JP7414694B2 - Analysis result judgment method, validity judgment support method, analysis result learning device, and analysis result judgment device - Google Patents

Description

The present invention relates to an analysis result determination method, a validity determination support method, an analysis result learning device, and an analysis result determination device.

Currently, various technologies using artificial intelligence (AI) technology are being developed for construction work. For example, a method has been developed that uses the running spectrum of seismograph records to automatically distinguish between low-frequency tremors and normal seismic motion signals with high precision (see Non-Patent Document 1). A running spectrum is one of the methods of displaying temporal changes in frequency components included in wave signals such as earthquakes. This is a method of displaying the duration and time change of the frequency components included in a signal as an image by showing the frequency components in the vertical direction of the image and displaying them in the horizontal direction according to the passage of time. This technology makes it possible to automatically distinguish between low-frequency tremors and normal seismic motion with high precision.

“Press release “Developing a new method to automatically and highly accurately distinguish seismic motion and low-frequency tremor signals using artificial intelligence””, [online], Japan Agency for Marine-Earth Science and Technology, [November 2020] Searched on the 2nd of each month], Internet <http://www.jamstec.go.jp/j/about/press_release/20190116/>

On the other hand, there are still many tasks that are still performed manually by humans, and further improvements in efficiency and accuracy are desired. For example, in order to obtain permits for nuclear power-related facilities, it is necessary to perform seismic analysis and evaluation and submit related materials (referred to as "quality assurance materials"). Currently, humans print out input data used in seismic analysis and evaluation, compare it with supporting documents, and check-marked documents are used as quality assurance documents. The work of checking this input data is enormous, especially since double checks and triple checks are performed.

From this perspective, the present invention provides an analysis result determination method, a validity determination support method, an analysis result learning device, and an analysis result determination device that can efficiently perform analysis quality confirmation work in seismic analysis work. .

The analysis result determination method according to the present invention is an analysis result determination method for determining the analysis result of earthquake response analysis. In the seismic response analysis, response time history waveforms at each node of the analysis model are calculated by dynamic analysis.
This analysis result determination method includes a learning step in which a learning device is subjected to machine learning using learning data, and a determination step in which an analysis result of an earthquake response analysis is determined by the learning device learned in the learning step. The learning process includes a learning data generation process and a learning processing process.
In the learning data generation step, image data of a response time history waveform obtained as an analysis result or a frequency characteristic of the response time history waveform, and analysis result judgment information indicating whether the response time history waveform is normal or abnormal. A set of data is generated as learning data.
In the learning processing step, the learning device is trained to output the analysis result judgment information by inputting the image data.
In the determination step, the analyzed image data is input to a trained learning device to determine whether the response time history waveform on which the image data is based is normal or abnormal.

In the analysis result determination method according to the present invention, a learning device is subjected to machine learning using learning data that includes a response time history waveform obtained as an analysis result or image data of the frequency characteristics of the response time history waveform. By using a learning device trained using the learning data, it is possible to judge whether the analysis results are normal or abnormal without relying on humans. Therefore, it is possible to improve the efficiency of business operations. In addition, image data that visualizes the response time history waveform and the frequency characteristics of the response time history waveform can be used to judge abnormal responses that may occur in dynamic problems even when viewed by humans. Association (labeling) with analysis result judgment information is relatively easy.

The learning data for machine learning includes first learning data in which the image data analyzed under correct analysis conditions is attached with analysis result judgment information indicating normality, and image data analyzed under incorrect analysis conditions. The learning data includes second learning data with analysis result judgment information indicating an abnormality. In the learning data generation process, one of the following analysis conditions is set: no damping input, incorrect damping input, incorrect input of material property values, incorrect change of physical properties to dynamic analysis, and incorrect range of lateral viscous boundary elements. The second learning data may be created by performing the following steps.
In this way, abnormal responses that may occur in dynamic problems can be generally satisfied.

