Rock shear strength prediction method and system based on improved machine learning algorithm

Abstract

The invention relates to the technical field of data analysis, in particular to a rock shear strength prediction method and system based on an improved machine learning algorithm. The invention discloses a rock shear strength prediction system based on an improved machine learning algorithm. The rock shear strength prediction system comprises a data acquisition module and a rock shear strength analysis module. According to the method, the shear strength of the rock is predicted through the longitudinal wave speed, the density, the uniaxial compressive strength and the uniaxial tensile strength which are easy to obtain, complex work of a triaxial test can be not needed, a physical embedded input layer is arranged in a set rock shear strength analysis model, input features are expanded through priori knowledge of a physical rule, and the shear strength of the rock is predicted. The problem of low prediction accuracy caused by data sparsity is solved, so that the prediction is more accurate; and a knowledge enhancement adjustment layer is also arranged, and feature enhancement is carried out by taking rock shear strength prediction modes in different scenes as priori knowledge, so that the accuracy of rock shear strength prediction can be improved on the premise of interpretability.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and particularly to a method and system for predicting the shear strength of rocks based on an improved machine learning algorithm. Background Art

[0002] The shear strength parameters of rocks (internal friction angle and cohesion) are crucial basic parameters in geotechnical engineering design. Traditionally, they are mainly obtained through triaxial compression tests or direct shear tests. However, these conventional testing methods have limitations in many aspects: First, during the test process, experimenters must closely observe the rock failure process under high-pressure environments, bearing significant safety risks, such as the pressure system failure may cause equipment explosion, high-speed ejection of rock fragments and other dangers; Second, the test results under field conditions are often uncertain due to equipment limitations, environmental interferences (such as temperature fluctuations, vibrations) and human operation deviations, reducing the accuracy and reliability of the parameters; Third, standard triaxial tests usually take a long time (a single test may take several days to complete), are costly (requiring professional equipment and technical personnel) and have strict requirements for sample preparation.

[0003] In engineering practice, especially in the preliminary investigation stage of projects, geotechnical engineers often face time constraints and limited budgets, and cannot conduct sufficient triaxial tests. For example, in the feasibility study stage of tunnels, slopes and underground projects, the design team often needs to quickly evaluate the rock mass stability and formulate preliminary support plans without complete triaxial test data. At this time, engineers have to rely on empirical formulas or simplified assumptions to estimate the internal friction angle and cohesion, and the uncertainty of this method significantly increases the engineering risk. Summary of the Invention

[0004] The present invention predicts the shear strength of rocks through easily obtainable longitudinal wave velocity, density, uniaxial compressive strength and uniaxial tensile strength, without the need for the complex work of triaxial tests. And a physical embedding input layer is set in the established rock shear strength analysis model to expand the input features through the prior knowledge of physical rules, solving the problem of low prediction accuracy caused by data sparsity. It is also constrained by physical rules to make the prediction more accurate; A knowledge enhancement adjustment layer is also set, and the shear strength prediction methods of rocks in different scenarios are used as prior knowledge for feature enhancement, which can improve the accuracy of rock shear strength prediction on the premise of interpretability; A scale mapping layer is also set to adaptively transform the scale difference between rock specimens and rock masses, solving the problem of significant differences in shear strength prediction between rock specimens and actual rock masses; In summary, the rock shear strength analysis model set in this application not only has high interpretability, but also improves the accuracy of rock shear strength prediction under the guidance of professional knowledge in the field of rock shear analysis.

[0005] The present invention provides a method for predicting the shear strength of rocks based on an improved machine learning algorithm, including: Sending the rock analysis vector and the scale analysis vector into the rock shear strength analysis model for processing, and outputting the rock shear strength data; The rock shear strength analysis model includes a physical embedding input layer, a knowledge enhancement adjustment layer, a scale mapping layer, and an output layer. The physical embedding input layer is used to obtain the theoretical rock shear strength data and the reference rock shear strength data based on physical rules and the rock analysis vector, and construct a physically embedded rock analysis vector based on the theoretical rock shear strength data and the reference rock shear strength data. The knowledge enhancement adjustment layer processes the physically embedded rock analysis vector through a number of expert analysis units respectively, outputs the corresponding expert-enhanced rock analysis vectors, and performs feature fusion based on all the expert-enhanced rock analysis vectors to construct a knowledge-enhanced rock analysis vector. The scale mapping layer is used to perform scale transformation on the knowledge-enhanced rock analysis vector based on the scale analysis vector to construct a scale-mapped rock analysis vector, so as to obtain a more accurate prediction effect of the rock shear strength. The output layer is used to perform a fully connected operation on the scale-mapped rock analysis vector and output the rock shear strength data.

