iteratively optimize the policy network parameters of the algorithm through the policy gradient update rule to obtain the optimized configuration-bond strength mapping parameters.

[0027] Specifically, step S4 is implemented as follows: the interface adhesion strength feature matrix output from S3 is flattened (the matrix elements are arranged in row-major order) into a one-dimensional vector (the vector length is equal to the total number of elements in the feature matrix), which serves as the input state for the configuration tensor-guided policy gradient algorithm; a configuration tensor (a second-order tensor composed of particle alignment direction and particle distribution density, where the tensor elements are the particle alignment direction angle and particle distribution density values, respectively) is introduced, and the two components of the configuration tensor are multiplied element-wise with the input state vector (each component is multiplied with all elements of the input state vector to obtain two expanded vectors, which are then concatenated into one vector) to form an expanded state vector (the vector length is twice the length of the input state vector); the algorithm's policy network uses three fully connected layers, with the number of neurons in the input layer equal to the length of the expanded state vector, and the number of neurons in the hidden layers set to 256 and 128 respectively (the activation function of the first hidden layer is ReLU, and the second is tanh), and the number of neurons in the output layer equal to the number of parameters to be updated. The output is a probability distribution of parameter update values ​​(each parameter update value corresponds to a probability value, and the sum of probabilities is 1). After selecting the parameter update direction with the highest probability according to the probability distribution, the reduction value of the interface bonding strength prediction error under this direction is calculated as a reward (the error is calculated using mean square error, and the reward value is the absolute value of the error reduction). The reward value is adjusted by combining the weights of the structure tensor components (the weight of particle arrangement direction ranges from 0.3 to 0.5, and the weight of particle distribution density ranges from 0.4 to 0.7, and the sum of the two weights is 1) (adjusted reward value = particle arrangement direction weight × error reduction value + particle distribution density weight × error reduction value). Then, the policy network parameters are iteratively updated using the gradient ascent method (the learning rate is set to 0.005-0.01, and the parameter values ​​are adjusted according to the gradient direction in each iteration to maximize the reward). The number of iterations is set to 50-100 times. The reward value is calculated after each iteration. The iteration stops when the change in reward value is less than 0.001 for 5 consecutive iterations, and the optimized structure-bonding strength mapping parameters are obtained. This step uses the configuration tensor as a guide and employs multiple rounds of iterative optimization via the policy gradient algorithm to ensure that the mapping parameters accurately reflect the influence of configuration characteristics on the interface bonding strength. At the same time, it avoids parameter optimization from getting stuck in local optima, improves the adaptability of parameters to actual interface characteristics, and provides optimized core parameter support for subsequent federated training.

[0028] S5. Input the mapping parameters obtained in S4 into the multimodal federated training platform. This platform performs federated training on the soil-rock interface data of multiple distributed nodes, and performs collaborative updates of model parameters without sharing the original data, generating intermediate model parameters after collaborative training.

[0029] Specifically, the implementation process of step S5 is as follows: The texture-bond strength mapping parameters obtained in S4 are uploaded to the multimodal federated training platform according to the preset data transmission protocol. This platform is deployed on the cloud server and includes 5-10 distributed nodes (each node corresponds to an engineering scenario, such as slope engineering node, tunnel engineering node, foundation engineering node, etc.). Each node stores soil-rock interface data under the corresponding engineering scenario (each node has 1000-5000 sets of samples, and each set of samples includes complete lithology, texture, interface bond parameters and corresponding measured shear strength values). The platform uses a federated averaging algorithm to first train the local model of each node for 10-20 rounds (stochastic gradient descent is used during local training, with a learning rate of 0.001-0.005 and a batch size of 32-64). During the training process, each node only updates the local model parameters based on the local data and does not upload the original data. Initial data is sent to the middle platform to ensure data privacy and security. After each round of local training, each node uploads its local model parameters (including weight matrix and bias vector) and the corresponding number of samples to the middle platform. The middle platform calculates the weight of each node's parameters based on the number of samples (weight = number of samples in the node / total number of samples in all nodes, with a weight range of 0.05-0.3), and performs a weighted summation on parameters of the same type in each node (e.g., the first-layer weight matrix of all nodes is weighted and summed) to obtain the global model parameters. The global model parameters are then distributed to each node, and each node replaces the corresponding parameters in its local model based on the global parameters before performing the next round of local training. This process of "local training - parameter upload - global aggregation - parameter distribution" is repeated for 5-8 rounds until the global model parameters converge (the mean square error of the global parameters in two adjacent rounds is less than 0.0001), generating the intermediate model parameters after collaborative training. This step, through federated training, enables the collaborative use of distributed multi-scenario data while fully protecting the data privacy of each node. It avoids the problem of insufficient model generalization ability caused by single-scenario data, allowing intermediate model parameters to adapt to the differences in soil-rock interface characteristics in different engineering scenarios. At the same time, it improves the model's adaptability to complex interface conditions and provides comprehensive parameter support for the construction of the final prediction model.

[0030] S6. Based on the intermediate model parameters obtained in S5, construct a soil-rock anisotropic interface shear strength prediction model. By adjusting the output layer parameters of the model, the model can output the corresponding shear strength prediction results according to the input soil-rock interface parameters.