The validity determination support method according to the present invention is a validity determination support method that supports determination of the validity of analysis conditions used in earthquake response analysis.
This validity judgment support method includes a drawing process of acquiring analysis result judgment information of all nodes judged by the above-described analysis result judgment method and drawing the analysis result judgment information in association with an analysis model.
In this way, it is possible to judge at a glance whether the analysis results for each node are normal or abnormal, making it easy to identify incorrect analysis conditions.

An analysis result learning device according to the present invention is an analysis result learning device that learns analysis results of earthquake response analysis. In the seismic response analysis, response time history waveforms at each node of the analysis model are calculated by dynamic analysis.
This analysis result learning device includes a learning data acquisition section and a learning processing section.
The learning data acquisition unit includes image data of a response time history waveform obtained as an analysis result or a frequency characteristic of the response time history waveform, and analysis result determination information indicating whether the response time history waveform is normal or abnormal. Obtain the set of as learning data.
The learning processing section has a learning device, and causes the learning device to perform machine learning so as to output the analysis result judgment information by inputting the image data.

In the analysis result learning device according to the present invention, the learning device is subjected to machine learning using learning data that includes a response time history waveform obtained as an analysis result or image data that is an image of the frequency characteristic of the response time history waveform. By using a learning device trained using the learning data, it is possible to judge whether the analysis results are normal or abnormal without relying on humans. Therefore, it is possible to improve the efficiency of business operations. In addition, image data that visualizes the response time history waveform and the frequency characteristics of the response time history waveform can be used to judge abnormal responses that may occur in dynamic problems even when viewed by humans. Association (labeling) with analysis result judgment information is relatively easy.

An analysis result determination device according to the present invention is an analysis result determination device that determines an analysis result of an earthquake response analysis. In the seismic response analysis, response time history waveforms at each node of the analysis model are calculated by dynamic analysis.
This analysis result determination device includes a determination data acquisition section and a determination processing section.
The determination data acquisition unit acquires a response time history waveform obtained as an analysis result or image data of a frequency characteristic of the response time history waveform.
The determination processing unit uses a trained learning device that has performed machine learning so as to input the image data and output a determination result indicating whether the response time history waveform on which the image data is based is normal or abnormal. have

In the analysis result determination device according to the present invention, the determination is made by a learning device that performs machine learning using learning data that includes a response time history waveform obtained as an analysis result or image data of the frequency characteristics of the response time history waveform. do. Therefore, it is possible to judge whether the analysis results are normal or abnormal without relying on humans, and it is possible to improve the efficiency of business operations.

According to the present invention, analysis quality confirmation work in seismic analysis work can be performed efficiently.

1 is a schematic configuration diagram of a validity determination support system according to an embodiment of the present invention. This is an example of an analytical model. These are examples of image data, in which (a) is an example of an acceleration time history image, and (b) is an example of a transfer function image that visualizes the frequency characteristics of the acceleration time history image. 3 is an illustration of a flowchart of a validity determination support method (including an analysis result determination method) according to an embodiment. These are examples of the types of analytical models to be trained, and (a) is a "stratified model", (b) is an "irregular model", (c) is a "fracture zone model", and (d) is an example of an "asymmetric model". It is. These are examples of acceleration time history images, (a) is an example of an acceleration time history image of a normal response, (b) is an example of an acceleration time history image of an abnormal response in which a pulse has occurred, and (c) is an example of an acceleration time history image of an initial This is an example of an acceleration time history image of an abnormal response in which variable vibration occurs. These are examples of transfer function images, (a) is an example of a transfer function image of a normal response, (b) is an example of a transfer function image of an abnormal response in which high-frequency noise has occurred, and (c) is an example of a transfer function image of an abnormal response in which high-frequency noise has occurred. This is an example of a transfer function image of an abnormal response that has occurred. FIG. 3 is an image diagram showing a processing flow of a determination step of the validity determination support method according to the embodiment.

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings as appropriate. The figures are only shown schematically to provide a thorough understanding of the invention. Therefore, the present invention is not limited to the illustrated example. In each figure, common or similar components are designated by the same reference numerals, and their overlapping explanations will be omitted.

