Damage degree predicting device, damage degree prediction method, trained model generation method and trained model

Abstract

To provide a technology that can output appropriate prediction results even if image data are provided as input data.SOLUTION: A damage degree prediction device 1 includes: memory means 11 for storing a trained model M1 generated by machine learning using training data in which image data showing the deterioration degree of each beam in a beam structure of a pier or bridge is arranged according to the position of the beam, and the damage state of the beam is set as an explanatory variable; input means 21 for inputting into the trained model M1 second image data obtained by resizing first image data showing the beam structure of a pier or bridge whose damage degree is to be predicted to a set number of pixels; and output means 31 for outputting damage state data corresponding to the damage state of the beam structure to be subjected to damage prediction obtained in response to input by the input means 21.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a technique for predicting damage in a beam structure of a pier or bridge. [Background technology]

[0002] Techniques for the construction or maintenance inspection of structures are known. In general, in order to predict or evaluate damage when an external force is applied to a pier beam structure, a detailed periodic inspection diagnosis described in Non-Patent Document 1 is performed, and a structural analysis is performed based on the result. However, since it takes time to obtain a remaining strength evaluation using the detailed periodic inspection result, the structure to be evaluated is likely to be subjected to a large external force during the evaluation, and if a large external force is received before the evaluation result is obtained, the evaluation result itself will lose value. In addition, in terms of time and cost, a detailed inspection itself may not be performed, or a detailed periodic inspection diagnosis may be performed but the remaining strength of the target structure may not be evaluated. In that case, since the current available strength cannot be grasped, repairs must be limited to restoration to the original shape, which is not a rational repair. For the above reasons, detailed periodic inspection diagnosis and structural analysis based on the result require cost and time, a method for relatively easily evaluating the remaining strength from the deterioration degree judgment result obtained from a general inspection diagnosis has been proposed (Non-Patent Document 2).

[0003] Furthermore, the present inventors previously proposed in Patent Document 1, as a damage prediction device for a pier or bridge, a device that uses a trained model that has been subjected to machine learning using teacher data in which input data are a map of the deterioration degree distribution or soundness judgment distribution in the beam structure of a pier or bridge, the planar shape of the beam structure, and the external force applied to the beam structure, and output data is information indicating the damage state of the beam structure obtained from the results of structural analysis when the external force is applied to the beam structure. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-143575 [Non-patent literature]

[0005] [Non-Patent Document 1] "Port Facility Maintenance and Management Technical Manual (Revised Edition)" edited by the Port and Harbor Bureau of the Ministry of Land, Infrastructure, Transport and Tourism, Coastal Technology Research Center, July 2018, p.296, p.317 [Non-Patent Document 2] K. Uno and M. Iwanami, "Application of a residual strength evaluation method using deterioration assessment results to an actual pier," Journal of Japan Society of Civil Engineers, Vol. 74, No. 2, pp. I_55-I_60, 2018. Summary of the Invention [Problem to be solved by the invention]

[0006] However, in the technology described in Patent Document 1, when a map of deterioration distribution or soundness judgment distribution and the planar shape of the beam structure are given as input data as image data, research by the present inventors has shown that when making a prediction for a pier or bridge of an unknown shape (i.e., a shape that has never been learned), the shape of the pier or bridge may be misrecognized, for example by outputting a shape that is close to the shape of piers or bridges that have been learned in the past.

[0007] In response to this, the present invention provides a technique that can output appropriate prediction results even when image data is given as input data. [Means for solving the problem]

[0008] The present invention provides a damage prediction device comprising: a memory means for storing a trained model generated by machine learning using training data in which image data representing the degree of deterioration of each beam in a beam structure of a pier or bridge is arranged according to the position of the beam, and the explanatory variable is image data in which the damage state of the beam is the objective variable; an input means for inputting into the trained model second image data in which first image data showing a beam structure of a pier or bridge whose damage degree is to be predicted is resized to a set number of pixels; and an output means for outputting damage state data corresponding to the damage state of the beam structure to be analyzed, obtained in accordance with the input by the input means.

[0009] The trained model may be generated by machine learning using image data in which an image representing the deterioration degree of each beam is arranged according to the position of the beam, and the image data is resized to have the set number of pixels as an explanatory variable.

