JP7423427B2 - Seismic motion evaluation model generation method, seismic motion evaluation model generation device, seismic motion evaluation method, and seismic motion evaluation device - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act (1) Published via telecommunications line, Proceedings of the 2019 Autumn Conference of the Seismological Society of Japan, https://confit. atlas. jp/guide/event/zisin2019/subject/S22-08/advanced, https://confit. atlas. jp/guide/event/zisin2019/subject/S22-09/advanced, August 16, 2020 (2) Presentation at the meeting, Seismological Society of Japan 2019 Autumn Conference, September 18, 2021 (3 ) Published in a publication, Abstracts of the 14th Annual Conference of the Japan Society of Earthquake Engineering (CD-ROM), September 18, 2021 (4) Presented at the 14th Annual Meeting of the Japan Society of Earthquake Engineering Conference, September 19, 2020 (5) Published via telecommunications line, Proceedings of the 2020 National Conference of the Japanese Society for Artificial Intelligence (34th), https://confit. atlas. jp/guide/event/jsai2020/subject/3Rin4-03/advanced? eventCode=jsai2020, May 22, 2020

The present invention relates to an earthquake motion evaluation model generation method, an earthquake motion evaluation model generation device, an earthquake motion evaluation method, and an earthquake motion evaluation device.

Technology that evaluates and predicts the characteristics of seismic motion due to earthquakes that are thought to occur in the future based on various data obtained regarding past earthquakes (slip rupture phenomena of seismic source faults) and seismic motion (ground shaking caused by earthquakes). Due to its usefulness, it is widely used in society for everything from analysis and interpretation of seismic motion characteristics to behavior prediction and structural design of buildings and structures, and even earthquake disaster prevention.

Previously, some data was selected by individual experts using their own judgment, and was analyzed using specialized methods to determine the maximum amplitude (maximum acceleration, maximum velocity, maximum displacement) of seismic motion caused by future earthquakes. etc.) and response spectrum (maximum response amplitude of a single mass point system when earthquake motion is input to a single mass point system with a certain natural period and attenuation constant), etc.). (often referred to as evaluation formulas, seismic motion prediction formulas, etc.) have been developed and utilized (for example, see Patent Document 1 and Non-Patent Document 1).

Japanese Patent Application Publication No. 2008-39446

Nobuyuki Morikawa and Hiroyuki Fujiwara, A New Ground Motion Prediction Equation for Japan Applicable up to M9 Mega-Earthquake, Journal of Disaster Research, Vol.8, No.5, 2013, p.878-888

Practical prediction formulas such as the distance attenuation formula developed and utilized through past studies depend on the data of earthquake observation records used in the study by individual experts, and they are not suitable for large-scale earthquakes or the largest earthquakes. Since earthquake motions are rare natural phenomena, there is an imbalance in the quality and quantity of past knowledge reflected in their prediction formulas. In addition, the increase in the amount of data that is accumulated and analyzed on a daily basis and the improvement in the information processing capacity of computers have the potential to greatly contribute to improving this problem, but each time new data is obtained, individual It takes an enormous amount of time and effort for experts to reexamine and reevaluate prediction formulas, so there are some limitations.

The present invention was made in consideration of these circumstances, and takes full advantage of the increase in data volume and the improvement in information processing ability of computers, while improving the characteristics of seismic motion caused by earthquakes that are expected to occur in the future. It is an object of the present invention to provide a seismic motion evaluation model generation method, a seismic motion evaluation model generation device, an earthquake motion evaluation method, and a seismic motion evaluation device that can evaluate and predict with accuracy.

The present invention solves the above problems, and a seismic motion evaluation model generation method according to an embodiment of the present invention includes:
A seismic motion evaluation model generation method for generating an earthquake motion evaluation model by machine learning using a computer, the method comprising:
For each combination of a plurality of actual earthquakes and at least one observation point where actual earthquake motion due to the plurality of actual earthquakes was observed, various earthquake motion characteristic parameters and earthquake motion observation records when the actual earthquake motion was observed are stored in association with each other. From a first database, acquire a plurality of first learning data consisting of the feature quantities and the objective variables, with the seismic motion characteristic parameters as the feature quantities and the seismic motion index obtained from the seismic motion observation record as the objective variable. a first acquisition step of
For each combination of a plurality of virtual earthquakes and at least one virtual observation point from which virtual seismic motions due to the plurality of virtual earthquakes are calculated by simulation, various earthquake motion characteristic parameters and earthquake motion calculations when the virtual earthquake motions are calculated by the simulation. From a second database that stores the results in association with each other, a second database consisting of the feature quantities and the objective variables is obtained, with the various seismic motion characteristic parameters as the feature quantities and the seismic motion index obtained from the seismic motion calculation results as the objective variable. a second acquisition step of acquiring a plurality of learning data;
The feature amount is calculated based on the plurality of first learning data acquired in the first acquisition step and the plurality of second learning data acquired in the second acquisition step. and a generation step of generating the seismic motion evaluation model as a learned model of the machine learning by learning the correlation of the target variables by the machine learning.

Furthermore, the seismic motion evaluation model generation device according to an embodiment of the present invention includes:
The computer is equipped with a control unit that executes each step included in the seismic motion evaluation model generation method.

Furthermore, the seismic motion evaluation method according to an embodiment of the present invention includes:
A seismic motion evaluation method that uses a computer to evaluate the characteristics of the seismic motion based on the seismic motion evaluation model generated by the seismic motion evaluation model generation method,
a reception step of receiving the seismic motion characteristic parameters to be predicted;
The various seismic motion characteristic parameters of the prediction target received in the reception process are input into the seismic motion evaluation model as the feature quantities, and based on the objective variables output from the seismic motion evaluation model, the seismic motion characteristic parameters of the prediction target are and a prediction step of predicting the seismic motion index corresponding to the seismic motion characteristic parameters.

Furthermore, the seismic motion evaluation device according to an embodiment of the present invention includes:
The computer is equipped with a control unit that executes each step included in the seismic motion evaluation method.

According to the seismic motion evaluation model generation device and the seismic motion evaluation device according to an embodiment of the present invention, the first learning data is based on actual earthquakes that occurred in the past, and the second learning data is based on the virtual earthquake calculated by simulation. By learning the correlation between feature values and objective variables using machine learning based on the learning data of It is possible to provide a seismic motion evaluation model that can highly accurately evaluate and predict the characteristics of seismic motion due to earthquakes that are expected to occur in the future, while making the most of improved capabilities.

Further, according to the seismic motion evaluation method and seismic motion evaluation device according to an embodiment of the present invention, by using the seismic motion evaluation model generated by the seismic motion evaluation model generation device and the seismic motion evaluation model generation method, individual specialized It is possible to evaluate and predict with high precision the characteristics of seismic motion caused by earthquakes that are expected to occur in the future, without relying on experience in the home.

1 is a schematic configuration diagram showing an example of a seismic motion evaluation system 1 according to a first embodiment of the present invention. FIG. 1 is a block diagram showing an example of a seismic motion evaluation system 1 according to a first embodiment of the present invention. FIG. 2 is a functional explanatory diagram showing an example of the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method according to the first embodiment of the present invention. 1 is a data configuration diagram showing an example of a database 10. FIG. It is a scatter diagram showing the correlation of each feature of a set of learning data 12 (learning data set A). FIG. 2 is a schematic diagram showing an overview of a gradient boosting tree. FIG. 2 is a functional explanatory diagram showing an example of the seismic motion evaluation device 4 and the seismic motion evaluation method according to the first embodiment of the present invention. An example of the output results when representing the seismic motion index as a distance attenuation characteristic is shown. (a) is the maximum acceleration PGA according to the moment magnitude Mw, and (b) is the pseudo velocity response spectrum pSv (1.0 s) according to the epicenter azimuth Λ. ). The degree of influence of each feature (earthquake characteristic parameters) on the objective variable (seismic motion index) is shown, (a) is Kanto model A5, (b) is Kanto model A6, (c) is Kanto model B5, (d) is It is a figure which each shows Kanto model B6. It is a scatter diagram showing the correlation between observed values obtained from seismic motion observation records and evaluation values obtained by seismic motion evaluation model 13 for each objective variable (seismic motion index) in Kanto model A6. The "evaluation value/observation value" ratio of each objective variable (earthquake motion index) is shown as a histogram, and (a) is a diagram showing Kanto model A5, and (b) is a diagram showing Kanto model A6, respectively. It is a figure which shows the average and dispersion of the "evaluation value/observation value" ratio of each objective variable (earthquake motion index) about Kanto models A5, A6, B5, and B6. It is a functional explanatory diagram showing an example of a seismic motion evaluation model generation device 3 and a seismic motion evaluation model generation method according to a second embodiment of the present invention. 3 is a scatter diagram showing the correlation between each feature of a set of learning data 12, in which (a) shows a learning data set S and (b) a learning data set T. FIG. It is a functional explanatory diagram showing an example of a seismic motion evaluation device 4 and a seismic motion evaluation method according to a second embodiment of the present invention. An example of an output result when expressing a seismic motion index as an azimuth characteristic is shown, (a) is a diagram showing a point model S4, and (b) is a diagram showing a point model T4, respectively. It is a figure which shows the degree of influence of each feature quantity (earthquake motion characteristic parameters) on the target variable (seismic motion index), and (a) shows the point model S4, and (b) shows the point model T4, respectively. For each objective variable (seismic motion index), the correlation between the observed value obtained from the seismic motion observation record and the evaluation value obtained from the seismic motion evaluation model 13 is shown. (a) shows the point model S4, and (b) shows the point model T4. It is a scatter diagram showing each. The "evaluation value/observation value" ratio of each objective variable (earthquake motion index) is shown as a histogram, and (a) is a diagram showing the point model S4, and (b) is a diagram showing the point model T4, respectively. It is a figure which shows the average and dispersion of the "evaluation value/observation value" ratio of each objective variable (earthquake motion index) about point model S4, T4. FIG. 2 is a schematic diagram showing the epicenter characteristics, propagation characteristics, and site characteristics of seismic motion at a plurality of nearby observation points. FIG. 3 is a diagram showing a first observation point group C in which a plurality of observation points c1 to c3 are grouped together, and a second observation point group D in which a plurality of observation points d1 to d3 are grouped together. It is a functional explanatory diagram showing an example of a seismic motion evaluation model generation device 3 and a seismic motion evaluation model generation method according to a third embodiment of the present invention. It is a functional explanatory diagram showing an example of a seismic motion evaluation device 4 and a seismic motion evaluation method according to a third embodiment of the present invention. It is a block diagram showing an example of seismic motion evaluation system 1 concerning a 4th embodiment of the present invention. It is a functional explanatory diagram showing an example of a seismic motion evaluation model generation device 3 and a seismic motion evaluation model generation method according to a fourth embodiment of the present invention. It is a scatter diagram showing the interpolation relationship between the first learning data 12A and the second learning data 12B. FIG. 3 is a diagram showing interdependence of each feature amount by cluster analysis. Various data formats of the epicenter azimuth Λ are shown, (a) the definition of each data format is shown, and (b) the degree of influence of each feature quantity (seismic motion characteristic parameters) on the target variable (earthquake motion index) in each data format. FIG.

