Regional earthquake risk assessment method based on deep learning proxy model

Abstract

The invention provides a regional earthquake risk assessment method based on a deep learning proxy model, and the method comprises the steps: preparing A parts of training earthquake data and B parts of verification earthquake data, the A parts of training seismic data are input into an image generator of a deep learning framework for iteration to generate a corresponding predicted seismic oscillation intensity distribution diagram and a regional seismic risk distribution diagram; comparing with a true value, and constructing a feature loss function and an adversarial loss function for evaluating the difference between the two values; continuously adjusting generator parameters based on the feature loss function and the adversarial loss function until the constructed loss function value tends to be stable, and completing the training of the constructed model; the verification data are input into the constructed model to verify the performance of the model, the to-be-detected seismic data are input into the trained model, then the to-be-detected seismic risk distribution diagram is output, and the damage level is judged, the method reduces errors generated by calculation two-stage analysis, and the regional seismic assessment efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a regional earthquake risk assessment method based on a deep learning agent model. Background Art

[0002] Infrastructure (such as transportation infrastructure) is one of the most widely affected entities in earthquake events and plays a key role in emergency response. For post-earthquake rapid emergency response planning, efficient regional earthquake risk assessment is needed. Since the relationship between ground motion and infrastructure damage cannot usually be expressed by mathematical expressions in analytical form, earthquake risk assessment mainly adopts event-based methods, in which earthquake hazards are modeled by simulating scenario earthquakes or various seismic events through Monte Carlo simulations as part of event-based probabilistic loss estimation models. It includes ground motion prediction and damage risk estimation. The traditional regional earthquake risk framework mainly includes two stages, such as Figure 1 Shown: Ground Motion Intensity (GMI) prediction and loss risk estimation. First, the earthquake scenario based on the earthquake source model is used as input to predict the GMI map. Then, the predicted GMI data is used as input to estimate the loss risk of the infrastructure of interest. However, a region usually involves a large number of infrastructure systems, which have specific geometry, materials, structural properties and soil conditions, thus requiring a large amount of data access and complex individual calculations. Traditional regional earthquake assessments are inefficient, labor-intensive, time-consuming, and difficult to meet the needs of rapid emergency response. At the same time, due to the use of a two-stage analysis method, the overall error is large. Summary of the invention

[0003] The purpose of the present invention is to provide a regional earthquake risk assessment method based on a deep learning agent model to solve the problems raised in the above background technology. To achieve the above purpose, the present invention provides the following technical solutions: A regional earthquake risk assessment method based on a deep learning agent model comprises the following steps: Prepare A sets of training earthquake data and B sets of verification earthquake data, wherein each set of training earthquake data includes training earthquake event parameters and corresponding training earthquake intensity parameters and distribution maps and real regional earthquake risk distribution maps, and the verification earthquake data includes verification earthquake event parameters and corresponding verification earthquake intensity parameters and real regional earthquake risk distribution maps; A set of training earthquake event parameters are input into the image generator of the deep learning framework for iterative output of the corresponding predicted seismic intensity distribution map and predicted regional earthquake risk distribution map; the predicted seismic intensity distribution map and the predicted regional earthquake risk distribution map and the corresponding real regional earthquake risk distribution map are input into the feature comparator of the deep learning framework, the feature comparator is used to compare the feature differences between the predicted seismic intensity distribution map and the predicted regional earthquake risk distribution map and the real regional earthquake risk distribution map, and construct a feature loss function to evaluate the difference between the predicted regional earthquake risk distribution map and the real regional earthquake risk distribution map at the feature level; at the same time, the predicted seismic intensity distribution map and the predicted regional earthquake risk distribution map and the real regional earthquake risk distribution map are input into the discriminator of the deep learning framework, and an adversarial loss function is constructed to evaluate the difference in distribution between the two, and then, the generator parameters are adjusted based on the feature loss function and the adversarial loss function until the constructed loss function value tends to be stable, and the generator learning model is obtained, and the B set of verification earthquake data is input into the constructed generator learning model to verify the effect of the generator learning model to obtain the trained generator; When an earthquake event or a simulated earthquake event occurs in a region, earthquake data to be tested is obtained, and the earthquake data to be tested is input into a trained generator, and then a distribution map of earthquake risk to be tested is output; Based on the earthquake risk distribution map to be tested, the corresponding damage level is determined according to the earthquake damage model.

[0004] Furthermore, the training earthquake data and verification earthquake data are obtained through real earthquake records or numerical simulations, and the training earthquake event parameters include magnitude, focal depth, earthquake center and rupture surface; the training seismic intensity parameters include peak acceleration, peak velocity, intensity and spectral acceleration.