<About the configuration of the validity judgment support system according to the embodiment>
With reference to FIG. 1, a validity determination support system 1 according to an embodiment will be described. FIG. 1 is a schematic configuration diagram of a validity determination support system 1 according to an embodiment.
The validity determination support system 1 is a system that supports determination of the validity of analysis conditions used in earthquake response analysis. The validity determination support system 1 mainly includes an analysis result learning device 10 and an analysis result determination device 20. Note that the analysis result learning device 10 and the analysis result determination device 20 can also be configured as one device.
The analysis result learning device 10 is a device that learns the relationship between a response time history waveform output by performing earthquake response analysis and an analysis result of the response time history waveform (for example, whether it is normal or abnormal). . The analysis result learning device 10 performs machine learning using, for example, information obtained by imaging response time history waveforms of nodes output as a result of earthquake response analysis.
The analysis result determination device 20 is a device that determines whether the response time history waveform is normal or abnormal based on the response time history waveform output by performing earthquake response analysis. The analysis result determination device 20 determines normality or abnormality using, for example, information obtained by imaging response time history waveforms of nodes output as a result of earthquake response analysis.

(Analysis result learning device)
As shown in FIG. 1, the analysis result learning device 10 includes an earthquake response analysis section 11, a learning data generation section 12, and a learning processing section 13. The earthquake response analysis section 11, learning data generation section 12, and learning processing section 13 are realized by program execution processing by a CPU (Central Processing Unit), a dedicated circuit, and the like. When these functions are realized by program execution processing, a program for realizing the functions is stored in a storage unit (not shown).

The earthquake response analysis unit 11 calculates the influence of an earthquake on a structure by numerical analysis using a virtual space, for example, calculates the influence of an earthquake on a structure by earthquake response analysis using the finite element method. .
An analysis model D and analysis conditions E are input to the earthquake response analysis section 11. The analysis model D is an abstraction (modeling) of the structure to be analyzed. The analytical model D may be either two-dimensional or three-dimensional. FIG. 2 shows an example of analysis model D. The analytical model D is divided into a plurality of regions (also called "elements", "mesh", etc.) (that is, the analytical model D is a collection of meshes (elements)). A mesh is a simple shape such as a triangle or a quadrilateral, and the vertices are called "nodes." The analysis conditions E include, for example, information regarding a structure to be analyzed, information regarding an earthquake applied as an external force, and the like.

The earthquake response analysis unit 11 receives, for example, the analysis model D and the analysis conditions E as input, and calculates an acceleration time history waveform at each node. The acceleration time history waveform is an example of a response time history waveform, and the response time history waveform analyzed by the earthquake response analysis unit 11 may be other than acceleration. The earthquake response analysis unit 11 may further obtain the frequency characteristics of the response time history waveform.
The earthquake response analysis unit 11 outputs image data F, which is an image of the calculated response time history waveform of each node and the frequency characteristics of the response time history waveform, to the learning data generation unit 12. The imaging method is not particularly limited. The image data F is, for example, a binary image (black and white image). The earthquake response analysis unit 11 creates, for example, an acceleration time history image that is an image of an acceleration time history waveform, or a transfer function image that is an image of the frequency characteristics of the acceleration time history image, and outputs it to the learning data generation unit 12. An example of the image data F is shown in FIG. FIG. 3(a) is an example of an acceleration time history image, and FIG. 3(b) is an example of a transfer function image that visualizes the frequency characteristics of the acceleration time history image. The acceleration time history image and the transfer function image may include a reference axis, but preferably do not include any other supplementary information. In other words, in Figure 3, in order to make it easier to understand the acceleration time history image and the transfer function image, explanations of the axes (“time”, “acceleration”, “frequency”, “transfer function”) and numerical values (“frequency” 0Hz" and "50Hz", and "0" and "20" indicating the transfer function), but it is best not to include this information in the actual acceleration time history image and transfer function image. (The same applies to FIGS. 6 and 7, which will be described later).
In this embodiment, it is assumed that the analysis function (earthquake analysis program that implements earthquake response analysis) in the earthquake response analysis section 11 has been sufficiently verified, and there are no calculation errors in the analysis. In other words, if correct analysis conditions are input, a normal analysis result (assuming an acceleration time history waveform here) at each node will always be obtained. Therefore, if there is an abnormality in the analysis results, it means that there was a mistake in the analysis conditions. Mistakes in analysis conditions include, for example, non-input of necessary information, incorrect input, and mistakes in prerequisite conditions.