Hereinafter, one embodiment of the present invention will be described with reference to the accompanying drawings.

(First embodiment)
FIG. 1 is a schematic configuration diagram showing an example of a seismic motion evaluation system 1 according to a first embodiment of the present invention. FIG. 2 is a block diagram showing an example of the seismic motion evaluation system 1 according to the first embodiment of the present invention.

The seismic motion evaluation system 1 includes an observation data providing device 2A that collects seismic motion observation records in which seismic motions caused by earthquakes are observed at a plurality of observation points, and provides the observation results to the outside; An analysis data providing device 2B that analyzes the epicenter and scale and provides the analysis results to the outside; an underground structure data providing device 2C that provides the underground structure parameters related to the underground structure used for evaluating and predicting seismic motion to the outside; and a data providing device. The seismic motion evaluation model generation device 3 learns data provided by the devices 2A to 2C by machine learning and generates the seismic motion evaluation model 13, and the seismic motion evaluation model 13 is generated based on the seismic motion evaluation model 13 generated by the seismic motion evaluation model generation device 3. It includes a seismic motion evaluation device 4 that evaluates and predicts seismic motion, and a network 5 that connects each device.

The network 5 communicates various data and signals by wireless communication or wired communication, and any communication standard is used. In this embodiment, the seismic motion evaluation model generation device 3 and the seismic motion evaluation device 4 are described as being separate devices, but they may be configured as one device (seismic motion processing device).

The observation data providing device 2A transmits three-component time history waveform data in the north-south direction, east-west direction, and up-down direction measured by seismometers (not shown) installed at multiple observation points when an earthquake occurs. Each is collected as a seismic motion observation record, and observation data including the seismic motion observation record and additional information such as the position of an observation point indicating the position of an observation point is provided to the outside. In this embodiment, the observation data is, for example, data provided by the strong motion observation network K-NET of the National Research and Development Agency, National Research Institute for Earth Science and Disaster Prevention (hereinafter referred to as "NIED"), and is particularly The following explanation assumes that data observed at 138 observation points (see Figure 1) installed in the six prefectures of Tokyo (Tokyo, Kanagawa, Chiba, Saitama, Ibaraki, Tochigi, and Gunma) are used.

The analysis data providing device 2B analyzes the epicenter and scale of the earthquake based on the seismic motion observation records, and provides, as the analysis results, moment magnitude, Japan Meteorological Agency magnitude, epicenter position, epicenter depth, earthquake type, fault type, and Analytical data including earthquake source mechanism solutions will be provided externally. In this embodiment, the analytical data providing device 2B will be described as using data provided by, for example, the broadband seismic observation network F-NET of the National Institute of Disaster Prevention Science or the Japan Meteorological Agency.

The underground structure data providing device 2C externally transmits underground structure data including, for example, earthquake bedrock depth, engineering bedrock depth, layer thickness, density, seismic wave propagation velocity, Q value, attenuation constant, etc. as underground structure parameters. Provided to. In this embodiment, the underground structure data providing device 2C will be described as using data provided by, for example, the Earthquake Hazard Station J-SHIS of the National Research Institute for Earth Science and Disaster Prevention.

Note that when an earthquake occurs, the data providing devices 2A to 2C may provide provided data regarding the earthquake to the seismic motion evaluation device 4 in real time, or upon receiving a request for data from the seismic motion evaluation device 4. , the seismic motion evaluation device 4 may be provided with data related to the request (data related to a plurality of earthquakes that have occurred in the past that meet predetermined conditions). Further, in this embodiment, the data providing devices 2A to 2C are described as being three separate devices, but they are not limited to this, and may be configured as a single device, or may be configured as a single device, or may be configured as a single device. may be further added.

(About the configuration of the seismic motion evaluation model generation device 3 and the processes involved in each part)
The seismic motion evaluation model generation device 3 generates, for example, a gradient boosting tree, a random forest, a neural network, a convolution neural By executing a machine learning algorithm such as a network (CNN), a seismic motion evaluation model 13 is generated as a machine learning trained model.

The seismic motion evaluation model generation device 3 is composed of a general-purpose or dedicated computer, and as shown in FIG. 31, a communication section 32 which is a communication interface with the network 5, an input section 33 composed of a keyboard, a mouse, etc., and a display section 34 composed of a display, a touch panel, etc.

The storage unit 30 includes a database 10 in which provided data provided by the data providing devices 2A to 2C is registered and updated, a seismic motion evaluation model 13 that is a trained model, and a seismic motion evaluation model generating device 3 that controls operations. A seismic motion evaluation model generation program 300 that implements a seismic motion evaluation model generation method is stored.

The control unit 31 functions as a DB management unit 310, an acquisition unit 311, and a generation unit 312 by executing the seismic motion evaluation model generation program 300. Note that the details of the functions of each part will be described later.

FIG. 3 is a functional explanatory diagram showing an example of the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method according to the first embodiment of the present invention.

(About the database management process by the DB management unit 310 and the database 10)
The DB management unit 310 manages the database 10 based on the provided data (observation data, analysis data, underground structure data) provided by the data providing devices 2A to 2C. Specifically, each time an earthquake occurs, seismic motion due to the earthquake is observed at at least one observation point, and the data providing devices 2A to 2C provide data related to the earthquake, the DB management unit 310 For each combination of an earthquake and at least one observation point where seismic motion due to the earthquake was observed, various seismic motion characteristic parameters and seismic motion observation records at the time the seismic motion was observed are associated and registered in the database 10.

FIG. 4 is a data configuration diagram showing an example of the database 10. The database 10 includes various earthquake motion characteristic parameters and earthquake motion observation records when the earthquake motion was observed, for each combination of multiple earthquakes that occurred in the past and at least one observation point where the earthquake motion due to the multiple earthquakes was observed. A plurality of pieces of seismic motion data 11 are registered and stored in association with each other.

Seismic motion characteristic parameters are various parameters that describe various characteristics of earthquake motion, and various characteristics of earthquake motion include, for example, source characteristics and propagation characteristics of earthquake motion, site characteristics, azimuth characteristics, and earthquake motion observation characteristics. further including. In this embodiment, the seismic motion characteristic parameters are obtained by classifying and recording provided data (observation data, analysis data, underground structure data) provided by the data providing devices 2A to 2C according to the above-mentioned seismic motion characteristics. It is. Below, the seismic source characteristics, propagation characteristics, site characteristics, azimuth characteristics, and seismic motion observation characteristics included in the seismic motion characteristic parameters will be explained.

The epicenter characteristics are, for example, magnitude (moment magnitude Mw, Japan Meteorological Agency magnitude, etc.), epicenter position (latitude lat_eq, longitude lon_eq), epicenter depth H, earthquake type (inland crustal earthquake, plate boundary earthquake, intraslab earthquake), Fault type Mech (normal fault, reverse fault, strike-slip fault), source mechanism solution (strike1, dip1 dip1, slip angle rake1), and conjugate solution of the source mechanism solution (strike2, dip2 dip2, slip angle rake2) At least one of the following. The epicenter characteristics according to this embodiment include moment magnitude Mw, epicenter position (latitude lat_eq, longitude lon_eq), epicenter depth H, earthquake type Type, fault type Mech, epicenter mechanism solution (Strike1, dip1, rake1), and epicenter These are the conjugate solutions (Strike2, dip2, rake2) of the mechanism solution.

The propagation characteristic is, for example, at least one of the epicenter distance X, the shortest fault distance, and the epicenter distance. The propagation characteristic according to this embodiment is the epicenter distance X, which is calculated as the distance between the observation point position and the epicenter position.

The site characteristics are, for example, at least one of observation point position (latitude lat_site, longitude lon_site), seismic basement depth, engineering basement depth, layer thickness, density, seismic wave propagation velocity, Q value, and attenuation constant. It is. The site characteristics according to this embodiment are observation point position (latitude lat_site, longitude lon_site), top layer S-wave velocity VS1, surface layer 10m average S-wave velocity AVS10, surface layer 30m average S-wave velocity AVS30, topography classification JCODE, S Wave velocity 700 m/s layer top depth D7, S wave velocity 1400 m/s layer top depth D17, S wave velocity 2100 m/s layer top depth D24, and earthquake basement depth D28 (=S wave velocity 2700 m/s) s layer top surface depth). In addition, if the surface layer 30m average S-wave velocity AVS30 cannot be obtained from the underground structure data of the observation point, the surface layer 30m average S-wave velocity AVS30 of the 250m mesh of the Earthquake Hazard Station J-SHIS of the National Research Institute for Earth Science and Disaster Prevention is substituted. If the 10 m average S-wave velocity AVS10 cannot be obtained from the underground structure data at the observation point, the 30 m average S-wave velocity AVS30 in the surface layer was substituted. In addition, the earthquake bedrock surface depth D28 is the bottom surface depth of the 28th layer of the deep ground model of the mesh that includes the target observation point position published by the Earthquake Hazard Station J-SHIS of the National Research Institute for Earth Science and Disaster Prevention. (equal to the top surface depth of the 29th layer with a P wave velocity of 5000 m/s and an S wave velocity of 2700 m/s).

The azimuth characteristic is, for example, the epicenter azimuth Λ, which indicates the azimuth in which the epicenter is located with respect to the observation point. Therefore, the epicenter azimuth Λ is calculated as the azimuth in which the epicenter position exists based on the observation point position. At this time, the epicenter azimuth Λ is determined clockwise with true north as 0°, and is a discontinuous amount with the true north as the boundary. Use pairs.

