A CPTU-based multi-source geological exploration data fusion method

Abstract

The present application belongs to the cross technical field of geotechnical engineering investigation and artificial intelligence, and specifically relates to a multi-source geological investigation data fusion method based on CPTU, comprising: obtaining CPTU data set and borehole data set and preprocessing; performing data set matching based on depth, and utilizing sliding window to extract CPTU center point features, window statistical features and multi-scale features extracted by wavelet transform to form a feature set; constructing a standardized training sample with the feature set and soil body parameter label; adopting multi-output random forest for model training, and applying the trained model to full-depth continuous profile prediction and uncertainty quantification. The present application solves the problems of data fragmentation and uncertainty quantification in traditional methods by fusing discrete borehole data and continuous CPTU data, establishes a fusion framework of multi-source heterogeneous data and soil body parameter mapping, and provides reliable prediction and decision support for marine geotechnical engineering, infrastructure construction and geological disaster assessment.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of geotechnical engineering investigation and artificial intelligence, specifically involving a multi-source geological investigation data fusion method based on CPTU. Background Technology

[0002] Marine geotechnical investigation is a prerequisite for marine resource development and marine engineering construction. The accuracy and reliability of the investigation results are related to the safety, stability, and life-cycle cost of marine engineering projects. Accurately obtaining the physical and mechanical parameters of marine soil (such as undrained shear strength, compression modulus, permeability coefficient, overconsolidation ratio, and fine particle content) is a prerequisite for the refined design, safety assessment, and long-term performance prediction of structures such as offshore platforms, subsea pipelines, and offshore wind power foundations.

[0003] Currently, the main methods for obtaining soil engineering property parameters rely on borehole sampling and laboratory geotechnical testing. While this method can directly obtain soil samples and provide relatively detailed physical and mechanical parameters, it has long been considered the "gold standard." However, its inherent limitations are amplified in the marine environment: firstly, offshore drilling operations are affected by harsh environmental conditions such as wind, waves, and currents, resulting in high economic costs and long operation cycles; secondly, the sampling process inevitably disturbs the soil samples, affecting the accuracy of the data. To overcome the limitations of borehole sampling, pore pressure static cone penetration testing (CPTU) technology, due to its advantages of high efficiency, economy, in-situ testing, and continuous testing, has become an indispensable core tool for marine geotechnical exploration.

[0004] However, these two types of core data differ fundamentally in spatial distribution, physical meaning, and data dimensions, and also have significant shortcomings in practical applications, mainly in the following aspects:

[0005] (1) CPTU continuous data and borehole discrete data are separated. The two have fundamental differences in spatial distribution, physical meaning and data dimensions, making it difficult to directly integrate and utilize them. Current engineering practice mainly relies on subjective experience to make rough and qualitative manual comparisons and stratigraphic divisions between the two, which is inefficient and inconsistent.

[0006] (2) Each borehole test value represents the properties of a certain volume of soil at its depth. Under the conditions of sparse borehole data and dense CPTU data, single-point or single-scale features are difficult to effectively represent the actual soil characteristics, resulting in limited information in feature expression. Moreover, there is a lack of methods to extract equivalent comprehensive features of the soil volume from high-density CPTU signals.

[0007] (3) Existing methods for interpreting soil parameters based on CPTU data usually only provide deterministic prediction results. They only output deterministic results and cannot quantify the confidence level of the prediction results, making it difficult to meet the requirements of high-risk decision-making in engineering for the credibility of the results. Summary of the Invention

[0008] This invention overcomes the above-mentioned defects and provides a multi-source geological exploration data fusion method based on CPTU. It solves the problems in the prior art, such as the separation of borehole discrete data and CPTU continuous data, which makes it difficult to directly integrate and utilize the data, the inability of a single feature to effectively represent the actual soil characteristics, resulting in a weak foundation for intelligent interpretation samples, and the inability to quantify the uncertainty of prediction results.

[0009] To achieve the above objectives, the present invention provides a multi-source geological exploration data fusion method based on CPTU, comprising the following steps:

[0010] S1. Obtain soil parameters from the CPTU dataset and borehole test data points, and perform fusion preprocessing;

[0011] S2. For each borehole test data point, search for the point with the smallest absolute depth difference in the CPTU depth sequence based on the depth value and use it as the matching anchor point.