[0010] The image data in which the image showing the deterioration degree of each beam is arranged according to the position of the beam and the beam shown in the image based on the first image data may each correspond to one pixel.

[0011] The input means may change the first image data by enlarging the first image data so that the second image data has a set number of pixels.

[0012] The input means may, when the first image data is enlarged to a determined size, and one pixel corresponds to a plurality of structural elements in the beam structure, modify the second image data to have a set number of pixels by setting the pixel value to the structural elements determined based on the proportion of the one pixel.

[0013] When the first image data is enlarged to a determined size, and one pixel corresponds to a plurality of structural elements in the beam structure, the input means may modify the second image data so that it has a set number of pixels by setting a structural element determined based on the type of each structural element included in the one pixel as a pixel value.

[0014] When the number of pixels of the beam shown in the image based on the first image data is different from the number of pixels of the beam shown in the image based on the second image data, the output means may determine a representative value of multiple pixel values ​​constituting the beam shown in the image based on the second image data as the damage state of the beam.

[0015] The representative value may be an average value of a plurality of pixel values ​​constituting a beam shown in an image based on the second image data.

[0016] The representative value may be a pixel value that is the most damaged pixel value among a plurality of pixel values ​​constituting a beam shown in an image based on the second image data.

[0017] The trained model may be generated by an algorithm using an autoencoder or a convolutional autoencoder that compresses input image data, retains features, and then performs a restoration process.

[0018] The trained model may be generated using as explanatory variables the external force conditions acting on a pier or bridge including the beam, data regarding the beam structure of the beam, and numerical data of the beam in the beam structure, and the input means may input the external force conditions acting on a pier or bridge including the beam structure for which the damage degree is to be predicted, and the output means may output damage state data for the external force conditions.

[0019] The numerical data of the beam may include at least one of the dimensions of the beam, the presence or absence of piles at both ends of the beam, the length or protruding length of the piles supporting the beam, or characteristic values ​​of the piles.

[0020] The data regarding the beam structure may include the number of rows of beams that constitute the pier or bridge. The external force conditions may include an earthquake, a dynamic load, and a static load. The damage state data may be data representing at least any one of the degree of damage to the beam, the area of ​​the damage, or the proportion of the area of ​​the damage.

[0021] The output means may output, as the damage state data, a damage degree distribution based on the damage classification in the beam structure, by mapping it on a diagram showing the beam structure.

[0022] The beam structure may be that of an n-layer bridge, the teaching data may include, as the explanatory variables, information indicating which layer the beam structure is in and information indicating the distance from other adjacent layers, and the input means may input information indicating which layer the beam structure is in in the beam structure and information indicating the distance from other adjacent layers.

[0023] The present invention also provides a damage prediction method implemented using the above-mentioned damage prediction device.

[0024] The present invention also provides a method for generating a learned model by machine learning using training data in which image data representing the degree of deterioration of each beam in a beam structure of a pier or bridge is arranged according to the position of the beam, and the damage state of the beam structure is the objective variable, and the method further provides an explanatory variable in which image data in which images representing the degree of deterioration of each beam are arranged according to the position of the beam, the image data being resized to a set number of pixels.

[0025] The trained model may be generated by an algorithm using an autoencoder or a convolutional autoencoder that compresses image data in which an image representing the deterioration degree of each beam is arranged according to the position of the beam, and performs a restoration process after retaining features.

[0026] The present invention also provides a trained model generated by the above-mentioned generation method. Effect of the Invention

[0027] According to the present invention, even if image data is given as input data, appropriate output data can be output, so that there is no need to convert it into numerical information, and prediction results can be obtained quickly. [Brief description of the drawings]