The learning data generation unit 12 generates learning data used in learning. The learning data here includes image data F that visualizes the response time history waveform or the frequency characteristics of the response time history waveform, and information indicating whether the response time history waveform is normal or abnormal (this information is referred to as "analysis result judgment information"). ). The analysis result determination information may be the result of human analysis and determination of response time history waveforms and frequency characteristics of each node. The administrator of the validity determination support system 1 sets analysis result determination information in the image data F using, for example, an input device (not shown). Roughly speaking, information indicating normality (analysis result judgment information) is attached to the response time history waveform etc. calculated as a result of analysis performed under the correct analysis conditions, and information indicating normality (analysis result judgment information) is attached to the response time history waveform calculated as a result of analysis performed under the correct analysis conditions. Information indicating an abnormality (analysis result judgment information) is attached to the response time history waveform and the like. It should be noted here that even if analysis is performed under incorrect analysis conditions, abnormal responses do not necessarily occur at all nodes. In other words, even if analysis is performed under incorrect analysis conditions, a normal response time history waveform may be obtained at a certain node, and an abnormal response time history waveform may be obtained at a certain node. The location and range of the node where the abnormal response occurs is related to, for example, the nature of the error in the analysis conditions. Note that the analysis result determination information may include information indicating what kind of abnormality has occurred. In other words, the analysis result determination information may be classified according to the type of abnormality, such as normal, abnormality a, abnormality b, etc. The learning data generation section 12 outputs learning data to the learning processing section 13.

The image data F that visualizes the response time history waveform or the frequency characteristics of the response time history waveform is combined with information indicating whether the response time history waveform is normal or abnormal (this information is called "analysis result judgment information"). The learning data generation unit 12 may obtain the registered information as the learning data by registering the information in the storage unit in advance in a state of . The storage unit may be provided in a device connected via a communication line (for example, a cloud system). The function of acquiring learning data stored in the storage unit may be referred to as a “learning data acquisition unit”. The learning data generation unit 12 acquires the response time history waveform of each node and the frequency characteristics of the response time history waveform from the earthquake response analysis unit 11, and the learning data generation unit 12 acquires the image data F used for learning. You may also create one.

The learning processing section 13 has a learning device, and uses the learning data generated by the learning data generation section 12 to perform machine learning on the learning device. A learning device is a learning system in machine learning that compares the results of classification, prediction, judgment, etc. based on the given data with the actual result that is the correct answer, and adjusts various parameters to derive good results. becomes possible. The learning device is, for example, a neural network, and this embodiment will also be described assuming a neural network. Note that the learning device is also called a "learning model" or "artificial intelligence (AI)."

Neural networks, as is well known, are an information processing method that was created for the purpose of making computers imitate the movements of the human brain. Neural networks are good at complex nonlinear processing, and have a learning function that eliminates the need to formulate the relationship between explanatory variables and objective variables. Neural networks generally consist of an input layer, a hidden layer, and an output layer. Explanatory variables are input to the input layer, and objective variables are output from the output layer. The middle layer plays the role of linking the two. There are no restrictions on the number of hidden layers and the number of neurons in each hidden layer, and they can be set arbitrarily.

The method of learning the learning device is not particularly limited, and for example, the learning device is trained using the error backpropagation method. Specifically, image data F constituting the learning data is input to the input layer, and the intermediate layer is adjusted based on the error between the result output from the output layer and the analysis result judgment information constituting the learning data. That is, the learning processing unit 13 causes the learning device to perform machine learning so as to output information (analysis result determination information) indicating whether the response time history waveform is normal or abnormal by inputting the image data F. The number of learning data to be trained is not particularly limited, and, for example, learning is terminated when the expected accuracy is reached. Note that a learning device (AI) corresponding to the type of abnormal response may be created (details will be described later).

