Civil engineering rock mass stability evaluation method and system

Abstract

The invention relates to the technical field of civil engineering rock mass stability evaluation, and discloses a civil engineering rock mass stability evaluation method and system. The method comprises the following steps: acquiring multi-modal geological exploration data and preprocessing the multi-modal geological exploration data into a standardized feature tensor; data driving feature extraction and physical attribute parameterization are respectively carried out through a double-path parallel deep neural network; generating a predicted value based on a physical mechanism in combination with a rock physical constitutive model; training a model by using a composite loss function comprising prediction loss and physical consistency loss; and finally generating a causal interpretation report for quantifying the contribution degree of each physical parameter. The system comprises a multi-modal data preprocessing module, a dual-path neural network model, a prediction and deduction module, a training optimization module and a causal interpretation generation module. According to the invention, high-precision, physically interpretable and engineering-credible rock mass stability intelligent evaluation is realized.

Description

Technical Field

[0001] This invention belongs to the field of civil engineering, specifically relating to methods and systems for assessing the stability of rock masses in civil engineering. Background Technology

[0002] In civil engineering, rock mass stability assessment is a core component in ensuring the safety of tunnels, slopes, and underground engineering projects. It relies on the accurate acquisition of rock mass mechanical parameters (such as elastic modulus, internal friction angle, and cohesion) and an understanding of their physical mechanisms. Traditional methods are mainly based on on-site borehole sampling, laboratory tests, or empirical formulas. While these methods have a certain physical basis, they suffer from problems such as high cost, limited coverage, and difficulty in reflecting the spatial heterogeneity of rock masses.

[0003] In recent years, deep learning technology has been introduced into the field of rock mass parameter inversion, achieving significant progress in efficiency and breadth by integrating multi-source data such as ground-penetrating radar, seismic wave velocity, and remote sensing imagery to realize end-to-end predictions. However, such models typically operate as "black boxes," only outputting numerical results without explaining the basis for the predictions. For example, they cannot explain "why the predicted value of the elastic modulus in a certain area is low," making it difficult for engineers to determine whether the results conform to the laws of rock mechanics. This severely weakens the credibility and adoptability of the model in critical engineering decisions.

[0004] Rock mass stability analysis heavily relies on classical mechanics theories such as the Mohr-Coulomb failure criterion, which clearly states that rock mass instability is jointly controlled by physical factors such as stress state, pore water pressure, and structural properties. An ideal AI-assisted system should not only predict parameters but also reveal the contribution mechanism of each input feature to the output result and map it to an understandable physical semantic level.

[0005] Existing deep learning methods generally lack the ability to couple with the basic equations of rock mechanics. They neither embed the constraints of differential equations into the model training process nor establish an attribution path between the prediction results and the physical failure mechanism, causing the model output to be detached from the engineering cognitive system.

[0006] Existing technologies for intelligent inversion of rock mass parameters generally suffer from three major drawbacks: First, model interpretation tools (such as LIME and Grad-CAM) only provide pixel-level or feature-level importance, and cannot be correlated with specific physical quantities in rock mechanics (such as pore water pressure and joint density). Secondly, the prediction results are not linked to stability criteria such as the Mohr-Coulomb criterion, making it difficult to determine whether parameter anomalies truly constitute a risk of instability. Third, the lack of a closed-loop system architecture encompassing "prediction-attribution-verification" makes it impossible to generate sensitivity heat maps and key factor annotations with engineering guidance significance. These problems are particularly prominent in stability assessments under complex geological conditions (such as high-water-pressure weak interlayers and fault fracture zones), easily leading to misjudgments or overly conservative designs. There is an urgent need to construct an assessment method that deeply integrates interpretable artificial intelligence and rock mechanics principles to achieve highly reliable, traceable, and verifiable intelligent diagnosis of rock mass stability. Summary of the Invention

[0007] To address the technical problem of low reliability in engineering decisions due to the lack of physical interpretability in the prediction of rock mechanics parameters by existing deep learning models, this invention provides a method and system for assessing rock mass stability in civil engineering. This invention aims to solve the aforementioned problem by constructing a deep neural network architecture that integrates physical mechanisms and data-driven approaches. The architecture includes a data-driven path for extracting high-dimensional nonlinear features from multimodal geological exploration data, and a physical parameterization path for generating rock mass parameters with clear physical meaning.

