Prediction method of mechanical properties of coarse-grained soil based on sequence model

Abstract

The present invention discloses a method for predicting the mechanical properties of coarse-grained soil based on a sequence model, which belongs to the field of computer application technology. The method includes step 1: collecting coarse-grained soil experimental data, recording the data characteristics according to the model requirements, and then organizing and filling the data to divide it into a training set and a test set; step 2: based on the data of step 1, training the machine learning model until convergence; step 3: based on the converged model of step 2, inputting the coarse-grained soil data to be predicted to obtain the prediction results; displaying the prediction curve through a visualization module, and then setting evaluation indicators to judge the final prediction effect. The present invention can accurately predict the mechanical characteristics of coarse-grained soil and save the manpower, material resources, and financial resources required for experiments; for buildings using coarse-grained soil as the main material, the results obtained can be used as a safety reference and have strong practicality.

Description

Technical Field

[0001] The present invention relates to the technical field of computer application, and in particular to a method for predicting mechanical properties of coarse-grained soil based on a sequence model. Background Art

[0002] Dams are essential structures for retaining and draining water during hydropower development. A dam failure, storing billions of tons of water, would pose a severe threat to people and the ecosystem downstream, potentially more severe than a natural flood. Therefore, ensuring dam safety has always been a paramount consideration in dam construction.

[0003] In recent years, most dams built in my country are earth-rock dams, and coarse-grained soil has become the primary construction material for these dams. Coarse-grained soil, a granular material composed of interlocking rock particles (such as gravel and boulders), is a common material for high earth-rock dam construction. It is widely distributed in nature and exhibits excellent engineering properties, including good compaction properties, high shear strength, strong permeability, minimal settlement deformation, and high fill density. Because coarse-grained soil lacks cohesion, earth-rock dams constructed with it rely on shear resistance generated by dilatation, friction, crushing, and particle reorganization between particles. This inter-particle shear resistance is crucial for the dam's stability. Therefore, accurate and reasonable prediction of the coarse-grained soil's strength is crucial before dam construction to ensure its safety.

[0004] To study the engineering and mechanical properties of coarse-grained soils, conventional triaxial compression tests are used in laboratories. Currently, the commonly used indoor triaxial testing apparatus in China typically has a diameter of 30 cm and a height of 60 cm. To ensure that the test results are not affected by the sample, the maximum sample particle size must be less than one-fifth of the specimen diameter, meaning the maximum allowable particle size is 60 mm. However, in actual engineering construction, large-scale coarse materials, such as sand, gravel, and rockfill, are widely used, with particle sizes ranging from 600 mm to 800 mm. Using oversized instruments is time-consuming, labor-intensive, and expensive, resulting in minimal experimental data. To conduct tests using conventional triaxial apparatus, the original material is scaled, and the results from this scaled-down test are then used to infer the mechanical properties of the prototype material. Scale effects are unavoidable in testing rockfill materials for earth-rockfill dams, and the mechanical properties of the prototype material obtained through this scaled-down method are not accurate enough. Therefore, how to use the mechanical properties of small-size coarse-grained soils obtained through conventional indoor triaxial testing to infer the mechanical properties of large-size coarse-grained soils has become an urgent challenge in rockfill material research.

[0005] Many scholars have constructed soil models that can describe the mechanical properties of coarse-grained soil through mathematical formulas, such as the elastoplastic constitutive model, the Duncan-Zhang model, and the nonlinear elastic shear-tensile KG model. These models have limitations in simulating the mechanical properties of coarse-grained soil: (1) The mathematical formulas of all models are derived based on a certain experimental data, which limits the model's predictive ability for other different tests. (2) The mathematical formulas are becoming more and more complex, and the number of parameters is also increasing. In recent years, machine learning has been very effective in solving nonlinear problems, and many effective algorithms have emerged. Combined with the situation of the triaxial compression test process, which itself contains time series characteristics, a machine learning algorithm with time series can be used to predict the experimental results.

[0006] In summary, in the current research on the engineering and mechanical properties of coarse-grained soils, experiments cannot fully reproduce the actual situation, and the experimental results after using the scaling effect cannot accurately reflect the actual situation. However, the use of machine learning methods combined with the temporal characteristics of the experiment itself can better solve the above problems. Summary of the Invention

[0007] The present invention aims to propose a method for predicting the mechanical properties of coarse-grained soil based on a sequence model, which is characterized by comprising the following steps:

[0008] Step 1: Collect coarse-grained soil experimental data, record the data characteristics according to the needs of the sequence model, organize and fill the data, and divide it into training set and test set;

[0009] Step 2: Based on the data from step 1, train the machine learning model until convergence;

[0010] Step 3: Based on the convergence model of step 2, input the coarse-grained soil data to be predicted and obtain the prediction results; display the prediction curve through the visualization module, and then set evaluation indicators to judge the final prediction effect.

[0011] The data characteristics in step 1 include experimental confining pressure, container diameter, maximum particle size, container height, relative density, dry density, porosity, fine particle size ratio, crushing rate and experimental gradation.

