Saturated clay shear strength prediction method based on machine learning and feature selection

Abstract

The invention relates to the technical field of test instruments, and discloses a machine learning and feature selection-based saturated clay shear strength prediction method, which comprises the following steps of: obtaining physical property parameters of saturated clay and corresponding shear strength data, and establishing an initial data set; performing data preprocessing on the initial data set to obtain a training set and a verification set; taking the screened training set and verification set as input features; constructing a machine learning model, and establishing a mapping relationship between the physical property parameters and the corresponding shear strength data; and utilizing a machine learning model to predict and obtain an anti-shearing strength value. According to the method, physical property parameters and corresponding anti-shearing strength data of saturated clay are collected, and a high-dimensional nonlinear mapping relation between the physical property parameters and the corresponding anti-shearing strength data is constructed by utilizing a feedforward neural network algorithm. Therefore, rapid and automatic prediction of the shear strength can be realized based on the easily obtained physical property parameters, the prediction efficiency is remarkably improved, and the limitation that a traditional test method is time-consuming and labor-consuming is broken through.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of testing instruments, in particular to a saturated clay shear strength prediction method based on machine learning and feature selection. BACKGROUND

[0002] The undrained shear strength of saturated clay, as a key parameter reflecting the mechanical properties of clay, is crucial for evaluating the bearing capacity of foundation, the stability of slope, and the interaction between soil and structure. Currently, the methods commonly used to determine the undrained shear strength of saturated clay include laboratory triaxial shear test, vane shear test, and empirical methods based on in-situ testing. However, although traditional test methods can obtain the undrained shear strength of saturated clay, they require a large amount of manpower, material resources, and time, and the maintenance cost of equipment is high, especially in small and medium-sized engineering or emergency survey scenarios, where the economy is difficult to meet the actual demand.

[0003] With the development of numerical analysis technology and the significant improvement of computer computing power, machine learning has gradually become a new method to solve complex engineering problems, especially nonlinear problems. Machine learning methods can effectively process massive soil parameters, identify nonlinear relationships, and provide accurate prediction results based on effective data.

[0004] Natural saturated clay often exhibits significant nonlinear response, such as strain softening effect and structural dependence, and traditional methods based on linear assumption analysis model are difficult to accurately model such behavior. Therefore, the present application proposes a saturated clay shear strength prediction method based on machine learning technology and feature selection method. SUMMARY

[0005] To overcome the shortcomings of the prior art, the present application provides a saturated clay shear strength prediction method based on machine learning and feature selection, which solves the problems of high cost, time-consuming, labor-intensive, and difficulty in meeting the economic needs of small and medium-sized engineering of traditional test methods, and the problem of inaccurate prediction of traditional analysis model due to the difficulty in accurately capturing the nonlinear mechanical behavior of saturated clay.

[0006] To achieve the above purpose, the present application realizes the following technical scheme: a saturated clay shear strength prediction method based on machine learning and feature selection, comprising the following steps:

[0007] Obtain the physical parameters of saturated clay and the corresponding shear strength data, and establish an initial data set;

[0008] Data preprocessing is performed on the initial data set to obtain a training set and a validation set;

[0009] Using a feature selection method, the filtered training set and validation set are used as input features;

[0010] Build a machine learning model and use the input features and corresponding shear strength data for training to establish a mapping relationship between physical property parameters and corresponding shear strength data;

[0011] The trained machine learning model is used to predict the shear strength value of the saturated clay according to the input physical property parameters of the saturated clay to be predicted.

[0012] Preferably, the physical property parameters include natural moisture content, porosity, liquid limit, plastic limit, liquid index, plasticity index, dry density and cohesion of saturated clay.

[0013] Preferably, the data preprocessing includes:

[0014] The initial data set is divided into a training set and a validation set, wherein 70% is the training set and 30% is the validation set, and the data in the initial data set are standardized based on the mean and standard deviation of the training set data.

[0015] Preferably, the standardization processing formula includes:

[0016]

[0017] Among them, x k and x k,nom are the measured values ​​before and after normalization respectively; μ k and σ k Represents the kth input variable x in the training data set k The mean and standard deviation of .

[0018] Preferably, the feature selection method includes the fusion of grey relational analysis method and maximum mutual information coefficient method, and the training set and the validation set are screened by analyzing the correlation between each physical parameter and shear strength.