[0008] This invention introduces a physical consistency constraint module. During network training, this module utilizes prior physical constitutive relations or empirical formulas from the field of rock mechanics as regularization terms to force the prediction results of the data-driven path to be consistent with the results calculated based on physical laws by the physical parameterization path. Ultimately, this invention not only outputs high-precision predicted values ​​of rock mechanics parameters but also generates a traceable causal explanation chain. This explanation chain explicitly quantifies the contribution of intermediate physical parameters such as porosity, fracture density, and water saturation to the final prediction results, thereby transforming the "black box" decision-making process of the deep learning model into a physically interpretable and engineering-reliable "grey box" analysis process.

[0009] This invention provides a method for assessing the stability of rock masses in civil engineering, which includes the following steps: Obtain multimodal geological exploration raw data for the rock mass area to be evaluated, including seismic wave exploration data, borehole core image data, and geological logging data; The multimodal geological exploration raw data is preprocessed, including time-frequency domain transformation of the seismic wave exploration data to obtain spectral features, edge detection and morphological analysis of the borehole core image data to extract the topological features of rock mass structural surfaces and fracture networks, and normalization of the geological logging data, ultimately forming a standardized multi-channel input feature tensor. The standardized multi-channel input feature tensor is input into a dual-path parallel deep neural network model, which includes a data-driven feature extraction path and a physical property parameterization path. Through the data-driven feature extraction path, multi-scale deep feature extraction is performed on the standardized multi-channel input feature tensor to generate a high-dimensional data-driven feature vector, which characterizes the potential nonlinear correlation of rock mass mechanical behavior. The standardized multi-channel input feature tensor is mapped through the physical property parameterization path to generate a set of core physical property parameters of rock mass with clear physical meaning. The core physical property parameters of rock mass include rock porosity, fracture network density, water saturation, mineral composition vector, and uniaxial compressive strength of rock. The high-dimensional data-driven feature vector is input into the first prediction head network to generate the first predicted value of the data-driven rock mechanics parameters. The core physical property parameters of the rock mass medium are substituted into a preset, differentiable rock physical constitutive model for forward calculation to generate a second predicted value of the rock mass mechanical parameters based on physical mechanism. Construct a composite loss function that includes a prediction loss term and a physical consistency loss term, wherein the prediction loss term is used to measure the difference between the first predicted value and the actual rock mass mechanical parameter value of the rock mass sample, and the physical consistency loss term is used to measure the difference between the first predicted value and the second predicted value; The composite loss function is used to perform end-to-end backpropagation training on the dual-path parallel deep neural network model, and the network weight parameters are iteratively optimized through gradient descent algorithm until the composite loss function converges to below a preset threshold. After the model training converges, a structured causal explanation report is generated using the core physical property parameters of the rock mass output by the physical property parameterization path and the first predicted value. The causal explanation report quantitatively reveals the specific contribution of each core physical property parameter to the final predicted value of the rock mass mechanical parameters.

[0010] In one embodiment of the present invention, the data-driven feature extraction path specifically consists of three parallel sub-networks targeting different data modalities: a temporal convolutional network for processing seismic wave exploration data after time-frequency domain transformation, a graph convolutional network for processing the topological features of rock mass structural surfaces and fracture networks, and a multilayer perceptron for processing normalized geological logging data. The output feature vectors of the three parallel sub-networks are weighted and concatenated through an attention fusion layer to form the high-dimensional data-driven feature vector. The temporal convolutional network employs a residual connection structure with multi-scale dilated convolutions to capture the propagation attenuation characteristics of seismic wave signals at different time scales. The graph convolutional network models the extracted fracture network topology as a graph data structure, where nodes represent fracture intersections and edges represent fracture segments. It aggregates neighborhood information through multilayer graph convolution operations to characterize the connectivity and spatial distribution complexity of the fracture network.

[0011] In one embodiment of the present invention, the physical property parameterization path is a fully connected neural network with multiple independent output heads, which correspond to rock porosity, fracture network density, water saturation, mineral composition vector, and uniaxial compressive strength of rock, respectively. Each output head employs a constrained activation function, which ensures that the output physical property parameter values ​​are within their physical valid range. Specifically, the output activation functions for rock porosity and water saturation are hard Z-shaped functions, ensuring that the output values ​​are within the range of 0 to 1, and the output activation function for fracture network density is a modified linear unit, ensuring that the output values ​​are non-negative.

