A machine learning-based method and system for predicting bridge seismic vulnerability

Abstract

The application discloses a bridge earthquake vulnerability prediction method and system based on machine learning, comprising: obtaining a set of earthquake parameters, constructing a nonlinear mapping model using a deep neural network, predicting vulnerability probability based on the nonlinear mapping model to obtain preliminary vulnerability probability data; obtaining bridge structure data to construct a bridge geometric model, analyzing the segmented material heterogeneity of the bridge geometric model, and if the heterogeneity exceeds a preset threshold, calculating segmented response data under seismic action; based on the segmented response data, analyzing the spatial difference vulnerability probability distribution of the bridge as a whole to obtain fine prediction results; if the deviation between the fine prediction results and the preliminary vulnerability probability exceeds a preset threshold, adjusting the nonlinear mapping model, outputting the updated vulnerability probability distribution, and obtaining the final vulnerability prediction results. The application significantly improves the accuracy and reliability of bridge earthquake vulnerability prediction, and has important significance for improving the seismic safety of bridge structures.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bridge seismic vulnerability prediction, and in particular to a bridge seismic vulnerability prediction method and system based on machine learning. BACKGROUND

[0002] Bridge seismic engineering is an important branch of civil engineering, which is related to infrastructure safety and social stability. Its research has a key significance for protecting life and property safety and maintaining the function of the transportation network. With the frequent occurrence of earthquake disasters, accurately predicting the seismic vulnerability of bridges has become an urgent need, which directly affects the efficiency of post-disaster rescue and recovery. However, the existing prediction methods have obvious deficiencies in accuracy and efficiency. Many traditional methods rely on simplified models or empirical formulas, which are difficult to fully capture the complex nonlinear relationship between ground motion parameters and bridge failure probability, resulting in large prediction errors. In addition, some methods do not fully consider the heterogeneity of bridge structures, ignoring the unique response characteristics of different parts in earthquakes, which limits the degree of refinement of the prediction.

[0003] In the prediction of bridge seismic vulnerability, the core challenge lies in how to effectively associate ground motion parameters with failure probability and how to deal with the complex characteristics of bridge structures. Ground motion parameters are diverse and coupled with each other, and traditional methods are difficult to build an efficient mapping model, limiting the prediction accuracy. At the same time, due to the differences in geometry, material distribution and connection method, bridge structures show significant spatial heterogeneity, and overall analysis often hides the vulnerability characteristics of local weak links. If feature extraction and prediction are not performed for segmented structures, the comprehensive results may be distorted. Therefore, how to learn the relationship between ground motion parameters and failure probability, build an efficient prediction model, and combine the segmented characteristics of bridge structures for refined prediction has become a key problem to improve the prediction accuracy of bridge seismic vulnerability. SUMMARY

[0004] To solve the above technical problems, the present application provides a bridge seismic vulnerability prediction method based on machine learning to overcome the limitations of traditional methods that ignore the heterogeneity of bridge structures and provide technical support for bridge seismic design and reinforcement.

[0005] To achieve the above purpose, the present application provides a bridge seismic vulnerability prediction method based on machine learning, comprising:

[0006] Obtain a set of seismic parameters, construct a nonlinear mapping model using a deep neural network, and predict the vulnerability probability based on the nonlinear mapping model to obtain preliminary vulnerability probability data;

[0007] Obtain bridge structure data to construct a bridge geometric model, analyze the segmented material heterogeneity of the bridge geometric model, and if the heterogeneity exceeds a preset threshold, calculate the segmented response data under seismic action;

[0008] Based on the segmented response data, the spatial difference vulnerability probability distribution of the whole bridge is analyzed to obtain a fine prediction result.

[0009] If the fine prediction result deviates from the preliminary vulnerability probability by more than a preset threshold, the nonlinear mapping model is adjusted, and an updated vulnerability probability distribution is output to obtain a final vulnerability prediction result.

[0010] Preferably, obtaining the set of seismic parameters comprises:

[0011] Obtaining multi-dimensional ground motion data from a ground motion parameter database and standardizing the data;

[0012] Using a principal component analysis algorithm to extract features from the standardized ground motion data set, separate key characteristic acceleration values, frequency distribution and duration, and obtain a set of seismic parameters.

