Knowledge-guided and data-driven integrated static cone penetration soil layer automatic division method

Abstract

This invention discloses an automatic soil layer segmentation method based on knowledge-guided and data-driven approaches, comprising: S1, collecting CPT penetration data and manual stratification results, and performing data preprocessing; S2, extracting local statistical features and morphological / relative relationship features using a multi-scale sliding window based on engineering experience, and constructing a temporal feature set; S3, constructing a multi-task learning framework based on a bidirectional long short-term memory network, setting a knowledge-driven rule layer, and generating soft labels and real labels for joint supervision; S4, training the model using the training set; S5, automatically segmenting the test set, fusing the prediction results of overlapping windows, and outputting the final soil layer segmentation and boundary point identification results. Through the deep integration of knowledge-guided and data-driven approaches, the accuracy and consistency of soil layer classification and boundary identification are significantly improved, exhibiting good interpretability and cross-site generalization ability, and is suitable for rapid profile generation in geotechnical engineering investigations.

Description

Technical Field

[0001] This invention relates to the field of soil layer division in geotechnical engineering, specifically to an automatic method for static cone penetration testing soil layer division that integrates knowledge guidance and data-driven approaches. Background Technology

[0002] Subsurface soil layer delineation is a core task in geotechnical engineering investigation, directly impacting the planning, design, and construction safety of structures and infrastructure. The cone penetration test (CPT) is widely used to obtain subsurface soil profiles due to its ease of operation, low error, high repeatability, and cost-effectiveness. CPT continuously collects parameters such as cone tip resistance and sidewall friction along depth, generating penetration curves that provide ample information for stratification. However, influenced by factors such as transport media and depositional environment, soil exhibits significant heterogeneity and spatial variability, resulting in high dispersion and strong fluctuations in CPT penetration curves, posing challenges to automatic stratification based on curve characteristics. Engineering practice typically relies on empirical manual interpretation and merging of thin layers, but this process is time-consuming, subjective, and lacks consistent standards, making it difficult to maintain consistency in large-scale projects. Therefore, there is an urgent need for an automatic identification method that combines objectivity, robustness, and universality to improve stratification consistency, reduce reliance on manual labor, and enhance the reliability of investigation results, thereby better supporting engineering design optimization and construction risk control.

[0003] In recent years, scholars both domestically and internationally have proposed various data-driven methods for automatic soil layer classification in the CPT (Concentrated Terrain Perspective) model, mainly including two paths: statistical analysis and machine learning. In terms of statistical analysis, Lin Jun et al. used intragroup correlation coefficients and modified Bartlett's test to conduct statistical analysis of CPT data, proposing a mechanical stratification method for identifying homogeneous soil layer boundaries, and verified its effectiveness in revealing primary and secondary stratification interfaces through measured data; Ching et al. used wavelet transform modulus maxima technology to detect soil layer boundary points. In terms of machine learning, Cai Guojun et al. constructed a CPT soil layer classification method based on cluster analysis; Cai et al. used a generalized regression neural network, with modified cone tip resistance and pore water pressure as inputs, to achieve automatic soil category classification; Carvalho et al. integrated multiple machine learning models and tested different feature combinations on several representative CPT datasets, systematically evaluating the performance of each algorithm in soil classification tasks.

[0004] The aforementioned methods, by learning from expert annotations or historical engineering data, can quickly output consistent and reproducible stratification results during the inference stage, significantly reducing human error and improving efficiency. However, existing methods often treat CPT samples as independent, rarely explicitly characterizing their continuity and sequence dependence along the depth direction. In fact, CPT penetration curves are essentially sequential data that evolves with depth, containing obvious spatial and stratigraphic evolution characteristics. Using deep neural networks (such as bidirectional recurrent networks) with time-dependent modeling capabilities to model penetration data is expected to more effectively uncover internal structures and improve stratification accuracy and boundary identification stability. At the same time, single data-driven paradigms often rely on data quality and scale, facing problems such as data sparsity, heterogeneity, and insufficient interpretability under complex working conditions, thus limiting generalization ability. Explicitly integrating geological exploration knowledge into the model in the form of rules and priors, constraining and guiding the learning process, can improve the model's interpretability and cross-site adaptability while ensuring efficiency. Therefore, a knowledge-data dual-driven automatic CPT soil layer division method has become a key development direction for promoting intelligent stratigraphic identification and engineering implementation. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes an automatic static cone penetration test (CPPT) layer segmentation method that integrates knowledge-guided and data-driven approaches. This method combines geological engineering empirical rules with temporal modeling using deep learning, improving layering accuracy and cross-site generalization ability while ensuring engineering interpretability.

