Intelligent prediction method for seepage characteristic parameters of earth and stone materials

Abstract

The invention relates to an intelligent prediction method for seepage characteristic parameters of earth and rocks, which comprises the following steps of: primarily selecting and processing an indoor seepage test result data set of the earth and rocks, constructing an interpretable machine learning model to quantify the influence degree of different characteristic parameters on the seepage characteristic parameters of the earth and rocks, and realizing dimensionality reduction and optimization of the data set. And establishing an intelligent combined prediction model fusing the bidirectional long-short-term memory neural network, a sparrow search algorithm and attention mechanism optimization, completing prediction model training based on the optimized data set, and inputting related characteristic parameters to realize intelligent prediction of soil and stone seepage characteristic parameters. According to the invention, engineering technicians can be assisted to complete rapid and intelligent prediction of seepage characteristic parameters (permeability coefficient and critical permeability drop) according to gradation characterization parameters, porosity, dry density and other basic characteristic parameters of soil and stone, and the test cost is reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of experimental research and application of physical and mechanical properties of hydraulic rock materials, and in particular relates to an intelligent prediction method for soil and rock material seepage characteristic parameters. Background Art

[0002] The seepage characteristic parameters (permeability coefficient and critical permeability gradient) of soil and rock materials are one of the key parameters for judging their engineering applicability. They are closely related to the seepage safety design, analysis and evaluation of the project, and are also the key content related to the safe operation of the project. The acquisition of soil and rock material seepage characteristic parameters usually requires indoor seepage tests or permeability deformation tests to obtain the corresponding permeability coefficient and permeability gradient. Traditional indoor seepage tests are easily affected by factors such as test equipment limitations and sampling and transportation links, which may result in errors in the test results. In addition, indoor tests have a long cycle and high costs, making it difficult to quickly obtain seepage characteristic parameters at the project site, which is not conducive to timely and effective analysis and judgment by engineering and technical personnel. In response to this, relevant researchers have proposed an empirical formula method based on statistical analysis of a large amount of test data, but it has problems such as weak applicability, limited prediction objects and insufficient prediction accuracy, so it is limited in application in actual engineering.

[0003] With the development of artificial intelligence technology, intelligent algorithms have gradually been applied to the fields of water conservancy and geotechnical engineering. Based on the known input parameters and the corresponding output data, a prediction model can be established to predict the values ​​of unknown parameters, which is of great significance for the evaluation and prediction of parameters such as the mechanical properties and permeability characteristics of rock and soil. At present, when predicting geotechnical material parameters based on intelligent algorithms, there are problems such as redundant model data sets, high data quality requirements, poor model generalization ability, and internal black boxes in the model. Therefore, it is necessary to explore the multi-dimensional data analysis advantages of new intelligent combination algorithms, and form an implementation method for predicting soil and rock seepage characteristic parameters based on basic characteristic parameters such as soil and rock grading characteristic parameters, pore structure parameters, and hydraulic condition parameters, so as to achieve rapid, accurate and intelligent prediction of soil and rock seepage characteristic parameters.

[0004] Based on the above analysis, the existing technology still has the following shortcomings: although the existing machine learning methods can improve the prediction accuracy, they are usually regarded as "black box" models, which make it difficult to explain the influence of various characteristic parameters on the seepage characteristics, and are not conducive to a deep understanding of the physical properties of the materials. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent prediction method for the seepage characteristic parameters of soil and rock materials, which can quickly and accurately predict the seepage characteristic parameters (including permeability coefficient and critical permeability gradient) of soil and rock materials based on their basic characteristic parameters, so as to facilitate engineering and technical personnel to quickly judge the applicability of soil and rock materials.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: An intelligent prediction method for soil-rock seepage characteristic parameters, comprising the following steps: S1. Collect and organize the soil and stone indoor seepage test results to obtain the original data set; S2. Draw a box plot of the data points of each characteristic parameter, determine and remove abnormal distribution data, and complete the preliminary selection of the original data set; S3. Build an interpretable machine learning model based on SHAP theory and LightGBM algorithm, and train and calculate the interpretable machine learning model based on the original data set after the preliminary selection; S4. Calculate the Shapley value of each characteristic parameter using an interpretable machine learning model to quantify the influence of each characteristic parameter on the seepage characteristics; S5. Perform data dimension reduction according to the impact level, remove relevant feature parameters, and complete the lightweighting of the data set; S6. Establish an intelligent prediction model for soil-rock seepage characteristic parameters, and determine input parameters and target parameters of the intelligent prediction model for the lightweight data set; S7. Determine the ratio of the training set and the test set of the intelligent prediction model, train the intelligent prediction model, calculate the performance index of the test set model, and complete the training of the intelligent prediction model; S8. Input the characteristic parameters required for prediction of a soil-rock sample into the intelligent prediction model, and use the established intelligent prediction model to predict and output the progressive seepage characteristic parameters.