[0031] Specifically, the implementation process of step S6 is as follows: Based on the intermediate model parameters obtained in S5, a soil-rock anisotropic interface shear strength prediction model is constructed. The model adopts a three-layer fully connected structure of "feature input layer - hidden layer - output layer". The number of neurons in the feature input layer is equal to the total dimension of the parameters collected in S1 (a total of 6 dimensions, corresponding to the proportion of lithological components, lithological density, particle arrangement direction, particle distribution density, interface cohesion magnitude, and interface cohesion area), and receives the standardized lithology, texture, and interface cohesion parameters; the hidden layer is set with 2 layers, with 128 neurons in the first layer and 64 neurons in the second layer, both using the weight matrix (dimensions of 6×128 and 128×64 respectively) and bias vector (dimensions of 128 and 64 respectively) from the intermediate model parameters, and the activation function is ReLU; the output layer is set with 1 neuron to output the shear strength prediction result, and the weight matrix (dimension of 64×1) and bias vector (dimension of 64×1) of the output layer are... The bias (range 0-5) needs to be adjusted separately. During the adjustment process, the mean square error between the model's predicted value and the measured value is used as the objective function (the error is calculated using the squared average of the differences between the predicted and measured values ​​of all training samples). The output layer parameters are optimized using the batch gradient descent method, with a learning rate of 0.001-0.01 and 30-60 iterations. The objective function value is calculated after each iteration. Optimization is stopped when the objective function value drops below 0.1 or the change in the objective function value is less than 0.001 after 5 consecutive iterations. After the parameters are adjusted, the model is validated. The validation set uses 200-500 sets of soil-rock interface data that were not used in the training (data from engineering areas different from the training scenario). During validation, the mean absolute error between the model's predicted value and the measured value is calculated to ensure that the error is controlled within 5%, so that the model can stably output the corresponding shear strength prediction results based on the input lithology, texture, and interface bonding parameters. This step, through targeted model structure design and output layer parameter optimization, balances the universality of intermediate model parameters with the accuracy of final prediction, enabling the constructed prediction model to be directly applied to actual engineering scenarios without additional adjustments to the core structure. This achieves efficient and accurate prediction of the shear strength of soil-rock anisotropic interfaces, meeting the strength parameter requirements of engineering design and safety assessment.

[0032] Preferably, the feature extraction process of the lithology-constrained Transformer network satisfies the following expression: ,in, This represents the lithology-texture-cohesion correlation feature vector. This represents the original feature matrix composed of the input lithological parameters, fabric parameters, and interface bonding parameters. This represents the multi-head attention calculation function. Presentation layer normalization operation, Represents the self-attention computation function. This represents the computation function of the feedforward neural network.

[0033] Specifically, during feature extraction using the lithology-constrained Transformer network, the lithological parameters, fabric parameters, and interface bonding parameters obtained in step S1 are first used to form an original feature matrix. The number of rows in this matrix corresponds to the number of samples (32-64 samples per processing run), and the number of columns corresponds to the total number of parameter dimensions (6 dimensions in total). During network operation, self-attention is first calculated on the original feature matrix. Attention weights between different parameter dimensions are calculated using a scaled dot product method, with weight values ​​controlled between 0 and 1 to ensure that key parameter correlations are highlighted. After the self-attention calculation, the result is added to the original feature matrix and layer normalization is performed. In the layer normalization operation, the mean and standard deviation are used to standardize the data. The mean is calculated as the parameter mean of all samples in the current layer, and a small value of 0.0001 is added when calculating the standard deviation to avoid the denominator being zero. The processed results are then input into a feedforward neural network. This network's hidden layers combine linear transformation with ReLU activation. The weight matrix dimensions of the linear transformation are set to 6×512 and 512×6, and the activation function threshold is 0, with values ​​less than 0 being zeroed out. The feedforward neural network output is then added to the layer-normalized self-attention result and normalized again, ultimately yielding a lithology-texture-cohesion correlation feature vector. The vector dimension is set to 256-512 dimensions based on network complexity. This vector accurately reflects the nonlinear correlation between multiple parameters, providing high-quality feature input for subsequent steps, ensuring the comprehensiveness and accuracy of feature extraction, and avoiding information loss caused by isolated single-parameter features.

[0034] Preferably, the process by which the heterogeneous interface adhesion strength mapping neural network outputs the interface adhesion strength feature matrix satisfies the following expression: ,in, This represents the characteristic matrix of interfacial bond strength. This represents the activation function. Representing nodes in a graph neural network The set of neighboring nodes, Represents a node With nodes Adjacency matrix elements between them This represents the weight matrix of a graph neural network. Representing nodes respectively ,node The initial eigenvectors.