[0028] [Figure 1] FIG. 1 is a block diagram illustrating a functional configuration of a damage level prediction device according to an embodiment. [Diagram 2] FIG. 1 is a block diagram illustrating a hardware configuration of a damage level prediction device according to an embodiment. [Diagram 3] FIG. 2 is a diagram illustrating a structure of a machine learning model used in a damage degree prediction device according to an embodiment. [Figure 4] 1 is a flowchart illustrating a process in a learning device for preparing teacher data to be used in a damage level prediction device according to an embodiment. [Diagram 5] FIG. 13 is a diagram illustrating an example of criteria for determining the degree of deterioration of a pier. [Figure 6] FIG. 1 is a diagram illustrating criteria for determining bridge soundness. [Figure 7] 1A and 1B are diagrams for explaining how to handle a beam structure in one embodiment. [Figure 8] A figure illustrating image data in which images representing the deterioration level of each beam are arranged according to the position of the beam, which is explanatory variables in machine learning or input data during prediction processing in a damage level prediction device according to one embodiment. [Figure 9] 11A and 11B are diagrams illustrating examples of information indicating a planar shape of a beam structure. [Figure 10] FIG. 1 is a diagram illustrating external force conditions on a pier. [Figure 11] FIG. 11 is a diagram illustrating a result of a structural analysis according to an embodiment. [Figure 12] 1 is a flowchart illustrating a learning process of a machine learning model in one embodiment. [Figure 13] FIG. 1 is a diagram for explaining a problem in a damage prediction device of the prior art. [Figure 14] FIG. 1 is a diagram for explaining a problem in a damage prediction device of the prior art. [Figure 15]13A and 13B are diagrams for explaining a process of enlarging the size of image data showing a beam structure so that the image data has a set number of pixels, in the damage level prediction device according to an embodiment. [Figure 16] 11A and 11B are diagrams for explaining nearest neighbor interpolation processing when image data showing a beam structure is enlarged in a damage degree prediction device according to an embodiment. [Figure 17] 11A and 11B are diagrams for explaining how pixel values ​​are determined when image data showing a beam structure is enlarged in a damage level prediction device according to an embodiment. [Figure 18] FIG. 1 is a diagram for explaining the processing of a convolutional autoencoder in a damage degree prediction device according to an embodiment. [Figure 19] 11 is a sequence chart illustrating a damage level prediction process according to an embodiment. [Figure 20] FIG. 13 is a diagram illustrating information indicating a predicted damage state in one embodiment. [Figure 21] FIG. 11 is a block diagram illustrating a functional configuration of a damage degree prediction device according to a modified example. [Figure 22] FIG. 13 is a diagram illustrating an example of a dimensionless distance indicating the position of a beam in a modified example. [Figure 23] 13A and 13B are diagrams illustrating examples of identification information given to each beam in a multi-layer beam structure according to a modified example. [Figure 24] FIG. 1 illustrates an irregular beam structure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0029] 1. Configuration FIG. 1 is a diagram illustrating a functional configuration of a damage prediction device 1 according to an embodiment. The damage prediction device 1 is a system that predicts damage when an external force is applied to a target structure. The target structure is a beam structure of a pier or bridge. Here, a "beam" refers to a rod-shaped structure that is placed between two piles (or supports) of the superstructure of a pier, or a structure in which one end of the structure is supported by one pile (or support), and is sometimes called a girder. The damage prediction device 1 predicts damage to the target structure using a trained model generated by machine learning without structural analysis of the structure.

[0030] The damage prediction device 1 is a client device or a user terminal that receives instructions or data input from a user and outputs information according to the input instructions or data. The damage prediction device 1 has a storage means 11, an input means 21, and an output means 31.

[0031] The storage means 11 stores various data and programs. In this embodiment, the data stored in the storage means 11 includes a trained model M1. The trained model M1 is a machine learning model that is trained by providing teacher data. The trained model M1 is a model that is trained and constructed using, for example, a neural network, a random forest, a decision tree, a support vector machine, or deep learning.

[0032] When constructing the trained model M1, the training data includes, as explanatory variables, at least image data in which images showing the deterioration degree of each beam in the beam structure of a pier or bridge are arranged according to the position of the beam. In addition, this training data includes, as objective variables, at least damage state data according to the damage state of the beam structure. This training data is data related to at least one pier or bridge. The trained model M1 is generated by a learning device 51.

[0033] The learning device 51 performs machine learning using teacher data to obtain a trained model M1. The learning device 51 has a storage means and a learning means. The storage means stores the teacher data and the trained model (or the model before learning). The learning means provides the model with the teacher data to perform machine learning to obtain a trained model. The learning device 51 outputs the trained model M1, which is the generated trained model, to the damage degree prediction device 1.