The seismic motion observation characteristic is, for example, the directional component Comp of the seismic motion indicating that the time history waveform data recorded as the seismic motion observation record is in one of the north-south direction, east-west direction, and up-and-down direction.

The seismic motion observation record is, for example, a time history waveform of three components of acceleration, velocity, and displacement in the north-south direction, east-west direction, and up-and-down direction. Seismic motion observation records are evaluated and analyzed using predetermined evaluation methods to obtain various seismic motion indicators. The seismic motion observation records according to this embodiment include 138 observation points installed in six prefectures in the Kanto region (Tokyo, Kanagawa, Chiba, Saitama, Ibaraki, Tochigi, and Gunma) of the strong motion observation network K-NET (Figure 1 This is a time history waveform for two horizontal components (north-south direction and east-west direction) observed at

The seismic motion index includes at least one of an amplitude characteristic, a periodic characteristic, and a temporal characteristic of the seismic motion.

The amplitude characteristic is at least one of maximum acceleration PGA, maximum velocity, and maximum displacement of earthquake motion obtained from earthquake motion observation records (time history waveforms of acceleration, velocity, and displacement). The amplitude characteristic according to this embodiment is the maximum acceleration PGA.

The period characteristic is a response value for at least one period in a response spectrum or Fourier spectrum obtained from seismic motion observation records (time history waveforms of acceleration, velocity, and displacement). The response spectrum is, for example, an acceleration response spectrum, a pseudo velocity response spectrum pSv, a velocity response spectrum, a displacement response spectrum, etc. for a predetermined damping constant (eg, 5%). Examples of the Fourier spectrum include an acceleration Fourier spectrum, a velocity Fourier spectrum, and a displacement Fourier spectrum. The periodic characteristics according to this embodiment are pseudo velocity response spectra pSv(0.1s) and pSv(0 .5s), pSv (1.0s), pSv (3.0s), and pSv (5.0s).

The temporal characteristic is, for example, a response duration for at least one period in a response duration spectrum obtained from seismic motion observation records (time history waveforms of acceleration, velocity, and displacement). The response duration spectrum is, for example, an acceleration response duration spectrum, a velocity response duration spectrum TSv, a displacement response duration spectrum, etc. for a predetermined damping constant (for example, 5%). The temporal characteristics according to this embodiment are speed response duration spectra TSv (0.1 s), TSv ( 0.5s), TSv (1.0s), TSv (3.0s), and TSv (5.0s). Note that the parameters that define the start and end of the response duration are p1=0.03 and p2=0.95.

(About the acquisition process by the acquisition unit 311 and the learning data 12)
As shown in FIG. 3, the acquisition unit 311 extracts feature quantities and seismic motion parameters from a plurality of pieces of seismic motion data 11 registered in the database 10, using seismic motion characteristic parameters as feature quantities and seismic motion indices obtained from seismic motion observation records as objective variables. A plurality of learning data 12 composed of objective variables are acquired. Note that the learning data 12 is data used as learning data (training data), verification data, and test data in supervised learning. Further, a set of learning data 12 made up of a plurality of learning data 12 is referred to as a learning data set.

At this time, the acquisition unit 311 selects, as the learning data 12, seismic motion data 11 based on an earthquake that satisfies at least one of a predetermined earthquake condition and a predetermined seismic motion condition. It is preferable that the seismic conditions and seismic motion conditions are set such that, for example, seismic motions with relatively large amplitudes are distributed in a well-balanced manner in the metropolitan area.

The predetermined earthquake conditions include the first to sixth earthquake conditions shown below. For example, the first earthquake condition is an earthquake that occurred during the period from 1996 to January 15, 2019, with a maximum seismic intensity of 4 or higher and a seismic intensity of 2 or higher in Chiyoda Ward, Tokyo. After excluding earthquakes where the number of locations where seismic intensities were recorded in the Kanto region is very small (about single digits), the second earthquake condition is that the bias in the epicenter position and the bias in the earthquake occurrence area are below a predetermined threshold. shall be. Regarding the earthquake scale (magnitude Mj according to the Japan Meteorological Agency), it is found that the earthquake is balanced when separated into three stages: 3.0<Mj≦5.0, 5.0<Mj≦5.5, and 5.5<Mj. This is the third earthquake condition. When multiple earthquakes occur in the same earthquake-generating area, the fourth earthquake condition is to prevent deviations in the epicenter depth from occurring. The fifth earthquake condition is that the moment magnitude Mw is determined by the broadband seismic observation network F-net. The sixth earthquake condition is to exclude earthquakes with extremely wide fault planes (for example, the 2011 Tohoku Pacific Coast Earthquake (Mj9.0)). A total of 74 earthquakes meet the above-mentioned first to sixth earthquake conditions.

Examples of the predetermined seismic motion conditions include the first to fifth seismic motion conditions shown below. For example, the first seismic motion condition is that the maximum acceleration of the three-component composite value when the three-component accelerations are combined is 1 cm/s ² or more. The second seismic motion condition is to exclude those with small amplitude and large noise (those with a small S/N ratio) in the seismic motion observation records. The third seismic motion condition is to exclude earthquake motion observation records for which there is insufficient recording time (insufficient head, insufficient tail, or insufficient both). The fourth seismic motion condition is to exclude cases in which different earthquakes, different events, etc. are mixed in the seismic motion observation records, making it difficult to analyze the records. Regarding the response spectrum and Fourier spectrum in the seismic motion index, the fifth seismic motion condition is to exclude those with large noise mainly in the band with a period of 0.05 to 10 seconds.

The acquisition unit 311 selects the seismic motion data 11 that satisfies the first seismic motion condition from among the seismic motions caused by a total of 74 earthquakes that satisfy the first to sixth earthquake conditions, whereby the learning selected by the acquisition unit 311 is performed. The set of training data 12 (hereinafter referred to as "learning data set A") consists of a total of 14104 (7052 records x two horizontal components) seismic motion data 11. In addition, the acquisition unit 311 selects the seismic motion data 11 that satisfies the first to fifth earthquake motion conditions from the earthquake motions caused by a total of 74 earthquakes that satisfy the first to sixth earthquake conditions. The set of learning data 12 selected by (hereinafter referred to as "learning data set B") consists of a total of 11,488 (5744 records x two horizontal components) seismic motion data 11.

Below, the feature quantities constituting the learning data 12 according to the present embodiment are moment magnitude Mw, epicenter depth H, epicenter distance X, and epicenter azimuth Λ( The following description assumes that there are six types of seismic motion characteristic parameters: the pair of sinΛ and cosΛ), the surface layer average S-wave velocity AVS30, and the earthquake basement depth (S-wave velocity 2700 m/s layer top depth) D28.

Further, the objective variables constituting the learning data 12 according to this embodiment are the 11 types of seismic motion indicators shown in FIG. ), pSv (1.0s), pSv (3.0s), pSv (5.0s), speed response duration spectrum TSv (0.1s), TSv (0.5s), TSv (1.0s), TSv (3.0s) and TSv (5.0s).

FIG. 5 is a scatter diagram showing the correlation between each feature of the set of learning data 12 (learning data set A). The characteristics of the training dataset A are that there are few records of small-scale earthquakes in the distant area, that there are no records of large-scale earthquakes in the vicinity, that most of the earthquakes are distributed at shallower epicenter depths of 100 km or less, and that there are few records of small earthquakes in the Kanto region. Although the epicenter azimuth Λ for each observation point is dispersed in all directions, one reason is that there are many cases in the northeast direction (epicenter azimuth Λ≒45°).

(Regarding the generation process by the generation unit 312 and the seismic motion evaluation model 13)
As shown in FIG. 3, the generation unit 312 generates a learned model by learning the correlation between the feature amounts and the objective variable by machine learning based on the plurality of learning data 12 acquired by the acquisition unit 311. A seismic motion evaluation model 13 is generated and stored in the storage unit 30. In this embodiment, a case will be described in which a gradient boosting decision tree is used as a machine learning algorithm in machine learning.

FIG. 6 is a schematic diagram showing an overview of a gradient boosting tree. A gradient boosting tree is a learning device that combines gradient boosting and decision trees. Gradient boosting is a method of building a strong learner (high-performance machine learning model) by combining multiple weak learners (low-performance machine learning models). Decision trees are a method of generating machine learning models that can perform classification and regression by performing conditional branching using the branching structure of trees. A gradient boosting tree, which is a combination of these two methods, is a method that combines multiple weak learners generated by a decision tree using gradient boosting.

In this embodiment, among the 11 types of seismic motion indicators, the number of data for the maximum acceleration PGA and the pseudo velocity response spectrum pSv is considered to decrease rapidly as the amplitude increases. Therefore, in this embodiment, in order to reduce the bias that occurs in the distribution of the objective variables, common logarithms (log ₁₀ PGA and log ₁₀ pSv) are used as the objective variable data for the maximum acceleration PGA and the pseudo velocity response spectrum pSv, respectively. It was to be. Additionally, as a loss function in the gradient boosting tree, we decided to apply the least squares method (normal distribution) to the maximum acceleration PGA and the pseudo velocity response spectrum pSv, and to apply the Poisson distribution to the velocity response duration spectrum TSv. .

Furthermore, in this embodiment, when generating the seismic motion evaluation model 13 using the gradient boosting tree, the generation unit 312 generates the seismic motion evaluation model 13 for each target variable (11 types of seismic motion indicators). That is, the first seismic motion evaluation model 13A has learned the correlation between the feature quantities (six types of seismic motion characteristic parameters) and the objective variable (maximum acceleration PGA = first type of seismic motion index); The third earthquake motion evaluation model 13B learns the correlation between the earthquake motion characteristic parameters) and the objective variable (pseudo velocity response spectrum pSv (0.1 s) = second type of earthquake motion index). Each of the evaluation models 13C to 11th seismic motion evaluation model 13K was generated, and a total of 11 seismic motion evaluation models 13A to 13K were generated. In the following description, it is assumed that the seismic motion evaluation model 13 includes a total of 11 seismic motion evaluation models 13A to 13K.

(About the configuration of the seismic motion evaluation device 4 and the processes involved in each part)
The seismic motion evaluation device 4 evaluates and predicts seismic motion based on the seismic motion evaluation model 13 generated by the seismic motion evaluation model generation device 3, and outputs the results to an output medium such as a display medium or a paper medium.