[0005] Furthermore, the generator is trained to map the training earthquake event parameters to the predicted regional seismic intensity distribution map and the regional earthquake risk distribution map.

[0006] Furthermore, the total loss function of the generator is defined as follows: L G =L featG + λL advG Among them, L featG is the feature loss, L advG is the adversarial loss, and λ is the weight parameter used to balance the feature loss and the adversarial loss.

[0007] Furthermore, the feature loss is obtained by training the feature comparator. The feature loss aims to minimize the difference between the generated image and the real image at the feature level. The mean square error (MSE) at the feature level is chosen to measure this difference.

[0008] Furthermore, the feature comparator adopts the pre-trained VGG19 model.

[0009] Furthermore, the adversarial loss is the output of discriminator training. The adversarial loss requires that the image generated by the generator can fool the discriminator so that the discriminator cannot distinguish between the generated image and the real image. The adversarial loss function is defined as follows:

[0010] in, is a real image, is the production image and N is the number of samples.

[0011] Furthermore, the earthquake damage model is the HAZUS earthquake model, and the damage levels include slight, moderate, extensive or complete damage. Beneficial Effects

[0012] The deep learning framework of the present invention maps end-to-end from earthquake parameters to regional seismic intensity or infrastructure damage risk maps. In order to achieve an overall regional earthquake risk assessment, the regional seismic intensity or damage risk map is used as a proxy output. Although this requires the proxy model to generate maps by capturing the extremely complex nonlinear relationship between earthquake source parameters and high-dimensional maps, the deep generative learning framework is used to generate seismic intensity and risk maps and residual learning by combining convolutional neural networks (CNNs), generative adversarial networks (GANs) and perceptual losses to suppress degradation problems and enhance the generalization ability of deep generative networks in earthquake risk prediction, overcoming the shortcomings of traditional methods that require a large amount of data access and complex individual calculations, while reducing the errors generated by the two-stage analysis of the calculation, and improving the efficiency and accuracy of regional earthquake assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a framework map of regional earthquake risk based on existing technology; Figure 2 It is a training flow chart of the deep learning framework of the present invention; Figure 3 It is a flow chart of application after deep learning framework training of the present invention; Figure 4 It is a structural diagram of the generator of the present invention; Figure 5 It is a structural diagram of the discriminator of the present invention; Figure 6 is a verification training example diagram of the present invention; Figure 7 This is an example diagram of a verification test of the present invention; Figure 8 This is an example diagram of the second verification test of the present invention. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0015] The present invention provides a regional earthquake risk assessment method based on a deep learning agent model, comprising the following steps: S1: Prepare A sets of training earthquake data and B sets of verification earthquake data, wherein each set of training earthquake data includes training earthquake event parameters and corresponding training earthquake intensity parameters and distribution maps and real regional earthquake risk distribution maps, and the verification earthquake data includes verification earthquake event parameters and corresponding verification earthquake intensity parameters and real regional earthquake risk distribution maps; Specifically, the training earthquake data and the verification earthquake data are obtained through real earthquake records or numerical simulations, and the training earthquake event parameters and the verification earthquake event parameters include magnitude, focal depth, earthquake center and rupture surface; the training seismic intensity parameters and the verification seismic intensity parameters include peak acceleration, peak velocity, intensity, and spectral acceleration.

[0016] S2: Figure 2 , each set of training earthquake event parameters is input into the image generator of the deep learning framework for iterative output of the corresponding predicted seismic intensity distribution map and the predicted regional earthquake risk distribution map; the predicted regional earthquake risk distribution map and the corresponding real regional earthquake risk distribution map are input into the feature comparator of the deep learning framework, the feature comparator is used to predict the features of the regional earthquake risk distribution map and the corresponding real regional earthquake risk distribution map, and evaluate the difference between the predicted regional earthquake risk distribution map and the real regional earthquake risk distribution map at the feature level, and then construct a feature loss function; the training earthquake event parameters, the corresponding predicted regional earthquake risk distribution map and the real regional earthquake risk distribution map are input into the discriminator of the deep learning framework to minimize the difference in distribution between the predicted real regional earthquake risk distribution map and the real regional earthquake risk distribution map, and then construct an adversarial loss function, adjust the generator parameters based on the feature loss function and the adversarial loss function until the constructed loss tends to be stable to obtain the constructed generator, and input B sets of verification earthquake data into the constructed generator to obtain the trained generator; Specifically, the generator is trained to map training earthquake event parameters to a predicted regional earthquake risk distribution map; The generator is a modified convolutional neural network (CNN), which needs to train the complex relationship between earthquake event parameters and the earthquake risk distribution map of the predicted area, and its function is to minimize the difference between the earthquake risk distribution map of the predicted area and the earthquake risk distribution map of the real area; the construction of the generator is as follows Figure 4 ; The definition and design of the loss function is crucial for the training of the generator. The loss function consists of two parts. The total loss function of the generator is defined as follows: L G =L featG + λL advG Among them, L featG is the feature loss, L advG is the adversarial loss, and λ is the weight parameter used to balance the feature loss and the adversarial loss.