[0006] As a preferred aspect, the physical embedding input layer obtains the theoretical rock shear strength data and the reference rock shear strength data based on physical rules and the rock analysis vector, and constructs a physically embedded rock analysis vector based on the theoretical rock shear strength data and the reference rock shear strength data, which specifically includes the following steps: Calculating the corresponding theoretical internal friction angle and theoretical cohesion based on the Mohr-Coulomb criterion, and forming the theoretical rock shear strength data by combining the theoretical internal friction angle and the theoretical cohesion; Traversing the historical rock analysis database, calculating the similarity between the historical rock analysis data in the historical rock analysis database and the rock analysis vector. If the similarity between the historical rock analysis data in the historical rock analysis database and the rock analysis vector is higher than the similarity threshold, taking the historical rock shear strength data corresponding to the historical rock analysis data as the reference rock shear strength data, and the reference rock shear strength data includes a reference internal friction angle and a reference cohesion; until all the historical rock analysis data in the historical rock analysis database have been traversed, outputting all the reference rock shear strength data; Concatenate the rock analysis vector, the theoretical rock shear strength data, and all reference rock shear strength data at the beginning and end to construct a reconstructed rock analysis vector, and construct a physical correlation matrix based on the reconstructed rock analysis vector. The physical correlation strength value between the $i$-th data and the $j$-th data in the reconstructed rock analysis vector is stored in the $i$-th row and $j$-th column of the physical correlation matrix, where $i = 1, 2, 3, \ldots, N$ and $j = 1, 2, 3, \ldots, N$, and $N$ is the total number of data in the reconstructed rock analysis vector; perform a self-attention mechanism on the reconstructed rock analysis vector based on the physical correlation matrix to strengthen the reconstructed rock analysis vector and construct a physically embedded rock analysis vector.

[0007] As a preferred aspect, perform a self-attention mechanism on the reconstructed rock analysis vector based on the physical correlation matrix to construct a physically embedded rock analysis vector, which specifically includes the following steps: perform matrix multiplication operations on the reconstructed rock analysis vector with the value weight matrix and the key weight matrix respectively to construct the corresponding rock analysis value vector $V$ and rock analysis key vector $K$, perform a matrix multiplication operation on the physical correlation matrix with the query weight matrix to construct the corresponding rock analysis query vector $Q$, and implement the self-attention mechanism through the following formula: $G = \text{softmax}(QK^T / \sqrt{D})$, where $G$ is the physically embedded rock analysis vector, $T$ is the matrix transpose operation, and $D$ is the dimension size of the rock analysis key vector $K$. T / D 0.5 ), where $G$ is the physically embedded rock analysis vector, $T$ is the matrix transpose operation, and $D$ is the dimension size of the rock analysis key vector $K$.

[0008] As a preferred aspect, the knowledge enhancement adjustment layer processes the physically embedded rock analysis vector through several expert analysis units respectively, outputs the corresponding expert-enhanced rock analysis vectors, and performs feature fusion based on all expert-enhanced rock analysis vectors to construct a knowledge-enhanced rock analysis vector, which specifically includes the following steps: For each expert analysis unit, perform the following steps: perform a forgetting gate calculation on the physically embedded rock analysis vector: $f = \sigma(W_U + b)$, where $f$ is the forgetting gate vector, $\sigma$ is the sigmoid function, $W$ is the forgetting gate weight matrix, $b$ is the forgetting gate bias value, and $U$ is the physically embedded rock analysis vector; perform an input gate calculation on the physically embedded rock analysis vector: $r = \sigma(W_U + b)$, where $r$ is the input gate vector, $W$ is the input gate weight matrix, $b$ is the input gate bias value, and complete the feature transfer through the following formula: $H = f * U + r$ t $= \sigma(W f U + b f ), where $f t $ is the forgetting gate vector, $\sigma$ is the sigmoid function, $W f $ is the forgetting gate weight matrix, $b f $ is the forgetting gate bias value, $U$ is the physically embedded rock analysis vector, perform an input gate calculation on the physically embedded rock analysis vector: $r t $= \sigma(W r U + b r ), where $r t $= is the input gate vector, $W r $ is the input gate weight matrix, $b r $ is the input gate bias value, and complete the feature transfer through the following formula: $H = f t $*U + rt *U, where H is the corresponding expert-enhanced rock analysis vector, and * is the dot product operation; The physical embedding rock analysis vector is sent to the expert weight analysis network for processing to obtain the expert weight corresponding to each expert analysis unit. The weighted fusion operation is performed on all expert-enhanced rock analysis vectors and their corresponding expert weights through the feature fusion unit in the knowledge enhancement adjustment layer to obtain the knowledge enhancement rock analysis vector.

