A Method for Predicting Marine Fine-Grained Soil Content Based on Multi-Source CPTU Data and Machine Learning

Abstract

This invention relates to the interdisciplinary field of geotechnical engineering investigation and artificial intelligence, specifically to a method for predicting the content of marine fine-grained soil based on multi-source CPTU data and machine learning. The method includes: acquiring CPTU data and corresponding geotechnical test component content data; constructing CPTU prototype vectors for geotechnical test points and calculating the cosine similarity weight between data points and prototype vectors, and the Gaussian distance weight between data points and the center of the window; generating multiple sliding windows centered on each test point and extracting high-dimensional feature vectors using dual weights; training a random forest regression model using the feature vectors; moving the window of the borehole to be predicted according to the step size and extracting features, inputting the extracted features into the trained model, and generating a continuous depth profile of the fine-grained soil component content. This invention breaks through the indirect qualitative discrimination mode of traditional empirical methods, realizing direct, quantitative, and continuous prediction of soil component content from CPTU data, and solving the technical problems of high cost, long cycle, discrete data, and inability to quantify traditional methods.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of geotechnical engineering investigation and artificial intelligence, specifically to a method for predicting the composition and content of marine fine-grained soil based on pore pressure static cone penetration test (CPTU) data and machine learning algorithms, which is particularly suitable for identifying soil composition and content in marine geotechnical engineering investigation. Background Technology

[0002] Marine geotechnical investigation serves as a fundamental frontier for marine resource development and marine engineering construction. The accuracy and reliability of its findings directly impact the safety, stability, construction costs, and operational lifespan of marine engineering projects. Marine sedimentary environments are dominated by fine-grained soils (clay, silt, and fine sand), whose composition and content distribution directly determine the engineering properties of the soil (such as strength, deformation, and permeability), thus affecting the stability and durability of various marine engineering structures. Even subtle changes in the composition and content of fine-grained soils often lead to qualitative changes in their engineering properties. For example, an increase in clay content significantly reduces soil permeability and increases its compressibility, while an increase in silt content directly affects the soil's liquefaction sensitivity and dynamic response characteristics. This sensitive correlation between composition and properties makes accurate prediction of the composition and content of fine-grained soils a prerequisite for the safe construction of marine engineering projects.

[0003] Currently, the traditional method for obtaining soil composition and content mainly relies on borehole sampling combined with laboratory geotechnical tests. Although this method can directly obtain soil samples and provide detailed physical and mechanical parameters, and has long been regarded as the "gold standard," its inherent shortcomings have been amplified in the marine environment, especially in the process of moving towards the deep sea: offshore drilling operations are affected by harsh environmental conditions such as wind, waves, and currents, resulting in high economic costs and long operation cycles; the sampling process inevitably disturbs the soil samples, affecting the accuracy of the data; more importantly, this method can only obtain data at discrete depth points, making it difficult to capture the continuous spatial variation of soil composition, and thus failing to meet the needs of fine characterization of complex strata.

[0004] Static cone penetration test (CPTU), as an efficient, economical, and data-continuous in-situ testing technique, has become a widely used method in marine engineering exploration. However, existing CPTU-based techniques are mostly focused on soil classification and identification. For example, Robertson's empirical chart method and the SBTn classification method mainly function to qualitatively or semi-quantitatively classify soil types based on CPTU parameters (such as cone tip resistance, side friction, and pore pressure). These methods have obvious limitations: first, they remain at the level of soil type identification and cannot directly and quantitatively predict the specific percentage content of clay, silt, and sand; second, these empirical relationships are mostly based on databases established in specific regions, with strong regional dependence and poor universality, making it difficult to directly extend to marine areas with different sedimentary environments. Summary of the Invention

[0005] To address the problems of high cost, long cycle time, large disturbance, strong data dispersion, and inability to directly, quantitatively, and continuously predict the composition and content of fine-grained soils based on CPTU data in current methods for detecting the composition and content of marine fine-grained soils, this invention proposes a method for predicting the content of marine fine-grained soils based on multi-source CPTU data and machine learning, which adopts the following scheme:

[0006] A method for predicting marine fine-grained soil content based on multi-source CPTU data and machine learning includes the following steps:

[0007] Step 1: Obtain a training dataset containing CPTU data and the corresponding depth of geotechnical test component content, and perform preprocessing;

[0008] Step 2, Prototype Vector Construction and Dual Weight Calculation: Based on the geotechnical test depth points in the training set, construct the CPTU prototype vector; for each CPTU data point, calculate its cosine similarity with the prototype vector as the first weight, and calculate its Gaussian distance from the center of the window as the second weight.

