Sand liquefaction discrimination method and system based on machine learning

Abstract

This invention provides a machine learning-based method and system for identifying sand liquefaction, belonging to the field of machine learning technology. The invention simultaneously acquires and cleans in-situ and ground motion parameters, combines them with historical liquefaction annotations to classify samples, and then constructs a class-balanced dataset by fusing oversampling and undersampling. Multidimensional mechanical and response features are extracted, an optimized gradient boosting tree model is trained, and this model is used to assess the probability of site liquefaction, achieving high-precision intelligent identification of sand liquefaction.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for determining sand liquefaction based on machine learning. Background Technology

[0002] Currently, the engineering community commonly uses empirical discrimination diagrams, modified standard penetration tests, standard cone penetration tests, or shear wave velocities as single indicators to assess the liquefaction potential of sand. These methods are based on post-earthquake measured data compiled between 1970 and 1990, and provide a "liquefaction / non-liquefaction" dichotomy through thresholds or empirical curves, and are still used in most design codes.

[0003] With the development of the Internet of Things and geological cloud platforms, in-situ test data and seismic records are growing exponentially. Meanwhile, machine learning (ML) can uncover nonlinear coupling relationships in multidimensional, large-sample data, improving prediction reliability. Domestic and international research has begun exploring the benefits of algorithms such as random forests, gradient boosting trees, and deep neural networks for liquefaction detection; some scholars have also utilized transfer learning and incremental learning to achieve cross-site applicability.

[0004] Current technical shortcomings include: sand liquefaction is affected by multiple factors including static, dynamic, and seepage, making it difficult to accurately express pore pressure development and stress history using a single index; there are far fewer measured cases of seismic liquefaction than non-liquefaction cases, leading to "minority class" distortion in machine learning models during the training phase, and the discrimination results are prone to underreporting liquefaction; existing machine learning research mostly focuses on simple feature inputs and lacks a systematic method to integrate shear wave velocity frequency domain information with dynamic indicators such as pore pressure dynamic increment. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a machine learning-based method and system for identifying sand liquefaction, which realizes intelligent identification of sand liquefaction by combining multi-source parameter fusion, sample balance optimization and high-precision model training, effectively overcoming the defects of traditional methods such as single feature, sample bias and insufficient identification accuracy.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A machine learning-based method for determining sand liquefaction includes:

[0008] In-situ and ground motion parameters are acquired synchronously at the sample site, and the in-situ and ground motion parameters are time-aligned, outlier removed, and dimensionless to obtain a cleaned dataset.

[0009] The cleaned dataset is divided into liquefied and non-liquefied categories based on preset historical liquefaction labels;

[0010] Perform synthetic minority oversampling on the minority class samples in the liquefaction class and the non-liquefaction class, and combine it with undersampling remixing to generate a class-balanced dataset;

[0011] On the class-balanced dataset, the cyclic stress ratio, effective stress path slope, pore water pressure growth gradient, shear wave velocity multi-scale wavelet energy, and cone-wave velocity coupling index are calculated to obtain the sample feature vector set.

[0012] The set of sample feature vectors and the corresponding liquefaction labels are input into the gradient boosting tree classification algorithm. Cross-validation is used to optimize the hyperparameters and train the algorithm to obtain the liquefaction discrimination model.

[0013] The real-time feature vector set corresponding to the site to be evaluated is input into the liquefaction discrimination model to obtain the liquefaction probability, and the sand liquefaction of the site to be evaluated is judged based on the liquefaction probability.

[0014] Preferably, the in-situ and ground motion parameters include: cone penetration resistance, side friction resistance, standard penetration blow count, shear wave velocity, groundwater level, instantaneous values ​​and continuous variation sequences of pore water pressure, peak ground acceleration and effective duration corresponding to the earthquake to be evaluated.

[0015] Preferably, the cleaned dataset is divided into liquefied and non-liquefied categories based on preset historical liquefaction labels, including:

[0016] An index is generated for each record in the cleaned dataset; the index includes at least the site identifier, borehole number, and depth range.

[0017] The system calls up a preset historical liquefaction labeling table; the historical liquefaction labeling table records the corresponding liquefaction observation results according to the site identifier, borehole number, and depth range;

[0018] For each record in the cleaned dataset, a matching liquefaction observation is retrieved from the historical liquefaction label table based on the corresponding index, and the matching liquefaction observation is converted into a binary label; in the binary label, liquefaction corresponds to the label value 1, and non-liquefaction corresponds to the label value 0;

[0019] According to the binary labels, records with a label value of 1 are aggregated to form a subset of the liquefaction class, and records with a label value of 0 are aggregated to form a subset of the non-liquefaction class.