The seismic motion evaluation device 4, like the seismic motion evaluation model generation device 3, is composed of a general-purpose or dedicated computer, and as shown in FIG. a control section 41 constituted by a processor such as, a communication section 42 which is a communication interface with the network 5, an input section 43 constituted by a keyboard, a mouse, etc., and a display section 44 constituted by a display, a touch panel, etc. Equipped with.

The storage unit 40 stores the seismic motion evaluation model 13 generated as a learned model by the seismic motion evaluation model generation device 3, and the seismic motion evaluation program 400 that controls the operation of the seismic motion evaluation device 4 to realize the seismic motion evaluation method. There is.

The control unit 41 functions as a reception unit 410, a prediction unit 411, and an output processing unit 412 by executing the seismic motion evaluation program 400. Note that the details of the functions of each part will be described later.

FIG. 7 is a functional explanatory diagram showing an example of the seismic motion evaluation device 4 and the seismic motion evaluation method according to the first embodiment of the present invention.

(Regarding the reception process by the reception unit 410)
The receiving unit 410 receives various characteristic parameters of seismic motion to be predicted. Specifically, the reception unit 410 receives, for example, the moment magnitude Mw, the epicenter position, and the epicenter depth H of an earthquake that the user of the seismic motion evaluation device 4 envisions as a prediction target (hereinafter referred to as a "supposed earthquake"). is received via the input unit 33, and also receives via the input unit 33 a prediction point position indicating the position of a prediction point at which it is desired to predict how much seismic motion will occur due to the hypothetical earthquake. Note that the predicted point position may be an arbitrary position or may be the same as the observation point position. Further, the predicted point positions may be plural, for example, each grid point at a predetermined grid interval (for example, 5 km interval).

Then, the reception unit 410 calculates the epicenter distance Based on this, the surface layer 30 m average S-wave velocity AVS30 and the earthquake bedrock depth (top surface depth of the S-wave velocity 2700 m/s layer) D28 are obtained for the predicted point position of the earthquake motion. Thereby, the reception unit 410 receives various earthquake motion characteristic parameters to be predicted (moment magnitude Mw of the assumed earthquake, epicenter depth H of the assumed earthquake, epicenter distance X between the hypocenter of the assumed earthquake and the predicted point position, predicted point position , the epicenter azimuth Λ indicating the direction in which the hypocenter position of the assumed earthquake exists, the surface layer 30m average S-wave velocity AVS30 at the predicted point position, and the earthquake bedrock depth D28) at the predicted point position are accepted.

(About the prediction process by the prediction unit 411)
The prediction unit 411 uses the earthquake motion characteristic parameters of the prediction target received by the reception unit 410 as feature quantities, and uses the earthquake motion evaluation model 13 (6 types of earthquake motion characteristic parameters as feature quantities and 11 types of earthquake motion indicators as objective variables). Based on the target variables output from the seismic motion evaluation model 13 by inputting them into a trained model that has learned the correlation between the two, seismic motion indices corresponding to the various seismic motion characteristic parameters to be predicted are predicted. At this time, if the reception unit 410 receives multiple earthquake motion characteristic parameters as prediction targets by accepting multiple assumed earthquakes or multiple prediction point positions, the prediction unit 411 accepts multiple earthquake motion characteristic parameters as prediction targets. By inputting each of the various characteristic parameters into the seismic motion evaluation model 13, the seismic motion index corresponding to each of the plurality of seismic motion characteristic parameters is predicted.

In this embodiment, since the seismic motion evaluation model 13 includes a total of 11 seismic motion evaluation models 13A to 13K generated for each objective variable (11 types of seismic motion indicators), the prediction unit 411 The characteristic parameters are input as feature quantities to a total of 11 seismic motion evaluation models 13A to 13K, and each of the objective variables (11 types of seismic motion indicators) output from each of the 11 seismic motion evaluation models 13A to 13K is predicted. The seismic motion index corresponding to various seismic motion characteristic parameters is predicted. That is, the maximum acceleration PGA is predicted using the first earthquake motion evaluation model 13A that has learned the correlation between the feature quantities (six types of earthquake motion characteristic parameters) and the maximum acceleration PGA (=the first type of earthquake motion index). , using the second seismic motion evaluation model 13B that has learned the correlation between the feature quantities (six types of seismic motion characteristic parameters) and the pseudo velocity response spectrum pSv (0.1 s) (=the second type of seismic motion index). By predicting each objective variable using the third earthquake motion evaluation model 13C to the eleventh earthquake motion evaluation model 13K, such as predicting the pseudo velocity response spectrum pSv (0.1 s), a total of 11 Earthquake motion evaluation models 13A to 13K were used to predict 11 types of earthquake motion indicators.

(About the output processing process by the output processing unit 412)
The output processing unit 412 outputs the seismic motion index predicted by the prediction unit 411 to a visible output medium. For example, when the output medium is a display medium such as the display unit 44, the output processing unit 412 generates display data (output data) to be displayed on the display medium, and outputs the generated display data to the display medium. Further, when the output medium is a paper medium, the output processing unit 412 generates print data (output data) for printing on the paper medium, and prints out the print data on the paper medium. Note that the output processing unit 412 may output the output data to, for example, a seismic motion prediction map creation system or a hazard map creation system, or to a disaster prevention system for public facilities, buildings, factories, etc. You can do it like this.

For example, the output processing unit 412 may be configured such that the seismic motion characteristics parameters to be predicted received by the reception unit 410 are a plurality of seismic motion characteristics parameters for one hypothetical earthquake, such that each grid point is a predicted point position. When the seismic motion index is a characteristic parameter, for example, a contour diagram or a , output to the output medium as a color-coded mesh diagram.

In addition, when the seismic motion characteristic parameters to be predicted received by the reception section 410 are a plurality of seismic motion characteristic parameters including a plurality of different azimuth characteristics centered around the prediction point position, the output processing section 412 outputs The values of a plurality of seismic motion indices in each direction predicted by the prediction unit 411 based on the seismic motion characteristic parameters of By assigning them to each direction, multiple seismic motion indicators are output to the output medium.

Furthermore, when the seismic motion characteristic parameters to be predicted received by the receiving section 410 are a plurality of earthquake motion characteristic parameters including different epicenter distances, the output processing section 412 outputs a By expressing the values of the plural seismic motion indices predicted by the prediction unit 411 as distance attenuation characteristics, the plural seismic motion indices are outputted to an output medium as shown in FIG. 8 (details will be described later).

Figure 8 shows an example of the output results when representing the seismic motion index as a distance attenuation characteristic, where (a) is the maximum acceleration PGA according to the moment magnitude Mw, and (b) is the pseudo velocity response spectrum pSv according to the epicenter azimuth Λ. (1.0 s).

The output result shown in Figure 8(a) is output when the assumed earthquake is an earthquake with an epicenter in the Ibaraki area and the predicted point position is "SIT006 (Chichibu)" (see Figure 1). This assumes shaking at a rock site due to a plate boundary earthquake. In addition, the average of the existing distance attenuation formula based on Non-Patent Document 1 is also output for reference, but in general, similar to the existing distance attenuation formula, the seismic motion evaluation model 13 is used to calculate the seismic motion index. can be said to have been predicted.

The output results shown in Figure 8(b) indicate that the assumed earthquakes are plate boundary earthquakes in the northeast, south, and southwest directions (Λ = 45°, 160°, 230°) and northwest direction (Λ = 320°). ) is output when it is an inland earthquake. Comparing these four hypothetical earthquakes, even if the moment magnitude Mw and epicenter distance It is the smallest, and the size relationship between the northeast and southwest directions is reversed when the epicenter distance X is around 50 km. In addition, for inland earthquakes in the northwest direction, the slope of distance attenuation is relatively small, and when the epicenter distance It has become smaller. These differences are the result of regional characteristics of the epicenter characteristics and differences in propagation characteristics being reflected in the seismic motion evaluation model 13. Furthermore, by considering the epicenter orientation Λ, detailed regional characteristics are reflected in the seismic motion evaluation model 13. This is thought to be the result of

(About the characteristics of seismic motion evaluation model 13)
Next, the characteristics of the seismic motion evaluation model 13 will be explained while comparing a plurality of seismic motion evaluation models 13 each generated by the seismic motion evaluation model generation device 3 using different learning data 12. At that time, regarding the plurality of seismic motion data 11 in which various seismic motion characteristic parameters and seismic motion observation records are associated, the seismic motion index obtained from the seismic motion observation record is set as the "observed value", and the various seismic motion characteristic parameters are input into the seismic motion evaluation model 13. The characteristics of the seismic motion evaluation model 13 will be explained using the "observed value" and the "evaluation value" with the seismic motion index obtained by the "evaluation value".

Here, "Kanto model A" is generated by machine learning using training dataset A, and "Kanto model A6" considering all the above six types as feature quantities and epicenter azimuth Λ (sinΛ, cosΛ) "Kanto Model A5" which takes into account only five types excluding ``Kanto Model A5'' was generated and compared. In addition, "Kanto model B" is generated by machine learning using learning dataset B, and "Kanto model B6" that takes all the above six types into consideration as feature quantities and the epicenter direction Λ (sinΛ, cosΛ) "Kanto Model B5", which takes into account only the five types excluding the above, was generated and compared.

(About the influence of each feature on the objective variable)
Figure 9 shows the degree of influence of each feature (seismic motion characteristic parameters) on the objective variable (earthquake motion index), where (a) is Kanto model A5, (b) is Kanto model A6, (c) is Kanto model B5, (d) is a diagram showing the Kanto model B6.

The degree of influence of each feature shown in Figure 9 is determined by selecting one feature from the learning data 12 (learning data set A or learning data set B) used for machine learning and shuffling only that data string. This indicates the degree to which the evaluation accuracy deteriorates when the target variable is evaluated by replacing it with a random data string and leaving the data strings of other feature values as they are. Therefore, if the evaluation accuracy has significantly deteriorated, the feature amount is considered to be important, and conversely, if the evaluation accuracy has not changed, the feature amount is considered to have a small influence on the evaluation.