[0019] Preferably, the fusion of the grey relational analysis method and the maximum mutual information coefficient method comprises:

[0020] (1) The grey relational analysis method specifically includes: constructing a dynamic relational model of a reference sequence and a comparison sequence, wherein the reference sequence represents shear strength, and the comparison sequence represents physical property parameters, and calculating the grey relational degree:

[0021] δ ij =|Y ij -X ij ∣;

[0022] Among them, Y ij represents the value in the reference sequence at the jth data point; X ij Represents the value at the jth data point, which is the value in the comparison sequence; δij value Y of the reference sequence ij value X of the comparison sequence ij the absolute difference at data point j;

[0023]

[0024] wherein, γ j represents the grey correlation degree coefficient of the reference sequence and the comparison sequence at the jth data point; ρ is the resolution coefficient; minminδ ij represents the minimum value of the absolute difference between the value Y of all reference sequences ij and the value X of the comparison sequence ij ; maxmaxδ ij represents the maximum value of the absolute difference between the value Y of all reference sequences ij and the value X of the comparison sequence ij ;

[0025] (2) The maximum mutual information coefficient method specifically includes:

[0026]

[0027] wherein, MIC(X, Y) is the maximum mutual information coefficient; X represents the comparison sequence; Y represents the reference sequence; I(X, Y) is the mutual information between X and Y; N is the sample quantity; k is the interval number for dividing the observation value range represented by X; and l is the interval number for dividing the observation value range of Y;

[0028] (3) The grey correlation degree analysis method and the MIC method are fused to perform weighted summation on the retention rate of the same feature:

[0029] F score = 0.5MIC + 0.5γ;

[0030] wherein, F score is the result of the weighted summation of the retention rate of the same feature, γ is the grey correlation degree coefficient, and MIC is the maximum mutual information coefficient.

[0031] Preferably, the machine learning model selects a Feedforward-Neural-Network type with adaptive regularization capability, specifically including:

[0032] input layer: according to the feature screening result;

[0033] hidden layer: designed as double layers, the first layer is designed as 32 nodes, adopts Rectified-Linear-Unit activation, and the second layer is designed as 16 nodes, adopts Leaky-ReLU activation;

[0034] Output layer: for nodes, directly output the prediction value of the shear strength of saturated clay.

[0035] Preferably, an artificial constraint layer is added after the input layer to constrain the lower limit of the prediction value by the shear strength formula of saturated clay in soil mechanics.

[0036] The present application provides a saturated clay shear strength prediction method based on machine learning and feature selection.

[0037] The present application has the following advantages:

[0038] 1. The present application can quickly and automatically predict the shear strength based on easily accessible physical parameters by collecting various physical parameters of saturated clay and corresponding shear strength experimental data, and using a feedforward neural network algorithm, thereby establishing a high-dimensional nonlinear mapping relationship between physical parameters and shear strength, and significantly improving the prediction efficiency and overcoming the time-consuming and labor-intensive limitations of traditional test methods.

[0039] 2. The present application introduces a feature engineering optimization link, specifically combining gray correlation analysis and maximum mutual information coefficient method to screen input physical parameters, effectively eliminating redundant or weakly correlated features, and retaining key features that contribute most to the target variable, and combining subsequent model iteration training and hyperparameter optimization, so that the machine learning model obtained by training can more accurately capture the multivariate and highly coupled internal laws between the physical parameters of clay and shear strength under complex geological conditions, and the artificial constraint layer can further ensure the physical reasonableness of the prediction results and enhance the robustness of the model.

[0040] 3. The present application can quickly and accurately provide prediction results of the shear strength of saturated clay based on the method, which has great potential in application scenarios such as preliminary evaluation of foundation in small and medium-sized engineering projects, temporary engineering, emergency survey, etc. By replacing or partially replacing traditional cumbersome and costly indoor or in-situ tests, the present application can effectively save the direct cost and indirect time cost of project in geotechnical investigation and testing, thereby helping to shorten the overall project period and improve the economic benefits of engineering construction. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The present application is a method flowchart. DETAILED DESCRIPTION

[0042] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described in the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0043] In order to better understand the present application, the above will be described in detail in combination with specific embodiments.