[0035] Specifically, during the execution of the heterogeneous interface bonding strength map neural network, the lithology-texture-bond correlation feature vector obtained in step S2 is first decomposed into 128-256 node features. Each node feature corresponds to the comprehensive parameter correlation information of a local region of the interface. An adjacency matrix is ​​constructed based on the three-dimensional geometric coordinates of the interface. The matrix dimension is consistent with the number of nodes (128×128 or 256×256). The matrix elements are determined according to the Euclidean distance between nodes. Elements corresponding to nodes with a distance less than 0.5m are set to 1, and those greater than or equal to 0m are set to 0, thus representing the correlation between nodes. During the execution of the network convolutional layer, the features of the neighboring nodes of each node are first linearly transformed by the weight matrix. The weight matrix dimension is set to 1×64 (the dimension of a single node feature is 1, and it becomes 64 dimensions after transformation). After transformation, the arithmetic mean of the neighboring node features is aggregated. The aggregation range is all neighboring nodes of each node (the number of neighboring nodes is controlled on average between 8 and 15). Simultaneously, a linear transformation is applied to the current node's own features using another weight matrix, also with a dimension of 1×64. The transformed features are then added to the aggregated neighbor features, and finally input into a ReLU activation function for non-linear transformation. The activation function maps feature values ​​less than 0 to 0, while leaving those greater than 0 unchanged. After processing through 3-5 convolutional layers, global average pooling is performed on the final features of all nodes, calculating the average of all node features to obtain a 64-128 dimension interface adhesion strength feature matrix. This matrix comprehensively reflects the distribution pattern of adhesion strength in different regions of the interface, providing accurate feature basis for subsequent parameter optimization and avoiding model bias caused by ignoring features in local areas.

[0036] Preferably, the parameter iterative optimization process of the gradient algorithm for constructing tensor-guided strategies satisfies the following expression: ,in, These represent the policy network parameters before and after optimization, respectively. Indicates the learning rate. Indicates the parameter gradient calculation, Indicating in strategy Expectation calculation below, Representing the trajectory Cumulative rewards Represents the structure tensor.

[0037] Specifically, during the execution of the configuration tensor-guided gradient algorithm, the interface adhesion strength feature matrix output in step S3 is first converted into a one-dimensional vector. The vector length is determined based on the dimension of the feature matrix (64 or 128), serving as the input state for the algorithm. A configuration tensor is then introduced, which consists of the particle alignment direction (angle range 0°-90°) and the particle distribution density (value range 10³-10⁻⁶). 5The algorithm consists of two tensors (each containing 128 or 256 neurons). Each component of the tensor is element-wise multiplied by the input state vector. Each component is multiplied by all elements of the input vector, resulting in two vectors of equal length. These two vectors are then concatenated to form an extended state vector (length 128 or 256). The input layer of the algorithm's policy network receives this extended state vector and performs a linear transformation through a first fully connected layer (256 neurons). The transformation weight matrix has a dimension of extended vector length × 256, and the activation function is ReLU, which sets values ​​less than 0 in the transformation result to zero. The output of the first layer is input to a second fully connected layer (128 neurons), with a weight matrix dimension of 256 × 128 and the activation function tanh, mapping the output values ​​to the range -1 to -1. The output of the second layer is input to the output layer (number of neurons equal to the number of parameters to be updated, set to 32-64), obtaining a probability distribution of parameter updates. The sum of the probability values ​​is 1, and the update direction with the highest probability is selected as the parameter adjustment basis for the current iteration. The mean squared error reduction of the predicted interface bonding strength under the updated direction is calculated as a reward. The reward value is adjusted by combining the weights of the structural tensor components (particle alignment direction 0.3-0.5, particle distribution density 0.4-0.7). The policy network parameters are then iteratively updated using the gradient ascent method (learning rate 0.005-0.01) for 50-100 iterations until the reward value stabilizes (the change is less than 0.001 for 5 consecutive iterations). The optimized mapping parameters are obtained, which can accurately reflect the influence of structural characteristics on bonding strength and avoid parameter optimization from getting trapped in local optima.

[0038] Preferably, the collaborative training process performed by the multimodal federated training platform satisfies the following expression: ,in, This represents the intermediate model parameters after co-training. Indicates the number of distributed nodes. Indicates the first The number of samples per node Indicates the first The model parameters are obtained by training each node locally.

[0039] Specifically, during the multimodal federated training platform's operation, it first receives configuration-bond strength mapping parameters uploaded by 5-10 distributed nodes. Each node corresponds to a different engineering scenario (slope, tunnel, etc.), and the uploaded parameters include a weight matrix (64×32 dimensions) and a bias vector (32 dimensions). The platform first verifies the node's identity (through a preset node key) and parameter format (ensuring the matrix dimension and numerical range meet the requirements, weights range from -1 to -1, and biases range from 0 to 5). After successful verification, it obtains the number of samples for each node (1000-5000 groups per node) and calculates the weight of each node (the number of node samples divided by the total number of samples across all nodes, with a value of 0.05-0.3). Weighted summation is performed on parameters of the same type for each node. For example, the weight matrix of node 1 is multiplied by its weight, and the weighted summation of the weight matrices of other nodes is added to obtain the global weight matrix. The bias vector is obtained similarly through weighted summation. Floating-point precision is used during weighted calculations, retaining 8 decimal places to avoid the accumulation of calculation errors. After obtaining the initial global parameters, the platform calculates the mean squared error between these parameters and the previous round of global parameters. If the error is less than 0.0001, the parameters are considered to have converged and are used as intermediate model parameters. If they do not converge, the initial global parameters are distributed to each node. The nodes update their local models based on these parameters and perform 10-20 rounds of local training (learning rate 0.001-0.005, batch size 32-64). The parameters are then uploaded again, and the collaborative training process is repeated for 5-8 rounds. The final intermediate model parameters can adapt to data from multiple scenarios, ensuring the model's generalization ability while avoiding privacy leaks caused by sharing the original data.

