Cptu soil layering method based on machine learning and uncertainty analysis

Abstract

The present application relates to a kind of CPTU soil layering method based on machine learning and uncertainty analysis, comprising the following steps, S1.CPTU data of marine soil and the soil drilling data of CPTU survey position adjacent are obtained;S2.classification model is trained;S3.CPTU data of target position is input into the classification model trained in S2, obtain classification result, and the classification result is evaluated using model accuracy evaluation index;S4.combining uncertainty analysis, the confidence of each classification result is evaluated, and uncertainty analysis result is obtained;S5.according to uncertainty analysis result, the number of soil layering and demarcation position are determined, and layering result is output.The method combines CPTU high-resolution data and machine learning algorithm, by introducing uncertainty analysis, the automatic layering method of class confidence optimization is optimized to the classification result, so as to obtain the soil layer division result with engineering reference value.

Description

Technical Field

[0001] This invention relates to the field of soil stratification technology in marine engineering surveys, and in particular to a CPTU soil stratification method based on machine learning and uncertainty analysis. Background Technology

[0002] Marine exploration is indispensable in marine engineering, infrastructure construction, and seabed resource development. Accurate soil stratification and classification provide essential data for design and construction. Pore pressure cone penetration testing (CPTU), as an efficient method for marine soil exploration, can quickly obtain high-resolution soil parameters by measuring parameters such as cone tip resistance, side friction, and pore water pressure, providing important data support for soil layer classification and identification.

[0003] However, CPTU soil stratification of marine soils faces even greater challenges. Due to the complexity and heterogeneity of marine soil layers, traditional CPTU classification methods (such as Robertson diagrams) often fail to accurately determine soil categories. Furthermore, although existing machine learning methods can classify soils, the resulting classifications often lead to overly fine subdivisions and an excessive number of layers, making it difficult to obtain stratification results that are effective in engineering and have practical application value. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the technical problem to be solved by the present invention is to provide a CPTU soil stratification method based on machine learning and uncertainty analysis. This method combines high-resolution CPTU data with machine learning algorithms to effectively classify marine soils. By introducing uncertainty analysis, an automatic stratification method with optimized prediction category confidence is used to optimize the classification results, thereby obtaining soil layer division results with engineering reference value.

[0005] This invention is accomplished using the following technical solution: a CPTU soil stratification method based on machine learning and uncertainty analysis, comprising the following steps:

[0006] S1. Obtain CPTU data of marine soil and drilling data of soil near the CPTU survey location to obtain soil categories at certain depths;

[0007] S2. Train the base model using the labeled data from the CPTU dataset in S1, and construct a semi-supervised learning model by combining it with the unlabeled data from the CPTU dataset;

[0008] S3. Input the target location CPTU data into the classification model trained in S2 to obtain the classification result. Evaluate the classification result using a model accuracy evaluation metric.

[0009] S4. Combine uncertainty analysis to evaluate the confidence level of each classification result, and quantitatively analyze the uncertainty of the prediction results to obtain the uncertainty analysis results;

[0010] S5. Based on the uncertainty analysis results, determine the number and boundary locations of soil layers, and output the layering results.

[0011] Furthermore, the specific steps of S2 are as follows:

[0012] S21. Train a basic random forest classification model based on labeled data, input unlabeled data into the model, and obtain the prediction results of each decision tree;

[0013] S22. The predicted category is determined by a voting mechanism, which is the category with the most votes.

[0014] S23. For each predicted sample, calculate the percentage of votes for its predicted category in all decision trees, and use this percentage as the classification confidence. If the classification confidence exceeds a predetermined threshold, the prediction result is considered reliable. If the confidence is lower than the threshold, the sample is skipped.

[0015] S24. Use reliable prediction results as pseudo-labels and add them to the training dataset to form a new training set;

[0016] S25. Through multiple rounds of iterative training, using labeled data and added pseudo-label data, the classification model is gradually optimized to obtain a well-trained classification model.

[0017] Furthermore, the CPTU data includes cone tip resistance, side friction resistance, and pore water pressure parameters. The labeled data is CPTU data with a corresponding soil type, and the unlabeled data is CPTU data without a corresponding soil type.

[0018] Furthermore, when training the base model with labeled data from the CPTU dataset in S2, the base learner is a random forest, with input parameters being tip resistance, side friction resistance, and pore water pressure, and output features being soil type.

[0019] Furthermore, the model accuracy evaluation metrics in S3 are precision, recall, and F1 score.

