A regional rapid earthquake damage prediction method based on machine learning and model update

Abstract

A regional rapid earthquake damage prediction method based on machine learning and model updating, which relates to the technical field of structural seismic safety assessment. Aiming at the problem that it is difficult to identify internal damage of structures in the existing technology, resulting in low accuracy of post-earthquake damage assessment, compared with the existing damage identification method based on remote sensing images, the damage discrimination of this application is more refined. By calculating the maximum inter-story response of each floor through the updated structural numerical model, the damage state (intact, slightly damaged, moderately damaged, severely damaged, destroyed) can be accurately discriminated, rather than simply judging whether the structure collapses, thereby improving the accuracy of post-earthquake damage assessment.

Description

Technical Field

[0001] The present invention relates to the technical field of structural seismic safety assessment, and specifically to a method for rapid regional earthquake damage prediction based on machine learning and model updating. Background Art

[0002] In the past few decades, earthquakes have had a significant impact on people's lives and property. Accurate and timely earthquake damage assessment of regional building groups is conducive to identifying vulnerable parts in cities before earthquakes and preparing emergency rescue supplies. Therefore, it is crucial to conduct accurate and timely pre-earthquake damage prediction and post-earthquake damage assessment for regional building groups. However, due to the time-consuming and laborious acquisition of detailed information of building groups at the regional scale, the establishment of finite element models, and the structural nonlinear time history analysis, how to conduct rapid and accurate regional earthquake damage assessment with limited information remains a major problem in the field of disaster prevention and mitigation. Existing research mainly focuses on improving accuracy and speed, but there are still some deficiencies.

[0003] For pre-earthquake damage prediction of regions, the current main methods include vulnerability analysis methods, capacity demand spectrum methods, time history analysis methods, and methods combined with artificial intelligence. The vulnerability analysis method has low accuracy because it only uses a few ground motion indicators (usually no more than three) to describe ground motion. The capacity demand spectrum method is difficult to consider the time domain characteristics of ground motion. Although the time history analysis method is accurate, it is time-consuming when calculating high-fidelity finite element models. To address this issue, existing research mainly solves the problems of accuracy and speed from the following two ideas: (1) simplification of the finite element model, and (2) training machine learning or neural network surrogate models for rapid calculation. For idea (1), the multi-degree-of-freedom lumped mass (MDOF) model is a commonly used simplified model. The time history analysis method based on multiple degrees of freedom simulates the deformation characteristics of each floor through shear springs or bending springs, greatly reducing the degrees of freedom of the structure. This method takes into account both computational efficiency and computational accuracy, but there are problems with accurately calibrating parameters such as structural stiffness, strength, and energy dissipation. For idea (2), using machine learning or neural networks to map the complex relationships between structures, ground motions, and structural responses is a commonly used method. Existing related research uses MDOF models or high-fidelity finite element models as structural samples. When using the MDOF model as a structural sample, there will also be problems with accurately calibrating parameters such as stiffness, strength, and energy dissipation. When using high-fidelity finite element models as structural samples, a large number of finite element models need to be established to consider different types and numbers of floors of structures. Due to the time-consuming and laborious aspects of obtaining structural detailed information, establishing finite element models, and calculating structural responses, it is difficult to use high-fidelity finite element models as structural samples at the regional scale.

[0004] For the post-earthquake damage assessment of regions, existing methods usually rely on damage identification methods based on remote sensing images to identify whether a structure has collapsed. However, this method is difficult to identify internal damage to the structure, resulting in a low discrimination accuracy rate. Summary of the Invention

[0005] The object of the present invention is to propose a rapid regional earthquake damage prediction method based on machine learning and model updating to address the problem that existing technologies have difficulty in identifying internal damage to structures, which leads to a low accuracy rate in post-earthquake damage assessment.