[0040] Preferably, the process by which the soil-rock anisotropic interface shear strength prediction model outputs the prediction results satisfies the following expression: ,in, This indicates the predicted shear strength. This represents the output layer weight matrix. This represents the hidden layer weight matrix. This represents the input soil-rock interface parameters. This represents the hidden layer bias vector. This represents the output layer bias vector. This represents the ReLU activation function.

[0041] Specifically, when constructing the soil-rock anisotropic interface shear strength prediction model, the network structure is first built based on the intermediate model parameters obtained in step S5. The hidden layer uses the weight matrix (dimensions 6×128, 128×64) and bias vector (dimensions 128, 64) from the intermediate parameters, and the activation function is ReLU (threshold 0). The input layer receives the standardized soil-rock interface parameters (6 dimensions: lithology, texture, and cohesion, with numerical ranges according to step S1), and performs feature transformation through the hidden layer. The first hidden layer converts the 6-dimensional input into 128-dimensional features, and the second layer converts it into 64-dimensional features. The output layer has one neuron with a weight matrix of dimension 64×1 (initial values ​​-0.5 to 0.5) and an initial bias value of 0 to 5, which needs to be optimized separately. During optimization, the mean squared error between the model's predicted and measured values ​​is used as the objective function (calculating the squared average of the differences between all training samples). Batch gradient descent is employed to update the output layer parameters. Each iteration uses 32-64 sets of samples to calculate the gradient, with a learning rate set to 0.001-0.01. Weights and biases are adjusted according to the gradient direction (decrease parameters for positive gradients and increase them for negative gradients). The iterations are performed 30-60 times. After each iteration, the objective function value is calculated. Optimization stops if the value is less than 0.1 or if the value changes less than 0.001 for five consecutive iterations. After optimization, when the model receives new interface parameters, feature transformations are performed on the input and hidden layers, ultimately outputting the shear strength prediction result from the output layer. This result accurately reflects the actual shear strength of the interface, avoiding prediction bias caused by unoptimized output layer parameters and meeting the precise strength assessment requirements of engineering projects.

[0042] Preferably, step S3 includes the following sub-steps: S31, performing node mapping processing on the lithology-facies-bond correlation feature vector, mapping each feature dimension in the correlation feature vector to a node in the heterogeneous interface bond strength graph neural network, forming an initial node feature set; S32, constructing an adjacency matrix of the graph neural network based on the geometric structure of the interface bond region, wherein the element values ​​in the adjacency matrix are determined according to the spatial distance between nodes and the correlation of bond strength; S33, inputting the initial node feature set and the adjacency matrix into the convolutional layer of the graph neural network, updating the node features through convolution operations, and capturing the local feature correlation between nodes; S34, performing global pooling operation on the node features output by the convolutional layer, integrating the feature information of all nodes, and generating an interface bond strength feature matrix.

[0043] Specifically, step S3 includes four sub-steps: In S31, during node mapping, the lithology-facies-cohesion correlation feature vector (dimensions 256-512) obtained in step S2 is split into equal-dimensional segments. Each dimension corresponds to a node in the heterogeneous interface bonding strength graph neural network. After splitting, an initial node feature set of 128-256 nodes is formed. The feature value of each node directly uses the feature vector value of the corresponding dimension, ensuring complete transmission of feature information. In S32, when constructing the adjacency matrix, based on the three-dimensional geometric coordinates of the interface bonding region (obtained through a laser scanner, with coordinate accuracy retained to three decimal places), the Euclidean distance between any two nodes is calculated. Nodes with a distance less than 0.5 meters are determined to be adjacent nodes, and those with a distance greater than or equal to 0.5 meters are determined to be non-adjacent nodes. Based on this, a 0-1 adjacency matrix is ​​generated (the matrix dimension is consistent with the number of nodes, and the positions of adjacent nodes are...). The elements are 1 (non-adjacent elements are 0), accurately representing the relationships between nodes; when performing convolution operations in S33, the initial node feature set and the adjacency matrix are input into the 3rd to 5th convolutional layers of the graph neural network. Each layer uses the mean aggregation function to take the arithmetic mean of the feature values ​​of all neighboring nodes (average number 8-15) of each node, and then uses the ReLU activation function (threshold 0) to perform a non-linear transformation on the aggregated features, mapping feature values ​​less than 0 to 0, while keeping those greater than 0 unchanged, thus realizing node feature updates and local association capture; when performing global pooling in S34, the arithmetic mean of all node features output by the convolutional layer (each node feature dimension is 64) is calculated and integrated into a 64-128 dimension interface adhesion strength feature matrix. This matrix comprehensively reflects the spatial distribution law of interface adhesion strength, providing accurate feature basis for subsequent parameter optimization and avoiding model deviation caused by the omission of local features.