[0009] As a preferred aspect, the knowledge enhancement rock analysis vector is scale-transformed based on the scale analysis vector through the scale mapping layer to construct the scale mapping rock analysis vector, which specifically includes the following steps: The knowledge enhancement rock analysis vector and the scale analysis vector are concatenated and then sent to the scale mapping network for processing to output the scale transformation matrix. Then, the dot product operation is performed on the knowledge enhancement rock analysis vector and the scale transformation matrix to construct the scale mapping rock analysis vector.

[0010] As a preferred aspect, the rock shear strength analysis model is trained, which specifically includes the following steps: Obtain a number of rock shear strength analysis training samples. The rock shear strength analysis training samples include rock analysis vectors, scale analysis vectors, and corresponding rock shear strength data. All rock shear strength analysis training samples are divided into several scenario rock shear strength analysis training sets according to the scenarios corresponding to the expert analysis units. For each scenario rock shear strength analysis training set, the following training is performed: Construct a pre-training model, which includes a physical embedding input layer, a knowledge enhancement adjustment layer, a scale mapping layer, and an output layer. Only one expert analysis unit is set in the knowledge enhancement adjustment layer. The pre-training model is trained through the scenario rock shear strength analysis training set, and the training target is the rock shear strength data in the rock shear strength analysis training samples. Obtain all the trained pre-training models, and after parallelly concatenating the expert analysis units in all the trained pre-training models, form a new knowledge enhancement adjustment layer with the expert weight analysis network and the feature fusion unit. Based on the new knowledge enhancement adjustment layer, construct a complete rock shear strength analysis model. All rock shear strength analysis training samples are combined into a rock shear strength analysis training set, and the complete rock shear strength analysis model is trained through the rock shear strength analysis training set, and the training target is the rock shear strength data in the rock shear strength analysis training samples.

[0011] The present invention also provides a rock shear strength prediction system based on an improved machine learning algorithm, including: A data acquisition module, which forms a rock analysis vector by combining the longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength corresponding to the rock specimen, and forms a scale analysis vector by combining the scale of the rock specimen and the scale corresponding to the rock mass; A rock shear strength analysis module is used to process the rock analysis vector and the scale analysis vector by inputting them into a rock shear strength analysis model, and output rock shear strength data.

[0012] The present invention has the following advantages: The present invention predicts the rock shear strength through easily obtainable longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength, without the need for the complex work of triaxial tests. A physical embedding input layer is set in the set rock shear strength analysis model, and prior knowledge of physical rules is used to expand the input features, solving the problem of low prediction accuracy caused by data sparsity. Constraints are also imposed through physical rules to make the prediction more accurate. A knowledge enhancement adjustment layer is also set, and the rock shear strength prediction methods in different scenarios are used as prior knowledge to perform feature enhancement, which can improve the accuracy of rock shear strength prediction on the premise of interpretability. A scale mapping layer is also set to adaptively transform the scale difference between rock specimens and rock masses, solving the problem of significant differences in corresponding shear strength predictions between rock specimens and actual rock masses. In summary, the rock shear strength analysis model set in this application not only has high interpretability but also improves the accuracy of rock shear strength prediction under the guidance of professional knowledge in the field of rock shear analysis. Description of the Drawings

[0013] Figure 1 It is a schematic structural diagram of the rock shear strength analysis model adopted in the embodiment of the present invention.

