A method and system for identifying seismic liquefaction of sandy soil based on SSA-CNN-SVM model coupling

Abstract

This invention discloses a method and system for identifying seismic liquefaction of sandy soil based on a coupled SSA-CNN-SVM model. This invention relates to the field of seismic safety assessment technology and includes the following steps: acquiring a historical sandy soil liquefaction sample dataset, including key influencing indicators and liquefaction state labels; simultaneously collecting geographical location information; extracting statistical feature parameters to calculate the liquefaction resistance index and seismic intensity index; performing weighted similarity measurement based on geographical feature parameters; dividing the sample area into multiple geological regions using a hierarchical clustering algorithm; within each region, calculating the Mahalanobis distance between sample areas; performing secondary clustering using a density-based clustering algorithm to identify clusters of samples with similar liquefaction mechanisms; and establishing a prediction model coupled with a deep learning model and optimization algorithm for each cluster to achieve liquefaction hazard assessment, significantly improving the accuracy and robustness of sandy soil seismic liquefaction identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of anti-seismic safety evaluation, in particular to a sand earthquake liquefaction discrimination method and system based on SSA-CNN-SVM model coupling. BACKGROUND

[0002] With the acceleration of urbanization and the increasing infrastructure construction, soil liquefaction phenomenon poses an increasing threat to buildings and infrastructure in natural disasters such as earthquakes, especially in sandy areas; sandy soil liquefaction refers to the phenomenon that saturated sandy soil loses its bearing capacity due to pressure changes under the action of vibration such as earthquakes, resulting in the soil becoming liquid. This phenomenon not only poses a great challenge to engineering construction and public safety, but also affects the efficiency of post-earthquake rescue and reconstruction work. Therefore, accurate discrimination and evaluation of sandy soil liquefaction is particularly important.

[0003] Currently, existing technologies mainly rely on traditional liquefaction discrimination methods such as empirical formulas, standard penetration tests, and soil tests. These methods analyze soil physical properties and historical seismic data to assess the soil's anti-liquefaction capacity. However, these methods often lack comprehensive consideration of multi-dimensional features, such as failing to effectively combine geographic information, anti-liquefaction strength, and vibration intensity, making it difficult to fully reflect the complex factors of soil liquefaction, resulting in insufficient accuracy in liquefaction discrimination under complex geological conditions.

[0004] The above information disclosed in the background section is only used to enhance the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide a sandy soil earthquake liquefaction discrimination method based on SSA-CNN-SVM model coupling to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions: A sandy soil earthquake liquefaction discrimination method based on SSA-CNN-SVM model coupling, the specific steps comprising: Obtain the geographic feature parameters, statistical feature parameters, key influence indicators and liquefaction state labels of each sample area, the sample area is a liquefaction area or a non-liquefaction area, determine the anti-liquefaction strength index representing the anti-liquefaction capacity of the sample area based on the key influence indicators and statistical feature parameters, and the vibration intensity index representing the seismic action intensity of the sample area; Taking the geographical characteristic parameters as the partition basis, a weighted similarity measure between each sample region in pairs on the geographical characteristic parameters is calculated, and a hierarchical clustering algorithm is used for clustering analysis on all sample regions to divide all sample regions into several geological regions; For any geological region, the Mahalanobis distance between each sample region in the geological region is calculated based on the anti-liquefaction strength index and the vibration intensity index, and a density-based spatial clustering algorithm is used for secondary clustering of the sample regions in the geological region to divide the sample regions in the geological region into several sample clusters with similar liquefaction mechanisms. For any sample cluster, an SSA-CNN-SVM coupled model is independently constructed to establish the mapping relationship between the key influence indicators and the liquefaction state labels of each sample region in the sample cluster, so as to evaluate the liquefaction risk of the region to be analyzed through the SSA-CNN-SVM coupled model.

[0007] Further, the key influence indicators include the current groundwater depth, standard penetration blow count, average particle size, non-uniformity coefficient, overburden effective pressure, dynamic shear stress ratio and historical seismic intensity of the sample region. The liquefaction state label includes a first liquefaction state, a second liquefaction state, a third liquefaction state and a non-liquefaction state, and the severity of soil liquefaction increases with the increase of the state label level. The key influence indicator data collected is preprocessed for missing values, specifically including numerical filling by K-nearest neighbor method.

[0008] Further, the geographical characteristic parameters include geological age coding, sedimentary environment coding and topographic and geomorphic type coding of the sample region. The statistical characteristic parameters include the current clay content and historical earthquake magnitude of the sample region. The calculation of the anti-liquefaction strength index is based on the statistical characteristic parameters and the key influence indicators, and is calculated according to the basic theory of soil mechanics, and the specific formula is: In the formula, is the anti-liquefaction strength index of the i-th sample region, is the current standard penetration blow count of the i-th sample region, is the current dynamic shear stress ratio of the i-th sample region, is the current clay content of the i-th sample region, and i is the index of the sample region. The method for calculating the vibration intensity index is based on the statistical characteristic parameters and the key influence indicators, and is calculated based on the basic theory of earthquake engineering, and the specific formula is: In the formula, is the vibration intensity index of the i-th sample area, is the average value of historical earthquake magnitudes in the i-th sample area, is the average value of historical earthquake intensities in the i-th sample area, is the current groundwater depth in the i-th sample area.

[0009] Furthermore, the specific logic for calculating the weighted similarity measure of geographical feature parameters between all sample areas is as follows: The entropy weight method is used to determine the weights of each geographical feature parameter. The specific logic is: Determine the proportion of the i-th sample area under the k-th geographical feature parameter, determine the entropy value of the k-th geographical feature parameter based on this proportion, calculate the weight of the k-th geographical feature parameter based on the entropy value of the k-th geographical feature parameter. The formula for calculating the proportion of the i-th sample area under the k-th geographical feature parameter is: In the formula, represents the proportion of the i-th sample area under the k-th geographical feature parameter, is the parameter value of the i-th sample area on the k-th geographical feature, is the total number of sample areas, and k is the index of the type of geographical feature parameter; The formula for determining the entropy value of the k-th geographical feature parameter is: In the formula, is the entropy value of the k-th geographical feature parameter; The formula for calculating the weight of the k-th geographical feature parameter is: In the formula, is the weight of the k-th geographical feature parameter; Based on the weights of geographical feature parameters, a weighted Euclidean distance of geographical feature parameters is constructed as the weighted similarity measure representing the geological similarity between sample areas. The formula for specifically calculating the weighted Euclidean distance is: In the formula, is the weighted Euclidean distance between the i-th sample area and the j-th sample area, is the parameter value of the i-th sample area on the k-th geographical feature, is the parameter value of the j-th sample area on the k-th geographical feature, j is the index of the sample area, and .