[0006] To achieve the above objectives, this invention provides an automatic method for dividing soil layers using static cone penetration testing, which integrates knowledge-guided and data-driven approaches. Includes the following steps: S1. Data Acquisition and Preprocessing: Collect static cone penetration data and quality-controlled manual stratification results; divide the training set and test set by borehole; scale the input parameters proportionally using a uniform scaling factor, limiting their value range to between 0 and 1, while maintaining the original numerical ratio, relative size, and relative positional relationship between different parameters. S2. Knowledge-driven feature extraction: Based on engineering experience, a multi-scale sliding window is used to extract features from the static penetration data. The extracted features include at least local statistical features and morphological / relative relationship features, forming an information-based time-series feature set. S3. Model Construction: Construct a multi-task learning framework with a bidirectional long short-term memory network as the main body. This framework simultaneously outputs the probability that the target depth point is a soil layer boundary point and the probability distribution of the corresponding soil layer category. Embed a knowledge-driven rule layer at the front end of the main network of the multi-task learning framework. The knowledge-driven rule layer is used to generate soft labels based on the feature-soil layer correspondence established based on engineering experience. The soft labels and the real labels from the manual stratification results jointly supervise the model training. S4. Model Training: Using the temporal features of the training set, and combining the real labels and the soft labels generated by the knowledge-driven rule layer, the multi-task learning framework constructed in step S3 is trained end-to-end to obtain the soil layer automatic segmentation model. S5. Automatic segmentation and result fusion: Input the temporal features of the test set into the trained automatic soil segmentation model to obtain preliminary segmentation results; fuse the results of the same depth point in multiple predictions to obtain the final soil segmentation and boundary point identification results.

[0007] More preferably, in step S2, the multi-scale sliding window includes at least a first sliding window for extracting local statistical features and a second sliding window for extracting morphological / relative relationship features, wherein the window length of the first sliding window is less than the window length of the second sliding window; the local statistical features include the mean, standard deviation and slope calculated for the cone tip resistance, sidewall friction and friction ratio respectively within the sliding window.

[0008] More preferably, the morphological / relative relationship features include at least the undulation F1 calculated based on the cone tip resistance curve, the peak intensity H, and the relative position R calculated based on the cone tip resistance and sidewall friction curve. mean With alternating frequency R cross .

[0009] Further preferably, in step S3, the knowledge-driven rule layer constructs rule response functions corresponding to each input feature based on engineering experience, which are used to characterize the contribution of different features to different soil layer categories; wherein, the cone tip resistance curve undulation... F Peak intensity of cone tip resistance curve H The alternation frequency of the cone tip resistance and sidewall friction curves R cross The corresponding rule-based response function employs a sigmoid-type learnable gating function, whose parameters are adaptively updated during model training; the relative positions of the cone tip drag and sidewall friction curves... R mean The corresponding rule response function adopts a condition adjustment function based on the indicator function.

[0010] Further preferably, the real labels are scaled proportionally, and the outputs of the rule response functions in the knowledge-driven rule layer are used to obtain the rule contribution coefficients of each feature to the corresponding soil layer category, and the scaled labels are modulated.

[0011] More preferably, in step S3, the loss function of the multi-task learning framework is a joint loss function, which includes at least a classification loss term generated by the real label supervision, a knowledge distillation loss term generated by the soft label supervision, and a boundary point detection loss term.

[0012] More preferably, in step S3, the multi-task learning framework includes 1 to 4 layers of Bi-LSTM, an attention layer, and a normalization layer connected in sequence; the output layer of the multi-task learning framework includes a boundary probability output branch using a Sigmoid activation function and a soil category probability output branch using a Softmax activation function.

[0013] More preferably, in step S5, the fusion process uses soft voting and / or probability-weighted averaging to fuse the soil layer category probability and boundary point probability of the same depth point in multiple overlapping prediction windows; and determines the final boundary point position according to a preset probability threshold.

