Coarse-grained soil shear strength prediction method based on deep learning

Abstract

The invention discloses a deep learning-based coarse-grained soil shear strength prediction method. The method comprises the following steps of constructing a standardized data set; a Seft-Net prediction model is constructed; dividing the standardized data set into a training set and a test set, inputting the training set into a Seft-Net prediction model for training to obtain a trained Seft-Net prediction model, and inputting the test set into the trained Seft-Net prediction model for testing to obtain a tested Seft-Net prediction model; key parameters of coarse-grained soil to be predicted are input into the tested Seft-Net prediction model, and a shear strength peak value qpeak is output; and according to the shear strength peak value qpeak, obtaining the shear strength indexes phi 0 and delta phi of the coarse-grained soil. According to the method, the problems of low learning efficiency and low prediction accuracy of an existing deep learning method are solved.

Description

Technical Field

[0001] This invention belongs to the technical field of geotechnical mechanics parameter prediction methods, specifically relating to a deep learning-based method for predicting the shear strength of coarse-grained soil. Background Technology

[0002] Coarse-grained soils, due to their advantages such as good compaction, high permeability, and easy availability, are widely used in large geotechnical structures such as earth-rock dams, slopes, revetments, and roadbeds. Their mechanical properties directly affect the safety and stability of these projects. The shear strength of coarse-grained soils is a key parameter characterizing their mechanical behavior, directly related to their maximum ability to resist shear failure under shear loads. It is crucial for assessing the stability and bearing capacity of coarse-grained soil-related projects and is an indispensable strength parameter in engineering design and construction decisions. Traditionally, the shear strength of coarse-grained soils has been determined primarily through laboratory testing methods such as large-scale triaxial tests and direct shear tests. While these methods provide accurate test results, the time-consuming and costly process limits their feasibility in rapid assessment and large-scale engineering applications.

[0003] With the development of data science and artificial intelligence technologies, machine learning methods have shown great application potential in the civil engineering field, solving the problems of high cost and long time consumption of traditional experimental methods. Some studies have attempted to apply traditional machine learning models, including shallow machine learning techniques such as Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Networks, to predict mechanical parameters, enabling modeling and prediction of small-scale datasets. However, when dealing with data with complex nonlinear coupling relationships between parameters, shallow machine learning methods usually require complex feature engineering to preprocess the input parameters and struggle to deeply mine the complex patterns hidden in the data. In contrast, deep learning can automatically extract features, has powerful nonlinear mapping capabilities, can effectively capture nonlinear relationships in data, and has significant advantages when processing large-scale data. However, directly applying general deep learning models to predict the shear strength of coarse-grained soils still faces challenges: the physical meaning and correlation of input parameters (such as particle size, void ratio, confining pressure, etc.) differ, and the model structure may not effectively capture the intrinsic relationship between the soil material's unique physical mechanisms and strength parameters, resulting in low prediction accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for predicting the shear strength of coarse-grained soil, thereby addressing the problem of low prediction accuracy in existing deep learning methods.

[0005] The technical solution adopted in this invention is a deep learning-based method for predicting the shear strength of coarse-grained soil, and the specific steps are as follows: Step 1: Construct a standardized dataset; Step 2, construct the Seet-Net prediction model; Step 3: Divide the standardized dataset into a training set and a test set. Input the training set into the Seft-Net prediction model for training to obtain the trained Seft-Net prediction model. Input the test set into the trained Seft-Net prediction model for testing to obtain the tested Seft-Net prediction model. Step 4: Input the key parameters of the coarse-grained soil to be predicted into the Seft-Net prediction model after testing, and output the peak shear strength. q peak ; Step 5, based on the peak shear strength q peak The shear strength index of coarse-grained soil was obtained. f 0 and Δ f .

[0006] The invention is further characterized by: In step 1, each coarse-grained soil sample in the standardized dataset contains 13 input features and 1 output feature; Input features include the uniaxial compressive strength of the parent rock. s r Regularity of particle shape r Soil particle density G s Initial porosity e 0. Coefficient of non-uniformity C u curvature coefficient C c Confining pressure s 3. Typical particle size D max , D i ,in, D max For the maximum particle size, D Indicates particle size, i This represents the cumulative percentage of soil particles smaller than particle size D, where i = 10, 30, 50, 60, 80. Among them, the uniaxial compressive strength of the parent rock s r Regularity of particle shape r Soil particle density G s These are particle property parameters; initial void ratio. e 0. Typical particle size D max , D i Inhomogeneity coefficient C ucurvature coefficient C c For soil properties; confining pressure s 3 is an external factor; The output characteristic is the peak shear strength. q peak .

[0007] In step 2, the Seet-Net prediction model consists of an input module, a feature transformation module, a feature extraction module, a regression head module, and an output module.

[0008] The specific processing procedure of the input module is as follows: the 13 input features of each coarse soil sample are combined into a 1×13 one-dimensional array, and then the resulting one-dimensional array is converted into a 3×13 two-dimensional matrix. The input features form a 1×13 one-dimensional array in the following order: particle property parameters, parent rock uniaxial compressive strength. s r Regularity of particle shape r Soil particle density G s Soil property parameters: initial void ratio e 0. Typical particle size D max , D 80 , D 60 , D 50 , D 30 , D 10 Inhomogeneity coefficient C u curvature coefficient C c External factors confining pressure s 3.