[0007] Preferably, in step S2, according to the box plot distribution of each characteristic parameter point, the data points outside the 5% quantile and 95% quantile interval are set as abnormal interval data, that is, abnormal data. The abnormal data is outside the stable distribution interval. If there is abnormal data under a certain characteristic parameter distribution, the group data corresponding to the characteristic parameter is eliminated.

[0008] Preferably, in step S3, the Python language compiler is used to call the LightGBM model to perform interpretable machine learning model training and testing, wherein the test set is 15% of the sample data, and the remaining sample data is used as a training set to complete the interpretable machine learning model training and build an interpretable machine learning model based on the SHAP-LightGBM algorithm.

[0009] Preferably, in step S4, the Shapley value calculation method is as follows: ; in, Represents the characteristic variables in the data set; subscript i Representative i Features; M is the total number of characteristic variables; S Is a collection There are 2 subsets of M-1 Possibility, indicating S The total number of elements in; Indicates that there are S The model output value when there are features, Representation characteristics i exist S Marginal contribution under the collection.

[0010] Preferably, in step S4, the calculation results of the Shapley values ​​of the characteristic parameters are sorted, and the sorting results are used to reflect the influence of different characteristic parameters on the permeability coefficient of gravel soil; the characteristic parameters include particle size d 10 , porosity ratio e And the dry density of the sample ρ d , particle size d 10 , porosity ratio e And the dry density of the sample ρ d The influence on the permeability coefficient of gravel soil samples is higher than other characteristic parameters.

[0011] Preferably, the processing process of step S5 is: taking the Shapley absolute mean of each characteristic variable as the classification data basis, taking the maximum Shapley value absolute mean as the benchmark, recorded as SP max , calculate the absolute average value of SHAP for the remaining feature variables sp The absolute average of the maximum Shapley value SP max Ratio sp r , and define sp r The range [0,0.05) indicates a slight impact, the range [0.05,0.5) indicates a moderate impact, and the range [0.5,1] indicates a significant impact.

[0012] Preferably, in step S6, the intelligent prediction model mainly includes a network training layer, a parameter optimization layer and an attention mechanism layer. The network training layer uses a BiLSTM model to construct a nonlinear mapping relationship between the relevant characteristic parameters of the gravel soil material and its permeability coefficient. The BiLSTM model uses an Adam gradient descent algorithm for parameter adjustment training; The parameter optimization layer mainly optimizes the three main parameters of the BiLSTM model: learning rate, batch size, and regularization coefficient; The parameter optimization algorithm uses the sparrow search algorithm to search for the optimal solution by updating the sparrow position. The update mechanism is as follows: Finder's location update: ; Where: x i,j Indicates i A sparrow in j The position element in the dimension, i =1~ N , j =1~ d ; t Indicates the current iteration number, t max is the preset maximum number of iterations; α is a uniform random number in (0,1], Q Represents a standard normal distribution random number; R 2 and ST They represent the warning value and the safety value respectively; When a new discoverer appears, a corresponding sparrow must be transformed into a joiner; the position update method of the joiner is as follows: ; Where: x b is the sparrow with the best fitness in the current population, x w is the sparrow with the worst fitness in the current population; i > n / 2, indicating that the fitness value is lower. i A follower has not obtained food and is in a very dangerous state. At this time, you need to continue searching in other areas; The population position is randomly initialized, and the fitness value of each sparrow in the validation set is calculated according to the fitness function. The optimal solution is further searched by updating the sparrow position. The optimal position obtained after the maximum number of iterations is the optimal hyperparameter required; the fitness function is as follows: ; Where: m is the number of validation set data samples, y i Output the true experimental value of the permeability coefficient for the validation set, y i ′ is the predicted value of the intelligent prediction model.