[0017] Preferably, the feature loss is obtained by training a feature comparator, and the feature loss aims to minimize the difference between the generated image and the real image at the feature level, and the mean square error (MSE) at the feature level is selected to measure this difference; the feature comparator adopts a pre-trained VGG19 model, which is a 19-layer convolutional neural network that has been pre-trained on a large number of images on the ImageNet database and can effectively extract image features; Secondly, the adversarial loss is output through discriminator training. The adversarial loss requires that the image generated by the generator can fool the discriminator so that the discriminator cannot distinguish between the generated image and the real image. The adversarial loss function is defined as follows:

[0018] in, is a real image, is the production image and N is the number of samples.

[0019] See Figure 5 , the discriminator receives two types of inputs at the same time: the regional earthquake risk distribution map and the training earthquake event parameters. The regional earthquake risk distribution map includes the predicted regional earthquake risk distribution map and the real regional earthquake risk distribution map (that is, the real value); the output is a single value from 0 to 1, which represents the probability that the input regional earthquake risk distribution map is real under the corresponding training earthquake event parameters. For the input of the training earthquake event parameters, a fully connected layer is used to encode it into a latent vector. For the input of the regional earthquake risk distribution map, a series of residual blocks are used for convolution. The residual blocks in the discriminator are similar to the residual blocks of the generator.

[0020] S3: Figure 3 , input the earthquake data to be tested into the trained generator and then output the earthquake risk distribution map to be tested; S4: Determine the corresponding damage level according to the earthquake damage model based on the earthquake risk distribution map to be tested; Specifically, the earthquake damage model is the HAZUS earthquake model, and the damage levels include slight, moderate, extensive or complete damage.

[0021] In order to verify the deep learning framework of the present invention, the present invention provides the following examples: Get data: Based on the training earthquake event parameters, the corresponding training earthquake intensity parameters are generated by the traditional earthquake framework of ground motion prediction equation (GMPE), spatial interpolation and fragility function. Here, the training earthquake intensity parameters and distribution maps and the real regional earthquake risk distribution maps are generated by ShakeMap software. A total of 100 sets of relevant data are generated in this verification scheme. Specifically, the focal depth in the training earthquake event parameters is set to vary between [20, 200] kilometers, the magnitude is set to vary between [5.5, 10.0], and other parameters remain unchanged or use default values. Using ShakeMap, the MMI and SA (1.0) values ​​on the regional grid are calculated, and the corresponding training earthquake intensity parameter map is generated using the specified visualization and color scheme. Since these features are uniformly distributed within a given range, there is no correlation between them. It should be noted that in practice, the more training earthquake event parameters are considered, the more data samples are usually required to train the corresponding high-dimensional alternative model to avoid overfitting problems and explore the interpolation and extrapolation capabilities of the framework. The generated 100 sets of related data are divided into a training set (70 sets of sample data), an interpolation test set #1 (10 sets of sample data), and an extrapolation test set #2 (20 sets of sample data). Among them, test set #1 is within the range of the training set data, while test set #2 is beyond the range of the training set data. Figure 6 ; Earthquake Risk Assessment: The selected area is 945 highway bridges in a specific area. Specifically, the processing steps are: Step 1: Amplify the training ground motion intensity according to the site category to take into account the influence of soil conditions; Step 2: Determine the classification of the bridge based on its descriptive parameters such as construction year, material type, bridge inclination, span, and bridge length; Step 3: According to the amplified training ground motion intensity parameters and bridge categories, the corresponding fragility function is used to predict the damage probability; Step 4: Calculate the probability of reaching a certain impairment status threshold (such as minor impairment, moderate impairment, severe impairment, and complete impairment).

[0022] The required bridge information is provided by the National Bridge Inventory (NBI) published by a certain country, and the relevant information can be obtained on its website, while the fragility model is provided by the Hazus Earthquake Model Technical Manual.

[0023] The above steps are also processed by ShakeMap software, and the training ground motion intensity generated by ShakeMap software is used to calculate the damage probability of all bridges. Specifically, the "minor damage state" is selected as the threshold example for calculation, and the above bridges and their minor damage probabilities are mapped to the study area to generate the relevant damage risk map (see Figure 6 Finally, to verify the output flexibility of the proposed deep learning (DL) framework, a combined map combining ground motion intensity and damage risk information was generated (see Figure 6 All these maps are stored as 256×256×3 pixel images as output data of the framework.

