Complex rock mass stability assessment method based on back propagation neural network

Abstract

The invention provides a complex rock mass stability assessment method based on a back propagation neural network, and aims to provide an efficient and accurate assessment means for rock mass stability analysis under complex geological conditions. According to the method, the rock mass numerical model is constructed, influence factors are determined, and the stability of the rock mass is quantitatively analyzed and evaluated by using the back propagation neural network, so that the multi-scale characteristics of the rock mass, the influence of an interlayer staggered belt and external loading conditions can be comprehensively considered, and the stability of the rock mass in different stress states can be accurately predicted; scientific basis is provided for rock engineering design and construction, and engineering risks are reduced.

Description

Technical Field

[0001] The invention belongs to the field of rock mechanics, and in particular relates to a complex rock mass stability assessment method based on a back propagation neural network. Background Art

[0002] Rock mass stability assessment is a basic and critical task in rock engineering and underground engineering, which is directly related to the safety, economy and reliability of engineering design. Traditional rock mass stability assessment methods usually rely on geological surveys, laboratory tests and empirical formulas. These methods can provide useful information to a certain extent, but they often have limitations when dealing with complex geological conditions and large-scale rock mass structures. Especially when facing rock masses containing interlayer slip zones, multi-scale joints and other complex geological structures, the accuracy and efficiency of these methods are greatly limited.

[0003] With the development of computer technology, numerical simulation methods have become an important means of rock mass stability analysis. By establishing a detailed geological model and using advanced numerical analysis techniques, the mechanical behavior and failure process of rock mass under different working conditions can be simulated. However, numerical simulation methods require a lot of computing resources and have high requirements on the accuracy of model parameters. These parameters are often difficult to obtain through direct measurement and need to be estimated through experience or indirect methods.

[0004] Artificial intelligence technology, especially machine learning and deep learning, provides a new perspective for rock mass stability assessment. Artificial neural networks, as a computational model that mimics the structure of human brain neurons, can identify complex nonlinear relationships by learning large amounts of data. Back propagation neural networks, as one of the classic algorithms, are widely used in geotechnical engineering due to their simple structure, efficient training, and good generalization ability. However, although back propagation neural networks have shown potential in rock mass stability assessment, they still face some challenges in practical applications. For example, how to construct a neural network structure suitable for rock mass stability assessment, how to select appropriate input parameters, how to ensure the quality and representativeness of training data, and how to improve the prediction accuracy and generalization ability of the model all need further research and resolution. Therefore, developing a back propagation neural network rock mass stability assessment method that can comprehensively consider geological conditions, engineering characteristics, and construction technology is of great significance for improving the scientificity and practicality of the assessment. Summary of the invention

[0005] The present invention provides a complex rock mass stability assessment method based on a back propagation neural network, characterized in that the method comprises the following steps: Step S1, establishing a rock mass numerical model: according to actual engineering geological conditions, a three-dimensional numerical model of the rock mass to be evaluated including layered structure and joint characteristics is established for the target engineering area; Step S2, determining influencing factors: collecting historical data related to rock mass stability and conducting statistical analysis on them to identify key factors that have a significant impact on rock mass stability, including mechanical parameters of the rock mass, stress state, and cavern excavation conditions; Step S3, establishing a back propagation neural network: based on the key influencing factors determined in step S2, constructing a back propagation neural network model, including an input layer, a hidden layer and an output layer, wherein the input layer receives parameter information of the rock mass, the hidden layer processes the input information through nonlinear transformation, and the output layer outputs the stability assessment result of the rock mass; Step S4, training the back propagation neural network: using the rock mass stability data set collected in step S2, the constructed back propagation neural network model is trained, and the connection weights and thresholds in the model are optimized by the gradient descent method and the back propagation algorithm, so that the evaluation results output by the network maintain a high consistency with the historical stability data; Step S5, evaluating rock mass stability: inputting relevant parameters of the three-dimensional rock mass numerical model of the rock mass to be evaluated into the trained back propagation neural network to generate a stability evaluation result of the rock mass.