[0014] Figure 2 It is a schematic structural diagram of the rock shear strength prediction system based on an improved machine learning algorithm adopted in the embodiment of the present invention. Detailed Embodiments

[0015] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0016] Embodiment 1, a rock shear strength prediction method based on an improved machine learning algorithm, includes: The longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength corresponding to the rock specimen are combined to form a rock analysis vector, and the scale of the rock specimen and the scale corresponding to the rock mass are combined to form a scale analysis vector. It should be noted that, in order to indirectly analyze the shear strength of the rock, the easily obtainable longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength are selected as inputs in this application, avoiding the complex work of triaxial tests, and the scale is characterized by the corresponding diameter. The rock analysis vector and the scale analysis vector are sent into the rock shear strength analysis model for processing, and the rock shear strength data is output. The rock shear strength data includes the internal friction angle and cohesion, and the internal friction angle and cohesion are key characteristic parameters for characterizing the shear strength of the rock; Here, a brief description of the acquisition methods corresponding to the longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength is given. Among them, the longitudinal wave velocity is obtained by placing an ultrasonic sound source in one hole and a receiver in another adjacent hole of the rock specimen, and calculating the longitudinal wave velocity according to the time when the ultrasonic wave passes through the rock specimen. The density of the rock specimen is measured through density measurement, and the uniaxial compressive strength and uniaxial tensile strength of the rock specimen are measured through a point load instrument; See Figure 1, the rock shear strength analysis model includes a physical embedding input layer, a knowledge enhancement adjustment layer, a scale mapping layer, and an output layer. The physical embedding input layer is used to calculate the rock analysis vector based on physical rules, construct the corresponding theoretical rock shear strength data, and obtain the reference rock shear strength data from the historical rock analysis database based on the rock analysis vector. The rock analysis vector is extended by the theoretical rock shear strength data and the reference rock shear strength data to construct a reconstructed rock analysis vector, and a corresponding physical association matrix is constructed to strengthen the reconstructed rock analysis vector to construct a physically embedded rock analysis vector. It should be noted here that since the rock analysis vector only has four data, there is a problem of data sparsity in the subsequent machine learning analysis process, resulting in poor actual analysis effects. Therefore, in the physical embedding input layer, the rock analysis vector is extended by the theoretical rock shear strength data and the reference rock shear strength data to increase the effective feature dimension, reduce the analysis error, and construct the theoretical rock shear strength data based on physical rules to establish physical relationship constraints between effective features. In addition to data-driven, physical constraint-driven is also realized based on the physical association matrix, improving the accuracy of rock shear strength prediction; the knowledge enhancement adjustment layer processes the physically embedded rock analysis vector through several expert analysis units respectively, outputs the corresponding expert-enhanced rock analysis vector, and performs feature fusion based on all expert-enhanced rock analysis vectors to construct a knowledge-enhanced rock analysis vector. The expert analysis units here correspond to the prediction of rock shear strength in different scenarios, such as scenarios where the rock is sandstone and its variants, granite, mudstone, dolomite, etc., or scenarios where the rock is under the action of groundwater or high in-situ stress. Since the prediction methods of the shear strength of rocks in different scenarios also vary with the scenarios, different expert analysis units are set in this application to fit the shear strength prediction in different scenarios, and the features in the physically embedded rock analysis vector are feature-enhanced with the expert analysis units as prior knowledge, which can improve the accuracy of rock shear strength prediction on the premise of interpretability; the scale mapping layer is used to perform scale transformation on the knowledge-enhanced rock analysis vector based on the scale analysis vector to construct a scale-mapped rock analysis vector. Since there are significant differences in scale between the rock specimen and the actual rock mass, there are also significant differences in the corresponding shear strength prediction. Therefore, it is necessary to perform scale transformation on the rock analysis vector to obtain a more accurate prediction effect of the rock shear strength; the output layer is used to perform a fully connected operation on the scale-mapped rock analysis vector and output the rock shear strength data; This application predicts the shear strength of rocks through easily obtainable longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength, eliminating the need for the complex work of triaxial tests. By setting a physical embedding input layer in the established rock shear strength analysis model and expanding the input features with prior knowledge of physical rules, it solves the problem of low prediction accuracy caused by data sparsity. Moreover, through physical rule constraints, the prediction becomes more accurate. A knowledge enhancement adjustment layer is also set up to strengthen features using the rock shear strength prediction methods in different scenarios as prior knowledge, improving the accuracy of rock shear strength prediction while maintaining interpretability. A scale mapping layer is set up to adaptively transform the scale difference between rock specimens and rock masses, solving the problem of significant differences in shear strength prediction between rock specimens and actual rock masses. In summary, the rock shear strength analysis model established in this application not only has high interpretability but also improves the accuracy of rock shear strength prediction under the guidance of professional knowledge in the field of rock shear analysis.