Comparing the Kanto model A5 shown in FIG. 9(a) and the Kanto model A6 shown in FIG. 9(b), the influence of the epicenter distance X is dominant in the short period, and the influence of the moment magnitude Mw is dominant in the long period. The degree of influence of moment magnitude Mw generally increased with period, except for the velocity response duration spectrum TSv (0.1 s), and the degree of influence of epicenter distance X decreased with period. The degree of influence of the epicenter depth H is relatively large on the short-period side, the degree of influence of the average S-wave velocity AVS30 in the surface layer 30m is relatively large in the velocity response duration spectrum TSv (1.0 s), and the degree of influence on the earthquake basement depth is relatively large. The degree of influence of the speed D28 was large when the period of the speed response duration spectrum TSv was 1 second or more. The degree of influence of the epicenter azimuth Λ is comparable to the epicenter depth H, the earthquake bedrock depth D28, and the average S-wave velocity AVS30 in the surface layer when sinΛ and cosΛ are considered together, and relatively, the period is less than 1 second. It became larger. Overall, it appears that the characteristics differ between the short-period side and the long-period side, with the period around 1 second as the boundary.

Comparing the Kanto model B5 shown in FIG. 9(c) and the Kanto model B6 shown in FIG. 9(d), the effect on the speed response duration spectrum TSv is easier to understand than the results of the prototype Kanto model A. The degree of influence of the moment magnitude Mw increased with the period, and the degree of influence of the epicenter distance X and the hypocenter depth H decreased with the period. The degree of influence of surface layer 30m average S-wave velocity AVS30 is relatively large with a period of 1 second, the degree of influence of earthquake basement depth D28 is large with a period of 1 second or more, and the degree of influence of epicenter direction Λ is relatively large. became large when the period was less than 1 second.

(About prediction accuracy of objective variable)
FIG. 10 is a scatter diagram showing the correlation between observed values obtained from seismic motion observation records and evaluation values obtained by the seismic motion evaluation model 13 for each target variable (seismic motion index) in Kanto model A6.

In Fig. 10, the maximum acceleration PGA, pseudo speed response spectra pSv (1.0s), pSv (3.0s), and speed response duration spectra TSv (1.0s), TSv (3.0s) are compared as objective variables. did. In both cases, the horizontal axis is the observed value, and the vertical axis is the evaluated value. Of the 12 learning data, the learning data used for machine learning is 64% of the total data, and the validation data used for model verification is 36% of the total data, and the evaluation value is equal to the observed value. The case is shown by a solid line, and the case where the observed value is double or half is shown by a broken line. The observed values were generally evaluated and modeled very well, with most of the estimated values falling within a factor of two and a half of the observed values. Even in light of conventional knowledge regarding seismic motion, this evaluation accuracy is considered to be high.

FIG. 11 shows the "evaluation value/observed value" ratio of each target variable (earthquake motion index) as a histogram, in which (a) shows the Kanto model A5, and (b) shows the Kanto model A6. It was found that the values of the "evaluation value/observed value" ratio had a regular distribution centered on 1, and the variation in the speed response duration spectrum TSv was smaller than that in the maximum acceleration PGA and the pseudo speed response spectrum pSv.

・
FIG. 12 is a diagram showing the average μ and variation σ of the “evaluation value/observation value” ratio of each objective variable (earthquake motion index) for Kanto models A5, A6, B5, and B6.

The average of the "evaluation value/observation value" ratio is approximately 1, and the common logarithm standard deviation of the "evaluation value/observation value" ratio is 0.17 to 0.0 for the maximum acceleration PGA and the pseudo velocity response spectrum pSv. 20. The speed response duration spectrum TSv was 0.10 to 0.12. The accuracy of this evaluation is high even in light of conventional knowledge regarding seismic motion. When the epicenter orientation Λ is taken into account as a feature (Kanto model A5 → A6), the average and dispersion of each objective variable are slightly improved, and it is quite important to consider the differences in ground motion characteristics due to the epicenter orientation Λ. It was also confirmed that Furthermore, when learning data set B was used instead of learning data set A (Kanto model A6→B6), the average and dispersion of the speed response duration spectrum TSv were slightly improved.

As described above, according to the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method according to the present embodiment, it is possible to make maximum use of the increase in the amount of data and the improvement in the information processing ability of computers, while predicting future occurrences. It is possible to provide a seismic motion evaluation model 13 that can evaluate and predict the characteristics of seismic motion caused by an earthquake with high accuracy.

Further, according to the seismic motion evaluation device 4 and the seismic motion evaluation method according to the present embodiment, by using the seismic motion evaluation model 13 generated by the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method, individual experts can It is possible to evaluate and predict with high precision the characteristics of seismic motion caused by earthquakes that are expected to occur in the future, without relying on experience.

Furthermore, if the seismic motion characteristic parameters include at least the azimuth characteristics of the seismic motion, the correlation between the azimuth characteristics of the seismic motion and the seismic motion index is learned by machine learning, so that the underground structure around the site differs depending on the direction. Since the influence on earthquake motion is reflected in the earthquake motion evaluation model 13, it is possible to provide the earthquake motion evaluation model 13 that can evaluate and predict the characteristics of earthquake motion with high accuracy.

In addition, if the seismic motion index includes at least the temporal characteristics of seismic motion, the correlation between various seismic motion characteristic parameters and the temporal characteristics of seismic motion is learned by machine learning, and the temporal characteristics of seismic motion due to earthquakes that are expected to occur in the future are learned by machine learning. It is possible to provide the seismic motion evaluation model 13 that can evaluate and predict with high precision the temporal characteristics of seismic motion as one of the characteristics of seismic motion without defining a prediction formula for evaluating and predicting.

(Second embodiment)
In the seismic motion evaluation system 1 according to the first embodiment, the seismic motion evaluation model generation device 3 generates one seismic motion evaluation model 13 (for example, Kanto models A5, A6, B5, B6) are generated, and the seismic motion evaluation device 4 uses one seismic motion evaluation model 13 to evaluate and predict seismic motion in the Kanto region. On the other hand, in the seismic motion evaluation system 1 according to the second embodiment, the seismic motion evaluation model generation device 3 generates the seismic motion evaluation model 13 for each observation point, and the seismic motion evaluation device 4 generates the seismic motion evaluation model for each observation point. The difference is that earthquake motion at each observation point is evaluated and predicted using 13. Other basic configurations and operations are similar to those of the first embodiment, so the following description will focus on the differences between the two.

(About seismic motion evaluation model generation device 3)
In this embodiment, the seismic motion evaluation model generation device 3 selects "SIT006 (Chichibu)" (Fig. 1 (see Fig. 1) and "TKY028 (Etchujima)" (see Fig. 1), which is an example of a site with deep foundations and soft surface ground, will be explained as generating seismic motion evaluation models 13 respectively. Here, the seismic motion evaluation model 13 at SIT006 (Chichibu) is referred to as "point model S", the seismic motion evaluation model 13 at TKY028 (Etchujima) is referred to as "point model T", and the set of learning data 12 necessary for machine learning of these is referred to as "point model S". These will be referred to as "learning data set S" and "learning data set T", respectively.

FIG. 13 is a functional explanatory diagram showing an example of the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method according to the second embodiment of the present invention.

The acquisition unit 311 acquires earthquakes that occurred during the period from 1996 to May 31, 2019, from a plurality of seismic motion data 11 registered in the database 10, as predetermined earthquake conditions, such as SIT006 (Chichibu) or TKY028 ( The maximum acceleration of the three-component composite value is 1 cm/s ² or more (first seismic motion condition), based on an earthquake that satisfies the observation of seismic motion on Etchu Island), and as an earthquake motion that satisfies the predetermined seismic motion conditions. And, the seismic motion data 11 based on the seismic motion that satisfies the fact that the moment magnitude Mw is obtained by the broadband seismic observation network F-net (fifth seismic motion condition) is used as the learning data 12 (learning data sets S, T). select. As a result, the learning data set S consists of a total of 1468 (734 records x 2 horizontal components) seismic motion data 11, and the learning data set T consists of a total of 1314 (657 earthquakes x 2 components) seismic motion data 11.

At this time, since the point models S and T are point-specific seismic motion evaluation models 13, the feature quantities that make up the learning data 12 are the surface These are four types of earthquake motion characteristic parameters, excluding the 30m average S-wave velocity AVS30 and the earthquake bedrock depth D28, the moment magnitude Mw, the epicenter depth H, the epicenter distance Note that the objective variables constituting the learning data 12 are the above-mentioned 11 types of seismic motion indicators, as in the first embodiment.

FIG. 14 is a scatter diagram showing the correlation between each feature of the set of learning data 12, in which (a) shows the learning data set S, and (b) shows the learning data set T. Training datasets S and T have less data on near-field earthquakes compared to training dataset A shown in Figure 5, but other than that point, they have qualitatively the same tendency as training dataset A. be.

As shown in FIG. 13, the generation unit 312 generates feature quantities (four types of earthquake motion characteristic parameters) based on the plurality of learning data 12 (learning data sets S, T) acquired by the acquisition unit 311. A seismic motion evaluation model 13 (point models S4, T4) is generated for each observation point by learning the correlation between the target variables (11 types of seismic motion indicators) using machine learning.

(About seismic motion evaluation device 4)
FIG. 15 is a functional explanatory diagram showing an example of the seismic motion evaluation device 4 and seismic motion evaluation method according to the second embodiment of the present invention.

The reception unit 410 receives various earthquake motion characteristic parameters to be predicted (the moment magnitude Mw of the assumed earthquake, the epicenter depth H of the assumed earthquake, the epicenter distance X between the hypocenter of the assumed earthquake and the predicted point position, and the predicted point position). As a reference, the epicenter direction Λ) indicating the direction in which the epicenter position of the assumed earthquake exists is accepted. In this embodiment, the predicted point position is SIT006 (Chichibu) or TKY028 (Etchujima).

The prediction unit 411 inputs the seismic motion characteristic parameters of the prediction target received by the reception unit 410 as feature quantities to the seismic motion evaluation model 13 (point models S4, T4) according to the prediction point position, thereby performing the seismic motion evaluation. Based on the target variables output from the model 13, seismic motion indicators corresponding to the seismic motion characteristic parameters to be predicted are predicted. At this time, the prediction unit 411 uses the point model S4 when the predicted point position is SIT006 (Chichibu), and uses the point model T4 when the predicted point position is TKY028 (Etchujima).

The output processing unit 412 outputs the seismic motion index predicted by the prediction unit 411 to a visible output medium. For example, when the seismic motion characteristic parameters to be predicted received by the reception section 410 are a plurality of seismic motion characteristic parameters including a plurality of different azimuth characteristics centered around the prediction point position, the output processing section 412 outputs the The values of a plurality of seismic motion indices in each direction predicted by the prediction unit 411 based on the seismic motion characteristic parameters of As shown in FIG. 16 (details will be described later), a plurality of seismic motion indices are outputted to an output medium by assigning them to each direction.