(Analysis result judgment device)
As shown in FIG. 1, the analysis result determination device 20 includes an earthquake response analysis section 21, a determination data acquisition section 22, and a determination processing section 23. The earthquake response analysis section 21, the judgment data acquisition section 22, and the judgment processing section 23 are realized by program execution processing by a CPU (Central Processing Unit), a dedicated circuit, and the like. When these functions are realized by program execution processing, a program for realizing the functions is stored in a storage unit (not shown).

The earthquake response analysis unit 21 calculates the influence of an earthquake on a structure by numerical analysis using a virtual space, for example, calculates the influence of an earthquake on a structure by earthquake response analysis using the finite element method. . An analysis model D and analysis conditions E are input to the earthquake response analysis section 21. The processing of the seismic response analysis section 21 is similar to the processing of the seismic response analysis section 11 in the learning process, so a detailed explanation will be omitted. The seismic response analysis unit 21 outputs image data F, which is an image of the calculated response time history waveform of each node and the frequency characteristics of the response time history waveform, to the determination data acquisition unit 22. The image data F is, for example, a binary image (black and white image). It is assumed that the analysis function (earthquake analysis program that implements earthquake response analysis) in the earthquake response analysis section 21 has been sufficiently verified, and there are no calculation errors in the analysis.

The determination data acquisition unit 22 acquires determination data used in determination. The determination data here is image data F that is an image of the response time history waveform or the frequency characteristics of the response time history waveform. The determination data acquisition unit 22 may acquire determination data from another device via a communication line. The determination data acquisition unit 22 outputs the determination data to the determination processing unit 23. Note that the judgment data acquisition unit 22 acquires information indicating the response time history waveform of each node and the frequency characteristics of the response time history waveform from the earthquake response analysis unit 21, and the judgment data acquisition unit 22 acquires the image data F used for judgment. You may also create one.

The determination processing unit 23 has a learned learning device (AI). The learned learning device is a learning device subjected to machine learning by the analysis result learning device 10. In other words, the trained learning device is trained to output information (analysis result determination information) indicating whether the acceleration time history waveform is normal or abnormal by inputting the image data F. The determination processing unit 23 outputs information (determination result) indicating whether the analyzed response time history waveform is normal or abnormal by inputting the determination data acquired by the determination data acquisition unit 22 into a trained learning device. In this embodiment, the determination processing section 23 outputs the determination results of all nodes to the drawing processing section 31.

The determination results (corresponding to analysis result determination information) of all nodes are input to the drawing processing unit 31. The drawing processing unit 31 displays the determination results in association with the nodes of the analysis model D. With this, for example, if analysis is performed under incorrect analysis conditions, the location and range of the node where an abnormal response has occurred can be confirmed at a glance. In this embodiment, a device in which the function of the drawing processing unit 31 is added to the analysis result determination device 20 will be referred to as a "validity determination support device 30."

<About the validity judgment support method (including the analysis result judgment method) according to the embodiment>
A validity determination support method (including an analysis result determination method) according to the embodiment will be described with reference to FIG. 4 (see FIGS. 1 to 3 as appropriate). FIG. 4 is an example of a flowchart of the validity determination support method (including the analysis result determination method) according to the embodiment. The validity judgment support method supports the judgment of the validity of analysis conditions used in earthquake response analysis.
As shown in FIG. 4, the validity determination support method includes a learning step (S10) and a determination step (S20). The learning step (S10) is a step of performing machine learning on the learning device using learning data. The determination step (S20) is a step of determining the analysis result of the earthquake response analysis using the trained learning device (AI) learned in the learning step.