[0017] Based on the physical embedding input layer, theoretical rock shear strength data and reference rock shear strength data are obtained based on physical rules and rock analysis vectors, and a physical embedding rock analysis vector is constructed based on the theoretical rock shear strength data and reference rock shear strength data. The specific steps are as follows: Calculate the corresponding theoretical internal friction angle and theoretical cohesion based on the Mohr-Coulomb criterion. The specific calculation method is: the theoretical internal friction angle φ_theory = arcsin[(UCS - UTS) / (UCS + UTS)], where UCS is the uniaxial compressive strength and UTS is the uniaxial tensile strength. The theoretical cohesion c_theory = UCS(1 - sinφ_theory) / (2cosφ_theory). The theoretical internal friction angle and theoretical cohesion are combined to form the theoretical rock shear strength data. Traverse the historical rock analysis database, calculate the similarity between the historical rock analysis data in the historical rock analysis database and the rock analysis vector. The specific calculation method can adopt the cosine similarity algorithm. Here, the historical rock analysis data refers to the longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength collected historically. If the similarity between the historical rock analysis data in the historical rock analysis database and the rock analysis vector is higher than the similarity threshold (the similarity threshold is set artificially), take the historical rock shear strength data corresponding to the historical rock analysis data as the reference rock shear strength data. The reference rock shear strength data includes the reference internal friction angle and the reference cohesion. It should be noted that the historical rock shear strength data corresponding to the historical rock analysis data here is obtained through triaxial tests or output by a rock shear strength analysis model, and the historical rock shear strength data has confidence; until all the historical rock analysis data in the historical rock analysis database have been traversed, output all the reference rock shear strength data. These reference rock shear strength data can provide prior knowledge of similar rock shear strength while expanding features; Concatenate the rock analysis vector, the theoretical rock shear strength data, and all the reference rock shear strength data end to end to construct a reconstructed rock analysis vector, and construct a physical association matrix based on the reconstructed rock analysis vector. The physical association matrix stores the physical association strength values between the data items in the reconstructed rock analysis vector. Here, the physical association strength value refers to the correlation between physical quantities. For example, there is a correlation between the longitudinal wave velocity and density in the elastic wave propagation theory, and there is a correlation between the uniaxial compressive strength and uniaxial tensile strength in the brittle material failure theory. The physical association strength value is determined by experts through the fuzzy comprehensive evaluation method. The value stored in the i-th row and j-th column of the physical association matrix is the physical association strength value between the i-th data and the j-th data in the reconstructed rock analysis vector, where i = 1, 2, 3, …, N, j = 1, 2, 3, …, N, and N is the total number of data in the reconstructed rock analysis vector. Here, the data refers to the longitudinal wave velocity, density, uniaxial compressive strength, uniaxial tensile strength, theoretical internal friction angle, theoretical cohesion, internal friction angle, and reference cohesion; perform the self-attention mechanism on the reconstructed rock analysis vector based on the physical association matrix to strengthen the reconstructed rock analysis vector and construct a physically embedded rock analysis vector; Perform the self-attention mechanism on the reconstructed rock analysis vector based on the physical association matrix to construct a physically embedded rock analysis vector, which specifically includes the following steps: Perform matrix multiplication operations on the reconstructed rock analysis vectors with the value weight matrix and the key weight matrix respectively to construct the corresponding rock analysis value vector V and rock analysis key vector K. Perform matrix multiplication operations on the physical association matrix and the query weight matrix to construct the corresponding rock analysis query vector Q. The self-attention mechanism is implemented through the following formula: G = softmax(QK T / D 0.5 ), where G is the physically embedded rock analysis vector, T is the matrix transpose operation, and D is the dimension size of the rock analysis key vector K. By performing the self-attention mechanism on the reconstructed rock analysis vectors through the physical association matrix, the data that conforms to the physical association relationship in the reconstructed rock analysis vectors can be strengthened, and the data that does not conform to the physical association relationship can be suppressed, making the process of predicting the shear strength of rocks more in line with physical constraints; The knowledge enhancement adjustment layer processes the physically embedded rock analysis vectors through several expert