[0006] The technical solution adopted by the present invention to solve the above technical problems is as follows:

[0007] A rapid regional earthquake damage prediction method based on machine learning and model updating, comprising the following steps:

[0008] Step 1: Establish a structural sample through a simplified multi-degree-of-freedom model, and then obtain the structural parameters of the structural sample, the maximum inter-story drift angle, and the ground motion parameters of the earthquake suffered by the structural sample. The structural parameters include easily obtainable structural parameters and non-easily obtainable structural parameters;

[0009] Step 2: Use the structural parameters and ground motion parameters as inputs and the maximum inter-story drift angle as the output to train a structural response prediction model;

[0010] Step 3: For the building to be evaluated, obtain the easily obtainable structural parameters of the building to be evaluated and the monitored value of the maximum inter-story drift angle. Then, obtain the ground motion parameters of the earthquake suffered by the building to be evaluated, and set the initial value of the non-easily obtainable structural parameters. Then, input the ground motion parameters of the earthquake suffered by the building to be evaluated, the easily obtainable structural parameters, and the initial value of the non-easily obtainable structural parameters into the trained structural response prediction model to obtain the predicted value of the maximum inter-story drift angle as the output. Use the monitored value of the maximum inter-story drift angle and the predicted value of the maximum inter-story drift angle to obtain an objective function, and use the non-easily obtainable structural parameters as the variables to be optimized. Use the particle swarm optimization algorithm to iteratively solve and obtain the optimal non-easily obtainable structural parameters;

[0011] Step 4: For the target earthquake scenario, obtain the corresponding ground motion parameters under this scenario, and input the ground motion parameters, the easily obtainable structural parameters obtained in Step 3, and the optimal non-easily obtainable structural parameters into the trained structural response prediction model to obtain the maximum inter-story drift angle as the output, and obtain the earthquake damage situation under the target earthquake scenario based on the output maximum inter-story drift angle.

[0012] Further, the structural response prediction model is an XGBoost model.

[0013] Further, the loss function of the XGBoost model is the MSE loss function, expressed as:

[0014]

[0015] Among them, m i represents the number of layers of the i-th sample, n represents the total number of samples, and MIDR ij represents the maximum inter-story drift angle label value of the j-th layer of the i-th sample, represents the predicted value of the maximum inter-story drift angle of the j-th layer of the i-th sample.

[0016] Furthermore, the specific steps of the first step are as follows:

[0017] Establish a structural sample through a simplified multi-degree-of-freedom model, then obtain the structural parameters of the structural sample, and use the structural parameters to establish a finite element model. Then, obtain the ground motion parameters of the earthquake suffered by the structural sample, and input the ground motion parameters into the finite element model to obtain the output maximum inter-story drift angle.

[0018] Furthermore, the objective function is expressed as:

[0019]

[0020] Among them, m represents the number of layers of the structure to be updated, and MIDR monitor,i represents the monitored value of the maximum inter-story drift angle of the i-th layer,

[0021] MIDR predict,i represents the predicted value of the maximum inter-story drift angle of the i-th layer.

[0022] Furthermore, the trained structural response prediction model is determined by the mean absolute error MAE, the mean relative error MAPE, the goodness of fit R 2 and the kernel density estimation KDE plot;

[0023] The mean absolute error MAE is expressed as:

[0024]

[0025] The mean relative error MAPE is expressed as:

[0026]

[0027] The goodness of fit R2 is expressed as:

[0028]

[0029] Among them, m i represents the number of layers of the i-th sample, n represents the total number of samples, and MIDR ij represents the maximum inter-story drift angle label value of the j-th layer of the i-th sample, Denotes the predicted maximum inter-story drift angle of the j-th layer of the i-th sample.

[0030] Furthermore, the easily obtained parameters of the structure are the number of stories and the story height.

[0031] Furthermore, the not easily obtained parameters of the structure are the period calculation coefficient γ1, the stiffness reduction coefficient α K and the hysteretic parameter τ.

[0032] Furthermore, the ground motion parameters include PGD, PGA, PGD, PGV / PGA, I a , HI, ASI, VSI, EPA, SED, CAV, A rms , V rms and Sa(ξ = 5%, T = 0.1s) to Sa(ξ = 5%, T = 3.0s).

[0033] Furthermore, the specific steps of Step 2 are as follows:

[0034] Using the structural parameters and ground motion parameters as inputs and the maximum inter-story drift angle as the output, training data is constructed, and the training data is divided into a training set, a validation set, and a test set in a ratio of 0.64:0.17:0.19. Then, the training of the structural response prediction model is completed using the training set, the validation set, and the test set.