[0006] Furthermore, step S1 includes the following steps: Step S101: Conducting engineering geological survey on the target engineering area, collecting and summarizing its stratum lithology, geological structure characteristics, and spatial distribution characteristics of the rock mass around the surge chamber in the area, and establishing a layered rock mass model containing interlayer dislocation zones; Step S102: according to the actual survey results, a corresponding type of multi-scale columnar jointed rock mass is established at a corresponding position of the layered rock mass model; Step S103: simulating the excavation of a surge chamber in the multi-scale columnar jointed rock mass model to generate a three-dimensional numerical model including the chamber structure; Step S104: importing the generated three-dimensional digital model into numerical analysis software, performing three-dimensional reconstruction and generating a final numerical model.

[0007] Furthermore, in step S2, the mechanical parameters include the elastic modulus and shear strength of the rock mass in the numerical model generated in step S1, the stress state includes the initial stress field or externally applied stress in the numerical model generated in step S1, and the rock mass excavation conditions include the number of caverns, excavation steps, loading conditions, the existence state of columnar joints, the existence state of interlayer dislocation zones, as well as the mechanical parameters of the rock mass and the mechanical properties of the interlayer dislocation zones.

[0008] Furthermore, step S3 includes the following steps: Step S301: constructing a topological structure of a back propagation neural network, including an input layer, an output layer and at least one hidden layer, wherein the input layer is provided with multiple sets of parameters for describing the mechanical properties of the rock mass, excavation conditions and stress states, and the output layer contains four neurons, corresponding to four stability states of the rock mass: stable, basically stable, unstable and very unstable, wherein the number of nodes in the hidden layer is configured according to the complexity of the input parameters to realize nonlinear mapping; Step S302: Before training, the input and output data are processed into an appropriate format. The input data includes the discrete values ​​of the parameters of the state of stress loading direction, cavern excavation state, columnar joints and interlayer dislocation zone. The output data represents the four stability states described in step S301 in the form of one-hot encoding, where the evaluation criteria of the stability state are based on the analysis results of cracks and stress states. For the stable state, there are no cracks around the cavern, and the maximum principal stress of the rock mass around the cavern is relatively small; for the basically stable state, cracks appear in the rock mass around the cavern but no cracks appear along the interlayer dislocation zone; for the unstable state, when through cracks are formed along the interlayer dislocation zone; for the very unstable state, when the minimum principal stress of the rock mass around the cavern reaches the tensile strength of the rock, tensile failure occurs on the joint surface inside the model.

[0009] Furthermore, step S4 includes the following steps: Step S401: Based on the three-dimensional numerical model of the rock mass to be evaluated established in step S1, the model is simulated under different loading conditions to generate multiple rock mass stability data sets, each data set including input layer parameters and corresponding output layer stability states.

[0010] Step S402: Divide the rock stability data set into a training set and a test set, use the gradient descent method and the back propagation algorithm to adjust the connection weights and thresholds of the back propagation neural network, and monitor the mean square error during the training process until the error converges and meets the set threshold, so as to obtain a back propagation neural network model that can accurately predict rock stability.

[0011] Furthermore, in step S5, relevant parameters of the three-dimensional rock mass numerical model to be evaluated are input into the trained back propagation neural network, and the stability state of the rock mass is judged according to the activation degree of the four neurons in the output layer described in step S301, and the influence of the strain at the top arch, the number of caverns, the excavation state of the caverns, the stress state, the existence state of columnar joints and interlayer dislocation zones on the stability of the rock mass is comprehensively considered.

[0012] The beneficial effects of the present invention are: 1. High accuracy: The back propagation neural network can handle complex nonlinear problems and accurately predict the stability of rock masses. 2. High efficiency: Compared with traditional methods, the evaluation method based on the back propagation neural network greatly shortens the evaluation time and improves work efficiency. 3. Strong adaptability: This method is suitable for rock stability evaluation under various complex geological conditions and has strong versatility and adaptability. 4. Strong scalability: As more data accumulates, the prediction accuracy of the back propagation neural network can be improved through further training. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flow chart provided by the present invention; Figure 2 The three-dimensional numerical model of the rock mass provided by the present invention; Figure 3 A schematic diagram of the back propagation neural network structure provided by the present invention; Figure 4 The error variation curve during the back propagation neural network training process provided by the present invention; Figure 5 The confusion matrix of the back propagation neural network training samples provided by the present invention; Figure 6 The confusion matrix of the numerical model for rock mass stability assessment provided by the present invention. DETAILED DESCRIPTION

[0014] To further understand the present invention, the present invention is described in more detail below in conjunction with the accompanying drawings and specific embodiments.