[0020] Preferably, a synthetic minority oversampling process is performed on the minority class samples in the liquefaction class and the non-liquefaction class, combined with undersampling remixing, to generate a class-balanced dataset, including:

[0021] Statistically calculate the subset D of the liquefaction class. L With the subset D of the non-liquefiable classN The number of samples, denoted as n. L With n N ;

[0022] The subset with the smaller sample size is identified as the minority class D. min The subset with the larger sample size is determined as the majority class D. maj ;

[0023] For D min Perform synthetic minority oversampling based on k-nearest neighbor interpolation, including:

[0024] a1) Regarding D min For each sample, calculate the k nearest neighbor samples in the feature space;

[0025] b1) Generate interpolation points between each sample and its nearest neighbor to construct a synthetic sample;

[0026] c1) Repeat steps a1) and b1) until the number of samples in the constructed synthetic sample reaches the target value n. tar ;

[0027] For the majority class D maj Perform undersampling remixing to obtain the processed majority class set D′. maj ;

[0028] The processed majority class set D′ maj With D′ min The samples are merged and their order is shuffled to form the class-balanced dataset.

[0029] Preferably, for most classes D maj Perform undersampling remixing to obtain the processed majority class set D′. maj ,include:

[0030] a2) Using the Tomek-link criterion, identify the match with D. min The majority class samples that constitute the boundary pairs;

[0031] b2) Remove majority class samples that constitute boundary pairs to refine class boundaries;

[0032] c2) If the number of samples in the majority class after deletion is still greater than n tar Then, randomly draw samples from the remaining samples according to a uniform distribution until the sample size equals n. tar The processed majority class set D′ is obtained. maj ;

[0033] The constructed synthetic sample is compared with the original minority class D. min Merge into a new balanced minority class set D′ min .

[0034] Preferably, the target value n tar The formula for calculating n is: tar =α·max(n L ,n N ); where α is the balance coefficient, and its value ranges from 0.8 to 1.0.

[0035] Preferably, on the class-balanced dataset, the cyclic stress ratio, effective stress path slope, pore water pressure growth gradient, shear wave velocity multi-scale wavelet energy, and cone-wave velocity coupling index are calculated to obtain a set of sample feature vectors, including:

[0036] For cone penetration resistance q c Side friction resistance f s Standard penetration test blow count (N), shear wave velocity (V) s Groundwater level depth z w Instantaneous value of pore water pressure u0, pore water pressure time series u(t), peak surface acceleration a max Effective duration t d Perform unit unification and dimension normalization;

[0037] According to formula u h =γ w z w Calculate the static pressure u at the measuring point h And according to the static water pressure u h The pore pressure increment sequence Δu(t) is obtained; where Δu(t) = u(t) - u h ;γ w The unit weight of water;

[0038] Using formula Calculate the cyclic stress ratio; where g is the gravitational acceleration constant;

[0039] During the effective duration t of the earthquake d Inside, with Δu(t) as the ordinate, Construct an effective stress path for the x-axis and use the least squares method to obtain the slope S of the effective stress path. EP ;in,

[0040] In t d The time derivative of Δu(t) with respect to the interval is calculated, and the maximum positive value of the time derivative is taken as the characteristic gradient G of the pore water pressure growth. u ;in,

[0041] Shear wave velocity depth curve V s (z) Perform three-level discrete wavelet decomposition and calculate the reconstructed signal energy E at each scale.j The multi-scale wavelet energy (E1, E2, E3) of the shear wave velocity is obtained; where E j =∑ z (V s,j (z)) 2 j = 1, 2, 3; z represents the depth position, and j represents the scale number of the wavelet decomposition;

[0042] According to the formula Calculate the cone-wave velocity coupling index I cv ;

[0043] The cyclic stress ratio (CSR) and the effective stress path slope (S) are used to calculate the cyclic stress ratio (CSR) and the effective stress path slope (S). EP pore water pressure growth gradient G u Multi-scale energy (E1, E2, E3), cone-wave velocity coupling index I cv With standard penetration test blow count N, groundwater level depth z w Peak ground acceleration a max Effective duration t d The sample feature vectors are combined in a fixed order to form a complete set of sample feature vectors.

[0044] A machine learning-based sand liquefaction discrimination system includes:

[0045] The in-situ parameter acquisition and cleaning unit is used to simultaneously acquire in-situ and ground motion parameters at the sample site, and to perform time alignment, outlier removal and dimensionless conversion on the in-situ and ground motion parameters to obtain a cleaned dataset.

[0046] The historical annotation and partitioning unit is used to divide the cleaned dataset into liquefied and non-liquefied categories based on preset historical liquefaction annotations.

[0047] The class balance processing unit is used to perform synthetic minority oversampling and undersampling remixing on the minority class samples in the liquefaction class and the non-liquefaction class to generate a class balanced dataset.

[0048] A multi-parameter feature construction unit is used to calculate the cyclic stress ratio, effective stress path slope, pore water pressure growth gradient, shear wave velocity multi-scale wavelet energy, and cone-wave velocity coupling index on the class-balanced dataset to obtain a set of sample feature vectors.

[0049] The model training and optimization unit is used to input the set of sample feature vectors and the corresponding liquefaction labels into the gradient boosting tree classification algorithm, optimize the hyperparameters using cross-validation and train to obtain the liquefaction discrimination model.

[0050] The site liquefaction discrimination unit is used to input the set of real-time feature vectors corresponding to the site to be evaluated into the liquefaction discrimination model to obtain the liquefaction probability, and to make sand liquefaction discrimination of the site to be evaluated based on the liquefaction probability.