[0044] Preferably, step S4 includes the following sub-steps: S41, converting the interface adhesion strength feature matrix into a feature format that can be processed by the policy gradient algorithm, and extracting the calibration feature values ​​in the feature matrix as the input state of the algorithm; S42, introducing a configuration tensor, fusing each component of the configuration tensor with the input state to form an extended state vector with configuration guidance information; S43, based on the extended state vector, generating an action probability distribution for parameter updates through the policy network, and selecting the parameter update direction according to the action probability distribution; S44, calculating the reward value corresponding to the parameter update direction, adjusting the reward weight in combination with the guidance effect of the configuration tensor, and updating the policy network parameters using the policy gradient formula.

[0045] Specifically, step S4 includes four sub-steps: In S41, during feature format conversion, the interface adhesion strength feature matrix (64-128 dimensions) output from step S3 is expanded into a one-dimensional vector in row-major order. The vector length is consistent with the total number of elements in the feature matrix (64 or 128), converting it into an input state format that can be processed by the configuration tensor-guided gradient algorithm, ensuring data compatibility with algorithm requirements; In S42, during state vector fusion, a configuration tensor (consisting of particle alignment direction 0°-90° and particle distribution density 10³-10⁻⁶) is introduced. 5 Composed of units per cubic meter), the two components of the tensor are multiplied element-wise with the input state vector. Each component is multiplied with all elements of the input vector to obtain two vectors of equal length. These two vectors are then concatenated to form an extended state vector of length 128 or 256, which is incorporated into the configuration guidance information. When generating action probabilities in S43, the extended state vector is input into the three fully connected layers of the policy network. The number of neurons in the input layer is consistent with the vector length. The first hidden layer (256 neurons) uses the ReLU activation function (threshold 0), the second hidden layer (128 neurons) uses the tanh activation function (output range -1-1), and the number of neurons in the output layer is 32-64 (the same as the number of parameters to be updated). The probability distribution of parameter update amounts (total probability 1) is generated, and the update direction with the highest probability is selected as the basis for parameter adjustment. When updating parameters in S44, the mean square error reduction value of the interface bonding strength prediction under the update direction is calculated as a reward. The reward value is adjusted in combination with the weights of the component of the structure tensor (particle arrangement direction 0.3-0.5, particle distribution density 0.4-0.7). The strategy network parameters are iteratively updated through the gradient ascent method (learning rate 0.005-0.01) for 50-100 iterations until the reward value changes less than 0.001 for 5 consecutive times. The optimized structure-bonding strength mapping parameters are obtained, which accurately reflect the influence of structure characteristics on bonding strength and avoid parameter optimization from getting trapped in local optima.

[0046] Preferably, step S5 includes the following sub-steps: S51, the multimodal federated training platform receives local model parameter update requests sent by each distributed node and verifies the node identity and model parameter format in the request; S52, after verification, the platform obtains the sample quantity information of each node and calculates the weight coefficient of the model parameters of each node based on the sample quantity; S53, the model parameters of each node are weighted and summed according to the weight coefficient to obtain the preliminary global model parameters; S54, the preliminary global model parameters are checked for consistency. If the parameters meet the preset consistency threshold, they are used as intermediate model parameters after collaborative training; otherwise, the weighted calculation is performed again.

[0047] Specifically, step S5 includes four sub-steps: S51 When performing request verification, the multimodal federated training platform receives local model parameter update requests sent by each distributed node (5-10, corresponding to different engineering scenarios), verifies the node identity through the preset node key (key length 256 bits, verification pass rate must reach 100%), and checks the model parameter format to ensure that the weight matrix dimension (64×32), bias vector dimension (32), and numerical range (weight -1-1, bias 0-5) meet the requirements, and rejects parameter requests with incorrect format; S52 When calculating weight coefficients, after verification, the platform obtains the number of samples for each node (1000-5000 groups for each node), calculates the weight coefficient of each node according to the formula "number of node samples / total number of samples for all nodes", and the coefficient value ranges from 0.05 to 0.3 to ensure that nodes with a large number of samples contribute more to the global parameters; S53 During the weighted summation of parameters, the weight matrix and bias vector of each node are multiplied by their corresponding weight coefficients, and then the weighted results of all nodes are summed element by element to obtain the preliminary global model parameters. The calculation process uses floating-point precision and retains 8 decimal places to avoid error accumulation. When S54 performs consistency verification, the middleware calculates the mean square error between the preliminary global parameters and the previous round of global parameters. If the error is less than 0.0001, the parameters are considered to have converged and are used as the intermediate model parameters after collaborative training. If they do not converge, the preliminary global parameters are distributed to each node. The nodes update their local models based on these parameters and perform 10-20 rounds of local training (learning rate 0.001-0.005, batch size 32-64) before uploading the parameters again. The collaborative training process is repeated for 5-8 rounds. The final intermediate model parameters can adapt to data from multiple scenarios, ensuring the model's generalization ability while avoiding privacy leaks caused by sharing the original data.