[0015] Taking the Baihetan Hydropower Station as an example, the surge chamber is located in the columnar jointed basalt. A large number of unfavorable structures such as cracks of different scales, dense columnar joints, and hidden joints are developed in the rock mass, resulting in a huge difference in the strength and deformation characteristics of the rock mass from the rock blocks. At the same time, there are many large structural surfaces such as weak interlayer dislocation zones and steep faults in the rock mass. During the engineering excavation process, the deformation and failure mode of the multi-scale columnar jointed rock mass with interlayer dislocation zones has a significant impact on the stability of the underground caverns, posing a huge challenge to the engineering design and construction. It is necessary to evaluate the stability of the multi-scale columnar jointed rock mass with interlayer dislocation zones.

[0016] A complex rock mass stability assessment method based on a back propagation neural network comprises the following steps: establishing a rock mass numerical model, determining influencing factors, establishing a back propagation neural network, training the back propagation neural network and assessing rock mass stability.

[0017] The specific steps are as follows: Step S1, establishing a rock mass numerical model: according to actual engineering geological conditions, a three-dimensional numerical model of the rock mass to be evaluated including layered structure and joint characteristics is established for the target engineering area.

[0018] S101: Conduct engineering geological survey on the left bank of Baihetan Hydropower Station, summarize its stratum lithology, geological structure and spatial distribution characteristics of the rock mass around the surge chamber, and establish a layered rock model containing the C2 interlayer dislocation zone.

[0019] S102: Establish corresponding types of multi-scale columnar jointed rock masses at corresponding positions of the layered rock mass.

[0020] S103: Excavate a surge chamber in the multi-scale columnar jointed rock mass to obtain a three-dimensional digital model.

[0021] S104: After the three-dimensional digital model is established, the digital model is imported into the numerical software to obtain the three-dimensional reconstructed numerical model as shown in FIG. Figure 2 As shown, columnar joints are represented by orange vertical lines, clearly showing their distribution and orientation; the area of ​​the C2 interlayer dislocation zone is marked by the white transparent shaded area surrounded by dotted lines, reflecting its position and influence range in the model.

[0022] Figure 2 A 3D numerical model with different numbers of caverns is presented, where: Figure 2 Figure (a) shows the numerical model including only chamber 1; Figure 2 (b) shows the numerical model including cave 1 and cave 2; Figure 2 Figure (c) shows the numerical model including cave 1, cave 2 and cave 3; Figure 2 Figure (d) shows the numerical model including cave 1, cave 2, cave 3 and cave 4.

[0023] Step S2, determining influencing factors: collecting historical data related to rock mass stability and performing statistical analysis on them to identify key factors that have a significant impact on rock mass stability, including mechanical parameters of the rock mass, stress state, and cavern excavation conditions.

[0024] S201: According to the rock mass excavation conditions, such as the number of caverns, excavation steps, loading conditions, the existence of columnar joints, the existence of interlayer dislocation zones, etc., as well as the mechanical parameters of the rock mass and the mechanical properties of the interlayer dislocation zones, analyze their influence on the stability of the rock mass and determine the input parameters of the neural network.

[0025] Step S3, establishing a back propagation neural network: Based on the key influencing factors determined in step S2, a back propagation neural network model is constructed, including an input layer, a hidden layer and an output layer, wherein the input layer receives parameter information of the rock mass, the hidden layer processes the input information through nonlinear transformation, and the output layer outputs the stability assessment result of the rock mass.

[0026] S301: Build a topological structure of a back-propagation neural network, including an input layer, an output layer and at least one hidden layer.