[0051] The present invention discloses the following technical effects:

[0052] This invention addresses the technical shortcomings of existing liquefaction discrimination methods, such as reliance on single parameters, imbalanced class samples, insufficient feature dimensions, and poor model generalization ability. By introducing joint acquisition and normalization of in-situ parameters and seismic motion parameters, a multi-source coupled liquefaction characteristic index (including cyclic stress ratio, stress path slope, pore pressure growth gradient, wavelet energy characteristics, and penetration-wave velocity coupling index) is constructed. Based on this, a sample equalization mechanism combining synthetic oversampling and undersampling remixing significantly alleviates the bias problem caused by insufficient liquefaction samples in model training. Furthermore, through cross-validation-driven hyperparameter optimization and gradient boosting tree algorithm training, the constructed liquefaction discrimination model exhibits stronger accuracy and stability, enabling precise quantification of sand liquefaction probability under different site conditions. This improves the model's adaptability and engineering practicality, overcoming the limitations of traditional empirical map-based methods and static machine learning models in dynamically integrating multi-source parameters and responding to complex site conditions. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0056] 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.

[0057] The purpose of this invention is to provide a method and system for judging sand liquefaction based on machine learning. It realizes intelligent judgment of sand liquefaction by combining multi-source parameter fusion, sample balance optimization and high-precision model training, and effectively overcomes the defects of traditional methods such as single feature, sample bias and insufficient judgment accuracy.

[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a machine learning-based method for determining sand liquefaction, comprising:

[0060] Step 100: Simultaneously acquire in-situ and ground motion parameters at the sample site, and perform time alignment, outlier removal, and dimensionless conversion on the in-situ and ground motion parameters to obtain the cleaned dataset;

[0061] Step 200: Divide the cleaned dataset into liquefied and non-liquefied classes based on the preset historical liquefaction labels;

[0062] Step 300: Perform synthetic minority oversampling on the minority class samples in both the liquefied and non-liquefied classes and combine it with undersampling remixing to generate a class-balanced dataset;

[0063] Step 400: Calculate the cyclic stress ratio, effective stress path slope, pore water pressure growth gradient, shear wave velocity multi-scale wavelet energy, and cone-wave velocity coupling index on the class-balanced dataset to obtain the sample feature vector set.

[0064] Step 500: Input the sample feature vector set and the corresponding liquefaction label into the gradient boosting tree classification algorithm, optimize the hyperparameters using cross-validation and train to obtain the liquefaction discrimination model;

[0065] Step 600: Input the set of real-time feature vectors corresponding to the site to be evaluated into the liquefaction discrimination model to obtain the liquefaction probability, and make the sand liquefaction discrimination of the site to be evaluated based on the liquefaction probability.

[0066] Preferably, the in-situ and ground motion parameters include: cone penetration resistance, side friction resistance, standard penetration blow count, shear wave velocity, groundwater level, instantaneous values ​​and continuous variation sequences of pore water pressure, peak ground acceleration and effective duration corresponding to the earthquake to be evaluated.

[0067] In the specific implementation process, the key to step 100 lies in the synchronous acquisition and structured integration of in-situ parameters and seismic motion parameters to construct high-quality data inputs that can be used for model discrimination. To achieve synchronization and physical coupling consistency, this invention uniformly schedules in-situ parameters such as cone penetration resistance, side friction resistance, standard penetration blow count, shear wave velocity, and groundwater level within the sampling time window, and ensures that these parameters are in the same time reference frame as the instantaneous values ​​and continuous variation sequences of pore water pressure by deploying data acquisition nodes with a clock unification mechanism. In addition, by connecting to the regional seismic monitoring platform or the seismometer equipment installed on the site, the peak surface acceleration and effective duration matching the seismic event to be evaluated are extracted and precisely matched to the same sampling section and depth number, realizing the construction of integrated temporal-spatial-layer structural data. This synchronous integration process solves the problems of disconnection between in-situ parameters and seismic parameters, inconsistent data fields, and inability to support dynamic feature extraction in the prior art, and is a prerequisite for ensuring the modeling of dynamic processes in liquefaction characterization.

[0068] After data acquisition, all raw parameters were immediately cleaned. Key steps included: First, frame-level time alignment of the multi-source data sequences was performed based on a unified timestamp to prevent time drift in pore pressure response, shear velocity profiles, and ground motion indicators. Second, for abrupt changes and abnormal drift segments in the pore pressure data, a sliding window statistical residual method was used to remove local erroneous values ​​while retaining liquefaction precursor features such as slow pore pressure increases to avoid accidentally deleting key response signals. Third, all parameters involved in subsequent modeling were dimensionless, standardized to the [0,1] interval using the min-max normalization method, and logarithmic domain transformation was performed on shear velocity and penetration parameters to accommodate the nonlinear feature mapping requirements of the model. This data standardization not only improved the comparability between samples but also provided numerical stability for parameter fusion and gradient iteration in subsequent feature construction.