[0010] Furthermore, the specific steps of using the agglomerative hierarchical clustering algorithm to perform cluster analysis on all sample regions include: treating each sample region as a separate cluster, calculating the weighted Euclidean distance between all clusters, merging the two clusters with the closest weighted Euclidean distance each time, and recalculating the weighted Euclidean distance between all clusters, repeating the merging operation until all samples are merged into one cluster to form a clustering tree, and cutting the clustering tree according to multiple preset similarity thresholds so that the clustering tree is cut into several cut clusters under each similarity threshold; Calculate the average contour coefficient of each cut cluster under each similarity threshold, select the similarity threshold corresponding to the maximum average contour coefficient as the optimal cut threshold, and take the cut clusters obtained by cutting the cluster tree with the optimal cut threshold as the precise cut clusters. Each precise cut cluster is a geological region, so as to divide all sample areas into several geological regions.

[0011] Furthermore, the logic underlying the calculation of Mahalanobis distances between sample regions within each geological region is as follows: The liquefaction resistance index and seismic intensity index of sample areas within any geological region are normalized to obtain standard values ​​for the liquefaction resistance index and seismic intensity index. Based on these standard values, the Mahalanobis distance between sample areas within the geological region is calculated using the following formula: In the formula, Let be the Mahalanobis distance between the p-th and q-th sample regions within the geological region. This is the inverse matrix of the geological region's covariance matrix, which includes the covariance matrices of the liquefaction resistance index and the seismic intensity index of all sample areas within the geological region. This is the standardized feature vector of the p-th sample region within the geological region, specifically a vector including the standard values ​​of the liquefaction resistance index and the vibration intensity index. Let p be the standardized feature vector of the q-th sample region within the geological region, and p and q be the indices of the sample regions within the geological region. ; The specific logic for secondary clustering is as follows: For any geological region, calculate the Mahalanobis distance between all sample regions within it, and perform clustering based on the Mahalanobis distance using a density-based spatial clustering algorithm. This includes: using the k-distance graph method to determine the neighborhood radius of the density-based spatial clustering algorithm, calculating the Mahalanobis distance from each sample region to its preset number of nearest neighbors, sorting and plotting the obtained nearest neighbor Mahalanobis distances, selecting the distance corresponding to the inflection point of the curve as the neighborhood radius, and setting the minimum number of samples. All sample regions are marked as unvisited. An unvisited sample region is randomly selected, and the number of unvisited sample regions within its neighborhood radius is checked to see if it is not less than the set minimum number of samples. If not, no new cluster is created. If so, a new cluster is created with the unvisited sample region as the core point, and all unvisited sample regions within the neighborhood radius of the core point are added to the newly created cluster. The sample regions in the new cluster are updated to visited. Another unvisited sample region is randomly selected, and the above operation is repeated until no more new clusters are built. Each new cluster is regarded as a group of sample clusters with similar liquefaction mechanisms.

[0012] Furthermore, the SSA-CNN-SVM coupled model specifically refers to the Sparrow Search algorithm-convolutional neural network-support vector machine coupled model, which uses the Sparrow Search algorithm to automatically and globally optimize the key hyperparameters of the convolutional neural network to obtain the optimal model initialization configuration. The key hyperparameters include the learning rate, batch size, and L2 regularization coefficient.

[0013] Furthermore, based on the optimal model initialization configuration, key hyperparameters of the convolutional neural network are set. This convolutional neural network is used for feature extraction. Key influence indicators of sample regions within the same sample cluster are used as training data, and these key influence indicators are input into the configured convolutional neural network. The convolutional neural network then extracts deep feature vectors of the input key influence indicators. The convolutional neural network comprises, in sequence, an input layer, a convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, a Dropout layer, and a fully connected layer; The logic behind extracting deep feature vectors is as follows: based on the standard values ​​of the anti-liquefaction intensity index and vibration intensity index of the cluster centers of each sample cluster, the weight response coefficients of the convolution kernel in the convolutional layer to the key influencing indicators are adaptively adjusted. For any sample cluster, a latent liquefaction feature map is output through the convolutional layer of the corresponding convolutional neural network. Based on the latent liquefaction feature map, a deep feature vector is output through a batch normalization layer, an activation function layer, a max pooling layer, a Dropout layer, and a fully connected layer. The latent liquefaction feature map is specifically represented as follows: In the formula, This represents the potential liquefaction feature map generated by the m-th convolutional kernel. Nonlinear gating to represent liquefaction risk For the u-th sample cluster, the weighted response coefficient of the m-th convolutional kernel to the 0-th key influencing indicator is given. This represents the 0th key impact indicator. This represents the bias term of the m-th convolutional kernel; m is the index of the convolutional kernel within the convolutional layer, o is the index of the key influencing metric, and u is the index of the sample cluster. The specific formula used to calculate the weighted response coefficient of the m-th convolutional kernel to the 0-th key influencing indicator for the u-th sample cluster is as follows: In the formula, These are the standard weighted response coefficients of the convolution kernel; Let u be the standard value of the cluster center vibration intensity index of the u-th sample cluster. The standard value of the liquefaction resistance index of the cluster center of the u-th sample cluster; The deep feature vector is used as the input to the support vector machine model, and the corresponding liquefaction state label is used as the output result for training. The key influencing indicators of the sandy soil to be identified are input into the trained prediction model, and the model outputs the corresponding liquefaction state label. The logic behind using the SSA-CNN-SVM coupled model to assess the liquefaction risk of the region under analysis is as follows: determine the geological region to which the region under analysis belongs, calculate the Mahalanobis distance between the region under analysis and the sample region, determine the sample cluster to which the region under analysis belongs, select the corresponding SSA-CNN-SVM coupled model, and use the key influencing indicators of the region under analysis as input to output the liquefaction status label of the region under analysis.