[0013] Preferably, obtaining bridge structure data to construct a bridge geometric model comprises:

[0014] Obtaining bridge structure data from a bridge structure database and cleaning the bridge structure data;

[0015] Extracting features from the cleaned bridge structure data to obtain geometric shape features, material distribution features and connection method features, and reducing dimensions through principal component analysis to obtain an optimized feature set;

[0016] Extracting feature boundary conditions from the optimized feature set through a finite element analysis algorithm;

[0017] Calculating the mechanical parameters of the optimized feature set according to the feature boundary conditions, and constructing a bridge geometric model according to the mechanical parameters.

[0018] Preferably, analyzing the segmented material heterogeneity of the bridge geometric model comprises:

[0019] Obtaining material data for each segment from the bridge geometric model, performing format unification and missing value filling to obtain a standardized material data set;

[0020] Grid division is performed on the standardized material data set, the geometric boundary is decomposed into calculation units, and boundary conditions are generated in combination with connection constraints to obtain a gridded unit;

[0021] Performing finite element calculation on the gridded unit to obtain segmented heterogeneity parameters;

[0022] The segmented heterogeneity parameters include material stiffness, damping characteristics and mass distribution data.

[0023] Preferably, calculating the segmented response data under the action of the earthquake comprises:

[0024] Based on the segment heterogeneity parameter, a long short-term memory network is used to calculate displacement distribution data, velocity distribution data, and stress distribution data of each segment structure under the action of an earthquake;

[0025] Segment response data is obtained based on the displacement distribution data, the velocity distribution data, and the stress distribution data.

[0026] Preferably, obtaining the fine prediction result comprises:

[0027] A machine learning classification algorithm is used to classify the segment response data to obtain a set of segment vulnerability probability distributions;

[0028] A spatial interpolation algorithm is used to perform spatial interpolation on the set of segment vulnerability probability distributions to generate a spatially differentiated vulnerability probability distribution of the bridge as a whole, thereby obtaining the fine prediction result.

[0029] Preferably, classifying the segment response data comprises:

[0030] The segment response data is preprocessed to obtain a set of segment structure features;

[0031] The set of segment structure features is classified based on a support vector machine to identify potential weak points and obtain weak point distribution data;

[0032] Based on the weak point distribution data, a probability density estimation algorithm is used to calculate the vulnerability probability of each weak point;

[0033] The vulnerability probability of each weak point and the stress distribution data are integrated to obtain the set of segment vulnerability probability distributions.

[0034] Preferably, the set of segment vulnerability probability distributions is spatially interpolated to generate a spatially differentiated vulnerability probability distribution of the bridge as a whole, thereby obtaining the fine prediction result, which comprises:

[0035] Local probability data is obtained from the set of segment vulnerability probability distributions;

[0036] Based on the local probability data, a data screening algorithm is used to screen high-probability regions to obtain a local high-risk data set;

[0037] Based on the local high-risk data set, a spatial interpolation algorithm is used to calculate probability transition values between regions to obtain a spatially differentiated vulnerability probability distribution of the bridge as a whole;

[0038] The spatially differentiated vulnerability probability distribution of the bridge as a whole is spatially smoothed and optimized to obtain the fine prediction result.

[0039] Preferably, if the fine prediction result deviates from the preliminary vulnerability probability by more than a preset threshold, the non-linear mapping model is adjusted, an updated vulnerability probability distribution is output, and a final vulnerability prediction result is obtained, including:

[0040] The difference between the fine prediction result and the preliminary vulnerability probability data is calculated by using a deviation detection algorithm, and a difference value is obtained.

[0041] If the difference value exceeds a preset threshold, a high-deviation region is screened by using a data screening method, and a high-deviation data set is obtained.

[0042] According to the high-deviation data set, the parameters of the non-linear mapping model are adjusted by using a parameter optimization algorithm, the mapping relationship is updated by iterative calculation, and an adjusted non-linear mapping model is obtained.

[0043] The seismic parameter set is re-input into the adjusted non-linear mapping model to obtain a vulnerability probability distribution, and a final vulnerability prediction result is obtained.