[0069] Preferably, the cleaned dataset is divided into liquefied and non-liquefied categories based on preset historical liquefaction labels, including:

[0070] An index is generated for each record in the cleaned dataset; the index includes at least the site identifier, borehole number, and depth range.

[0071] The system calls up a preset historical liquefaction labeling table; the historical liquefaction labeling table records the corresponding liquefaction observation results according to the site identifier, borehole number, and depth range;

[0072] For each record in the cleaned dataset, a matching liquefaction observation is retrieved from the historical liquefaction label table based on the corresponding index, and the matching liquefaction observation is converted into a binary label; in the binary label, liquefaction corresponds to the label value 1, and non-liquefaction corresponds to the label value 0;

[0073] According to the binary labels, records with a label value of 1 are aggregated to form a subset of the liquefaction class, and records with a label value of 0 are aggregated to form a subset of the non-liquefaction class.

[0074] In the specific implementation of this invention, the key to step 200 lies in establishing a one-to-one correspondence between the cleaned data and historical liquefaction observations, achieving accurate classification between liquefied and non-liquefied categories. To this end, this invention first generates a structured index for each cleaned data record. This index includes at least a site identifier, borehole number, and depth range, used to uniquely identify the spatial origin of each sample. Then, a preset historical liquefaction labeling table is invoked. This labeling table is a tag library formed based on post-earthquake surveys, in-situ tests, or remote sensing identification, which records the actual liquefaction occurrence at each measuring point according to the same site identifier, borehole number, and depth range fields. Through the alignment of the above index fields, the system can quickly locate the corresponding liquefaction observation results of the cleaned sample in the historical labeling table, ensuring the accuracy of the classification and the traceability of the project.

[0075] After retrieving the labeled results, this invention uniformly transforms the matched liquefaction observations into a standard binary classification label system, where liquefaction is assigned a value of 1 and non-liquefaction is assigned a value of 0. All labeled data records are clustered according to their label values: samples with a label of 1 are assigned to the liquefaction subset for subsequent feature construction and model training to learn liquefaction patterns; samples with a label of 0 are assigned to the non-liquefaction subset for modeling the non-liquefaction discrimination boundary. This partitioning process does not rely on manual intervention and is entirely based on automatic matching of structured indexes and historical labels, ensuring the objectivity and standardization of the label transfer process, and providing the necessary categorical foundation for subsequent sample resampling and feature difference analysis.

[0076] Preferably, a synthetic minority oversampling process is performed on the minority class samples in the liquefaction class and the non-liquefaction class, combined with undersampling remixing, to generate a class-balanced dataset, including:

[0077] Statistically calculate the subset D of the liquefaction class. L With the subset D of the non-liquefiable class N The number of samples, denoted as n. L With n N ;

[0078] The subset with the smaller sample size is identified as the minority class D. min The subset with the larger sample size is determined as the majority class D. maj ;

[0079] For D min Perform synthetic minority oversampling based on k-nearest neighbor interpolation, including:

[0080] a1) Regarding Dmin For each sample, calculate the k nearest neighbor samples in the feature space;

[0081] b1) Generate interpolation points between each sample and its nearest neighbor to construct a synthetic sample;

[0082] c1) Repeat steps a1) and b1) until the number of samples in the constructed synthetic sample reaches the target value n. tar ;

[0083] For the majority class D maj Perform undersampling remixing to obtain the processed majority class set D′. maj ;

[0084] The processed majority class set D′ maj With D′ min The samples are merged and their order is shuffled to form the class-balanced dataset.

[0085] In a specific implementation of this invention, to address the problem of significant differences in the number of liquefied and non-liquefied samples leading to class bias during model training, a joint reconstruction strategy of oversampling of minority class samples and undersampling of majority class samples is preferably introduced to generate a class-balanced dataset. First, the number of samples in the liquefied and non-liquefied subsets obtained in the previous step is counted, and the class with fewer samples is identified as the "minority class," and the class with more samples is identified as the "majority class." Then, for the minority class subset, a synthetic sample generation process based on nearest neighbor interpolation in the feature space is executed: for each original sample in the minority class, several nearest neighbor samples in the feature space are calculated; then, new sample points are generated by interpolation between the sample and its neighboring samples, simulating the potential but unobserved sample distribution. This process is repeated until the number of synthetic samples reaches the set target number, thereby expanding the distribution range of minority class samples and enhancing their contribution to model training.

[0086] After the minority class is expanded and synthesized, an undersampling remixing operation is performed on the majority class subset. Specifically, samples near inter-class boundaries in the majority class are preferentially removed to reduce interference from ambiguous regions. If the number of majority class samples is still significantly higher than the target value, the size of the majority class is further reduced proportionally through random downsampling. Finally, the processed majority class sample set is merged with the expanded minority class sample set, and the overall sample order is shuffled to form a new class-balanced dataset. This data reconstruction mechanism not only achieves class balance in terms of quantity but also maintains the continuity and representativeness of feature distribution through spatial interpolation.

