Prediction method of mechanical properties of large-size coarse-grained soil based on progressive ensemble learning

Abstract

The present invention discloses a method for predicting the mechanical properties of large-size coarse-grained soil based on progressive ensemble learning, which belongs to the field of computer application technology. The method comprises the following steps: Step 1: Collecting data from triaxial experiments, classifying, filtering and completing the data of coarse-grained soils of various particle sizes; Step 2: Based on the data set classified by the maximum particle size obtained in Step 1, training the model group in sequence; Step 3: According to the model group in Step 2, combined with the characteristic data of the coarse-grained soil to be predicted, the prediction result of its mechanical properties is obtained. The method of the present invention not only has better performance in predicting the mechanical properties of coarse-grained soil with a smaller maximum particle size, but also can predict the mechanical properties of coarse-grained soil with a larger maximum particle size, and the result has good accuracy and generalization.

Description

Technical Field

[0001] The invention relates to the field of computer application technology, and in particular to a method for predicting mechanical properties of large-size coarse-grained soil based on progressive ensemble learning. Background Art

[0002] Coarse-grained soil refers to a mixture of coarse particles such as boulders, crushed stones (or gravel), stone chips, and stone powder. It is the main building material for earth-rock dams, so it is extremely important to master its mechanical properties for the design of earth-rock dams. There are many factors that affect the properties of coarse-grained soil, among which the proportion of particles of various sizes, called gradation, is one of the most important influencing factors. In the actual coarse-grained soil prototype gradation, the maximum particle size can reach 600-800mm, while the maximum particle size supported by general indoor triaxial instruments is only 60mm, so its mechanical properties cannot be directly measured.

[0003] At present, the only way to reduce the gradation in engineering is to use some methods, called scaling, which is generally to scale it down to a maximum particle size of 60 mm for testing, and then infer the characteristics of the prototype gradation from this data. This will bring inevitable errors, which is called the scaling effect. At present, in the process of inferring the mechanical properties of large-size soil materials on site from the characteristics of test soil materials, traditional mathematical formula models are generally used. This type of formula model is based on certain assumptions, has a large number of parameters, and the parameters come from a small number of test results, and does not have good accuracy, interpretability, and generalization.

[0004] In summary, in the study of mechanical properties of large-size coarse-grained soil, there is a lack of methods based on a large amount of real data that can accurately and generally predict the mechanical properties of coarse-grained soil according to parameters such as gradation, dry density, porosity, etc. Summary of the invention

[0005] The purpose of the present invention is to propose a method for predicting the mechanical properties of large-size coarse-grained soil based on progressive ensemble learning, which is characterized by comprising the following steps:

[0006] Step 1: Collect the data of triaxial test, classify, filter and complete the data of coarse-grained soil of each particle size;

[0007] Step 2: Based on the data set classified by the maximum particle size obtained in step 1, the model groups are trained in sequence;

[0008] Step 3: Based on the model group in step 2, combined with the characteristic data of the coarse-grained soil to be predicted, the prediction results of its mechanical properties are obtained.

[0009] The step 1 includes the following sub-steps:

[0010] Step A1: Obtain triaxial test data, group and classify them according to the maximum particle size, and determine the input characteristic parameters and output labels of the data;

[0011] Step A2: First, review the data grouped and classified in step A1, and exclude the data that does not meet the triaxial test standards; then normalize the data according to the proposed standards, and fill in some missing values ​​with geotechnical formulas; finally, fill in the missing values ​​of the input feature parameters with the average value of the corresponding category data set, and use hyperbola fitting to fill and make all output labels fall within the specified range.

[0012] The step 2 includes the following sub-steps:

[0013] Step B1: The data set with the smallest maximum particle size obtained by processing in step 1 is used to train the XGBoost model, and the obtained model group for predicting the maximum particle size is used as the base model group;

[0014] Step B2: First, the data set of the maximum particle size is updated by the base model group, and a neural network model is trained based on the data set and added to the model group; then a new base model is trained with the new label of the data set, and the new base model is pruned and restored; finally, the i-th base model is added to the i-1-th model group as the i-th model group to predict the mechanical properties of coarse-grained soil with the i-th maximum particle size; the above process is repeated until the corresponding maximum particle size of the model group meets the requirements.

