Method for predicting shear strength of structural surfaces with different roughness in three-dimensional stress state

Abstract

The invention belongs to the field of mechanical calculation simulation, and particularly relates to a method for predicting shear strength of structural surfaces with different roughness in a three-dimensional stress state. The method comprises the following steps: firstly, preparing a standard rock sample and a silicone rubber sample, splitting the standard rock sample along a splitting line to form a split rock sample with a rough surface, scanning the rough wall surface of the split rock sample, constructing a three-dimensional model, and rendering to generate a two-dimensional image; then, under the three-dimensional stress condition, a shear strength test is carried out, and the shear strength of the structural surface is measured. And finally, associating the two-dimensional image with the shear strength to construct a data set, adopting a GoogLeNet deep convolutional neural network as a reference model, performing parameter adjustment through Bayesian optimization, optimizing a training process in combination with K-fold cross validation and an early stop strategy, evaluating model performance by using a test set, and continuously iterating to improve prediction precision. According to the method, the nonlinear mapping relation between the images of the structural surfaces with different roughness and the shear strength is established, and the mechanical behaviors such as the shear strength of the structural surfaces can be predicted.

Description

Technical Field

[0001] The invention belongs to the field of mechanical calculation simulation, and in particular relates to a method for predicting the shear strength of structural surfaces with different roughness in a three-dimensional stress state. Background Art

[0002] The rock mass is composed of continuous media of complete rock blocks and discontinuous characteristics of structural surfaces, and the overall rock mass presents typical discrete medium characteristics. As the main transmission path of mechanical effects between rock blocks, the shear strength of structural surfaces under three-dimensional stress state is a key mechanical parameter that determines the stability of rock masses. Most of the existing structural surface shear strength estimation models are based on direct shear tests of structural surfaces. However, when the rock mass structural surface is in a three-dimensional stress environment different from the direct shear test, the mechanical conditions will cause the shear mechanical behaviors of the structural surface, such as peak strength, dilatation-closure characteristics of the structural surface, and gap width, to change from the direct shear mechanical behaviors, which in turn limits the applicability of the structural surface shear strength estimation model, rock mass stability evaluation method, and disaster prevention and control countermeasures established around the direct shear behavior of the structural surface. Therefore, it is urgent to establish a structural surface shear strength estimation model suitable for three-dimensional stress states to overcome the limitations of existing models in complex mechanical environments.

[0003] The core of the structural surface shear strength estimation model lies in "how to evaluate the shear strength based on the roughness of the structural surface", that is, focusing on reflecting the influence of the two internal factors of the structural surface's own mechanical properties and roughness on the shear strength. P The roughness of the structural surface is characterized by morphological characteristic parameters such as 3D images, and the shear strength model is constructed through these parameters. However, this type of method has obvious limitations: first, a single or a small number of morphological parameters are difficult to fully characterize the geometric complexity of the structural surface, which can easily lead to the loss of key information; second, the existing models are mostly based on linear or simple nonlinear relationships, which cannot fully reflect the complex mapping relationship between roughness and shear strength under three-dimensional stress conditions; in addition, the roughness of different structural surfaces is different, which limits the applicability and prediction accuracy of the existing models. Therefore, it is urgent to introduce a more advanced method that can comprehensively extract and characterize the roughness information of the structural surface, accurately capture the nonlinear relationship between the roughness of the structural surface and the shear strength under the three-dimensional stress state, and then construct a more accurate and robust shear strength prediction model to provide reliable theoretical support and technical guarantee for the strength assessment of the structural surface under complex working conditions. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for predicting the shear strength of structural surfaces with different roughness under three-dimensional stress states. The method adopts the GoogLeNet deep convolutional neural network as the basic model, and through a multi-layer convolution module and a feature fusion mechanism, realizes the comprehensive extraction of complex roughness features, establishes a nonlinear mapping relationship between the images of structural surfaces with different roughness and the shear strength, and then accurately predicts the shear mechanical behavior such as the shear strength of the structural surface under the three-dimensional stress state, thereby providing reliable data support for engineering design and safety assessment.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solution: a method for predicting the shear strength of structural surfaces with different roughness in a three-dimensional stress state, comprising the following steps:

[0006] Step 1: Prepare standard rock sample Sam1: Use a rock coring machine, a rock cutting machine and a polishing machine to process the collected rock blocks to prepare a standard cylindrical rock sample Sam1.