[0044] A bridge seismic vulnerability prediction system based on machine learning, comprising:

[0045] A ground motion parameter acquisition module is configured to construct a non-linear mapping model by using a deep neural network, perform vulnerability probability prediction based on the non-linear mapping model, and obtain preliminary vulnerability probability data.

[0046] A bridge model construction module is configured to obtain bridge structure data to construct a bridge geometric model, analyze the segmented material heterogeneity of the bridge geometric model, and calculate segmented response data under seismic action if the heterogeneity exceeds a preset threshold.

[0047] A fine result prediction module is configured to analyze the spatial difference vulnerability probability distribution of the bridge as a whole based on the segmented response data, and obtain a fine prediction result.

[0048] An update prediction module is configured to determine whether the fine prediction result deviates from the preliminary vulnerability probability by more than a preset threshold, and if the threshold is exceeded, the non-linear mapping model is adjusted, an updated vulnerability probability distribution is output, and a final vulnerability prediction result is obtained.

[0049] Compared with the prior art, the present application has the following advantages and technical effects:

[0050] The application discloses a bridge seismic vulnerability prediction method based on machine learning, obtains a set of seismic parameters, adopts a deep neural network to construct a nonlinear mapping model, performs vulnerability probability prediction based on the nonlinear mapping model to obtain preliminary vulnerability probability data, obtains bridge structure data to construct a bridge geometric model, analyzes the segmented material heterogeneity of the bridge geometric model, calculates segmented response data under seismic action, analyzes the spatial difference vulnerability probability distribution of the whole bridge to obtain fine prediction results, adjusts the nonlinear mapping model if the deviation between the fine prediction results and the preliminary vulnerability probability exceeds a preset threshold, outputs updated vulnerability probability distribution to obtain final vulnerability prediction results. The method combines the nonlinear mapping model to perform vulnerability prediction, analyzes the segmented material heterogeneity of the bridge geometric model to obtain segmented response data, and then performs fine prediction, and the spatial interpolation algorithm used in the fine prediction adopts Kriging interpolation, which can generate the spatial difference vulnerability probability distribution of the whole bridge, provides fine prediction results for the seismic design and maintenance of the bridge, overcomes the limitation of ignoring the heterogeneity of the bridge structure in the traditional method, compares and analyzes the fine prediction results and the preliminary vulnerability probability, adjusts the nonlinear mapping model, and significantly improves the precision and reliability of the bridge seismic vulnerability prediction. BRIEF DESCRIPTION OF DRAWINGS

[0051] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application, and their

[0052] Figure 1 FIG. 1 is a flowchart of a bridge seismic vulnerability prediction method based on machine learning according to an embodiment of the application. DETAILED DESCRIPTION

[0053] It should be noted that the embodiments and features in the application can be combined with each other without conflict. The application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0054] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0055] As shown in FIG. 1, the embodiment proposes a bridge seismic vulnerability prediction method based on machine learning, which comprises: Figure 1 As shown in FIG. 1, the embodiment proposes a bridge seismic vulnerability prediction method based on machine learning, which comprises:

[0056] Obtaining a seismic parameter set, constructing a nonlinear mapping model by using a deep neural network, and predicting a vulnerability probability based on the nonlinear mapping model to obtain preliminary vulnerability probability data;

[0057] Obtaining bridge structure data to construct a bridge geometric model, analyzing the sectional material heterogeneity of the bridge geometric model, and calculating sectional response data under seismic action if the heterogeneity exceeds a preset threshold;

[0058] Based on the sectional response data, analyzing the spatial difference vulnerability probability distribution of the bridge as a whole to obtain a fine prediction result;

[0059] If the deviation between the fine prediction result and the preliminary vulnerability probability exceeds a preset threshold, adjusting the nonlinear mapping model, outputting an updated vulnerability probability distribution, and obtaining a final vulnerability prediction result.

[0060] Further, obtaining a seismic parameter set includes: obtaining multi-dimensional ground motion data from a ground motion parameter database, and performing standardization processing on the data; using a principal component analysis algorithm to extract features from the standardized processed ground motion data set, separate key feature acceleration values, frequency distribution and duration, and obtain the seismic parameter set.