[0027] like Figure 3 As shown, the input layer contains 25 parameters, represented by p1 to p25, which are defined as follows: Parameters 1 to 20 represent the Z-direction strains at 20 uniformly distributed points on the top arches of caverns 1 to 4, where: parameters 1 to 5 correspond to cavern 1; parameters 6 to 10 correspond to cavern 2; parameters 11 to 15 correspond to cavern 3; parameters 16 to 20 correspond to cavern 4; parameter 21 represents the loading direction of stress; parameter 22 represents the excavation stage of the cavern; parameter 23 represents the total number of caverns in the model; parameter 24 represents the existence status of columnar joints; and parameter 25 represents the existence status of the C2 interlayer dislocation zone.

[0028] Hassoun's related research shows that if the number of nodes contained in the hidden layer is large enough, the three-layer back-propagation neural network can realize any complex nonlinear mapping. Therefore, a hidden layer is set. According to the Kolmogorov theorem, the hidden layer contains 54 (2×25+4) neurons to realize complex nonlinear mapping.

[0029] The output layer sets four neurons, representing the four stability states of the rock mass: stable, basically stable, unstable and very unstable, represented by s1, s2, s3 and s4 respectively.

[0030] S302: Processing the input and output data into an appropriate format before training, specifically including: Input data: Parameter 21 represents the loading direction of stress, where the X direction and Y direction are represented by the values ​​1 and 2 respectively; Parameter 22 represents the excavation status of the cavern, where unexcavated, first step excavation, and second step excavation are represented by values ​​0, 1, and 2, respectively; Parameter 23 represents the number of chambers in the model, ranging from 1 to 4; Parameter 24 represents the existence of columnar joints, where the absence and existence are represented by values ​​0 and 1 respectively; Parameter 25 represents the existence state of the C2 interlayer dislocation zone, where the absence and existence are represented by the values ​​0 and 1 respectively.

[0031] Output data: One-hot encoding is used to represent the four stability states: [1 0 0 0] represents a stable state, [0 10 0] represents a basically stable state, [0 0 1 0] represents an unstable state, and [0 0 0 1] represents a very unstable state.

[0032] The evaluation criteria for stability status are as follows: When there are no cracks around the cavern, the maximum principal stress of the rock mass around the cavern is relatively small, and the rock mass is in a stable state; When cracks appear in the rock mass around the cavern, but no cracks appear along the C2 interlayer dislocation zone, the rock mass is in a basically stable state; When through cracks are formed along the C2 interlayer dislocation zone, the rock mass is in an unstable state; When the minimum principal stress of the rock mass around the cavern reaches the tensile strength of the rock and causes tensile failure on the joint surface, the rock mass is in a very unstable state.

[0033] Step S4, training the back propagation neural network: using the rock stability data set collected in step S2, the constructed back propagation neural network model is trained, and the connection weights and thresholds in the model are optimized by the gradient descent method and the back propagation algorithm, so that the evaluation results output by the network maintain a high degree of consistency with the historical stability data.

[0034] S401: Using the established three-dimensional numerical model, different models are simulated under different loading conditions, and a total of 296 data sets are obtained. Each data set includes input layer parameters, such as rock excavation conditions, etc., and corresponding output layer parameters, that is, known stability states.

[0035] S402: The collected 296 rock mass stability data sets are divided into training sets and test sets, of which the training set accounts for 80% and the test set accounts for 20%. During the training process, the gradient descent method and back propagation algorithm are used to adjust the connection weights and thresholds of each layer of the neural network, and the mean square error during the training process is monitored until the network output is consistent with the actual stability data with the minimum error. Finally, the parameters of the back propagation neural network that accurately reflects the stability of the rock mass can be obtained.

[0036] The error and parameter update gradients during training are as follows: Figure 4 As shown, Figure 4Figure (a) is the relationship curve between the training set mean square error, test set mean square error and the number of training times of the back propagation neural network. The mean square error is the expected value of the square difference between the predicted value and the true value. The smaller the mean square error value, the better the accuracy of the prediction model. It can be seen that the mean square error of the test set reaches a small value of 0.03 and remains stable after 120 training times. After that, the MSE of the training set continues to decrease with the increase in the number of training times, indicating that the error between the stability predicted by the back propagation neural network and the actual stability is getting smaller and smaller. Figure 4 Figure (b) shows the relationship between the parameter update gradient and the number of training times. As the training progresses, the parameter update gradient in the back propagation neural network is constantly adjusted. As the number of training times increases, the parameter update gradient first increases and then gradually decreases. After 120 training times, it reaches a low value and remains stable, indicating that the model training is about to be completed. When the number of training times is 329, the parameter update gradient reaches the set threshold and the model training is completed. At this time, the parameters after the back propagation neural network training can be used to predict the stability of the cave in the model.