[0009] The specific processing procedure of the feature transformation module is as follows: the two-dimensional matrix is ​​divided into three groups according to particle attribute parameters, soil attribute parameters, and external factors. The group of particle attribute parameters is processed by the first Shortcut to obtain result A. The group of soil attribute parameters is input into the convolutional layer for processing to obtain result B. The group of external factors is processed by the second Shortcut to obtain result C. Result A, result B, and result C are concatenated to obtain result D.

[0010] The feature extraction module consists of an inlet stream, eight intermediate streams, and an outlet stream; The ingress stream consists of a first convolutional layer, a first ReLU activation function, a second convolutional layer, a first depthwise separable convolutional module, a second depthwise separable convolutional module, and a third depthwise separable convolutional module, with the result D serving as the input to the first convolutional layer. The first depthwise separable convolutional module consists of a second ReLU activation function, a first depthwise convolutional layer, a third ReLU activation function, a second depthwise convolutional layer, a first max pooling layer, and a first pointwise convolutional layer; the output of the second convolutional layer serves as the input to the second ReLU activation function and the first pointwise convolutional layer, and the output of the max pooling layer is concatted with the output of the first pointwise convolutional layer. The second depthwise separable convolutional module consists of a fourth ReLU activation function, a third depthwise convolutional layer, a fifth ReLU activation function, a fourth depthwise convolutional layer, a second max pooling layer, and a second pointwise convolutional layer; the result of the first concat operation is used as the input to the fourth ReLU activation function and the second pointwise convolutional layer, and the output of the second max pooling layer and the output of the second pointwise convolutional layer are subjected to a second concat operation; The third depthwise separable convolutional module consists of a sixth ReLU activation function, a fifth depthwise convolutional layer, a seventh ReLU activation function, a sixth depthwise convolutional layer, a third max pooling layer, and a third pointwise convolutional layer; the result of the second concat operation is used as the input to the sixth ReLU activation function and the third pointwise convolutional layer, and the output of the third max pooling layer is concatted with the output of the third pointwise convolutional layer. Each intermediate stream consists of a seventh deep convolutional layer, an eighth deep convolutional layer, a ninth deep convolutional layer, and a fourth concat operation. For the first intermediate stream, the result of the third concat operation is used as the input of the seventh deep convolutional layer, and the output of the ninth deep convolutional layer is combined with the result of the third concat operation to perform the fourth concat operation. For the second to eighth intermediate streams, the result of the fourth concat operation of the previous intermediate stream is used as the input of the seventh deep convolutional layer of the current intermediate stream, and the output of the ninth deep convolutional layer of the current intermediate stream is combined with the result of the fourth concat operation of the previous intermediate stream to perform the fourth concat operation of the current intermediate stream. The outflow consists of the tenth deep convolutional layer, the eleventh deep convolutional layer, the fourth max pooling layer, the fifth concat operation, the twelfth deep convolutional layer, the thirteenth deep convolutional layer, and the global average pooling layer. The result of the fourth concat operation of the eighth intermediate flow is combined with the output result of the fourth max pooling layer for the fifth concat operation.

[0011] The regression head module consists of a first fully connected layer, an eighth ReLU activation function, and a second fully connected layer. The output of the feature extraction module serves as the input to the first fully connected layer.

[0012] In step 3, mean absolute error and mean squared error are used as loss functions during the training process.

[0013] The specific process of step 5 is as follows: Step 5.1, based on each peak shear strength predicted in Step 4 q peak Draw the Mohr's circle and strength envelope for each confining pressure, and obtain the corresponding internal friction angle for each confining pressure based on the strength envelope. f ; Step 5.2, since the strength envelope of coarse-grained soil is non-linear, it is expressed as: (1) In the formula, f Indicates the angle of internal friction. p a Indicates atmospheric pressure. s 3 represents confining pressure. f 0 and △ f These represent the intercept and slope of the approximate straight line in logarithmic coordinates when the internal friction angle varies with confining pressure; Step 5.2, based on the internal friction angle obtained in Step 5.1 f ,draw f and s 3 / p a The relationship diagram is used to obtain the shear strength index of coarse-grained soil. f 0 and Δ f .

[0014] The beneficial effects of this invention are as follows: This invention provides a deep learning-based method for predicting the shear strength of coarse-grained soil. It categorizes coarse-grained soil indices into three types: particle properties, soil properties, and external factors. Feature transformation is performed through a feature transformation module to improve feature transformation efficiency. Furthermore, the improved depthwise separable convolution used in the feature extraction module has a multi-scale receptive field and a multi-branch architecture, aiming to learn and fuse multi-level hierarchical features and abstract high-dimensional features to capture the inherent potential nonlinear mapping relationship of the data, thereby improving prediction accuracy. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the Seft-Net prediction model in the method of the present invention; Figure 2 This is a schematic diagram of the feature extraction module in the method of the present invention; Figure 3 This is the convergence curve of the loss function (MSE) as a function of the training cycle during the training of the Seft-Net prediction model in the method of this invention; Figure 4 This is a scatter plot showing the relationship between actual and predicted values ​​during the training phase of this invention. Figure 5 This is a scatter plot showing the relationship between actual and predicted values ​​during the testing phase of this invention. Figure 6 Material 1 in Embodiment 6 of the present invention f and s 3 / p a Relationship diagram; Figure 7 Material 2 in Embodiment 6 of the present invention f and s 3 / p a Relationship diagram.