(Learning process (S10))
The learning process (S10) mainly includes a preliminary preparation (earthquake response analysis) process (S11), a pre-processing process (S12), a labeling process (S13), and a learning processing process (S14).
“Advance preparation process (S11)”
In the preliminary preparation step, first, abnormal responses that may occur in a dynamic problem are determined. In this embodiment, five abnormal responses are "E1: high frequency noise", "E2: overdamping", "E3: pulse", "E4: initial abnormal vibration", and "E5: displacement drift". The five listed here can generally explain abnormal responses in dynamic problems. In order to generate these abnormal responses, incorrect analysis conditions are intentionally set. Here, errors in analysis conditions include (1) not inputting attenuation, (2) incorrectly inputting attenuation (for example, inputting "3" instead of "3% (0.03)"), (3) material property values (deformation (4) incorrectly changing physical properties to dynamic analysis (for example, changing from self-weight analysis); (5) incorrectly entering the range of lateral viscous boundary elements.
Perform the analysis under the correct analysis conditions and output the acceleration time history waveform of "E0: normal response" at all nodes. In addition, analysis was performed using the incorrect analysis conditions (1) to (5) above, and "E1: High frequency noise", "E2: Overdamping", "E3: Pulse", "E4: Initial abnormal vibration", " E5: Output the acceleration time history waveform where the abnormal response of "displacement drift" occurred.
In this embodiment, (A) a learning device (AI) that can determine (detect) "E1: high frequency noise", (B) a learning device (AI) that can determine (detect) "E2: overdamping", (C) Learning device (AI) that can determine (detect) “E3: Pulse”, (D) Learning device (AI) that can determine (detect) “E4: Initial abnormal vibration”, (E) Determine “E5: Displacement drift” We will create five trained learning machines (AI) that can (detect). In order to create these five learning devices, 50 cases of input data were prepared for each, and analysis was performed using these input data. Each case is a case in which, for example, the analysis model, input physical property values, and seismic motion are changed. As a result, the amount of learning data for each learning device (A) to (E) is "number of analysis data (here, 50") x number of nodes (for example, 2,000) = approximately 100,000. Various analytical models may be used, such as a "stratified model,""irregularmodel,""fracture zone model," or "asymmetric model." Figure 5 shows the types of analytical models to be trained. FIG. 5(a) is an example of a "stratified model", (b) is an "irregular model", (c) is a "fracture zone model", and (d) is an example of an "asymmetric model". It is best to use various types of seismic motion, for example, ``ocean trench type'' and ``inland type'' for each type of ground.

"Pre-processing step (S12)"
Image the response time history waveform. The image data in this embodiment is of the following two types.
(1) Acceleration time history image (2) Transfer function image The reason for using "acceleration time history image" and "transfer function image" is as follows. There are two ways to learn AI learning data: the method of learning the data itself (raw data) and the method of learning images. After trying learning using various raw data and learning using various images (such as running spectra), the inventor focused on ``which data should be learned to most improve the accuracy rate?'' did. In addition, we considered it important to use images that humans can judge as abnormal for abnormal responses that may occur in dynamic problems (for example, image creation in a "frequency band" specific to dynamic problems). high-frequency noise, etc.). As a result, we adopted two types: an "acceleration time history image" and a "transfer function image."
FIG. 6 illustrates an acceleration time history image. 6(a) is an example of an acceleration time history image of a normal response, (b) is an example of an acceleration time history image of an abnormal response in which a pulse has occurred, and (c) is an example of an acceleration time history image in an abnormal response in which an initial fluctuation vibration has occurred. This is an example of an acceleration time history image.
FIG. 7 illustrates a transfer function image. FIG. 7(a) is an example of a transfer function image of a normal response, (b) is an example of a transfer function image of an abnormal response in which high-frequency noise has occurred, and (c) is an example of a transfer function image in an abnormal response in which over-attenuation has occurred. This is an example of a function image.
Note that a transfer function image is particularly effective for creating AI for determining abnormal responses of "E1: high frequency noise" and "E2: overdamping". In addition, acceleration time history images are particularly effective for creating AI for determining abnormal responses of "E3: Pulse" and "E4: Initial Abnormal Vibration." Further, when creating an acceleration time history image, it is preferable to create an image without setting upper limits for "acceleration" and "time". To create a transfer function image, it is best to set the range of "transfer function" and "frequency" in advance and create an image (for example, set the transfer function in the range of "0" to "20" and set the frequency as "0Hz" to It is best to image it as a range of 50Hz).