[0007] Step 2: Prepare silicone rubber sample Sam2 and conduct tests: Use silicone rubber to cast a cylindrical silicone rubber sample Sam2 with a thickness of h and a diameter the same as that of the standard rock sample Sam1. After curing, perform stress-strain curve tests on the silicone rubber sample Sam2 under various confining pressure conditions.

[0008] Step 3: Prepare split rock sample Sam4: Use a universal testing machine and a splitting fixture to split the standard rock sample Sam1 along the height direction, so that it breaks along different splitting lines into two left and right rock samples Sam3 with different roughness structural surfaces, and cut off the rock block with a thickness of h at the upper or lower end of the rock sample Sam3 to form a split rock sample Sam4.

[0009] Preferably, Sam1 is broken into two left and right rock samples Sam3 along different cleavage lines, and rock blocks with a thickness of h at the upper end or the lower end of the left and right rock samples Sam3 are respectively cut off, and the inner wall surfaces of the left and right rock samples Sam3 are respectively scanned, that is, if a rock block with a thickness of h is cut off at the upper end position of the left rock sample, a rock block with a thickness of h is cut off at the lower end position of the right rock sample.

[0010] More preferably, during splitting, a thin steel bar or an inverted triangle indenter or other tool should be used to apply the splitting line load, and a flat indenter should be avoided.

[0011] Step 4: Constructing the three-dimensional morphology model D1: Use a morphology scanner to scan the walls on both sides of the structural surface of the split rock sample Sam4 to generate point cloud data, and encrypt the point cloud data through the nearest neighbor interpolation method to construct the three-dimensional morphology model D1.

[0012] Preferably, when the collected morphological discrete point cloud data contains noise points other than the structural surface, it is necessary to perform a removal process.

[0013] Step 5: Export the two-dimensional image file D2 of the structural surface: adjust the viewing angle and range, apply surface rendering technology to visualize the roughness of the structural surface, express the roughness through color gradient, and export the two-dimensional image file D2 of the structural surface.

[0014] Step 6: Prepare test sample Sam5: Wrap a plastic dura mater on the outside of the cut end of the split rock sample Sam4 to aggregate the left and right rock samples, ensure that the plastic dura mater is flush with the end surface, and fill the plastic dura mater cavity with silicone rubber of thickness h to supplement the removed rock block, thus forming a test sample Sam5 with the same diameter and length as the standard cylindrical rock sample Sam1.

[0015] Step 7: Conduct a three-dimensional stress state shear strength test: After the test sample Sam5 is cured, a three-dimensional stress state shear strength test is conducted on the test sample Sam5 using the confining pressure conditions used for the silicone rubber sample Sam2 in the second step, and the load and displacement changes during the shear process are recorded to draw a shear load-shear displacement curve.

[0016] Preferably, when the test sample Sam5 is subjected to a three-dimensional stress state shear strength test, the maximum shear displacement does not exceed the silicone rubber thickness h.

[0017] The shear strength test of the test specimen Sam5 under three-dimensional stress state was carried out under different confining pressure conditions. The purpose was to ensure that the neural network could accurately capture the nonlinear mapping relationship between the structural surface roughness and the shear strength by controlling the change of lateral pressure.

[0018] Step 8: Calculate the shear strength τ of structural surfaces with different roughness ts :According to the shear load-shear displacement curve drawn in the seventh step, the peak shear load is obtained. According to the peak shear load and the triaxial compression axial load of the silicone rubber sample Sam2, the shear strength τ of the structural surfaces with different roughness is calculated. ts , where each shear strength τ ts The calculation formula is expressed as formula (1), τ ts =(F′-F) / S(1), where F′ refers to the force exerted on the lower part of the left rock block and the upper part of the right rock block when the shear stress of the structural surface in the three-dimensional stress state reaches the peak value, F refers to the triaxial compression axial load value of the silicone rubber sample Sam2, and S is the shear action area of ​​the left and right rock blocks.

[0019] The triaxial compressive axial load of the silicone rubber sample Sam2 refers to the corresponding force value of the silicone rubber sample Sam2 when the structural surface reaches the peak shear stress under the same confining pressure conditions.

[0020] Since the calculation formula of the structural surface shear stress F0 is formula (2), F+F′=F0(2), where F′ refers to the force exerted on the lower part of the left rock block and the upper part of the right rock block when the structural surface shear stress reaches the peak value in the three-dimensional stress state, and F refers to the triaxial compression axial load value of the silicone rubber sample Sam2.