[0087] Preferably, for most classes D maj Perform undersampling remixing to obtain the processed majority class set D′. maj ,include:

[0088] a2) Using the Tomek-link criterion, identify the match with D. min The majority class samples that constitute the boundary pairs;

[0089] b2) Remove majority class samples that constitute boundary pairs to refine class boundaries;

[0090] c2) If the number of samples in the majority class after deletion is still greater than n tar Then, randomly draw samples from the remaining samples according to a uniform distribution until the sample size equals n. tar The processed majority class set D′ is obtained. maj ;

[0091] The constructed synthetic sample is compared with the original minority class D. min Merge into a new balanced minority class set D′ min .

[0092] Preferably, the target value n tar The formula for calculating n is: tar =α·max(n L ,n N ); where α is the balance coefficient, and its value ranges from 0.8 to 1.0.

[0093] In a preferred embodiment of the present invention, to further improve the quality of the discrimination boundary in the sample space, an undersampling remixing mechanism based on a boundary criterion is introduced for the majority class samples. Specifically, a neighborhood distance criterion is first used to analyze the sample distribution, identifying those samples in the majority class that form boundary pairs with minority class samples. These samples belong to different classes from their nearest neighbors and are the closest to each other. These majority class boundary samples are prone to causing classification confusion, so they are preferentially deleted after identification, thereby refining the classification boundary and reducing the interference of the discrimination ambiguity band on model training. This step improves the stability and robustness of the class boundaries learned by the model by structurally eliminating unstable samples.

[0094] After deleting boundary majority class samples, if the number of majority class samples is still significantly higher than the target value, further uniform random sampling is performed from the remaining samples until the sample size matches the target value, forming a processed majority class set. Simultaneously, the synthetic minority class samples obtained through oversampling are merged with the original minority class sample set to generate a balanced minority class set. The target sample size is calculated based on the maximum class sample size and a set of preset balance coefficients. These balance coefficients control the relative proportion of samples between classes, typically chosen between 80% and 100%, preserving the representativeness of the data distribution structure while avoiding learning bias caused by class imbalance. These are core parameters for constructing a high-confidence training sample set. This joint strategy effectively solves the problem of majority class over-dominance in model training.

[0095] Preferably, on the class-balanced dataset, the cyclic stress ratio, effective stress path slope, pore water pressure growth gradient, shear wave velocity multi-scale wavelet energy, and cone-wave velocity coupling index are calculated to obtain a set of sample feature vectors, including:

[0096] For cone penetration resistance q c Side friction resistance f s Standard penetration test blow count (N), shear wave velocity (V) s Groundwater level depth z w Instantaneous value of pore water pressure u0, pore water pressure time series u(t), peak surface acceleration a max Effective duration t d Perform unit unification and dimension normalization;

[0097] According to formula u h =γ w z w Calculate the static pressure u at the measuring point h And according to the static water pressure u h The pore pressure increment sequence Δu(t) is obtained; where Δu(t) = u(t) - u h ;γ w The unit weight of water;

[0098] Using formula Calculate the cyclic stress ratio; where g is the gravitational acceleration constant;

[0099] During the effective duration t of the earthquake d Inside, with Δu(t) as the ordinate, Construct an effective stress path for the x-axis and use the least squares method to obtain the slope S of the effective stress path. EP ;in,

[0100] In t dThe time derivative of Δu(t) with respect to the interval is calculated, and the maximum positive value of the time derivative is taken as the characteristic gradient G of the pore water pressure growth. u ;in,

[0101] Shear wave velocity depth curve V s (z) Perform three-level discrete wavelet decomposition and calculate the reconstructed signal energy E at each scale. j The multi-scale wavelet energy (W1, W2, E3) of the shear wave velocity is obtained; where E j =∑ z (V s,j (z)) 2 j = 1, 2, 3; z represents the depth position, and j represents the scale number of the wavelet decomposition;

[0102] According to the formula Calculate the cone-wave velocity coupling index I cv ;

[0103] The cyclic stress ratio (CSR) and the effective stress path slope (S) are used to calculate the cyclic stress ratio (CSR) and the effective stress path slope (S). EP pore water pressure growth gradient G u Multi-scale energy (E1, E2, E3), cone-wave velocity coupling index I cv With standard penetration test blow count N, groundwater level depth z w Peak ground acceleration a max Effective duration t d The sample feature vectors are combined in a fixed order to form a complete set of sample feature vectors.

[0104] In a preferred embodiment of the present invention, to ensure the physical consistency and numerical comparability of the feature vector construction, the various original parameters in the class-balanced dataset are first standardized and dimensionless. Specifically, this includes converting parameters such as cone penetration resistance, side friction resistance, standard penetration blow count, shear wave velocity, groundwater level depth, pore water pressure and their time series, peak ground acceleration, and effective duration into international standard units, and mapping them to a fixed numerical range using the max-min normalization method. Based on this, the static pore pressure at each measuring point is calculated using the groundwater level depth and the unit weight of water, and the static pore pressure is subtracted from the pore water pressure time series to obtain a pore pressure increment series reflecting the seismic disturbance effect, which serves as the basis for subsequent dynamic index extraction.