[0009] Step 3, Multiple sliding window feature extraction: Multiple sliding windows are generated with each geotechnical test depth point as the center; for the data in each window, a weighted calculation is performed using a combination of the first weight and the second weight to extract a high-dimensional feature vector;

[0010] Step 4: Train a random forest regression model using the high-dimensional feature vector and the corresponding geotechnical test component content;

[0011] Step 5: For the borehole to be predicted, move the window along the depth by a preset step size, perform the feature extraction process of steps S2 and S3 at each depth point, and input it into the trained model to predict the continuous depth profile of fine soil composition.

[0012] Furthermore, before step 4, a feature selection step is included: calculating the importance of each feature in the high-dimensional feature vector based on the random forest algorithm, and selecting a subset of key features for model training according to a preset threshold.

[0013] Further, in step 2, the cosine similarity weight is obtained by calculating and normalizing the cosine similarity between the neighboring CPTU data point vector Vi and the prototype vector Pi. The cosine similarity calculation formula is as follows:

[0014] cos_sim ) .

[0015] Furthermore, the Gaussian distance calculation formula in step 2 is as follows:

[0016] Where d = |current point depth - window center depth| (m) is the standard deviation parameter (m) of the Gaussian function.

[0017] Furthermore, in step 1, the CPTU data includes: cone tip resistance q t Side friction resistance f s pore water pressure u2, friction ratio F r pore pressure ratio B q Effective stress of the overlying layer .

[0018] Furthermore, the high-dimensional feature vector in step 3 is 49-dimensional, and it is constructed in the following way:

[0019] Based on the combined weights, calculate q t f s u2, F r B q , Normalized pore water pressure These 7 parameters include the weighted mean, weighted standard deviation, maximum value, minimum value, range, and median, totaling 42 statistical characteristics; calculate q. t The slope of the weighted linear trend varying with depth is used as a morphological feature; q is calculated. t and f s The weighted correlation coefficient is used as one correlation feature; the mean of the first weight, the mean of the second weight, the mean of the combined weight, the total number of data points in the window, and the number of valid points with a combined weight greater than 0.1 are calculated as a total of 5 data quality features.

[0020] Furthermore, the feature is that the preset step size is 0.05 meters; after obtaining the predicted value of the fine-grained soil component content at each depth point, normalization processing is performed so that the sum of each content is 100%.

[0021] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0022] 1. This invention fundamentally changes the indirect, qualitative model of test parameters, soil type identification, and empirical estimation in traditional empirical methods (such as the Robertson chart method). For the first time, it constructs a direct, quantitative mapping relationship from multi-source CPTU data to soil component content. Through the random forest algorithm, this invention autonomously learns the complex nonlinear relationship between key parameters such as cone tip resistance, side friction, and pore water pressure and the content of clay, silt, and sand particles. This achieves end-to-end accurate prediction from test signals to component percentages, eliminating information loss and subjective bias in intermediate stages and realizing precise quantification of component content.

[0023] 2. This invention innovatively introduces a dual-weight calculation mechanism to achieve adaptive feature focusing. By emphasizing data points similar to the prototype vector through similarity weights, it improves the target relevance of feature extraction. By generating multiple overlapping windows around each experimental depth point, a single experimental point is transformed into multiple training samples, enabling the training model to have position-invariant feature recognition capabilities, while also solving the problem of scarce geotechnical test data. A 49-dimensional multi-angle feature vector covering statistical, morphological, correlation, and data quality features is constructed, and key features are automatically selected using random forests to ensure the model's generalization ability under different marine geological conditions.