[0044] Please refer to the accompanying drawings Figure 1 The embodiment of the present application provides a saturated clay shear strength prediction method based on machine learning and feature selection, which comprises the following steps:

[0045] Obtain the physical parameters of saturated clay and the corresponding shear strength data, and establish an initial data set;

[0046] In this embodiment, the construction of the initial data set strictly follows the principles of scientificity and practicality, which is specifically described as follows:

[0047] The data mainly comes from detailed engineering investigation reports. Generally, these reports contain a large amount of soil test data obtained in actual engineering, which has high reliability. By systematically collecting and organizing the relevant data in these reports, the authenticity and consistency of the data set can be ensured. As an option, verified test data from multiple reliable sources can also be integrated to expand the size and coverage of the data set.

[0048] In order to ensure the learning effect of the model and the accuracy of subsequent prediction, the embodiment of the present application suggests collecting at least 500 groups of soil property parameter samples of saturated clay. A larger data set is usually helpful for machine learning model to better learn the complex patterns and nonlinear relationships in the data, thereby improving the generalization ability of the model.

[0049] In some engineering practices, a variety of saturated clay physical parameters can be collected. For example, the parameters that can be concerned include natural moisture content, void ratio, liquid limit, plastic limit, liquidity index, plasticity index, dry density, and even cohesion. However, in one specific embodiment of the present application, in order to construct a model for predicting shear strength, the selected input physical parameters and the target shear strength parameters have a specific correspondence, which is detailed as follows:

[0050] Specifically, each group of data samples collected contains the following input physical property parameters (Input-Physical-Property-Parameters), which are the basic features for subsequent prediction of shear strength, and thus a database can be established:

[0051] Specifically, according to the engineering investigation report, more than 500 groups of soil property parameters of saturated clay are collected, and the parameters include natural moisture content, void ratio, liquid limit, plastic limit, liquidity index, plasticity index, dry density, cohesion and internal friction angle, as shown in Table 1:

[0052] Table 1: Input and output parameter set of saturated clay

[0053] Parameter Symbol Unit Code Natural water content ω ％ X1 Pore ratio e - X2 Liquid limit L ]]> ​ ％ X3 Plastic limit p ]]> ​ ％ X4 Liquidity index I L ]] - X5 Plasticity index I P ]] - X6 Dry density d ]]> ​ g / cm3 X7 Cohesion c kPa [Y1] Internal friction angle φ ° [Y2]

[0054] In one possible implementation, in order to enhance the regularity and consistency within the data set, it is recommended to aggregate the saturated clay sample data of the same region or similar geological units. This can effectively reduce the influence of soil property variation caused by regional differences, so that the model can more systematically learn and reflect the internal relationship between the physical property parameters of saturated clay and its shear strength.

[0055] The final initial data set is a structured data set, each record of which contains a complete set of physical property parameters and corresponding shear strength parameters. This initial data set is the direct input for subsequent data preprocessing, feature selection (for example, using the fusion strategy of the gray correlation degree analysis method and the MIC method discussed above to evaluate the importance of each physical property parameter to the prediction target), and machine learning model construction and training. The collected specific physical property parameters and shear strength indicators are screened by domain knowledge and are considered to be key factors that affect and characterize the mechanical behavior of saturated clay.

[0056] The initial data set is preprocessed to obtain a training set and a validation set;

[0057] In this embodiment, after obtaining the input and output parameter set of saturated clay, the saturated clay sample data of the same region is aggregated to construct a saturated clay data set. The data set has regional consistency and can effectively ignore the influence of regional characteristics, and more systematically reflects the internal relationship between the physical property parameters of saturated clay and its shear strength. Then the data set is divided, 70% of which is the training set and 30% of which is the validation set, and all the data sets are standardized, and the processing method is:

[0058]

[0059] wherein, x k and x k,nom are the measured values before and after normalization respectively; μ k and σ k represent the mean and standard deviation of the kth input variable x k in the training data set.

[0060] It is particularly emphasized that, in order to avoid data leakage and ensure the fairness of model generalization performance evaluation, the mean and standard deviation used in the above standardization process must be calculated based on the training set data only.

[0061] Subsequently, using the statistics calculated from the training set, the corresponding variables in the entire initial data set (i.e., including the training set itself and an independent validation set) are standardized and converted.