[0048] The lithology-constrained Transformer network is the core network structure in this invention for extracting multi-parameter correlation features of the soil-rock interface. Essentially, it adds a lithology parameter constraint mechanism to the traditional Transformer network, specifically adapting to the feature fusion of multi-dimensional parameters such as lithology, texture, and interface bonding. In implementation, lithology parameters such as lithological composition ratio and lithological density obtained in step S1, texture parameters such as particle orientation and particle density, and interface bonding parameters such as interface bonding force and interface bonding area are first integrated into an original feature matrix. Then, a multi-head attention mechanism consisting of 8-12 attention heads is used to progressively mine the correlation between different parameters, with each attention head focusing on capturing a set of parameter correlations. The network has 6-10 encoder layers, each undergoing self-attention calculation, feedforward neural network nonlinear transformation, and layer normalization processing, ultimately outputting a 256-512 dimensional lithology-texture-bonding correlation feature vector. The role of this network is to overcome the limitations of single-parameter feature extraction, strengthen the constraint effect of lithology parameters on other parameters, and ensure that the extracted features accurately reflect the nonlinear intrinsic relationships between multiple parameters. Its significance lies in providing high-quality feature input for subsequent learning of interface adhesion strength features, avoiding model training bias caused by isolated features or missing correlations, and laying the foundation for improving the overall prediction accuracy of the model.

[0049] The heterogeneous interface bond strength graph neural network is a network structure used to learn the spatial distribution characteristics of bond strength at soil-rock interfaces. It is specifically designed for the association of local features and the integration of global features at heterogeneous interfaces. In implementation, the associated feature vector output by the lithology-constrained Transformer network is first split into 128-256 node features with equal dimensions, each node corresponding to a local region of the interface. Then, a 0-1 adjacency matrix is ​​constructed based on the three-dimensional geometric coordinates of the interface, classifying nodes less than 0.5 meters apart as adjacent nodes to represent the association between nodes. Subsequently, 3-5 convolutional layers are used, employing a mean aggregation function to average the features of adjacent nodes for each node, combined with a ReLU activation function for nonlinear transformation, updating the node features. Finally, a global average pooling operation is performed to integrate all node features into a 64-128 dimensional interface bond strength feature matrix. The network's function is to accurately capture the local characteristics and spatial distribution patterns of interface bond strength, solving the problem that traditional networks struggle to adapt to interface heterogeneity. Its significance lies in providing characteristic basis that can reflect the overall change law of interface bonding strength for subsequent parameter optimization, avoiding the model's insufficient adaptation to interface characteristics due to the omission of local features, and ensuring the accuracy of subsequent parameter optimization and model construction.

[0050] The structure tensor-guided policy gradient algorithm is the core algorithm in this invention for optimizing the structure-bond strength mapping parameters. It introduces a structure tensor guiding factor based on the traditional policy gradient algorithm, achieving precise matching between parameter optimization and structure characteristics. In implementation, the feature matrix output by the heterogeneous interface bond strength map neural network is first converted into a one-dimensional input state. Then, a structure tensor composed of particle arrangement direction and particle distribution density is introduced, and the tensor components are fused with the input state vector to form an extended state vector. The algorithm's policy network uses three fully connected layers, and the parameter update direction is determined through activation function processing and probability distribution generation. The reduction in prediction error under the update direction is calculated as a reward, and the reward is adjusted in conjunction with the weights of the structure tensor components. The network parameters are updated iteratively 50-100 times using the gradient ascent method until the reward value stabilizes, resulting in the optimized mapping parameters. The purpose of this algorithm is to ensure that the mapping parameters accurately reflect the influence of structure characteristics on interface bond strength, avoiding parameter optimization from getting trapped in local optima. Its significance lies in providing optimized core parameters for multimodal federated training, ensuring that the subsequently trained model can fully adapt to the differences in structural characteristics, and improving the model's predictive adaptability to the interface shear strength under different structural conditions.

[0051] The multimodal federated training platform is a core platform for achieving distributed collaborative training and data privacy protection, specifically designed for integrating soil-rock interface data from multiple engineering scenarios. In implementation, the platform is deployed on a cloud server, connecting 5-10 distributed nodes corresponding to different engineering scenarios. Each node uploads parameters optimized using a tensor-guided gradient algorithm. The platform first verifies the node's identity and parameter format using a node key. Then, based on the 1000-5000 samples per node, it calculates node weight coefficients (ranging from 0.05 to 0.3). A federated averaging algorithm is used to weight and sum the parameters of each node to obtain preliminary global parameters. After consistency verification (convergence is determined by a mean squared error less than 0.0001), the converged global parameters are used as intermediate model parameters. If convergence fails, parameters are distributed to the nodes, and the "local training - parameter upload - global aggregation" process is repeated 5-8 times. The platform's role is to achieve collaborative utilization of multi-source distributed data without sharing the original data, balancing data privacy protection and improved model generalization capabilities. Its significance lies in solving the problems of low utilization rate and high risk of privacy leakage in traditional centralized training, enabling the model to adapt to the interface characteristics of different engineering scenarios, and providing key support for the final construction of a highly generalizable shear strength prediction model.

