A two-stage underground station seismic damage intelligent prediction method

Abstract

This invention relates to a two-stage intelligent prediction method for seismic damage to underground railway stations, comprising: S1: using a first-stage deep learning prediction model to predict the samples to be predicted, obtaining the first-stage prediction result; S2: calculating the prediction confidence and classification interval based on the prediction probability and the preliminary prediction result of the seismic damage state, and determining whether the first-stage prediction result meets the preset reliability conditions. If it does, the first-stage prediction result is used as the final seismic damage state prediction result; if it does not, proceed to step S3; S3: using a numerical analysis model to perform a verification calculation on the underground railway station to be evaluated, and determining the final seismic damage state prediction result based on the verification calculation result. This invention introduces a second-stage triggering mechanism based on prediction confidence and classification interval, further verifying and analyzing low-confidence samples and class boundary samples, reducing the risk of misjudgment and omission under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of seismic safety assessment of underground structures, and in particular to a two-stage intelligent prediction method for seismic damage to underground stations. Background Technology

[0002] Underground stations are a crucial component of urban rail transit systems and a key node in urban lifeline engineering. Under earthquake loads, underground stations may suffer damage such as cracking, yielding, localized damage, and overall degradation of their load-bearing capacity. The ability to quickly and accurately identify the damage status of underground stations after an earthquake directly impacts emergency response, traffic restoration, secondary disaster prevention, and urban operational safety. Therefore, establishing a method for predicting earthquake damage to underground stations suitable for rapid post-earthquake assessment scenarios has significant engineering application value.

[0003] Existing methods for assessing seismic damage to underground railway stations mainly include field investigation and numerical analysis. Field investigation methods typically rely on manual inspection and expert judgment, and are greatly affected by factors such as the site environment, traffic conditions, and operational safety, making it difficult to complete large-scale rapid identification within a short period after an earthquake. Numerical analysis methods, especially those based on refined dynamic analysis models, can accurately simulate the response process of underground railway stations under seismic loading and determine the damage status based on parameters such as inter-story drift angles, structural internal forces, or damage indices, exhibiting high reliability. However, these methods usually require the establishment of complex dynamic analysis models, resulting in large computational loads and long processing times, making it difficult to meet the needs of rapid, batch prediction after an earthquake.

[0004] In recent years, machine learning and deep learning methods have been gradually applied to the field of earthquake damage prediction for underground railway stations. By establishing a mapping relationship between input information and earthquake damage status, rapid prediction of earthquake damage status of underground railway stations can be achieved. However, existing methods still have shortcomings in terms of input information organization and the reliability of results under complex working conditions.

[0005] On the one hand, some methods use only a few ground motion intensity indices or their reduced-dimensional features as input, making it difficult to fully preserve complete ground motion information. Other methods, while using complete ground motion time histories or ground motion images as input, typically build models around a single input channel, failing to adequately consider important factors affecting the evolution of seismic damage in underground stations, such as site conditions. Related research indicates that using only partial ground motion indices limits prediction accuracy, while directly using complete time histories significantly increases model complexity and data requirements. On the other hand, existing single-stage deep learning recognition methods typically use the model output directly as the final recognition result. When the sample to be identified is near the category boundary, or when the actual working conditions exceed the coverage of the training samples, the model is prone to misjudgment. For post-earthquake assessment of underground stations, directly classifying low-reliability samples as safe may lead to the omission of high-risk conditions, affecting subsequent handling decisions.

[0006] Therefore, existing technologies suffer from two main problems: firstly, high-precision analysis methods are time-consuming and difficult to directly apply to rapid batch prediction after earthquakes; secondly, deep learning models suffer from insufficient organization of input information and inadequate reliability of results under complex conditions. How to construct a deep learning prediction model that can directly output earthquake damage status, while ensuring recognition efficiency and improving the reliability of judgments for low-confidence samples and samples under complex conditions, remains a pressing technical problem to be solved in this field. Summary of the Invention

[0007] This invention aims to address the problems of long analysis time for high precision and insufficient reliability of single-stage deep learning prediction results in existing underground railway station seismic damage prediction methods. This invention provides a two-stage intelligent seismic damage prediction method for underground railway stations based on deep learning. This method rapidly predicts the seismic damage state of underground railway stations through a first-stage deep learning prediction model and calculates a prediction reliability index based on the prediction probability corresponding to each seismic damage category. When the first-stage prediction result meets the preset reliability conditions, the prediction result is directly output; when the first-stage prediction result does not meet the preset reliability conditions, a second-stage analysis is triggered for further verification. This improves the reliability and engineering applicability of the final seismic damage state prediction result while ensuring prediction efficiency.