[0035] The beneficial effects of the present invention are:

[0036] Compared with the existing damage identification method based on remote sensing images, the damage discrimination of this application is more refined. By calculating the maximum inter-story response of each layer using the updated structural numerical model, the damage state (intact, slightly damaged, moderately damaged, severely damaged, destroyed) can be accurately discriminated, rather than simply judging whether the structure has collapsed, thereby improving the accuracy of post-earthquake damage assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Is the overall flowchart of this application;

[0038] Figure 2 Is the schematic diagram of sample generation;

[0039] Figure 3 Is the magnitude and frequency distribution of the selected ground motion 1;

[0040] Figure 4 Is the magnitude and frequency distribution of the selected ground motion 2;

[0041] Figure 5 Is the logical diagram of the parameter composition of the structural sample;

[0042] Figure 6 Is the division of the sample;

[0043] Figure 7 It is a schematic diagram of the XGBoost model;

[0044] Figure 8 It is a kernel density estimation diagram of the maximum inter-story drift angle of each layer of the test set;

[0045] Figure 9 It is a schematic diagram of the regional structure model update;

[0046] Figure 10 It is a comparison of different damage states of the structure before and after the model update in Case 1;

[0047] Figure 11 It is a comparison of different damage states of the structure before and after the model update in Case 2. Specific implementation manner

[0048] It should be particularly noted that, without conflict, the various implementation manners disclosed in this application can be combined with each other.

[0049] Specific implementation manner one: Refer to Figure 1 This implementation manner is specifically described as follows. A regional rapid earthquake damage prediction method based on machine learning and model update described in this implementation manner includes the following steps:

[0050] Step 1: Establish a structural sample through a simplified multi-degree-of-freedom model (MDOF), and then obtain the structural parameters, maximum inter-story drift angle of the structural sample, and ground motion parameters of the earthquake suffered by the structural sample. The structural parameters include easily obtained structural parameters and non-easily obtained structural parameters;

[0051] Step 2: Use the structural parameters and ground motion parameters as inputs and the maximum inter-story drift angle as the output to train a structural response prediction model;

[0052] Step 3: For the building to be evaluated, obtain the easily obtained structural parameters and the monitored value of the maximum inter-story drift angle of the building to be evaluated. Then, obtain the ground motion parameters of the earthquake suffered by the building to be evaluated, and set the initial value of the non-easily obtained structural parameters. Then, input the ground motion parameters, easily obtained structural parameters, and the initial value of the non-easily obtained structural parameters of the earthquake suffered by the building to be evaluated into the trained structural response prediction model to obtain the predicted value of the maximum inter-story drift angle as the output. Use the monitored value of the maximum inter-story drift angle and the predicted value of the maximum inter-story drift angle to obtain an objective function, and use the non-easily obtained structural parameters as variables to be optimized. Use the particle swarm optimization algorithm to iteratively solve to obtain the optimal non-easily obtained structural parameters;

[0053] Step 4: For the target earthquake scenario, obtain the corresponding ground motion parameters under this scenario, and input the ground motion parameters, the easily obtained structure parameters obtained in Step 3, and the optimal non-easily obtained structure parameters into the trained structure response prediction model to obtain the output maximum inter-story drift ratio, and obtain the earthquake damage situation under the target earthquake scenario based on the output maximum inter-story drift ratio.

[0054] Model updating technology can reverse-infer the unknown parameters of a structure through structural responses such as natural vibration frequency, vibration mode, structural acceleration, velocity, and displacement time history. This technology provides a feasible way to accurately obtain the model of regional structures. However, since it is very time-consuming to perform model updating using time history analysis, it is difficult to achieve rapid regional structure model updating and cannot meet the requirements of rapid regional earthquake damage assessment. Existing research shows that using a surrogate model to map the relationship between the structure, ground motion, and structural response can quickly calculate the structural response. Therefore, if the surrogate model and regional model updating can be combined, it can meet the accurate and rapid earthquake damage assessment of regional structures.