[0020] Furthermore, the hyperparameters of the semi-supervised algorithm include three types: random forest decision tree depth (max_depth), number of random forest decision trees (n_estimators), and confidence threshold (threshold). The selection of random forest decision tree depth (max_depth) and number of random forest decision trees (n_estimators) is optimized through Bayesian parameter tuning and cross-validation. Among them, random forest decision tree depth (max_depth) and number of random forest decision trees (n_estimators) are used in the supervised model random forest to control the depth and number of decision trees, while confidence threshold is used to control the generation of pseudo-labels.

[0021] Furthermore, in S5, the predicted classification results from S3 and the uncertainty analysis results from S4 are combined to automatically determine the optimal boundary position and realize soil stratification based on the confidence of the predicted category. The final soil stratification result is obtained through multiple iterations.

[0022] Furthermore, the automatic stratification based on predicted class confidence optimization includes the following steps:

[0023] 1) Divide the target area into two soil layers and traverse the candidate boundary locations;

[0024] 2) Within each sub-interval, the average confidence level of each soil layer is calculated, and the soil category corresponding to the highest confidence level is determined as the soil category of that layer;

[0025] 3) When the two stratified categories are different, the target value is the overall confidence weighted mean of the combination of the boundary positions; when the stratified categories are the same, the target value is the difference between the confidence means of the two intervals.

[0026] 4) Finally, the boundary position corresponding to the maximum value of the objective function is selected as the stratification result of this iteration.

[0027] Furthermore, the iterative process is configured with parameters such as a maximum number of iterations, a minimum layer thickness, and a minimum confidence level. When the maximum number of iterations is reached, the iteration stops. If the soil layer thickness in a certain interval is lower than the minimum layer thickness in a certain iteration, the subdivision of that interval is stopped, and the current stratification result is retained. If the average confidence level of a certain soil layer is lower than the minimum confidence level, it is considered an uncertain soil layer. After the specified number of iterations, adjacent soil layers of the same soil type are considered as the same soil layer. Thin layers and uncertain soil layers are removed to obtain the final soil stratification result.

[0028] This invention achieves marine soil classification based on CPTU data by combining the high-resolution characteristics of CPTU data with a random forest classification model. A semi-supervised learning algorithm is used to fully utilize unlabeled data, enhancing the model's robustness and predictive performance. Uncertainty analysis is introduced to quantitatively evaluate the classification results. An automatic stratification method with confidence optimization accurately determines soil layer boundaries and effectively eliminates excessively thin and uncertain layers, thus outputting stratification results with engineering reference value. This method can provide a scientific and reliable basis for soil layer division in marine engineering design and construction. Attached Figure Description

[0029] Figure 1 This is a flowchart of a CPTU soil stratification method based on machine learning and uncertainty analysis.

[0030] Figure 2 This is a schematic diagram of the CPTU4 parameters in an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram of the CPTU6 parameters in an embodiment of the present invention;

[0032] Figure 4 This is a schematic diagram of the soil classification results of CPTU4 and CPTU6 models in an embodiment of the present invention;

[0033] Figure 5 This is a schematic diagram illustrating the uncertainty analysis results in an embodiment of the present invention;

[0034] Figure 6 This is a schematic diagram comparing the stratification results with the drilling results in an embodiment of the present invention. Detailed Implementation

[0035] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0036] Reference Figure 1-6 As shown, this invention provides a CPTU soil stratification method based on machine learning and uncertainty analysis, comprising the following steps:

[0037] S1. Obtain CPTU data of marine soil and drilling data of soil near the CPTU survey location to obtain soil type at a certain depth. The CPTU data includes parameters such as cone tip resistance, side friction resistance and pore water pressure.

[0038] S2. Train the base model using the labeled data from the CPTU dataset in S1, and then construct a semi-supervised learning model using the unlabeled data from the CPTU dataset. The specific steps of the semi-supervised learning model are as follows:

[0039] S21. A basic random forest classification model is trained based on labeled data. Unlabeled data is then input into the model to obtain the prediction results for each decision tree. The labeled data refers to CPTU data with corresponding soil categories, while the unlabeled data refers to CPTU data without corresponding soil categories. When training the basic random forest classification model with labeled CPTU data, the base learner is a random forest, the input parameters are tip resistance, side friction resistance, and pore water pressure, and the output feature is the soil category.

[0040] S22. The predicted category is determined by a voting mechanism, which is the category with the most votes.

[0041] S23. For each predicted sample, calculate the percentage of votes for its predicted class across all decision trees, and use this percentage as the classification confidence score. If the classification confidence score exceeds a predetermined threshold, the prediction result is considered reliable; if the confidence score is below the threshold, the sample is skipped.