[0014] More preferably, in step S1, the static penetration data includes at least the cone tip resistance and sidewall friction directly measured by the dual-bridge static penetration probe, and the friction ratio calculated based on the cone tip resistance and sidewall friction.

[0015] More preferably, in step S1, the preprocessing further includes: performing unit unification processing on the static penetration data, and processing outliers and missing values.

[0016] The beneficial effects of this invention are as follows: The present invention proposes a knowledge-guided and data-driven method for automatic static cone penetration testing of soil layers. By organically combining engineering experience rules with deep learning modeling, it achieves efficient and accurate identification of soil layer categories and boundary points, and has the following significant technical effects: 1. Improved Layering Accuracy and Consistency: By extracting features that combine local statistics and overall morphology through a multi-scale sliding window, and utilizing a bidirectional long short-term memory network (Bi-LSTM) for temporal modeling of the penetration curve, the structural dependencies in the depth direction are fully explored, significantly improving the accuracy of soil layer classification and boundary identification. Experiments show that this method outperforms traditional random forests and deep neural networks in both soil layer classification and boundary point detection, and the layering results are more stable and reliable.

[0017] 2. Enhance model interpretability and engineering credibility: By introducing a knowledge-driven rule layer at the front end of the main network, engineering experience is embedded into the model training process in the form of rule response functions and soft labels, making the discrimination boundary more consistent with the actual stratigraphic distribution pattern, improving the transparency and interpretability of the model, and making it easier for engineers to understand and adopt.

[0018] 3. Excellent cross-site generalization ability: The joint supervision mechanism of soft labels and real labels, as well as the knowledge-guided feature extraction strategy, effectively reduce the model's dependence on specific site data, enabling it to maintain high recognition accuracy and adaptability in unseen engineering projects, and making it suitable for large-scale exploration applications under different geological conditions.

[0019] 4. Significantly improves exploration efficiency and quality control: This method realizes end-to-end automatic division from data to profile, significantly reducing the time and subjective differences of manual interpretation, which is conducive to the rapid and consistent generation of stratigraphic profiles in large-scale projects, and provides efficient and reliable technical support for geotechnical engineering design and construction. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the steps of the method of the present invention; Figure 2 This is a comparison chart of the short-range feature extraction results of a static cone penetration test (CPT) curve before and after processing; Figure 3 This is a schematic diagram of the multi-task learning framework of the method of the present invention; Figure 4 This is a comparison chart of the soil layer classification accuracy of the method of this invention with that of Random Forest (RF) and Deep Neural Network (DNN); Figure 5 This is a comparison chart of the boundary point identification accuracy of the method of the present invention with that of Random Forest (RF) and Deep Neural Network (DNN); Figure 6 This is a schematic diagram of the soil layer division results based on static cone penetration test data from an engineering project that did not participate in the training, using the method of the present invention. Detailed Implementation

[0021] 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.

[0022] Example 1 Please see Figure 1This invention provides a technical solution: an automatic method for dividing static cone penetration soil layers by integrating knowledge guidance and data-driven approaches, the steps of which are as follows: Step 1: Collect raw CPT data and perform data preprocessing. The specific process is as follows: (1) Collect raw CPT penetration data, including borehole number and cone tip resistance directly measured by the double-bridge static cone penetration probe. q c and sidewall friction f s And the friction ratio calculated based on the two. R f And the results of manual stratification after quality control.

[0023] (2) Unit unification: convert the units of cone tip resistance and side wall friction to MPa and 10kPa respectively.

[0024] (3) Perform proportional scaling processing. Apply a uniform scaling factor to scale the cone tip resistance, side wall friction and friction ratio proportionally, so that the values ​​of each parameter are limited to the range of 0 to 1, and the original numerical proportional relationship, relative size relationship and relative position relationship between the three remain unchanged. (4) Divide the training set and the test set according to the borehole number in a preset ratio. The test set ratio can be selected from 10% to 30% to simulate the application scenario of the model at a new exploration point and avoid leakage of context information.

[0025] Step 2: Knowledge-Driven Feature Extraction: Combining engineering stratification experience, a multi-scale sliding window is used to extract features from the static cone penetration test data to obtain an information-based temporal feature set that can characterize the mechanical properties and curve morphology of the soil layer. The multi-scale sliding window includes at least a first sliding window and a second sliding window, wherein the first sliding window is used to extract local statistical features, and the second sliding window is used to extract morphological / relative relationship features. In this embodiment, the first sliding window is also called a short window, and the second sliding window is also called a long window. The specific process is as follows.