[0016] Figure 8 This is a schematic diagram of the feature transformation module, integrated structure, shortcut connection, and convolutional structure in Embodiment 7 of the present invention; Figure 9 The ablation experiment results are for the feature transformation module, integrated structure, shortcut connection, and convolutional structure in Embodiment 7 of the present invention. Figure 10 The ablation experiment results of the feature extraction module, the original Xception module, and the ResNet module in Embodiment 7 of the present invention; Figure 11 This is a sorting diagram of the input feature importance analysis in Embodiment 7 of the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 This invention presents a deep learning-based method for predicting the shear strength of coarse-grained soil, the specific steps of which are as follows: Step 1: Construct a standardized dataset; Step 2, construct the Seet-Net prediction model; Step 3: Divide the standardized dataset into a training set and a test set. Input the training set into the Seft-Net prediction model for training to obtain the trained Seft-Net prediction model. Input the test set into the trained Seft-Net prediction model for testing to obtain the tested Seft-Net prediction model. Step 4: Input the key parameters of the coarse-grained soil to be predicted into the Seft-Net prediction model after testing, and output the peak shear strength. q peak ; Step 5, based on the peak shear strength q peak The shear strength index of coarse-grained soil was obtained. f 0 and Δ f .

[0019] Example 2 Based on Example 1, in step 1, each coarse-grained soil sample in the standardized dataset contains 13 input features and 1 output feature; Input features include the uniaxial compressive strength of the parent rock. s r Regularity of particle shape r Soil particle density G s Initial porosity e 0. Coefficient of non-uniformity C u Curvature coefficient C c Confining pressure s 3. Typical particle size D max , D i ,in, D max For the maximum particle size, D Indicates particle size, i This represents the cumulative percentage of soil particles smaller than particle size D, where i = 10, 30, 50, 60, 80; confining pressure s There should be at least 3 data points; Among them, the uniaxial compressive strength of the parent rock s r Regularity of particle shape r Soil particle density G s These are particle property parameters; initial void ratio. e 0. Typical particle size D max , D i Inhomogeneity coefficient C u Curvature coefficient C c For soil properties; confining pressure s 3 is an external factor; The output characteristic is the peak shear strength. q peak The number of output features is related to the confining pressure. s 3. Corresponding number of data points.

[0020] The data in the standardized dataset was obtained through 91 research projects, including 63 academic papers, 22 engineering reports, and 6 indoor experiments, covering 1,108 sets of test results for 334 different coarse-grained soil materials, and 54,294 data points were collected manually.

[0021] Example 3 Based on Example 2, in step 2, as follows: Figure 1 As shown, the Seet-Net prediction model consists of an input module, a feature transformation module, a feature extraction module, a regression head module, and an output module.

[0022] The specific processing procedure of the input module is as follows: the 13 input features of each coarse soil sample are combined into a 1×13 one-dimensional array, and then the resulting one-dimensional array is converted into a 3×13 two-dimensional matrix; The input features form a 1×13 one-dimensional array in the following order: particle property parameters, parent rock uniaxial compressive strength. s r Regularity of particle shape r Soil particle density G s Soil property parameters: initial void ratio e 0. Typical particle size D max , D 80 , D 60 , D 50 , D 30 , D 10 Inhomogeneity coefficient C u Curvature coefficient C c External factors confining pressure s 3.

[0023] The specific processing procedure of the feature transformation module is as follows: the two-dimensional matrix is ​​divided into three groups according to particle property parameters, soil property parameters, and external factors. The group of particle property parameters is processed by the first Shortcut to obtain result A. The group of soil property parameters is input into the convolutional layer for processing to obtain result B. The group of external factors is processed by the second Shortcut to obtain result C. Result A, result B, and result C are concatenated to obtain result D. The inherent physical laws of different groups of parameters are considered to reduce inter-group interference, and in particular, it can effectively alleviate the multicollinearity problem between parameters. like Figure 2 As shown, the feature extraction module consists of an inlet stream, eight intermediate streams, and an outlet stream; The ingress stream consists of a first convolutional layer, a first ReLU activation function, a second convolutional layer, a first depthwise separable convolutional module, a second depthwise separable convolutional module, and a third depthwise separable convolutional module. The result D is used as the input of the first convolutional layer, the output of the first convolutional layer is used as the input of the first ReLU activation function, and the output of the first ReLU activation function is used as the input of the second convolutional layer.

[0024] The first convolutional layer has a kernel size of 3×3 and 32 output channels, while the second convolutional layer has a kernel size of 3×3 and 64 output channels.

[0025] The first depthwise separable convolutional module consists of a second ReLU activation function, a first depthwise convolutional layer, a third ReLU activation function, a second depthwise convolutional layer, a first max pooling layer, and a first pointwise convolutional layer. The output of the second convolutional layer serves as the input to the second ReLU activation function and the first pointwise convolutional layer. The output of the second ReLU activation function serves as the input to the first depthwise convolutional layer. The output of the first depthwise convolutional layer serves as the input to the third ReLU activation function. The output of the third ReLU activation function serves as the input to the second depthwise convolutional layer. The output of the second depthwise convolutional layer serves as the input to the max pooling layer. The output of the max pooling layer and the output of the first pointwise convolutional layer undergo a first concat operation. The kernel size of the first depthwise convolutional layer is 3×3, and the number of output channels is 128. The kernel size of the second depthwise convolutional layer is 3×3, and the number of output channels is 128. The kernel size of the first max pooling layer is 3×3, and the kernel size of the first pointwise convolutional layer is 1×1.