[0015] The input characteristic parameters in step A1 include maximum particle size, test confining pressure, container diameter, container height, relative density, dry density, void ratio and gradation.

[0016] The output label in step A1 is deviatoric stress or volumetric strain.

[0017] The calculation formula for the new label of the data set in step B2 is as follows:

[0018] y new i =y i -y pred i-1

[0019] Among them, y new i is the new label of the dataset with the i-th largest particle size; y i is the original label of the dataset with the i-th largest particle size; y pred i-1 It is the prediction result of the i-1th model group for the label of the i-th maximum particle size data set.

[0020] The calculation formula of the prediction result in step 3 is as follows:

[0021] y pred i =y pred i-1 +y pi

[0022]

[0023] Among them, y pred i is the predicted value of the mechanical properties of the i-th model group for the i-th maximum particle size coarse-grained soil sample; y pred i-1 is the predicted value of the mechanical properties of the i-th maximum particle size coarse-grained soil sample by the i-1th model group; y p i is the predicted value of the i-th base model; y p n is the predicted value of the nth base model.

[0024] The beneficial effects of the present invention are:

[0025] Compared with the traditional formula model method, the method of the present invention not only has better performance in predicting the mechanical properties of coarse-grained soil with a smaller maximum particle size, but can also predict the mechanical properties of coarse-grained soil with a larger maximum particle size, and the results have good accuracy and generalization; in addition, the mechanical properties of coarse-grained soils with various particle sizes can be predicted as needed, and the method has high practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of the method for predicting mechanical properties of large-size coarse-grained soil based on progressive ensemble learning of the present invention;

[0027] Figure 2 A pruning graph of the neural network model of the present invention;

[0028] Figure 3 Construct a flow chart for the model of the present invention;

[0029] Figure 4 Constructing an example diagram for the model of the embodiment of the present invention;

[0030] Figure 5 This is a display diagram of the missing label filling of the present invention;

[0031] Figure 6 This is a prediction diagram of the maximum particle size of XGB of the embodiment of the present invention being 60 mm;

[0032] Figure 7 This is the prediction diagram of the maximum particle size of XGB+NNs of the embodiment of the present invention being 80 mm;

[0033] Figure 8 This is the prediction diagram of the maximum particle size of XGB+NNs of the embodiment of the present invention being 160 mm;

[0034] Fig. 9This is a prediction diagram of the maximum particle size of XGB+NNs of the embodiment of the present invention being 200 mm. DETAILED DESCRIPTION

[0035] The present invention proposes a method for predicting mechanical properties of large-size coarse-grained soil based on progressive ensemble learning, which is further described below in conjunction with the accompanying drawings and specific embodiments.

[0036] This method is based on a large number of triaxial test data at home and abroad. The data of coarse-grained soils of various particle sizes are classified, filtered, and completed, and then the base model groups corresponding to each particle size are trained in sequence. Finally, the characteristic data of the coarse-grained soil to be predicted are combined to obtain the prediction results of its mechanical properties. Figure 1 The flowchart of the method for predicting mechanical properties of large-size coarse-grained soil based on progressive ensemble learning of the present invention comprises the following steps:

[0037] Step 1: Collect a large amount of triaxial test data from home and abroad, and classify, filter and complete the data of coarse-grained soils of various particle sizes; specifically, it includes the following sub-steps:

[0038] Sub-step A1: Obtain triaxial test data from home and abroad, and group and classify them by the maximum particle size. Since the biggest constraint in the measurement of mechanical properties of coarse-grained soil materials is their maximum particle size, the data is classified by the maximum particle size. The data format is determined by research, and the maximum particle size is divided into 10mm, 20mm, 40mm, 60mm, 80mm, 100mm, 130mm, 160mm and 200mm, and the data is classified according to the principle of proximity based on the maximum particle size. The input characteristic parameters of the data should include: maximum particle size, test confining pressure, container diameter, container height, relative density, dry density, void ratio and gradation, where gradation is the weight proportion of particles of different particle sizes in coarse-grained soil materials, which are divided into 0-5mm, 5-10mm, 10-20mm, 20-40mm, 40-60mm, 60-80mm, 80-100mm, 100-130mm, 130-160mm and 160-200mm, a total of 10 items, and the output label of the data should be deviatoric stress. That is, the final data set consists of data records in the above format, as shown in Tables 1 and 2.