[0026] (1) First sliding window feature extraction: For each curve, a sliding window with a length of 2 to 5 and a sliding step of 1 to 3 is used along the depth direction based on the number of sampled data points. The mean, standard deviation and slope of the cone tip resistance, side wall friction and friction ratio are calculated in the sliding window, and a total of 9 statistical features are obtained, which not only preserves the intensity information, but also reflects the fluctuation amplitude and change trend.

[0027] (2) Second sliding window feature extraction: In order to make full use of the correspondence between "curve shape and soil type" summarized by engineers in practice, a second sliding window with a window length of 8 to 15 is used along the depth direction to extract mesoscale geometric features from the static cone penetration curve. The shape / relative relationship features mainly include the undulation of the cone tip resistance curve. F Peak intensity of cone tip resistance curve H The relative positions of the cone tip resistance and sidewall friction curves R mean With alternating frequency R cross The calculation formula is as follows:

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] In the formula, N This represents the number of sampling points in the sliding window. q i For the first i The cone tip resistance value at each sampling point; q p,j For the first j The peak cone resistance of a significant peak; a significant peak refers to a peak significance greater than the threshold. thr The local maxima of the threshold, where the threshold is , and These are the maximum and minimum values ​​of the cone tip resistance within the current sliding window, respectively; K This represents the number of significant peaks within the current sliding window. q L,j , q R,j For the first j The cone-shaped resistance at the nearest trough or inflection point to the left or right of a significant peak; fs i , qc i For the first i Sidewall friction and cone tip resistance at each sampling point; To prevent the use of tiny constants with a denominator of zero, a value of 0.00001 is used; For the first iThe relative difference between sidewall friction and cone tip resistance at each sampling point; This is a sign function; it takes the value 1 when the input is greater than 0, 0 when it is equal to 0, and -1 when it is less than 0. δ (·) is an indicator function; it takes the value 1 if the condition inside the parentheses is met, and 0 otherwise.

[0034] (3) Dataset construction The length of the data segment used to construct the time series samples is set to 20-80 sampling data points, and the sample sliding step size is set to 1-10 sampling data points. On each static penetration curve, the corresponding feature sequence and its artificial hierarchical label are extracted based on the above data segment to construct the model training dataset to meet the requirements of the Bi-LSTM neural network for the time series input data format.

[0035] Step 3: Construct a multi-task learning framework based on a Bidirectional Long Short-Term Memory (Bi-LSTM) network, simultaneously outputting the probability distribution of the target depth point as the boundary point and its corresponding soil layer category; set a knowledge-driven rule layer at the front end of the main network, establish the correspondence between features and soil layers based on engineering experience, generate soft labels and jointly supervise them with the real labels to guide the judgment boundary to conform to the stratigraphic distribution pattern. The specific process is as follows: (1) Construction of a multi-task learning framework The network adopts a bidirectional LSTM structure, which consists of 1 to 4 Bi-LSTMs, a multi-head attention layer, a batch normalization layer, and a dual-output layer at the end. The dual-output layer uses the sigmoid and softmax activation functions respectively to output the probability distribution of the target depth sampling point becoming the boundary point and its corresponding soil layer category.

[0036] The dual output layer operates on the temporal features in a sampling-per-position sharing manner. The soil layer category probability output branch uses a fully connected layer and a Softmax activation function to output the soil layer category probability point by point, while the boundary point probability output branch uses a fully connected layer and a Sigmoid activation function to output the boundary point probability point by point.

[0037] (2) Set up a knowledge-driven rule layer a. Extracting soil layer division rules Cone tip resistance curve undulation F The smaller the value of 1, the greater the probability that the soil layer is silt, silty clay, clay, or silty clay; peak strength of cone tip resistance. H The larger the value, the greater the probability of a fine sand layer; R mean>0 indicates that the sidewall friction is usually on the right side of the cone tip resistance curve, and the probability of it being silt, silty clay, clay, or silty clay is relatively high. The more intersections there are between the sidewall friction and the cone tip resistance curve, the greater the probability of it being plain fill, miscellaneous fill, hydraulic fill, silt, silty sand, or fine sand.