[0061] Specifically, in the present embodiment, a certain seismic monitoring station records 1000 groups of seismic events, each group containing three-axis acceleration data, with a sampling frequency of 100 Hz and a duration of 60 seconds. The multi-dimensional ground motion data is obtained from the ground motion parameter database, and the data is cleaned by using the standardization processing technology, which can remove noise and outliers; the principal component analysis algorithm is used for feature extraction, assuming that the data set contains acceleration peak value, frequency distribution, duration and other features. The principal component analysis projects high-dimensional data into low-dimensional space through linear transformation, retaining the main information;

[0062] Further, a deep neural network is used to construct a nonlinear mapping model.

[0063] Specifically, in the present embodiment, the model architecture includes a stem module and a body module; the stem module uses multi-path deep separable convolution and linear transformation to quickly reduce the time dimension of the seismic record and reduce the subsequent computational overhead; the body module adopts a four-stage design, which simultaneously focuses on local and global features through a multi-head self-attention mechanism; training and optimization: in the training process, data augmentation methods play an important role in model performance. Randomly splitting the data set, such as 8:1:1, into training set, test set and validation set. At the same time, attention should be paid to the balance between model depth and computational complexity, and the risk of overfitting; the model can be used for seismic detection, seismic phase picking, first motion polarity classification, magnitude estimation, back azimuth estimation and epicenter distance estimation, etc. Various seismic monitoring tasks;

[0064] Further, the acquiring bridge structure data to construct the bridge geometric model comprises: acquiring bridge structure data from the bridge structure database, performing data cleaning on the bridge structure data; performing feature extraction on the cleaned bridge structure data to acquire geometric shape features, material distribution features and connection mode features, and obtaining an optimized feature set through principal component analysis dimension reduction; extracting feature boundary conditions from the optimized feature set through a finite element analysis algorithm; calculating mechanical parameters of the optimized feature set according to the feature boundary conditions, and constructing the bridge geometric model according to the mechanical parameters.

[0065] Specifically, in the embodiment, the data cleaning includes filtering outliers and noises: ensuring the accuracy and reliability of the data through filtering and data calibration; ensuring the time consistency and accuracy of different sensor data through data synchronization and calibration; the geometric shape features of the feature extraction include bridge span arrangement, main beam form, bridge deck width, etc.; the material distribution features include elastic modulus, Poisson's ratio and the like of different component materials; the connection mode features include pier type, abutment form, etc., and a bridge geometric model is simulated based on the extracted bridge geometric features.

[0066] Further, the analyzing the segmented material heterogeneity of the bridge geometric model comprises: acquiring material data of each segment from the bridge geometric model, performing format unification and missing value filling to obtain a standardized material data set; performing grid division on the standardized material data set to decompose geometric boundaries into calculation units, generating boundary conditions in combination with connection constraints to acquire grid units; performing finite element calculation on the grid units to acquire segmented heterogeneity parameters; the segmented heterogeneity parameters include material stiffness, damping characteristics and mass distribution data.

[0067] Specifically, in the embodiment, the material data of each segment is extracted from the bridge geometric model, and the data includes elastic modulus, Poisson's ratio, density and the like of the material; the standardized material data set is divided into grids to decompose geometric boundaries into calculation units. An adaptive grid division method, such as a material quadtree or octree-based division method, can be used to ensure that the grid can accurately reflect the material heterogeneity; in combination with the connection constraints, corresponding boundary conditions are generated for each grid unit, and the boundary conditions include fixed supports, applied loads and the like; the grid units are calculated using a finite element analysis software (such as ANSYS or ABAQUS);

[0068] Further, if the heterogeneity exceeds a preset threshold, segmented response data under the action of an earthquake is calculated; if the segmented material heterogeneity does not exceed the threshold, preliminary vulnerability probability data is directly output.

[0069] Further, the segmented response data under the earthquake action is calculated, including: based on the segmented heterogeneity parameters, using a long short-term memory network to calculate displacement distribution data, velocity distribution data, stress distribution data of each segment structure under the earthquake action; and based on the displacement distribution data, the velocity distribution data and the stress distribution data, the segmented response data is obtained.