"Labeling process (S13)"
Each training data image is labeled as "normal" or "abnormal". The number of analyzes consisting of correct input data and the number of analyzes consisting of incorrect input data in the ``advance preparation step (S11)'' are adjusted so that the ratio of normal and abnormal is approximately 50:50. Note that labeling is performed keeping in mind that even if analysis is performed with incorrect input data, all nodes do not become abnormal. The labeling method may be based on engineering judgment, for example, a human determines "normal" and "abnormal" by looking at the image. Note that the reason for setting the ratio of normal to abnormal at approximately 50:50 is to avoid overfitting.

"Learning process step S14"
Through supervised learning using labeled image data, the acceleration time history waveform output from seismic response analysis and the analysis results (for example, whether it is normal or abnormal) of the acceleration time history waveform are Let the learning device (AI) learn the relationship. For example, a learning device (AI) is trained using a convolutional neural network.

(Judgment step (S20))
The determination step (S20) shown in FIG. 4 mainly includes a preliminary preparation (earthquake response analysis) step (S21), a pre-processing step (S22), a determination processing step (S23), and a drawing processing step (S24). have The advance preparation (seismic response analysis) step (S21) and the pre-processing step (S22) are the same as the advance preparation step (S11) and the pre-processing step S12 in the learning step (S10).
"Advance preparation process (S21), pre-processing process (S22)"
In the preliminary preparation process, analysis is performed using an earthquake analysis program and the acceleration time history waveforms of all nodes are output.
In the pre-processing step, the acceleration time history waveform is imaged. The image data in this embodiment is of two types: an "acceleration time history image" and a "transfer function image."

“AI (judgment processing) step (S23)”
In the AI (judgment processing) step, the pre-processed image data is input to a learned learning device (AI), and it is determined whether the response results of all nodes are "normal" or "abnormal." The judgment result by AI is, for example, a value between "0" and "1.0" is output for all nodes, and if it is less than "0.5", it is "normal", and if it is "0.5" or more, it is "abnormal". It is determined that In this embodiment, (A) a learning device (AI) that can determine (detect) "E1: high frequency noise", (B) a learning device (AI) that can determine (detect) "E2: overdamping", (C) Learning device (AI) that can determine (detect) “E3: Pulse”, (D) Learning device (AI) that can determine (detect) “E4: Initial abnormal vibration”, (E) Determine “E5: Displacement drift” Since we have created five learning devices (AI) capable of (detection), image data is input to all of them, and each judgment result is output.

“Drawing processing step (S24)”
In the drawing processing step, the judgment results output in the AI (judgment processing) step are inputted into, for example, a drawing tool, and a screen is displayed on the display unit on which the positions of the nodes judged as "abnormal" can be seen. Note that instead of displaying the node positions determined to be "abnormal," or in addition to displaying the node positions determined to be "normal," node positions determined to be "normal" may be displayed. It is preferable that the display of the determination result be associated with the analysis model D that performed the analysis. When outputting the determination results using five AIs, the five determination results may be output as separate images, or each determination result may be output as one image. FIG. 8 shows the flow of a series of processes in the determination step (S20) described so far.

As described above, the validity judgment support system 1 according to the present embodiment uses learning data that includes image data of the response time history waveform obtained as an analysis result or the frequency characteristics of the response time history waveform. Make the learning device perform machine learning. By using a learning device trained using the learning data, it is possible to judge whether the analysis results are normal or abnormal without relying on humans. Therefore, it is possible to improve the efficiency of business operations.
In addition, image data that visualizes the response time history waveform and the frequency characteristics of the response time history waveform can be used to judge abnormal responses that may occur in dynamic problems even when viewed by humans. Association (labeling) with analysis result judgment information is relatively easy.
Further, in the validity judgment support system 1 according to the present embodiment, analysis result judgment information (judgment result) is drawn in association with the analysis model. Therefore, it is possible to judge at a glance whether the analysis results for each node are normal or abnormal, making it easy to identify incorrect analysis conditions.