[0013] Preferably, the attention mechanism layer adopts a multi-head attention mechanism, which is used in the hidden layer update stage under the multi-input features of the BiLSTM model. The output of the hidden layer of the BiLSTM model is used as the input vector of the attention mechanism layer, and the Attention weight matrix is ​​calculated based on the softmax activation function. Then, the multi-head attention is merged, and the weight matrix is ​​added for linear transformation to obtain the final model output.

[0014] Preferably, in step S7, for the division of the model training set and the test set, 85% of the data samples are used as the training set, and the remaining 15% of the data samples are used as the validation set, and the training set and the test set are randomly divided according to the total data samples.

[0015] An intelligent prediction system for soil-rock material seepage characteristic parameters adopts the intelligent prediction method for soil-rock material seepage characteristic parameters. The present invention can achieve the following beneficial effects: (1) The interpretable model constructed by the present invention when predicting the seepage characteristic parameters of soil and rock can quantify the influence of each characteristic parameter on the permeability coefficient and critical permeability gradient of soil and rock, and explain the relationship between the basic physical characteristic parameters such as soil and rock grading curve characterization parameters, porosity, dry density and its seepage characteristic parameters from the perspective of data analysis, thus making up for the black box model defects of traditional machine learning algorithms.

[0016] (2) The interpretable model method provided by the present invention can optimize the soil and rock model data set, reduce the data dimension, and reduce the model complexity, thereby improving the model calculation efficiency.

[0017] (3) The intelligent prediction method for seepage characteristic parameters of soil and rock materials provided by the present invention has the advantages of high prediction accuracy, strong generalization ability and high calculation efficiency. It makes up for the shortcomings of traditional empirical formulas in predicting seepage characteristic parameters of gravel soil samples, such as narrow applicability and insufficient accuracy. It is convenient for engineering and technical personnel to quickly predict the seepage characteristic parameters of soil and rock materials on site based on their basic physical characteristic parameters such as grading parameters, porosity and dry density, and also reduces the test cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is an implementation flow chart of an intelligent prediction method for soil and rock material seepage characteristic parameters provided by an embodiment of the present invention; Figure 2 It is the structure of each layer of the prediction model in an intelligent prediction method for soil and rock material seepage characteristic parameters provided by an embodiment of the present invention; Figure 3It is a ranking result of the influence degree of different characteristic parameters on the permeability coefficient in an intelligent prediction method of soil and rock seepage characteristic parameters provided by an embodiment of the present invention; Figure 4 It is a diagram of the results of training a permeability coefficient prediction model of an intelligent prediction method for soil and rock material seepage characteristic parameters provided by an embodiment of the present invention.

[0019] Figure 5 It is a comparison diagram of permeability coefficient prediction results of an intelligent prediction method for soil and rock material seepage characteristic parameters provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The preferred solution is Figures 1 to 5 As shown in FIG. 1 , an intelligent prediction method for seepage characteristic parameters of soil and rock materials is described by taking the prediction of the permeability coefficient of gravel soil in soil and rock materials as an example, and combining the accompanying drawings to illustrate the embodiment of the present invention. Figure 1 As shown, the following steps are included: Step 1: Collect and organize the basic physical property parameters of gravel soil materials and indoor seepage test data sets in related projects. In this embodiment, a total of 509 sets of gravel soil test data are obtained. There are 20 characteristic parameters in the data set, and the target parameter is the permeability coefficient, as shown in Table 1.

[0021] Table 1 Feature parameters contained in the dataset

[0022] The characterization parameters of each gradation curve in the above table can be obtained from the gradation curve obtained from the particle screening test, and the dry density and porosity ratio are determined by corresponding indoor tests.

[0023] Step 2: For the collected original data set, draw a box plot of each feature parameter, and calculate its mean, median, 5% quantile and 95% quantile.

[0024] Furthermore, according to the box plot distribution of each characteristic parameter point, the data points outside the 5% quantile and 95% quantile interval are set as abnormal interval data. These data are outside the stable distribution interval. If there are abnormal data under a characteristic parameter distribution, the corresponding group data of the characteristic parameter will be eliminated. From the source of the data, the comprehensiveness and applicability of the collected data can be guaranteed, and the influence of a few super-large particle size gradation sample data on the subsequent model prediction results can be avoided.