[0024] 3. This invention generates continuous composition curves with a step size of 0.05m through a sliding window mechanism. It can clearly show the subtle changes in soil composition in complex strata such as thin interlayers and gradient layers, and generate continuous soil composition profiles along the depth, overcoming the discreteness defects of traditional methods. It expands point-like laboratory data into linear continuous profiles, realizing the leap from sampling statistics to full profile analysis, and providing key geological information for the design of offshore platform pile foundations and the selection of submarine pipeline routes. Attached Figure Description

[0025] The invention will now be further described with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart of the method for predicting the composition of marine fine-grained soil based on multi-source CPTU data and machine learning, as described in an embodiment of the present invention.

[0027] Figure 2 This is a comparison chart of the predicted and actual values ​​of clay, silt and sand content in fine-grained soil based on the test set in this invention.

[0028] Figure 3 This is a predicted profile of soil composition and content from borehole BH001 of this invention.

[0029] Figure 4 This is a predicted profile of soil composition and content from borehole BH002 of this invention.

[0030] Figure 5 This is a continuity analysis of the predicted profiles of boreholes BH001 and BH002 in this invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1: This example proposes a method for predicting the content of fine-grained marine soils based on multi-source CPTU data and machine learning. This method constructs a prototype vector for each test point based on CPTU test data and indoor geotechnical test data. A multi-sliding window mechanism (20 windows, window size 0.15m, step size 0.05m) is used to extract statistical features, morphological features, correlation features, and data quality features of each window along the depth direction. A dual weighting mechanism integrating cosine similarity weight and Gaussian distance weight is designed to weight the contribution of each data point within the window. A random forest regression algorithm is used to train a quantitative prediction model from multi-dimensional CPTU features to clay content, silt content, and sand content, aiming to minimize the weighted mean square error. Feature selection is optimized through feature importance analysis, and cross-validation is used to evaluate the model's generalization ability, ultimately achieving continuous profile prediction along the depth. A prediction continuity index and gap analysis are introduced to evaluate the reliability of the prediction results. Specifically, this includes:

[0033] S1. Conduct CPTU field penetration tests, collect CPTU data from multiple boreholes and laboratory geotechnical test data, and construct a multi-source heterogeneous dataset. CPTU data includes depth h and cone tip resistance q. t Side friction resistance f s pore water pressure u2, friction ratio F r pore pressure ratio B q Effective stress of the overlying layer Normalized pore water pressure Parameters such as depth h, clay content, silt content, and sand content are included in the indoor geotechnical test data.

[0034] S2. Data preprocessing and standardization, including sorting, deduplication, and handling of missing values, and outlier detection and removal based on the 3σ criterion to ensure data quality. Geotechnical mechanics derived parameters are calculated from CPTU data to construct a complete feature system. Simultaneously, the sum of clay, silt, and sand content in the laboratory geotechnical test data is verified to be 100% to ensure data validity. The processed data is then divided into training and testing sets.

[0035] Outlier detection is performed using the Laida criterion (3σ). The Laida criterion is based on a normal distribution, and data falling within the range of 3σ indicates an outlier. The probability within the specified range is approximately 99.73%, and beyond that range... ± The probability of data points appearing within this range is only 0.27%, and the calculation formula includes:

[0036]

[0037]

[0038] in, The mean, The standard deviation is 1 / 3, and the valid range of the data is 1 / 3. .

[0039] S3. Prototype Vector Construction and Similarity-Distance Weight Calculation: Based on the training set, the CPTU data of each geotechnical test depth point is constructed into a prototype vector P. i prototype vector P i Features include cone tip resistance q t Side friction resistance f s pore water pressure u2, friction ratio F r pore pressure ratio B q Effective stress of the overlying layer Normalized pore water pressure Key features include: the prototype vector represents the CPTU feature fingerprint of the soil at the test point, used for subsequent similarity calculation. Prototype vector P i = .

[0040] For each data point within the window, calculate the neighboring data point vector V. i With prototype vector P i The cosine similarity is used as the first weight, and the Gaussian formula is used to calculate the Gaussian distance between adjacent data points and the geotechnical test depth points as the second weight; the two are multiplied to obtain the comprehensive double weight.