[0014] This invention also provides a sand seismic liquefaction discrimination system based on SSA-CNN-SVM model coupling, used to execute the above-mentioned sand seismic liquefaction discrimination method based on SSA-CNN-SVM model coupling, including: The information acquisition module is used to acquire the geographical feature parameters, statistical feature parameters, key impact indicators and liquefaction status labels of each sample area. The sample area is either a liquefied area or a non-liquefied area. Based on the key impact indicators and statistical feature parameters, the liquefaction resistance index, which characterizes the liquefaction resistance of the sample area, and the seismic intensity index, which characterizes the seismic intensity at the sample area, are determined. The preliminary regional division module is used to calculate the weighted similarity measure between each sample region on the geographical feature parameters based on the geographical feature parameters, and to perform cluster analysis on all sample regions using the agglomerative hierarchical clustering algorithm to divide all sample regions into several geological regions. The liquefaction mechanism analysis module is used to calculate the Mahalanobis distance between sample areas within any geological region based on the liquefaction resistance index and vibration intensity index, and to perform secondary clustering of the sample areas within the geological region using a density-based spatial clustering algorithm, so as to divide the sample areas within the geological region into several sample clusters with similar liquefaction mechanisms. The risk assessment module is used to independently construct an SSA-CNN-SVM coupled model for any sample cluster, in order to establish the mapping relationship between key impact indicators and liquefaction status labels of each sample area within the sample cluster, thereby assessing the liquefaction risk of the area to be analyzed through the SSA-CNN-SVM coupled model.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This scheme effectively addresses the problem of insufficient model generalization ability caused by neglecting regional geological differences in traditional liquefaction discrimination methods by incorporating geographical location information and employing a two-level clustering partitioning strategy. Specifically, firstly, geographical feature parameters are used as the first-level partitioning basis, and an agglomerative hierarchical clustering algorithm is used to divide the overall sample area into several geological regions, ensuring that sample areas within the same region have similar geological backgrounds. Then, for each region, the liquefaction resistance index and seismic intensity index, which reflect liquefaction resistance and seismic intensity, are used as the second-level partitioning basis, and a density-based spatial clustering algorithm is used to identify sample clusters with similar liquefaction mechanisms. This layer-by-layer refined partitioning method ensures that the sample areas within each cluster are highly consistent in their causal mechanisms, laying a foundation of data homogeneity for the subsequent establishment of targeted prediction models and avoiding model bias caused by mixed training of sample areas across regions and mechanisms. Furthermore, this scheme considers the liquefaction resistance index and the seismic intensity index during the feature construction stage. These two indices quantitatively characterize the physical mechanism of liquefaction from two dimensions: the soil's own liquefaction resistance and the intensity of external seismic loads, respectively. Combined with two-level clustering partitioning, the distribution of sample areas within each cluster is more concentrated on these two indices. Building upon this foundation, this scheme independently constructs a prediction model for each sample cluster, coupled with a deep learning model (CNN) and optimization algorithm (SSA, i.e., Sparrow Search algorithm) and support vector machine (SVM). Since sample areas within the same cluster share similar geological conditions and liquefaction mechanisms, model training can fully exploit local feature patterns, significantly reducing the interference of sample area heterogeneity on model performance. Simultaneously, by using SSA to optimize the hyperparameters of the CNN-SVM, the network structure and kernel function parameters can be adaptively adjusted, further improving the model's fitting accuracy and generalization ability on local data. Compared to existing single global models, this scheme achieves refined evaluation through partitioned modeling and a tailored approach for each region, significantly improving the accuracy and robustness of sandy soil seismic liquefaction detection. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 A graph showing the comparison between model recall and precision; Figure 3 A comparison chart of model training accuracy; Figure 4 This is a schematic diagram of the overall system structure of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0019] Example: Please see Figures 1-3 The present invention provides a technical solution: A method for identifying seismic liquefaction of sandy soil based on a coupled SSA-CNN-SVM model, comprising the following steps: Step 1: Obtain the geographical feature parameters, statistical feature parameters, key impact indicators and liquefaction status labels of each sample area. The sample area is either a liquefied area or a non-liquefied area. Based on the key impact indicators and statistical feature parameters, determine the liquefaction resistance index, which characterizes the liquefaction resistance of the sample area, and the seismic intensity index, which characterizes the seismic intensity of the sample area.

[0020] The key influencing indicators include the current groundwater depth, standard penetration test blow count, average grain size, uniformity coefficient, overlying effective pressure, dynamic shear stress ratio, and historical seismic intensity of the sample area. The specific methods for obtaining the key impact indicators include: The current groundwater depth is obtained by monitoring the groundwater level through on-site drilling or observation wells, or by obtaining hydrological data in geological reports to understand the changes in the groundwater depth in the region. The current earthquake intensity is obtained by accessing the local earthquake monitoring agency and acquiring historical earthquake event records within a preset historical interval, including epicenter location, magnitude, and intensity. The intensity values ​​for different regions are then calculated using the earthquake intensity correction method and an earthquake intensity table. The preset historical interval is the time range of 5-10 years prior to the current moment. The current standard penetration test blow count is based on the completed geological exploration report and existing standard penetration test data. The current average particle size is obtained by conducting sieving tests or laser particle size analysis on soil samples to obtain particle size distribution data, and then calculating the average particle size. By referring to relevant geological or soil research literature, typical particle size information of soil in a specific area can be obtained. The current non-uniformity coefficient is obtained from the particle size distribution curve through particle size analysis. The specific formula used to calculate the non-uniformity coefficient is as follows: In the formula, , where D60 and D10 are the particle diameters of 60% and 10% of the particle size in the particle size distribution curve, respectively; The current effective overburden pressure is calculated by drilling to obtain soil profiles and unit weight data of each layer, or by using geological models and combining known soil layer thickness and density to estimate the effective overburden pressure.

[0021] The liquefaction status labels include Level 1 liquefaction, Level 2 liquefaction, Level 3 liquefaction, and no liquefaction. The severity of liquefaction in sandy soil increases sequentially with the increase of the status label level. For the collected key impact indicator data, missing value preprocessing is performed, specifically including numerical imputation using the K-nearest neighbor method. The K-nearest neighbor method is a conventional existing technology, and the specific steps are not described in detail here.

[0022] Feature parameters for subsequent partitioning are extracted and constructed from the data of each sample area. These feature parameters include geographic feature parameters and statistical feature parameters. Based on the statistical feature parameters, the liquefaction resistance and seismic intensity of the corresponding sample area are analyzed, and the liquefaction resistance index and seismic intensity index are calculated.

[0023] The geographic feature parameters include the geological age code, sedimentary environment code, and topographic type code of the sample area; The specific method for obtaining geological age codes is as follows: refer to the geological maps of the country or region and relevant geological research literature to understand the geological age division of the study area, extract the age information of different soil layers from the regional geological reports obtained by the local geological survey department, and convert them into a coding format.

[0024] The specific method for obtaining the sedimentary environment code is as follows: Based on geological and sedimentological literature, analyze the sedimentary environment of the study area, which can usually be divided into rivers, lakes, oceans, etc. During the drilling process, observe the sediment types and bedding characteristics to determine the sedimentary environment and code it. Topographic and geomorphic type coding uses satellite remote sensing imagery and topographic maps to identify different topographic and geomorphic features (such as mountains, plains, hills, etc.), uses GIS software to analyze regional topographic data, obtains corresponding geomorphic classifications, and then encodes them.

[0025] The statistical characteristic parameters include the current clay content and historical earthquake magnitude of the sample area; earthquake magnitude information of a specific area is extracted using an established earthquake database; clay content data known in geological or soil survey reports is referenced, or laboratory analysis is conducted after collecting soil samples on site, and the clay content (particle size less than 0.002 mm) is tested by sieving or sedimentation methods.