[0012] In one embodiment of the present invention, the preset, differentiable rock physics constitutive model is a combination of the Gassmann equation and the Kuster-Toksöz model based on the equivalent medium theory. Specifically, firstly, the equivalent bulk modulus and shear modulus of the rock skeleton are calculated using the mineral composition vector through the Voigt-Reuss-Hill average model. Next, the equivalent elastic modulus of the dry rock, including pores and fractures, is calculated using the rock porosity and fracture network density parameters through the Kuster-Toksöz model. Finally, the water saturation parameter is substituted into the Gassmann equation to complete the fluid substitution calculation, ultimately obtaining the second predicted value of the rock mechanics parameters based on the physical mechanism. The entire calculation process is implemented through tensor operations to ensure end-to-end differentiability within the neural network framework.

[0013] In one embodiment of the present invention, the specific mathematical expression of the composite loss function is as follows: the total loss equals the first weighting coefficient multiplied by the prediction loss term, plus the second weighting coefficient multiplied by the physical consistency loss term. The prediction loss term is specifically the root mean square error between the first predicted value and the actual rock mass mechanical parameter values ​​of the rock mass sample. The physical consistency loss term is specifically the smoothing L1 loss between the first predicted value and the second predicted value. The first weighting coefficient and the second weighting coefficient are hyperparameters, and their numerical ratio determines the trade-off between data fitting accuracy and adherence to physical laws during model training.

[0014] As one embodiment of the present invention, the specific steps for generating a structured causal explanation report include: calculating the partial derivative of the first predicted value of the data-driven rock mechanics parameter with respect to each core physical property parameter of the rock mass medium, i.e., the gradient value; normalizing the absolute value of the gradient value to obtain the sensitivity coefficient of each core physical property parameter to the prediction result; storing and outputting the structured causal explanation report in the form of key-value pairs, where the key is the name of the core physical property parameter and the value is the specific predicted value of the parameter and its corresponding sensitivity coefficient, thereby providing quantitative decision-making basis for engineering technicians.

[0015] This invention provides a rock mass stability assessment system for civil engineering projects, comprising: The multimodal data acquisition and preprocessing module is used to acquire seismic wave exploration data, borehole core image data and geological logging data of the rock mass area to be evaluated, and to perform time-frequency domain transformation, fracture network topology extraction and data normalization operations to generate standardized multi-channel input feature tensors. The dual-path parallel deep neural network model is configured to receive the standardized multi-channel input feature tensor and perform parallel processing; the dual-path parallel deep neural network model includes a data-driven feature extraction path for generating high-dimensional data-driven feature vectors characterizing the potential nonlinear correlation of rock mass mechanical behavior, and a physical property parameterization path for generating a set of core physical property parameters of rock mass medium with clear physical meaning. The data-driven prediction module contains a first prediction head network, which receives the high-dimensional data-driven feature vector and outputs the first predicted value of the data-driven rock mechanics parameters. The physical mechanism deduction module has a differentiable rock physical constitutive model embedded in it. It is used to receive the core physical property parameters of the rock mass medium and output the second predicted value of the rock mass mechanical parameters based on the physical mechanism. The model training and optimization module internally defines a composite loss function that includes a prediction loss term and a physical consistency loss term. During the training phase, it is used to iteratively optimize the network weights of the dual-path parallel deep neural network model based on the first predicted value, the second predicted value, and the true value of the rock mass sample through the backpropagation algorithm. The causal explanation generation module is used to quantify the contribution of each physical property parameter to the final prediction result by calculating the gradient of the first predicted value relative to the core physical property parameters of the rock mass after the model has been trained, and to generate a structured causal explanation report.

[0016] In one embodiment of the present invention, the multimodal data acquisition and preprocessing module includes a seismic wave data processing unit, which uses a short-time Fourier transform algorithm to convert the original time-domain seismic wave signal into a time-spectrum map. The multimodal data acquisition and preprocessing module also includes a core image analysis unit, which integrates a Hough transform algorithm for detecting linear fractures and a region growing algorithm for segmenting irregular structural surfaces, ultimately generating a fracture network adjacency matrix characterizing the location, direction, length, and aperture of the fractures.

[0017] In one embodiment of the present invention, the dual-path parallel deep neural network model is deployed on a computing device integrated with a tensor processing unit. The data-driven feature extraction path and the physical property parameterization path share an initial embedding layer network, which is used to map input features from different modalities to a unified high-dimensional semantic space, and then branch them to their respective dedicated processing networks to promote early interaction of cross-modal information.