[0025] In this embodiment, after the initial data selection of the original data set, 460 sets of stable data remain for subsequent analysis.

[0026] Step 3: The 460 groups of sample data after the preliminary selection are used as the data set, with 20 characteristic parameters as input and permeability coefficient as output, forming the basic interpretable machine learning model data; Furthermore, the Python language compiler was used to call the LightGBM model for model training and testing. The test set was 15% of the sample data, a total of 69 groups, and the remaining sample data was used as the training set to complete the model training and build an interpretable machine learning model based on the SHAP-LightGBM algorithm. Step 4: Use SHAP theory to calculate the Shapley value of each characteristic parameter input in the LightGBM model on the permeability coefficient, that is, to quantify the degree of influence. The calculation method of Shapley value is as follows:

[0027] in, Represents the subscript of the characteristic variable in the data set, i Representative i Features, M is the total number of feature variables. S Is a collection There are 2 subsets of M-1 Possibility, indicating S The total number of elements in . Indicates that there are S The model output value when there are features, Representation characteristics i exist S The marginal contribution of the set. The previous weights are obtained by permutations and combinations, representing the probability of the element in the set.

[0028] Furthermore, the parameters are sorted according to the calculation results of the Shapley values ​​of the characteristic parameters, such as Figure 2 As shown in Figure 2, the sorting results can significantly reflect the influence of different characteristic parameters on the permeability coefficient of gravel soil. d 10 , porosity ratio e And the dry density of the sample ρ d The influence on the permeability coefficient of gravel soil samples is significantly higher than other characteristic parameters.

[0029] Step 5: To facilitate the subsequent establishment of a prediction model for the seepage characteristic parameters of gravel soil based on relevant characteristic parameters, the influence degree of each characteristic parameter is classified. The classification criteria are as follows: The absolute mean value of Shapley of each characteristic variable is used as the classification data basis, and the absolute mean value of the maximum Shapley value is used as the benchmark, which is recorded as SP max, calculate the absolute average value of SHAP for the remaining feature variables sp The absolute average of the maximum Shapley value SP max Ratio sp r , and define sp r The range [0,0.05) indicates a slight impact, the range [0.05,0.5) indicates a moderate impact, and the range [0.5,1] indicates a significant impact.

[0030] According to the above classification standards, the classification results of the influence degree of each characteristic parameter on the permeability coefficient of gravel soil in this embodiment are as follows in Table 2: Table 2 Feature parameters contained in the dataset Target variable Important impact General Impact Subtle impact Permeability coefficient <![CDATA[Characteristic particle size d 10 , void ratio e, dry density ρ d > <![CDATA[Characteristic particle size d 17 、d 20 、d 30 、d 40 、d 50 、d 60 、d 70 、d 80 、d 90 、d 100 , coefficient of uniformity Cu, coefficient of curvature Cc, area S of grading curve, content P of characteristic particle size d＞5 、P 2＞d＞0.075 > <![CDATA[Characteristic particle size d9, characteristic particle size content P d＜0.075 > Furthermore, according to the classification results in the above table, the characteristic particle size of the minor impact category d 9. Characteristic particle size content P d<0.075 Two characteristic parameters were eliminated, and the data set contained 18 characteristic parameters after dimensionality reduction; Step 6: Figure 3 The present invention provides a soil and rock material seepage characteristic parameter intelligent prediction model structure. The intelligent prediction model mainly includes a network training layer, a parameter optimization layer and an attention mechanism layer.

[0031] The network training layer uses the BiLSTM model to construct a nonlinear mapping relationship between the relevant characteristic parameters of gravel soil and its permeability coefficient. In this embodiment, the BiLSTM model uses the Adam gradient descent algorithm for parameter training, setting the maximum number of training times to 100 times, the learning rate reduction factor to 0.02, the initial learning rate to 0.1, the number of samples selected before each parameter adjustment, and the initial batch size to 16.

[0032] The parameter optimization layer mainly optimizes the three main parameters of the BiLSTM model: learning rate, batch size, and regularization coefficient. In this embodiment, the optimization ranges of each parameter are set as learning rate: [0.001, 0.1], batch size: [16, 64], and regularization coefficient: [0.001, 0.1].