[0008] To achieve the above objectives, the technical solution of the present invention includes: A two-stage intelligent prediction method for seismic damage to underground railway stations includes: S1: The first-stage deep learning prediction model is used to predict the sample to be predicted, and the first-stage prediction results are obtained. The first-stage prediction results include the prediction probability and preliminary prediction results of the damage status corresponding to each damage category. S2: Calculate the prediction confidence and classification interval based on the prediction probability and the preliminary prediction results of the earthquake damage status, and determine whether the prediction results of the first stage meet the preset reliability conditions. If they do, the prediction results of the first stage are used as the final earthquake damage status prediction results. If they do not meet the conditions, proceed to step S3. S3: A numerical analysis model is used to verify the calculation of the underground station to be evaluated, and the final prediction result of the seismic damage status is determined based on the verification calculation results.

[0009] A further improvement of the present invention is that the first-stage deep learning prediction model is used to jointly predict multi-source heterogeneous input data to obtain the first-stage prediction result; the input data includes seismic motion data and static site-related parameters.

[0010] A further improvement of the present invention is that the first-stage deep learning prediction model includes a dynamic excitation feature extraction branch, a static site constraint encoding branch, and a multi-source heterogeneous feature coupling and fusion unit. The dynamic excitation feature extraction branch is configured to convert ground motion data into two-dimensional ground motion images and extract deep spatial feature vectors that characterize the distribution patterns of ground motion amplitude, spectrum and duration through a deep convolutional structure. The static site constraint encoding branch is configured to convert the static site-related parameters into numerical feature vectors, and then convert them into site feature vectors characterizing the site stiffness and spatial constraint effects through a fully connected mapping network structure. The multi-source heterogeneous feature coupling and fusion unit is configured to perform nonlinear interactive splicing and dimensionality reduction mapping between the deep spatial feature vector and the site feature vector to construct a hidden layer representation vector that characterizes the comprehensive nonlinear effect of soil-structure dynamic interaction. The classification output unit is configured to output the predicted probability corresponding to each damage category based on the hidden layer representation vector.

[0011] A further improvement of the present invention is that the process of acquiring the dataset used in the training of the first-stage deep learning prediction model includes: The spectral representation-random function method is used to generate random ground motion acceleration time histories belonging to the same probability complete set, ensuring that different ground motion samples have a consistent probability distribution, thus obtaining ground motion data; Set static site-related parameters, including shear wave velocity and burial depth; Several site conditions are formed based on the aforementioned static site-related parameters; For each site condition, a dynamic analysis model for the underground station is established. The dynamic analysis model for the underground station includes the geometric information, soil parameter information, material parameter information and boundary condition information of the underground station. The soil parameters and site conditions in the underground station dynamic analysis model are configured according to the site conditions, so that the underground station dynamic analysis model can characterize the dynamic response characteristics of the underground station under different site conditions. The ground motion data and static site-related parameters are input into the underground station dynamic analysis model to conduct nonlinear time history analysis and extract the seismic damage response index of the underground station. Each working condition sample is labeled according to the preset seismic damage criteria to form a seismic damage status label; The ground motion data, static site parameters, and corresponding seismic damage status labels are correlated to form a training sample set.

[0012] A further improvement of the present invention is that the seismic damage response index includes one or more of the following: inter-story drift angle, structural internal force, and component damage index.

[0013] A further improvement of the present invention is that the first-stage deep learning prediction model employs a loss function during the training process, wherein the loss function is the sparse classification cross-entropy loss function.