analysis units respectively, outputs the corresponding expert-enhanced rock analysis vectors, and performs feature fusion based on all the expert-enhanced rock analysis vectors to construct the knowledge-enhanced rock analysis vector. The specific steps are as follows: For each expert analysis unit, perform the following steps: Calculate the forgetting gate for the physically embedded rock analysis vector: f t = σ(W f U + b f ), where f t is the forgetting gate vector, σ is the sigmoid function, W f is the forgetting gate weight matrix, b f is the forgetting gate bias value, and U is the physically embedded rock analysis vector. Calculate the input gate for the physically embedded rock analysis vector: r t = σ(W r U + b r ), where r t is the input gate vector, W r is the input gate weight matrix, b r is the input gate bias value. Complete the feature transfer through the following formula: H = f t * U + r t * U, where H is the corresponding expert-enhanced rock analysis vector, and * is the dot product operation. Through the forgetting gate and the input gate, the features that are not to be concerned and the features that need to be retained can be determined respectively in the scenario corresponding to the expert analysis unit, so as to realize the feature analysis process of the corresponding scenario; Send the physical embedded rock analysis vector into the expert weight analysis network for processing to obtain the expert weight corresponding to each expert analysis unit. Here, the expert weight analysis network is established based on a multi-layer perceptron. The obtained expert weight can represent the proportion of the corresponding rock sample's scenario. Perform a weighted fusion operation on all expert-enhanced rock analysis vectors and their corresponding expert weights through the feature fusion unit in the knowledge enhancement adjustment layer to obtain the knowledge-enhanced rock analysis vector; Perform a scale transformation on the knowledge-enhanced rock analysis vector based on the scale analysis vector through the scale mapping layer to construct the scale-mapped rock analysis vector, which specifically includes the following steps: Concatenate the knowledge-enhanced rock analysis vector and the scale analysis vector and send them into the scale mapping network for processing to output the scale transformation matrix. Here, the scale mapping network is also established based on a multi-layer perceptron. Then perform a dot product operation on the knowledge-enhanced rock analysis vector and the scale transformation matrix to construct the scale-mapped rock analysis vector. Process the knowledge-enhanced rock analysis vector through the scale transformation matrix to map it to the scale of the rock mass, which can make the prediction of the rock shear strength more in line with the actual scale of the rock mass; Train the rock shear strength analysis model, which specifically includes the following steps: Obtain several rock shear strength analysis training samples, which include rock analysis vectors, scale analysis vectors, and the corresponding rock shear strength data. Here, the rock analysis vectors, scale analysis vectors, and the corresponding rock shear strength data are measured by the operator according to the actual triaxial test. Divide all rock shear strength analysis training samples into several scenario rock shear strength analysis training sets according to the scenarios corresponding to the expert analysis units. For each scenario rock shear strength analysis training set, perform the following training: construct a pre-trained model, which includes a physical embedded input layer, a knowledge enhancement adjustment layer, a scale mapping layer, and an output layer. And only one expert analysis unit is set in the knowledge enhancement adjustment layer. Train the pre-trained model through the scenario rock shear strength analysis training set, and the training target is the rock shear strength data in the rock shear strength analysis training sample. Judge whether the training conditions are met. The training conditions generally mean that the accuracy of the pre-trained model meets the expectation. If the training conditions are met, output the trained pre-trained model; otherwise, continue to train the pre-trained model through the scenario rock shear strength analysis training set; Obtain all the trained pre-trained models, and after parallel splicing the expert analysis units in all the trained pre-trained models, form a new knowledge enhancement adjustment layer together with the expert weight analysis network and the feature fusion unit. Based on the new knowledge enhancement adjustment layer, construct a complete rock shear strength analysis model. Combine all the rock shear strength analysis training samples to form a rock shear strength analysis training set, and train the complete rock shear strength analysis model through the rock shear strength analysis training set. The training target is the rock shear strength data in the rock shear strength analysis training samples. Determine whether the training conditions are met. The training conditions generally mean that the accuracy of the rock shear strength analysis model meets the expectations. If the training conditions are met, output the trained rock shear strength analysis model; otherwise, continue to train the complete rock shear strength analysis model through the rock shear strength analysis training set.