[0033] The parameter optimization algorithm uses the sparrow search algorithm to search for the optimal solution by updating the sparrow position. The update mechanism is as follows: Finder's location update:

[0034] Where: x i,j Indicates i A sparrow inj The position element in the dimension, i =1~ N , j =1~ d ; t Indicates the current iteration number, t max is the preset maximum number of iterations; α is a uniform random number in (0,1], Q Represents a standard normally distributed random number. R 2 and ST They represent the warning value and the safety value respectively.

[0035] When a new discoverer appears, a corresponding sparrow must be transformed into a joiner. The position update method of the joiner is as follows:

[0036] Where: x b is the sparrow with the best fitness in the current population, x w is the sparrow with the worst fitness in the current population. i > n / 2, indicating that the fitness value is lower. i A follower has not obtained food and is in a very dangerous state. At this time, it is necessary to continue searching other areas.

[0037] The population position is randomly initialized, and the fitness value of each sparrow in the validation set is calculated according to the fitness function. The optimal solution is further searched by updating the sparrow position. The optimal position obtained after the maximum number of iterations is the optimal hyperparameter required. In this embodiment, the number of sparrow populations is set to 30, and the maximum number of iterations is 10. The ratios of discoverers, followers, and guards are 0.3, 0.5, and 0.2, and the safety threshold is 0.7. The fitness function is as follows:

[0038] Where: m is the number of validation set data samples, y i Output the true experimental value of the permeability coefficient for the validation set, y i ′ is the predicted value of the intelligent prediction model.

[0039] The attention mechanism adopted in this embodiment is a multi-head attention mechanism, which is mainly used in the hidden layer update stage under the multi-input features of the BiLSTM model. The output of the hidden layer of the BiLSTM model is used as the input vector of the attention mechanism layer, and the Attention weight matrix is ​​calculated based on the softmax activation function. After that, the multi-head attention is merged and the weight matrix is ​​added for linear transformation to obtain the final model output.

[0040] Step 7: The 460 sets of data collected in this embodiment are used as the model training data set. The data set is reliable data after the feature parameter optimization is completed by the interpretable machine learning model in Step 5, that is, data dimension reduction is performed; Furthermore, for the division of model training set and test set, 85% of the data samples are used as the training set, and the remaining 15% of the data samples are used as the validation set, and the training set and the test set are randomly divided according to the total data samples; The optimized model parameters finally obtained in this embodiment are: learning rate: 0.058, batch size: 24, regularization coefficient: 0.001; The comparison results of the test set data in this embodiment are as follows Figure 4 As shown, the predicted value of the permeability coefficient in the test set data and the actual test value show a linear regression relationship close to 45°. The data regression error band is evenly distributed and narrow, and the predicted value of the data point basically falls within the 95% model prediction band, indicating that the model training effect in this embodiment is better and can be used for further prediction applications.

[0041] Step 8: In this embodiment, 15 groups of new data of gravel soil test samples that have not participated in model training are selected for prediction application. Among the 15 groups of data, the 18 characteristic parameters after dimension reduction in Step 5 are used as input, and the output target parameter is the permeability coefficient; Furthermore, the intelligent prediction model network that has been trained in Step 6 is called to output the permeability coefficients of 15 groups of gravel soil test samples; In this embodiment, the permeability coefficient output result obtained by using the soil and rock material seepage characteristic parameter intelligent prediction model provided by the present invention is as follows: Figure 5 As shown, the predicted values ​​of the permeability coefficients in the selected 15 groups of engineering samples have a good linear regression relationship with the experimental values, and the prediction results of most samples are within the error band, indicating that the intelligent prediction model provided by the present invention has good generalization ability and high stability.