[0026] The logic behind calculating the liquefaction resistance index is as follows: based on statistical characteristic parameters and key influencing indicators, the liquefaction resistance index is comprehensively calculated according to the basic theory of soil mechanics. The specific formula used is as follows: In the formula, Let i be the liquefaction resistance index of the i-th sample region. Let be the current standard penetration test (SPT) hit count for the i-th sample region. The dynamic shear stress ratio of the i-th sample region is given. Let i be the current clay content of the i-th sample region, where i is the index of the sample region. It should be noted that the liquefaction resistance index This value reflects the resistance of the sample area to liquefaction; the larger the value, the stronger the resistance to liquefaction. Current Standard Penetration Stroke Count The standard penetration test (SPT) number reflects the density and strength of the soil layer; the higher the SPT blow count, the denser the soil, the higher the soil strength, and the stronger its resistance to liquefaction. It is directly proportional to the liquefaction resistance index; Current dynamic shear stress ratio The effective vertical stress reflects the influence of soil gravity and overlying pressure on liquefaction resistance. A higher effective vertical stress results in higher external shear stress on the soil and weaker liquefaction resistance. Therefore, the denominator reflects the influence of effective vertical stress on the soil. A higher dynamic shear stress ratio indicates a lower ability of the soil to withstand external loads and a weakened liquefaction resistance. A constant is then introduced. and This is to appropriately scale and normalize the effect of effective stress, so as to avoid its effect being too large or too small; Current clay content This represents the percentage of clay particles by mass in the soil. A higher clay content indicates stronger soil cohesion and a lower likelihood of liquefaction. Therefore, it is introduced... As a correction factor, it reflects the influence of clay particles in the soil on their resistance to liquefaction. The higher the value, the smaller the correction factor, and the lower the liquefaction resistance index. The larger the value, the stronger the soil's resistance to liquefaction.

[0027] The specific method for calculating the seismic intensity index is as follows: based on statistical characteristic parameters, combined with key influencing indicators, and based on the basic theory of earthquake engineering, the seismic intensity index is calculated using the following formula: middle, Let be the vibration intensity index of the i-th sample region. Let be the average historical earthquake magnitude of the i-th sample region. Let be the historical average earthquake intensity of the i-th sample area. Let be the current groundwater depth in the i-th sample area.

[0028] It should be noted that the first Vibration intensity index of each sample area The index is used to quantitatively characterize the magnitude of the liquefaction driving force that a site may bear under seismic action. The larger the index, the stronger the triggering ability of the earthquake on sand liquefaction and the greater the impact on the liquefaction potential of the soil. Historical earthquake magnitude average The magnitude is used to describe the amount of energy released by an earthquake. The higher the magnitude, the greater the energy of the earthquake and the stronger the shaking effect on the soil. Historical average earthquake intensity The intensity of an earthquake's impact on the Earth's surface and engineering structures is usually related to peak ground acceleration and response spectrum characteristics; the higher the intensity, the more violent the ground movement. Reflecting the magnitude of the original driving force of an earthquake, magnitude and intensity together determine the shear strength of the ground motion on the soil layer. The multiplication of the two is an empirical combination that can more comprehensively take into account both the energy magnitude (magnitude) and the actual degree of damage (intensity) of the earthquake compared to using either one alone. No. The current groundwater level depth in a sample area is the vertical distance from the surface to the groundwater level. Shallower depths make saturated sand layers more prone to liquefaction, directly affecting soil saturation and vibration attenuation. Greater depths cause vibration intensity to gradually decrease as it propagates through the soil, reducing the likelihood of liquefaction; the effect of groundwater level on liquefaction is not linear. At very shallow depths (0-2 meters), a slight increase in depth rapidly reduces the likelihood of liquefaction. At greater depths, further increases in depth gradually weaken the inhibitory effect on liquefaction. (Exponential function) It can well describe nonlinear relationships, coefficients It is an empirical constant obtained by fitting a large amount of historical liquefaction data.

[0029] Step 2: Based on the geographic feature parameters, calculate the weighted similarity measure between each pair of sample areas on the geographic feature parameters, and use the agglomerative hierarchical clustering algorithm to perform cluster analysis on all sample areas to divide all sample areas into several geological regions.

[0030] The logic behind calculating the weighted similarity measure of all sample regions across geographic feature parameters is as follows: The entropy weight method is used to determine the weight of each geographic feature parameter. Specifically, the logic is: determine the proportion of the i-th sample region under the k-th geographic feature parameter; determine the entropy value of the k-th geographic feature parameter based on this proportion; calculate the weight of the k-th geographic feature parameter based on its entropy value; and the formula for calculating the proportion of the i-th sample region under the k-th geographic feature parameter is as follows: In the formula, This represents the proportion of the i-th sample region under the k-th geographical feature parameter. Let be the parameter value of the i-th sample region for the k-th geographical feature. is the total number of sample areas, and k is the index of the types of geographic feature parameters; It should be noted that this indicates the first... The sample region in the first The proportion of a sample region to the total number of samples under a given geographic feature parameter reflects the relative importance of that feature. This is achieved by considering the parameter values ​​for each sample region. Divide by the sum of parameter values ​​from all sample regions to obtain the first... The sample region in the first The relative proportion of various geographical characteristic parameters; The formula used to determine the entropy value of the kth geographic feature parameter is as follows: In the formula, Let be the entropy value of the k-th geographical feature parameter; It should be noted that entropy is used to measure the uncertainty or diversity of information. The larger the entropy value, the more uniform the distribution of the geographic feature parameter, and the greater the amount of information. The entropy formula contains... The term is used to measure the contribution of each sample region's weight, and the overall uncertainty of this feature is calculated in a weighted manner. Used to standardize the entropy value, making its value within the interval This facilitates the comparison of different characteristics.

[0031] The formula used to calculate the weight of the kth geographic feature parameter is: In the formula, The weight of the k-th geographical feature parameter; It should be noted that the smaller the entropy value, the more concentrated the distribution of the feature and the relatively less information it contains. Therefore, this feature should be given a higher weight in the weighted similarity measure. The larger the entropy value, the richer the information it contains and the relatively lower its weight.

[0032] A weighted Euclidean distance is constructed based on the weights of geographical feature parameters, serving as a weighted similarity measure characterizing the geological similarity between sample regions. The specific formula for calculating the weighted Euclidean distance is as follows: In the formula, The weighted Euclidean distance between the i-th sample region and the j-th sample region is... Let be the parameter value of the i-th sample region for the k-th geographical feature. The parameter value of the j-th sample region for the k-th geographic feature, where j is the index of the sample region, and .

[0033] The specific steps of clustering all sample regions using the agglomerative hierarchical clustering algorithm include: treating each sample region as a separate cluster, calculating the weighted Euclidean distance between all clusters, merging the two clusters with the closest weighted Euclidean distance each time, and recalculating the weighted Euclidean distance between all clusters, repeating the merging operation until all samples are merged into one cluster to form a clustering tree, and cutting the clustering tree according to multiple preset similarity thresholds so that the clustering tree is cut into several cut clusters under each similarity threshold; Calculate the average contour coefficient of each cut cluster under each similarity threshold, select the similarity threshold corresponding to the maximum average contour coefficient as the optimal cut threshold, and take the cut clusters obtained by cutting the cluster tree with the optimal cut threshold as the precise cut clusters. Each precise cut cluster is a geological region, so as to divide all sample areas into several geological regions. The specific steps include: calculating the average contour coefficient of the current cut cluster when selecting each similarity threshold; observing the changing trend of the average contour coefficient; selecting the similarity threshold that maximizes the average contour coefficient as the optimal cut threshold; cutting the clustering tree based on the optimal cut threshold; and the corresponding cut cluster is the best clustering result.

