Method for predicting dynamic shear modulus of saturated coral sand based on interpretable artificial intelligence

Abstract

The invention discloses a saturated coral sand dynamic shear modulus prediction method based on interpretable artificial intelligence, and the method comprises the steps: preparing a saturated coral sand sample, carrying out the testing, generating a data set, building a saturated coral sand dynamic shear modulus prediction model through a CNN-TCN-Attention combined neural network, and training the model; and performing explanatory analysis on data of the prediction model by using an SHAP method. The method focuses on the particularity of the saturated coral sand soil texture and adopts a resonant column test to obtain a large amount of indoor test data, the problem of data scarcity in the deep learning prediction field is solved, meanwhile, the SHAP method can effectively solve the black box problem existing in a machine learning model, the interpretability of the prediction model is enhanced, and the prediction precision is improved. And the influence rule of each input variable on the output variable can be determined. The explainable deep learning model is proposed for the first time to achieve accurate prediction of the dynamic shear modulus of the saturated coral sand, and the reliability of the model is verified based on additional tests.

Description

Technical Field

[0001] The invention relates to the field of coral sand performance research and digital modeling, and in particular to a method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence. Background Art

[0002] Coral sand is a special marine sediment mainly distributed in tropical waters between 30° north and south latitude, accounting for about 40% of the seabed area. In recent years, with the vigorous development of offshore oil and wind energy resources and the construction of island and reef projects, many marine structures have been built on coral sand sites. These projects may cause immeasurable losses to offshore infrastructure when facing dynamic load risks such as earthquakes, waves and hurricanes. As an important component of island and coral reef foundation engineering, the dynamic deformation characteristics of coral sand are crucial to the safety of island construction projects.

[0003] The dynamic shear modulus of soil is a basic representation of the dynamic characteristics of soil and an important parameter required for site seismic response analysis. As an extremely important parameter in the seismic safety evaluation and dynamic response analysis of such projects, the rationality of its selection is directly related to the safety and economy of the engineering structure. Therefore, the research on the performance analysis and prediction of the dynamic shear modulus of saturated coral sand is of great significance to improving the safety and stability of island and reef engineering structures.

[0004] The resonant column test and the bending element test are two commonly used methods for testing the dynamic shear modulus G. The resonant column test obtains the resonant frequency of the sample by adjusting the excitation frequency, and indirectly calculates the shear wave velocity V of the sample based on the wave theory. s , and finally get 10 -6 ~10 -4 The dynamic shear modulus G of the shear strain magnitude. The bending element test directly tests the time it takes for the shear wave to propagate through the specimen, and then obtains the shear wave velocity V of the test specimen. s , thus obtaining the maximum dynamic shear modulus of soil G max The two test methods have their own emphasis. The data from a single test cannot fully and accurately reflect the mechanical properties of the soil. A combination of multiple tests is required to jointly reflect the characteristic parameters of the soil sample.

[0005] However, the combination of multiple tests will generate a large amount of test data, and the test data may have problems, which makes scientific research complicated and cumbersome. In addition, the prediction models and function forms proposed by a large number of studies are different, with problems such as specific applicable objects, single research conditions and reliance on multiple parameters. It is impossible to obtain the dynamic shear modulus in the entire strain range through one experimental device. It is necessary to obtain the dynamic shear modulus corresponding to different strains through empirical formula fitting. The existing empirical formula is only valid for a local range and cannot fully fit the entire range of the test test.

[0006] At present, there are also many cases of using ML (Machine Learning) models to study the performance of coral sand. However, on the one hand, traditional machine learning models are difficult to achieve accurate real-time predictions for high-dimensional nonlinear problems. On the other hand, they do not consider the interpretability of the prediction algorithm and cannot effectively reveal the contribution of each input variable to the output variable and the interaction between variables.

[0007] Therefore, how to make an accurate judgment on the influencing factors of the dynamic shear modulus and predict the dynamic shear modulus by using the correlation and degree of influence between the influencing factors is a technical problem that needs to be solved urgently. Summary of the invention

[0008] In order to address the shortcomings of the prior art, the present invention provides a method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence. A resonant column test is used to obtain a large amount of indoor test data to fill the data scarcity problem in the field of deep learning prediction. At the same time, a global explanatory analysis method is used to solve the black box problem existing in the machine learning model, enhance the interpretability of the prediction model, and determine the influence of each input variable on the output variable, thereby greatly improving the reliability of the prediction model.