[0042] In summary, the present invention mainly includes three aspects of technical content: a preliminary selection scheme for basic data sets of soil-rock seepage test, a method for quantifying the influence of multiple characteristic variables on soil-rock seepage characteristic parameters, and an intelligent prediction model for soil-rock seepage characteristic parameters. The above three aspects of technical solutions are important components of the present invention. Among them: The present invention provides a preliminary selection scheme for a soil-rock seepage test basic data set, including: Collect and organize the indoor seepage test data of soil and rock materials in the project, and establish a basic data set for predicting the seepage characteristic parameters of soil and rock materials. The basic data set is characterized by containing 20 characteristic parameters including gradation curve characterization parameters, porosity and dry density, in addition to two target parameters, namely permeability coefficient and critical permeability drop. Use box-type data distribution charts to statistically analyze the distribution characteristics of the collected data and obtain key information on the location and dispersion of relevant data; The data outside the 5% quantile to the 95% quantile were taken as the abnormal interval, and the abnormal sample data outside the stable distribution interval were eliminated to complete the preliminary selection of the basic data set for soil-rock seepage test; A second aspect of the present invention provides a method for quantifying the influence of multiple characteristic parameters on soil and rock material seepage characteristic parameters, comprising: Build an interpretable machine learning model based on SHAP theory and LightGBM (lightweight gradient boosting algorithm), and train the interpretable model on the data set after the preliminary selection; According to the interpretable model trained above, the Shapley values ​​of each characteristic parameter under the interpretable model are calculated and sorted. The Shapley value represents the influence of different characteristic parameters on the seepage characteristic parameters of soil and rock materials. The larger the Shapley value corresponding to a characteristic parameter, the greater the influence of the characteristic parameter on the seepage characteristic parameters. The calculation method of the above Shapley value is:

[0043] in, Represents the subscript of the characteristic variable in the data set, i Representative i Features, M is the total number of feature variables. S Is a collection There are 2 subsets of M-1 Possibility, indicating S The total number of elements in . Indicates that there are S The model output value when there are features, Representation characteristics i exist S The marginal contribution of the set. The previous weights are obtained by permutations and combinations, representing the probability of the element in the set.

[0044] Furthermore, the influence degree of each characteristic parameter is classified, and the classification criteria are as follows: The absolute mean value of Shapley of each characteristic variable is used as the classification data basis, and the absolute mean value of the maximum Shapley value is used as the benchmark, which is recorded as SP max , calculate the absolute average value of SHAP for the remaining feature variables sp The absolute average of the maximum Shapley value SP max Ratio sp r , and define sp r The range [0,0.05) indicates a slight impact, the range [0.05,0.5) indicates a moderate impact, and the range [0.5,1] indicates a significant impact.

[0045] Select feature parameters according to the above classification criteria to achieve dimensionality reduction of model data sets, thereby reducing model complexity and computational costs; The third aspect of the present invention provides an intelligent prediction model for soil and rock material seepage characteristic parameters, comprising: Select the model data set after dimensionality reduction, determine the number of corresponding input characteristic parameters under different prediction targets, and the prediction targets in the present invention are respectively permeability coefficient and critical permeability drop; Determine the proportion of training set and test set data involved in predictive model training; Establish an intelligent prediction model, which mainly contains a network training layer, a parameter optimization layer, and an attention mechanism layer; The network training layer in the intelligent prediction model of the present invention adopts a bidirectional long short-term memory neural network (BiLSTM) as the core component of the prediction model, and a bidirectional long short-term memory neural network is used to construct a nonlinear mapping relationship between soil and stone material related characteristic parameters and their seepage characteristic parameters. The characteristic of this neural network is that it can make full use of all the information in the data and capture the intrinsic characteristics of the data more comprehensively. In the data training process, both past data and future data have an impact on the current data. By integrating the state parameters of the forward and backward long short-term memory neural networks and superimposing them in an output layer, the network can analyze the data from both forward and backward dimensions at the same time, thereby improving the utilization rate of the data and the accuracy of the prediction.

[0046] The parameter optimization layer in the intelligent prediction model of the present invention adopts the sparrow search algorithm, which mainly optimizes the three main parameters of the learning rate, batch size, and regularization coefficient in the bidirectional long short-term memory neural network used in the present invention. The optimization method is as follows: The population position is randomly initialized, and the fitness value of each sparrow in the test set is calculated according to the fitness function. Then, the optimal solution is searched by updating the sparrow position. The update mechanism is as follows: Finder's location update:

[0047] Where: x i,j Indicates i A sparrow in j The position element in the dimension, i =1~ N , j =1~ d ; t Indicates the current iteration number, t max is the preset maximum number of iterations; α is a uniform random number in (0,1], Q Represents a standard normally distributed random number. R 2 and ST They represent the warning value and the safety value respectively.