[0034] Step 3: For any geological region, based on the liquefaction resistance index and vibration intensity index, calculate the Mahalanobis distance between sample regions within the geological region, and use a density-based spatial clustering algorithm to perform secondary clustering of the sample regions within the geological region, so as to divide the sample regions within the geological region into several sample clusters with similar liquefaction mechanisms.

[0035] The logic underlying the calculation of Mahalanobis distances between sample regions within each geological region is as follows: The liquefaction resistance index and seismic intensity index of sample areas within any geological region are normalized to obtain standard values ​​for the liquefaction resistance index and seismic intensity index. Based on these standard values, the Mahalanobis distance between sample areas within the geological region is calculated using the following formula: In the formula, Let be the Mahalanobis distance between the p-th and q-th sample regions within the geological region. This is the inverse matrix of the geological region's covariance matrix, which includes the covariance matrices of the liquefaction resistance index and the seismic intensity index of all sample areas within the geological region. This is the standardized feature vector of the p-th sample region within the geological region, specifically a vector including the standard values ​​of the liquefaction resistance index and the vibration intensity index. Let p be the standardized feature vector of the q-th sample region within the geological region, and p and q be the indices of the sample regions within the geological region. ; It should be noted that this formula is used to calculate the Mahalanobis distance between sample areas within a geological region to measure the similarity between sample areas. Mahalanobis distance is a measure that considers the correlation between variables and the distribution of data. It is used to analyze the similarity of the liquefaction resistance index and the vibration intensity index of different sample areas. The smaller the Mahalanobis distance, the more similar the characteristics between the samples. in It is a covariance matrix calculated based on the liquefaction resistance index and vibration intensity index of all sample areas within the geological region, and then processed by the inverse of the covariance matrix. Mahalanobis distance can effectively adjust the weights of different features; The specific logic for secondary clustering is as follows: For any geological region, calculate the Mahalanobis distance between all sample regions. Based on the Mahalanobis distance, perform clustering using a density-based spatial clustering algorithm, including: using the k-distance graph method to determine the neighborhood radius of the density-based spatial clustering algorithm; calculating the Mahalanobis distance from each sample region to its preset number of nearest neighbors; sorting and plotting the obtained nearest neighbor Mahalanobis distances; selecting the distance corresponding to the inflection point of the curve as the neighborhood radius; and setting the minimum number of samples. Specific steps include: selecting an appropriate... The value is usually determined based on the size and features of the dataset, typically chosen as 5% to 10% of the total number of samples. For each sample region, the value is calculated to the nearest neighbor. The Mahalanobis distances of each sample region are recorded to form a distance list. For each sample region, the nth region is selected from the distance list. The smallest value is used as the k-distance of the sample region. The k-distances of all sample regions are collected into an array, and the data are sorted according to the k-distance values. A k-distance plot is then drawn, with the horizontal axis representing the index of the sample region and the vertical axis representing the corresponding k-distance value. In the k-distance plot, the changes in the curve are observed, and the inflection points are identified. The k-distance value corresponding to the inflection point is determined as the neighborhood radius. Based on the neighborhood radius, a spatial clustering algorithm is used for clustering. All sample regions are marked as unvisited. An unvisited sample region is randomly selected, and the number of unvisited sample regions within its neighborhood radius is checked to see if it is not less than the set minimum number of samples. If not, no new cluster is created. If so, a new cluster is created with the unvisited sample region as the core point, and all unvisited sample regions within the neighborhood radius of the core point are added to the newly created cluster. The sample regions in the new cluster are updated to visited. Another unvisited sample region is randomly selected, and the above operation is repeated until no more new clusters are built. Each new cluster is regarded as a group of sample clusters with similar liquefaction mechanisms.

[0036] Step 4: For any sample cluster, independently construct an SSA-CNN-SVM coupled model to establish the mapping relationship between key influencing indicators and liquefaction status labels of each sample area within the sample cluster. In this way, the SSA-CNN-SVM coupled model is used to assess the liquefaction risk of the area to be analyzed.

[0037] The SSA-CNN-SVM coupled model specifically refers to the Sparrow Search algorithm-convolutional neural network-support vector machine coupled model, which uses the Sparrow Search algorithm to automatically and globally optimize the key hyperparameters of the convolutional neural network, including the learning rate, batch size, and L2 regularization coefficient, to obtain the optimal model initialization configuration.

[0038] Based on the optimal model initialization configuration, the key hyperparameters of the convolutional neural network are set. The convolutional neural network is used for feature extraction. The key influence indicators of the sample regions in the same sample cluster are used as training data. The key influence indicators are used as inputs to the configured convolutional neural network. The deep feature vectors of the input key influence indicators are extracted by the convolutional neural network. The deep feature vectors are used as inputs to the support vector machine model. The corresponding liquefaction state label is used as the output result for training. The key influence indicators of the sand to be identified are input into the trained prediction model, and the model outputs the corresponding liquefaction state label.

[0039] The specific steps for training the prediction model include: normalizing the key influencing indicators to eliminate dimensional differences, and dividing them into a training set for model training and a test set for performance evaluation according to a preset ratio; the ratio of the training set to the test set is between 7:3 and 9:1, preferably 8:2; the training set is used for model training and hyperparameter optimization, and the test set is used to independently verify the generalization performance of the model. The division process uses random sampling to ensure the consistency of the distribution characteristics of the two sets of data. The key hyperparameters of the convolutional neural network feature extraction module are automatically and globally optimized using the sparrow search algorithm. The key hyperparameters include at least the learning rate, batch size, and L2 regularization coefficient, in order to obtain the optimal model initialization configuration. The parameters of the SSA algorithm are as follows: the number of sparrows is 30-80, preferably 50; the maximum number of iterations is 15-30, preferably 20; and the hyperparameter optimization range is specifically the learning rate [1e-4, 1e-2], batch size [16, 128], and L2 regularization coefficient [1e-6, 1e-3].

[0040] The parameter values ​​of the convolutional neural network feature extraction module are as follows: the number of convolutional kernels ranges from 8 to 32, preferably 16; the kernel size ranges from 1×1 to 3×3, preferably 1×1; the activation function can be ReLU or LeakyReLU, preferably ReLU; the dropout rate of the dropout layer ranges from 0.3 to 0.7, preferably 0.5; the number of neurons in the fully connected layer ranges from 64 to 256, preferably 128; the optimizer can be SGDM or Adam, preferably SGDM; the initial learning rate ranges from 1e-4 to 5e-4, preferably 0.0005; the L2 regularization parameter ranges from 5e-4 to 2e-3, preferably 0.001; and based on the initial hyperparameters, the SGDM optimizer is used during training to continuously optimize the hyperparameters.

[0041] The parameter values ​​of the support vector machine classification module are as follows: the kernel function can be either the radial basis function (RBF) or the polynomial kernel function, with the radial basis function being preferred; the penalty coefficient c ranges from 3 to 10, with 5 being preferred; and the kernel function parameter g ranges from 0.05 to 0.3, with 0.1 being preferred.