[0034] In FIG. 1, input means 21 accepts input of input data to be input to trained model M1. Storage means 11 provides the input data input via input means 21 to trained model M1 and obtains a corresponding output result. The input data input to trained model M1 includes at least image data showing a beam structure of a pier or bridge that is the target of damage prediction. The output result (i.e., the prediction result) obtained from trained model M1 in response to this input data includes at least damage state data corresponding to the damage state of the beam structure that is the target of damage prediction. Output means 31 outputs the damage state data of the beam structure that is the target of damage prediction obtained by storage means 11.

[0035] FIG. 2 is a diagram illustrating a hardware configuration of the damage prediction device 1. The damage prediction device 1 is a computer device, such as a personal computer, having a CPU (Central Processing Unit) 101, a memory 102, a storage 103, an input device 104, and a display device 105. The CPU 101 is a control device that executes a program to perform various calculations and controls other hardware elements in the damage prediction device 1. The memory 102 is a main storage device that functions as a work area when the CPU 101 executes a program, and includes, for example, a RAM. The storage 103 is a non-volatile storage device that records various programs and data, and includes, for example, an SSD (Solid State Drive) or an HDD (Hard Disk Drive). The input device 104 is a device for a user to input instructions or data to the damage prediction device 1, and includes, for example, at least one of a touch screen, a keyboard, and a microphone. The display device 105 is a device that presents information to a user, and includes, for example, an OLED (Organic Light Emitting Diode) display or an LCD (Liquid Crystal Display).

[0036] In this example, the programs stored in the storage 103 include a program (hereinafter referred to as a "client program") for causing a computer device to function as the damage degree prediction device 1. The CPU 101 executes the client program, thereby implementing the functions of the damage degree prediction device 1 in the computer device. When the CPU 101 is executing the client program, at least one of the memory 102 and the storage 103 is an example of a storage means 11. The input device 104 is an example of an input means 21. The display device 105 is an example of an output means 31.

[0037] Although the hardware configuration of the learning device 51 (FIG. 1) is not shown, it is a computer device having the same hardware configuration as the damage degree prediction device 1. The learning device 51 may be implemented in the same computer device as the damage degree prediction device 1, or may be implemented in a computer device other than the damage degree prediction device 1, for example, a server device on a network.

[0038] 2.Operation Before describing the operation of the damage degree prediction device 1, an overview of the machine learning model used in this embodiment will be given.

[0039] FIG. 3 is a diagram illustrating the structure of a machine learning model according to an embodiment, illustrating a neural network. In this example, the machine learning model 90 has an input layer 91, an intermediate layer 92, and an output layer 93. In this example, when generating a trained model M1, in addition to (1) image data in which an image showing the deterioration degree of each beam in a beam structure of a pier or bridge is arranged according to the position of the beam, (2) external force conditions acting on a pier or bridge including the beam structure, (3) data related to the beam structure of the beam, and (4) numerical data of the beam in the beam structure are input to the input layer 91. Damage state data showing the damage state of the beam structure obtained from the structural analysis result of the beam structure is output to the output layer 93.

[0040] The damage state data includes at least one of information on the extent of damage, the degree of damage, and the percentage of damage. More specifically, the information indicating the damage state includes, for example, a map of the distribution of the degree of damage.

[0041] The intermediate layer 92 is a layer located between the input layer 91 and the output layer 93, and includes parameters for obtaining data to be output via the output layer 93 from data input via the input layer 91. When generating the trained model M1, the parameters are trained using the input data input to the input layer 91 as described above and training data having data indicating the damage state of each beam in the beam structure as output data. In this example, the "damage state" includes at least one of the damage degree, damage distribution, and damage area ratio.

[0042] The operations of the damage prediction device 1 and the learning device 51 are described below. The operations of the damage prediction device 1 and the learning device 51 are roughly divided into three stages: preparation of teacher data, generation of a trained model M1 (i.e., learning), and prediction using the trained model M1. Each of the processes of preparation of teacher data, learning, and prediction is described below. In the following, functional elements such as each device and each means may be described as the subject of processing, which means that a hardware element such as the CPU 101 executes processing in cooperation with other hardware elements according to a program. In addition, a pier is used as an example of a target structure in the following description.