[0070] Specifically, in the embodiment, the long short-term memory network is constructed and trained; the network structure: an LSTM network is constructed, the input is the acceleration time history of ground motion and the heterogeneity parameters of the structure, and the output is the response data such as displacement, velocity and stress of the structure. A stacked structure can be used, for example, 3 one-way LSTM layers and 1 fully connected layer; the training data: using historical earthquake data and corresponding structure response data as the training set. The data needs to be divided into subsequences to reduce the training difficulty and time consumption; the training process: using mean square error as the loss function, using Adam optimizer for training, and preventing overfitting through dropout operation;

[0071] Further, obtaining the fine prediction result includes: using a machine learning classification algorithm to classify the segmented response data to obtain a segmented vulnerability probability distribution set; and using a spatial interpolation algorithm to spatially interpolate the segmented vulnerability probability distribution set to generate a spatial difference vulnerability probability distribution of the whole bridge, and obtaining the fine prediction result.

[0072] Specifically, in the embodiment, the machine learning classification algorithm selects a random forest algorithm, which has good robustness to high-dimensional data and noise. The standardized segmented response data is used as the input to train the selected classification model. The goal of the model is to classify the response data of each segment structure into different vulnerability levels (such as low, medium and high). Through the prediction result of the model, the vulnerability probability distribution of each segment is obtained. The spatial interpolation algorithm uses Kriging interpolation, which can consider spatial autocorrelation and is suitable for cases where data is unevenly distributed. Specifically, the segmented vulnerability probability distribution set is taken as the known data points, and the selected spatial interpolation algorithm is used for interpolation according to the spatial layout of the bridge. For example, when using Kriging interpolation, the semi-variogram function needs to be modeled first, and then interpolation is performed according to the minimum variance estimation principle. Continuous distribution is generated: through spatial interpolation, the discrete segmented vulnerability probability distribution is expanded to the whole space of the bridge, and the spatial difference vulnerability probability distribution of the whole bridge is generated.

[0073] Further, classifying the segmented response data includes: preprocessing the segmented response data to obtain a segmented structure feature set; classifying the segmented structure feature set based on a support vector machine to identify potential weak points and obtain weak point distribution data; calculating the vulnerability probability of each weak point using a probability density estimation algorithm based on the weak point distribution data; integrating the vulnerability probability of each weak point and the stress distribution data to obtain a segmented vulnerability probability distribution set;

[0074] Specifically, in this embodiment, a suitable kernel function is selected according to the distribution of data. For linearly separable data, a linear kernel can be used; for nonlinearly separable data, a Gaussian kernel can be used; the segmented structural feature set is used as input to train the SVM classifier. The parameters of the classifier, such as the penalty parameter C and the parameters of the kernel function, are optimized by cross-validation and other methods, the trained SVM classifier is used to classify each segment, and potential weak points are identified; and the vulnerability probability of each weak point and the stress distribution data are integrated to obtain the vulnerability probability distribution of each segment.

[0075] Further, the segmented vulnerability probability distribution set is spatially interpolated to generate a spatially differentiated vulnerability probability distribution of the bridge as a whole, and a fine prediction result is obtained, including: obtaining local probability data from the segmented vulnerability probability distribution set; based on the local probability data, a data screening algorithm is used to screen high-probability areas to obtain a local high-risk data set; based on the local high-risk data set, a spatial interpolation algorithm is used to calculate the probability transition value between each area to obtain a spatially differentiated vulnerability probability distribution of the bridge as a whole; and the spatially differentiated vulnerability probability distribution of the bridge as a whole is spatially smoothed and optimized to obtain a fine prediction result.

[0076] Specifically, in this embodiment, the vulnerability probability data set of a certain bridge segment contains 100 units, the probability value ranges from 0.1 to 0.5, and the screening threshold is set to 0.3. Therefore, units with a probability greater than 0.3 are marked as high-risk.

[0077] Further, if the deviation between the fine prediction result and the preliminary vulnerability probability exceeds a preset threshold, the nonlinear mapping model is adjusted, and an updated vulnerability probability distribution is output to obtain a final vulnerability prediction result, including: using a deviation detection algorithm to calculate the difference between the fine prediction result and the preliminary vulnerability probability data to obtain a difference value; if the difference value exceeds the preset threshold, a high-deviation area is screened by a data screening method to obtain a high-deviation data set; according to the high-deviation data set, the parameters of the nonlinear mapping model are adjusted by a parameter optimization algorithm, the mapping relationship is updated by iterative calculation, and an adjusted nonlinear mapping model is obtained; the seismic parameter set is re-input into the adjusted nonlinear mapping model to obtain the final vulnerability prediction result.