[0105] Subsequently, based on normalization and incremental calculations, several key engineering seismic response characteristics were extracted. Among them, the cyclic stress ratio was calculated based on the relative relationship between cone penetration resistance, side friction resistance, and peak ground acceleration, characterizing the proportional strength of seismic shear loading relative to the soil's shear capacity; the effective stress path slope was constructed over the effective duration of the earthquake, using pore pressure increment as the ordinate and the shear strength ratio based on cone penetration resistance and side friction resistance as the abscissa, establishing a stress path diagram and fitting its slope using the least squares method to reflect the stress evolution trend; the pore water pressure growth gradient was obtained by differentiating the incremental sequence over the effective time interval and extracting the maximum positive growth rate, capturing the key dynamic characteristic of rapid pore pressure accumulation before liquefaction.

[0106] Furthermore, to incorporate stratigraphic scale characteristics and nonlinear response information, this invention performs a three-level discrete wavelet decomposition on the depth distribution curve of shear wave velocity, extracting the reconstructed signal energy at each scale as features to form a multi-scale wavelet energy set reflecting changes in the stiffness of deep structures. In addition, a coupling index is constructed using cone penetration resistance and shear wave velocity to evaluate the synergistic effect between the relative compaction of the soil layer and wave propagation capacity, supplementing the description of the comprehensive response capability that is difficult to characterize individually. Finally, the above five main features are combined with Standard Penetration Blows, groundwater depth, peak ground acceleration, and effective duration into a complete feature vector, ensuring that each dimension of the feature possesses both physical interpretability and classification ability, providing a stable and reliable data foundation for subsequent model training.

[0107] In the specific implementation of this invention, the core of step 500 lies in constructing a liquefaction discrimination model with high discrimination accuracy and generalization ability, thereby completing the mapping from feature vectors to liquefaction probabilities. To this end, the set of sample feature vectors constructed in the preceding steps and their corresponding liquefaction labels are used as supervised learning inputs and fed into the gradient boosting tree classification algorithm for training. This process relies not only on the physical representation ability of the feature space but also on the adaptive optimization of the model structure and parameter configuration. To avoid underfitting or overfitting, this invention introduces a cross-validation mechanism to jointly tune multiple key hyperparameters, including the learning rate, the number of trees, the maximum depth of the trees, and the subsample ratio. All parameter combinations are evaluated through analysis, and the optimal performance result is used as the final parameter configuration.

[0108] Specifically, the training process employs at least five-fold cross-validation, dividing the sample data into multiple subsets, which are used alternately as training and validation data to systematically evaluate the stability and accuracy of the model under different parameter combinations. This invention preferably uses the area under the receiver operating characteristic (ROC) curve as the performance metric, balancing classification accuracy and sample distribution balance. Finally, the parameter set with the best average performance is selected, and the final liquefaction discrimination model is retrained on all data based on this parameter set. Its structure and model weights are then permanently stored for subsequent real-time liquefaction probability inference. The above training and optimization process provides reliable support for the model's practical application and is an indispensable technical step, a key guarantee for achieving efficient and reliable liquefaction discrimination.

[0109] In the implementation of this invention, step 600 aims to achieve intelligent identification of the liquefaction risk of sandy soil at the site to be assessed based on the trained liquefaction discrimination model. First, for the site to be assessed, the aforementioned feature extraction process is repeated to collect in-situ and seismic parameters in real time, including cone penetration resistance, side friction resistance, standard penetration blow count, shear wave velocity, groundwater level, pore water pressure, peak ground acceleration, and effective duration. Time alignment, normalization, and multi-feature calculation are then performed to form a consistent real-time feature vector set. This feature vector set maintains consistency with the training phase in parameter type, extraction method, and numerical distribution, ensuring the timeliness and adaptability of the model input.

[0110] After inputting the aforementioned real-time feature vectors into the trained liquefaction discrimination model, the model outputs the corresponding liquefaction probability value to quantify the likelihood of soil liquefaction under the current working conditions. This invention compares this liquefaction probability with a preset risk threshold. When the liquefaction probability is higher than or equal to the threshold, a liquefaction determination result is output; otherwise, it is determined as non-liquefaction. This discrimination process relies entirely on the feature-label mapping relationship learned within the model for reasoning, offering advantages such as fast response speed, no human intervention required, and online deployment. It is suitable for various application scenarios such as pre-earthquake assessment, rapid epicenter early warning, and post-earthquake verification, significantly improving the accuracy of liquefaction discrimination and the efficiency of engineering applications.

[0111] Corresponding to the above methods, such as Figure 2 As shown, this embodiment also provides a sand liquefaction discrimination system based on machine learning, including:

[0112] The in-situ parameter acquisition and cleaning unit is used to simultaneously acquire in-situ and ground motion parameters at the sample site, and to perform time alignment, outlier removal and dimensionless conversion on the in-situ and ground motion parameters to obtain a cleaned dataset.

[0113] The historical annotation and partitioning unit is used to divide the cleaned dataset into liquefied and non-liquefied categories based on preset historical liquefaction annotations.

[0114] The class balance processing unit is used to perform synthetic minority oversampling and undersampling remixing on the minority class samples in the liquefaction class and the non-liquefaction class to generate a class balanced dataset.