[0026] The second depthwise separable convolutional module consists of a fourth ReLU activation function, a third depthwise convolutional layer, a fifth ReLU activation function, a fourth depthwise convolutional layer, a second max-pooling layer, and a second pointwise convolutional layer. The result of the first concat operation serves as the input to the fourth ReLU activation function and the second pointwise convolutional layer. The output of the fourth ReLU activation function serves as the input to the third depthwise convolutional layer. The output of the third depthwise convolutional layer serves as the input to the fifth ReLU activation function. The output of the fifth ReLU activation function serves as the input to the fourth depthwise convolutional layer. The output of the fourth depthwise convolutional layer serves as the input to the second max-pooling layer. The output of the second max-pooling layer and the output of the second pointwise convolutional layer undergo a second concat operation. The kernel size of the third depthwise convolutional layer is 3×3, and the number of output channels is 256. The kernel size of the fourth depthwise convolutional layer is 3×3, and the number of output channels is 256. The kernel size of the second max-pooling layer is 3×3, and the kernel size of the second pointwise convolutional layer is 1×1.

[0027] The third depthwise separable convolutional module consists of a sixth ReLU activation function, a fifth depthwise convolutional layer, a seventh ReLU activation function, a sixth depthwise convolutional layer, a third max pooling layer, and a third pointwise convolutional layer. The result of the second concat operation is used as the input to the sixth ReLU activation function and the third pointwise convolutional layer. The output of the sixth ReLU activation function is used as the input to the fifth depthwise convolutional layer. The output of the fifth depthwise convolutional layer is used as the input to the seventh ReLU activation function. The output of the seventh ReLU activation function is used as the input to the sixth depthwise convolutional layer. The output of the sixth depthwise convolutional layer is used as the input to the third max pooling layer. The output of the third max pooling layer and the output of the third pointwise convolutional layer are then concatted together.

[0028] The kernel size of the fifth deep convolutional layer is 3×3, and the number of output channels is 728. The kernel size of the sixth deep convolutional layer is 3×3, and the number of output channels is 728. The kernel size of the second max pooling layer is 3×3, and the kernel size of the third pointwise convolutional layer is 1×1.

[0029] Each intermediate stream consists of a seventh, eighth, and ninth deep convolutional layer, followed by a fourth concat operation. The output of the seventh deep convolutional layer serves as the input to the eighth deep convolutional layer, and the output of the eighth deep convolutional layer serves as the input to the ninth deep convolutional layer. For the first intermediate stream, the result of the third concat operation serves as the input to the seventh deep convolutional layer, and the output of the ninth deep convolutional layer is combined with the result of the third concat operation to perform a fourth concat operation. For the second to eighth intermediate streams, the result of the fourth concat operation of the previous intermediate stream serves as the input to the seventh deep convolutional layer of the current intermediate stream, and the output of the ninth deep convolutional layer of the current intermediate stream is combined with the result of the fourth concat operation of the previous intermediate stream to perform a fourth concat operation of the current intermediate stream.

[0030] The kernel size of the seventh, eighth, and ninth deep convolutional layers is 3×3, and the number of output channels is 728.

[0031] The outflow consists of the tenth deep convolutional layer, the eleventh deep convolutional layer, the fourth max pooling layer, the fifth concat operation, the twelfth deep convolutional layer, the thirteenth deep convolutional layer, the global average pooling layer, and the fourth pointwise convolutional layer. The result of the fourth concat operation of the eighth intermediate flow is used as the input to the fourth pointwise convolutional layer and the tenth deep convolutional layer. The output of the fourth pointwise convolutional layer is concatted with the output of the fourth max pooling layer. The output of the tenth deep convolutional layer is used as the input to the eleventh deep convolutional layer. The output of the eleventh deep convolutional layer is used as the input to the fourth max pooling layer. The result of the fifth concat operation is used as the input to the twelfth deep convolutional layer. The output of the twelfth deep convolutional layer is used as the input to the thirteenth deep convolutional layer. The output of the thirteenth deep convolutional layer is used as the input to the global average pooling layer. The output of the global average pooling layer is used as the input to the regression head module.

[0032] The kernel size of the 10th deep convolutional layer is 3×3, and the number of output channels is 728. The kernel size of the 11th deep convolutional layer is 3×3, and the number of output channels is 1024. The kernel size of the 4th max pooling layer is 3×3. The kernel size of the 12th deep convolutional layer is 3×3, and the number of output channels is 1536. The kernel size of the 13th deep convolutional layer is 3×3, and the number of output channels is 2048.

[0033] In the depthwise convolution stage of the feature extraction module, each channel of the input feature map is independently convolved. Subsequently, the result of the depthwise convolution is convolved point-by-point using a 1×1 convolution kernel to achieve linear combination between channels. This process significantly reduces the number of model parameters and computational complexity, improves the model's computational efficiency and inference speed, and effectively maintains the model's feature representation ability.

[0034] depthwise convolution D-Conv It can be expressed as the following formula:

[0035] Pointwise convolution P-Conv It can be expressed as the following formula:

[0036] In the formula, W It is the weight matrix of the convolution kernel; x It is the input of the convolutional layer, ( i, j () represents the coordinates of the output feature map; m , n and k The three dimensions of the convolution kernel are height, width, and the channel dimension of the input feature map. W p The weight matrix of the point convolution kernel. Wd This is the weight matrix for the depthwise convolution; Depthwise separable convolution DP-Conv The entire process can be represented by the following formula: .

[0037] The regression head module consists of a first fully connected layer, an eighth ReLU activation function, and a second fully connected layer. The output of the feature extraction module serves as the input to the first fully connected layer.

[0038] After the feature extraction module extracts high-dimensional features of 13 coarse-grained soil properties, these feature maps are passed to two fully connected layers serving as the regression head module. The purpose of this transformation is to reduce the dimensionality of the features extracted from the feature extraction module while preserving the key information about the coarse-grained soil properties they contain. Through this dimensionality reduction mapping, the fully connected layers transform the learned high-dimensional features into peak shear strength. q peak The final prediction.