[0042] S24. Use reliable prediction results as pseudo-labels and add them to the training dataset to form a new training set.

[0043] S25. Through multiple rounds of iterative training, using labeled data and added pseudo-label data, the classification model is gradually optimized to obtain a well-trained classification model.

[0044] In this semi-supervised learning method, semi-supervised algorithm hyperparameters are used. These hyperparameters include three types: random forest decision tree depth (max_depth), number of random forest decision trees (n_estimators), and confidence threshold (threshold). The selection of these hyperparameters is optimized using Bayesian parameter tuning and cross-validation. Specifically, the random forest decision tree depth (max_depth) and the number of random forest decision trees (n_estimators) are used in the supervised random forest model to control the depth and number of decision trees. The confidence threshold is used to control the generation of pseudo-labels.

[0045] S3. Input the target location CPTU data into the classification model trained in S2 to obtain the classification results. Evaluate the classification results in S2 using model accuracy evaluation metrics, which include precision, recall, and F1 score.

[0046] S4. Combine uncertainty analysis to evaluate the confidence level of each classification result and quantitatively analyze the uncertainty of the prediction result to obtain the uncertainty analysis results. Specifically, uncertainty analysis evaluation refers to calculating the percentage of votes for each classification result across all decision trees and using this percentage as the classification confidence level.

[0047] S5. Based on the uncertainty analysis results, determine the number and boundary locations of soil layers, and output the layering results.

[0048] Specifically, combining the predicted classification results in S3 and the uncertainty analysis results in S4, an automatic stratification method based on predicted category confidence optimization automatically determines the optimal boundary position and achieves soil stratification. After multiple iterations, the final soil stratification result is obtained. The iteration process is configured with parameters such as a maximum number of iterations, a minimum layer thickness, and a minimum confidence level. Iteration stops when the maximum number of iterations is reached. If the soil layer thickness in a certain interval is lower than the minimum layer thickness in a given iteration, the subdivision of that interval is stopped, and the current stratification result is retained. If the average confidence level of a soil layer is lower than the minimum confidence level, it is considered an uncertain soil layer. After the specified number of iterations, adjacent soil layers of the same soil category are considered as the same soil layer. Thin layers and uncertain soil layers are removed to obtain the final soil stratification result.

[0049] The above-mentioned automatic stratification method based on predicted class confidence optimization includes the following steps:

[0050] 1) Divide the target area into two soil layers and traverse the candidate boundary locations;

[0051] 2) Within each sub-interval, the average confidence level of each soil layer is calculated, and the soil category corresponding to the highest confidence level is determined as the soil category of that layer;

[0052] 3) When the two stratified categories are different, the target value is the overall confidence weighted mean of the combination of the boundary positions; when the stratified categories are the same, the target value is the difference between the confidence means of the two intervals.

[0053] 4) Finally, the boundary position corresponding to the maximum value of the objective function is selected as the stratification result of this iteration.

[0054] Example: Reference Figure 2-6 As shown, the CPTU soil stratification method based on machine learning and uncertainty analysis is applied to a certain marine tunnel project, and the steps are as follows:

[0055] S1. Seven sets of CPTU data obtained during the exploration phase of a marine tunnel project were collected. Each set of CPTU data outputs a set of cone tip resistance, side friction resistance, and pore water pressure at 0.01-meter intervals along the depth direction. Combined with drilling results within 3 meters around the CPTU exploration points, soil type information at certain depths was obtained. The main soil types at the target site include silty clay, silt, and silty clay.

[0056] S2. Train a basic random forest classification model based on labeled data from CPTU, and construct a semi-supervised learning model by combining it with unlabeled data from CPTU.

[0057] Specifically, the 7 sets of CPTU data collected in S1 were distinguished by whether they were labeled or not, resulting in a total of 2,635 labeled data sets and 33,016 unlabeled data sets.

[0058] When training the basic random forest classification model with labeled data, five sets of CPTU data (CPTU1, CPTU2, CPTU3, CPTU5, CPTU7) and their corresponding soil categories for the target site are used for training, while the remaining two sets of CPTU data (CPTU4, CPTU6) are used as target locations for model validation. (Parameters for CPTU4 and CPTU6 are as follows...) Figure 2 and Figure 3 (As shown)

[0059] The semi-supervised learning model uses three hyperparameters: the depth of the random forest decision tree (max_depth), the number of random forest decision trees (n_estimators), and the confidence threshold (threshold). The depth and number of random forest decision trees (max_depth) are optimized using Bayesian hyperparameter tuning and cross-validation. In Bayesian hyperparameter tuning, the search range for the depth (max_depth) is 3 to 10, the search range for the number of random forest decision trees (n_estimators) is 50 to 200, and the cross-validation fold count is 10 folds. The tuning results show that the optimal hyperparameter combination is max_depth = 3 and n_estimators = 187. Additionally, based on experience, threshold = 0.90 is set.