[0018] Example 2, a rock shear strength prediction system based on an improved machine learning algorithm, see Figure 2 , including: A data acquisition module that forms a rock analysis vector from the longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength corresponding to the rock specimen, and forms a scale analysis vector from the scale of the rock specimen and the scale corresponding to the rock mass. It should be noted that, in order to indirectly analyze the rock shear strength, in this application, the easily obtainable longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength are selected as inputs, avoiding the complex work of triaxial tests, and the scale is characterized by the corresponding diameter. A rock shear strength analysis module for sending the rock analysis vector and the scale analysis vector into the rock shear strength analysis model for processing, and outputting rock shear strength data. The rock shear strength data includes the internal friction angle and cohesion, and the internal friction angle and cohesion are key characteristic parameters characterizing the rock shear strength.

[0019] It should be understood that for those of ordinary skill in the art, improvements or transformations can be made according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention. The parts not described in detail in this specification belong to the prior art well-known to those of ordinary skill in the art.

Claims

1. A method for predicting the shear strength of rock based on an improved machine learning algorithm, characterized in that, Including: Input the rock analysis vector and the scale analysis vector into the rock shear strength analysis model for processing, and output the rock shear strength data; The rock shear strength analysis model includes a physical embedding input layer, a knowledge enhancement adjustment layer, a scale mapping layer, and an output layer. The physical embedding input layer is used to obtain the theoretical rock shear strength data and the reference rock shear strength data based on physical rules and the rock analysis vector, and construct a physically embedded rock analysis vector based on the theoretical rock shear strength data and the reference rock shear strength data. The knowledge enhancement adjustment layer processes the physically embedded rock analysis vector through several expert analysis units respectively, outputs the corresponding expert-enhanced rock analysis vectors, and performs feature fusion based on all the expert-enhanced rock analysis vectors to construct a knowledge-enhanced rock analysis vector. The scale mapping layer is used to perform scale transformation on the knowledge-enhanced rock analysis vector based on the scale analysis vector to construct a scale-mapped rock analysis vector to obtain a more accurate prediction effect of the rock shear strength. The output layer is used to perform a fully connected operation on the scale-mapped rock analysis vector and output the rock shear strength data.

2. The rock shear strength prediction method based on an improved machine learning algorithm according to claim 1, wherein Based on physical rules and the rock analysis vector, the physical embedding input layer obtains the theoretical rock shear strength data and the reference rock shear strength data, and constructs a physically embedded rock analysis vector based on the theoretical rock shear strength data and the reference rock shear strength data. The specific steps are as follows: Calculate the corresponding theoretical internal friction angle and theoretical cohesion based on the Mohr-Coulomb criterion, and form the theoretical rock shear strength data with the theoretical internal friction angle and the theoretical cohesion; Traverse the historical rock analysis database, calculate the similarity between the historical rock analysis data in the historical rock analysis database and the rock analysis vector. If the similarity between the historical rock analysis data in the historical rock analysis database and the rock analysis vector is higher than the similarity threshold, use the historical rock shear strength data corresponding to the historical rock analysis data as the reference rock shear strength data. The reference rock shear strength data includes a reference internal friction angle and a reference cohesion; until all the historical rock analysis data in the historical rock analysis database have been traversed, output all the reference rock shear strength data; Concatenate the rock analysis vector, the theoretical rock shear strength data, and all the reference rock shear strength data end to end to construct a reconstructed rock analysis vector, and construct a physical association matrix based on the reconstructed rock analysis vector. The physical association strength value between the i-th data and the j-th data in the reconstructed rock analysis vector is stored in the i-th row and j-th column of the physical association matrix, where i = 1, 2, 3, …, N, j = 1, 2, 3, …, N, and N is the total number of data in the reconstructed rock analysis vector; perform a self-attention mechanism on the reconstructed rock analysis vector based on the physical association matrix to strengthen the reconstructed rock analysis vector and construct a physically embedded rock analysis vector.

3. The rock shear strength prediction method based on the improved machine learning algorithm according to claim 2, characterized in that, Perform self-attention mechanism on the reconstructed rock analysis vector based on the physical correlation matrix to construct a physically embedded rock analysis vector, which specifically includes the following steps: perform matrix multiplication operations on the reconstructed rock analysis vector with the value weight matrix and the key weight matrix respectively to construct the corresponding rock analysis value vector V and rock analysis key vector K, perform matrix multiplication operation on the physical correlation matrix with the query weight matrix to construct the corresponding rock analysis query vector Q, and implement the self-attention mechanism through the following formula: G = softmax(QK T / D 0.5 ), where G is the physically embedded rock analysis vector, T is the matrix transpose operation, and D is the dimension size of the rock analysis key vector K.