[0039] The output module is used to output the output of the second fully connected layer of the regression head module.

[0040] Example 4 Based on Example 3, in step 3, mean absolute error (MAE) and mean squared error (MSE) are used as loss functions during training, and mean absolute percentage error (MAPE) and root mean square error (RMSE) are used as evaluation indicators for comparison results. The parameters of the Seft-Net prediction model are adjusted through optimization techniques such as stochastic gradient descent (SGD) and backpropagation to improve the accuracy of regression prediction.

[0041] During training, the initial learning rate is set to 0.001, and an exponential decay strategy is adopted (e.g., multiply by 0.98 every 10 epochs); the weight decay coefficient is set to 0.99; the batch size can be set to 1 or a smaller value depending on the computing resources; and the training epochs are set to around 3000.

[0042] The ratio of training set to test set is 5:1.

[0043] The prediction performance of the Seft-Net prediction model on the test set is shown in Table 1. Table 1

[0044] The results are consistent with the general trend of deep learning models, namely that the error observed during the testing phase is higher than that observed during the training phase. Figure 3 This demonstrates the stable convergence of the loss function during training.

[0045] like Figure 4 and Figure 5 As shown, the proximity of a point to the diagonal represents the accuracy of the prediction. The results show that, except for a few outliers, most points are closely distributed around the diagonal, indicating that the model achieves high prediction accuracy on both the training and test sets. This result not only demonstrates the ability of the Seft-Net prediction model to capture the complex nonlinear relationship between coarse-grained soil physical properties and peak intensity, but also showcases its potential in practical applications.

[0046] Example 5 Based on Example 4, the specific process of step 5 is as follows: Step 5.1, based on each peak shear strength predicted in Step 4 q peak Draw the Mohr's circle and strength envelope for each confining pressure, and obtain the corresponding internal friction angle for each confining pressure based on the strength envelope. f ; Step 5.2, since the strength envelope of coarse-grained soil is non-linear, it is expressed as: (1) In the formula, f Indicates the angle of internal friction. p a Indicates atmospheric pressure. s 3 represents confining pressure. f 0 and △ f These represent the intercept and slope of the approximate straight line in logarithmic coordinates when the internal friction angle varies with confining pressure; Step 5.2, based on the internal friction angle obtained in Step 5.1 f ,draw f and s 3 / p a The relationship diagram is used to obtain the shear strength index of coarse-grained soil. f 0 and Δ f .

[0047] Example 6 Step 1: Construct a standardized dataset; In the standardized dataset, each coarse-grained soil sample contains 13 input features and 1 output feature; Input features include the uniaxial compressive strength of the parent rock. s r Regularity of particle shape r Soil particle density G s Initial porosity e 0. Coefficient of non-uniformity C u curvature coefficient C c Confining pressure s 3. Typical particle size D max , D i ,in, D max For the maximum particle size, D Indicates particle size, i This represents the cumulative percentage of soil particles smaller than particle size D, where i = 10, 30, 50, 60, 80. Uniaxial compressive strength of parent rock s r Regularity of particle shape r Soil particle density G s These are particle property parameters; initial void ratio. e 0. Typical particle size D max , D i Inhomogeneity coefficient C u curvature coefficient C c For soil properties; confining pressure s 3 is an external factor; The output characteristic is the peak shear strength. q peak ; Step 2, construct the Seet-Net prediction model; The Seft-Net prediction model consists of an input module, a feature transformation module, a feature extraction module, a regression head module, and an output module. The specific processing procedure of the input module is as follows: the 13 input features of each coarse soil sample are combined into a 1×13 one-dimensional array, and then the resulting one-dimensional array is converted into a 3×13 two-dimensional matrix. The input features form a 1×13 one-dimensional array in the following order: particle property parameters, parent rock uniaxial compressive strength. s r Regularity of particle shape r Soil particle density G s Soil property parameters: initial void ratio e 0. Typical particle size D max , D 80 , D 60 , D 50 , D 30 , D 10 Inhomogeneity coefficientC u curvature coefficient C c External factors confining pressure s 3; The specific processing procedure of the feature transformation module is as follows: the two-dimensional matrix is ​​divided into three groups according to particle attribute parameters, soil attribute parameters, and external factors. The group of particle attribute parameters is processed by the first Shortcut to obtain result A. The group of soil attribute parameters is input into the convolutional layer for processing to obtain result B. The group of external factors is processed by the second Shortcut to obtain result C. Result A, result B, and result C are concatenated to obtain result D. The feature extraction module consists of an inlet stream, eight intermediate streams, and an outlet stream; The ingress stream consists of a first convolutional layer, a first ReLU activation function, a second convolutional layer, a first depthwise separable convolutional module, a second depthwise separable convolutional module, and a third depthwise separable convolutional module, with the result D serving as the input to the first convolutional layer. The first depthwise separable convolutional module consists of a second ReLU activation function, a first depthwise convolutional layer, a third ReLU activation function, a second depthwise convolutional layer, a first max pooling layer, and a first pointwise convolutional layer; the output of the second convolutional layer serves as the input to the second ReLU activation function and the first pointwise convolutional layer, and the output of the max pooling layer is concatted with the output of the first pointwise convolutional layer. The second depthwise separable convolutional module consists of a fourth ReLU activation function, a third depthwise convolutional layer, a fifth ReLU activation function, a fourth depthwise convolutional layer, a second max pooling layer, and a second pointwise convolutional layer; the result of the first concat operation is used as the input to the fourth ReLU activation function and the second pointwise convolutional layer, and the output of the second max pooling layer and the output of the second pointwise convolutional layer are subjected to a second concat operation; The third depthwise separable convolutional module consists of a sixth ReLU activation function, a fifth depthwise convolutional layer, a seventh ReLU activation function, a sixth depthwise convolutional layer, a third max pooling layer, and a third pointwise convolutional layer; the result of the second concat operation is used as the input to the sixth ReLU activation function and the third pointwise convolutional layer, and the output of the third max pooling layer is concatted with the output of the third pointwise convolutional layer. Each intermediate stream consists of a seventh deep convolutional layer, an eighth deep convolutional layer, a ninth deep convolutional layer, and a fourth concat operation; the output of the ninth deep convolutional layer is combined with the result of the third concat operation to perform the fourth concat operation; for the second to eighth intermediate streams, the result of the fourth concat operation of the previous intermediate stream is used as the input of the seventh deep convolutional layer of the current intermediate stream, and the output of the ninth deep convolutional layer of the current intermediate stream is combined with the result of the fourth concat operation of the previous intermediate stream to perform the fourth concat operation of the current intermediate stream. The outflow consists of the tenth deep convolutional layer, the eleventh deep convolutional layer, the fourth max pooling layer, the fifth concat operation, the twelfth deep convolutional layer, the thirteenth deep convolutional layer, and the global average pooling layer. The result of the fourth concat operation of the eighth intermediate flow is combined with the output result of the fourth max pooling layer to perform the fifth concat operation. The regression head module consists of a first fully connected layer, an eighth ReLU activation function, and a second fully connected layer. The output of the feature extraction module serves as the input to the first fully connected layer. Step 3: Divide the standardized dataset into a training set and a test set (in a ratio of 5:1). Input the training set into the Seft-Net prediction model for training to obtain the trained Seft-Net prediction model. Input the test set into the trained Seft-Net prediction model for testing to obtain the tested Seft-Net prediction model. During training, mean absolute error and root mean square error are used as loss functions; mean absolute percentage error (MAPE) and root mean square error (RMSE) are used as evaluation metrics for comparison results; the parameters of the Seft-Net prediction model are adjusted using optimization techniques such as stochastic gradient descent (SGD) and backpropagation to improve the accuracy of regression prediction.