[0012] S3. Using the matched anchor point as the center, set a dynamic sliding window along the depth direction, and extract three types of features within the window: a) the CPTU dataset corresponding to the anchor point as the center point feature; b) the statistical features of the CPTU data within the window; c) the CPTU multi-scale features of the window.

[0013] S4. Construct a feature set that integrates local context information from the three types of features described in S3. Associate the feature set with the soil parameter labels corresponding to the anchor points to construct a multi-dimensional feature vector and form standardized training samples.

[0014] Furthermore, it also includes step S5, whereby the standardized training samples are used for subsequent training of a multi-output random forest to jointly predict multiple soil parameters; the trained model is then applied to the complete CPTU penetration curve to generate a complete prediction profile including the predicted mean, standard deviation, confidence interval, and information entropy.

[0015] Furthermore, the CPTU dataset includes original parameters and derived parameters. The original parameters include tip resistance, side friction resistance, and pore water pressure. The derived parameters include normalized tip resistance, friction ratio, and pore pressure ratio.

[0016] Furthermore, in step S3, the CPTU multi-scale features are extracted using wavelet transform, as follows:

[0017] S3c1: For the cone tip resistance signal within the sliding window, multi-level wavelet decomposition is performed using the db4 wavelet basis function to obtain the detail coefficients of each level;

[0018] S3c2: For each level of detail coefficients array, when it is valid, calculate the energy and standard deviation of the detail coefficients for that level;

[0019] S3c3: Calculate the total energy and average standard deviation based on the detail coefficients of all effective layers;

[0020] S3c4: The energy and standard deviation of the detail coefficients of each layer, together with the total energy and average standard deviation, constitute the CPTU multiscale feature.

[0021] Furthermore, the wavelet decomposition calculation formula is as follows:

[0022] ,

[0023] Where J is the number of decomposition layers, A J D is the approximate coefficient of layer J. j These are the detail coefficients of layer j;

[0024] The formula for calculating the standard deviation is as follows:

[0025] ,

[0026] in, Let be the mean of the detail coefficients at the j-th layer. It is the number of coefficients in this layer.

[0027] Furthermore, in step S3, the statistical characteristics of the CPTU data within the window are multiple statistics of the cone tip resistance signal calculated within the window.

[0028] Furthermore, in step S3c1, before performing wavelet decomposition, it is determined whether the length of the cone tip resistance signal meets the preset decomposition layer requirement; if it does not meet the requirement, the decomposition layer is automatically adjusted to the maximum safe number of layers allowed by the current signal length.

[0029] Furthermore, the sliding window is a symmetrical sliding window with a fixed number of points centered on the matching anchor point.

[0030] Furthermore, in step S4, after constructing a feature set for each drilling point, the percentage of effective features in the feature set is calculated. Only when the percentage exceeds 50% is the data point retained for training sample construction.

[0031] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0032] This invention constructs a fusion framework of "multi-source heterogeneous data-soil parameter mapping" by performing depth alignment and feature fusion of discrete borehole test data and continuous pore pressure static cone penetration test data. Compared to the traditional method of "using discrete borehole data and continuous CPTU data separately," this invention constructs a multi-dimensional feature space of "borehole-cone tip resistance-side friction resistance-pore water pressure," achieving spatial correlation and information complementarity between discrete borehole test data and continuous CPTU test data. Through the nearest neighbor depth matching algorithm, it solves the problems of inconsistent sampling intervals and depth benchmarks between borehole and CPTU data in traditional methods, eliminating information loss caused by data fragmentation. A sliding window method is used to extract multi-dimensional features of each window along the depth direction, including original / derived parameter features, window statistical features, and wavelet transform features, successfully transforming discrete borehole test points into training samples with spatial context. An optimized random forest multi-output regression algorithm is used to establish a quantitative prediction model from the CPTU multi-source feature space to seven soil parameters: clay content, silt content, sand content, liquid limit, plastic limit, plasticity index, and undrained shear strength. Ultimately, this achieves intelligent fusion of borehole and CPTU data and the construction of a sample library.

[0033] Furthermore, this invention constructs an uncertainty quantification system based on information entropy. Through probability analysis, confidence interval estimation, and visualization assessment of prediction reliability, it establishes a standardized process from data acquisition to result output, solving the problems of inconsistent processing procedures and poor comparability of results in traditional methods, and ensuring the reliability of the model under different geological conditions.