[0014] A further improvement of the present invention is that the process of determining whether the prediction result of the first stage meets the preset reliability conditions includes: S21: Take the maximum value among the predicted probabilities of each damage level output by the first-stage deep learning prediction model as the prediction confidence level. C : S22: Transform the predicted probabilities into a predicted probability vector, and use the difference between the largest and second largest components as the classification margin. ; S23: Predict confidence level Compared with the preset prediction confidence threshold Compare and divide the intervals With respect to the preset classification interval threshold The results are compared to determine whether to proceed to the second stage of analysis; when the prediction confidence level... Not less than the preset prediction confidence threshold And the classification interval Not less than the preset classification interval threshold At that time, it is determined that the current first-stage prediction result meets the preset reliability conditions.

[0015] A further improvement of the present invention is that the prediction confidence threshold T c The classification interval threshold is set to 0.80. T d Set to 0.20.

[0016] A further improvement of the present invention is that, in step S2, if the earthquake damage category of the first-stage prediction result is a high-risk earthquake damage category, the first-stage prediction result is directly determined to be inconsistent with the preset reliability conditions, and the process proceeds to step S3.

[0017] A further improvement of the present invention is that, in step S3, the numerical analysis model is a dynamic analysis model of an underground station.

[0018] Compared with existing technologies, this method jointly inputs seismic motion image information and static site-related parameters in the first stage, enabling it to simultaneously characterize the combined impact of seismic motion characteristics and site conditions on the seismic damage evolution of underground stations, thereby improving the accuracy of seismic damage prediction. This invention introduces a second-stage triggering mechanism based on prediction confidence and classification interval, further verifying and analyzing low-confidence samples and class boundary samples, reducing the risk of misjudgment and omission under complex conditions, and improving the stability and reliability of the final prediction results. Compared to methods relying solely on refined numerical analysis, this invention significantly improves the efficiency of rapid post-earthquake assessment; compared to single-stage deep learning prediction methods, this invention balances prediction efficiency and result reliability, making it more suitable for rapid post-earthquake assessment and emergency decision-making scenarios for underground stations. Attached Figure Description

[0019] Figure 1 This is a flowchart of the two-stage intelligent prediction method for seismic damage to underground stations according to the present invention; Figure 2 This is a schematic diagram of the first stage of the deep learning prediction model. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0021] like Figure 1 As shown, an embodiment of the present invention provides a two-stage intelligent prediction method for seismic damage to underground railway stations, which includes the following steps: S1: The first-stage deep learning prediction model is used to predict the sample to be predicted, and the first-stage prediction results are obtained. The first-stage prediction results include the prediction probability and preliminary prediction results of the damage status corresponding to each damage category. S2: Calculate the prediction confidence and classification interval based on the prediction probability and the preliminary prediction results of the earthquake damage status, and determine whether the prediction results of the first stage meet the preset reliability conditions. If they do, the prediction results of the first stage are used as the final earthquake damage status prediction results. If they do not meet the conditions, proceed to step S3. S3: A numerical analysis model is used to verify the calculation of the underground station to be evaluated, and the final prediction result of the seismic damage status is determined based on the verification calculation results.

[0022] like Figure 2As shown, the first-stage deep learning prediction model is used to jointly predict multi-source heterogeneous input data to obtain the first-stage prediction results. The input data of the first-stage deep learning prediction model includes ground motion data and static site-related parameters; the ground motion data is in the form of ground motion acceleration time history data; the static site-related parameters include the shear wave velocity and burial depth around the underground station to be predicted.

[0023] To jointly process multi-source heterogeneous input data, the first-stage deep learning prediction model includes a dynamic excitation feature extraction branch, a static site constraint encoding branch, a multi-source heterogeneous feature coupling and fusion unit, and an output unit.

[0024] The dynamic excitation feature extraction branch is configured to convert ground motion data into two-dimensional ground motion images and extract deep spatial feature vectors that characterize the distribution patterns of ground motion amplitude, spectrum and duration through a deep convolutional structure.

[0025] The static site constraint encoding branch is configured to convert the static site-related parameters into numerical feature vectors, and then into site feature vectors characterizing the site stiffness and spatial constraint effects through a fully connected mapping network structure.

[0026] The multi-source heterogeneous feature coupling and fusion unit is configured to perform nonlinear interactive splicing and dimensionality reduction mapping between the deep spatial feature vector and the site feature vector to construct a hidden layer representation vector that characterizes the comprehensive nonlinear effect of soil-structure dynamic interaction.