[0042] The method utilizes a sparrow search algorithm to automatically optimize the key hyperparameters of the convolutional neural network feature extraction module. Specific steps include: setting the number of sparrows to 50 and the maximum number of iterations to 20; optimizing the learning rate within the range of [1e-4, 1e-2], the batch size within the range of [16, 128], and the L2 regularization coefficient within the range of [1e-6, 1e-3]. Wherein, the state vector of the r-th sparrow individual A set of core hyperparameters for liquefaction prediction to be optimized is specifically represented as follows: in, Let represent the hyperparameter combination of the convolutional neural network liquefaction prediction model to be optimized, represented by the r-th sparrow. The specific vector is [learning rate, batch size, regularization coefficient]. For learning rate, For batch processing size, is the L2 regularization coefficient; r is the index of the individual sparrow; To prevent the algorithm from searching for invalid parameter regions or causing the model to diverge, a three-dimensional solution space constraint boundary set is established based on the distribution characteristics of sand liquefaction data, specifically represented as follows: in, Represents the set of constraint boundaries in the three-dimensional solution space. This represents the physical lower bound of hyperparameters to prevent underfitting of the model. The physical upper bound of the hyperparameters is defined to prevent the model from diverging or overfitting. The lower and upper bounds of the hyperparameters are set for the learning rate, batch size, and regularization coefficient, respectively, as defined below: For the learning rate constraint interval: For batch processing size constraint range: The physical lower and upper bounds for the regularization coefficients are defined similarly; It should be noted that sand liquefaction data exhibits a high degree of nonlinearity; if If the gradient descent step size of the model is too large, it is prone to oscillation or even divergence near the optimal solution (i.e., missing the optimal liquefaction discrimination critical point); if The convergence speed is too slow and it is easy to get stuck in local saddle points; this interval ensures that the model can accurately capture the gradient changes of highly sensitive indicators such as earthquake intensity. set up This ensures sufficient sample diversity within the batch (e.g., including both liquefied and non-liquefied samples) to guarantee accurate gradient update direction; [Settings are missing from the original text.] This utilizes the effect of random noise in small batches, which helps the model escape local extreme points caused by complex geological formations.

[0043] Given that geotechnical investigation data often contains noise (such as human error in standard penetration test blow counts), we introduce... It can suppress the model from overfitting to noise while retaining the weights of key geological features, preventing the model from losing its ability to generalize to new sites due to "rote memorization" of training data.

[0044] The Sparrow Search algorithm searches for the optimal hyperparameter combination based on model performance evaluated during the training process. Specifically, it uses K-fold cross-validation to evaluate model performance, where K=5. An average cross-validation error function is constructed to determine the optimal hyperparameter combination. The specific expression of the average cross-validation error function is as follows: In the formula, Let be the average cross-validation error function of the hyperparameter combination. This is the validation set data for the Kth round of validation. The prediction model trained excluding the validation set data, For the K rounds of validation, the key impact metrics within the validation set are... This represents the predicted liquefaction state label output by the model when the input to the prediction model is a key impact indicator within the validation set of K rounds of validation. These are the actual liquefaction state labels corresponding to the key impact indicators within the validation set during the K-round validation process. For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The model predicts the local test error in round K; K is the validation round index. This formula represents the average expected error rate of the model in predicting unknown geological conditions across 5 different training-validation cycles; only when The combination of hyperparameters that reaches the global minimum This combination of hyperparameters was considered the optimal hyperparameter combination for soil liquefaction in the region. The convolutional neural network comprises, in sequence, an input layer, a convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, a Dropout layer, and a fully connected layer; Input layer: Receives an input tensor of shape [7,1,1], corresponding to 7 indices.

[0045] Each sample is represented as a three-dimensional tensor of shape [7,1,1]; a one-dimensional convolutional kernel slides over a feature sequence of length 7 to capture local correlation patterns between different evaluation metrics.

[0046] Convolutional layer: 16 1×1 convolutional kernels are used for feature extraction, followed by a batch normalization layer and activated by the ReLU function; The logic behind extracting deep feature vectors is as follows: based on the standard values ​​of the anti-liquefaction intensity index and vibration intensity index of the cluster centers of each sample cluster, the weight response coefficients of the convolution kernel in the convolutional layer to the key influencing indicators are adaptively adjusted. For any sample cluster, a latent liquefaction feature map is output through the convolutional layer of the corresponding convolutional neural network. Based on the latent liquefaction feature map, a deep feature vector is output through a batch normalization layer, an activation function layer, a max pooling layer, a Dropout layer, and a fully connected layer. The latent liquefaction feature map is specifically represented as follows: In the formula, This represents the potential liquefaction feature map generated by the m-th convolutional kernel. Nonlinear gating to represent liquefaction risk For the u-th sample cluster, the weighted response coefficient of the m-th convolutional kernel to the 0-th key influencing indicator is given. This represents the 0th key impact indicator. This represents the bias term of the m-th convolutional kernel; m is the index of the convolutional kernel within the convolutional layer, o is the index of the key influencing metric, and u is the index of the sample cluster. The specific formula used to calculate the weighted response coefficient of the m-th convolutional kernel to the 0-th key influencing indicator for the u-th sample cluster is as follows: In the formula, These are the standard weighted response coefficients of the convolution kernel; Let u be the standard value of the cluster center vibration intensity index of the u-th sample cluster. The standard value of the liquefaction resistance index of the cluster center of the u-th sample cluster; It should be noted that the optimized weights reflect the contribution of this key influencing indicator to the liquefaction triggering mechanism, while the liquefaction resistance index reflects the soil's ability to resist liquefaction under external forces such as earthquakes; the higher the index value, the stronger the liquefaction resistance. This index can be used to assess the liquefaction resistance characteristics of sample clusters. Utilizing the properties of the natural logarithm, the relative change in liquefaction resistance with respect to vibration intensity is introduced; thus, when... When the resistance to liquefaction is improved (i.e., the resistance to liquefaction is enhanced), the weighted response coefficient will increase accordingly, but the rate of increase will gradually slow down; this is to avoid overcompensation of the weighted response coefficient. The seismic intensity index is used to quantitatively characterize the liquefaction driving force generated during an earthquake. The larger the index, the more significant the impact of the earthquake on soil liquefaction, and the greater its contribution to the accuracy of the prediction model in predicting soil liquefaction. Therefore, it is proportional to the weight response coefficient of the key impact indicators, making the prediction model focus on training data with significant characteristics.

[0047] Pooling layer: Dimensionality reduction is performed using a max pooling layer with a step size of 2.

[0048] Dropout layer: Set the dropout rate to 0.5, randomly discard some neurons to prevent overfitting.

[0049] Fully connected layer: After flattening the feature layer, the high-order feature vector is output through the fully connected layer.

[0050] The logic behind using the SSA-CNN-SVM coupled model to assess the liquefaction risk of the region under analysis is as follows: determine the geological region to which the region under analysis belongs, calculate the Mahalanobis distance between the region under analysis and the sample region, determine the sample cluster to which the region under analysis belongs, select the corresponding SSA-CNN-SVM coupled model, and use the key influencing indicators of the region under analysis as input to output the liquefaction status label of the region under analysis.