FIG. 16 shows an example of the output results when representing the seismic motion index (pseudo velocity response spectrum pSv and velocity response duration spectrum TSv) as azimuth characteristics, where (a) is the point model S4, and (b) is the point model T4. FIG.

The output results shown in Figure 16 are based on multiple seismic motion characteristic parameters obtained by changing the epicenter azimuth Λ at 5° intervals under the conditions of moment magnitude Mw = 6, epicenter depth H = 10 km, and epicenter distance X = 180 km. The values of pseudo velocity response spectra pSv at 5° intervals predicted based on the above are output. In FIG. 16(a), the value of the pseudo velocity response spectrum pSv in each direction is standardized by the maximum value with SIT006 (Chichibu) as the reference (center), and then expressed as the distance from the center. In FIG. 16(b), the value of the pseudo velocity response spectrum pSv in each direction with TKO0028 (Etchujima) as the reference (center) is standardized by the maximum value, and then expressed as the distance from the center.

(About the characteristics of the seismic motion evaluation model 13 for each location)
Next, regarding the characteristics of the seismic motion evaluation model 13 for each location, we will discuss the location model S4 generated by machine learning using the learning dataset S, and the location model T4 generated by machine learning using the learning dataset T. This will be explained by comparing.

FIG. 17 shows the degree of influence of each feature quantity (earthquake motion characteristic parameters) on the objective variable (earthquake motion index), in which (a) shows the point model S4, and (b) shows the point model T4.

In the point model S4 shown in FIG. 17(a), the degree of influence on the maximum acceleration PGA and the short-period pseudo velocity response spectrum pSv is that the moment magnitude Mw and the epicenter distance X are equivalent, and the influence of the epicenter distance X is slightly greater. However, apart from that, the moment magnitude Mw has the greatest influence. The degree of influence of the epicenter azimuth Λ is greater on the shorter period side. The degree of influence of the epicenter azimuth Λ is larger on the velocity response duration spectrum TSv than on the pseudo velocity response spectrum pSv, exceeds the epicenter depth H, and is equal to or greater than the epicenter distance X, and the Kanto models A5 and A6 shown in FIG. This result greatly exceeds that of . The influence of the epicenter depth H is small.

In the point model T4 shown in FIG. 17(b), the influence of the epicenter distance X is dominant in the short period, and the influence of the moment magnitude Mw is dominant in other periods. The degree of influence of the moment magnitude Mw generally increased with the period, and the degree of influence of the epicenter distance X decreased with the period. The degree of influence of the epicenter azimuth Λ is larger on the short-period side; in particular, the speed response duration spectrum TSv is larger than the maximum acceleration PGA and the pseudo-velocity response spectrum pSv, and for periods other than 0.1 seconds, it exceeds the epicenter distance On the other hand, the moment magnitude exceeded Mw. The influence of the epicenter depth H is small. In the model T4 at a deep point in the foundation, the degree of influence of the moment magnitude Mw on the short-period velocity response duration spectrum TSv is relatively small compared to the model S4 in a shallow point in the foundation.

Figure 18 shows the correlation between the observed values obtained from the seismic motion observation records and the evaluation values obtained by the seismic motion evaluation model 13 for each objective variable (seismic motion index), where (a) is the point model S4, and (b) is the It is a scatter diagram showing each point model T4. The observed values were generally evaluated and modeled very well, with most of the estimated values falling within a factor of two and a half of the observed values.

FIG. 19 shows the "evaluation value/observed value" ratio of each target variable (earthquake motion index) as a histogram, in which (a) shows the point model S4, and (b) shows the point model T4. It was found that the values of the "evaluation value/observed value" ratio had a regular distribution centered on 1, and the variation in the speed response duration spectrum TSv was smaller than that in the maximum acceleration PGA and the pseudo speed response spectrum pSv.

FIG. 20 is a diagram showing the average and dispersion of the "evaluation value/observation value" ratio of each target variable (earthquake motion index) for the point models S4 and T4. The average of the "evaluation value/observation value" ratio is approximately 1, and the common log standard deviation of the "evaluation value/observation value" ratio is approximately 0.2 or more for the maximum acceleration PGA and the pseudo velocity response spectrum pSv. , the speed response duration spectrum TSv was a little over 0.1.

As described above, according to the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method according to the present embodiment, the correlation between various seismic motion characteristic parameters and seismic motion indices is learned for each observation point by machine learning. Since uncertain factors that affect seismic motion due to differences in the underground structure at each observation point are excluded, it is possible to provide the seismic motion evaluation model 13 that can evaluate and predict the characteristics of seismic motion with high accuracy.

Further, according to the seismic motion evaluation device 4 and the seismic motion evaluation method according to the present embodiment, by using the seismic motion evaluation model 13 for each observation point generated by the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method, individual It is possible to evaluate and predict with high precision the characteristics of seismic motion caused by earthquakes that are expected to occur in the future, without relying on the experience of experts.

(Third embodiment)
In the seismic motion evaluation system 1 according to the second embodiment, the seismic motion evaluation model generation device 3 generates the seismic motion evaluation model 13 (for example, point models S4, T4) for each observation point, and the seismic motion evaluation device 4 The seismic motion evaluation model 13 for each observation point is used to evaluate and predict the seismic motion at each observation point. In contrast, in the seismic motion evaluation system 1 according to the third embodiment, the seismic motion evaluation model generation device 3 generates the seismic motion evaluation model 13 for each observation point group in which a plurality of observation points are grouped according to predetermined classification criteria. However, the difference is that the seismic motion evaluation device 4 evaluates and predicts seismic motion in each observation point group using a seismic motion evaluation model 13 for each observation point group. Other basic configurations and operations are similar to those of the first and second embodiments, so the following description will focus on the differences between the two.

FIG. 21 is a schematic diagram showing the epicenter characteristics, propagation characteristics, and site characteristics of seismic motion at a plurality of nearby observation points.

As in the case of the first to third observation points shown in Figure 21, cases where multiple observation points are close to each other, cases where the observation points are very close to each other, cases where the underground structure around the observation points is changing suddenly, etc. However, the source and propagation characteristics of seismic motion can be considered to be almost the same between multiple nearby observation points, and the seismic motion input to the foundation directly beneath each observation point is considered to be almost the same. . In addition, regarding site characteristics, it is considered possible to judge the similarity between multiple nearby observation points by comparing seismic motion observation records and seismic motion indices obtained from seismic motion observation records.

Therefore, the predetermined classification criteria include the first to third classification criteria shown below. For example, the first classification criterion is to classify a plurality of observation points into the same observation point group when the distance between the plurality of observation points is smaller than a predetermined reference value. The second classification criterion is to classify multiple observation points into the same observation point group when the degree of similarity when comparing subsurface structural parameters at each of the multiple observation points is higher than a predetermined standard value. It is. The third classification criterion is that each seismic motion index obtained from each seismic motion observation record (time history waveform data) or each seismic motion observation record observed at a plurality of observation points is, for example, GOF (Goodness of Fit). ), the plurality of observation points are classified into the same observation point group when they are more similar than a predetermined reference value.

FIG. 22 is a diagram showing a first observation point group C in which a plurality of observation points c1 to c3 are grouped and a second observation point group D in which a plurality of observation points d1 to d3 are grouped. be. By grouping the plurality of observation points c1 to c3 and d1 to d3 according to predetermined classification criteria, a first observation point group C and a second observation point group D are formed as shown in FIG. It is formed. Here, the seismic motion evaluation model 13 at the first observation point group C is referred to as "group model C," and the seismic motion evaluation model 13 at the second observation point group D is referred to as "group model D." The training data 12 will be referred to as "learning data set C" and "learning data set D."

FIG. 23 is a functional explanatory diagram showing an example of the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method according to the third embodiment of the present invention.

The acquisition unit 311 acquires, from the database 10, seismic motion data 11 based on earthquakes in which seismic motion was observed at each of the observation points c1 to c3 forming the first observation point group, as learning data 12 (learning data set C). Select as. The acquisition unit 311 also acquires, from the database 10, seismic motion data 11 based on earthquakes in which seismic motion was observed at each of the observation points d1 to d3 forming the second observation point group. D).

The generation unit 312 generates feature quantities (four types of earthquake motion characteristic parameters) and objective variables (eleven types of A seismic motion evaluation model 13 (group models C4, D4) is generated for each observation point group by learning the correlation between the seismic motion index) using machine learning.

FIG. 24 is a functional explanatory diagram showing an example of the seismic motion evaluation device 4 and seismic motion evaluation method according to the third embodiment of the present invention.

The prediction unit 411 inputs the earthquake motion characteristic parameters of the prediction target received by the reception unit 410 as feature quantities to the earthquake motion evaluation model 13 (group models C4 and D4) according to the prediction point position, thereby performing the earthquake motion evaluation. Based on the target variables output from the model 13, seismic motion indicators corresponding to the seismic motion characteristic parameters to be predicted are predicted.

At this time, the prediction unit 411 determines whether the prediction point position included in the earthquake motion characteristic parameters to be predicted is inside each observation point c1 to c3 or a line connecting each observation point c1 to c3 (for example, the prediction point position shown in FIG. 22). In the case of e1), group model C4 is used, and when the predicted point position is inside each observation point d1 to d3 or the line connecting each observation point d1 to d3 (for example, predicted point e2 shown in FIG. 22), uses group model D4. Note that if the predicted point position is between the first observation point group C and the second observation point group D, as shown by predicted point e3 in FIG. 22, the prediction is made using the group model C4. The seismic motion index predicted using the group model D4 and the seismic motion index predicted using the group model D4 may be averaged, or may be divided proportionally according to distance or similarity of underground structure parameters.

As described above, according to the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method according to the present embodiment, the correlation between various seismic motion characteristic parameters and seismic motion indicators is learned for each observation point group by machine learning. This makes it easy to secure learning data (teacher data) and reduces variation in the learning data, making it possible to provide a seismic motion evaluation model that can evaluate and predict seismic motion characteristics with high accuracy. can.

Further, according to the seismic motion evaluation device 4 and the seismic motion evaluation method according to the present embodiment, by using the seismic motion evaluation model for each observation point group generated by the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method, individual It is possible to evaluate and predict with high precision the characteristics of seismic motion caused by earthquakes that are expected to occur in the future, without relying on the experience of experts.