[0021] Preferably, substitute formula (2) into formula (1) in step S8, and the shear strength τ ts Further expressed as formula (3), τ ts =(F0-2F) / S(3), where F0 is the structural surface shear stress, F refers to the triaxial compression axial load value of the silicone rubber sample Sam2, and S is the shear action area of ​​the left and right rock blocks.

[0022] Step 9: Construct the image and mechanical strength data set D3 of the structural surface: Repeat steps 1 to 5 to obtain multiple split rock samples Sam4 with different rough surfaces, and export the two-dimensional image file D2 corresponding to each split rock sample Sam4; Repeat steps 6 to 8 to obtain the shear strength τ of each test sample Sam5 ts ; The two-dimensional image file D2 of each structural surface and its corresponding shear strength τ ts The data set D3 containing images and mechanical strength is constructed.

[0023] Step 10: Train the GoogLeNet model: randomly divide the data set D3 into training set D31 and test set D32, and use the training set D31 to train the computational model.

[0024] The GoogLeNet deep convolutional neural network is used as the basic model, and the hyperparameters are automatically tuned through the Bayesian optimization algorithm to obtain the optimal parameter combination.

[0025] K-fold cross-validation training is performed on the training set D31. The training set is divided into K subsets. K-1 subsets are randomly selected for training each time, and the remaining 1 subset is used for validation. The cycle is executed until the average loss reaches the predetermined standard, and the training of the calculation model is completed.

[0026] Preferably, the GoogLeNet deep convolutional neural network is used as the basic model, and its overall architecture includes an input layer, a convolution layer, a maximum pooling layer, an Inception module, a global average pooling layer, a fully connected layer and an output layer, wherein the Inception module consists of four branches: 1x1 convolution, 3x3 convolution, 5x5 convolution and a pooling layer.

[0027] More preferably, when the hyperparameters of the GoogLeNet model are optimized by the Bayesian optimization algorithm, the hyperparameters that need to be optimized include: the convolution kernel size, step size and padding method of the convolution layer, the pooling window size and step size of the pooling layer, and the number of filters of each branch in the Inception module.

[0028] Preferably, when the hyperparameters of the GoogLeNet model are optimized by the Bayesian optimization algorithm, the Gaussian process is used as the proxy function, the expected improvement EI is used as the acquisition function, and the mean square error MSE of the validation set is used as the objective function.

[0029] Preferably, when the mean square error (MSE) of the objective function validation set reaches the minimum value, the corresponding hyperparameter combination is the optimal hyperparameter combination. The calculation formula is shown in formula (4): Where n is the number of samples; y n is the true value corresponding to the nth sample; is the predicted value corresponding to the nth sample.

[0030] Preferably, step S10 is executed repeatedly until the average loss reaches a predetermined standard, and the root mean square error RMSE is used as the loss function. The calculation formula of the average loss is shown in formula (5): Where n is the number of samples; y n is the true value corresponding to the nth sample; is the predicted value corresponding to the nth sample.

[0031] Preferably, an “early stopping strategy” is used to dynamically monitor the average loss of the validation set. When the model does not achieve performance improvement in several consecutive rounds, training is automatically stopped to ensure that training ends at the best time.

[0032] Step 11: Use the test set D32 to test the trained GoogLeNet model, evaluate the model's prediction performance, and quantify its performance by calculating the accuracy index.

[0033] Preferably, the calculation accuracy index adopts the determination coefficient R 2 , which is calculated as shown in formula (6), (6), where n is the number of samples in the test set, y i is the true value of the internal friction angle of the test set sample, is the predicted value of the sample internal friction angle, is the average value of the true value of the internal friction angle of the sample.

[0034] Step 12: Perform error analysis based on the test results in step 11 to evaluate the deviation between the predicted value and the actual value. When the prediction error exceeds expectations, make improvements by optimizing hyperparameters, expanding the data set, and adjusting the model structure. Repeat steps 10 and 11 until the model performance reaches the expected standard.

[0035] Preferably, when evaluating, the accuracy index determination coefficient R 2 When the value of is greater than or equal to 0.90, the model performance is considered to have reached the expected standard and no further optimization is required; 2 If it is less than 0.90, then make improvements by adjusting hyperparameters, expanding the data set, or optimizing the model structure, and repeat steps 10 and 11 until the accuracy index determination coefficient R 2 The value of is greater than or equal to 0.90.