4. The rock shear strength prediction method based on the improved machine learning algorithm according to claim 3, characterized in that The knowledge-enhanced adjustment layer processes the physically embedded rock analysis vector through a number of expert analysis units respectively, outputs the corresponding expert-enhanced rock analysis vectors, and performs feature fusion based on all the expert-enhanced rock analysis vectors to construct a knowledge-enhanced rock analysis vector. The specific steps are as follows: For each expert analysis unit, perform the following steps: Calculate the forget gate for the physically embedded rock analysis vector: f t = σ(W f U + b f ), where f t is the forget gate vector, σ is the sigmoid function, W f is the forget gate weight matrix, b f is the forget gate bias value, and U is the physically embedded rock analysis vector. Calculate the input gate for the physically embedded rock analysis vector: r t = σ(W r U + b r ), where r t is the input gate vector, W r is the input gate weight matrix, b r is the input gate bias value. Complete the feature transfer through the following formula: H = f t * U + r t * U, where H is the corresponding expert enhanced rock analysis vector, and * is the dot product operation; Send the physically embedded rock analysis vector into the expert weight analysis network for processing to obtain the expert weight corresponding to each expert analysis unit. Perform a weighted fusion operation on all the expert-enhanced rock analysis vectors and their corresponding expert weights through the feature fusion unit in the knowledge-enhanced adjustment layer to obtain the knowledge-enhanced rock analysis vector.

5. The method for predicting the shear strength of rock based on an improved machine learning algorithm according to claim 4, characterized in that, Perform a scale transformation on the knowledge-enhanced rock analysis vector based on the scale analysis vector through the scale mapping layer to construct a scale-mapped rock analysis vector. The specific steps are as follows: Concatenate the knowledge-enhanced rock analysis vector and the scale analysis vector and send them into the scale mapping network for processing to output a scale transformation matrix. Then perform a dot product operation on the knowledge-enhanced rock analysis vector and the scale transformation matrix to construct the scale-mapped rock analysis vector.

6. The rock shear strength prediction method based on the improved machine learning algorithm according to claim 5, characterized in that, Train the rock shear strength analysis model. The specific steps are as follows: Obtain a number of rock shear strength analysis training samples. The rock shear strength analysis training samples include rock analysis vectors, scale analysis vectors, and the corresponding rock shear strength data. Divide all the rock shear strength analysis training samples into several scenario rock shear strength analysis training sets according to the scenarios corresponding to the expert analysis units. For each scenario rock shear strength analysis training set, perform the following training: Construct a pre-trained model. The pre-trained model includes a physically embedded input layer, a knowledge-enhanced adjustment layer, a scale mapping layer, and an output layer. And only one expert analysis unit is set in the knowledge-enhanced adjustment layer. Train the pre-trained model through the scenario rock shear strength analysis training set, and the training target is the rock shear strength data in the rock shear strength analysis training samples. Obtain all the trained pre-trained models. After parallelly concatenating the expert analysis units in all the trained pre-trained models, form a new knowledge-enhanced adjustment layer together with the expert weight analysis network and the feature fusion unit. Construct a complete rock shear strength analysis model based on the new knowledge-enhanced adjustment layer. Form a rock shear strength analysis training set with all the rock shear strength analysis training samples. Train the complete rock shear strength analysis model through the rock shear strength analysis training set, and the training target is the rock shear strength data in the rock shear strength analysis training samples.

7. A rock shear strength prediction system based on an improved machine learning algorithm, characterized in that, The system applies the rock shear strength prediction method based on the improved machine learning algorithm described in any one of claims 1-6 above, including: A data acquisition module that forms a rock analysis vector from the longitudinal wave velocity, density, uniaxial compressive strength, and uniaxial tensile strength corresponding to the rock specimen, and forms a scale analysis vector from the scale of the rock specimen and the scale corresponding to the rock mass. A rock shear strength analysis module for sending the rock analysis vector and the scale analysis vector into the rock shear strength analysis model for processing and outputting the rock shear strength data.

Images