[0048] During training, the initial learning rate is set to 0.001, and an exponential decay strategy is adopted (e.g., multiply by 0.98 every 10 epochs); the weight decay coefficient is set to 0.99; the batch size can be set to 1 or a smaller value depending on the computing resources; and the training epochs are set to around 3000.

[0049] Step 4: Input the key parameters of the coarse-grained soil to be predicted into the Seft-Net prediction model after testing, and output the peak shear strength. q peak ; The key parameter data of the coarse-grained soil to be predicted are shown in Table 2. Table 2

[0050] Peak shear strength output qpeak The data is shown in Table 3; Table 3

[0051] Step 5, based on the peak shear strength q peak The shear strength index of coarse-grained soil was obtained. f 0 and Δ f ; The specific process is as follows: Step 5.1, based on each peak shear strength predicted in Step 4 q peak Draw the Mohr's circle and strength envelope for each confining pressure, and obtain the corresponding internal friction angle for each confining pressure based on the strength envelope. f ; Step 5.2, since the strength envelope of coarse-grained soil is non-linear, it is expressed as: (1) In the formula, f Indicates the angle of internal friction. p a Indicates atmospheric pressure. s 3 represents confining pressure. f 0 and △ f These represent the intercept and slope of the approximate straight line in logarithmic coordinates when the internal friction angle varies with confining pressure; Step 5.2, based on the internal friction angle obtained in Step 5.1 f , drawing as Figure 6 and Figure 7 shown f and s 3 / p a The relationship diagram is used to obtain the shear strength index of coarse-grained soil. f 0 and Δ f , f 0 and Δ f The values ​​are shown in Table 4; Table 4

[0052] Example 7 This invention also includes ablation experiments, whereby the specific impact of adding, deleting, or modifying individual components of the model was evaluated on its performance. The aim of this method is to identify the frameworks or features that contribute most to the model's accuracy and to reveal potential areas for improvement.

[0053] (1) Analysis of the impact of input feature order; In the ablation experiment, the impact of input feature order on the performance of the Seft-Net prediction model was explored. Given that there are multiple possible permutations and combinations of the 13 input features, this invention selected five representative permutations based on the pre-training results for experimentation and in-depth analysis. The input order is shown in Table 5: Table 5

[0054] The impact of the input feature order on the experimental results is shown in Table 6. Table 6

[0055] The results in Table 6 show that the predictive performance of the Seft-Net prediction model is optimized when variables with similar attributes are grouped together, as in experiment number 1. Further analysis indicates that changing the order of variables within a single category, as shown in experiment number 2, has a limited impact on the overall model performance. However, when the order of variables from different categories is swapped, as in items 3 to 5, the model performance is significantly affected. In this final test, the input variables were predicted using the order shown in experiment number 1.