[0037] The confusion matrix of the training set and the test set is as follows Figure 5 As shown, Figure 5 Figure (a) shows the confusion matrix of the training set. Figure 5 Figure (b) shows the confusion matrix of the test set, which is used to reflect the prediction accuracy of rock mass stability. In the confusion matrix, the horizontal axis represents the target class of the real stability of the rock mass, and the vertical axis represents the output class of the predicted stability of the rock mass. Figure 5 Taking the four values ​​in the first column of (a) as an example, the number 93 means that 93 training data with s1 stability were successfully predicted as s1 stability by the back propagation neural network. The remaining three numbers 0 mean that 0 training data with s1 stability were not successfully predicted as s2, s3 or s4 stability by the back propagation neural network. The values ​​on the diagonal, i.e., green squares, represent accurate predictions of rock stability. According to the confusion matrix, it can be seen that 0 out of 222 training data and 7 out of 74 test data predicted incorrect stability, which means that the trained back propagation neural network has a good prediction accuracy of about 97.6%. In addition, no data was predicted as non-adjacent levels of stability. It should be noted that among the 15 training data with s4 stability, 4 were predicted as s3, which may be caused by the two states of crack extension and stress in the model, that is, the development of cracks along the C2 interlayer dislocation zone and the stress in the rock mass reaching its tensile strength occurred almost simultaneously. In general, the trained back propagation neural network showed good performance and was able to accurately distinguish the stability of the four types of rock.

[0038] Step S5, evaluating rock mass stability: inputting relevant parameters of the three-dimensional rock mass numerical model of the rock mass to be evaluated into the trained back propagation neural network to generate a stability evaluation result of the rock mass.

[0039] The new rock sample parameters are input into the trained back-propagation neural network, and the stability state of the rock mass is judged according to the activation degree of the four neurons in the output layer.

[0040] A new numerical model without the C2 interlayer slip zone is created, and the numerical calculation results of each step of the model are converted into an appropriate format and evaluated by the trained back-propagation neural network. Figure 6 is the confusion matrix of the numerical model. It can be seen that when the simulation proceeds to step 4, that is, step 4, all excavation steps are completed, and the stability of the model is s1, corresponding to the position (1, 1) in the matrix. There is a tensile stress concentration phenomenon in the columnar joints, interlayer joint surfaces and caverns. When the stress in the Y direction is loaded to loading step 29, that is, step 29, the tensile stress in the rock mass around cavern No. 1 exceeds its tensile strength. At this time, the stability of the model is s4, corresponding to the position (4, 4) in the matrix, and the tensile stress is distributed in the columnar joints around the four caverns.

[0041] Therefore, the trained back propagation neural network model shows good accuracy in evaluating the stability of the rock mass, and also considers the influence of the strain at the crown arch, the number of caverns, the excavation status of the caverns, the stress state, the existence status of columnar joints, and the existence status of the C2 interlayer dislocation zone on the stability of the rock mass.