[0027] The classification output unit is configured to output the predicted probability corresponding to each damage category based on the hidden layer representation vector.

[0028] In one specific embodiment, the deep convolutional structure in the dynamic excitation feature extraction branch can use VGG16 as the image processing backbone network. In another specific embodiment, the kernel size of the convolutional layer can be 3×3, the stride can be 1, and the window size of the pooling layer can be 2×2, the stride can be 2. The fully connected mapping network structure of the static site constraint encoding branch is a multilayer perceptron network, which includes eight layers from input to output: an input layer, a first fully connected layer, a first activation layer, a second fully connected layer, a second activation layer, and an output layer. In one specific embodiment, the output dimension of the first fully connected layer can be set to 32, the output dimension of the second fully connected layer can be set to 16, and the activation function can be the ReLU function. In this embodiment, when the seismic damage status is a multi-classification task, the Softmax activation function can be used to output the probability distribution of each seismic damage category.

[0029] The process of acquiring the dataset used in the training of the first-stage deep learning prediction model includes: The spectral representation-random function method is used to generate random ground motion acceleration time histories belonging to the same probability complete set, ensuring that different ground motion samples have a consistent probability distribution, thus obtaining ground motion data. In a specific embodiment, this step uses the spectral representation-random function method to generate random ground motion acceleration time histories with peak ground accelerations of 0.1g, 0.2g, 0.3g, 0.4g, 0.6g, and 0.9g, to ensure the consistency of the generated ground motion samples in terms of statistical characteristics and to provide representative input data for subsequent model training.

[0030] Static site-related parameters are set, including shear wave velocity and burial depth; in one specific embodiment, the static site-related parameters include the shear wave velocity of a single soil mass. V s and the burial depth of underground stations H The shear wave velocity of a single soil mass can be taken as 80 m / s, 160 m / s, 240 m / s, 320 m / s, 400 m / s, or 500 m / s, and the burial depth of the underground station can be set as 1 m, 3 m, 5 m, 7 m, or 10 m.

[0031] Different site conditions are formed based on the aforementioned static site parameters. In one specific embodiment, multiple sets of site conditions are formed by combining different shear wave velocities and different burial depths, resulting in a total of 30 sets of site conditions. Each set of site conditions corresponds to a specific set of soil parameters and burial depth conditions, used to characterize the dynamic response characteristics of underground stations under different site environments.

[0032] For different site conditions, a dynamic analysis model for the underground station is established. The dynamic analysis model for the underground station includes at least the geometric information, soil parameter information, material parameter information, and boundary condition information of the underground station. In this embodiment, the dynamic analysis model for the underground station is a finite element dynamic analysis model, which is used to simulate the nonlinear dynamic response process of the underground station under different seismic motions and different site conditions.

[0033] The soil parameters and site conditions in the underground station dynamic analysis model are configured according to different site conditions, so that the underground station dynamic analysis model can characterize the dynamic response characteristics of the underground station under different site conditions.

[0034] The ground motion data and static site-related parameters are input into the underground station dynamic analysis model to conduct nonlinear time history analysis and extract the seismic damage response index of the underground station. The seismic damage response index includes one or more of the following: inter-story drift angle, structural internal force, and component damage index. In this embodiment, the maximum inter-story drift angle is selected as the main seismic damage judgment index.

[0035] Each working condition sample is labeled according to a preset seismic damage criterion to form a seismic damage status label. In one specific embodiment, the seismic damage level of each sample is classified according to the extracted maximum inter-story drift angle and the preset seismic damage criterion. The seismic damage level classification is shown in Table 1.

[0036] Table 1 Classification of Earthquake Damage Levels for Underground Stations The ground motion data, static site parameters, and corresponding seismic damage status labels are correlated to form a training sample set.

[0037] After obtaining the training sample set, the training set, validation set, and test set are divided according to a preset ratio; the first-stage deep learning prediction model is trained; the first-stage deep learning prediction model is trained based on the training set, and the training is monitored and parameters are adjusted based on the validation set; the performance of the first-stage deep learning prediction model after training is evaluated based on the test set.