Furthermore, observation point groups are classified and grouped according to predetermined classification criteria (distance between observation points, similarity of subsurface structural parameters, etc.), thereby eliminating uncertain factors that may be caused by differences in subsurface structure on seismic motion. Therefore, it is possible to provide an earthquake motion evaluation model that can evaluate and predict earthquake motion characteristics with high accuracy.

(Fourth embodiment)
In the seismic motion evaluation system 1 according to the first embodiment, the seismic motion evaluation model generation device 3 acquires learning data 12 from the database 10 based on earthquakes that have occurred in the past, and generates the seismic motion evaluation model 13. On the other hand, in the seismic motion evaluation system 1 according to the fourth embodiment, the seismic motion evaluation model generation device 3 has a first database 10A based on actual earthquakes that occurred in the past, and a first database 10A based on virtual earthquakes calculated by simulation. The difference is that learning data 12 is acquired from the database 10B of No. 2, respectively, and a seismic motion evaluation model 13 is generated. Other basic configurations and operations are similar to those of the first embodiment, so the following description will focus on the differences between the two.

FIG. 25 is a block diagram showing an example of the seismic motion evaluation system 1 according to the fourth embodiment of the present invention.

The seismic motion simulation device 6 calculates a virtual earthquake and a virtual seismic motion due to the virtual earthquake by executing a simulation according to a predetermined simulation method, and provides the simulation conditions and simulation results to the outside as simulation data. At this time, any method can be adopted as the predetermined simulation method, and a plurality of methods may be adopted.

Note that when receiving a data request from the seismic motion evaluation device 4, the seismic motion simulation device 6 generates simulation data related to the request (if simulation conditions are received from the seismic motion evaluation device 4, the simulation is performed based on the simulation conditions). The seismic motion evaluation device 4 may be provided with simulation data including simulation results when executed. Furthermore, the seismic motion simulation device 6 may provide simulation data by executing a simulation on its own device, or may provide simulation data when a simulation is executed by another device.

The storage unit 30 of the seismic motion evaluation model generation device 3 includes, in addition to the seismic motion evaluation model 13 and the seismic motion evaluation model generation program 300, a first database in which provided data provided by the data providing devices 2A to 2C are registered and updated. 10A (corresponding to the database 10 according to the first embodiment), and a second database 10B in which simulation data provided by the seismic motion simulation device 6 is registered and updated.

By executing the seismic motion evaluation model generation program 300, the control unit 31 controls the first DB management unit 310A, the second DB management unit 310B, the first acquisition unit 311A, the second acquisition unit 311B, and the generation 312. The first DB management section 310A and the first acquisition section 311A correspond to the DB management section 310 and the acquisition section 311, respectively, according to the first embodiment.

FIG. 26 is a functional explanatory diagram showing an example of the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method according to the fourth embodiment of the present invention.

The first DB management unit 310A (first DB management process) stores various earthquake motion characteristic parameters and earthquake motion observation records when actual earthquake motion is observed, based on the provided data provided by the data providing devices 2A to 2C. It is associated and registered in the first database 10A. Therefore, in the first database 10A, for each combination of a plurality of actual earthquakes and at least one observation point where actual earthquake motions due to the plurality of actual earthquakes were observed, various seismic motion characteristics when the actual earthquake motions were observed are stored. A plurality of seismic motion data 11 in which parameters and seismic motion observation records are associated are registered and stored.

The first acquisition unit 311A (first acquisition step) uses earthquake motion characteristic parameters as feature quantities from a plurality of earthquake motion data 11 registered in the first database 10A, and uses earthquake motion indexes obtained from earthquake motion observation records as feature quantities. As variables, a plurality of first learning data 12A composed of feature amounts and objective variables are acquired.

The second DB management unit 310B (second DB management process) calculates various earthquake motion characteristic parameters and earthquake motion calculation results when the virtual earthquake motion is calculated by simulation, based on the simulation data provided by the earthquake motion simulation device 6. It is associated and registered in the second database 10B. Therefore, in the second database 10B, for each combination of a plurality of virtual earthquakes and at least one virtual observation point where the virtual earthquake motion due to the plurality of virtual earthquakes is calculated as the simulation seismic motion calculation result, the virtual earthquake motion is simulated. A plurality of pieces of seismic motion data 11 in which various seismic motion characteristic parameters and seismic motion calculation results calculated by are associated are registered and stored.

The second acquisition unit 311B (second acquisition step) uses various earthquake motion characteristic parameters as feature quantities from the plurality of earthquake motion data 11 registered in the second database 10B, and uses earthquake motion indices obtained from the earthquake motion calculation results as the purpose. As variables, a plurality of second learning data 12B composed of feature amounts and objective variables are acquired.

At this time, the second acquisition unit 311B acquires the second learning data 12B from the second database 10B so that the distribution of the seismic motion characteristic parameters of the actual earthquake motion is interpolated by the distribution of the seismic motion characteristic parameters of the virtual earthquake motion. It is preferable to obtain more than one.

FIG. 27 is a scatter diagram showing the interpolation relationship between the first learning data 12A and the second learning data 12B. The scatter diagram in FIG. 27 shows the correlation between moment magnitude Mw (horizontal axis) and epicenter distance This shows that.

In the first learning data 12A acquired from the first database 10A, the second acquisition unit 311B detects data-deficient portions (for example, areas surrounded by broken lines in FIG. The seismic motion characteristic parameters of the inner part) are searched to see if they are registered in the second database 10B, and a second learning set is created based on the seismic motion characteristic parameters of the hypothetical earthquake motion corresponding to the data-missing part. Data 12B is obtained from the second database 10B. At that time, if the second learning data 12B corresponding to the data missing part is not registered in the second database 10B, the second acquisition unit 311B sets simulation conditions based on the data missing part. The second learning data 12B corresponding to the missing data portion may be obtained from the seismic motion simulation device 6 by requesting the seismic motion simulation device 6 to do so. Although the distribution of the seismic motion characteristic parameters is based on the moment magnitude Mw and the epicenter distance X in the example shown in FIG. 27, the distribution is not limited to this. , it is sufficient if the distribution has at least two characteristics as variables.

Furthermore, the seismic motion data 11 based on a plurality of simulation methods may be registered in the second database 10B, and in that case, the second acquisition unit 311B may register seismic motion data 11 based on a plurality of simulation methods A plurality of pieces of second learning data 12B may be acquired from the registered second database 10B. At that time, the second acquisition unit 311B may acquire the second learning data 12B so as to interpolate the distribution of seismic motion characteristic parameters between each simulation method, or The performance may be compared and the earthquake motion data 11 based on a simulation method suitable for various earthquake motion characteristic parameters may be selected as the second learning data 12B.

As shown in FIG. 26, the generation unit 312 (generation process) generates a plurality of pieces of first learning data 12A acquired by the first acquisition unit 311A and a plurality of pieces of first learning data 12A acquired by the second acquisition unit 311B. The seismic motion evaluation model 13 is generated as a learned model by learning the correlation between the feature amount and the objective variable by machine learning based on the second learning data 12B.

The seismic motion evaluation device 4 evaluates and predicts seismic motion based on the seismic motion evaluation model 13 generated by the seismic motion evaluation model generation device 3, and outputs the results to an output medium. Note that the seismic motion evaluation device 4 is configured similarly to the first embodiment, so detailed explanation will be omitted.

As described above, according to the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method according to the present embodiment, the first learning data 12A based on actual earthquakes that occurred in the past and the virtual earthquake calculated by simulation Correlations between various seismic motion characteristic parameters and seismic motion indices are learned by machine learning using the second learning data 12B based on . This makes it easy to secure learning data (teacher data) and reduces bias in the learning data, thereby providing an earthquake motion evaluation model 13 that can evaluate and predict the characteristics of earthquake motion with high accuracy. be able to.

Further, according to the seismic motion evaluation device 4 and the seismic motion evaluation method according to the present embodiment, by using the seismic motion evaluation model 13 generated by the seismic motion evaluation model generation device 3 and the seismic motion evaluation model generation method, individual experts can It is possible to evaluate and predict with high precision the characteristics of seismic motion caused by earthquakes that are expected to occur in the future, without relying on experience.

Furthermore, the second learning data 12B is selected so as to complement earthquake motion characteristic parameters that have not occurred as an actual earthquake in the past and cannot be prepared with the first learning data 12A. Since the bias in the distribution of the learning data is reduced, it is possible to provide the seismic motion evaluation model 13 that can evaluate and predict the characteristics of seismic motion with high accuracy.

(Other embodiments)
Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and can be modified as appropriate without departing from the technical idea of the present invention.

For example, in the second embodiment, the seismic motion evaluation model generation device 3 creates two point models S4 and T4, and in the third embodiment, the seismic motion evaluation model generation device 3 creates two group models C4, The explanation has been made assuming that D4 is created. On the other hand, the seismic motion evaluation model generation device 3 may generate the seismic motion evaluation model 13 for each of three or more observation points or for each group of three or more observation points. Further, the fourth embodiment may be combined with the second embodiment or the third embodiment, and the seismic motion evaluation model generation device 3 uses the first database 10A based on actual earthquakes that occurred in the past, and the simulation The learning data 12 may be obtained from the second database 10B based on the virtual earthquake calculated by the method, and the seismic motion evaluation model 13 may be generated for each observation point or for each observation point group.

When a new earthquake occurs and a seismic motion observation record is observed, the seismic motion evaluation model generation device 3 according to each of the above embodiments updates the database 10 with seismic motion data 11 based on the new earthquake, and updates the database 10 with seismic motion data 11 based on the new earthquake. The seismic motion evaluation model 13 may be automatically regenerated based on the database 10. Then, the seismic motion evaluation device 4 according to each of the embodiments described above may automatically re-evaluate and re-predict the characteristics of the seismic motion based on the regenerated seismic motion evaluation model 13.

(About feature selection)
In each of the above embodiments, the feature quantities constituting the learning data 12 are described as being 6 or 4 types of earthquake motion characteristic parameters selected from the 25 types of earthquake motion characteristic parameters shown in FIG. , any seismic motion characteristic parameters from the 25 types of earthquake motion characteristic parameters may be selected and combined as features, or other earthquake motion characteristic parameters other than the 25 types of earthquake motion characteristic parameters may be further combined as features. Good too. At this time, a cluster analysis (see FIG. 28 described later) of each feature quantity may be performed, and based on the analysis results, seismic motion characteristic parameters to be used as feature quantities may be selected. For example, if the seismic motion characteristics are classified into multiple clusters, the seismic motion characteristic parameters that represent each cluster may be selected as the feature values, or if there are multiple seismic motion characteristic parameters that are highly interdependent. In this case, representative earthquake motion characteristic parameters may be selected as feature quantities.