[0041] Cosine similarity weight calculation includes:

[0042] a. For CPTU data point i at the geotechnical test depth, obtain its characteristic prototype vector P. i ;

[0043] b. For a neighboring CPTU data point i, obtain its vector V. i ;

[0044] c. Calculate the adjacent data point vector V i With the geotechnical test depth point vector P i Cosine similarity;

[0045] cos_sim )

[0046] d. Calculate the normalized similarity.

[0047] The formula for calculating Gaussian distance is as follows:

[0048]

[0049] Where d = |current point depth - window center depth| (m) is the standard deviation parameter (m) of the Gaussian function.

[0050] Double weighting = similarity weight × Gaussian distance weight.

[0051] This invention innovatively introduces a dual-weighting calculation mechanism to achieve adaptive feature focusing. This mechanism simultaneously integrates feature space similarity and physical space proximity. Cosine similarity weighting enhances the feature matching degree with the prototype vector, improving the target relevance of feature extraction; Gaussian distance weighting highlights the contribution of data points near the window center, ensuring the local representativeness of the features. The synergistic effect of the dual weights automatically identifies key data points within the window, effectively suppressing noise and outlier interference, significantly improving feature quality and model robustness, while also taking into account the continuity of strata in the depth direction.

[0052] S4. Multiple sliding window feature extraction: For each geotechnical test depth point, 20 overlapping sliding windows are generated around its depth. The center depth of each window is uniformly distributed within ±0.15m of the test point depth. Each window is 0.15m in size. All CPTU data points within the window are extracted, and the window features are calculated. The steps are as follows:

[0053] a. Cosine similarity weight: The cosine similarity between this point and the prototype vector;

[0054] b. Distance weight: A Gaussian function of the absolute difference between the depth of this point and the depth of the window center;

[0055] c. Combined weight = similarity weight × distance weight.

[0056] Window features include four dimensions:

[0057] 1. Statistical characteristic calculation: Calculate each parameter (q) t f s u2, F r B q , , The weighted statistical characteristics of (weighted mean, weighted standard deviation, maximum value, minimum value, range, median);

[0058] 2. Morphological characteristic calculation: Calculate the cone tip resistance q t Correlation characteristics of depth-weighted linear trend (slope) (depth-weighted mean, q) t (weighted mean, covariance, variance)

[0059] 3. Calculation of correlation characteristics: Calculate the weighted correlation coefficient (covariance, standard deviation, correlation coefficient) between qt and fs.

[0060] 4. Data quality feature calculation: mean of similarity weight, mean of distance weight, mean of combined weight, total number of data points, number of valid points (number of points with a combined weight greater than 0.1).

[0061] By performing multi-dimensional feature calculations on the data within a window, a comprehensive feature vector containing 49 dimensions is constructed for each window. This vector system integrates four types of features: 42 statistical features covering 7 key parameters, used to describe the central tendency and dispersion of the parameter distribution; 1 morphological feature, used to reflect the changing trend of the parameter with depth; 1 correlation feature, used to reveal the intrinsic mechanical relationship between parameters; and 5 data quality features, used to provide a self-assessment of feature reliability.

[0062] This invention utilizes a high-dimensional feature system, a dual-weighting mechanism, and a multiple sliding window design to work synergistically. By generating multiple overlapping windows around each test point, it successfully transforms sparse indoor test data into rich training samples. This not only effectively solves the bottleneck of scarce geotechnical test data but also enables the model to possess position-invariant feature recognition capabilities. This multi-angle, self-evaluating feature representation method fundamentally overcomes the limitations of traditional single-feature descriptions.

[0063] S5. Feature selection: A random forest feature importance algorithm is used to select the most important features for the prediction target (clay, silt, and sand content) from the above 49 features based on a set threshold, reducing redundancy and improving the model's generalization ability, as detailed below:

[0064] a. Based on the training set data, use the random forest model to calculate the importance of each feature;

[0065] b. Feature importance is a built-in metric in random forests, based on the reduction of Gini impurity or the reduction of average error;

[0066] c. Sort the features by importance and select the features whose importance is greater than the threshold (0.005);

[0067] d. If the number of selected features is less than 10, then at least the top 10 most important features should be selected.