[0036] The beneficial effects of the present invention are as follows: (1) Through the multi-layer convolution module and feature fusion mechanism of the GoogLeNet deep convolutional neural network, it is possible to efficiently capture the details of the structural surface at different scales, comprehensively characterize its complex geometric morphology, and avoid the omission of important information. (2) It can accurately capture the complex nonlinear mapping relationship between the roughness of the structural surface and the shear strength, thereby realizing the accurate prediction of the shear mechanical behavior of the structural surface under three-dimensional stress conditions, and providing reliable support for engineering design and safety assessment. (3) It has excellent generalization ability and can adapt to various types of structural surfaces. Regardless of how their geometric features or material properties change, the model can maintain stable and reliable prediction performance. (4) Through the automated prediction process, human intervention is reduced, errors caused by subjective experience are avoided, and the consistency and reliability of the results are improved. (5) It has efficient computing performance, supports real-time prediction and rapid decision-making, and can provide timely data support for engineering design, construction and safety management, reduce construction risks, and ensure project safety. (6) The framework has good scalability and foresight, and other advanced neural networks can be integrated in the future to further optimize performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the algorithm flow chart of the present invention;

[0038] Figure 2 This is a structural diagram of applying an axial load to the test sample Sam5 during the shear strength test in the present invention;

[0039] Figure 3 In the present invention, when no shear force is applied during the shear strength test Figure 2 Force diagram of middle section A;

[0040] Figure 4 In the shear strength test of the present invention, after the shear force is applied Figure 2Force diagram of middle section A;

[0041] Figure 5 The original structural diagram of the walls on both sides of the structural surface of the split rock sample Sam4 in the present invention;

[0042] Figure 6 A point cloud data map generated after scanning in the present invention;

[0043] Figure 7 is the encrypted point cloud data map in the present invention;

[0044] Figure 8 A three-dimensional morphology model constructed in the present invention;

[0045] Fig. 9 A two-dimensional image file of the structure surface derived in the present invention;

[0046] Fig.10 This is a schematic diagram of the GoogLeNet model architecture. DETAILED DESCRIPTION

[0047] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the structure of the present invention.

[0048] Embodiment 1: Figure 1 As shown in the flowchart in, a method for predicting the shear strength of structural surfaces with different roughness in a three-dimensional stress state comprises the following steps:

[0049] Step 1: Prepare standard rock sample Sam1: Use a rock coring machine, a rock cutting machine and a polishing machine to process the collected rock blocks to prepare a Φ50×100 (mm) standard cylindrical rock sample Sam1.

[0050] Step 2: Prepare silicone rubber sample Sam2 and conduct tests: Use silicone rubber to cast a cylindrical silicone rubber sample Sam2 with a thickness of 10 mm and a diameter of 50 mm. After 8 days of curing, stress-strain curve tests are conducted on the silicone rubber sample Sam2 under four confining pressure conditions of 3 MPa, 5 MPa, 7 MPa and 9 MPa.

[0051] Step 3: Prepare the split rock sample Sam4: Use a universal testing machine and a splitting fixture to split the standard rock sample Sam1 in the height direction, so that different splitting lines along the height direction are broken into two left and right rock samples Sam3 with different roughness structural surfaces, and cut off the rock block with a thickness of 10mm at the upper or lower end of the rock sample Sam3 to form the split rock sample Sam4. During the splitting, an inverted triangle pressure head is used to apply a linear load to ensure that the stress is concentrated on the splitting line to avoid the use of a flat pressure head causing unevenness of the force-bearing surface. The original pattern of the walls on both sides of the structural surface of the split rock sample Sam4 is as follows: Figure 5shown.

[0052] Step 4: Construct a 3D topographic model D1: Use a topographic scanner to scan the two side walls of the split rock sample Sam4 structure surface to generate point cloud data, such as Figure 6 As shown; point cloud data is encrypted by nearest neighbor interpolation method, as shown Figure 7 As shown; construct a three-dimensional morphology model D1, such as Figure 8 shown.

[0053] In this step, the filtering method based on the distance threshold is used to remove abnormal points that are significantly deviated from the structural surface, and the outlier points with low local density are identified and deleted through the statistical outlier analysis method. After the noise point removal is completed, the nearest neighbor interpolation method is used to encrypt the processed point cloud data, and the discrete point spacing is set to 0.25mm. Finally, a three-dimensional morphology model D1 that meets the accuracy requirements is generated, as shown in Figure 1. Figure 8 shown.