[0060] Finally, through multiple rounds of iterative training, a total of 1052 sets of unlabeled data were regarded as pseudo-labels with high confidence and added to the training set, thus forming an optimized classification model and obtaining a well-trained classification model.

[0061] S3. Input the target location CPTU data into the classification model trained in S3 to obtain the classification result. Use the model accuracy evaluation index to evaluate the classification result in S2.

[0062] The model's predictive performance is shown in Table 1. The data shows that the model performs best in classifying silty clay, with both precision and recall at 0.84 and an F1 score of 0.86. However, due to the smaller sample size, the model's performance in classifying silt is relatively weaker, with a precision of 0.63 and an F1 score of 0.71. Overall, the model achieves an average precision and recall of 0.81 across all classes, demonstrating good classification performance.

[0063] Table 1

[0064] Soil categories Sample size accuracy Recall rate F1 score silty clay 214 0.84 0.74 0.79 silt 155 0.63 0.81 0.71 Silty clay 519 0.84 0.84 0.86 overall 888 0.82 0.81 0.81

[0065] S4. Combine uncertainty analysis to assess the confidence level of each classification result and quantitatively analyze the uncertainty of the prediction results.

[0066] Table 2 shows the classification accuracy of the model across different confidence intervals. The results indicate a significant correlation between the model's prediction confidence and the actual accuracy; higher confidence levels correlate with higher classification accuracy. This further validates the effectiveness of uncertainty analysis in evaluating soil classification results. Classification results at the two target locations (e.g., ...) Figure 4 As shown in the figure, the preliminary stratification results along the depth direction are relatively detailed, with a large number of soil layers and poor distribution regularity. This indicates that the classification results need further optimization through uncertainty analysis and post-processing to obtain stratification results with greater engineering application value. Uncertainty analysis (such as...) Figure 5 As shown in the diagram, we can further understand the reliability of the classification results. Taking CPTU4 as an example, its shallow soil is mainly composed of silt with high confidence, while the shallow soil of CPTU6 is mainly composed of silty clay with high confidence. This confidence-based analysis not only improves the interpretability of the classification results, but also provides support for subsequent stratification optimization.

[0067] Table 2

[0068] Confidence interval Sample size accuracy 0.9 to 1.0 10 1.00 0.7 to 0.9 531 0.92 0.5 to 0.7 310 0.70 0 to 0.5 46 0.26

[0069] S5. Combining the classification results with uncertainty analysis, the automatic stratification method based on the prediction category confidence optimization automatically determines the optimal boundary position and realizes soil stratification. After multiple iterations, the final soil stratification result is obtained.

[0070] The iterative process was set with a maximum of 3 iterations, a minimum layer thickness of 0.5 meters, and a minimum confidence level of 0.4. The layering depth range, soil type, objective function, and mean confidence level for each iteration in CPTU4 and CPTU6 are shown in Tables 3 and 4.

[0071] Table 3

[0072]

[0073] Table 4

[0074]

[0075] After a specified number of iterations, soil layers with the same soil type in adjacent layers are considered as the same soil layer. Thin layers and uncertain soil layers are removed to obtain the final soil stratification result.