[0038] In the training process of the first-stage deep learning prediction model, in each round, the samples in the training set are input into the first-stage deep learning prediction model to obtain the prediction results. The prediction results include the prediction probability corresponding to each earthquake damage category. The error between the prediction results and the sample labels is calculated through the loss function module. The parameters of the first-stage deep learning prediction model are updated using the Adam optimizer, and the learning rate is adaptively adjusted. The training process is monitored based on the validation set, and the first-stage deep learning prediction model after training is determined.

[0039] In one specific embodiment, the loss function used in the training process of the first-stage deep learning prediction model is the sparse classification cross-entropy loss function. The initial learning rate can be set to 0.0001, and the learning rate can be dynamically adjusted using the Reduce LR On Plateau mechanism, while overfitting is prevented through the Early Stopping mechanism.

[0040] In actual prediction, the ground motion data of the sample to be predicted comes from the measured ground motion data and the actual static site parameters of the station to be predicted.

[0041] In step S2, the process of determining whether the prediction results of the first stage meet the preset reliability conditions includes: S21: Take the maximum value among the predicted probabilities of each damage level output by the first-stage deep learning prediction model as the prediction confidence level. C : in, p 1 , p 2 ,p n These represent the predicted probabilities for the corresponding earthquake damage levels. The confidence levels are calculated. C This can determine the degree of confidence the model has in predicting the current sample.

[0042] S22: Transform the predicted probabilities into a predicted probability vector, and use the difference between the largest and second largest components as the classification margin. Its expression is: in: It is the largest component in the prediction probability vector. It is the second largest component in the prediction probability vector.

[0043] S23: Predict confidence level Compared with the preset prediction confidence threshold Compare and divide the intervals With respect to the preset classification interval threshold The results are compared to determine whether to proceed to the second stage of analysis. The criterion for this judgment is: when the prediction confidence level... Not less than the preset prediction confidence threshold And the classification interval Not less than the preset classification interval threshold If the current prediction result is deemed to have high reliability, then for samples that meet the preset reliability conditions, the second stage of analysis is skipped, and the prediction result of the first stage deep learning prediction model is directly used as the final earthquake damage state prediction result.

[0044] In one specific embodiment, the prediction confidence threshold T c The classification interval threshold is set to 0.80. T d The threshold is set to 0.20. During implementation, the threshold can be adjusted according to the training sample size, the method of classifying earthquake damage levels, the distribution characteristics of model output, and the needs of engineering applications, and is not limited to the specific value mentioned above.

[0045] In some implementations, when the damage category of the first-stage prediction result belongs to a preset high-risk damage category, the second-stage analysis is still triggered even if the prediction confidence and classification interval meet preset reliability conditions, in order to improve the reliability of the high-risk sample determination results. The first-stage prediction result includes the prediction probability corresponding to each damage category, and the damage category of the first-stage prediction result is the damage category with the highest prediction probability. In this embodiment, the damage categories corresponding to labels 3 and 4 in Table 1 are high-risk damage categories.

[0046] After the second stage analysis is completed, the final earthquake damage state prediction result is output. In some implementations, if an unsolvable anomaly occurs during the second stage analysis, the first stage prediction result can be recalled by setting a fault tolerance mechanism or a rollback mechanism.