[0034] The full-profile continuous prediction and uncertainty quantification system of this invention expands "point-like" test data into "continuous profile + uncertainty assessment", providing a reliable basis for key engineering decisions such as offshore platform pile foundation design, submarine pipeline route selection, and submarine geological hazard assessment. Attached Figure Description

[0035] Figure 1 This is a flowchart of a multi-source geological exploration data fusion method based on CPTU according to the present invention;

[0036] Figure 2 This is a schematic diagram of the continuous prediction profile of clay particles, powder particles, sand particles, and liquid limit in the whole borehole of the present invention;

[0037] Figure 3 This is a schematic diagram of a continuous prediction profile of the plastic limit, plasticity index, and undrained shear strength of the fully drilled borehole according to the present invention.

[0038] Figure 4 This is a schematic diagram of the quantitative profile of the continuous prediction uncertainty of clay particles, powder particles, sand particles, and liquid limit in the whole borehole of the present invention;

[0039] Figure 5This is a schematic diagram of the quantitative profile of the continuous prediction uncertainty of the plastic limit, plasticity index, and undrained shear strength of the fully drilled borehole according to the present invention. Detailed Implementation

[0040] The technical solutions in 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.

[0041] like Figure 1 As shown, this invention proposes a multi-source geological exploration data fusion method based on CPTU, as detailed below:

[0042] I. Conduct CPTU field penetration tests and laboratory geotechnical tests to obtain the original CPTU parameters (cone tip resistance q). t Side friction resistance f s Pore ​​water pressure u2) and derived parameters (normalized cone tip resistance Q) t Friction ratio F r , pore pressure ratio B q Based on the borehole soil samples, the physical and mechanical parameters of the soil obtained from the geotechnical tests are obtained (clay content, silt content, sand content, liquid limit, plastic limit, plasticity index, and undrained shear strength).

[0043] II. Preprocessing for fusing borehole data and CPTU data: Based on depth, the borehole dataset (depth-soil parameter labels) and the CPTU dataset (depth-cone tip resistance q) are processed. t Side friction resistance f s The pore water pressure u2 curve undergoes automatic depth identification and alignment, data cleaning, and robust standardization. The robust standardization calculation formula is as follows:

[0044] ,

[0045] Where Q2 is the median of feature x, and IQR is the interquartile range of feature x.

[0046] Third, construct a multi-dimensional feature engineering method. Match the geotechnical test data points from the boreholes to the CPTU depth sequence based on their depth values. Using each matching point as the center point, an adaptive sliding window method is employed to extract multi-dimensional features within the window, including the CPTU center point features corresponding to the anchor point, the statistical features of the CPTU data within the window, and the multi-scale features of the CPTU within the window. Construct a local contextual information feature set, as follows:

[0047] ,

[0048] The first row contains CPTU center point features, the second row contains statistical features of CPTU data within the window, and the third row contains CPTU multi-scale features within the window.

[0049] Specifically, (1) the geotechnical test data points of each borehole are matched to the CPTU depth sequence according to the depth value using the nearest neighbor matching algorithm, the absolute depth difference is calculated, and the CPTU point with the smallest difference is selected as the matching point. The borehole point is anchored to the point with the closest physical position on the CPTU profile. With the matching point as the center, the symmetrical sliding window method with 11 points is preferred to extract the CPTU center point features within the window range. At the same time, the validity of the matching points also needs to be verified: (1) check whether the CPTU depth points after matching exceed the valid data range, and automatically remove those borehole points whose depth exceeds the CPTU measurement range to avoid invalid matching; (2) after extracting the CPTU center point features of the matching point, the validity of these feature values ​​will be checked immediately. Only when more than 30% of the CPTU parameters at the matching point are non-null values ​​is the matching considered valid. This verification mechanism directly filters out invalid matching caused by depth misalignment or missing data, ensuring the reliability of anchoring.

[0050] In this embodiment, the CPTU center point is characterized by cone tip resistance q. t Side friction resistance f s Pore ​​water pressure u2, normalized cone tip resistance Q t Friction ratio F r , pore pressure ratio B q ;

[0051] (2) Calculate q within the window t Several statistics of the signal, including the mean K mean Standard deviation K std Minimum value K min Maximum value K max Range K range Skewness K skewness kurtosis K kurtosis These statistics collectively characterize the central tendency, dispersion, and distribution pattern of the soil mechanical response within the window, and are more representative of the overall behavior of a "soil unit" than a single center point value.