[0076] In an embodiment of the present invention, the soil stratification results of CPTU4 are as follows: at depths of 9.30 m to 19.50 m, the soil is determined to be silt and considered as a single soil layer; at depths of 19.50 m to 20.39 m, the soil is determined to be silty clay, but due to low confidence, it is considered an uncertain soil layer and discarded; at depths of 20.39 m to 28.75 m, the soil is re-determined to be silt; and at depths of 28.76 m to 68.30 m, the soil is determined to be silty clay and considered as a complete soil layer. Based on the above analysis, this method divides the soil at the location of CPTU4 into two main soil layers: a silt layer at depths of 9.30 m to 28.76 m and a silty clay layer at depths of 28.76 m to 68.30 m. Similarly, in the soil stratification of CPTU6, three soil layers were identified: a silt layer at a depth of 4.00 m to 6.42 m, a silty clay layer at a depth of 6.42 m to 27.66 m, and a silty clay layer at a depth of 27.66 m to 72.00 m. This was compared with actual drilling results (e.g., Figure 6 As shown in the figure, the stratification results of CPTU4 are highly consistent with the drilling results, and the stratification boundary positions are accurate, verifying the reliability and accuracy of this method. CPTU6 correctly identified the soil layers in most locations, only being misclassified as silt layers at shallower depths. Overall, this method can effectively achieve soil stratification, exhibiting good classification performance, and by eliminating uncertain and excessively thin layers, it generates stratification results with greater engineering reference value.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A CPTU soil layering method based on machine learning and uncertainty analysis, characterized in that: The method comprises the following steps: S1. Obtain CPTU data of marine soil and drilling data of soil adjacent to the CPTU survey position to obtain soil categories at partial depths; S2. Train a basic model using labeled data of the CPTU data in S1, and construct a semi-supervised learning model in combination with unlabeled data of the CPTU data, and the specific steps are as follows: S21. Train a basic random forest classification model based on the labeled data, input the unlabeled data into the model, and obtain the prediction results of each decision tree; S22. Determine the prediction category as the category with the most votes through a voting mechanism; S23. For each prediction sample, calculate the proportion of the number of votes of the prediction category in all decision trees, and take the proportion as the classification confidence. If the classification confidence exceeds a predetermined threshold, the prediction result is considered reliable. If the confidence is lower than the threshold, the sample is skipped; S24. Add the reliable prediction result as a pseudo-label to the training data set to form a new training set; S25. Gradually optimize the classification model by multiple rounds of iterative training using the labeled data and the added pseudo-label data to obtain a trained classification model; S3. Input the CPTU data of the target position into the trained classification model in S2 to obtain a classification result, and evaluate the classification result using a model accuracy evaluation index; S4. Evaluate the confidence of each classification result in combination with uncertainty analysis, and quantitatively analyze the uncertainty of the prediction result to obtain an uncertainty analysis result; S5. Determine the number and boundary position of the soil layering according to the uncertainty analysis result, and output the layering result. In S5, the prediction classification result in S3 and the uncertainty analysis result in S4 are combined, the automatic layering based on the prediction category confidence is optimized to automatically determine the best boundary position and realize soil layering, and the final soil layering result is obtained through multiple iterations. The maximum number of iterations, the minimum layer thickness and the minimum confidence are set as parameters. When the maximum number of iterations is reached, the iteration stops. If the soil layer thickness in a certain interval is lower than the minimum layer thickness in a certain iteration, the subdivision of the interval is stopped, and the current layering result is retained. If the average confidence of a soil layer is lower than the minimum confidence, the soil layer is considered as an uncertain soil layer. After a specified number of iterations, the soil layers with the same soil category are considered as the same soil layer, and the thin layer and the uncertain soil layer are removed to obtain the final soil layering result. The automatic layering based on the prediction category confidence comprises the following steps: 1) Divide the target area into two soil layers, and traverse the candidate boundary position; 2) In each sub-interval, calculate the average confidence of each layer of soil, and determine the soil category corresponding to the maximum confidence as the soil category of the layer; 3) When the layering categories are different, the target value is the weighted average value of the overall confidence of the boundary position combination. When the layering categories are the same, the target value is the difference between the average confidence values of the two intervals; 4) Finally, select the boundary position corresponding to the maximum value of the target function as the layering result of this iteration.

2. The method of claim 1, wherein the method further comprises: The CPTU data includes cone tip resistance, side friction resistance and pore water pressure parameters, the labeled data is CPTU data corresponding to soil categories, and the unlabeled data is CPTU data without corresponding soil categories.

3. The CPTU soil layering method based on machine learning and uncertainty analysis according to claim 1 or 2, characterized in that S2 When the labeled data of the CPTU data is used to train the basic model, the basic learner is a random forest, the input parameters are cone tip resistance, side friction resistance and pore water pressure, and the output features are soil categories.

4. The method of claim 1, wherein the method further comprises: The model accuracy evaluation index in S3 is accuracy, recall rate and F1 score.

5. The method of claim 1 or 2, wherein the method further comprises: The semi-supervised algorithm hyperparameters include three kinds of hyperparameters: random forest decision tree depth max_depth, random forest decision tree number n_estimators and confidence threshold threshold. The random forest decision tree depth max_depth and the random forest decision tree number n_estimators are selected by optimizing the Bayesian parameter tuning method and cross-validation, wherein the random forest decision tree depth max_depth and the random forest decision tree number n_estimators are used in the supervised model random forest to control the decision tree depth and the decision tree number, and the confidence threshold threshold is used to control the generation of pseudo labels.

Images