[0047] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A two-stage intelligent prediction method for seismic damage to underground railway stations, characterized in that, include: S1: The first-stage deep learning prediction model is used to predict the sample to be predicted, and the first-stage prediction results are obtained. The first-stage prediction results include the prediction probability and preliminary prediction results of the damage status corresponding to each damage category. S2: Calculate the prediction confidence and classification interval based on the prediction probability and the preliminary prediction results of the earthquake damage status, and determine whether the prediction results of the first stage meet the preset reliability conditions. If they do, the prediction results of the first stage are used as the final earthquake damage status prediction results. If they do not meet the conditions, proceed to step S3. S3: A numerical analysis model is used to verify the calculation of the underground station to be evaluated, and the final prediction result of the seismic damage status is determined based on the verification calculation results; The first-stage deep learning prediction model includes a dynamic excitation feature extraction branch, a static site constraint encoding branch, a multi-source heterogeneous feature coupling and fusion unit, and a classification output unit; The dynamic excitation feature extraction branch is configured to convert ground motion data into two-dimensional ground motion images and extract deep spatial feature vectors that characterize the distribution patterns of ground motion amplitude, spectrum and duration through a deep convolutional structure. The static site constraint encoding branch is configured to convert the static site-related parameters into numerical feature vectors, and then convert them into site feature vectors characterizing the site stiffness and spatial constraint effects through a fully connected mapping network structure. The multi-source heterogeneous feature coupling and fusion unit is configured to perform nonlinear interactive splicing and dimensionality reduction mapping between the deep spatial feature vector and the site feature vector to construct a hidden layer representation vector that characterizes the comprehensive nonlinear effect of soil-structure dynamic interaction. The classification output unit is configured to output the predicted probability corresponding to each damage category based on the hidden layer representation vector. The process of acquiring the dataset used in the training of the first-stage deep learning prediction model includes: The spectral representation-random function method is used to generate random ground motion acceleration time histories belonging to the same probability complete set, ensuring that different ground motion samples have a consistent probability distribution, thus obtaining ground motion data; Set static site-related parameters, including shear wave velocity and burial depth; Several site conditions are formed based on the aforementioned static site-related parameters; For each site condition, a dynamic analysis model for the underground station is established. The dynamic analysis model for the underground station includes the geometric information, soil parameter information, material parameter information and boundary condition information of the underground station. The soil parameters and site conditions in the underground station dynamic analysis model are configured according to the site conditions, so that the underground station dynamic analysis model can characterize the dynamic response characteristics of the underground station under different site conditions. The ground motion data and static site-related parameters are input into the underground station dynamic analysis model to conduct nonlinear time history analysis and extract the seismic damage response index of the underground station. Each working condition sample is labeled according to the preset seismic damage criteria to form a seismic damage status label; The ground motion data, static site parameters, and corresponding seismic damage status labels are correlated to form a training sample set.

2. The intelligent prediction method for seismic damage of a two-stage underground station according to claim 1, characterized in that, The first-stage deep learning prediction model is used to jointly predict multi-source heterogeneous input data to obtain the first-stage prediction result; the input data includes seismic motion data and static site-related parameters.

3. The intelligent prediction method for seismic damage of a two-stage underground station according to claim 1, characterized in that, The seismic damage response indicators include one or more of the following: inter-story drift angle, structural internal force, and component damage indicators.

4. The intelligent prediction method for seismic damage of a two-stage underground station according to claim 1, characterized in that, The first-stage deep learning prediction model uses a loss function during training, which is the sparse classification cross-entropy loss function.

5. The intelligent prediction method for seismic damage of a two-stage underground station according to claim 1, characterized in that, The process of determining whether the prediction results of the first stage meet the preset reliability conditions includes: S21: Take the maximum value among the predicted probabilities of each damage level output by the first-stage deep learning prediction model as the prediction confidence level. C : S22: Transform the predicted probabilities into a predicted probability vector, and use the difference between the largest and second largest components as the classification margin. ; S23: Predict confidence level Compared with the preset prediction confidence threshold Compare and divide the intervals With respect to the preset classification interval threshold The results are compared to determine whether to proceed to the second stage of analysis; when the prediction confidence level... Not less than the preset prediction confidence threshold And the classification interval Not less than the preset classification interval threshold At that time, it is determined that the current first-stage prediction result meets the preset reliability conditions.

6. The intelligent prediction method for seismic damage of a two-stage underground station according to claim 5, characterized in that, The prediction confidence threshold T c The classification interval threshold is set to 0.

80. T d Set to 0.

20.

7. The intelligent prediction method for seismic damage of a two-stage underground station according to claim 1, characterized in that, In step S2, if the earthquake damage category of the first-stage prediction result is a high-risk earthquake damage category, it is directly determined that the first-stage prediction result does not meet the preset reliability conditions, and the process proceeds to step S3.

8. The intelligent prediction method for seismic damage of a two-stage underground station according to claim 1, characterized in that, In step S3, the numerical analysis model is a dynamic analysis model of an underground station.

Images