[0012] The step 1 includes the following sub-steps:

[0013] Step A1: Using every 0.3% change in axial strain as the segmentation criterion, the deviatoric stress and axial strain curves were divided into 50 points. The deviatoric stress variation during the experiment, when the axial strain varied from 0% to 15%, was presented as a point set, with the axial strain used as a column feature. The gradation curve was segmented by particle size less than 5 mm, 5 mm to 10 mm, and 10 mm to 20 mm, with the proportion of each particle size in the total material used as a feature, and the deviatoric stress used as the data label.

[0014] Step A2: Remove outliers and fill in missing values ​​using the mean filling method or civil engineering formula; for missing labels, select a nonlinear fitting function based on the R value and the image to fit;

[0015] Step A3: Divide the curve obtained in each experiment in step A1 into 50 point sets, with each curve as a set of data, and divide all collected data into training set and test set.

[0016] The ratio of training set to test set is 8:2.

[0017] The nonlinear fitting function in step A2 is:

[0018]

[0019] Where Dr is the relative density, ρ d is the dry density of fill, ρ dmin is the minimum dry density of the material, ρ dmax is the maximum dry density of the material.

[0020] The step 2 includes the following sub-steps:

[0021] Step B1: Set the sliding window according to the mechanical characteristic curve of the coarse-grained soil to be predicted;

[0022] Step B2: Select the bidirectional LSTM model as the main model and use the post-attention mechanism to improve the prediction effect. By using the self-attention mechanism to enhance the importance of key features, the most important features that affect the experimental results are obtained.

[0023] Step B3: Select a loss function and perform iterative training to obtain a converged model.

[0024] The step size of the sliding window in step B1 is 3.

[0025] The beneficial effects of the present invention are:

[0026] The present invention can accurately predict the mechanical characteristics of coarse-grained soil and save the manpower, material resources and financial resources required for the experiment; for buildings using coarse-grained soil as the main material, the obtained results can be used as a safety reference and have strong practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a model architecture diagram of the method for predicting mechanical properties of coarse-grained soil based on a sequence model according to the present invention;

[0028] Figure 2 is the fitting curve graph obtained using the best fitting function;

[0029] Figure 3 Visualize weight distribution diagrams for the most important features affecting triaxial compression test results;

[0030] Figure 4 The prediction curve graph displayed by the visualization module. DETAILED DESCRIPTION

[0031] The present invention proposes a method for predicting the mechanical properties of coarse-grained soil based on a sequence model, which will be further described below with reference to the accompanying drawings and specific embodiments.

[0032] Figure 1 This is a model architecture diagram of the sequence model-based method for predicting the mechanical properties of coarse-grained soil. This prediction method uses indoor triaxial compression test data to obtain the required features and labels. After preprocessing the data, it is divided into a training set and a test set. The training set is input into the constructed model. After iterative training, a trained model is obtained. The test set is then input into the trained model to obtain the final prediction value. Specifically, the following steps are included:

[0033] Step 1: Collect coarse-grained soil experimental data. Organize and fill in the data. Step 1 includes the following sub-steps:

[0034] Sub-step A1: Collect triaxial compression test data for coarse-grained soil from major domestic and international research institutes and projects. Presented here are experimental data from static property tests and stress-strain calculations of concrete-faced rockfill dam materials at the Hanshan Reservoir, China Institute of Water Resources and Hydropower Research. Based on the model's requirements, experimental data such as the experimental confining pressure, container diameter, maximum particle size, container height, relative density, dry density, void ratio, fine particle size fraction, crushing rate, and experimental gradation are recorded as features. The deviatoric stress (q) versus axial strain (εa) curve is segmented at every 0.3% change in axial strain (εa). This mechanical characteristic curve is divided into 50 points, representing the change in deviatoric stress for axial strain changes of 0% to 15% during the experiment. The axial strain is used as a column of features. The gradation curve is segmented based on particle size, with the proportion of the total material less than 5mm, 5mm-10mm, and 10mm-20mm being used as features. Deviatoric stress is used as a data label. The extracted data set is shown in Table 1:

[0035] Table 1

[0036]

[0037] Sub-step A2: Clean and fill the data according to the needs of the model. Remove outliers and use the mean filling method or civil engineering formula to fill in the missing values. For missing labels, a nonlinear fitting function is used here. The best fitting function is selected based on the R value and the image. The fitting effect is as follows: Figure 2 The parameters are shown in Table 2, where the points are original data and the curves are fitted curves. The civil engineering formula is:

[0038]

[0039] Among them, Dr is relative density, ρ d ——Dry density of fill, ρ dmin ——Minimum dry density of material, ρ dmax ——The maximum dry density of the material.

[0040] Table 2

[0041]

[0042] Sub-step A3: Divide the training set into a test set. Based on Sub-step A1, the curve obtained from each experiment can be divided into 50 point sets. Each curve is considered a data set, and all collected data is divided into a training set and a test set in a ratio of 8:2.