[0052] (3) Perform multi-scale wavelet transform algorithm on the qt signal within the window to extract wavelet transform features. Perform 3-level "db4" wavelet decomposition on the qt signal within the fixed window. There is a minimum requirement for the length of the input signal during decomposition. Therefore, before wavelet decomposition, the input signal (i.e., the qt signal within the window) will be decomposed. tThe algorithm determines the number of decomposition levels based on the sequence length. If the minimum requirement for a 3-level decomposition is met, wavelet decomposition is performed. Otherwise, the maximum allowed decomposition level for the current signal length is automatically calculated, and the number of decomposition levels is dynamically reduced to this safe value. This ensures the algorithm will not crash when faced with window signals of any length, achieving robustness and guaranteeing smooth execution of the entire process and high feature extraction coverage. The wavelet decomposition and standard deviation calculation formulas are as follows:

[0053] ,

[0054] Where J is the number of decomposition layers, A J D is the approximate coefficient of layer J. j These are the detail coefficients of layer j;

[0055] ,

[0056] in, Let be the mean of the detail coefficients at the j-th layer. This is the number of coefficients in that layer;

[0057] After wavelet decomposition, the length of the detail coefficient array is checked layer by layer. Only when the detail coefficient array of a certain layer is not empty is the energy (sum of squared coefficients) and standard deviation of that layer calculated. The extracted wavelet transform features include q t Signal 3-layer detail factor energy 3-layer detail coefficient standard deviation Total Energy Mean standard deviation For invalid layers, the corresponding feature terms will be assigned a specific identifier value.

[0058] The multi-scale wavelet transform algorithm employed in this invention deeply integrates the multi-scale analysis capabilities of wavelet transform into the specific engineering problem of CPTU soil parameter prediction through dynamic adaptive adjustment of the number of layers, layer-by-layer safe computation mechanism, engineering meaning mapping model, and fixed efficient processing strategy. This technical solution is not only robust, computationally efficient, and interpretable, but also generates enhanced features with clear physical meaning, thereby significantly improving the model's ability to identify and predict complex geological structures, especially thin-layer structures.

[0059] IV. Constructing a multi-dimensional feature vector based on the fusion of borehole and CPTU data: The extracted CPTU center point features, window statistical features, and wavelet transform features, along with the corresponding seven soil parameter labels, are used to construct a labeled CPTU dataset as training samples. This transforms a single discrete test point into multiple training samples with spatial context, effectively addressing the sparsity problem of geotechnical test samples. The constructed feature vector has a 21-dimensional dimension, comprehensively describing the multi-scale characteristics of the soil. The feature vector is shown below:

[0060] ,

[0061] In the left-hand label, the first row represents CPTU center point features, the second row represents window statistical features, and the third row represents wavelet transform features. The right-hand label shows the following data in order: the first row represents clay content, silt content, and sand content; the second row represents liquid limit and plastic limit; and the third row represents undrained shear strength S. u Plasticity Index I p .

[0062] After constructing a feature set for each borehole point containing all CPTU center point features, window statistical features, and wavelet features, the percentage of valid (non-NaN) features in this feature set is first calculated. Only when this percentage exceeds 50% is the data point retained for subsequent training. This means that if a window signal around a matching point causes a large number of feature calculation failures (such as generating NaNs) due to severe noise interference, the sample will be automatically discarded. Although this strategy does not directly repair noise, it effectively prevents low-quality, unreliable feature data from entering the machine learning model, ensuring the cleanliness of the training data from the source.

[0063] This invention proposes a borehole data anchoring and dynamic sliding window feature extraction mechanism, achieving spatial alignment and contextual fusion of discrete borehole data and continuous CPTU data. This mechanism combines precise anchoring in the depth direction with local feature extraction in the spatial dimension: a nearest neighbor matching algorithm accurately locates geotechnical test points to the CPTU depth sequence, establishing a reliable correspondence between discrete and continuous data, ensuring the accuracy of feature extraction location; and based on an adaptive sliding window strategy, it extracts center point features, statistical features, and multi-scale wavelet features within the window to construct a local contextual feature set characterizing the soil properties surrounding the anchoring point. The synergy between the anchoring mechanism and window features effectively integrates the accuracy of discrete test data with the density of continuous test data, significantly improving the model's ability to represent spatial variation patterns in the stratigraphy and providing a reliable spatial benchmark for multi-source data fusion.