[0043] 2-1. Preparation of training data In machine learning in the learning device 51, structural analysis is performed on piers or bridges of various sizes, each of which is given multiple analysis conditions (e.g., conditions related to deterioration distribution and external forces). For the structural analysis, known structural analysis software (e.g., "Engineer's Studio" by FORUM8 Co., Ltd.) is used. As the result of the structural analysis, data on the damage distribution, the degree of damage, and the damage area ratio, which is the area ratio of damage to the total area of ​​the pier or bridge superstructure, are obtained as the damage state. These data are used as training data.

[0044] 4 is a flowchart illustrating a process of preparing teacher data in the learning device 51. First, the learning device 51 acquires image data in which images representing the deterioration degree of each beam in the beam structure of a pier or bridge are arranged according to the position of the beam (step S101). The images representing the deterioration degree of each beam are created based on the result of a deterioration degree determined by a person observing the actual beam structure at the pier or bridge site. The images representing the deterioration degree of each beam may also be created based on the result of a deterioration degree determined based on an image of the beam photographed at the site.

[0045] FIG. 5 is a diagram illustrating an example of a criterion for judging the deterioration degree of a pier (cited from Non-Patent Document 1). In this example, a criterion is set for each of the side and bottom parts of the beam. For both the side and bottom parts, four categories of deterioration degree are defined: a, b, c, and d. Deterioration degree a is the most severe (or highest) state of deterioration, and deterioration degree d is the least severe or no deterioration (or lowest) state of deterioration. A beam structure is divided into a plurality of beams, and the deterioration degree of each beam is judged. The learning device 51 recognizes the bottom and side parts of the image of a target beam in the beam structure, and judges the deterioration degree of the beam by checking it against the judgment criterion. In this example, the learning device 51 judges the deterioration degrees of the bottom and side parts of the beam, respectively, and adopts the heavier deterioration degree as the deterioration degree of the beam. For example, if the deterioration degree of the bottom is c, and the deterioration degree of one of the two sides is b and the other is d, the deterioration degree of the beam is b.

[0046] If the target structure is a bridge (e.g., a road bridge) rather than a pier, an indicator of soundness is used instead of the degree of deterioration.

[0047] FIG. 6 is a diagram illustrating criteria for judging the soundness of bridge structures (cited from "Guidelines for Periodic Inspection of Road Bridges, Ministry of Land, Infrastructure, Transport and Tourism, Road Bureau, February 2019"). In this example, four categories, I, II, III, and IV, are defined for the soundness of road bridges. Category I is the healthiest (or least deteriorated) state, and category IV is the most deteriorated state. The learning device 51 recognizes an image of a target beam in a beam structure and judges the soundness of the beam by checking it against the judgment criteria. Note that the judgment criteria for the deterioration degree and soundness illustrated in FIG. 5 and FIG. 6 are merely examples, and other criteria may be used. The categories of deterioration degree and soundness are not limited to four, and may be divided into five or more or three or less.

[0048] Returning to Fig. 4, the person judging the deterioration degree judges the deterioration degree of each beam based on the judgment criteria as described above, images representing the deterioration degree of the beams are created based on the judgment, and image data is created in which the images are arranged according to the position of each beam in the beam structure. Then, in step S101, the learning device 51 acquires the created image data.

[0049] The image data acquired by the learning device 51 will be described below. Fig. 7 is a diagram for explaining how to handle a beam structure. Fig. 7(A) is a diagram illustrating a beam structure. In this diagram, gray rectangles represent beams (or girders), black squares represent piles, and white rectangles represent floorboards. In an actual beam structure, the beams, piles, and floorboards each have different sizes, but here, for the purpose of facilitating handling as data or facilitating illustration, they are all standardized to squares of the same size.