[0042] The above-mentioned specific implementation methods are used to explain the present invention and are only preferred embodiments of the present invention, rather than limiting the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A complex rock mass stability assessment method based on back propagation neural network, characterized by: The method comprises the following steps: Step S1, establishing a rock mass numerical model: according to actual engineering geological conditions, a three-dimensional numerical model of the rock mass to be evaluated including layered structure and joint characteristics is established for the target engineering area; Step S2, determining influencing factors: collecting historical data related to rock mass stability and conducting statistical analysis on them to identify key factors that have a significant impact on rock mass stability, including mechanical parameters of the rock mass, stress state, and cavern excavation conditions; Step S3, establishing a back propagation neural network: based on the key influencing factors determined in step S2, constructing a back propagation neural network model, including an input layer, a hidden layer and an output layer, wherein the input layer receives parameter information of the rock mass, the hidden layer processes the input information through nonlinear transformation, and the output layer outputs the stability assessment result of the rock mass; Step S4, training the back propagation neural network: using the rock mass stability data set collected in step S2, the constructed back propagation neural network model is trained, and the connection weights and thresholds in the model are optimized by the gradient descent method and the back propagation algorithm, so that the evaluation results output by the network maintain a high consistency with the historical stability data; Step S5, evaluating rock mass stability: inputting relevant parameters of the three-dimensional rock mass numerical model of the rock mass to be evaluated into the trained back propagation neural network to generate a stability evaluation result of the rock mass.

2. The complex rock mass stability assessment method based on back propagation neural network according to claim 1 is characterized by: Step S1 includes the following steps: Step S101: Conducting engineering geological survey on the target engineering area, collecting and summarizing its stratum lithology, geological structure characteristics, and spatial distribution characteristics of the rock mass around the surge chamber in the area, and establishing a layered rock mass model containing interlayer dislocation zones; Step S102: according to the actual survey results, a corresponding type of multi-scale columnar jointed rock mass is established at a corresponding position of the layered rock mass model; Step S103: simulating the excavation of a surge chamber in the multi-scale columnar jointed rock mass model to generate a three-dimensional numerical model including the chamber structure; Step S104: importing the generated three-dimensional digital model into numerical analysis software, performing three-dimensional reconstruction and generating a final numerical model.

3. The complex rock mass stability assessment method based on back propagation neural network according to claim 1 is characterized in that: In step S2, the mechanical parameters include the elastic modulus and shear strength of the rock mass in the numerical model generated in step S1, the stress state includes the initial stress field or externally applied stress in the numerical model generated in step S1, and the rock mass excavation conditions include the number of caverns, excavation steps, loading conditions, the existence state of columnar joints, the existence state of interlayer dislocation zones, as well as the mechanical parameters of the rock mass and the mechanical properties of the interlayer dislocation zones.

4. The complex rock mass stability assessment method based on back propagation neural network according to claim 1 is characterized in that: Step S3 includes the following steps: Step S301: constructing a topological structure of a back propagation neural network, including an input layer, an output layer and at least one hidden layer, wherein the input layer is provided with multiple sets of parameters for describing the mechanical properties of the rock mass, excavation conditions and stress states, and the output layer contains four neurons, corresponding to four stability states of the rock mass: stable, basically stable, unstable and very unstable, wherein the number of nodes in the hidden layer is configured according to the complexity of the input parameters; Step S302: Before training, the input and output data are processed into an appropriate format, wherein the input data includes the discrete values ​​of the parameters of the existence state of stress loading direction, cavern excavation state, columnar joints and interlayer dislocation zones, and the output data represents the four stability states described in step S301 in the form of one-hot encoding, wherein the evaluation criteria of the stability state are based on the analysis results of the crack and stress state.

5. The complex rock mass stability assessment method based on back propagation neural network according to claim 1 is characterized by: Step S4 includes the following steps: Step S401: Based on the three-dimensional numerical model of the rock mass to be evaluated established in step S1, the model is simulated under different loading conditions to generate multiple rock mass stability data sets, each data set including input layer parameters and corresponding output layer stability states; Step S402: Divide the rock mass stability data set into a training set and a test set, use the gradient descent method and the back propagation algorithm to adjust the connection weights and thresholds of the back propagation neural network, and monitor the mean square error during the training process until the error converges and meets the set threshold.

6. The complex rock mass stability assessment method based on back propagation neural network according to claim 1 is characterized in that: In step S5, the relevant parameters of the three-dimensional rock mass numerical model to be evaluated are input into the trained back propagation neural network, and the stability state of the rock mass is judged according to the activation degree of the four neurons in the output layer described in step S301, and the influence of the strain at the top arch, the number of caverns, the excavation state of the caverns, the stress state, the existence state of columnar joints and interlayer dislocation zones on the stability of the rock mass is comprehensively considered.

Images