FIG. 28 is a diagram showing the interdependence of each feature amount by cluster analysis. As a result of cluster analysis of each feature in the learning data set A, the 25 types of seismic motion characteristic parameters were classified into three clusters, as shown in FIG. 28. That is, there are three regions with relatively large amounts of mutual information, as zoned on the diagonal line in FIG.

The first one is located at the upper right of Fig. 28 and contains the epicenter position (latitude lat_eq, longitude lon_eq), moment magnitude Mw, epicenter depth H, source mechanism solution (Strike1, dip1, rake1), and the conjugate solution of the source mechanism solution ( Strike2, dip2, rake2), earthquake type Type, fault type Mech, and epicenter distance X. The second is the observation point position located in the center (latitude lat_site, longitude lon_site), surface layer 10m average S-wave speed AVS10, surface layer 30m average S-wave speed AVS30, top layer S-wave speed VS1, S-wave speed 700m /s layer top depth D7, S wave velocity 1400 m/s layer top depth D17, S wave velocity 2100 m/s layer top depth D24, seismic basement depth D28, and microtopographic classification JCODE. ”. The third is the "epicenter orientation cluster" located at the lower left and consisting of epicenter orientations (sinΛ, cosΛ). Note that the directional component Comp (in this embodiment, the north-south direction component or the east-west direction component) of the seismic motion located in the upper right corner does not belong to any cluster and is independent.

The source clusters and site clusters were classified into source characteristics and site characteristics, respectively. The epicenter distance X, which is a propagation characteristic, was aggregated into the epicenter cluster, but the mutual information between the epicenter distance That's what I found out.

(About the epicenter azimuth Λ)
In each of the above embodiments, the pair of sin Λ and cos Λ (hereinafter referred to as "case C10") is applied to the epicenter azimuth Λ, which is one of the feature quantities, but other data formats may be used as the epicenter azimuth Λ. may be applied.

Figure 29 shows various data formats of the epicenter azimuth Λ, (a) shows the definition of each data format, and (b) shows each feature value (seismic motion characteristic parameters) for the target variable (earthquake motion index) in each data format. ) is a diagram showing the degree of influence of Note that FIG. 29(b) shows the results when using the point model T4 generated by machine learning using the learning data set T.

As for the data format of the feature quantity of the epicenter azimuth Λ, cases C11 and C12 apply a continuous quantity using trigonometric functions like case C10, case C11 applies only sinΛ, and case C12 applies a continuous quantity using trigonometric functions. , cosΛ only was applied. Further, in case C01, ΛS, which defines due north as 0° in the epicenter direction, is applied as is, and in case C02, ΛN, in which true south is defined as 0° in the epicenter direction, is applied as is. In cases C10-12, C01, and C02, the epicenter azimuth Λ is given as a continuous quantity for all directions, but for C01 for the northern earthquake and C02 for the southern earthquake, the epicenter is near the azimuth 0°. The orientation value becomes discontinuous.

Furthermore, the data format of the feature quantity of the epicenter azimuth Λ does not give a continuous quantity corresponding to the epicenter azimuth Λ, as in cases C10-12, C01, and C02, but as a data format that provides a continuous quantity corresponding to the epicenter azimuth Λ, as in cases C21-C24, and C33. Among a plurality of azimuth divisions (for example, 12 azimuths) obtained by dividing the azimuth into a predetermined number of divisions, the azimuth division corresponding to the epicenter azimuth Λ may be given as a discrete quantity. For case C24, four sets of numbers corresponding to the number of times N, S, E, and W appear in the four directions in the character string of case C22 are applied, and for case C33, case C23 is changed to 20 directions.

The degree of influence of each feature in each case has the same tendency in each case as shown in FIG. is relatively small in the five cases C21 to C24 and C33, where Λ is given by a discrete quantity using azimuth divisions, and relatively high in cases C11 to C12, C01, and C02, where the epicenter azimuth Λ is given by a continuous quantity. Do you get it.

(About the program)
In the above embodiment, the seismic motion evaluation model generation program 300 and the seismic motion evaluation program 400 have been described as being stored in the storage units 30 and 40, respectively, but they are stored in installable or executable format files on a CD-ROM, It may also be provided recorded on a computer-readable recording medium such as a DVD or a USB memory. Furthermore, the seismic motion evaluation model generation program 300 and the seismic motion evaluation program 400 may be provided by being stored on a computer connected to a network such as the Internet and downloaded via the network.

1...Earthquake motion evaluation system,
2A...Observation data providing device, 2B...Analysis data providing device,
2C...Underground structure data providing device,
3... Earthquake motion evaluation model generation device, 4... Earthquake motion evaluation device,
5...Network, 6...Seismic motion simulation device,
10...Database, 10A...First database, 10B...Second database,
11...Earthquake data,
12...Learning data, 12A...First learning data, 12B...Second learning data,
13, 13A-13K...Earthquake motion evaluation model,
30...Storage unit, 31...Control unit, 32...Communication unit, 33...Input unit, 34...Display unit,
40...Storage unit, 41...Control unit, 42...Communication unit, 43...Input unit, 44...Display unit,
300...Seismic motion evaluation model generation program,
310...DB management unit, 310A...first DB management unit, 310B...second DB management unit,
311... Acquisition unit, 311A... First acquisition unit, 311B... Second acquisition unit, 312... Generation unit,
400...Seismic motion evaluation program,
410... Reception unit, 411... Prediction unit, 412... Output processing unit

Claims (6)

A seismic motion evaluation model generation method for generating an earthquake motion evaluation model by machine learning using a computer, the method comprising:
For each combination of a plurality of actual earthquakes and at least one observation point where actual earthquake motion due to the plurality of actual earthquakes was observed, various earthquake motion characteristic parameters and earthquake motion observation records when the actual earthquake motion was observed are stored in association with each other. From a first database, acquire a plurality of first learning data consisting of the feature quantities and the objective variables, with the seismic motion characteristic parameters as the feature quantities and the seismic motion index obtained from the seismic motion observation record as the objective variable. a first acquisition step of
For each combination of a plurality of virtual earthquakes and at least one virtual observation point from which virtual seismic motions due to the plurality of virtual earthquakes are calculated by simulation, various earthquake motion characteristic parameters and earthquake motion calculations when the virtual earthquake motions are calculated by the simulation. From a second database that stores the results in association with each other, a second database consisting of the feature quantities and the objective variables is obtained, with the various seismic motion characteristic parameters as the feature quantities and the seismic motion index obtained from the seismic motion calculation results as the objective variables. a second acquisition step of acquiring a plurality of learning data;
The feature amount is calculated based on the plurality of first learning data acquired in the first acquisition step and the plurality of second learning data acquired in the second acquisition step. and a generation step of generating the seismic motion evaluation model as a learned model of the machine learning by learning the correlation of the target variables by the machine learning,
The second acquisition step includes:
acquiring a plurality of the second learning data from the second database so that the distribution of the seismic motion characteristic parameters of the actual earthquake motion is interpolated by the distribution of the seismic motion characteristic parameters of the virtual earthquake motion;
Earthquake motion evaluation model generation method. The seismic motion characteristic parameters are:
including at least two characteristics of the seismic source characteristics, propagation characteristics, site characteristics, and azimuth characteristics for the actual earthquake motion and the virtual earthquake motion,
The distribution of the seismic motion characteristic parameters is as follows:
A distribution in which at least two of the characteristics included in the seismic motion characteristic parameters are variables;
The seismic motion evaluation model generation method according to claim 1 . A seismic motion evaluation model generation method for generating an earthquake motion evaluation model by machine learning using a computer, the method comprising:
For each combination of a plurality of actual earthquakes and at least one observation point where actual earthquake motion due to the plurality of actual earthquakes was observed, various earthquake motion characteristic parameters and earthquake motion observation records when the actual earthquake motion was observed are stored in association with each other. From a first database, acquire a plurality of first learning data consisting of the feature quantities and the objective variables, with the seismic motion characteristic parameters as the feature quantities and the seismic motion index obtained from the seismic motion observation record as the objective variable. a first acquisition step of
For each combination of a plurality of virtual earthquakes and at least one virtual observation point from which virtual seismic motions due to the plurality of virtual earthquakes are calculated by simulation, various earthquake motion characteristic parameters and earthquake motion calculations when the virtual earthquake motions are calculated by the simulation. From a second database that stores the results in association with each other, a second database consisting of the feature quantities and the objective variables is obtained, with the various seismic motion characteristic parameters as the feature quantities and the seismic motion index obtained from the seismic motion calculation results as the objective variables. a second acquisition step of acquiring a plurality of learning data;
The feature amount is calculated based on the plurality of first learning data acquired in the first acquisition step and the plurality of second learning data acquired in the second acquisition step. and a generation step of generating the seismic motion evaluation model as a learned model of the machine learning by learning the correlation of the target variables by the machine learning,
The second acquisition step includes:
A plurality of pieces of second learning data are obtained from the second database that stores the various seismic motion characteristic parameters and the seismic motion calculation results in association with each other when the virtual seismic motion is calculated by the simulation based on a plurality of simulation methods. get,
Earthquake motion evaluation model generation method. A seismic motion evaluation method that uses a computer to evaluate the characteristics of the seismic motion based on the seismic motion evaluation model generated by the seismic motion evaluation model generation method according to any one of claims 1 to 3 ,
a reception step of receiving the seismic motion characteristic parameters to be predicted;
The various seismic motion characteristic parameters of the prediction target received in the reception process are input into the seismic motion evaluation model as the feature quantities, and based on the objective variables output from the seismic motion evaluation model, the seismic motion characteristic parameters of the prediction target are a prediction step of predicting the seismic motion index corresponding to the seismic motion characteristic parameters;
Earthquake motion evaluation method. A computer,
comprising a control unit that executes each step included in the seismic motion evaluation model generation method according to any one of claims 1 to 3 ;
Earthquake motion evaluation model generator. A computer,
comprising a control unit that executes each step included in the seismic motion evaluation method according to claim 4 ;
Earthquake motion evaluation device.

Images