[0050] FIG. 7(B) shows a standardized beam structure. In this beam structure, the white square in the upper left of the figure is set as the origin, the x-axis is set in the left-right direction on the paper surface (or display surface), and the y-axis is set in the up-down direction on the paper surface (or display surface), and the direction toward the right on the paper surface (or display surface) is the x-direction, and the direction toward the bottom is the y-direction. Hereinafter, the x-direction and y-direction in other figures will be described in the same manner as in this description, and for example, the beam located third from the left and second from the top will be represented as [3,2]. FIG. 7(A) also shows arrows indicating the directions corresponding to the x-direction and y-direction in FIG. 7(B). Note that the x-direction and y-direction shown by the arrows in FIG. 7 indicate the directions of the x-axis and y-axis, but the position where the two arrows intersect does not indicate the origin of the coordinate axes.

[0051] Fig. 8 is a diagram in which the deterioration degree judgment results are mapped for each beam of the beam structure standardized as in Fig. 7(B). That is, image S in Fig. 8 is an image represented by image data created based on the results of judging the deterioration degree of each beam at the pier or bridge site. Hereinafter, each of the standardized square areas representing beams, piles, and floorboards as shown in Fig. 7(B) will be referred to as a pixel. Each of the square areas representing beams, piles, and floorboards as shown in Fig. 8 is also a pixel.

[0052] In image S shown in Fig. 8, letters a to d indicating the degree of deterioration are displayed in each pixel corresponding to a beam. Learning device 51 acquires image data indicating image S as an explanatory variable. That is, image S in Fig. 8 shows an example of image data in which images indicating the degree of deterioration of each beam are arranged according to the position of the beam, acquired by learning device 51 in step S101 in Fig. 4. Learning device 51 acquires the degree of deterioration (a to d) of each beam shown in image S as an explanatory variable.

[0053] Returning to FIG. 4, the learning device 51 then acquires data on the beam structure of each beam (step S102). The data on the beam structure includes the number of beam rows constituting the pier or bridge. The number of beam rows indicates, for example, the number of beam rows provided in the normal direction of the pier or bridge. Note that the process of step S102 does not have to be executed, and the data on the beam structure is not necessarily required for machine learning in the learning device 51.

[0054] Next, the learning device 51 acquires numerical data of each beam (step S103). The numerical data of the beam includes at least one of the following: the beam dimensions, the presence or absence of piles at both ends of the beam, the length or protruding length of the piles supporting the beam, and the characteristic values ​​of the piles. Note that the process of step S103 does not have to be executed, and the numerical data of the beams is not necessarily required for machine learning in the learning device 51.

[0055] 9A and 9B are diagrams illustrating information indicating the planar shape of a beam structure, where FIG. 9A is a diagram illustrating position data of the beams, and FIG. 9B is a diagram illustrating numerical data of the beams.

[0056] 9A, identification numbers are assigned to the vertical and horizontal beams, respectively. These identification numbers indicate serial numbers that are position data acquired by the learning device 51 in step S104 of FIG.

[0057] In FIG. 9(A), the x-axis is set in the left-right direction of the paper (or display surface), the y-axis is set in the up-down direction, and the z-axis is set in the vertical direction to the paper (or display surface), and the direction toward the right on the paper (or display surface) is set as the x-direction, the direction toward the bottom is set as the y-direction, and the direction toward the vertical upward direction to the paper (or display surface) is set as the z-direction. Note that the direction toward the left may be set as the x-direction, the direction toward the top may be set as the y-direction, or the direction toward the vertical downward direction may be set as the z-direction. In FIG. 9(A), a circle with a "·" inside it indicating the z-direction indicates an arrow pointing from the back to the front of the paper (or display surface). Note that the x-direction, y-direction, and z-direction indicated by the arrows in FIG. 9(A) indicate the directions of the x-axis, y-axis, and z-axis, but the position where the arrows intersect does not indicate the origin of the coordinate axes.

[0058] The table in Figure 9(B) contains information indicating the size of the target beam structure (the number of beams in each direction) and information indicating the dimensions of each beam. This beam structure is composed of seven beams in the x direction and five beams in the y direction. For example, beam number 101 has a length of 4.40 m in the x direction, a length of 0.50 m in the y direction, and a height of 0.50 m in the z direction. These values ​​indicate the numerical data of the beams.

[0059] Returning to FIG. 4, the learning device 51 acquires the external force conditions to be applied to the target beam structure (step S104).