[0038] b. Establish the rule response function The following rule response function is introduced to adjust the rule contribution of different features to the corresponding soil layer categories:

[0039]

[0040]

[0041]

[0042] In the formula, , , and These represent the regular response values ​​corresponding to the undulation of the cone tip resistance curve, the peak intensity, the alternation frequency of the cone tip resistance and sidewall friction curves, and their relative position characteristics, respectively, and are used to characterize the regular contribution of each feature to the corresponding soil layer category. σ (·) represents the Sigmoid function; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. k f1 , k h , k rcross This is a learnable feature amplification factor used to adjust sensitivity; the initial value is preferably set between 1 and 20. τ f1 , τ h , τ rcross The corresponding learnable feature threshold is used to control the response range, and the initial value is preferably set between 0 and 1; k mean Relative position features R mean The adjustment coefficient corresponding to the conditional rule.

[0043] c. Soft tag construction The real label vector y true After performing one-hot encoding, it is scaled according to a certain ratio. κ The preferred value is between 0 and 1, and the corresponding soil layer category is modulated according to the rule contribution coefficient output by each rule response function to obtain a soft label:

[0044] In the formula, 1 fill This is the mask vector for the soil layer category corresponding to the feature; The Sigmoid function is defined as follows: ; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0045] (3) Loss function design For the dual-task learning framework, a joint loss function is designed; the loss function for the soil classification task is a weighted combination of real label supervision and soft label distillation. The loss function for boundary point detection is The final joint function is The calculation formula is as follows:

[0046]

[0047]

[0048] In the formula, CE(·) is the sparse cross-entropy loss; KL(·) is the Kullback-Leibler divergence; y true This represents the true label vector for soil layer categories; This is the predicted probability vector for soil layer categories output by the model. The weighting factor is set to 0.7; The predicted probability of the dividing point; The actual label for the dividing point.

[0049] Step 4, Model Training: Using the information-based temporal feature set carrying prior geological information formed by the training set, the dual-task framework from Step 3 is trained end-to-end to obtain the automatic soil layer segmentation model; the training process can use an adaptive gradient optimization algorithm (such as Adam, Nadam, etc.), and the learning rate is preferably 10. -4 ~10 -2 Within this range, the batch size is preferably between 16 and 128, and the number of training rounds can be adjusted between 20 and 100 depending on the size of the dataset.

[0050] Step 5, Automatic Soil Layer Segmentation and Result Fusion: The generated time-series feature set is input into the trained model to obtain preliminary soil layer segmentation results. For multiple prediction results of the same depth point in different overlapping windows, fusion processing (such as soft voting, probability-weighted averaging, etc.) can be used to reduce local prediction fluctuations and improve the overall consistency and stability of the segmentation results. The fusion strategy in this step is not limited to the above methods and can be flexibly selected according to actual needs.

[0051] In several embodiments, the following evaluation methods may be used to determine the accuracy and stability of the soil layer division results: (1) Soil layer multi-category classification index For each type of soil layer c Treating this category as positive and the rest as negative, calculate the precision for this category. c Recall rate c and F1 value F 1c And give the macro average. F The specific calculation formula for the indicator is as follows:

[0052]

[0053]

[0054]

[0055] In the formula, c For soil layer category index, c =1,2,3…, C Precision c Recall c The first c The classification accuracy and recall rate of soil-like layers; F 1c For the first c The F1 score for soil types is the harmonic mean of the classification precision and recall for that category; Macro-F1 is the macro average of the F1 scores for all categories. TP c For the truth c And the prediction is c The number of samples; FP c For the true or false c But the prediction is c The number of samples; FN c For the truth c But the prediction is not cThe number of samples; C This represents the number of categories.

[0056] (2) Boundary point detection Let B be the set of artificially defined boundaries, and A be the set of model predictions. Given a tolerance... θ (e.g., 10–30 cm), for each a∈A, if there exists a unique b ∈ B Make |ab|≤ θ Then it is recorded as TP, otherwise as FP; unmatched b ∈ B Let FN be the boundary point detection evaluation index.