[0115] A multi-parameter feature construction unit is used to calculate the cyclic stress ratio, effective stress path slope, pore water pressure growth gradient, shear wave velocity multi-scale wavelet energy, and cone-wave velocity coupling index on the class-balanced dataset to obtain a set of sample feature vectors.

[0116] The model training and optimization unit is used to input the set of sample feature vectors and the corresponding liquefaction labels into the gradient boosting tree classification algorithm, optimize the hyperparameters using cross-validation and train to obtain the liquefaction discrimination model.

[0117] The site liquefaction discrimination unit is used to input the set of real-time feature vectors corresponding to the site to be evaluated into the liquefaction discrimination model to obtain the liquefaction probability, and to make sand liquefaction discrimination of the site to be evaluated based on the liquefaction probability.

[0118] The beneficial effects of this invention are as follows:

[0119] (1) This invention introduces multi-source measured parameters such as cone penetration resistance, side friction resistance, shear wave velocity, groundwater level, pore water pressure and ground motion to construct a structured and spatiotemporally unified input dataset, which breaks through the problem of single parameters and local response of existing liquefaction discrimination methods, and realizes a more comprehensive and representative modeling basis for the liquefaction potential of sandy soil.

[0120] (2) The present invention combines a sample balancing mechanism of minority class oversampling and majority class undersampling and remixing, which solves the model training bias problem caused by the imbalance of positive and negative sample ratios in liquefaction discrimination, significantly improves the model’s ability to identify minority class liquefaction samples, and provides a guarantee for the accurate discrimination of high-risk areas of earthquake liquefaction.

[0121] (3) This invention designs a variety of engineering mechanics and structural response features, including cyclic stress ratio, effective stress path slope, pore pressure growth gradient, wavelet energy characteristics and cone-wave velocity coupling index, and constructs a feature system with physical interpretability and distinguishability, which effectively enhances the clarity of the model's discrimination boundary and the stability of its generalization.

[0122] (4) This invention adopts a gradient boosting tree model based on cross-validation optimization and supports the rapid input of the real-time feature vector of the site to be evaluated into the model for liquefaction probability calculation and risk assessment. It realizes an efficient, accurate and automated liquefaction identification process and is applicable to various engineering scenarios such as earthquake pre-assessment and disaster emergency response.

[0123] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0124] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for discriminating sand liquefaction based on machine learning, characterized by, include: In-situ and ground motion parameters are acquired synchronously at the sample site, and the in-situ and ground motion parameters are time-aligned, outlier removed, and dimensionless to obtain a cleaned dataset. The cleaned dataset is divided into liquefied and non-liquefied categories based on preset historical liquefaction labels; Perform synthetic minority oversampling on the minority class samples in the liquefaction class and the non-liquefaction class, and combine it with undersampling remixing to generate a class-balanced dataset; On the class-balanced dataset, the cyclic stress ratio, effective stress path slope, pore water pressure growth gradient, shear wave velocity multi-scale wavelet energy, and cone-wave velocity coupling index are calculated to obtain the sample feature vector set. The set of sample feature vectors and the corresponding liquefaction labels are input into the gradient boosting tree classification algorithm. Cross-validation is used to optimize the hyperparameters and train the algorithm to obtain the liquefaction discrimination model. The real-time feature vector set corresponding to the site to be evaluated is input into the liquefaction discrimination model to obtain the liquefaction probability, and the sand liquefaction of the site to be evaluated is judged based on the liquefaction probability. The in-situ and ground motion parameters include: cone penetration resistance, side friction resistance, standard penetration blow count, shear wave velocity, groundwater level, instantaneous values ​​and continuous variation sequences of pore water pressure, peak ground acceleration and effective duration corresponding to the earthquake to be evaluated. On the class-balanced dataset, the cyclic stress ratio, effective stress path slope, pore water pressure growth gradient, shear wave velocity multi-scale wavelet energy, and cone-wave velocity coupling index are calculated to obtain a set of sample feature vectors, including: For cone penetration resistance Side friction resistance Standard penetration test number Shear wave velocity Groundwater level depth Instantaneous value of pore water pressure Pore ​​water pressure time series Peak ground acceleration Effective duration Perform unit unification and dimension normalization; According to the formula Calculate the static pore pressure at the measuring point And according to the static water pressure Obtain the pore pressure increment sequence ;in, ; The unit weight of water; Using formula Calculate the cyclic stress ratio; where, It is the gravitational acceleration constant; During the effective duration of the earthquake Inside, with For the ordinate, Construct an effective stress path for the x-axis and obtain the slope of the effective stress path using the least squares method. ;in, ; exist interval pairs Calculate the time derivative and take the maximum positive value of the time derivative as the characteristic gradient of pore water pressure growth. ;in, ; Shear wave velocity depth curve Perform three-level discrete wavelet decomposition and calculate the reconstructed signal energy at each scale. The multi-scale wavelet energy of shear wave velocity was obtained. ;in, ; Indicates depth position. Indicates the scale number of the wavelet decomposition; According to the formula Calculate the cone-wave velocity coupling index ; Cyclic stress ratio (CSR) and effective stress path slope pore water pressure growth gradient Multiscale energy Cone-wave velocity coupling index With Standard Penetration Strokes Groundwater level depth Peak ground acceleration Effective duration The sample feature vectors are combined in a fixed order to form a complete set of sample feature vectors; When performing time alignment, outlier removal, and dimensionless processing on the in-situ and ground motion parameters, cone penetration resistance, side friction resistance, standard penetration blow count, shear wave velocity, and groundwater level are uniformly scheduled within the sampling time window, and are placed in the same time reference system as the instantaneous values ​​and continuous variation sequences of pore water pressure. Peak surface acceleration and effective duration matching the earthquake event to be evaluated are extracted and paired to the same sampling section and depth number. Frame-level time alignment is performed on the multi-source data sequence based on a unified timestamp. The sliding window statistical residual judgment method is used to remove abrupt changes and abnormal drift segments in the pore pressure data, while retaining the liquefaction precursor characteristic segments of slowly rising pore pressure. Logarithmic domain transformation is performed on the shear wave velocity and penetration parameters.