[0018] As one embodiment of the present invention, the causal explanation generation module further includes a visualization interface generation unit. The visualization interface generation unit renders the structured causal explanation report into an interactive dashboard. This dashboard displays the sensitivity coefficients of each core physical property parameter in the form of a bar chart and provides a drill-down analysis function, allowing users to select a specific core physical property parameter and trace back to the specific area or feature in the original geological exploration data that caused the parameter's anomaly. This achieves complete visual tracing from the final evaluation result to the original exploration evidence.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention designs a neural network architecture that uses both data-driven and physical mechanism parallel paths, and introduces physical consistency loss for constraint, ensuring that the model prediction results are not only highly accurate at the data level, but also compliant at the physical mechanism level.

[0020] 2. This invention overcomes the fatal flaw of traditional deep learning models operating as "black boxes," whose predictions lack physical basis. The invention's unique causal explanation generation module can quantitatively reveal the influence weights of internal physical parameters on the final rock mass stability assessment results, providing engineering decision-makers with clear, transparent, and auditable judgment criteria. This significantly enhances the model's practicality and credibility in high-risk fields such as civil engineering safety assessments.

[0021] 3. This invention integrates the rock physics constitutive model as a differentiable component of the neural network, realizing a deep coupling between prior physical knowledge and data-driven learning. Compared with a simple data-driven model, this invention reduces the dependence on training data and exhibits stronger robustness and generalization ability in scenarios with small samples or poor data quality. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall technical solution architecture of the civil engineering rock mass stability assessment method and system proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the deep neural network that integrates physical mechanisms and data-driven dual paths in this invention; Figure 3 This is a logical flowchart of the multimodal geological exploration data preprocessing and feature construction in this invention; Figure 4 This is a logical framework diagram of the parallel connection and attention fusion of multimodal sub-networks in the data-driven feature extraction path in this invention; Figure 5 This is a schematic diagram of the joint deduction framework of the physical property parameterization path and the differentiable rock physical constitutive model in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal and the causal explanation generation module in this invention. Detailed Implementation

[0023] Please refer to the attached document. Figures 1 to 6 This invention provides a method and system for assessing the stability of rock masses in civil engineering, aiming to address the lack of physical interpretability in the prediction of rock mechanics parameters using deep learning models. While traditional data-driven models can achieve high-precision fitting, their internal decision-making logic cannot establish a clear mapping relationship with the physical laws of rock mechanics, making it difficult for engineers to trust their output results. To overcome this deficiency, this invention constructs a deep neural network architecture that integrates physical mechanisms and data-driven approaches. While ensuring prediction accuracy, it forces the model output to conform to the constitutive relationship of rock physics and generates a traceable and quantifiable causal explanation chain.

[0024] The method for assessing the stability of rock masses in civil engineering includes the following steps: Acquire multimodal geological exploration raw data of the rock mass area to be evaluated; perform data preprocessing on the multimodal geological exploration raw data to form a standardized multi-channel input feature tensor; The standardized multi-channel input feature tensor is input into a dual-path parallel deep neural network model; a high-dimensional data-driven feature vector is generated through the data-driven feature extraction path; A set of core physical property parameters of rock mass with clear physical meaning are generated through the physical property parameterization path; the high-dimensional data-driven feature vector is input into the first prediction head network to generate the first predicted value of data-driven rock mass mechanical parameters; The core physical property parameters of the rock mass medium are substituted into a preset differentiable rock physical constitutive model to generate a second predicted value of rock mass mechanical parameters based on physical mechanism; a composite loss function including prediction loss term and physical consistency loss term is constructed. The composite loss function is used to perform end-to-end backpropagation training on the dual-path parallel deep neural network model; After the model training converges, a structured causal explanation report is generated using the core physical property parameters of the rock mass output by the physical property parameterization path and the first predicted value.

[0025] In the aforementioned steps, acquiring multimodal geological exploration raw data for the rock mass area to be evaluated specifically includes: acquiring three-dimensional seismic reflection wave signals covering the target area using seismic wave exploration equipment, and recording the time-domain waveform sequence of each receiving point under different excitation sources; obtaining continuous core samples through drilling operations, and using a high-resolution industrial camera to capture segmented images of the core surface, forming a pixel-level clear core image sequence; and simultaneously recording five types of logging curve data—natural gamma ray intensity, resistivity, acoustic transit time, density, and neutron porosity—along the borehole axis using geological logging instruments. These three types of data together constitute the multimodal geological exploration raw dataset, characterizing the rock mass state from three dimensions: macroscopic wavefield response, microscopic structural morphology, and in-situ physical properties.