[0057]

[0058]

[0059] In the formula, TP To predict the number of boundary points that successfully match the manual boundary points within a given tolerance range; FP The number of predicted boundary points that failed to match the artificial boundary points; FN The number of manually matched boundary points that were not predicted; Precision; Recall. The F1 score for boundary point detection is the harmonic mean of the precision and recall of boundary point detection.

[0060] This invention addresses the challenge of balancing accuracy, interpretability, and cross-site adaptability when performing soil layer delineation based on static cone penetration test (CPPT) data. It proposes a knowledge- and data-driven automatic soil profile delineation method. Guided by geological engineering experience, this method extracts features at multiple scales and establishes a knowledge-driven rule layer before the temporal modeling framework. This generates soft labels and jointly supervises them with real-world annotations, constraining the model to learn discrimination boundaries that conform to stratigraphic distribution patterns. This enables efficient and accurate identification of soil layer categories and boundary points. Compared to purely data-driven methods, this invention significantly improves accuracy, stability, and interpretability, and possesses good cross-site generalization capabilities, making it suitable for large-scale engineering applications.

[0061] Example 2: This invention proposes an automatic method for dividing soil layers using static cone penetration testing, which integrates knowledge guidance and data-driven approaches. The implementation process of the technical solution provided by this invention will be explained in detail below using specific engineering data.

[0062] The implementation process of the method of the present invention is illustrated by the following engineering data: The data comes from geotechnical engineering investigation projects of multiple rail transit stations, square foundation pits and large maintenance bases in North China. The data is obtained using a double-bridge static cone penetrometer with a sampling interval of 0.1 m.

[0063] Step 1: Collect 265 sets of static cone penetration test data, totaling 102,016 data points, covering 12 soil types, namely plain fill, miscellaneous fill, dredged fill, silt, clayey silt, clay, silty clay, silty clayey clay, silty clay, silt, silty sand and fine sand, numbered in the above order, with 0 corresponding to plain fill.

[0064] The units for cone tip resistance and sidewall friction are standardized to MPa and 10 kPa, respectively, and the cone tip resistance, sidewall friction, and friction ratio are scaled proportionally using a uniform scaling factor. In this embodiment, the values ​​are uniformly divided by 100 to limit their range to 0 to 1, while maintaining the original numerical ratio, relative size, and relative positional relationship between the three.

[0065] In the dataset partitioning, 30 complete static penetration curves were randomly selected from 265 as the test set, and the remaining 235 were used as the training set. This partitioning method can simulate the application of the model at entirely new exploration sites, effectively avoiding overly optimistic assessments caused by the leakage of contextual information from the same borehole, and making the evaluation results closer to actual engineering needs.

[0066] Step 2: For each curve, a sliding window with a length of 3 and a step size of 1 is used. Within the window, the mean, standard deviation, and slope of the cone tip resistance, sidewall friction, and friction ratio are calculated, resulting in a total of 9 statistical characteristics. Figure 2 This is the result of short-range feature extraction for a CPT curve.

[0067] Mesoscale geometric features were extracted using a window of length 10, mainly including the undulation of the cone tip drag curve. F Peak intensity of cone tip resistance curve H The relative positions of the cone tip resistance and sidewall friction curves R mean With alternating frequency R cross The calculation formula is the same as that in Example 1.

[0068] Set the sliding window length to 50 and the sliding step size to 1. Extract the long and short window features and their corresponding manually hierarchical labels from each CPT curve to construct the model dataset. The dataset shape is { N batch 50 N feature} N batchThe number of dataset groups. N feature The number of input features.

[0069] Step 3: Construct a multi-task learning framework. The framework structure, from top to bottom, consists of: two Bi-LSTMs, one multi-head attention layer, one batch normalization layer, and a dual-output layer at the end. A schematic diagram of the framework is shown below. Figure 3 As shown.

[0070] A knowledge-driven rule layer is set up at the front end of the main network. The construction method of the knowledge-driven rule layer is the same as in Example 1. According to the characteristics of this project, the relevant parameters in the rule response function are initialized: k f1 =-10、 k h =8、 k rcross =8、 k mean =0.6、 τ f1 =0.6、 τ h =0.4、 τ rcross =0.4. Scaling ratio in building soft tags. κ The value was set to 0.6, and the final label calculation formula is as follows:

[0071] In the formula, y soft This is a soft label vector modulated by knowledge rules; y true This represents the true label vector for soil layer categories; F 1. H , R cross , R mean These are the characteristics of fluctuation intensity, local peak prominence, relative position crossover frequency, and relative response intensity, respectively. For the Sigmoid function; For indicator functions; 1 fill This is the mask vector for the soil layer category corresponding to the feature.