[0062] This approach ensures the "unknown" nature of the validation set during model evaluation, thus more truly reflecting the model's performance on future new data. If the statistics of the entire data set are used for standardization, the information of the validation set will be indirectly "leaked" to the training process, leading to an overly optimistic evaluation of the model's performance.

[0063] Therefore, through the above data set division and standardization process, the original saturated clay data set is converted into high-quality data suitable for direct use by machine learning models, providing standardized input for the subsequent feature selection module and laying a solid foundation for the construction, training, and effective validation of machine learning models.

[0064] Using feature selection methods, the filtered training set and validation set are used as input features;

[0065] In this embodiment, after completing the division and standardization of the initial data set, the next step is to use effective feature selection methods to select the key features that have the most influence on predicting the shear strength from the numerous pre-processed saturated clay physical parameters. Not all collected physical parameters contribute equally important to the target variable (shear strength); some parameters may have weak correlation with the target variable, or even be redundant information or noise. Introducing these irrelevant features not only increases the complexity of the subsequent machine learning model, prolongs the training time, but also may reduce the prediction accuracy and generalization ability of the model. Therefore, feature selection is an important link to improve model performance and enhance model interpretability, which directly determines the quality of the feature subset input into the machine learning model.

[0066] Therefore, to achieve efficient and robust feature screening, a strategy of combining Grey-Relational-Analysis (GRA) and Maximal-Information-Coefficient (MIC) is adopted. This fusion method aims to comprehensively utilize the advantages of the two different evaluation indicators to more comprehensively measure the linear and nonlinear correlation between each physical property parameter and the shear strength, thereby avoiding the one-sidedness that may be brought by a single method and selecting the optimal feature subset. The screening process is mainly based on the training set data to avoid the early leakage of validation set information, and then the screening results are applied to the training set and the validation set to ensure the consistency of the feature dimensions of the two data sets.

[0067] Specifically, the detailed process of feature selection is as follows:

[0068] Application of Grey Relational Analysis (GRA):

[0069] Grey relational analysis is a multi-factor statistical analysis method that judges the closeness of the correlation between factors according to the similarity of the geometric shape of the sequence curve. The more similar the curve, the greater the correlation between the corresponding sequences, and vice versa.

[0070] Therefore, (1) the grey relational analysis method specifically includes: constructing a dynamic correlation model of the reference sequence and the comparison sequence, the reference sequence representing the shear strength, the comparison sequence representing the physical property parameter, and calculating the grey correlation degree:

[0071] δ ij =∣Y ij -X ij ∣;

[0072] Where Y ij represents the value in the reference sequence at the jth data point, X ij represents the value in the comparison sequence at the jth data point, and δ ij represents the absolute difference between the value Y \j of the reference sequence and the value X ij of the comparison sequence at the jth data point.

[0073]

[0074] Where γ j represents the grey correlation coefficient of the reference sequence and the comparison sequence at the jth data point, ρ is the resolution coefficient, minminδ ij represents the minimum value of the absolute difference between all reference sequence values Y ij and comparison sequence values X ij , and maxmaxδ ijY represents the value of the reference sequence ij X represents the value of the comparison sequence ij between the absolute difference values;

[0075] (2) The maximum mutual information coefficient (MIC) is an advanced method for measuring the strength and type of the relationship between two variables, which has good detection ability for linear and nonlinear relationships between variables, and has universality and balance, so the maximum mutual information coefficient method specifically includes:

[0076]

[0077] wherein MIC(X, Y) is the maximum mutual information coefficient; X represents the comparison sequence; Y represents the reference sequence; I(X, Y) is the mutual information between X and Y; N is the sample number; k is the number of intervals into which the range of observation values represented by X is divided; and l is the number of intervals into which the range of observation values of Y is divided;

[0078] (3) To obtain more robust feature evaluation, the correlation measures obtained by the above two methods are fused, and the gray correlation degree analysis method and the MIC method are fused to perform weighted summation on the retention rate of the same feature:

[0079] F score = 0.5MIC + 0.5γ;

[0080] wherein F score is the result of weighted summation of the same feature retention rate, γ is the gray correlation degree coefficient, and MIC is the maximum mutual information coefficient.

[0081] Therefore, by calculating the weighted summation result of each physical property parameter, the importance of all original physical property parameters can be ranked.

[0082] Then, according to a preset threshold (for example, selecting features with a weighted summation result greater than a certain value) or selecting a few features with a high ranking of the weighted summation result as the final key input features.