2. The machine learning-based method for determining sand liquefaction according to claim 1, characterized in that, The cleaned dataset is divided into liquefied and non-liquefied classes based on preset historical liquefaction labels, including: An index is generated for each record in the cleaned dataset; the index includes at least the site identifier, borehole number, and depth range. The system calls up a preset historical liquefaction labeling table; the historical liquefaction labeling table records the corresponding liquefaction observation results according to the site identifier, borehole number, and depth range; For each record in the cleaned dataset, a matching liquefaction observation is retrieved from the historical liquefaction label table based on the corresponding index, and the matching liquefaction observation is converted into a binary label; in the binary label, liquefaction corresponds to the label value 1, and non-liquefaction corresponds to the label value 0; According to the binary labels, records with a label value of 1 are aggregated to form a subset of the liquefaction class, and records with a label value of 0 are aggregated to form a subset of the non-liquefaction class.

3. The machine learning-based method for determining sand liquefaction according to claim 1, characterized in that, Perform synthetic minority oversampling on the minority class samples in the liquefaction class and the non-liquefaction class, combined with undersampling remixing, to generate a class-balanced dataset, including: Statistical subset of the liquefaction class A subset of the non-liquefiable class The sample size is denoted as . and ; The subset with the smaller sample size is identified as the minority class. The subset with the larger sample size is identified as the majority class. ; right Execution based on Synthetic minority oversampling using nearest neighbor interpolation includes: a1) For each sample, calculate in the feature space The nearest neighbor samples; b1) Generate interpolation points between each sample and its nearest neighbor to construct a synthetic sample; c1) Repeat steps a1) and b1) until the number of samples in the constructed synthetic sample reaches the target value. ; For the majority class Perform undersampling remixing to obtain the processed majority class set. ; The processed majority class set and The samples are merged and their order is shuffled to form the class-balanced dataset.

4. The machine learning-based method for determining sand liquefaction according to claim 3, characterized in that, For the majority class Perform undersampling remixing to obtain the processed majority class set. ,include: a2) Using the Tomek-link criterion, identify the... The majority class samples that constitute the boundary pairs; b2) Remove majority class samples that constitute boundary pairs to refine class boundaries; c2) If the number of samples in the majority class after deletion is still greater than Then, randomly draw samples from the remaining samples according to a uniform distribution until the sample size equals... The processed majority class set is obtained. ; The constructed synthetic sample is compared with the original minority class. Merge into a new balanced minority class set .

5. The machine learning-based method for determining sand liquefaction according to claim 3, characterized in that, target value The calculation formula is: ;in, This is the balance coefficient, and its value range is... .

6. A machine learning-based sand liquefaction discrimination system, characterized in that, The system for implementing the method as described in any one of claims 1 to 5 comprises: The in-situ parameter acquisition and cleaning unit is used to simultaneously acquire in-situ and ground motion parameters at the sample site, and to perform time alignment, outlier removal and dimensionless conversion on the in-situ and ground motion parameters to obtain a cleaned dataset. The historical annotation and partitioning unit is used to divide the cleaned dataset into liquefied and non-liquefied categories based on preset historical liquefaction annotations. The class balance processing unit is used to perform synthetic minority oversampling and undersampling remixing on the minority class samples in the liquefaction class and the non-liquefaction class to generate a class balanced dataset. A multi-parameter feature construction unit is used to calculate the cyclic stress ratio, effective stress path slope, pore water pressure growth gradient, shear wave velocity multi-scale wavelet energy, and cone-wave velocity coupling index on the class-balanced dataset to obtain a set of sample feature vectors. The model training and optimization unit is used to input the set of sample feature vectors and the corresponding liquefaction labels into the gradient boosting tree classification algorithm, optimize the hyperparameters using cross-validation and train to obtain the liquefaction discrimination model. The site liquefaction discrimination unit is used to input the set of real-time feature vectors corresponding to the site to be evaluated into the liquefaction discrimination model to obtain the liquefaction probability, and to make sand liquefaction discrimination of the site to be evaluated based on the liquefaction probability.

Images