[0072] For the dual-task learning framework, a joint loss function is designed, and the design steps are the same as in Example 1.

[0073] Step 4: Using the information-based temporal feature set carrying prior information about the strata formed by the training set, the dual-task framework of Step 3 is trained end-to-end to obtain the automatic soil layer segmentation model; the number of hidden units is 256, the number of multi-head attention mechanisms is 2, the attention head dimension is 64, the optimizer is Nadam, the learning rate is 0.01, and the number of training iterations is 50.

[0074] Step 5: Input the time-series feature set generated from the test set of this project into the trained model to obtain preliminary soil layer division results. For the predicted probability of the same depth point in each overlapping window, a soft-voting fusion strategy is adopted: the predicted probability of each soil layer is weighted and summed according to the number of window overlaps, and the category with the highest cumulative probability is taken as the final predicted type for that depth point; when the probability of a boundary point at that depth point exceeds 0.6, it is determined as a boundary point. The above fusion method used in this project can effectively smooth local prediction fluctuations and improve the stability of the profile division results.

[0075] By performing the above steps, the collected CPT data can be used to construct a knowledge-data dual-driven soil layer division model, and the target site can be automatically divided into soil layers to obtain a clear and highly accurate profile division result.

[0076] To verify the accuracy and stability of the method of this invention, it was compared with two classic data-driven soil layering methods—Random Forest (RF) and Deep Neural Network (DNN). Please refer to [link to relevant documentation]. Figure 4 and Figure 5 The training and test sets are divided along the entire static cone penetration test (CPT) curve, rather than by random shuffling. This strategy more realistically reflects the model's stratification performance at novel exploration points. Therefore, while the numerical accuracy of the test results may be slightly lower than that of conventional methods, the evaluation results are closer to actual engineering scenarios. Experimental results show that the knowledge-data dual-driven method of this invention significantly outperforms the aforementioned traditional machine learning methods in both soil layer classification accuracy and boundary point identification performance, verifying the effectiveness of the "knowledge-guided + data-driven" joint strategy in improving model accuracy, stability, and engineering adaptability.

[0077] To further evaluate the model's generalization ability, static cone penetration test data from an engineering project that was not used in the training—the Tianjin-Baoding New Railway Project—were used to conduct soil layer division tests. For example... Figure 6 As shown, the method of this invention is highly consistent with the results of manual classification in terms of stratigraphic boundary point identification and soil layer naming, and can effectively reproduce the logic of engineering experience in classification. It not only performs stably in known areas, but also demonstrates good generalization ability and engineering applicability in new areas, further reflecting the effectiveness of integrating prior geological knowledge with deep models.

[0078] This invention provides a new technique for soil profile division, which effectively improves the efficiency and accuracy of division based on static cone penetration test data. Its calculation results can provide reliable technical support for the planning, design and construction of geotechnical engineering.

[0079] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can make various modifications, equivalent substitutions or improvements without departing from the spirit and substance of the present invention, and these should all be included within the protection scope of the present invention.

Claims

1. An automatic method for dividing soil layers in static cone penetration tests, integrating knowledge-guided and data-driven approaches. Its features are, Includes the following steps: S1. Data Acquisition and Preprocessing: Acquire static cone penetration data and quality-controlled manual stratification results; The training set and the test set are divided into units based on the borehole. The input parameters are scaled proportionally using a uniform scaling factor, and their values ​​are limited to between 0 and 1. The original numerical ratio, relative size, and relative positional relationship between different parameters are kept unchanged. S2. Knowledge-driven feature extraction: Based on engineering experience, a multi-scale sliding window is used to extract features from the static penetration data. The extracted features include at least local statistical features and morphological / relative relationship features, forming an information-based time-series feature set. S3. Model Construction: Construct a multi-task learning framework with a bidirectional long short-term memory network as the main body. This framework simultaneously outputs the probability that the target depth point is a soil layer boundary point and the probability distribution of the corresponding soil layer category. Embed a knowledge-driven rule layer at the front end of the main network of the multi-task learning framework. The knowledge-driven rule layer is used to generate soft labels based on the feature-soil layer correspondence established based on engineering experience. The soft labels and the real labels from the manual stratification results jointly supervise the model training. S4. Model Training: Using the temporal features of the training set, and combining the real labels and the soft labels generated by the knowledge-driven rule layer, the multi-task learning framework constructed in step S3 is trained end-to-end to obtain the soil layer automatic segmentation model. S5. Automatic segmentation and result fusion: Input the temporal features of the test set into the trained automatic soil segmentation model to obtain preliminary segmentation results; fuse the results of the same depth point in multiple predictions to obtain the final soil segmentation and boundary point identification results.