[0068] S6. Model training and evaluation: The random forest regression algorithm is used, combined with 10-fold cross-validation. Multiple metrics, such as the coefficient of determination R, are calculated on the test set. 2 Mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), etc.

[0069] S6. Model training and evaluation: The random forest regression algorithm is used, combined with 10-fold cross-validation. Multiple metrics, such as the coefficient of determination R, are calculated on the test set. 2The parameters include mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). The model training and evaluation steps are as follows:

[0070] a. Train using a selected subset of features;

[0071] b. Random forest parameters: 100 decision trees, maximum depth of 15 layers, 5 samples per node, and 42 random seeds;

[0072] c. Perform 10-fold cross-validation;

[0073] d. Train the model using sample weights;

[0074] Total weight = Distance weight × Similarity weight × Dual weight × Data quality weight;

[0075] e. Make predictions on the test set and then de-standardize them to obtain the predicted values ​​of the original dimensions;

[0076] f. Calculate the performance indicators of each objective variable.

[0077] S7. Continuous Depth Profile Prediction: For each borehole, the predicted depth is generated in steps of 0.05m. For each depth, window data is extracted, window features are calculated, and a trained model is used for prediction. The prediction results are de-standardized, ensuring that the sum of the three components (clay, silt, and sand) is 100%. The prediction steps are as follows:

[0078] a. Extract CPTU data within a 0.3m window centered on that point;

[0079] b. Construct the prototype vector for the current depth (based on the nearest data point within the window);

[0080] c. Calculate window features (using the same dual-weighting mechanism as the training phase);

[0081] d. Use the trained model to predict the content of clay, silt, and sand particles;

[0082] e. Normalize the prediction results to ensure that the sum of the three is 100%.

[0083] S8. Post-processing and visualization output of prediction results, the steps are as follows:

[0084] a. Performance evaluation, showing a comparison of indicators such as R², MAE, RMSE, and MAPE, see Table 1;

[0085] b. A comparison chart of predicted and actual values ​​of clay, silt, and sand content in fine-grained soil on the test set is shown below. Figure 2 ;

[0086] c. Display the curves showing the changes in clay, silt, and sand content with depth, respectively, and overlay the verification from laboratory test points (see [link]). Figure 3 and Figure 4 ;

[0087] d. Continuity Analysis: Calculate the percentage of continuity of the predicted profile. Continuity = (Actual number of predicted points / Theoretical number of predicted points) × 100% (See...) Figure 5 ;

[0088] Table 1 shows the calculation results of the test set data for the method of this invention.

[0089] Element R² MAE (%) RMSE (%) MAPE (%) Continuity (%) clay 0.89 3.2 4.1 8.5 92.3 Powder 0.87 3.8 4.9 9.2 91.7 sand 0.91 2.9 3.7 7.8 93.1

[0090] This invention achieves precise capture of soil spatial variability through a multi-sliding window mechanism, establishes a dual weighting mechanism that integrates cosine similarity and Gaussian distance, effectively balances the relationship between feature similarity and spatial continuity, and uses a random forest algorithm to construct a robust mapping relationship from CPTU multi-source parameters to soil composition content.

[0091] Compared with traditional methods, this invention achieves a technological leap from "qualitative discrimination" to "quantitative prediction" and from "discrete point evaluation" to "continuous profile analysis," providing a revolutionary solution for marine geotechnical engineering investigation. Figure 3 As shown, the model's predicted values ​​on the test set are highly consistent with the laboratory measured values. The coefficients of determination (R²) for clay, silt, and sand are 0.89, 0.87, and 0.91, respectively, verifying the accuracy of the quantitative prediction. Figure 4 and Figure 5 Further evidence demonstrates that the predicted profiles of this invention maintain over 90% continuity across multiple boreholes, accurately identifying the interface locations of thin soil layers and clearly revealing the continuous variation patterns of component content in gradually changing strata. This technological breakthrough provides an innovative solution for marine geotechnical engineering investigation, offering reliable technical support for marine engineering structure design and safety evaluation through high-precision quantitative prediction and continuous profile analysis.