[0009] The embodiment of the present invention provides the following solution:

[0010] The embodiment of the present invention provides a method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence, and the method comprises the following steps:

[0011] Step 1, preparing saturated coral sand sample;

[0012] Step 2: Use saturated specimens to conduct tests and generate a data set. The influencing factors of each sample in the data set include relative density, confining pressure, fine particle content and strain, and the influencing target is the dynamic shear modulus;

[0013] Step 3: Use CNN-TCN-Attention (convolutional neural network-time domain convolutional network-attention mechanism) combined neural network to establish a dynamic shear modulus prediction model for saturated coral sand;

[0014] Step 4: Use the data set to input the CNN-TCN-Attention prediction model for training, optimize the parameters of the saturated coral sand dynamic shear modulus prediction model, and obtain the optimized saturated coral sand dynamic shear modulus prediction model;

[0015] Step 5: Use the SHAP method to perform explanatory analysis on the data of the optimized saturated coral sand dynamic shear modulus prediction model.

[0016] In an optional embodiment, step 1 adopts dry static pressure method and saturation treatment to prepare a saturated coral sand sample.

[0017] In an optional embodiment, the test in step 2 adopts a resonant column test.

[0018] In an optional embodiment, the saturated coral sand dynamic shear modulus prediction model described in step three includes a CNN (Convolutional Neural Networks) module, a TCN (Temporal Convolutional Networks) residual module and an attention mechanism module.

[0019] In an optional embodiment, the CNN module includes 6 convolutional layers, 6 pooling layers, and 1 fully connected layer, and the TCN residual module includes a one-dimensional fully convolutional network layer, a causal convolutional layer, a dilated convolutional layer, and a residual connection layer.

[0020] In an optional embodiment, the processing flow of the saturated coral sand dynamic shear modulus prediction model is as follows:

[0021] S3.1, taking the influencing factors in the data set as input, performing convolution, pooling and activation processing through the CNN module to extract deep features;

[0022] S3.2, the deep features enter the TCN residual module for temporal processing to obtain the temporal deep features;

[0023] S3.3. The time-series deep features enter the attention mechanism module, and the feature vector with attention weight is obtained and output as the prediction result.

[0024] In an optional embodiment, in the data set, 70% of the samples are used as a training data set, and the remaining 30% of the samples are used as a testing data set.

[0025] The beneficial effects of the present invention based on its technical solution are:

[0026] According to the resonant column test, the present invention obtains the resonant frequency of the coral sand sample, so that the dynamic shear modulus G of the soil can be calculated by one-dimensional wave theory. Then, the variation law of the dynamic shear modulus G with the material parameters and the external loading condition parameters is obtained, and the parameters with significant influence are selected as the input variables of deep learning. Then, the CNN-TCN-Attention model is constructed by combining the deep feature mining advantages of the convolutional neural network and the effective feature screening of the time convolutional network and the attention mechanism to improve the prediction accuracy. The accurate prediction of the dynamic shear modulus G of coral sand is achieved through a large number of data sets generated by indoor experiments. Finally, on the basis of the hybrid intelligent model, the SHAP (SHapley Additive exPlanation, global explanatory analysis method) method is used to analyze the contribution of the input variables to the prediction of the dynamic shear modulus of saturated coral sand, and the interaction and correlation between the input variables are compared and analyzed. Compared with the previous technology, the present invention pays attention to the particularity of the saturated coral sand soil and uses the resonant column test to obtain a large amount of indoor test data to fill the data scarcity problem in the field of deep learning prediction. In addition, the SHAP method can effectively solve the black box problem of machine learning models, enhance the interpretability of prediction models, and determine the influence of each input variable on the output variable. This paper proposes for the first time an interpretable deep learning model to accurately predict the dynamic shear modulus of saturated coral sand, and verifies the reliability of the model based on additional experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 A schematic flow chart of a method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence provided by the present invention.

[0029] Figure 2 It is a schematic diagram of the hierarchical structure of the dynamic shear modulus prediction model for saturated coral sand.

[0030] Figure 3 It is a schematic diagram of the Pearson correlation matrix of sample space distribution.