[0064] V. Train a multi-objective random forest model using the labeled CPTU dataset to establish a nonlinear mapping relationship from the multi-dimensional feature space to seven soil parameters. The specific steps are as follows:

[0065] a. Divide the constructed dataset into an 80% training set and a 20% test set;

[0066] b. The random forest model optimized using a multi-output regressor, with the following parameters set: 300 decision trees, maximum depth of 30 layers, minimum split sample size of 5, minimum leaf node sample size of 2, 42 random seeds, and 80% of the samples used per tree.

[0067] c. Perform 5-fold cross-validation on the training set using R. 2 RMSE and MAE are used to evaluate model performance.

[0068] VI. Apply the trained multi-objective prediction model to the complete CPTU penetration curve to perform continuous profile prediction of soil parameters at full depth, generating a complete prediction profile including the predicted mean, standard deviation, confidence interval, and information entropy, thus completing the intelligent fusion of borehole and CPTU data. The specific steps are as follows:

[0069] a. Extract CPTU features and wavelet transform features from borehole data points. Slide along the entire CPTU depth sequence with a fixed step size and repeat the feature extraction process (center point features + window statistical features + wavelet features). For boundary points, use an asymmetric window (5 points before / after) for feature extraction. For center points, use a symmetric window for adjustment and extraction to construct a full-depth feature matrix.

[0070] b. Merge the multi-dimensional feature vectors with the soil parameter features to construct a complete labeled CPTU dataset;

[0071] c. Apply the dataset to the trained multi-objective random forest model to perform full borehole prediction and obtain the predicted values ​​of 7 soil parameters at each depth point.

[0072] d. Combining the predicted mean and standard deviation, calculate the 95% confidence interval (mean ± 1.96 × standard deviation) based on the normal distribution assumption. Treat the predicted value of each tree as a distribution, discretize it, and calculate its information entropy as another measure of uncertainty.

[0073] This invention achieves a leap from discrete point estimation to continuous profile interpretation in soil parameter prediction by fusing full-depth continuous profile prediction with uncertainty quantification, while simultaneously quantifying the uncertainty of the prediction results. This mechanism integrates high-resolution continuous prediction with multi-dimensional uncertainty assessment: by applying a trained multi-objective random forest model to the complete CPTU penetration curve, it achieves full-depth borehole slip prediction, generating a continuous soil parameter profile with the same resolution as the original test data, thus overcoming the discreteness limitations of traditional interpolation methods.

[0074] VII. Visualize the prediction results, as follows:

[0075] a. Comparison of actual and predicted values: This chart shows the model's prediction performance on the test set, with metrics such as R², RMSE, and MAE given in Table 1.

[0076] b. Multi-parameter profile: Draw a continuous profile of 7 soil parameters along the depth and superimpose borehole test points for verification;

[0077] c. Uncertainty cloud map: The distribution characteristics of uncertainty in the depth-parameter space are shown by scatter plot cloud map.

[0078] Table 1 shows the calculation results of the test set data for the method of this invention.

[0079]

[0080] This invention achieves intelligent integration of discrete borehole data and continuous CPTU data through a multi-source data fusion framework. It captures the multi-scale characteristics of soil through multi-dimensional feature engineering, establishes a prediction model using an optimized random forest multi-output regression algorithm, and quantifies the uncertainty of the prediction results using information entropy. Compared with traditional methods, this invention represents a technological leap from "discrete point evaluation" to "continuous profile interpretation," from "single predicted value" to "prediction + uncertainty," and from "human experience interpretation" to "automated intelligent analysis."

[0081] Figure 2 , Figure 3 The comparison between predicted values ​​(continuous curves in the figure) and measured values ​​(scattered points in the figure) of seven soil parameters, including clay content and liquid limit, is presented. Table 1 shows that the predicted R² values ​​for all seven parameters exceed 0.8, demonstrating the high accuracy of the "CPTU feature-soil parameter" mapping relationship established by multi-source data fusion and the random forest model. The generated continuous predicted profile along the borehole depth clearly shows that this invention can expand sparse "point-like" experimental data into high-resolution "continuous" parameter curves and can finely characterize thin interlayers (local curve fluctuations), empirically demonstrating the leap from "discrete sampling" to "continuous interpretation" and solving the data fragmentation problem.