2. The method for automatic segmentation of static cone penetration test soil layers integrating knowledge guidance and data-driven approaches according to claim 1, characterized in that, In step S2, the multi-scale sliding window includes at least a first sliding window for extracting local statistical features and a second sliding window for extracting morphological / relative relationship features, wherein the window length of the first sliding window is less than the window length of the second sliding window; the local statistical features include the mean, standard deviation and slope of the cone tip resistance, sidewall friction and friction ratio calculated within the sliding window respectively.

3. The method for automatic segmentation of static cone penetration test soil layers integrating knowledge guidance and data-driven approaches according to claim 1, characterized in that, The morphological / relative relationship features include at least the undulation F1 and peak intensity H calculated based on the cone tip drag curve, and the relative position R calculated based on the cone tip drag and sidewall friction curves. mean With alternating frequency R cross .

4. The method for automatic segmentation of static cone penetration test soil layers integrating knowledge guidance and data-driven approaches according to claim 1, characterized in that, In step S3, the knowledge-driven rule layer constructs rule response functions corresponding to each input feature based on engineering experience, which are used to characterize the contribution of different features to different soil layer categories; among them, the cone tip resistance curve undulation. F Peak intensity of cone tip resistance curve H The alternation frequency of the cone tip resistance and sidewall friction curves R cross The corresponding rule-based response function employs a sigmoid-type learnable gating function, whose parameters are adaptively updated during model training; the relative positions of the cone tip drag and sidewall friction curves... R mean The corresponding rule response function adopts a condition adjustment function based on the indicator function.

5. The method for automatic segmentation of static cone penetration test soil layers integrating knowledge guidance and data-driven approaches according to claim 4, characterized in that, In step 3, the real labels are scaled proportionally, and the output of each rule response function in the knowledge-driven rule layer is used to obtain the rule contribution coefficient of each feature to the corresponding soil layer category, and the scaled labels are modulated.

6. The method for automatic segmentation of static cone penetration test soil layers integrating knowledge guidance and data-driven approaches according to claim 1, characterized in that, In step S3, the loss function of the multi-task learning framework is a joint loss function, which includes at least a classification loss term generated by the real label supervision, a knowledge distillation loss term generated by the soft label supervision, and a boundary point detection loss term.

7. The method for automatic segmentation of static cone penetration test soil layers integrating knowledge guidance and data-driven approaches according to claim 1, characterized in that, In step S3, the multi-task learning framework includes 1 to 4 layers of Bi-LSTM, an attention layer, and a normalization layer connected in sequence; the output layer of the multi-task learning framework includes a boundary probability output branch using the Sigmoid activation function and a soil category probability output branch using the Softmax activation function.

8. The method for automatic segmentation of static cone penetration test soil layers integrating knowledge guidance and data-driven approaches according to claim 1, characterized in that, In step S5, the fusion process uses soft voting and / or probability-weighted averaging to fuse the soil layer category probabilities and boundary point probabilities of the same depth point in multiple overlapping prediction windows; and determines the final boundary point position according to a preset probability threshold.

9. The method for automatic segmentation of static cone penetration test soil layers integrating knowledge guidance and data-driven approaches according to claim 1, characterized in that, In step S1, the static penetration data includes at least the cone tip resistance and sidewall friction directly measured by the dual-bridge static penetration probe, and the friction ratio calculated based on the cone tip resistance and sidewall friction.

10. The method for automatic segmentation of static cone penetration test soil layers integrating knowledge guidance and data-driven approaches according to claim 1, characterized in that, In step S1, the preprocessing further includes: performing unit unification processing on the static penetration data, and processing outliers and missing values.

Images