[0031] Figure 4 It is a schematic diagram of the regression distribution between the predicted value and the actual value.

[0032] Figure 5 It is a schematic diagram of the global SHAP value swarm and feature importance.

[0033] Figure 6 It is a schematic diagram of feature interaction.

[0034] Figure 7 It is a schematic diagram of the mutual influence relationship between different characteristic parameters. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention belong to the scope of protection of the embodiments of the present invention.

[0036] This embodiment provides a method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence. Figure 1 , the method comprises the following steps:

[0037] Step 1: Prepare saturated coral sand sample.

[0038] In this embodiment, the test uses coral sand from a certain island reef in the Nansha Islands. The coral sand particles are mainly composed of aragonite and high-magnesium calcite, and CaCO3 accounts for more than 90%, which is a calcareous sand. Its specific gravity is measured to be 2.80. Generally, coral sand particles with a diameter less than 0.075 mm are regarded as fine particles, and the remaining particles are regarded as sand particles. The coral sand is placed in a 0.075 mm standard sieve for vibration, and fine particles and sand particles of different masses are uniformly mixed in different proportions to obtain coral sand soil samples with different fine sand contents Fc of 0-30%. The coral sand soil samples are treated by dry static pressure method, and then pre-saturated, fully saturated and consolidated to obtain saturated coral sand samples.

[0039] Step 2: Use saturated specimens to conduct tests and generate a data set. The influencing factors of each sample in the data set include relative density, confining pressure, fine particle content and strain, and the influencing target is the dynamic shear modulus.

[0040] This embodiment adopts a resonant column test, in which vibration is applied to a cylindrical saturated coral sand sample, the vibration frequency is changed to make it resonate, and the corresponding dynamic shear modulus G is determined by the corresponding natural frequency. During the experiment, the relative density, confining pressure, and fine particle content of the saturated coral sand sample are changed in turn, and the dynamic elastic modulus under different strains is recorded to form a data set.

[0041] Step 3: Use the CNN-TCN-Attention combined neural network to establish a prediction model for the dynamic shear modulus of saturated coral sand, referring to Figure 2, the prediction model of dynamic shear modulus of saturated coral sand includes a CNN module, a TCN residual module and an attention mechanism module. In this embodiment, the CNN module includes 6 convolutional layers, an activation layer, 6 pooling layers, and 1 fully connected layer. The TCN residual module includes a TCN convolutional layer, a weight normalization layer, a ReLU layer, a Dropout layer, and a 1×1 convolution kernel residual module layer. The attention mechanism module includes a permute layer, an Eense layer, and a Multiply layer. The number of convolution kernels of the 6 convolutional layers is 16, 32, 32, 64, 64, and 64, respectively, and the area size of the 6 pooling layers is 2×1. Set the Dropout layer parameter to 0.3, the learning rate to 0.01, and the number of training times to 100.

[0042] The processing flow of the dynamic shear modulus prediction model for saturated coral sand is as follows:

[0043] S3.1, taking the influencing factors in the data set as input, performing convolution, pooling and activation processing through the CNN module to extract deep features;

[0044] S3.2, the deep features enter the TCN residual module for temporal processing to obtain the temporal deep features;

[0045] S3.3. The time-series deep features enter the attention mechanism module, and the feature vector with attention weight is obtained and output as the prediction result.

[0046] Step 4: Use the data set to input the CNN-TCN-Attention prediction model for training, optimize the parameters of the saturated coral sand dynamic shear modulus prediction model, and obtain the optimized saturated coral sand dynamic shear modulus prediction model.

[0047] In this embodiment, a saturated sample is tested to generate a data set with a total of 660 samples. The influencing factors are shown in the following table:

[0048]

[0049] The Pearson correlation matrix of sample space distribution is as follows Figure 3 As shown in the figure, the highest correlation is only 0.029, which shows that there is no high correlation between any two variables, thus proving the independence of the variables. This means that the database of input design combination samples is reasonable and reliable for developing artificial intelligence models.