[0078] Specifically, in this embodiment, 50 high-deviation units are detected in a certain bridge segment, the deviation value ranges from 0.15 to 0.3, and the screening condition is that the deviation value is greater than 0.2 and located near the bridge connection. Finally, 10 high-deviation units are extracted. The high-deviation data set composed of these units provides a precise target for subsequent model optimization.

[0079] Embodiment 2

[0080] In this embodiment, a bridge seismic vulnerability prediction system based on machine learning is proposed, comprising: a ground motion parameter acquisition module, configured to construct a nonlinear mapping model by using a deep neural network, to perform vulnerability probability prediction based on the nonlinear mapping model, and to obtain preliminary vulnerability probability data; a bridge model construction module, configured to obtain bridge structure data to construct a bridge geometric model, to analyze the segmented material heterogeneity of the bridge geometric model, and to calculate segmented response data under seismic action if the heterogeneity exceeds a preset threshold; a fine result prediction module, configured to analyze the spatial difference vulnerability probability distribution of the bridge as a whole based on the segmented response data, and to obtain fine prediction results; and an update prediction module, configured to judge whether the deviation between the fine prediction results and the preliminary vulnerability probability exceeds a preset threshold, to adjust the nonlinear mapping model if the threshold is exceeded, to output updated vulnerability probability distribution, and to obtain final vulnerability prediction results. The system proposed in the present application fully considers the complexity and dynamics of seismic vulnerability prediction, and can provide strong support for the seismic performance evaluation of bridges.

[0081] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A machine learning based bridge seismic vulnerability prediction method, characterized by, The method comprises the following steps: Obtain a seismic parameter set, construct a nonlinear mapping model using a deep neural network, and perform vulnerability probability prediction based on the nonlinear mapping model to obtain preliminary vulnerability probability data; Obtain bridge structure data to construct a bridge geometric model and analyze the segmented material heterogeneity of the bridge geometric model: Obtain material data for each segment from the bridge geometric model, perform format unification and missing value filling to obtain a standardized material data set, perform mesh division on the standardized material data set, decompose geometric boundaries into calculation units, generate boundary conditions in combination with connection constraints, and obtain meshed units; Perform finite element calculation on the meshed units to obtain segmented heterogeneity parameters; the segmented heterogeneity parameters include material stiffness, damping characteristics, and mass distribution data; If the heterogeneity exceeds a preset threshold, calculate segmented response data under seismic action: Based on the segmented heterogeneity parameters, use a long short-term memory network to calculate displacement distribution data, velocity distribution data, and stress distribution data of each segment under seismic action; and obtain segmented response data based on the displacement distribution data, the velocity distribution data, and the stress distribution data; Based on the segmented response data, analyze the spatial difference vulnerability probability distribution of the bridge as a whole to obtain a refined prediction result: use a machine learning classification algorithm to classify the segmented response data to obtain a segmented vulnerability probability distribution set; use a spatial interpolation algorithm to perform spatial interpolation on the segmented vulnerability probability distribution set to generate a spatial difference vulnerability probability distribution of the bridge as a whole, and obtain a refined prediction result; If the refined prediction result deviates from the preliminary vulnerability probability by more than a preset threshold, adjust the nonlinear mapping model and output an updated vulnerability probability distribution to obtain a final vulnerability prediction result. 2.The machine learning based bridge seismic vulnerability prediction method of claim 1, wherein, Obtaining a seismic parameter set comprises: Obtaining multi-dimensional ground motion data from a ground motion parameter database and performing standardization processing on the data; Using a principal component analysis algorithm to extract features from the standardized ground motion data set, separate key feature acceleration values, frequency distribution, and duration, and obtain a seismic parameter set. 3.The machine learning based bridge seismic vulnerability prediction method of claim 1, wherein, Obtaining bridge structure data to construct a bridge geometric model comprises: Obtaining bridge structure data from a bridge structure database and performing data cleaning on the bridge structure data; Extracting features from the cleaned bridge structure data to obtain geometric shape features, material distribution features, and connection mode features, and reducing dimensions through principal component analysis to obtain an optimized feature set; Extracting feature boundary conditions from the optimized feature set through a finite element analysis algorithm; Calculating mechanical parameters of the optimized feature set according to the feature boundary conditions and constructing a bridge geometric model according to the mechanical parameters. 4.The machine learning based bridge seismic vulnerability prediction method of claim 1, wherein, Classifying the segmented response data comprises: Preprocessing the segmented response data to obtain a segmented structure feature set; Classifying the segmented structure feature set based on a support vector machine to identify potential weak points and obtain weak point distribution data; Using a probability density estimation algorithm to calculate the vulnerability probability of each weak point based on the weak point distribution data; Integrating the vulnerability probability of each weak point and the stress distribution data to obtain a segmented vulnerability probability distribution set. 5.The machine learning based bridge seismic vulnerability prediction method of claim 1, wherein, The segmented vulnerability probability distribution set is spatially interpolated to generate a spatially differentiated vulnerability probability distribution of the bridge as a whole, and a fine prediction result is obtained, including: Obtaining local probability data from the segmented vulnerability probability distribution set; Based on the local probability data, a data screening algorithm is used to screen high-probability areas to obtain a local high-risk data set; Based on the local high-risk data set, a spatial interpolation algorithm is used to calculate the probability transition value between each area to obtain a spatially differentiated vulnerability probability distribution of the bridge as a whole; The spatially differentiated vulnerability probability distribution of the bridge as a whole is spatially smoothed and optimized to obtain a fine prediction result. 6.The machine learning based bridge seismic vulnerability prediction method of claim 1, wherein, If the fine prediction result deviates from the preliminary vulnerability probability by more than a preset threshold, the nonlinear mapping model is adjusted, and an updated vulnerability probability distribution is output to obtain a final vulnerability prediction result, including: A deviation detection algorithm is used to calculate the difference between the fine prediction result and the preliminary vulnerability probability data to obtain a difference value; If the difference value exceeds the preset threshold, a high-deviation area is screened out by a data screening method to obtain a high-deviation data set; According to the high-deviation data set, the parameters of the nonlinear mapping model are adjusted by a parameter optimization algorithm, the mapping relationship is updated by iterative calculation, and an adjusted nonlinear mapping model is obtained; The set of seismic parameters is input into the adjusted nonlinear mapping model to obtain a final vulnerability prediction result.