[0055] This application aims to solve the problem that it is difficult for traditional regional earthquake damage assessment methods to achieve rapid and accurate assessment of a large number of buildings in a region. A method combining machine learning and model updating is proposed for accurate and rapid calculation of structural responses in a region, and the structural responses include the maximum inter-story drift ratio of each floor of the structure. The specific structural type in the present invention is a concrete frame structure, but the idea framework can be used for any other structural types, such as steel frame structures, masonry structures, shear wall structures, wood structures, etc.

[0056] This application is more accurate compared with existing vulnerability analysis methods and capacity demand spectrum methods. Time history analysis is used to obtain structural response samples, and 43 ground motion indicators are used to describe ground motion information more accurately.

[0057] This application has a significant efficiency advantage compared with existing multi-degree-of-freedom structural time history analysis methods. The XGBoost model is used to establish the mapping relationship between the structure, ground motion, and structural response, which greatly accelerates the calculation speed of structural responses.

[0058] This application has a significant efficiency advantage compared with existing time history analysis-based model updating methods. Existing time history analysis-based model updating methods are difficult to quickly update the models of large-scale regional structures. The present invention uses the XGBoost model to map the nonlinear relationship between the structure, ground motion, and structural response, and uses the XGBoost model to calculate the structural response in each iteration of model updating instead of time history analysis, thereby achieving rapid updating of large-scale regional structures.

[0059] The overall flowchart of this application is as Figure 1 shown. The specific implementation steps are as follows:

[0060] Step 1: The main content of this step is the generation of samples, and its schematic diagram is as Figure 2 shown. Specifically, it includes the selection of ground motions and the choice of parameter IMs, the generation of structural samples, and the calculation of structural responses. The detailed content of Step 1 is as follows:

[0061] Step 1.1: Select two groups of ground motions, which are used for the training of the structural response prediction model XGBoost model and the model update of regional structures respectively. The first group of ground motions includes 3022 ground motions from 132 earthquake events, and the second group of ground motions includes 70 days of ground shaking from 7 earthquake events. To ensure the reliability of the present invention, the above two groups of ground motions come from different earthquake events. The magnitude and frequency distribution of the two selected groups of ground motions are as Figure 3 shown. In addition, 43 ground motion parameters are selected according to existing research, as shown in Table 1. For the convenience of subsequent model training, the ground motions are downsampled at a sampling frequency of 50 Hz, and 3000 points are collected for each ground motion. Since large ground motions may cause damage to structures, which is exactly what the research concerns. Therefore, the selected ground motions are amplitude-modulated, and the amplitude modulation coefficient is 4. If the PGA after amplitude modulation exceeds 2.0 g, then this ground motion will not be amplitude-modulated.

[0062] Step 1.2: Determine the fortification intensity DI, site category, and design earthquake grouping of the target area. Select the multi-degree-of-freedom shear model (MDOF) as the simplified model of the structure in the target area. Select the bilinear backbone curve and the single-parameter (τ) hysteretic model to simulate the inter-story force-displacement relationship of the MDOF model, and select Rayleigh damping as the damping model of the structure. Assume that the mass, stiffness, stiffness reduction coefficient, and hysteretic parameters of each layer are equal. Assume that the mass of each layer is unit 1. The parameters that determine the seismic performance of the structure are summarized into the following two categories: The first category is the easily obtained parameters ESP, such as the number of stories and story height of the structure; the second category is the non-easily obtained parameters, such as the period calculation coefficient γ1 and the stiffness reduction coefficient α K 、hysteretic parameter τ. Determine the ranges of the easily obtained parameters ESP and the non-easily obtained parameters DSP according to existing experience, as shown in Table 2 below. Then, 6500 combinations of the easily obtained parameters ESP and the non-easily obtained parameters DSP of the structural samples are obtained through Latin hypercube sampling.

[0063] For each parameter combination of the structural samples, the initial stiffness k0 in the bilinear backbone curve is calculated through formula (1), where the first-order mode vector Φ1 is obtained through modal analysis. I is the relative mass matrix, as shown in formula (2) specifically. A is the relative stiffness matrix, as shown in formula (3) specifically. The design strength F of the i-th layer is obtained through the bottom shear method di , and the yield strength F of the structure is obtained through the product of the design strength F di and the yield ratio Ω y (taking 1.1).yi Then, a corresponding degree-of-freedom shear model is established by using the finite element software OpenSees.