[0050] 70% of the samples are set as the training data set in the model training process, and the remaining 30% are used as the test data set. The predicted results are as follows Figure 4As shown in the figure, the R2 of the training set of the CNN-TCN-Attention model is 0.985 and 0.981, which are accurate enough for optimization. The p-value is used to test the difference in residual distribution between the training set and the test set, indicating whether the prediction error comes from the same distribution. A p-value less than 0.05 indicates that there is a significant difference in the residual distribution, indicating that there is a difference in model performance. If the p-value is greater than or equal to 0.05, the assumption that the residual distribution is the same cannot be denied, indicating that the model performance is similar. The p-value of the training set and the test set in this paper is 0.07, indicating that the performance of the two is close, and the model does not have overfitting, which further verifies the reliability of the algorithm.

[0051] Step 5: Use the SHAP method to perform explanatory analysis on the data of the optimized saturated coral sand dynamic shear modulus prediction model.

[0052] (1) Global explanatory analysis:

[0053] Reference Figure 5 , the characteristic parameters are sorted from top to bottom according to their importance in contributing to the dynamic shear modulus. The positive and negative SHAP values ​​on the horizontal axis represent the positive and negative impact on the output results. The larger the absolute value, the greater the impact. The color represents the value of the characteristic parameter, blue represents a lower value, and red represents a higher value. Therefore, Figure 5 It can not only show the importance of characteristic parameters, but also reveal the influence trend of characteristic changes on dynamic shear modulus. Figure 5 It can be seen that confining pressure is the most important parameter affecting the dynamic shear modulus. The greater the confining pressure, the greater the positive SHAP value, and the greater the positive effect on the dynamic shear modulus. Similarly, the SHAP value of relative density gradually becomes positive as the parameter value increases, which has a positive effect on the dynamic shear modulus. Strain and fine particle content have a negative effect, and the negative effect on the dynamic shear modulus increases with the increase of the parameter value, indicating that strain and fine particle content are negatively correlated with the dynamic shear modulus.

[0054] (2) Feature interaction analysis

[0055] Reference Figure 6 , SHAP interaction values ​​expand on the global interpretation, breaking down the contributions into main effects and interaction effects. You can use these values ​​to highlight and visualize interactions in your data. It can also be a useful tool to understand how your model makes predictions, Figure 6 (a) is a diagram of the interaction values, which enable fast and accurate pairwise interaction calculations of the parameters in the model. This will return a matrix for each prediction, with the main effects on the diagonal and the interaction effects outside the diagonal. These values ​​often reveal the interaction. In order to see how it affects the prediction of each sample, looking at all the predictions at the same time can understand the model's response to different features. Figure 6(b) The heat map shown. The left y-axis is the ranking of important features. Features are sorted from large to small according to their influence. The right y-axis is its visualization. The color depth in the image represents the size of the SHAP value, that is, the impact of the value under this feature on the model. The darker the color, the larger the absolute value of the SHAP value, and the greater the impact on the model. The top is the visualization of the prediction results of the model under these values. 460 data were selected from the training set for visualization, as shown in Figure 6 (b) Heatmaps can be used to effectively explore and understand the importance of features and the relationships between features in complex models, thereby improving the interpretability and explainability of the model. Figure 6 (a) The interaction value plot shows the interaction between the features. Specifically, the main effects of relative density and confining pressure are both positive, while fines content and strain are both negative. At high relative density, the value of the dynamic shear modulus of saturated coral sand will increase with increasing confining pressure, while the dynamic shear modulus will decrease with increasing strain. At high confining pressure, the value of the dynamic shear modulus will decrease with increasing fines content and strain. Figure 6 (b) It can be seen that high predictions (high values ​​in f(x) on the right) are associated with high relative density and high confining pressure (red). Low predictions (low values ​​in f(x) on the right) are associated with low relative confining pressure and low strain (blue). Heatmap of the situation and interaction value Figure 1 To.

[0056] (3) Interaction analysis between input variables:

[0057] Reference Figure 7 Schematic diagram of the mutual influence relationship between different characteristic parameters. Figure 7 (a), 7(b), 7(c), and 7(d) show the changes of SHAP value with the changes of relative density, confining pressure, fine grain content, and strain, respectively. The general rule of the influence of the four input parameters on the dynamic shear modulus of saturated coral sand is that the increase of relative density and confining pressure will increase the predicted value of dynamic shear modulus; the increase of fine grain content and strain will reduce the predicted value of dynamic shear modulus. Figure 7 It can be clearly seen in (a) that the relative density is proportional to the dynamic shear modulus. Figure 7 (d) It can be seen that the SHAP value of strain decreases with the increase of strain, and the decrease is rapid before 1e-4, and slows down after 1e-4. In short, the dynamic shear modulus of saturated coral sand can be improved by appropriately increasing the relative density and confining pressure values ​​and reducing the fine particle content and strain values. The strain value is controlled as much as possible above 1e-4, which can better maintain the dynamic shear modulus of saturated coral sand.