[0039] Table 1 Characteristics table

[0040]

[0041] Table 2 Label table

[0042]

[0043] This example selects data sets with maximum particle sizes of 60 mm, 80 mm, 160 mm, and 200 mm, among which the data volume of 60 mm is relatively large, while the data volume of other parts is relatively small. The specific situation of the data sets is summarized in Table 3:

[0044] Table 3 Particle size data distribution table

[0045]

[0046] Sub-step A2: Filter and complete the coarse-grained soil data. Review the data classified in sub-step A1, exclude the data that does not meet the triaxial drainage test standards, and then normalize the data according to the proposed standards, and then fill in some missing values ​​with geotechnical formulas related to triaxial tests. Some formulas are shown below:

[0047]

[0048] Dr——relative density, ρ d ——Dry density of fill, ρ dmin ——Minimum dry density of material, ρ dmax ——Maximum dry density of the material.

[0049]

[0050] ρ d ——Dry density of fill, ρ——density, ω——water content, G s ——Specific gravity of soil particles, ρ w ——water density, e——porosity.

[0051] Then, the missing values ​​are filled with the average value of the corresponding category data set features. The effect of partial data before and after filling is shown in Table 4. For the label output curve, the best nonlinear fitting function is used to fit and fill so that all labels are within the specified range. The effect of partial data before and after filling is shown in Figure 5 , some data are shown in Table 5.

[0052] Table 4 Partial data missing filling table

[0053]

[0054]

[0055] Table 5 Label missing filling data

[0056]

[0057] Step 2: Import the output of step 1, and train the model groups based on the data set classified by the maximum particle size;

[0058] Step 2 includes the following sub-steps:

[0059] Sub-step B1: From the data set processed in step 1, select the data set with the smallest maximum particle size, that is, the data set with a maximum particle size of 60 mm, and use this to train a 60 mm model group. That is, the 60 mm base model is trained using the data set with a maximum particle size of 60 mm, and a separate model group with a maximum particle size of 60 mm is formed. In addition, the inventors used the prediction results of the 60 mm model group, combined with the mean square error and relative error to evaluate the model's fitting effect on each particle size data, which were 22.11285 and 1.28951995% respectively. The results are as follows: Figure 6 shown.

[0060] Sub-step B2: For the model group corresponding to the second maximum particle size of 80 mm, use the previous maximum particle size of 60 mm model group to calculate the label of this maximum particle size 80 mm data set. The new label calculation formula is:

[0061] y neww 80 =y 80 -y pred 60

[0062] y new 80 ——New label with maximum particle size of 80mm; 80 ——Original label with maximum particle size of 80 mm; prea 60 ——Prediction values ​​of the 60mm model group using the 80mm data set as input.

[0063] Based on the features of the data set and the calculated new labels, the 80mm base model is trained. Here, the model is a neural network model. The training stops after reaching the set maximum number of iterations, and the model is compressed, i.e., the model is pruned and the training is resumed. The structure of the 80mm base model is as follows: Figure 2 As shown in , the number of neurons in the input layer is 16, the number of neurons in the two hidden layers are 45 and 20 respectively, and the output layer consists of 1 neuron. Figure 3 and Figure 4 As shown in the figure, the compression process is an iteration of model pruning and recovery training, and the number of iterations is manually set. After the model is trained, it is pruned, that is, about 60% of the weights with smaller weights are discarded, and then recovery training is performed, that is, the model weights are fine-tuned to restore the model performance. This process is repeated and stopped after the maximum number of iterations is reached to obtain the 80mm base model. Model compression makes the model structure simpler without losing performance, so as to improve the generalization ability and reduce the risk of overfitting caused by the small amount of data. Finally, the 80mm base model is added to the original 60mm model group to form a model group corresponding to the maximum particle size of 80mm. This model group is used to predict the mechanical properties of coarse-grained soil with a maximum particle size of 80mm. The calculation formula of the prediction result is as follows:

[0064] y pred 80 =y pred 60 +y p 80

[0065] y pred 80 =y p 60 +y p 80

[0066] y pred 60 ——The predicted value of the mechanical properties of the coarse-grained soil sample with a maximum particle size of 80 mm by the 60 mm model group; pred 80 ——The predicted value of the mechanical properties of the 80mm model group for the coarse-grained soil sample with a maximum particle size of 80mm; p 80 ——Prediction value of 80mm base model; y p 60 ——Prediction value of 60mm base model.

[0067] Repeat sub-step B2: For the model group corresponding to the third maximum particle size of 160 mm, use the previous maximum particle size of 80 mm model group to calculate the label of this maximum particle size of 160 mm data set. The new label calculation formula is:

[0068] y new 160 =y 160 -y pred 80

[0069] y new 160 ——New label with maximum particle size of 160mm; 160 ——Original label with maximum particle size of 160 mm; pred 80 ——Prediction values ​​of the 80mm model group using the 160mm data set as input.

[0070] Based on the features of the data set and the calculated new labels, the 160mm base model is trained. Here, the model is a neural network model. The training stops after reaching the set maximum number of iterations, and the model is compressed, i.e., the model is pruned and the training is resumed. The structure of the 160mm base model is as follows: Figure 2As shown, the number of neurons in the input layer is 16, the number of neurons in the two hidden layers is 40 and 35 respectively, and the output layer consists of 1 neuron. The compression process is an iteration of model pruning and recovery training, and the number of iterations is manually set. After the model is trained, it is pruned, that is, about 60% of the weights with smaller weights are discarded, and then recovery training is performed, that is, training and fine-tuning the model weights to restore the model performance. This process is repeated and stopped after the maximum number of iterations is reached to obtain the 160mm base model. Model compression makes the model structure simpler without losing performance, so as to improve the generalization ability and reduce the risk of overfitting caused by the small amount of data. Finally, the 160mm base model is added to the original 80mm model group to form a model group corresponding to the maximum particle size of 160mm. This model group is used to predict the mechanical properties of coarse-grained soil with a maximum particle size of 160mm. The calculation formula for the prediction result is as follows:

[0071] y pred 160 =y pred 80 +y p 160

[0072] y pred 160 ——The predicted value of mechanical properties of coarse-grained soil samples with a maximum particle size of 160 mm by the 160 mm model group; pred 80 ——The predicted value of the mechanical properties of the coarse-grained soil sample with a maximum particle size of 160 mm by the 80 mm model group; p 160 ——Prediction value of 160mm base model.

[0073] Repeat sub-step B2: For the fourth model group corresponding to the maximum particle size of 200 mm, use the previous model group with the maximum particle size of 160 mm to calculate the label of this maximum particle size of 200 mm data set. The new label calculation formula is:

[0074] y new 200 =y 200 -y pred 160

[0075] y new 200 ——New label with maximum particle size of 200mm; 200 ——Original label with maximum particle size of 200 mm; pred 160 ——Prediction values ​​of the 160mm model group using the 200mm data set as input.

[0076] Based on the features of the data set and the calculated new labels, the 200mm base model is trained. Here, the model is a neural network model. The training stops after reaching the set maximum number of iterations, and the model is compressed, i.e., the model is pruned and the training is resumed. The structure of the 200mm base model is as follows: Figure 2 As shown in the figure, the number of neurons in the input layer is 16, the number of neurons in the two hidden layers is 45 and 30 respectively, and the output layer consists of 1 neuron. The compression process is an iteration of model pruning and recovery training, and the number of iterations is manually set. After the model is trained, it is pruned, that is, about 60% of the weights with smaller weights are discarded, and then recovery training is performed, that is, training and fine-tuning the model weights to restore the model performance. This process is repeated and stopped after the maximum number of iterations is reached to obtain a 200mm base model. Model compression makes the model structure simpler without losing performance, so as to improve the generalization ability and reduce the risk of overfitting caused by less data. Finally, the 200mm base model is added to the original 160mm model group to form a model group corresponding to the maximum particle size of 200mm. This model group is used to predict the mechanical properties of coarse-grained soil with a maximum particle size of 200mm. The calculation formula for the prediction result is as follows:

[0077] y pred 200 =y pred 160 +y p 200

[0078] y pred 200 ——The predicted value of the mechanical properties of the 200mm model group for coarse-grained soil samples with a maximum particle size of 200mm; pred 160 ——The predicted value of mechanical properties of coarse-grained soil samples with a maximum particle size of 200 mm by the 160 mm model group; p 200 ——Prediction value of 200mm base model group.