[0048] When a new discoverer appears, a corresponding sparrow must be transformed into a joiner. The position update method of the joiner is as follows:

[0049] Where: x b is the sparrow with the best fitness in the current population, x w is the sparrow with the worst fitness in the current population. i > n / 2, indicating that the fitness value is lower. i A follower has not obtained food and is in a very dangerous state. At this time, it is necessary to continue searching other areas.

[0050] The optimal position obtained after reaching the maximum number of iterations is the optimal parameter sought; The fitness function in the present invention is as follows:

[0051] Where: m is the number of test set data samples, yi Output the true value of the variable for the test set, yi ′ is the model prediction value.

[0052] In addition to the above technical solutions, the present invention adds an attention mechanism layer to the intelligent prediction model, which has the following characteristics: it is mainly used in the hidden layer update stage under multi-input features of the bidirectional long short-term memory neural network, and the output of the hidden layer of the bidirectional long short-term memory neural network is used as the input vector of the attention mechanism layer, and the attention weight matrix is ​​calculated based on the softmax activation function.

[0053] The calculation method of the above attention mechanism is as follows: By introducing the task-related query vector Query ( Q ), calculate its difference with the key vector Key ( K ) and accordingly the value vector Value ( V ) are weighted and the attention value is obtained.

[0054] Information input: Q , K , V Input model, X is the input weight vector

[0055] Calculating attention distribution α :Through calculation Q and K Dot product is used to calculate the correlation, and softmax is used to calculate the score. Q = K = V=X , the attention weight is calculated by the softmax activation function,

[0056] Where: α i is the attention probability distribution, s ( x i , q ) is the attention scoring mechanism. The calculation method used in this paper is:

[0057] In the formula: add The scaling factor prevents the dot product result from being too large, which makes the gradient after calculation by the softmax activation function very small, which is not conducive to back propagation.

[0058] Information Weighted Average: Attention Distribution α i To explain the context query, i The degree of attention the information receives.

[0059] .

[0060] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. An intelligent prediction method for soil and rock seepage characteristic parameters, characterized in that The following steps are involved: S1. Collect and organize the soil and stone indoor seepage test results to obtain the original data set; S2. Draw a box plot of the data points of each characteristic parameter, determine and remove abnormal distribution data, and complete the preliminary selection of the original data set; S3. Build an interpretable machine learning model based on SHAP theory and LightGBM algorithm, and train and calculate the interpretable machine learning model based on the original data set after the preliminary selection; S4. Calculate the Shapley value of each characteristic parameter using an interpretable machine learning model to quantify the influence of each characteristic parameter on the seepage characteristics; S5. Perform data dimension reduction according to the impact level, remove relevant feature parameters, and complete the lightweighting of the data set; S6. Establish an intelligent prediction model for soil-rock seepage characteristic parameters, and determine input parameters and target parameters of the intelligent prediction model for the lightweight data set; S7. Determine the ratio of the training set and the test set of the intelligent prediction model, train the intelligent prediction model, calculate the performance index of the test set model, and complete the training of the intelligent prediction model; S8. Input the characteristic parameters required for prediction of a soil-rock sample into the intelligent prediction model, and use the established intelligent prediction model to predict and output the progressive seepage characteristic parameters.

2. The intelligent prediction method of soil and rock material seepage characteristic parameters according to claim 1 is characterized by: In step S2, according to the box plot distribution of each characteristic parameter point, the data points outside the 5% quantile and 95% quantile interval are set as abnormal interval data, that is, abnormal data. The abnormal data is outside the stable distribution interval. If there is abnormal data under a certain characteristic parameter distribution, the corresponding group data of the characteristic parameter will be eliminated.

3. The intelligent prediction method of soil and rock material seepage characteristic parameters according to claim 1 is characterized by: In step S3, the Python language compiler is used to call the LightGBM model to perform interpretable machine learning model training and testing, where the test set is 15% of the sample data, and the remaining sample data is used as the training set to complete the interpretable machine learning model training and build an interpretable machine learning model based on the SHAP-LightGBM algorithm.