Although the embodiments of the present invention have been described above, the present invention is not limited thereto, and can be implemented without changing the spirit of the claims.

For example, in the embodiment, it is assumed that earthquake response analysis is performed using the finite element method, but the response analysis is not limited to this. Instead of the analytical model based on the finite element method, a mass point system model (also called a "skewer dumpling model") or the like may be used.
Further, although the type of structure is not specified in the embodiment, AI may be created for each type of structure (for example, above-ground structure, underground structure, etc.).
Further, in the embodiment, learning is performed using an "acceleration time history waveform", but the learning is not limited to acceleration, and learning may be performed using a "response time history waveform" other than acceleration.

1 Validity judgment support system 10 Analysis result learning device 11 Earthquake response analysis section 12 Learning data generation section (learning data acquisition section)
13 Learning processing section 20 Analysis result judgment device 21 Earthquake response analysis section 22 Judgment data acquisition section 23 Judgment processing section 30 Validity judgment support device 31 Drawing processing section

Claims (5)

An analysis result determination method for determining analysis results of earthquake response analysis,
The seismic response analysis calculates the response time history waveform at each node of the analysis model by dynamic analysis,
A learning process in which the learning device performs machine learning using the learning data;
a determination step of determining the analysis result of the earthquake response analysis using the learning device learned in the learning step;
The learning process is
A set of a response time history waveform obtained as an analysis result or image data of the frequency characteristics of the response time history waveform, and analysis result judgment information indicating whether the response time history waveform is normal or abnormal is generated as learning data. a learning data generation process,
a learning processing step of learning the learning device to output the analysis result judgment information by inputting the image data,
In the determination step, inputting the analyzed image data into a trained learning device determines whether the response time history waveform on which the image data is based is normal or abnormal.
An analysis result determination method characterized by the following. The learning data for machine learning includes first learning data in which the image data analyzed under correct analysis conditions is attached with analysis result judgment information indicating normality, and image data analyzed under incorrect analysis conditions. and second learning data with analysis result judgment information indicating an abnormality.
In the learning data generation process, one of the following analysis conditions is set: no damping input, incorrect damping input, incorrect input of material property values, incorrect change of physical properties to dynamic analysis, and incorrect range of lateral viscous boundary elements. creating said second learning data by performing;
The analysis result determination method according to claim 1, characterized in that: A validity judgment support method for supporting judgment of the validity of analysis conditions used in earthquake response analysis, the method comprising:
A drawing processing step of acquiring analysis result judgment information of all nodes judged by the analysis result judgment method according to claim 1 or claim 2, and drawing the analysis result judgment information in association with an analysis model.
A validity judgment support method characterized by the following. An analysis result learning device for learning analysis results of earthquake response analysis,
The seismic response analysis calculates the response time history waveform at each node of the analysis model by dynamic analysis,
A set of a response time history waveform obtained as an analysis result or image data of the frequency characteristics of the response time history waveform, and analysis result judgment information indicating whether the response time history waveform is normal or abnormal is obtained as learning data. a learning data acquisition unit,
a learning processing unit that has a learning device and performs machine learning on the learning device so as to output the analysis result judgment information by inputting the image data;
An analysis result learning device characterized by: An analysis result determination device for determining analysis results of earthquake response analysis,
The seismic response analysis calculates the response time history waveform at each node of the analysis model by dynamic analysis,
a determination data acquisition unit that acquires image data of a response time history waveform obtained as an analysis result or a frequency characteristic of the response time history waveform;
a determination processing unit having a trained learning device that performs machine learning so as to input the image data and output a determination result indicating whether the response time history waveform on which the image data is based is normal or abnormal; Equipped with
An analysis result determination device characterized by:

Images