[0054] Step 5: Export the two-dimensional image file D2 of the structural surface: Based on the completed three-dimensional morphology model D1, adjust the viewing angle and range to a top view, and observe the structural surface at a vertical angle. Use surface rendering technology to visualize the roughness of the structural surface, and use color gradients to intuitively display the roughness of different areas. The darker the color, the higher the roughness, and the lighter the color, the lower the roughness. After rendering, export the top view visualization result of the structural surface as a two-dimensional image file D2 in RGB format, adjust the length and width of the image to 1024 pixels, and set the number of color channels to 3, such as Fig. 9 shown.

[0055] Step 6: Prepare test sample Sam5: Wrap the plastic hard film on the outside of the cut end of the split rock sample Sam4, so that the left and right rock samples are aggregated together, and ensure that the plastic hard film is flush with the end surface to avoid gaps affecting subsequent operations. Fill the plastic hard film cavity with a thickness of 10mm layer by layer. After each layer is filled, it can be properly vibrated or tapped to expel residual bubbles, ensure that the silicone rubber is tightly combined with the end surface of the rock sample, ensure that the filling process is uniform and free of bubbles, and form a flat and stable end surface to prepare for subsequent experiments. After the silicone rubber is filled, let the sample stand and solidify to form a test sample Sam5 with the same diameter and length as the standard cylindrical rock sample Sam1.

[0056] Step 7: After curing the test sample Sam5 and ensuring that the sample is stable, the confining pressure conditions used for the silicone rubber sample Sam2 in the second step are used, that is, the test sample Sam5 is subjected to a three-dimensional stress state shear strength test under four confining pressure conditions of 3MPa, 5MPa, 7MPa and 9MPa, and the load and displacement changes during the shear process are recorded, and the shear load-shear displacement curve is drawn. During the test, the maximum shear displacement is set to 10mm.

[0057] The shear strength test of the test specimen Sam5 under three-dimensional stress state was carried out under different confining pressure conditions. The purpose was to ensure that the neural network could accurately capture the nonlinear mapping relationship between the structural surface roughness and the shear strength by controlling the change of lateral pressure.

[0058] The principle and method of the shear strength test method of the structural surface under the whole three-dimensional stress state are as follows Figures 2 to 4 As shown. Figure 2 Under the three-dimensional stress state, axial loads are applied to the upper and lower ends of the standard rock sample Sam1, causing it to break along section A, forming two left and right rock samples Sam3. The upper end surface of the left rock sample is cut off at a height of h, and the lower section of the right rock sample is cut off at a height of h, and filled with silica gel at a height of h to make it consistent with the appearance of the standard rock sample Sam1. Figure 3 , the upper end of the left rock sample is the filled silica gel end and the lower end is the rock block end, while the upper end of the right rock sample is the rock block end and the lower end is the filled silica gel end, then the load F is applied to the silica gel end of the left and right rock samples, and the load F′ is applied to the rock block end. Figure 3 , the shear strength test of the test sample Sam5 under three-dimensional stress state was carried out under different confining pressure conditions, the load and displacement changes during the shear process were recorded, and the shear load-shear displacement curve was drawn. During the test, the maximum shear displacement was set to 10mm.

[0059] Step 8: Calculate the shear strength τ of structural surfaces with different roughness ts :According to the shear load-shear displacement curve drawn in the seventh step, the peak shear load is obtained. According to the peak shear load and the triaxial compression axial load F of the silicone rubber sample Sam2, the shear strength τ of the structural surfaces with different roughness is calculated. ts , where each shear strength τ ts The calculation formula is expressed as formula (1),

[0060] τ ts =(F′-F) / S(4), where F′ refers to the force exerted on the lower part of the left rock block and the upper part of the right rock block when the shear stress of the structural surface in the three-dimensional stress state reaches the peak value, F refers to the triaxial compression axial load value of the silicone rubber sample Sam2, and S is the shear action area of ​​the left and right rock blocks.

[0061] The triaxial compressive axial load of the silicone rubber sample Sam2 refers to the corresponding force value of the silicone rubber sample Sam2 when the structural surface reaches the peak shear stress under the same confining pressure conditions.