[0083] In summary, after the above feature selection step, the original training set and validation set containing all physical property parameters will be "filtered", that is, only the physical property parameter columns selected as key features and the corresponding shear strength target variable columns are retained. These training set and validation set after feature dimensionality reduction will be used as direct input data for subsequent machine learning model construction and training. This not only reduces the complexity of model learning, but also helps to improve the prediction performance and interpretability of the model.

[0084] A machine learning model is constructed and trained using the input features and corresponding shear strength data to establish a mapping relationship between the physical property parameters and the corresponding shear strength data.

[0085] In this embodiment, the construction and training process of the machine learning model aims to achieve high precision and good generalization ability. The specific technical implementation is described as follows:

[0086] The selection of the machine learning model is crucial for the success of the prediction task.

[0087] In a specific implementation of the present application, a feedforward neural network (FNN) is selected as the core machine learning model. The FNN is selected because of its powerful nonlinear fitting ability, which can effectively handle the complex interactions between soil parameters. At the same time, the selected FNN type is said to have adaptive regularization ability, which helps to prevent overfitting of the model during training, thereby improving the model's performance on unseen data.

[0088] Specifically, the network structure of the feedforward neural network model is designed as follows:

[0089] Input layer (Input-Layer):

[0090] The design of this layer is directly based on the results of the feature selection in the previous step. The number of nodes (or dimensions) of the input layer is completely consistent with the number of key physical parameters finally retained after feature selection. Each node corresponds to a selected physical parameter and is responsible for receiving the measurement value of the parameter from the outside after standardization.

[0091] Artificial constraint layer (Artificial-Constraint-Layer):

[0092] In some embodiments, in order to incorporate prior knowledge in the field of soil mechanics into the model and improve the physical reasonableness of the prediction results, an artificial constraint layer is specially added after the input layer and before the conventional hidden layer.

[0093] The main function of this artificial constraint layer is to impose a reasonable lower bound constraint on the subsequent prediction values of the model through known basic formulas or empirical laws in soil mechanics about the shear strength of saturated clay. For example, according to the principles of soil mechanics, the shear strength should not be lower than a certain theoretical value or empirical lower bound under certain conditions. This constraint layer ensures that the shear strength prediction value of the model output does not violate these basic physical or empirical limits through specific mechanisms (possibly through activation function adjustment, loss function modification, or layer operation), thereby making the prediction results more practically meaningful.

[0094] Hidden layer (Hidden-Layers):

[0095] The hidden layers are the core part of FNN to learn complex nonlinear features. In one preferred embodiment of the present application, the hidden layers are designed as a two-layer structure to balance the expressiveness and complexity of the model.

[0096] The first hidden layer: designed to contain 32 neuron nodes. This layer adopts Rectified-Linear-Unit (ReLU) as the activation function. The ReLU activation function is defined as:

[0097] f(x) = max(0, x);

[0098] The advantage is that it can alleviate the problem of gradient disappearance, speed up model convergence, and introduce nonlinearity. Wherein, x is the net input of the neuron.

[0099] The second hidden layer: designed to contain 16 neuron nodes. This layer adopts Leaky-ReLU as the activation function. The Leaky-ReLU activation function is defined as:

[0100]

[0101] Wherein, x is the net input of the neuron; a is a preset leakage coefficient.

[0102] Output layer (Output-Layer):

[0103] The output layer is responsible for generating the final prediction result. In this embodiment, the output layer is designed to contain a unique node.

[0104] This node directly outputs the predicted value of the shear strength of saturated clay. Since the shear strength is a continuous numerical value, this indicates that the model is constructed as a regression model. The output layer usually adopts a linear activation function, or does not use an explicit activation function, and directly outputs the weighted sum.

[0105] The training process of the model is to adjust the weight and bias parameters inside the neural network using the training set data screened and preprocessed in the foregoing steps (i.e., the screened key physical property parameters as input, and the corresponding saturated clay shear strength data as target output). The goal of training is to minimize the difference between the model prediction value and the true shear strength value, usually by optimizing a predefined loss function (such as the mean square error loss function) to achieve parameter updates using algorithms such as backpropagation and gradient descent optimizers (such as Adam, SGD, etc.)