[0052] like Figure 2As shown, a method for constructing a predictive model for the shear strength of a soil-rock heterogeneous interface is presented. This method is implemented through different units, including: a soil-rock interface multi-parameter acquisition unit, which is connected to lithological parameter detection equipment, fabric parameter measurement equipment, and interface bond parameter testing equipment, and is used to receive and integrate the lithological parameters, fabric parameters, and interface bond parameters acquired by each device; a lithological constraint feature extraction unit, which is connected to the soil-rock interface multi-parameter acquisition unit and has a built-in lithological constraint Transformer network, used to extract correlation features from the acquired parameters; and an interface bond strength map neural computation unit, which is connected to the lithological constraint feature extraction unit, and uses a heterogeneous interface... The interface bond strength map neural network processes the associated features and outputs an interface bond strength feature matrix. A configuration-guided parameter optimization unit, connected to the interface bond strength map neural computation unit, optimizes the parameters corresponding to the feature matrix using a configuration tensor-guided gradient algorithm. A federated collaborative training unit, connected to the configuration-guided parameter optimization unit, constructs a multimodal federated training platform for collaborative updating of model parameters across distributed nodes. A shear strength prediction model construction unit, connected to the federated collaborative training unit, adjusts the output layer parameters based on the collaboratively updated model parameters to construct a model capable of outputting prediction results for the shear strength of soil-rock anisotropic interfaces.

[0053] A method for constructing a predictive model for the shear strength of soil-rock anisotropic interfaces is proposed. This method employs a step-by-step integrated architecture, sequentially involving lithology-constrained networks, neural computation of interface bond strength maps, fabric-guided parameter optimization, and multimodal federated training. This approach achieves comprehensive mining and efficient utilization of multi-dimensional parameters of the soil-rock interface, accurately capturing the complex relationships between lithology, fabric, and interface bond parameters, avoiding the limitations of single technical steps. During parameter processing and model training, targeted constraint mechanisms and optimization strategies are incorporated. Iterative optimization of key parameters enhances the model's adaptability to interface characteristics. Simultaneously, the federated training mode balances data utilization and privacy protection, enabling the model to adapt to different engineering scenarios without relying on extensive field experimental data, significantly improving its practicality and applicability.

[0054] This method addresses the issues of inaccurate feature extraction from single-network structures and the inability of existing models to reflect key influencing factors of shear strength. By leveraging the synergistic effect of lithology-constrained networks and neural computation of interface bond strength maps, coupled with a feature fusion strategy, it fully explores the nonlinear correlations between multiple parameters, thereby improving the accuracy of feature extraction. Furthermore, to address the problems of low data utilization and insufficient generalization ability in existing models, this method utilizes a multimodal federated training platform to achieve collaborative training of distributed node data without sharing the original data. This avoids the limitations of centralized training while meeting data privacy protection requirements. Simultaneously, it incorporates a configuration-guided parameter optimization algorithm to further improve the model's efficiency in utilizing multi-source data. The final prediction model balances cost control, prediction accuracy, and scenario adaptability, overcoming the shortcomings of traditional methods and existing models.

[0055] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent 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 constructing a predictive model for the shear strength of a soil-rock anisotropic interface, characterized in that, Includes the following steps: S1. Obtain the lithological parameters, fabric parameters and interface bonding parameters of the soil-rock anisotropic interface. The lithological parameters include the proportion of lithological components and the degree of lithological compaction. The fabric parameters include the particle arrangement direction and particle distribution density. The interface bonding parameters include the magnitude of the interface bonding force and the area of ​​the interface bonding region. S2. Input the parameters obtained in S1 into the lithology-constrained Transformer network, and use the multi-head attention mechanism of the network to extract the features of the relationship between lithology parameters, fabric parameters, and interface bonding parameters to obtain the lithology-fabrication-bond correlation feature vector. S3. Input the associated feature vector obtained in S2 into the heterogeneous interface bonding strength graph neural network. Based on the node embedding method of the graph neural network, learn the bonding strength features of each node in the interface bonding region and output the interface bonding strength feature matrix. S4. Input the feature matrix output from S3 into the configuration tensor to guide the policy gradient algorithm. Using the configuration tensor as the guiding factor, iteratively optimize the policy network parameters of the algorithm through the policy gradient update rule to obtain the optimized configuration-bond strength mapping parameters. S5. Input the mapping parameters obtained in S4 into the multimodal federated training platform. This platform performs federated training on the soil-rock interface data of multiple distributed nodes, and performs collaborative updates of model parameters without sharing the original data, generating intermediate model parameters after collaborative training. S6. Based on the intermediate model parameters obtained in S5, construct a soil-rock anisotropic interface shear strength prediction model. By adjusting the output layer parameters of the model, the model can output the corresponding shear strength prediction results according to the input soil-rock interface parameters.

2. The method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface according to claim 1, characterized in that, The feature extraction process of the lithology-constrained Transformer network satisfies the following expression: ,in, This represents the lithology-texture-cohesion correlation feature vector. This represents the original feature matrix composed of the input lithological parameters, fabric parameters, and interface bonding parameters. This represents the multi-head attention calculation function. Presentation layer normalization operation, Represents the self-attention computation function. This represents the computation function of the feedforward neural network.