[0082] Figure 3 , Figure 4 A quantitative profile of the uncertainty in the prediction of seven soil parameters, including clay content and liquid limit, is presented, representing the uncertainty values ​​at each CPTU prediction point and corresponding to locations of significant formation variations. This upgrades the traditional single prediction value to a "predicted value + reliability assessment," providing crucial risk warning information for engineering decisions and resolving the problem of missing prediction results. In summary, Figures 2 to 5 The invention's method demonstrates significant advantages in achieving intelligent fusion of multi-source data, constructing a reliable sample library, and producing practical engineering results from the dimensions of accuracy, continuity, and uncertainty.

[0083] This invention constructs a complete output system encompassing predicted mean, fluctuation range, and reliability assessment by integrating information entropy calculation, confidence interval estimation, and standard deviation analysis within a learning framework. The synergistic effect of continuous prediction and uncertainty quantification simultaneously provides high-precision parameter profiles and clear risk warning information, significantly enhancing the engineering practical value of intelligent interpretation results and providing geological parameter support for marine geotechnical engineering design that combines continuity, accuracy, and reliability.

[0084] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for fusing multi-source geological exploration data based on CPTU, characterized in that, Includes the following steps: S1. Obtain soil parameters from the CPTU dataset and borehole test data points, and perform fusion preprocessing; S2. For each borehole test data point, search for the point with the smallest absolute depth difference in the CPTU depth sequence based on the depth value and use it as the matching anchor point. S3. Using the matched anchor point as the center, set a dynamic sliding window along the depth direction, and extract three types of features within the window: a) the CPTU dataset corresponding to the anchor point as the center point feature; b) the statistical features of the CPTU data within the window; c) the CPTU multi-scale features of the window. The CPTU multi-scale features are extracted using wavelet transform, as follows: S3c1: For the cone tip resistance signal within the sliding window, multi-level wavelet decomposition is performed using the db4 wavelet basis function to obtain the detail coefficients of each level. The wavelet decomposition calculation formula is as follows: ， Where J is the number of decomposition layers, A J D is the approximate coefficient of layer J. j These are the detail coefficients of layer j; S3c2: For each level of detail coefficients array, when it is valid, calculate the energy and standard deviation of the detail coefficients for that level. The standard deviation is calculated using the following formula: ， in, Let be the mean of the detail coefficients at the j-th layer. This is the number of coefficients in that layer; S3c3: Calculate the total energy and average standard deviation based on the detail coefficients of all effective layers; S3c4: The energy and standard deviation of the detail coefficients of each layer, together with the total energy and average standard deviation, constitute the CPTU multiscale feature; S4. Construct a feature set that integrates local context information from the three types of features described in S3. Associate the feature set with the soil parameter labels corresponding to the anchor points to construct a multi-dimensional feature vector and form standardized training samples.

2. The method for fusing multi-source geological exploration data based on CPTU according to claim 1, characterized in that, It also includes step S5, whereby the standardized training samples are used for subsequent training of a multi-output random forest to jointly predict multiple soil parameters; the trained model is then applied to the complete CPTU penetration curve to generate a complete prediction profile including the predicted mean, standard deviation, confidence interval, and information entropy.

3. The method for fusing multi-source geological exploration data based on CPTU according to claim 1, characterized in that, The CPTU dataset includes raw parameters and derived parameters. The raw parameters include tip resistance, side friction resistance, and pore water pressure. The derived parameters include normalized tip resistance, friction ratio, and pore pressure ratio.

4. The method for fusing multi-source geological exploration data based on CPTU according to claim 1, characterized in that, In step S3, the statistical characteristics of the CPTU data within the window are multiple statistics of the cone tip resistance signal calculated within the window.

5. The method for fusing multi-source geological exploration data based on CPTU according to claim 1, characterized in that, In step S3c1, before performing wavelet decomposition, it is determined whether the length of the cone tip resistance signal meets the preset decomposition layer requirement; if not, the decomposition layer is automatically adjusted to the maximum safe number of layers allowed by the current signal length.

6. The method for fusing multi-source geological exploration data based on CPTU according to claim 1, characterized in that, The sliding window is a symmetrical sliding window centered on the matching anchor point and with a fixed number of points.

7. The method for fusing multi-source geological exploration data based on CPTU according to claim 1, characterized in that, In step S4, after constructing a feature set for each drilling point, the percentage of effective features in the feature set is calculated. Only when the percentage exceeds 50% is the data point retained for training sample construction.

Images