7. A machine learning based bridge seismic vulnerability prediction system characterized in that, Including: The seismic parameter acquisition module is used to construct a nonlinear mapping model using a deep neural network, perform vulnerability probability prediction based on the nonlinear mapping model, and obtain preliminary vulnerability probability data; The bridge model construction module is used to obtain bridge structure data to construct a bridge geometric model, analyze the segmented material heterogeneity of the bridge geometric model, obtain segmented material data from the bridge geometric model, perform format unification and missing value filling to obtain a standardized material data set, perform grid division on the standardized material data set, decompose geometric boundaries into calculation units, generate boundary conditions combined with connection constraints, and obtain a gridded unit; The gridded unit is subjected to finite element calculation to obtain segmented heterogeneity parameters; the segmented heterogeneity parameters include material stiffness, damping characteristics, and mass distribution data; If the heterogeneity exceeds a preset threshold, calculate the segmented response data under seismic action: based on the segmented heterogeneity parameters, use a long short-term memory network to calculate the displacement distribution data, velocity distribution data, and stress distribution data of each segment under seismic action; and obtain segmented response data based on the displacement distribution data, velocity distribution data, and stress distribution data; The fine result prediction module is used to analyze the spatially differentiated vulnerability probability distribution of the bridge as a whole based on the segmented response data to obtain a fine prediction result: a machine learning classification algorithm is used to classify the segmented response data to obtain a segmented vulnerability probability distribution set; and a spatial interpolation algorithm is used to spatially interpolate the segmented vulnerability probability distribution set to generate a spatially differentiated vulnerability probability distribution of the bridge as a whole, and obtain a fine prediction result; An updating prediction module is configured to determine whether the deviation between the fine prediction result and the preliminary vulnerability probability exceeds a preset threshold value, adjust the non-linear mapping model if the deviation exceeds the threshold value, output an updated vulnerability probability distribution, and obtain a final vulnerability prediction result.

Images