[0051] Please see Figure 4 The present invention also provides a sand seismic liquefaction discrimination system based on SSA-CNN-SVM model coupling, used to execute the above-mentioned sand seismic liquefaction discrimination method based on SSA-CNN-SVM model coupling, characterized in that it includes: The information acquisition module is used to acquire the geographical feature parameters, statistical feature parameters, key impact indicators and liquefaction status labels of each sample area. The sample area is either a liquefied area or a non-liquefied area. Based on the key impact indicators and statistical feature parameters, the liquefaction resistance index, which characterizes the liquefaction resistance of the sample area, and the seismic intensity index, which characterizes the seismic intensity at the sample area, are determined. The preliminary regional division module is used to calculate the weighted similarity measure between each sample region on the geographical feature parameters based on the geographical feature parameters, and to perform cluster analysis on all sample regions using the agglomerative hierarchical clustering algorithm to divide all sample regions into several geological regions. The liquefaction mechanism analysis module is used to calculate the Mahalanobis distance between sample areas within any geological region based on the liquefaction resistance index and vibration intensity index, and to perform secondary clustering of the sample areas within the geological region using a density-based spatial clustering algorithm, so as to divide the sample areas within the geological region into several sample clusters with similar liquefaction mechanisms. The risk assessment module is used to independently construct an SSA-CNN-SVM coupled model for any sample cluster, in order to establish the mapping relationship between key impact indicators and liquefaction status labels of each sample area within the sample cluster, thereby assessing the liquefaction risk of the area to be analyzed through the SSA-CNN-SVM coupled model.

[0052] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0053] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0054] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0055] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for identifying seismic liquefaction of sandy soil based on a coupled SSA-CNN-SVM model, characterized in that, The specific steps include: Geographic feature parameters, statistical feature parameters, key impact indicators and liquefaction status labels of each sample area are obtained. The sample area is either a liquefied area or a non-liquefied area. Based on the key impact indicators and statistical feature parameters, the liquefaction resistance index, which characterizes the liquefaction resistance of the sample area, and the seismic intensity index, which characterizes the seismic intensity of the sample area, are determined. Based on the aforementioned geographic feature parameters, the weighted similarity measure between each pair of sample regions on the geographic feature parameters is calculated, and the agglomerative hierarchical clustering algorithm is used to perform cluster analysis on all sample regions to divide all sample regions into several geological regions. For any geological region, the Mahalanobis distance between sample regions within the geological region is calculated based on the liquefaction resistance index and the vibration intensity index. Then, a density-based spatial clustering algorithm is used to perform secondary clustering of the sample regions within the geological region, so as to divide the sample regions within the geological region into several sample clusters with similar liquefaction mechanisms. For any sample cluster, an independent SSA-CNN-SVM coupled model is constructed to establish the mapping relationship between key influencing indicators and liquefaction status labels of each sample region within the sample cluster. In this way, the liquefaction risk of the region to be analyzed is assessed through the SSA-CNN-SVM coupled model.

2. The method for determining seismic liquefaction of sandy soil based on SSA-CNN-SVM model coupling according to claim 1, characterized in that, The key influencing indicators include the current groundwater depth, standard penetration test blow count, average grain size, uniformity coefficient, overlying effective pressure, dynamic shear stress ratio, and historical seismic intensity of the sample area. The liquefaction status labels include Level 1 liquefaction, Level 2 liquefaction, Level 3 liquefaction, and no liquefaction. The severity of liquefaction in sandy soil increases sequentially with the increase of the status label level. For the collected key impact indicator data, missing values ​​were preprocessed, specifically including numerical imputation using the K-nearest neighbor method.

3. The method for determining seismic liquefaction of sandy soil based on SSA-CNN-SVM model coupling according to claim 2, characterized in that, The geographic feature parameters include the geological age code, sedimentary environment code, and topographic type code of the sample area; The statistical characteristic parameters include the current clay content and historical earthquake magnitude of the sample area; The logic behind calculating the liquefaction resistance index is as follows: based on statistical characteristic parameters and key influencing indicators, the liquefaction resistance index is comprehensively calculated according to the basic theory of soil mechanics. The specific formula used is as follows: In the formula, Let i be the liquefaction resistance index of the i-th sample region. Let be the current standard penetration test (SPT) hit count for the i-th sample region. The dynamic shear stress ratio of the i-th sample region is given. Let i be the current clay content of the i-th sample region, where i is the index of the sample region. The specific method for calculating the seismic intensity index is as follows: based on statistical characteristic parameters, combined with key influencing indicators, and based on the basic theory of earthquake engineering, the seismic intensity index is calculated using the following formula: In the formula, Let be the vibration intensity index of the i-th sample region. Let be the average historical earthquake magnitude of the i-th sample area. Let be the historical average earthquake intensity of the i-th sample area. Let be the current groundwater depth in the i-th sample area.

4. The method for determining seismic liquefaction of sandy soil based on SSA-CNN-SVM model coupling according to claim 3, characterized in that, The logic behind calculating the weighted similarity measure of all sample regions across geographic feature parameters is as follows: The entropy weight method is used to determine the weight of each geographic feature parameter. Specifically, the logic is: determine the proportion of the i-th sample region under the k-th geographic feature parameter; determine the entropy value of the k-th geographic feature parameter based on this proportion; calculate the weight of the k-th geographic feature parameter based on its entropy value; and the formula for calculating the proportion of the i-th sample region under the k-th geographic feature parameter is as follows: In the formula, This represents the proportion of the i-th sample region under the k-th geographical feature parameter. Let be the parameter value of the i-th sample region for the k-th geographical feature. is the total number of sample areas, and k is the index of the types of geographic feature parameters; The formula used to determine the entropy value of the kth geographic feature parameter is as follows: In the formula, Let be the entropy value of the k-th geographic feature parameter; The formula used to calculate the weight of the kth geographic feature parameter is: In the formula, The weight of the k-th geographical feature parameter; A weighted Euclidean distance is constructed based on the weights of geographical feature parameters, serving as a weighted similarity measure characterizing the geological similarity between sample regions. The specific formula for calculating the weighted Euclidean distance is as follows: In the formula, The weighted Euclidean distance between the i-th sample region and the j-th sample region is... Let be the parameter value of the i-th sample region for the k-th geographical feature. The parameter value of the j-th sample region for the k-th geographic feature, where j is the index of the sample region, and .

5. The method for determining seismic liquefaction of sandy soil based on SSA-CNN-SVM model coupling according to claim 4, characterized in that, The specific steps of clustering all sample regions using the agglomerative hierarchical clustering algorithm include: treating each sample region as a separate cluster, calculating the weighted Euclidean distance between all clusters, merging the two clusters with the closest weighted Euclidean distance each time, and recalculating the weighted Euclidean distance between all clusters, repeating the merging operation until all samples are merged into one cluster to form a clustering tree, and cutting the clustering tree according to multiple preset similarity thresholds so that the clustering tree is cut into several cut clusters under each similarity threshold; Calculate the average contour coefficient of each cut cluster under each similarity threshold, select the similarity threshold corresponding to the maximum average contour coefficient as the optimal cut threshold, and take the cut clusters obtained by cutting the cluster tree with the optimal cut threshold as the precise cut clusters. Each precise cut cluster is a geological region, so as to divide all sample areas into several geological regions.