[0064]

[0065]

[0066] Step 1.3: Combine the 3022 ground motions in the ground motion dataset 1 with 6500 structural combinations to obtain 19,643,000 groups of samples. Perform time history analysis on the above samples to obtain 19,643,000 groups of structural responses. The damage degree of the structure is roughly judged by the maximum inter-story drift ratio (MIDR) of each floor. Therefore, the maximum inter-story drift ratio of each floor obtained by nonlinear time history analysis is selected as the output value of the future structural response prediction model.

[0067] Step 2: Select 5 structural parameters (number of stories, story height (ESP), period calculation coefficient γ1, stiffness reduction coefficient α K , hysteretic parameter τ (DSP)) and 43 ground motion parameters (as shown in Table 1) as the input of the XGBoost model. The schematic diagram of the XGBoost model is as Figure 7 shown. Therefore, the input of the XGBoost algorithm contains 48 parameters, and the input shape of each sample is [1, 48]. Select the inter-story drift ratio of each floor of the structure as the output. For the convenience of training, for the structures with less than 10 stories, their structural responses are filled with zeros to make the output shape of each sample [1, 10]. The zero-filled part is removed when actually calculating the loss function.

[0068] The loss function of the model adopts the MSE loss function, as shown in formula (4) specifically.

[0069]

[0070] Divide the obtained 19,643,000 groups of structural responses into a training set, a validation set, and a test set at a ratio of 0.64:0.17:0.19. The schematic diagram of the data division is as Figure 6As shown in the figure. The hyperparameters of the XGBoost model include Estimator numbers, Maximum depth of a tree, the subsample ratio of columns, Minimum sum of instance weight, gamma, learning rate. Their approximate ranges are determined through existing research and preliminary tests, as shown in Table 3 specifically. Then, the optimal hyperparameter combination is obtained through the Bayesian optimization algorithm (with 20 initial random iterations and 100 optimization steps). During the hyperparameter optimization process, the MSE of the validation set is used as the loss function, as shown in Equation (5). The optimal hyperparameter combination is shown in Table 5.

[0071]

[0072] Finally, the performance of the model is comprehensively evaluated through the mean absolute error MAE (see Equation (6)), mean relative error MAPE (see Equation (7)), goodness of fit R 2 (see Equation (8)), kernel density estimation KDE (see Figure 8 ).

[0073]

[0074] Step 3: First, obtain 43 ground motion parameters IMs and 2 easily obtained structure parameters ESP at the structure to be updated, and assume the not easily obtained structure parameters DSP of the structure. Then, input the ground motion parameters, easily obtained structure parameters, and the assumed not easily obtained parameters into the trained structure response prediction model to obtain the predicted value of the structure response and calculate the objective function. The objective function is shown in Equation (9). When the number of iterations does not meet the termination condition, update the not easily obtained structure parameters. Otherwise, terminate the update and output the updated not easily obtained structure parameters. First, obtain 43 ground motion parameters IMs and 2 easily obtained structure parameters ESP at the structure to be updated, and assume 3 not easily obtained structure parameters DSP of the structure. Then, input the ground motion parameters, easily obtained structure parameters, and the assumed not easily obtained parameters into the trained structure response prediction model to obtain the predicted value of the structure response and calculate the objective function. Then, use the particle swarm optimization algorithm to iteratively solve the not easily obtained structure parameters DSP. When the number of iterations meets the termination condition, stop the iteration and output the updated not easily obtained structure parameters DSP. Calculate 43 parameters IMs of the ground motion that may occur. Finally, input the 43 parameters IMs of the ground motion, 2 easily obtained structure parameters ESP, and the updated not easily obtained structure parameters DSP into the trained structure response prediction model to calculate the maximum inter-story drift angle of each layer of the structure and determine whether the structure is damaged.