[0062] Since the calculation formula of the structural surface shear stress F0 is formula (2), F+F′=F0(5), where F′ refers to the force exerted on the lower part of the left rock block and the upper part of the right rock block when the structural surface shear stress reaches the peak value in the three-dimensional stress state, and F refers to the triaxial compression axial load value of the silicone rubber sample Sam2.

[0063] Preferably, substitute formula (2) into formula (1) in step S8, and the shear strength τ ts Further expressed as formula (3), τ ts =(F0-2F) / S(6), where F0 is the structural surface shear stress, F refers to the triaxial compression axial load value of the silicone rubber sample Sam2, and S is the shear action area of ​​the left and right rock blocks.

[0064] Therefore, we can directly use formula (3) to calculate the shear strength τ of different roughness structural surfaces by substituting the peak shear load F0 and the triaxial compression axial load F of the silicone rubber sample Sam2: ts .

[0065] Step 9: Construct the image of the structural surface and the mechanical strength data set D3: Construct the image of the structural surface and the mechanical strength data set D3: Repeat steps 1 to 5 to obtain multiple split rock samples Sam4 with different rough surfaces obtained along different splitting lines, and export the two-dimensional image file D2 corresponding to each split rock sample Sam4; Repeat steps 6 to 8 to obtain the shear strength τ of each test sample Sam5 ts ; The two-dimensional image file D2 of each structural surface and its corresponding shear strength τ ts The data set D3 containing images and mechanical strength is constructed.

[0066] Step 10: Train the GoogLeNet model: Fig.10 As shown, the data set D3 is randomly divided into a training set D31 and a test set D32, and the training set D31 is used to train the computing model, wherein the division ratio of the training set D31 to the test set D32 is 4:1.

[0067] The GoogLeNet deep convolutional neural network is used as the basic model, which consists of 1 input layer, 3 convolutional layers, 3 maximum pooling layers, 9 Inception modules, 1 global average pooling layer, 1 fully connected layer and 1 output layer. The key parameters of the model are automatically tuned using the Bayesian optimization algorithm, including the convolution kernel size, step size and padding method of the convolution layer, the window size and step size of the pooling layer, and the number of filters in each branch of the Inception module.

[0068] The mean square error (MSE) of the validation set is used as the objective function. When the MSE reaches the minimum value, the corresponding hyperparameter combination is the optimal combination. The calculation formula is as follows: Where n is the number of samples; y n is the true value corresponding to the nth sample; is the predicted value corresponding to the nth sample.

[0069] A 5-fold cross-validation training was performed on the training set D31. The training set was divided into 5 subsets. Four subsets were randomly selected for training each time, and the remaining one subset was used for validation. The cycle was executed until the average loss reached the predetermined standard or there was no improvement for 10 consecutive rounds. The training of the computational model was completed.

[0070] The training process is dynamically monitored based on the loss of the validation set. The root mean square error (RMSE) is used as the loss function. The calculation formula for the average loss is shown in formula (5): Where n is the number of samples; y n is the true value corresponding to the nth sample; is the predicted value corresponding to the nth sample.

[0071] If the average loss is lower than 0.05 or no longer improves significantly within 10 consecutive rounds, training will be automatically stopped to prevent overfitting.

[0072] Step 11: Use the test set D32 to test the trained GoogLeNet model, evaluate the model's prediction performance, and quantify its performance by calculating the accuracy index. The accuracy index is calculated using the determination coefficient R 2 , which is calculated as shown in formula (6), In the formula, n is the number of samples in the test set, y i is the true value of the internal friction angle of the test set sample, is the predicted value of the internal friction angle of the sample, is the average value of the true value of the internal friction angle of the sample.

[0073] Step 12: Perform error analysis based on the test results in step 11 to evaluate the deviation between the predicted value and the actual value. When the prediction error exceeds expectations, make improvements by optimizing hyperparameters, expanding the data set, and adjusting the model structure. Repeat steps 10 and 11 until the model performance reaches the expected standard.

[0074] When evaluating, when the accuracy index determines the coefficient R 2 When the value of is greater than or equal to 0.90, the model performance is considered to have reached the expected standard and no further optimization is required; 2 If it is less than 0.90, then make improvements by adjusting hyperparameters, expanding the data set, or optimizing the model structure, and repeat steps 10 and 11 until the accuracy index determination coefficient R 2 The value of is greater than or equal to 0.90.