[0106] In an embodiment of the present application, the hyperparameters are optimized using a Random-Grid-Search method. This method randomly selects a certain number of candidate hyperparameter combinations from the full distribution of possible values of the pre-defined hyperparameters (e.g. learning rate, batch size, regularization parameter, alpha value of Leaky-ReLU, number of hidden layer nodes, etc.).

[0107] For each selected hyperparameter combination, an FNN model is independently trained, and its prediction performance is evaluated using the reserved validation set data (e.g. in terms of root mean square error RMSE, mean absolute error MAE, or coefficient of determination R 2 The final set of hyperparameters that enables the model to perform best on the validation set is selected by comparing the performance of the models under different hyperparameter combinations.

[0108] That is, after the model training and hyperparameter optimization are completed, the final determined model is comprehensively validated using the reserved 30% validation set that has not been "seen" by the model during the entire training and optimization process.

[0109] This step aims to objectively evaluate the generalization ability and actual prediction accuracy of the model. Only when the performance indicators (e.g. accuracy, stability) of the model on the validation set reach the pre-set engineering application standards or acceptable levels, is the machine learning model considered to be constructed and trained, and can be subsequently deployed and applied for predicting the shear strength of new saturated clay samples.

[0110] Using the trained machine learning model, the shear strength value of the input saturated clay is predicted based on the input physical parameters.

[0111] In this embodiment, after successfully constructing and completing the training and hyperparameter optimization of the machine learning model, the weights and biases parameters between the internal neurons of each layer have been adjusted to a state that can better fit the relationship between the input (physical parameters) and output (shear strength) in the training data.

[0112] This optimized network structure and all learned parameter sets collectively constitute a nonlinear mapping function from the input physical parameter space to the output shear strength space.

[0113] Generally, this mapping relationship is not an explicit, simple mathematical formula, but is embedded in the multi-layer nonlinear transformation of the neural network.

[0114] Specifically, when a set of selected physical parameters (which have been standardized in the same way as during training) are input into the first layer (input layer) of the network, the data propagates forward layer by layer.

[0115] First, the data can flow through the artificial constraint layer, which imposes preliminary rational guidance on the potential prediction range according to soil mechanics principles.

[0116] Subsequently, the data enters the first hidden layer (for example, containing 32 nodes, using ReLU activation function), where the first nonlinear transformation is performed.

[0117] Then, the output of the first hidden layer is used as the input of the second hidden layer (for example, containing 16 nodes, using Leaky-ReLU activation function), for further feature extraction and nonlinear transformation.

[0118] Finally, these highly abstracted and transformed information is passed to the output layer (for example, a single node), which integrates all the information and produces a single numerical output.

[0119] This entire calculation path from the initial physical parameter input to the final numerical output, as well as the combined effect of all weights, biases, and activation functions in the path, is the mapping relationship between physical parameters and shear strength established by the invention. This mapping relationship is the result of data-driven learning and can capture complex coupling rules that traditional empirical formulas cannot express.

[0120] Once the feedforward neural network model is trained and verified to achieve the expected prediction accuracy and generalization ability, it can be deployed to predict the shear strength of new, unknown saturated clay samples.

[0121] In one possible application scenario, when the shear strength of a saturated clay at a certain location needs to be predicted, a series of physical parameters of the clay sample must first be collected.

[0122] The collected physical parameters must be consistent with the feature set used during model training, i.e., those key physical parameters selected by the aforementioned feature selection method.

[0123] Next, a crucial step is to preprocess the newly collected physical parameter data in exactly the same way as during the training phase.

[0124] Specifically, this means that the mean and standard deviation calculated on the original training set must be used to standardize the new input data. This is key to ensuring consistency in the distribution of model input data and avoiding prediction bias.

[0125] The physical property parameter data after standardization can be fed as input to the input layer of the trained feedforward neural network model.

[0126] The data is forward propagated in the network according to the previously established mapping relationship: through the input layer, the possible artificial constraint layer, the double hidden layer (the neurons in each layer are weighted and summed and processed through the specific activation function ReLU or Leaky-ReLU), and finally reaches the output layer.

[0127] The only node of the output layer calculates a numerical value. The numerical value is the shear strength value predicted by the machine learning model for the current input saturated clay sample.