[0075]

[0076] Then, the verification of the regional structure model update was carried out through two cases. Specifically as follows:

[0077] Case 1: Select a region in southwestern China with a fortification intensity of 8 degrees, site category II, and the second seismic design group. Assume that there are 36 concrete frame structures in this region, and their corresponding MDOF models are used as the actual monitored structures. The studied region is relatively small, and the attenuation of ground motion is not considered temporarily. The number of stories of the 36 selected monitored structures ranges from 1 to 10, with story heights of 3.3m, 3.6m, and 3.9m. The period calculation coefficient γ1 is set to 0.15, the stiffness reduction coefficient α k is set to 0.05, and the hysteretic parameter τ is set to 0.8. 70 ground motions are selected as the input. To verify the effectiveness of the model update, the ground motions are divided into a training set and a test set at a ratio of 2:8. Here, the ground motion training set represents the ground motions that have occurred and been recorded, and the test set represents the ground motions that may occur in the future. When using the particle swarm optimization algorithm for model update, 100 groups of initial DSP parameter values are selected, and 100 particle swarm optimization processes are executed. Finally, the optimal group of DSP parameters is selected, and the overall update time is 11.00 minutes. To show the effect of the model update, the DSP parameters of the structure before and after the update are input into the trained XGBoost model to obtain the maximum inter-story drift angle of each story, and then the corresponding damage state is obtained through the HAZUS report. The comparison of the proportion of different structural damage states before and after the update is as Figure 10 shown, and the specific data comparison is shown in Table 5. It can be seen that the errors of the proportion of different structural damage states after the update are all within 5%, indicating that this method has high accuracy.

[0078] Case 2: Similarly, select a region in southwestern China with a fortification intensity of 8 degrees, site category II, and the second seismic design group. The fiber models of 54 concrete frame structures are used as the actual monitored structures. The number of stories of the 54 selected monitored structures are 2 stories, 5 stories, and 8 stories, with story heights of 3.3m, 3.6m, and 3.9m. 70 ground motions are selected as the input. To verify the effectiveness of the model update, the ground motions are divided into a training set and a test set at a ratio of 2:8. When using the particle swarm optimization algorithm for model update, 100 groups of initial DSP parameter values are selected, and 100 particle swarm optimization processes are executed. Finally, the optimal group of DSP parameters is selected, and the overall update time is 11.19 minutes. To show the effect of the model update, the DSP parameters of the structure before and after the update are input into the trained XGBoost model to obtain the maximum inter-story drift angle of each story, and then the corresponding damage state is obtained through the HAZUS report. The comparison of the proportion of different structural damage states before and after the update is as Figure 11As shown, the specific data comparison is shown in Table 6. It can be seen that the errors of the proportions of different structural damage states after updating are all within 5%, indicating that this method is highly accurate.

[0079] Table 1: Selected ground motion indices

[0080]

[0081]

[0082] Table 2: Range of structural parameters

[0083]

[0084] Table 3: Hyperparameter range and data type of the XGBoost model

[0085]

[0086] Table 4: Optimal hyperparameter combination

[0087]

[0088]

[0089] Table 5: Comparison of the proportions and errors of different structural damage states after updating in Case 1

[0090]

[0091] Table 6: Comparison of the proportions and errors of different structural damage states after updating in Case 2

[0092]

[0093] It should be noted that the specific implementation manners are only explanations and illustrations of the technical solutions of the present invention, and the scope of the right protection cannot be limited thereby. Those that are only partial changes made according to the claims and the specification of the present invention should still fall within the protection scope of the present invention.