6. The method for determining seismic liquefaction of sandy soil based on SSA-CNN-SVM model coupling according to claim 5, characterized in that, The logic underlying the calculation of Mahalanobis distances between sample regions within each geological region is as follows: The liquefaction resistance index and seismic intensity index of sample areas within any geological region are normalized to obtain standard values ​​for the liquefaction resistance index and seismic intensity index. Based on these standard values, the Mahalanobis distance between sample areas within the geological region is calculated using the following formula: In the formula, Let be the Mahalanobis distance between the p-th and q-th sample regions within the geological region. This is the inverse matrix of the geological region's covariance matrix, which includes the covariance matrices of the liquefaction resistance index and the seismic intensity index of all sample areas within the geological region. This is the standardized feature vector of the p-th sample region within the geological region, specifically a vector including the standard values ​​of the liquefaction resistance index and the vibration intensity index. Let p be the standardized feature vector of the q-th sample region within the geological region, and p and q be the indices of the sample regions within the geological region. ; The specific logic for secondary clustering is as follows: For any geological region, calculate the Mahalanobis distance between all sample regions within it, and perform clustering based on the Mahalanobis distance using a density-based spatial clustering algorithm. This includes: using the k-distance graph method to determine the neighborhood radius of the density-based spatial clustering algorithm, calculating the Mahalanobis distance from each sample region to its preset number of nearest neighbors, sorting and plotting the obtained nearest neighbor Mahalanobis distances, selecting the distance corresponding to the inflection point of the curve as the neighborhood radius, and setting the minimum number of samples. All sample regions are marked as unvisited. An unvisited sample region is randomly selected, and the number of unvisited sample regions within its neighborhood radius is checked to see if it is not less than the set minimum number of samples. If not, no new cluster is created. If so, a new cluster is created with the unvisited sample region as the core point, and all unvisited sample regions within the neighborhood radius of the core point are added to the newly created cluster. The sample regions in the new cluster are updated to visited. Another unvisited sample region is randomly selected, and the above operation is repeated until no more new clusters are built. Each new cluster is regarded as a group of sample clusters with similar liquefaction mechanisms.

7. The method for determining seismic liquefaction of sandy soil based on SSA-CNN-SVM model coupling according to claim 6, characterized in that: The SSA-CNN-SVM coupled model specifically refers to the Sparrow Search algorithm-convolutional neural network-support vector machine coupled model. The Sparrow Search algorithm is used to automatically and globally optimize the key hyperparameters of the convolutional neural network to obtain the optimal model initialization configuration. The key hyperparameters include learning rate, batch size and L2 regularization coefficient.

8. The method for determining seismic liquefaction of sandy soil based on SSA-CNN-SVM model coupling according to claim 7, characterized in that: Based on the optimal model initialization configuration, the key hyperparameters of the convolutional neural network are set. The convolutional neural network is used for feature extraction. The key influence indicators of the sample regions in the same sample cluster are used as training data. The key influence indicators are used as inputs and input into the configured convolutional neural network. The deep feature vectors of the input key influence indicators are extracted by the convolutional neural network. The convolutional neural network comprises, in sequence, an input layer, a convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, a Dropout layer, and a fully connected layer; The logic behind extracting deep feature vectors is as follows: based on the standard values ​​of the anti-liquefaction intensity index and vibration intensity index of the cluster centers of each sample cluster, the weight response coefficients of the convolution kernel in the convolutional layer to the key influencing indicators are adaptively adjusted. For any sample cluster, a latent liquefaction feature map is output through the convolutional layer of the corresponding convolutional neural network. Based on the latent liquefaction feature map, a deep feature vector is output through a batch normalization layer, an activation function layer, a max pooling layer, a Dropout layer, and a fully connected layer. The latent liquefaction feature map is specifically represented as follows: In the formula, This represents the potential liquefaction feature map generated by the m-th convolutional kernel. Nonlinear gating to represent liquefaction risk For the u-th sample cluster, the weighted response coefficient of the m-th convolutional kernel to the 0-th key influencing indicator is given. This represents the 0th key impact indicator. This represents the bias term of the m-th convolutional kernel; m is the index of the convolutional kernel within the convolutional layer, o is the index of the key influencing metric, and u is the index of the sample cluster. The specific formula used to calculate the weighted response coefficient of the m-th convolutional kernel to the 0-th key influencing indicator for the u-th sample cluster is as follows: In the formula, These are the standard weighted response coefficients of the convolution kernel; Let u be the standard value of the cluster center vibration intensity index of the u-th sample cluster. The standard value of the liquefaction resistance index of the cluster center of the u-th sample cluster; The deep feature vector is used as the input to the support vector machine model, and the corresponding liquefaction state label is used as the output result for training. The key influencing indicators of the sandy soil to be identified are input into the trained prediction model, and the model outputs the corresponding liquefaction state label. The logic behind using the SSA-CNN-SVM coupled model to assess the liquefaction risk of the region under analysis is as follows: determine the geological region to which the region under analysis belongs, calculate the Mahalanobis distance between the region under analysis and the sample region, determine the sample cluster to which the region under analysis belongs, select the corresponding SSA-CNN-SVM coupled model, and use the key influencing indicators of the region under analysis as input to output the liquefaction status label of the region under analysis.

9. A sand seismic liquefaction discrimination system based on SSA-CNN-SVM model coupling, used to execute the sand seismic liquefaction discrimination method based on SSA-CNN-SVM model coupling as described in any one of claims 1-8, characterized in that, include: The information acquisition module is used to acquire the geographical feature parameters, statistical feature parameters, key impact indicators and liquefaction status labels of each sample area. The sample area is either a liquefied area or a non-liquefied area. Based on the key impact indicators and statistical feature parameters, the liquefaction resistance index, which characterizes the liquefaction resistance of the sample area, and the seismic intensity index, which characterizes the seismic intensity at the sample area, are determined. The preliminary regional division module is used to calculate the weighted similarity measure between each sample region on the geographical feature parameters based on the geographical feature parameters, and to perform cluster analysis on all sample regions using the agglomerative hierarchical clustering algorithm to divide all sample regions into several geological regions. The liquefaction mechanism analysis module is used to calculate the Mahalanobis distance between sample areas within any geological region based on the liquefaction resistance index and vibration intensity index, and to perform secondary clustering of the sample areas within the geological region using a density-based spatial clustering algorithm, so as to divide the sample areas within the geological region into several sample clusters with similar liquefaction mechanisms. The risk assessment module is used to independently construct an SSA-CNN-SVM coupled model for any sample cluster, in order to establish the mapping relationship between key impact indicators and liquefaction status labels of each sample area within the sample cluster, thereby assessing the liquefaction risk of the area to be analyzed through the SSA-CNN-SVM coupled model.

Images