[0026] The specific operations for data preprocessing of the original multimodal geological exploration data are as follows: For seismic wave exploration data, the short-time Fourier transform algorithm is used to convert it from the time domain to the time-frequency domain representation. The window length is set to 128 sampling points, the overlap rate is 50%, and finally a two-dimensional time-frequency spectrum is generated, with the horizontal axis being the time index, the vertical axis being the frequency component, and the pixel value representing the energy amplitude of the corresponding time-frequency unit.

[0027] For borehole core image data, Gaussian filtering is first performed to suppress imaging noise. Then, the Canny edge detection operator is applied to extract strong gradient boundaries in the image, followed by Hough transform to identify linear fracture structures. Simultaneously, a region growing algorithm is used to segment irregular structural areas. Finally, all detected fractures are abstracted into a graph structure, where nodes are defined as fracture intersections or endpoints, and edges are defined as fracture segments connecting two nodes. Each edge has three attributes: orientation angle, length, and aperture, thus constructing a fracture network adjacency matrix. For geological logging data, min-max normalization is performed independently on each type of logging curve, linearly mapping the original values ​​to the 0-1 interval to ensure comparability of data with different physical dimensions. After the above processing, the time-spectrum image, fracture network adjacency matrix, and normalized logging curves are concatenated along the channel dimension to form a unified multi-channel input feature tensor, which serves as the input to the subsequent neural network model.

[0028] The standardized multi-channel input feature tensors are input into a dual-path parallel deep neural network model. This model consists of two parallel processing paths: a data-driven feature extraction path and a physical property parameterization path. Both paths share an initial embedding layer network, which comprises two fully connected layers, each containing 512 neurons with a modified linear unit activation function, used to map heterogeneous input features to a unified high-dimensional semantic space. After mapping, the feature tensors are branched off to their respective dedicated paths for independent processing.

[0029] The data-driven feature extraction path consists of three parallel sub-networks targeting different data modalities. The first sub-network is a temporal convolutional network specifically designed for processing time-spectral maps. This network contains four levels of residual blocks, each consisting of two stacked dilated convolutional layers with dilation rates of 1, 2, 4, and 8, respectively. The kernel size is fixed at 3, and the number of channels is 64, 128, 256, and 512, respectively. Through this multi-scale dilated convolutional structure, this sub-network effectively captures the attenuation and dispersion characteristics of seismic wave signals at different propagation distances. The second sub-network is a graph convolutional network specifically designed for processing the fracture network adjacency matrix. This network contains three graph convolutional layers, each aggregating first-order neighborhood information using a mean aggregation strategy, with hidden units of 64, 128, and 256, respectively.

[0030] Graph convolution operations update the embedding vector of each node through a message-passing mechanism, and finally perform global average pooling on all node embeddings to obtain a feature vector representing the overall topological complexity of the fracture network. The third sub-network is a multilayer perceptron, specifically designed for processing normalized logging curves. This network contains three fully connected layers with 256, 512, and 256 neurons respectively, all using rectified linear units as activation functions. The output feature vectors of all three sub-networks have a dimension of 256 and are then input to an attention fusion layer. This attention fusion layer calculates the weight coefficients of each sub-network output using the following formula: in, For the first The output feature vectors of each subnetwork For learnable weight matrix, For bias vector This is the context vector. The weighted feature vectors are concatenated and then compressed into 128 dimensions through a fully connected layer, forming the final high-dimensional data-driven feature vector.

[0031] The physical property parameterization path is a fully connected neural network with five independent output heads. The main body of the network consists of four fully connected layers, each containing 1024 neurons, with a modified linear unit (MRU) activation function. The five output heads correspond to rock porosity, fracture network density, water saturation, mineral component vector, and uniaxial compressive strength of the rock, respectively. Specifically, the rock porosity output head uses a hard Z-shaped function as its activation function, ensuring that the output value strictly lies within the closed interval of 0 to 1; the water saturation output head also uses a hard Z-shaped function to constrain its physical validity; the fracture network density output head uses a modified linear unit (MRU) to ensure that the output value is non-negative; the mineral component vector output head contains multiple neurons, corresponding to major mineral components such as quartz, feldspar, mica, and calcite, and its output is normalized using a soft maximum function to ensure that the sum of the proportions of each component is 1. All output parameters have clear rock physical meanings, constituting a set of core physical property parameters for the rock mass.