4. The intelligent prediction method of soil and rock material seepage characteristic parameters according to claim 1, characterized in that: In step S4, the Shapley value is calculated as follows: ； in, Represents the characteristic variables in the data set; subscript i Representative i Features; M is the total number of characteristic variables; S Is a collection There are 2 subsets of M-1 Possibility, indicating S The total number of elements in; Indicates that there are S The model output value when there are features, Representation characteristics i exist S Marginal contribution under the collection.

5. The intelligent prediction method of soil and rock material seepage characteristic parameters according to claim 4 is characterized by: In step S4, the calculation results of Shapley values ​​of the characteristic parameters are sorted, and the sorting results are used to reflect the influence of different characteristic parameters on the permeability coefficient of gravel soil; the characteristic parameters include particle size d 10 , porosity ratio e And the sample dry density ρ d , particle size d 10 , porosity ratio e And the sample dry density ρ d The influence on the permeability coefficient of gravel soil samples is higher than other characteristic parameters.

6. The intelligent prediction method of soil and rock material seepage characteristic parameters according to claim 1, characterized in that: The processing process of step S5 is: taking the Shapley absolute mean of each characteristic variable as the classification data basis, taking the maximum Shapley value absolute mean as the benchmark, recorded as SP max , calculate the absolute average value of SHAP for the remaining feature variables sp The absolute average of the maximum Shapley value SP max Ratio sp r , and define sp r The range [0,0.05) indicates a slight impact, the range [0.05,0.5) indicates a moderate impact, and the range [0.5,1] indicates a significant impact.

7. The intelligent prediction method of soil and rock material seepage characteristic parameters according to claim 1, characterized in that: In step S6, the intelligent prediction model mainly includes a network training layer, a parameter optimization layer and an attention mechanism layer. The network training layer uses the BiLSTM model to construct a nonlinear mapping relationship between the relevant characteristic parameters of the gravel soil material and its permeability coefficient. The BiLSTM model uses the Adam gradient descent algorithm for parameter adjustment training; The parameter optimization layer mainly optimizes the three main parameters of the BiLSTM model: learning rate, batch size, and regularization coefficient; The parameter optimization algorithm uses the sparrow search algorithm to search for the optimal solution by updating the sparrow position. The update mechanism is as follows: Finder's location update: ； Where: x i,j Indicates i A sparrow in j The position element in the dimension, i =1~ N , j =1~ d ; t Indicates the current iteration number, t max is the preset maximum number of iterations; α is a uniform random number in (0,1], Q Represents a standard normal distribution random number; R 2 and ST They represent the warning value and the safety value respectively; When a new discoverer appears, a corresponding sparrow must be transformed into a joiner; the position update method of the joiner is as follows: ； Where: x b is the sparrow with the best fitness in the current population, x w is the sparrow with the worst fitness in the current population; when i > n / 2, indicating that the fitness value is lower. i A follower has not obtained food and is in a very dangerous state. At this time, you need to continue searching in other areas; The population position is randomly initialized, and the fitness value of each sparrow in the validation set is calculated according to the fitness function. The optimal solution is further searched by updating the sparrow position. The optimal position obtained after the maximum number of iterations is the optimal hyperparameter required; the fitness function is as follows: ； Where: m is the number of validation set data samples, y i Output the true experimental value of the permeability coefficient for the validation set, y i ′ is the predicted value of the intelligent prediction model.

8. The intelligent prediction method of soil and rock material seepage characteristic parameters according to claim 7, characterized in that: The attention mechanism layer adopts a multi-head attention mechanism, which is used in the hidden layer update stage under the multi-input features of the BiLSTM model. The output of the hidden layer of the BiLSTM model is used as the input vector of the attention mechanism layer. The Attention weight matrix is ​​calculated based on the softmax activation function. Then, the multi-head attention is merged and the weight matrix is ​​added for linear transformation to obtain the final model output.

9. The intelligent prediction method of soil and rock material seepage characteristic parameters according to claim 1, characterized in that: In step S7, for the division of the model training set and test set, 85% of the data samples are used as the training set, and the remaining 15% of the data samples are used as the validation set, and the training set and the test set are randomly divided according to the total data samples.

10. An intelligent prediction system for soil and rock seepage characteristic parameters, characterized in that: An intelligent prediction method for soil and rock material seepage characteristic parameters according to any one of claims 1-9 is adopted.

Images