[0058] In order to provide additional evidence for the effectiveness of the CNN-TCN-Attention prediction model proposed in this study, four models, including CNN, TCN, CNN-TCN and TCN-Attention, were used to build a prediction model to predict the dynamic shear modulus, and the results were compared with the CNN-TCN-Attention prediction results. The prediction effect of each model on the training set and test set is shown in the following table.

[0059]

[0060] Whether in training or testing sets, the CNN-TCN-Attention prediction model has higher prediction accuracy than the other four traditional models. 2 The value reached 0.985, and the R 2 The value reaches 0.981. The prediction accuracy of the CNN-TCN-Attention model is the highest, and the prediction accuracy of the CNN model is the weakest. The accuracy of the CNN-TCN model is higher than that of the TCN because the prediction accuracy and feature extraction ability of the traditional TCN network model are poor, and the deep feature mining advantage of CNN has greatly improved the prediction accuracy of the model. It is more suitable for rolling bearing residue. In addition, by comparing the algorithm results, it can be seen that the prediction results of CNN-TCN-Attention are better than those of CNN-TCN, the final values ​​of different errors are lower, and the prediction results are more accurate. This is because the CNN-TCN-Attention model can effectively extract key feature information and increase the correlation between the model input and output through the combination of TCN neural network and attention mechanism. It shows higher accuracy in dealing with the nonlinear problem of predicting the dynamic shear modulus of saturated coral sand. In short, the CNN-TCN-Attention model proposed in this paper has better performance in data prediction tasks, and the accuracy is significantly improved compared with the traditional neural network model.

[0061] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (modules, systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0063] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0064] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0065] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0066] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence, characterized in that: The method comprises the following steps: Step 1, preparing saturated coral sand sample; Step 2: Use saturated coral sand samples to conduct experiments and generate a data set. The influencing factors of each sample in the data set include relative density, confining pressure, fine particle content and strain, and the influencing target is the dynamic shear modulus; Step 3: Use the CNN-TCN-Attention combined neural network to establish a prediction model for the dynamic shear modulus of saturated coral sand; Step 4: using the data set to input the saturated coral sand dynamic shear modulus prediction model for training, optimizing the parameters of the saturated coral sand dynamic shear modulus prediction model, and obtaining an optimized saturated coral sand dynamic shear modulus prediction model; Step 5: Use the SHAP method to perform explanatory analysis on the data of the optimized saturated coral sand dynamic shear modulus prediction model.

2. The method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence according to claim 1, characterized in that: Step 1: Prepare saturated coral sand samples by dry static pressure method and saturation treatment.

3. The method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence according to claim 1, characterized in that: The experiment described in step 2 uses a resonant column experiment.

4. The method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence according to claim 1, characterized in that: The saturated coral sand dynamic shear modulus prediction model described in step three includes a CNN module, a TCN residual module and an attention mechanism module.

5. The method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence according to claim 4, characterized in that: The CNN module includes 6 convolutional layers, 6 pooling layers, and 1 fully connected layer, and the TCN residual module includes a one-dimensional fully convolutional network layer, a causal convolutional layer, a dilated convolutional layer, and a residual connection layer.

6. The method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence according to claim 4, characterized in that: The processing flow of the saturated coral sand dynamic shear modulus prediction model is as follows: S3.1, taking the influencing factors in the data set as input, performing convolution, pooling and activation processing through the CNN module to extract deep features; S3.2, the deep features enter the TCN residual module for temporal processing to obtain the temporal deep features; S3.

3. The time-series deep features enter the attention mechanism module, and the feature vector with attention weight is obtained and output as the prediction result.

7. The method for predicting the dynamic shear modulus of saturated coral sand based on explainable artificial intelligence according to claim 1, characterized in that: In the data set, 70% of the samples are used as training data sets, and the remaining 30% of the samples are used as testing data sets.

Images