[0032] The high-dimensional data-driven feature vector is input into the first prediction head network, which consists of three fully connected layers with 256, 128, and 1 neurons respectively. The last layer has no activation function and directly outputs a scalar value, namely the first predicted value of the data-driven rock mass mechanical parameters. Here, the rock mass mechanical parameters specifically refer to the rock mass elastic modulus, with units of gigapascals.

[0033] The core physical property parameters of the rock mass are substituted into a pre-defined differentiable rock physics constitutive model for forward calculation. This model, based on the equivalent medium theory, consists of a three-level calculation module connected in series: the Voigt-Reuss-Hill average model, the Kuster-Toksöz model, and the Gassmann equation. First, based on the mineral composition vector and the standard elastic constants of each mineral, the equivalent bulk modulus of the rock skeleton is calculated using the Voigt-Reuss-Hill average model. With shear modulus : ; in, and These are the Voigt upper limit and Reuss lower limit of the bulk modulus, respectively. and Similarly. Secondly, utilizing the porosity of the rock mass. With the density of the fracture network The equivalent elastic modulus of dry rock was calculated using the Kuster-Toksöz model. This model treats pores as spherical inclusions and fissures as oblate spheroidal inclusions, assigning different aspect ratio parameters to each. The effective medium equation is solved using a self-consistent approximation method to obtain the bulk modulus of the dry rock. With shear modulus Finally, the water saturation level was determined. Substituting into the Gassmann equation, the fluid substitution calculation is completed: ; in, The bulk modulus of the pore fluid is obtained by weighted averaging of the bulk moduli of water and air according to water saturation. Finally, based on... and Calculate the elastic modulus of saturated rock mass This serves as the second predicted value for rock mechanics parameters based on physical mechanisms. The entire calculation process is implemented entirely using tensor operations, ensuring end-to-end differentiation within the automatic differentiation framework.

[0034] The composite loss function is constructed, and its mathematical expression is as follows: ; in, The prediction loss term is defined as the root mean square error between the first predicted value and the true elastic modulus of the rock mass sample. The physical consistency loss term is defined as the smoothing L1 loss between the first and second predicted values. and The hyperparameters are set to 1 and 0.5 respectively, to balance the optimization objectives of data fitting and physical compliance.

[0035] The composite loss function was used to train the entire dual-path parallel deep neural network model through end-to-end backpropagation. The training process was performed on a computing device equipped with a tensor processing unit, with a batch size of 32. The Adam algorithm was selected as the optimizer, and the initial learning rate was set to 0.001, decaying by 10% every 10 training epochs. Training continued until the value of the composite loss function on the validation set decreased by less than 0.0001 for five consecutive epochs, at which point the model was considered converged.

[0036] After the model training converges, the causal interpretation generation module is activated. This module calculates the partial derivatives of the first predicted value with respect to each core physical property parameter of the rock mass medium, i.e.: ; in, The first prediction driven by data. For the first One core physical property parameter. Then, for all gradient values... Normalization is performed: ; Receive sensitivity coefficient The values ​​range from 0 to 1, with a total of 1. The causal explanation report for the structure is organized in key-value pairs and includes the following fields: rock porosity (value: 0.18, sensitivity: 0.35), fracture network density (value: 2.4 fractures / meter, sensitivity: 0.42), water saturation (value: 0.65, sensitivity: 0.15), mineral composition vector (quartz: 0.55, feldspar: 0.30, mica: 0.10, calcite: 0.05, sensitivity: 0.08), and uniaxial compressive strength of rock (value: 60 MPa, sensitivity: 0.20, score: 12). The report clearly indicates that fracture network density is the dominant factor leading to the current low elastic modulus in the region, contributing 42%, providing a quantitative basis for engineering support design.

[0037] The civil engineering rock mass stability assessment system includes a multimodal data acquisition and preprocessing module, a dual-path parallel deep neural network model, a data-driven prediction module, a physical mechanism deduction module, a model training and optimization module, and a causal interpretation generation module. The multimodal data acquisition and preprocessing module is responsible for executing the aforementioned data acquisition and preprocessing processes, outputting standardized multi-channel input feature tensors. The dual-path parallel deep neural network model is deployed on the tensor processing unit, receiving input tensors and outputting high-dimensional data-driven feature vectors and core physical property parameters of the rock mass medium in parallel.

[0038] The data-driven prediction module includes a first prediction head network that generates the first predicted value. The physical mechanism deduction module solidifies a differentiable rock physical constitutive model, receives physical property parameters, and outputs a second predicted value. The model training and optimization module calls a composite loss function during the training phase to drive backpropagation and update the network weights. The causal interpretation generation module calculates gradients and generates a structured report during the inference phase. Its built-in visualization interface generation unit renders the report as an interactive dashboard, allowing users to drill down to the corresponding anomaly areas in the original seismic profile or core image, achieving complete traceability from prediction results to geological evidence.