[0056] (2) Verify the effectiveness of the feature transformation module: Compare the feature transformation module, integrated structure, shortcut connection, and convolutional structure of this invention, among which, for example... Figure 8 As shown, the integrated structure specifically involves dividing the input features into three groups based on particle attribute parameters, soil attribute parameters, and external factors. These three groups are then directly input into the feature extraction module. The shortcut connection specifically involves dividing the input features into three groups based on particle attribute parameters, soil attribute parameters, and external factors. Each group undergoes shortcut processing, and the three processing results are concatenated before being input into the feature extraction module. The convolutional structure specifically involves dividing the input features into three groups based on particle attribute parameters, soil attribute parameters, and external factors. Each group undergoes a convolution operation, and the three processing results are concatenated before being input into the feature extraction module. For example... Figure 9 As shown, the results indicate that the ensemble structure (i.e., feature extraction directly without variable grouping) suffers from performance degradation because the fully connected layers struggle to converge effectively due to the lack of input variable grouping. In contrast, the shortcut connection structure significantly improves the model's accuracy. Furthermore, the convolutional structure exhibits similar performance. These findings prompted further exploration, and by fusing shortcut connections and convolutional structures, the feature transformation module of this invention ultimately achieves optimal performance.

[0057] (3) Verify the performance of the feature extraction module: such as Figure 10As shown, the feature extraction module of this invention is compared with the original Xception module and the ResNet module. The ResNet module improves its feature learning ability by increasing the network depth, while Xception and the feature extraction module of this invention adopt different strategies. They effectively capture multi-scale features of the data by expanding the depth and width of the network. The Xception and the feature extraction module of this invention utilize depthwise separable convolution technology and multi-scale fusion convolution operations, which not only improves the efficiency of parameter use but also enhances the model's learning of feature richness. Experimental results show that the Xception and the feature extraction module of this invention significantly outperform the ResNet module, thanks to their advantages in multi-scale feature extraction. The Conception module retains more information and achieves a certain performance improvement by concatenating the feature maps of residuals and shortcut connections, rather than simply adding them.

[0058] (4) Input Feature Importance Analysis: To resolve the conflict between model performance and interpretability, a perturbation-based approach can be used to evaluate feature importance. This method assesses feature importance by masking or replacing specific features in the samples and comparing changes in model performance. A significant performance drop usually indicates that the altered features are crucial to the model's predictions.

[0059] The model is first trained on the training set with selected variables masked, and then used to predict the same masked variables on the test set. Performance metrics before and after masking are compared. MAE and MAPE This allows for the assessment of the importance of each variable. Feature masking can be used to analyze the importance of each input variable. For example... Figure 11 As shown, the Seft-Net model utilizes 13 input features and determines the confining pressure through comparison and ranking. s 3 is the factor affecting the peak shear strength. q peak Predicting the most important variable. Shielding confining pressure. s 3 leads to MAPE It increased significantly to 69.552%, and other variables such as s r , e 0、 D 80 , C u , r , D 30 , D 10 , C c , D max , Gs , D 60 and D 50 The importance of each parameter decreases sequentially. This result aligns with engineering experience, enhancing the model's reliability. The predicted value can be directly used for dam slope stability calculations, dam stress-deformation analysis, etc., providing key parameters for design schemes and significantly shortening the design cycle.

[0060] Example 8 The Seft-Net prediction model of this invention is compared with fully connected neural networks (FCN), random forests (RF), and support vector machines (SVM). The comparison results are shown in Tables 7 and 8.

[0061] Table 7

[0062] As shown in Table 7, the Seft-Net prediction model exhibited superior prediction performance during the training phase. Despite the low coefficient of determination (... R 2 The results are similar, but the Seet-Net prediction model performs better on other error metrics, including mean absolute error (MAE). MAE ), root mean square error ( RMSE ) and mean absolute percentage error ( MAPE The error is significantly lower than that of FCN, RF, and SVM. This indicates that the Seft-Net prediction model has a higher fit and prediction accuracy on the training data.

[0063] Table 8

[0064] As shown in Table 8, during the testing phase, the Seft-Net prediction model also demonstrated the best prediction performance. R 2 The value is the highest, and in MAE , RMSE and MAPE In terms of error metrics, the Seft-Net prediction model outperforms traditional machine learning models by more than 30%. This result further confirms the Seft-Net prediction model's ability to generalize to unknown data.

Claims

1. A method for predicting the shear strength of coarse-grained soil based on deep learning, characterized in that, The specific steps are as follows: Step 1: Construct a standardized dataset; Step 2, construct the Seet-Net prediction model; Step 3: Divide the standardized dataset into a training set and a test set. Input the training set into the Seft-Net prediction model for training to obtain the trained Seft-Net prediction model. Input the test set into the trained Seft-Net prediction model for testing to obtain the tested Seft-Net prediction model. Step 4: Input the key parameters of the coarse-grained soil to be predicted into the Seft-Net prediction model after testing, and output the peak shear strength. q peak ; Step 5, based on the peak shear strength q peak The shear strength index of coarse-grained soil was obtained. φ 0 and Δ φ .

2. The method for predicting the shear strength of coarse-grained soil based on deep learning according to claim 1, characterized in that, In step 1, each coarse-grained soil sample in the standardized dataset contains 13 input features and 1 output feature; Input features include the uniaxial compressive strength of the parent rock. σ r Regularity of particle shape ρ Soil particle density G s Initial porosity e 0. Coefficient of non-uniformity C u Curvature coefficient C c Confining pressure σ 3. Typical particle size D max , D i ,in, D max For the maximum particle size, D Indicates particle size, i This represents the cumulative percentage of soil particles smaller than particle size D, where i = 10, 30, 50, 60, 80. Among them, the uniaxial compressive strength of the parent rock σ r Regularity of particle shape ρ Soil particle density G s These are particle property parameters; initial void ratio. e 0. Typical particle size D max , D i Inhomogeneity coefficient C u Curvature coefficient C c For soil properties; confining pressure σ 3 is an external factor; The output characteristic is the peak shear strength. q peak .