Claims

1. A rapid regional earthquake damage prediction method based on machine learning and model update, characterized in that It includes the following steps: Step 1: Establish a structural sample through a simplified multi-degree-of-freedom model, and then obtain the structural parameters of the structural sample, the maximum inter-story drift angle, and the ground motion parameters of the earthquake suffered by the structural sample. The structural parameters include easily obtained structural parameters and non-easily obtained structural parameters; Step 2: Use the structural parameters and ground motion parameters as inputs, and the maximum inter-story drift angle as the output to train the structural response prediction model; Step 3: For the building to be evaluated, obtain the easily obtained structural parameters and the monitored value of the maximum inter-story drift angle of the building to be evaluated. Then, obtain the ground motion parameters of the earthquake suffered by the building to be evaluated, and set the initial value of the non-easily obtained structural parameters. Then, input the ground motion parameters, the easily obtained structural parameters, and the initial value of the non-easily obtained structural parameters of the earthquake suffered by the building to be evaluated into the trained structural response prediction model to obtain the predicted value of the maximum inter-story drift angle as the output. Use the monitored value of the maximum inter-story drift angle and the predicted value of the maximum inter-story drift angle to obtain the objective function, and use the non-easily obtained structural parameters as the variables to be optimized. Use the particle swarm optimization algorithm to iteratively solve to obtain the optimal non-easily obtained structural parameters; Step 4: For the target earthquake scenario, obtain the corresponding ground motion parameters in this scenario, and input the ground motion parameters, the easily obtained structural parameters obtained in Step 3, and the optimal non-easily obtained structural parameters into the trained structural response prediction model to obtain the maximum inter-story drift angle as the output, and obtain the earthquake damage situation in the target earthquake scenario according to the output maximum inter-story drift angle.

2. A method for rapid regional earthquake damage prediction based on machine learning and model update according to claim 1, characterized in that The structural response prediction model is an XGBoost model.

3. A method for rapid regional earthquake damage prediction based on machine learning and model update according to claim 2, characterized in that The loss function of the XGBoost model is the MSE loss function, expressed as: where m i represents the number of layers of the i-th sample, n represents the total number of samples, and MIDR ij represents the maximum inter-story drift angle label value of the j-th layer of the i-th sample, represents the predicted value of the maximum inter-story drift angle of the j-th layer of the i-th sample.

4. A method for rapid regional earthquake damage prediction based on machine learning and model update according to claim 1, characterized in that The specific steps of Step 1 are: Establish a structural sample through a simplified multi-degree-of-freedom model, then obtain the structural parameters of the structural sample, establish a finite element model using the structural parameters, then obtain the ground motion parameters of the earthquake suffered by the structural sample, and input the ground motion parameters into the finite element model to obtain the maximum inter-story drift angle as the output.

5. A rapid regional earthquake damage prediction method based on machine learning and model update according to claim 1, characterized in that The objective function is expressed as: Among them, m represents the number of layers of the structure to be updated, and MIDR monitor,i represents the monitoring value of the maximum inter-story drift angle of the i-th layer, and MIDR predict,i represents the predicted value of the maximum inter-story drift angle of the i-th layer.

6. The regional rapid earthquake damage prediction method based on machine learning and model update according to claim 1, characterized in that The trained structural response prediction model is determined by the mean absolute error (MAE), mean absolute percentage error (MAPE), goodness of fit R 2 and kernel density estimation (KDE) plots; The mean absolute error MAE is expressed as: The mean relative error MAPE is expressed as: The goodness of fit R2 is expressed as: where m i represents the number of layers of the i-th sample, n represents the total number of samples, and MIDR ij represents the maximum inter-story drift angle label value of the j-th layer of the i-th sample, and represents the predicted value of the maximum inter-story drift angle of the j-th layer of the i-th sample.

7. A method for rapid regional earthquake damage prediction based on machine learning and model update according to claim 1, characterized in that The easily obtained structural parameters are the number of floors and the floor height.

8. A method for rapid regional earthquake damage prediction based on machine learning and model update according to claim 1, characterized in that The non-easily obtained parameters of the structure are the period calculation coefficient γ1 and the stiffness reduction coefficient α K and the hysteretic parameter τ.

9. A method for rapid regional earthquake damage prediction based on machine learning and model update according to claim 1, characterized in that The ground motion parameters include PGD, PGA, PGD, PGV / PGA, I a , HI, ASI, VSI, EPA, SED, CAV, A rms , V rms and Sa(ξ = 5%, T = 0.1s) to Sa(ξ = 5%, T = 3.0s).

10. A method for rapid regional earthquake damage prediction based on machine learning and model update according to claim 1, characterized in that The specific steps of Step 2 are: Use the structural parameters and ground motion parameters as inputs, and the maximum inter-story drift angle as the output to construct training data, and divide the training data into a training set, a validation set, and a test set in a ratio of 0.64:0.17:0.

19. Then, use the training set, the validation set, and the test set to complete the training of the structural response prediction model.

Images