[0039] This embodiment achieves a unified approach of high-precision prediction and highly reliable interpretation in rock mass stability assessment through the aforementioned methods and systems. The model not only outputs the elastic modulus value but also reveals its physical origin, solving the "black box" problem of deep learning and significantly improving the scientific rigor and reliability of engineering decisions.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the stability of rock masses in civil engineering, characterized in that, include: Data preprocessing is performed on the original multimodal geological exploration data of the rock mass area to be evaluated to form a standardized multi-channel input feature tensor; The standardized multi-channel input feature tensor is input into a dual-path parallel deep neural network model, which includes a data-driven feature extraction path and a physical property parameterization path. Through the data-driven feature extraction path, multi-scale deep feature extraction is performed on the standardized multi-channel input feature tensor to generate a high-dimensional data-driven feature vector, which characterizes the potential nonlinear correlation of rock mass mechanical behavior. The standardized multi-channel input feature tensor is mapped through the physical property parameterization path to generate a set of core physical property parameters of rock mass with clear physical meaning. The core physical property parameters of rock mass include rock porosity, fracture network density, water saturation, mineral composition vector, and uniaxial compressive strength of rock. The high-dimensional data-driven feature vector is input into the first prediction head network to generate the first predicted value of the data-driven rock mechanics parameters. The core physical property parameters of the rock mass medium are substituted into a preset, differentiable rock physical constitutive model for forward calculation to generate a second predicted value of the rock mass mechanical parameters based on physical mechanism. Construct a composite loss function that includes a prediction loss term and a physical consistency loss term, wherein the prediction loss term is used to measure the difference between the first predicted value and the actual rock mass mechanical parameter value of the rock mass sample, and the physical consistency loss term is used to measure the difference between the first predicted value and the second predicted value; The composite loss function is used to perform end-to-end backpropagation training on the dual-path parallel deep neural network model, and the network weight parameters are iteratively optimized by the gradient descent algorithm until the composite loss function converges to below a preset threshold. After the model training converges, a structured causal explanation report is generated using the core physical property parameters of the rock mass output by the physical property parameterization path and the first predicted value. The causal explanation report quantitatively reveals the specific contribution of each core physical property parameter to the final predicted value of the rock mass mechanical parameters.

2. The method for assessing the stability of rock mass in civil engineering according to claim 1, characterized in that, The multimodal geological exploration raw data includes seismic wave exploration data, borehole core image data, and geological logging data; The multimodal geological exploration raw data is preprocessed, including performing time-frequency domain transformation on the seismic wave exploration data to obtain spectral features, performing edge detection and morphological analysis on the borehole core image data to extract the topological features of rock mass structural surfaces and fracture networks, and performing normalization processing on the geological logging data, ultimately forming a standardized multi-channel input feature tensor.

3. The method for assessing the stability of rock mass in civil engineering according to claim 1, characterized in that, Data preprocessing is performed on the raw multimodal geological exploration data, including: The seismic wave exploration data was converted into a two-dimensional time-spectrum map using a short-time Fourier transform algorithm. Gaussian filtering, Canny edge detection, Hough transform, and region growing algorithms are applied to the borehole core image data to generate a fracture network adjacency matrix that characterizes the location, orientation, length, and aperture of the fractures. Each type of logging curve in the geological logging data is independently subjected to min-max normalization, and the original values ​​are linearly mapped to the interval between 0 and 1. The two-dimensional time-frequency spectrum, the fracture network adjacency matrix, and the normalized logging curves are spliced ​​together according to the channel dimension to form the standardized multi-channel input feature tensor.

4. The method for assessing the stability of rock mass in civil engineering according to claim 3, characterized in that, The data-driven feature extraction path specifically consists of three parallel sub-networks: The temporal convolutional network used to process the two-dimensional time spectrogram contains multiple levels of residual blocks, each level of residual block is composed of two dilated convolutional layers stacked together, and the dilation rate increases progressively. The graph convolutional network used to process the adjacency matrix of the fracture network models the fracture network as a graph structure, where nodes represent fracture intersections or endpoints, edges represent fracture segments, and neighborhood information is aggregated through multi-layer graph convolution operations. Multilayer sensing machine for processing normalized logging curves; The output feature vectors of the three parallel sub-networks are weighted and concatenated through an attention fusion layer to form the high-dimensional data-driven feature vector.