3. The method for predicting the shear strength of coarse-grained soil based on deep learning according to claim 1, characterized in that, In step 2, the Seet-Net prediction model consists of an input module, a feature transformation module, a feature extraction module, a regression head module, and an output module.

4. The method for predicting the shear strength of coarse-grained soil based on deep learning according to claim 3, characterized in that, The specific processing procedure of the input module is as follows: the 13 input features of each coarse soil sample are combined into a 1×13 one-dimensional array, and then the resulting one-dimensional array is converted into a 3×13 two-dimensional matrix; The input features form a 1×13 one-dimensional array in the following order: particle property parameters, parent rock uniaxial compressive strength. σ r Regularity of particle shape ρ Soil particle density G s Soil property parameters: initial void ratio e 0. Typical particle size D max , D 80 , D 60 , D 50 , D 30 , D 10 Inhomogeneity coefficient C u Curvature coefficient C c External factors confining pressure σ 3.

5. The method for predicting the shear strength of coarse-grained soil based on deep learning according to claim 4, characterized in that, The specific processing procedure of the feature transformation module is as follows: the two-dimensional matrix is ​​divided into three groups according to particle attribute parameters, soil attribute parameters, and external factors. The group of particle attribute parameters is processed by the first Shortcut to obtain result A. The group of soil attribute parameters is input into the convolutional layer for processing to obtain result B. The group of external factors is processed by the second Shortcut to obtain result C. Result A, result B, and result C are concatenated to obtain result D.

6. The method for predicting the shear strength of coarse-grained soil based on deep learning according to claim 5, characterized in that, The feature extraction module consists of an inlet stream, eight intermediate streams, and an outlet stream; The ingress stream consists of a first convolutional layer, a first ReLU activation function, a second convolutional layer, a first depthwise separable convolutional module, a second depthwise separable convolutional module, and a third depthwise separable convolutional module, with the result D serving as the input to the first convolutional layer. The first depthwise separable convolutional module consists of a second ReLU activation function, a first depthwise convolutional layer, a third ReLU activation function, a second depthwise convolutional layer, a first max pooling layer, and a first pointwise convolutional layer; the output of the second convolutional layer serves as the input to the second ReLU activation function and the first pointwise convolutional layer, and the output of the max pooling layer is concatted with the output of the first pointwise convolutional layer. The second depthwise separable convolutional module consists of a fourth ReLU activation function, a third depthwise convolutional layer, a fifth ReLU activation function, a fourth depthwise convolutional layer, a second max pooling layer, and a second pointwise convolutional layer; the result of the first concat operation is used as the input to the fourth ReLU activation function and the second pointwise convolutional layer, and the output of the second max pooling layer and the output of the second pointwise convolutional layer are subjected to a second concat operation; The third depthwise separable convolutional module consists of the sixth ReLU activation function, the fifth depthwise convolutional layer, the seventh ReLU activation function, the sixth depthwise convolutional layer, the third max pooling layer, and the third pointwise convolutional layer. The result of the second concat operation is used as the input to the sixth ReLU activation function and the third pointwise convolutional layer. The output of the third max pooling layer and the output of the third pointwise convolutional layer are used for the third concat operation. Each intermediate stream consists of a seventh deep convolutional layer, an eighth deep convolutional layer, a ninth deep convolutional layer, and a fourth concat operation; the output of the ninth deep convolutional layer is combined with the result of the third concat operation to perform the fourth concat operation; for the second to eighth intermediate streams, the result of the fourth concat operation of the previous intermediate stream is used as the input of the seventh deep convolutional layer of the current intermediate stream, and the output of the ninth deep convolutional layer of the current intermediate stream is combined with the result of the fourth concat operation of the previous intermediate stream to perform the fourth concat operation of the current intermediate stream. The outflow consists of the tenth deep convolutional layer, the eleventh deep convolutional layer, the fourth max pooling layer, the fifth concat operation, the twelfth deep convolutional layer, the thirteenth deep convolutional layer, and the global average pooling layer. The result of the fourth concat operation of the eighth intermediate flow is combined with the output result of the fourth max pooling layer for the fifth concat operation.

7. The method for predicting the shear strength of coarse-grained soil based on deep learning according to claim 5, characterized in that, The regression head module consists of a first fully connected layer, an eighth ReLU activation function, and a second fully connected layer. The output of the feature extraction module serves as the input to the first fully connected layer.

8. The method for predicting the shear strength of coarse-grained soil based on deep learning according to claim 1, characterized in that, In step 3, mean absolute error and mean squared error are used as loss functions during the training process.

9. The method for predicting the shear strength of coarse-grained soil based on deep learning according to claim 1, characterized in that, The specific process of step 5 is as follows: Step 5.1, based on each peak shear strength predicted in Step 4 q peak Draw the Mohr's circle and strength envelope for each confining pressure, and obtain the corresponding internal friction angle for each confining pressure based on the strength envelope. φ ; Step 5.2, since the strength envelope of coarse-grained soil is non-linear, it is expressed as: （1） In the formula, φ Indicates the angle of internal friction. p a Indicates atmospheric pressure. σ 3 represents confining pressure. φ 0 and △ φ These represent the intercept and slope of the approximate straight line in logarithmic coordinates when the internal friction angle varies with confining pressure; Step 5.2, based on the internal friction angle obtained in Step 5.1 φ ,draw φ and σ 3 / p a The relationship diagram is used to obtain the shear strength index of coarse-grained soil. φ 0 and Δ φ .

Images