[0043] Step 2: Based on the data from step 1, train the machine learning model until convergence. Step 2 includes the following sub-steps:

[0044] Sub-step B1: Set a sliding window based on the mechanical characteristic curve of the coarse-grained soil to be predicted. Because the mechanical characteristic curve to be predicted has uniform variations, it is interpreted as sequential data. Set the sliding step size to 3. Since the training set contains 240 sets of experimental data, after the sliding step size segmentation in Sub-step B1, the training set has a total of 11,280 sets of data and 14 sets of features.

[0045] Sub-step B2: Select the bidirectional LSTM model as the main model and use the post-attention mechanism to improve prediction performance. The self-attention mechanism is used to enhance the importance of key features.

[0046] Calculate the output matrix v of the attention layer according to the formula

[0047] q=inputs·w t

[0048] k=inputs·w x

[0049] e=tanh(q+k)·w a

[0050] v = softmax(e)·inputs

[0051] Among them, w x , w t , w aIt is obtained by linear transformation of the feature part in the inputs and the manually set hyperparameter units.

[0052] During the calculation process, the content vector with assigned attention weights can be obtained, and the weight value a=softmax(e) assigned to each feature can be obtained.

[0053] Visualizing the weight matrix can reveal the most important features that affect the experimental results in the triaxial compression test, such as Figure 3 shown.

[0054] Sub-step B3: Determine the loss function as REMSE, and the calculation formula is as follows:

[0055]

[0056] After iterative training, a trained model is finally obtained.

[0057] Step 3: Based on the model in step 2, input the coarse-grained soil data to be predicted and obtain the prediction results. Step 3 includes the following sub-steps:

[0058] Sub-step C1: The test set obtained from sub-step A3 has a total of 59 groups of experimental data. After the test set is split according to the sliding step size of 3, there are a total of 2773 groups of data, which are sequentially input into the model to obtain the final prediction value.

[0059] Sub-step C2: Display the prediction curve through the visualization module and set the evaluation index to judge the final prediction effect. Use x to represent the actual value and the dot to represent the predicted value. Figure 4 shown.

[0060] This example uses collected triaxial compression test data from coarse-grained soil to organize it into the input features required by a machine learning model. This data is then used to train the model until convergence, resulting in a model that can predict the engineering and mechanical properties of coarse-grained soil. The test selected multiple sets of non-training data samples, and the obtained results were compared with the original experimental results to derive multiple evaluation criteria. The results demonstrate that this example can accurately predict the mechanical properties of coarse-grained soil, demonstrating its high practical value.

[0061] This embodiment is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for predicting mechanical properties of coarse-grained soil based on a sequence model, characterized in that: The steps include: Step 1: Collect coarse-grained soil experimental data, record data features according to the needs of the sequence model, organize and fill the data, and divide them into training and test sets; the data features in step 1 include experimental confining pressure, container diameter, maximum particle size, container height, relative density, dry density, void ratio, fine particle size ratio, crushing rate, and experimental gradation; The step 1 includes the following sub-steps: Step A1: Using every 0.3% change in axial strain as the segmentation criterion, the deviatoric stress and axial strain curves were divided into 50 points. The deviatoric stress changes during the experiment when the axial strain changes from 0% to 15% were displayed as a point set, with the axial strain used as a column feature. The gradation curve was segmented by particle size less than 5 mm, greater than or equal to 5 mm and less than 10 mm, and greater than or equal to 10 mm and less than or equal to 20 mm. The proportion of each particle size in the total material was used as a feature, with the deviatoric stress used as the data label. Step A2: Remove outliers and fill in missing values ​​using the mean filling method or civil engineering formula; for missing labels, select a nonlinear fitting function based on the R value and the image to fit; The nonlinear fitting function in step A2 is: Where Dr is the relative density, ρ d is the dry density of fill, ρ dmin is the minimum dry density of the material, ρ dmax is the maximum dry density of the material; Step A3: Divide the curve obtained in each experiment in step A1 into 50 point sets, with each curve as a set of data, and divide all the collected data into a training set and a test set; Step 2: Based on the data from step 1, train the machine learning model until convergence; The step 2 includes the following sub-steps: Step B1: Set the sliding window according to the mechanical characteristic curve of the coarse-grained soil to be predicted; Step B2: Select the bidirectional LSTM model as the main model and use the post-attention mechanism to improve the prediction effect. By using the self-attention mechanism to enhance the importance of key features, the most important features that affect the experimental results are obtained. Step B3: Select a loss function and perform iterative training to obtain a converged model; Step 3: Based on the convergence model of step 2, input the coarse-grained soil data to be predicted and obtain the prediction results; display the prediction curve through the visualization module, and then set evaluation indicators to judge the final prediction effect.

2. The method for predicting mechanical properties of coarse-grained soil based on a sequence model according to claim 1 is characterized in that: The ratio of training set to test set is 8:

2.

3. The method for predicting mechanical properties of coarse-grained soil based on a sequence model according to claim 1 is characterized in that: The step size of the sliding window in step B1 is 3.

Images