5. The method for assessing the stability of rock mass in civil engineering according to claim 4, characterized in that, The attention fusion layer calculates the weight coefficients of each sub-network output, and the weight calculation formula is as follows: ； in, For the first The output feature vectors of each subnetwork For learnable weight matrix, For bias vector This is the context vector. The weighted feature vectors are concatenated and then compressed into 128 dimensions through a fully connected layer, forming the final high-dimensional data-driven feature vector.

6. The method for assessing the stability of rock mass in civil engineering according to claim 1, characterized in that, The physical property parameterization path is a fully connected neural network with multiple independent output heads, which correspond to rock porosity, fracture network density, water saturation, mineral composition vector, and uniaxial compressive strength of rock, respectively. Among them, the output activation function for rock mass porosity and water saturation is a hard Z-shaped function to ensure that the output value is within the closed interval of 0 to 1; The output activation function of the fracture network density is a modified linear unit to ensure that the output value is non-negative. The output of the mineral component vector is normalized by a soft maximum function to ensure that the sum of the proportions of each mineral component is 128-dimensional.

7. The method for assessing the stability of rock mass in civil engineering according to claim 1, characterized in that, The pre-defined, differentiable rock physics constitutive model is a combination of the Gassmann equation based on the equivalent medium theory and the Kuster-Toksöz model. Its calculation process includes: The equivalent bulk modulus and shear modulus of the rock skeleton were calculated using the mineral composition vector through the Voigt-Reuss-Hill average model. Using the porosity of the rock mass and the density of the fracture network, the equivalent elastic modulus of the dry rock containing pores and fractures is calculated by the Kuster-Toksöz model. Substituting the water saturation parameter into the Gassmann equation, the fluid substitution calculation is completed, and the second predicted value of the rock mass mechanical parameter based on the physical mechanism is obtained.

8. The method for assessing the stability of rock mass in civil engineering according to claim 1, characterized in that, The mathematical expression for the composite loss function is: ； in, The prediction loss term is defined as the root mean square error between the first predicted value and the true elastic modulus of the rock mass sample. The physical consistency loss term is defined as the smoothing L1 loss between the first and second predicted values. and This is a hyperparameter.

9. The method for assessing the stability of rock mass in civil engineering according to claim 1, characterized in that, The specific steps for generating the structured causal explanation report include: Calculate the partial derivative of the first predicted value of the data-driven rock mechanics parameters with respect to each core physical property parameter of the rock mass medium; The absolute values ​​of the partial derivatives are normalized to obtain the sensitivity coefficients of each core physical property parameter to the prediction results. The structured causal explanation report is stored and output in key-value pairs, where the key is the name of the core physical property parameter and the value is the specific predicted value of the parameter and its corresponding sensitivity coefficient.

10. A rock mass stability assessment system for civil engineering, characterized in that, include: The multimodal data acquisition and preprocessing module is used to acquire seismic wave exploration data, borehole core image data and geological logging data of the rock mass area to be evaluated, and to perform time-frequency domain transformation, fracture network topology extraction and data normalization operations to generate standardized multi-channel input feature tensors. The dual-path parallel deep neural network model is configured to receive the standardized multi-channel input feature tensor and perform parallel processing; the dual-path parallel deep neural network model includes a data-driven feature extraction path for generating high-dimensional data-driven feature vectors characterizing the potential nonlinear correlation of rock mass mechanical behavior, and a physical property parameterization path for generating a set of core physical property parameters of rock mass medium with clear physical meaning. The data-driven prediction module contains a first prediction head network, which receives the high-dimensional data-driven feature vector and outputs the first predicted value of the data-driven rock mechanics parameters. The physical mechanism deduction module has a differentiable rock physical constitutive model embedded in it. It is used to receive the core physical property parameters of the rock mass medium and output the second predicted value of the rock mass mechanical parameters based on the physical mechanism. The model training and optimization module internally defines a composite loss function that includes a prediction loss term and a physical consistency loss term. During the training phase, it is used to iteratively optimize the network weights of the dual-path parallel deep neural network model based on the first predicted value, the second predicted value, and the true value of the rock mass sample through the backpropagation algorithm. The causal explanation generation module is used to quantify the contribution of each physical property parameter to the final prediction result by calculating the gradient of the first predicted value relative to the core physical property parameters of the rock mass after the model has been trained, and to generate a structured causal explanation report.

Images