[0075] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting the shear strength of structural surfaces with different roughness in a three-dimensional stress state, characterized in that: The following steps are involved: S1: Preparation of standard rock sample Sam1: Use a rock coring machine, a rock cutting machine and a polishing machine to process the collected rock blocks to prepare a standard cylindrical rock sample Sam1; S2: Prepare silicone rubber sample Sam2 and conduct tests: Use silicone rubber to cast a cylindrical silicone rubber sample Sam2 with a thickness of h and a diameter the same as that of the standard rock sample Sam1. After curing, perform stress-strain curve tests on the silicone rubber sample Sam2 under various confining pressure conditions; S3: Preparation of split rock sample Sam4: Use a universal testing machine and a splitting fixture to split the standard rock sample Sam1 along the height direction, so that it is broken along different splitting lines into two left and right rock samples Sam3 with different roughness structural surfaces, and cut off the rock block with a thickness of h at the upper end or the lower end of the rock sample Sam3 to form a split rock sample Sam4; S4: Constructing a three-dimensional morphological model D1: Scanning the two side walls of the structure surface of the split rock sample Sam4 with a morphological scanner to generate point cloud data, and encrypting the point cloud data with the nearest neighbor interpolation method to construct a three-dimensional morphological model D1; S5: Export the two-dimensional image file D2 of the structural surface: adjust the viewing angle and range, apply surface rendering technology to visualize the roughness of the structural surface, express the roughness through color gradient, and export the two-dimensional image file D2 of the structural surface; S6: preparing the test sample Sam5: wrapping the outer side of the cut end of the split rock sample Sam4 with a plastic hard film so that the plastic hard film is flush with the end surface, and filling the plastic hard film cavity with a silicone rubber with a thickness of h to supplement the removed rock block, thereby forming the test sample Sam5; S7: Perform a three-dimensional stress state shear strength test: After the test sample Sam5 is cured, a three-dimensional stress state shear strength test is performed on the test sample Sam5 using the confining pressure conditions used for the silicone rubber sample Sam2 in step S2, and the load and displacement changes during the shearing process are recorded to draw a shear load-shear displacement curve; S8: Calculate the shear strength τ of structural surfaces with different roughness ts :According to the shear load-shear displacement curve drawn in step S7, the peak shear load is obtained, and according to the peak shear load and the triaxial compression axial load of the silicone rubber sample Sam2, the shear strength τ of the structural surfaces with different roughness is calculated ts , where each shear strength τ ts The calculation formula is expressed as formula (1), τ ts =(F′-F) / S(1), where F′ refers to the force on the lower part of the left rock block and the upper part of the right rock block when the shear stress of the structural surface in the three-dimensional stress state reaches the peak value, F refers to the triaxial compression axial load value of the silicone rubber sample Sam2, and S is the shear action area of ​​the left and right rock blocks; S9: Construct the image and mechanical strength data set D3 of the structural surface: Repeat steps S1 to S5 to obtain multiple split rock samples Sam4 with different rough surfaces, and export the two-dimensional image file D2 corresponding to each split rock sample Sam4; Repeat steps S6 to S8 to obtain the shear strength τ of each test sample Sam5 ts ; The two-dimensional image file D2 of each structural surface and its corresponding shear strength τ ts Correlation, constructing a dataset D3 containing images and mechanical strength; S10: Training the GoogLeNet model: randomly divide the dataset D3 into a training set D31 and a test set D32, and use the training set D31 to train the computational model. The GoogLeNet deep convolutional neural network is used as the basic model, and the hyperparameters are automatically tuned through the Bayesian optimization algorithm to obtain the optimal parameter combination. Perform K-fold cross-validation training on the training set D31, divide the training set into K subsets, randomly select K-1 subsets for training each time, and use the remaining 1 subset for validation. Repeat this process until the average loss reaches the predetermined standard, and the training of the computational model is completed. S11: Test the trained GoogLeNet model using the test set D32 to evaluate the model’s prediction performance and quantify its performance by calculating the accuracy index; S12: Perform error analysis based on the test results in step S11 to evaluate the deviation between the predicted value and the actual value. When the prediction error exceeds expectations, make improvements by optimizing hyperparameters, expanding the data set, and adjusting the model structure. Repeat steps S10 and S11 until the model performance reaches the expected standard.

2. A method for predicting shear strength of structural surfaces with different roughness in three-dimensional stress states according to claim 1, characterized in that: When the morphological discrete point cloud data collected in step S4 contains noise points other than the structural surface, it needs to be eliminated.