[0024] Based on the data formed above, the obtained training set is input into the deep learning framework for training to form a trained generator, and then the test set is input into the trained generator to output the MMI map, SA (1.0) map, bridge damage risk map, and a combined map containing SA (1.0) and damage risk. The map generated based on the traditional earthquake risk framework is regarded as the real result. Figure 7 The example interpolation prediction result of an earthquake with a focal depth of 100 kilometers and a magnitude of 9.0 is shown. To further study the extrapolation prediction ability of the model, predictions are made for test set #2 which is outside the range of the training data. Figure 8 An example extrapolated prediction result for a magnitude 7.0 earthquake with a focal depth of 200 km is shown. Whether interpolated or extrapolated, the predicted earthquake risk map is consistent with the real map generated by the traditional earthquake risk framework.

[0025] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. The protection scope of the present invention shall be subject to the protection scope of the claims. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A regional earthquake risk assessment method based on a deep learning agent model, characterized in that: The following steps are involved: Prepare A sets of training earthquake data and B sets of verification earthquake data, wherein each set of training earthquake data includes training earthquake event parameters and corresponding training earthquake intensity parameters and distribution maps and real regional earthquake risk distribution maps, and the verification earthquake data includes verification earthquake event parameters and corresponding verification earthquake intensity parameters and real regional earthquake risk distribution maps; The A-part training earthquake event parameters are all input into the image generator of the deep learning framework for iterative output of the corresponding predicted seismic intensity distribution map and predicted regional earthquake risk distribution map; the predicted seismic intensity distribution map and the predicted regional earthquake risk distribution map and the corresponding real regional earthquake risk distribution map are input into the feature comparator of the deep learning framework, the feature comparator is used to compare the feature differences between the predicted seismic intensity distribution map and the predicted regional earthquake risk distribution map and the real regional earthquake risk distribution map, and construct a feature loss function to evaluate the difference between the predicted regional earthquake risk distribution map and the real regional earthquake risk distribution map at the feature level; at the same time, the predicted seismic intensity distribution map and the predicted regional earthquake risk distribution map and the real regional earthquake risk distribution map are input into the discriminator of the deep learning framework, and an adversarial loss function is constructed to evaluate the difference in distribution between the two. The generator parameters are adjusted based on the feature loss function and the adversarial loss function until the constructed loss function value tends to be stable, and the generator learning model is obtained. The B-part verification earthquake data is input into the constructed generator learning model, and the effect of the generator learning model is verified to obtain the trained generator; When an earthquake event or a simulated earthquake event occurs in a region, earthquake data to be tested is obtained, and the earthquake data to be tested is input into a trained generator, and then a distribution map of earthquake risk to be tested is output; The corresponding damage level is determined based on the earthquake risk distribution map to be tested and the earthquake damage model.

2. A regional earthquake risk assessment method based on a deep learning agent model according to claim 1, characterized in that: The training earthquake data and verification earthquake data are obtained through real earthquake records or numerical simulations. The training earthquake event parameters include magnitude, focal depth, earthquake center and rupture surface; the training seismic intensity parameters include peak acceleration, peak velocity, intensity and spectral acceleration.

3. A regional earthquake risk assessment method based on a deep learning agent model according to claim 1, characterized in that: The generator is trained to map training earthquake event parameters to predicted regional seismic intensity distribution maps and regional earthquake risk distribution maps.

4. A regional earthquake risk assessment method based on a deep learning agent model according to claim 3, characterized in that: The total loss function of the generator is defined as follows: THE G =L featG + λL advG Among them, L featG is the feature loss, L advG is the adversarial loss, and λ is the weight parameter used to balance the feature loss and the adversarial loss.

5. A regional earthquake risk assessment method based on a deep learning agent model according to claim 4, characterized in that: The feature loss is obtained by training the feature comparator. The feature loss aims to minimize the difference between the generated image and the real image at the feature level. The mean square error at the feature level is chosen to measure this difference.

6. A regional earthquake risk assessment method based on a deep learning agent model according to claim 5, characterized in that: The feature comparator adopts the pre-trained VGG19 model.

7. A regional earthquake risk assessment method based on a deep learning agent model according to claim 1, characterized in that: The adversarial loss is the output of the discriminator training, and the adversarial loss function is defined as follows: ； in, is a real image, is the production image and N is the number of samples.

8. A regional earthquake risk assessment method based on a deep learning agent model according to claim 1, characterized in that: The earthquake damage model is the HAZUS earthquake model, and the damage levels include slight, moderate, extensive or complete damage.

Images