3. The method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface according to claim 1, characterized in that, The process by which the heterogeneous interface adhesion strength mapping neural network outputs the interface adhesion strength feature matrix satisfies the following expression: ,in, This represents the characteristic matrix of interfacial bond strength. This represents the activation function. Representing nodes in a graph neural network The set of neighboring nodes, Represents a node With nodes Adjacency matrix elements between them This represents the weight matrix of a graph neural network. Representing nodes respectively ,node The initial eigenvectors.

4. The method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface according to claim 1, characterized in that, The parameter iterative optimization process of the gradient algorithm for constructing tensor-guided strategies satisfies the following expression: ,in, These represent the policy network parameters before and after optimization, respectively. Indicates the learning rate. Indicates the parameter gradient calculation, Indicating in strategy Expectation calculation below, Representing the trajectory Cumulative rewards Represents the structure tensor.

5. The method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface according to claim 1, characterized in that, The collaborative training process performed by the multimodal federated training platform satisfies the following expression: ,in, This represents the intermediate model parameters after co-training. Indicates the number of distributed nodes. Indicates the first The number of samples per node Indicates the first The model parameters are obtained by training each node locally.

6. The method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface according to claim 1, characterized in that, The process by which the soil-rock anisotropic interface shear strength prediction model outputs prediction results satisfies the following expression: ,in, This indicates the predicted shear strength. This represents the output layer weight matrix. This represents the hidden layer weight matrix. This represents the input soil-rock interface parameters. This represents the hidden layer bias vector. This represents the output layer bias vector. This represents the ReLU activation function.

7. The method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface according to claim 1, characterized in that, S3 includes the following sub-steps: S31, performing node mapping processing on the lithology-facies-bond correlation feature vector, mapping each feature dimension in the correlation feature vector to a node in the heterogeneous interface bond strength graph neural network, forming an initial node feature set; S32, constructing an adjacency matrix of the graph neural network based on the geometric structure of the interface bond region, where the element values ​​are determined according to the spatial distance between nodes and the correlation of bond strength; S33, inputting the initial node feature set and the adjacency matrix into the convolutional layer of the graph neural network, updating the node features through convolution operations, and capturing the local feature correlation between nodes; S34, performing global pooling on the node features output by the convolutional layer, integrating the feature information of all nodes, and generating an interface bond strength feature matrix.

8. The method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface according to claim 1, characterized in that, S4 includes the following sub-steps: S41, converting the interface adhesion strength feature matrix into a feature format that can be processed by the policy gradient algorithm, and extracting the calibration feature values ​​in the feature matrix as the input state of the algorithm; S42, introducing a configuration tensor, fusing each component of the configuration tensor with the input state to form an extended state vector with configuration guidance information; S43, based on the extended state vector, generating an action probability distribution for parameter updates through the policy network, and selecting the parameter update direction according to the action probability distribution; S44, calculating the reward value corresponding to the parameter update direction, adjusting the reward weight in combination with the guidance effect of the configuration tensor, and updating the policy network parameters using the policy gradient formula.

9. The method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface according to claim 1, characterized in that, S5 includes the following steps: S51, the multimodal federated training platform receives local model parameter update requests sent by each distributed node and verifies the node identity and model parameter format in the request; S52, after verification, the platform obtains the sample quantity information of each node and calculates the weight coefficient of the model parameters of each node based on the sample quantity; S53, the model parameters of each node are weighted and summed according to the weight coefficient to obtain the preliminary global model parameters; S54, the preliminary global model parameters are checked for consistency. If the parameters meet the preset consistency threshold, they are used as intermediate model parameters after collaborative training; otherwise, the weighted calculation is performed again.

10. A method for constructing a predictive model for the shear strength of a soil-rock anisotropic interface according to any one of claims 1-9, characterized in that, This method is implemented through different units, including: The soil-rock interface multi-parameter acquisition unit is connected to lithological parameter detection equipment, fabrication parameter measurement equipment and interface bonding parameter testing equipment. It is used to receive and integrate the lithological parameters, fabrication parameters and interface bonding parameters collected by each device. The lithological constraint feature extraction unit is connected to the soil-rock interface multi-parameter acquisition unit and has a built-in lithological constraint Transformer network for performing correlation feature extraction on the acquired parameters. The interface bond strength map neural computation unit is connected to the lithological constraint feature extraction unit. It processes the associated features through the heterogeneous interface bond strength map neural network and outputs the interface bond strength feature matrix. The configuration guidance parameter optimization unit is connected to the interface adhesion strength map neural computing unit, and the configuration tensor guidance strategy gradient algorithm is used to optimize the parameters corresponding to the feature matrix. The federated collaborative training unit is connected to the configuration guidance parameter optimization unit to build a multimodal federated training platform for collaborative updating of model parameters across distributed nodes. The shear strength prediction model building unit is connected to the federated collaborative training unit. Based on the collaboratively updated model parameters, the output layer parameters are adjusted to build a model that can output the prediction results of the shear strength of the soil-rock anisotropic interface.

Images