[0128] The predicted value can be directly used in practical applications such as engineering evaluation and design parameter selection, providing fast and economical technical support for related decision-making. In this way, the present application realizes intelligent and automated prediction of the shear strength of saturated clay.

[0129] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made hereto without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the shear strength of saturated clay based on machine learning and feature selection, characterized by: The following steps are involved: Obtain the physical properties of saturated clay and the corresponding shear strength data to establish an initial data set; Perform data preprocessing on the initial data set to obtain the training set and validation set; Using feature selection method, the screened training set and validation set are used as input features; Build a machine learning model and use the input features and corresponding shear strength data for training to establish a mapping relationship between physical property parameters and corresponding shear strength data; The trained machine learning model is used to predict the shear strength value of the saturated clay according to the input physical property parameters of the saturated clay to be predicted.

2. The method for predicting the shear strength of saturated clay based on machine learning and feature selection according to claim 1, characterized in that: The physical property parameters include natural moisture content, porosity, liquid limit, plastic limit, liquid index, plastic index, dry density and cohesion of saturated clay.

3. The method for predicting the shear strength of saturated clay based on machine learning and feature selection according to claim 1, wherein: The data preprocessing includes: The initial data set is divided into a training set and a validation set, wherein 70% is the training set and 30% is the validation set, and the data in the initial data set are standardized based on the mean and standard deviation of the training set data.

4. The method for predicting the shear strength of saturated clay based on machine learning and feature selection according to claim 3, wherein: The standardization processing formula includes: Among them, x k and x k,nom are the measured values ​​before and after normalization respectively; μ k and σ k Represents the kth input variable x in the training data set k The mean and standard deviation of .

5. The method for predicting the shear strength of saturated clay based on machine learning and feature selection according to claim 1, wherein: The feature selection method includes the fusion of grey relational analysis method and maximum mutual information coefficient method, and the training set and the validation set are screened by analyzing the correlation between each physical parameter and shear strength.

6. The method for predicting the shear strength of saturated clay based on machine learning and feature selection according to claim 5, characterized in that: The fusion of the grey relational analysis method and the maximum mutual information coefficient method includes: (1) The grey relational analysis method specifically includes: constructing a dynamic relational model of a reference sequence and a comparison sequence, wherein the reference sequence represents shear strength, and the comparison sequence represents physical property parameters, and calculating the grey relational degree: d ij =∣Y ij -X ij ∣; Among them, Y ij represents the value in the reference sequence at the jth data point; X ij Represents the value at the jth data point, which is the value in the comparison sequence; δ ij The value Y representing the reference sequence ij The value X of the comparison sequence ij The absolute difference at data point j; Among them, γ j represents the grey correlation coefficient between the reference sequence and the comparison sequence at the jth data point; ρ is the resolution coefficient; minminδ ij Represents the value Y in all reference sequences ij The value X of the comparison sequence ij The minimum absolute difference between ij Represents the value Y in all reference sequences ij The value X of the comparison sequence ij The maximum absolute difference between (2) The maximum mutual information coefficient method specifically includes: Where MIC(X,Y) is the maximum mutual information coefficient; X represents the comparison sequence; Y represents the reference sequence; I(X,Y) is the mutual information between X and Y; N is the number of samples; k is the number of intervals into which the range of observations represented by X is divided; l is the number of intervals into which the range of observations represented by Y is divided; (3) Combining the grey relational analysis method and the MIC method, the retention rate of the same feature is weighted summed: F score =0.5MIC+0.5γ; Among them, F score is the weighted sum of the retention rates of the same feature, γ is the grey relational coefficient, and MIC is the maximum mutual information coefficient.

7. The method for predicting the shear strength of saturated clay based on machine learning and feature selection according to claim 1, characterized in that: The machine learning model is a Feedforward-Neural-Network type with adaptive regularization capability, specifically including: Input layer: filter results based on features; Hidden layer: Designed as a two-layer, the first layer is designed with 32 nodes and uses Rectified-Linear-Unit activation, and the second layer is designed with 16 nodes and uses Leaky-ReLU activation; Output layer: It is a node that directly outputs the predicted value of the shear strength of saturated clay.

8. The method for predicting the shear strength of saturated clay based on machine learning and feature selection according to claim 5, characterized in that: An artificial constraint layer is added after the input layer to constrain the lower limit of the predicted value by using the shear strength formula of saturated clay in soil mechanics.

Images