3. A method for predicting shear strength of structural surfaces with different roughness in three-dimensional stress states according to claim 1, characterized in that: When the three-dimensional stress state shear strength test is performed on the test sample Sam5 in step S7, the maximum shear displacement does not exceed the thickness h of the silicone rubber.

4. The method for predicting the shear strength of structural surfaces with different roughness in a three-dimensional stress state according to claim 1, characterized in that: In step S8, the triaxial compressive axial load of the silicone rubber sample Sam2 refers to the force value of the corresponding silicone rubber sample Sam2 when the structural surface reaches the peak shear stress under the same confining pressure condition.

5. A method for predicting shear strength of structural surfaces with different roughness in three-dimensional stress states according to claim 4, characterized in that: Since the calculation formula of the structural surface shear stress F0 is formula (2), F+F′=F0(2), where F′ refers to the force exerted on the lower part of the left rock block and the upper part of the right rock block when the structural surface shear stress reaches the peak value in the three-dimensional stress state, and F refers to the triaxial compression axial load value of the silicone rubber sample Sam2. Substitute formula (2) into formula (1) in step S8, and the shear strength τ ts Further expressed as formula (3), τ ts =(F0-2F) / S(3), where F0 is the structural surface shear stress, F refers to the triaxial compression axial load value of the silicone rubber sample Sam2, and S is the shear action area of ​​the left and right rock blocks.

6. The method for predicting shear strength of structural surfaces with different roughness in three-dimensional stress states according to claim 1, characterized in that: The GoogLeNet deep convolutional neural network is used as the basic model. Its overall architecture includes input layer, convolution layer, maximum pooling layer, Inception module, global average pooling layer, fully connected layer and output layer. The Inception module consists of four branches: 1x1 convolution, 3x3 convolution, 5x5 convolution and pooling layer.

7. The method for predicting shear strength of structural surfaces with different roughness in three-dimensional stress states according to claim 1, characterized in that: When the hyperparameters of the GoogLeNet model are optimized by the Bayesian optimization algorithm in step S10, the hyperparameters that need to be optimized include: the convolution kernel size, step size and padding method of the convolution layer, the pooling window size and step size of the pooling layer, and the number of filters of each branch in the Inception module.

8. The method for predicting shear strength of structural surfaces with different roughness in three-dimensional stress states according to claim 1, characterized in that: In step S10, when the hyperparameters of the GoogLeNet model are optimized by the Bayesian optimization algorithm, the Gaussian process is used as the proxy function, the expected improvement EI is used as the acquisition function, and the mean square error MSE of the validation set is used as the objective function. When the mean square error (MSE) of the objective function validation set reaches the minimum value, the corresponding hyperparameter combination is the optimal hyperparameter combination. The calculation formula is shown in formula (4): Where n is the number of samples; y n is the true value corresponding to the nth sample; is the predicted value corresponding to the nth sample, In step S10, the loop is executed until the average loss reaches the predetermined standard, and the root mean square error RMSE is used as the loss function. The calculation formula of the average loss is shown in formula (5): (5), where n is the number of samples; y n is the true value corresponding to the nth sample; is the predicted value corresponding to the nth sample, The "early stopping strategy" is used to dynamically monitor the average loss of the validation set. When the model does not achieve performance improvement in several consecutive rounds, training is automatically stopped to ensure that training ends at the best time.

9. The method for predicting shear strength of structural surfaces with different roughness in three-dimensional stress states according to claim 1, characterized in that: The accuracy index calculated in step S11 is determined by the coefficient of determination R 2 , which is calculated as shown in formula (6), In the formula, n is the number of samples in the test set, y i is the true value of the internal friction angle of the test set sample, is the predicted value of the sample internal friction angle, is the average value of the true value of the internal friction angle of the sample.

10. The method for predicting shear strength of structural surfaces with different roughness in three-dimensional stress states according to claim 1, characterized in that: In the evaluation of step S12, when the accuracy index determination coefficient R 2 When the value of is greater than or equal to 0.90, the model performance is considered to have reached the expected standard and no further optimization is required; 2 If it is less than 0.90, then make improvements by adjusting hyperparameters, expanding the data set, or optimizing the model structure, and repeat S10 and S11 until the accuracy index determination coefficient R 2 The value of is greater than or equal to 0.90.

Images