[0092] In summary, this invention provides a paradigm-innovative method for predicting the composition and content of fine-grained soil in marine geotechnical engineering, deeply integrating artificial intelligence technology with geotechnical engineering expertise. In engineering applications, this method can achieve continuous profile generation on an hourly basis, far exceeding the efficiency of the traditional weekly borehole sampling + laboratory testing model. In terms of technological innovation, it establishes for the first time a direct quantitative mapping relationship from multiple CPTU parameters to the content of fine-grained soil components, overcoming the limitations of qualitative judgment in traditional empirical chart methods. Regarding the quality of the results, the provided continuous profile data offers high-resolution input for numerical simulation, parameter inversion, and engineering design, significantly improving the accuracy and reliability of marine engineering design.

[0093] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for predicting the content of marine fine-grained soils based on multi-source CPTU data and machine learning, characterized in that, Includes the following steps: Step 1: Obtain a training dataset containing CPTU data and the corresponding depth geotechnical test component content, and perform preprocessing. The CPTU data includes: cone tip resistance q. t Side friction resistance f s pore water pressure u2, friction ratio F r pore pressure ratio B q Effective stress of the overlying layer ; Step 2, Prototype Vector Construction and Dual Weight Calculation: Based on the geotechnical test depth points in the training set, construct the CPTU prototype vector; for each CPTU data point, calculate its cosine similarity with the prototype vector as the first weight, and calculate its Gaussian distance from the center of the window as the second weight. Step 3, Multiple Sliding Window Feature Extraction: Multiple sliding windows are generated centered on each geotechnical test depth point; for the data within each window, a weighted calculation is performed using a combination of the first and second weights to extract a high-dimensional feature vector; the high-dimensional feature vector is 49-dimensional and is constructed in the following way: Based on the combined weights, calculate q t f s u2, F r B q , Normalized pore water pressure These 7 parameters include the weighted mean, weighted standard deviation, maximum value, minimum value, range, and median, totaling 42 statistical characteristics; calculate q. t The slope of the weighted linear trend varying with depth is used as a morphological feature; q is calculated. t and f s The weighted correlation coefficient is used as one correlation feature; the mean of the first weight, the mean of the second weight, the mean of the combined weight, the total number of data points in the window, and the number of valid points with a combined weight greater than 0.1 are calculated as a total of 5 data quality features. Step 4: Train a random forest regression model using the high-dimensional feature vector and the corresponding geotechnical test component content; Step 5: For the borehole to be predicted, move the window along the depth by a preset step size, perform the feature extraction process of steps S2 and S3 at each depth point, and input it into the trained model to predict the continuous depth profile of fine soil composition.

2. The method for predicting marine fine-grained soil content based on multi-source CPTU data and machine learning according to claim 1, characterized in that, Before step 4, a feature selection step is also included: calculating the importance of each feature in the high-dimensional feature vector based on the random forest algorithm, and selecting a subset of key features for model training according to a preset threshold.

3. The method for predicting marine fine-grained soil content based on multi-source CPTU data and machine learning according to claim 1, characterized in that, In step 2, the cosine similarity weight is obtained by calculating and normalizing the cosine similarity between the neighboring CPTU data point vector Vi and the prototype vector Pi. The cosine similarity calculation formula is as follows: cos_sim ) .

4. The method for predicting marine fine-grained soil content based on multi-source CPTU data and machine learning according to claim 1, characterized in that, The formula for calculating the Gaussian distance in step 2 is as follows: Where d = |current point depth - window center depth|, is the standard deviation parameter of the Gaussian function.

5. The method for predicting marine fine-grained soil content based on multi-source CPTU data and machine learning according to claim 1, characterized in that, The preset step size is 0.05 meters; after obtaining the predicted value of the fine-grained soil composition at each depth point, normalization is performed so that the sum of each content is 100%.

Images