[0079] The construction process of step 1 and step 2 of this embodiment is as follows Figure 4 shown.

[0080] Step 3 uses the model group trained in step 2 to import the coarse-grained soil material information to be predicted in the characteristic parameter format, obtains the prediction results and displays them in a drawing.

[0081] Start with the 60mm model group corresponding to sub-step 1, such as Figure 6 As shown in the figure, the 80 mm model group, 160 mm model group, and 200 mm model group are obtained in turn. The fitting ability of each model group for the mechanical properties of the coarse-grained soil with the corresponding maximum particle size is as follows Figure 7 , Figure 8 , Fig. 9 The pseudo code of sub-step B2 is shown in Table 6.

[0082] Table 6 Pseudo code for sub-step B2

[0083]

[0084] Compared with the traditional formula model method, this embodiment not only has better performance in predicting the mechanical properties of coarse-grained soil with a smaller maximum particle size, but can also predict the mechanical properties of coarse-grained soil with a larger maximum particle size, which is not available in the traditional formula model, and the results have good accuracy and generalization.

[0085] This embodiment is only a preferred specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for predicting mechanical properties of large-size coarse-grained soil based on progressive ensemble learning, characterized in that: The following steps are involved: Step 1: Collect the data of triaxial test, classify, filter and complete the data of coarse-grained soil of each particle size; The step 1 includes the following sub-steps: Step A1: Obtain triaxial test data, group and classify them according to the maximum particle size, and determine the input characteristic parameters and output labels of the data; The input characteristic parameters in step A1 include maximum particle size, test confining pressure, container diameter, container height, relative density, dry density, void ratio and gradation; The output label in step A1 is deviatoric stress or volumetric strain; Step A2: First, review the data grouped and classified in step A1, and exclude the data that does not meet the triaxial test standards; then normalize the data according to the proposed standards, and fill in some missing values ​​with geotechnical formulas; finally, fill in the missing values ​​of the input characteristic parameters with the average value of the corresponding category data set, and fill in all output labels with hyperbolic fitting to reach the specified range; Step 2: Based on the data set classified by the maximum particle size obtained in step 1, the model groups are trained in sequence; The step 2 includes the following sub-steps: Step B1: The data set with the smallest maximum particle size obtained by step 1 is used to train the XGBoost model, and the obtained model group for predicting the maximum particle size is used as the base model group; Step B2: First, the data set of the maximum particle size is updated by the base model group, a neural network model is trained based on the data set and added to the model group; then a new base model is trained with the new label of the data set, and the new base model is pruned and restored; finally, the obtained i-th base model is added to the i-1-th model group as the i-th model group to predict the mechanical properties of the coarse-grained soil with the i-th maximum particle size; the above process is repeated until the corresponding maximum particle size of the model group meets the requirements; The calculation formula for the new label of the data set in step B2 is as follows: ， in, is the new label of the dataset with the i-th largest particle size; is the original label of the dataset with the i-th largest particle size; is the prediction result of the i-1th model group for the label of the i-th maximum particle size data set; Step 3: Based on the model group in step 2, combined with the characteristic data of the coarse-grained soil to be predicted, the prediction results of its mechanical properties are obtained; The calculation formula of the prediction result in step 3 is as follows: ， in, is the predicted value of the mechanical properties of the i-th model group for the i-th coarse-grained soil sample with the largest particle size; is the predicted value of the mechanical properties of the i-th maximum particle size coarse-grained soil sample by the i-1th model group; is the predicted value of the i-th base model; is the predicted value of the nth base model.

Images