[0060] FIG. 10 is a diagram illustrating an example of external force conditions at a pier. Note that the external force conditions at a bridge may be defined separately (examples are omitted). The types of external forces are not limited to those shown here. The rows of this table indicate external force conditions, and the columns indicate explanatory variables. In this example, five types of explanatory variables are used: static (sea ← land), static (sea → land), L1, L2-1, and L2-2. "Static (sea ← land)" indicates, for example, the tractive force of a specific ship. "Static (sea → land)" indicates, for example, the docking force of a specific ship. "L1" indicates level 1 earthquake motion, that is, an earthquake with a high probability of occurring during the service life. "L2" indicates level 2 earthquake motion, that is, an earthquake with a low probability of occurring during the service life but that releases a large amount of energy, and L2-1 indicates type I: level 2 earthquake motion of plate boundary type (for example, the Tohoku Pacific Ocean Earthquake, etc.). L2-2 indicates type II: level 2 earthquake motion of inland direct hit type (for example, the Hyogo-ken Nanbu Earthquake, etc.). In addition, the static (sea ← land) and static (sea → land) listed as external force conditions are not only used as the towing and docking forces of a specified ship, but are also used as static design loads, making it possible to check whether damage that should not occur on a sound pier occurs. In addition, the external force conditions L1 and L2 can also be used as design loads depending on the importance of the pier being designed.

[0061] In the table of FIG. 10, the value of each cell is either "0" or "1". A value of "0" indicates that the external force is not applied, and a value of "1" indicates that the external force is applied. For example, the external force condition "static (sea ← land)" indicates that only the towing force of a ship, for example, is applied to the target beam structure, and no other external forces are applied. Note that in this example, only an external force condition in which a single type of external force is applied is shown, but an external force condition in which multiple types of external forces are applied may also be used. Specifically, for example, an external force condition in which a static (sea ← land) external force and a level 1 earthquake motion are applied simultaneously may be used.

[0062] 4, the learning device 51 acquires a structural analysis result assuming that an external force indicated by the external force condition acquired in step S104 is applied to the beam and beam structure based on the data acquired in steps S101 to S103 (step S105). The structural analysis may be performed by the learning device 51 itself or by a device other than the learning device 51.

[0063] FIG. 11 is a diagram illustrating the results of the structural analysis. In this example, the damage state data of the results of the structural analysis includes the degree of damage, the distribution of the degree of damage, and the area of ​​the damage or the percentage of the area of ​​the damage. In one example, the results of the structural analysis are mapped onto an image of a standardized beam structure. In this map, all beams have a common size, and may not include information indicating the actual size of the beam structure, or each beam may be displayed at a scale corresponding to its actual size.

[0064] The damage area ratio means the ratio of the number of damaged beams to the number of beams in the beam structure (or the ratio of damaged and undamaged beams), or the area ratio of the beams in the beam structure that have a specified damage to the total area of ​​the beams in the beam structure. Alternatively, the damage area ratio may mean the area ratio of the beams in the beam structure that have a specified damage to the total area of ​​the pier superstructure.

[0065] In this map, the degree of damage to each beam is displayed based on a predefined damage classification, with color, hatching, or text information to indicate the type of damage. Since color cannot be displayed in Figure 11, hatching is used instead. For example, crack damage is represented in yellow, yield damage in red, and ultimate damage in black. As these damages are displayed on a map of the beam structure, this figure can be said to show the extent and degree of damage.

[0066] This map is converted into information including pairs of beam identification information and damage levels. Alternatively, as a result of the structural analysis, information including pairs of beam identification information and damage levels is output in addition to or instead of the map in FIG.

[0067] The learning device 51 repeatedly performs the processes of steps S101 to S105 in Fig. 4 for each of a plurality of piers having various shapes and sizes, and further for the beams that constitute each pier, and accumulates data. The data thus accumulated is used as teacher data in the learning process described below.

[0068] In the learning process by the learning device 51, in the above-mentioned teacher data, the explanatory variables are the image data in which images showing the deterioration degree of each beam in the beam structure obtained in step S101 of Figure 4 are arranged according to the position of the beam, the data related to the beam structure obtained in step S102, the numerical data of the beams obtained in step S103, and the external force conditions obtained in step S104, and the structural analysis results obtained in step S105 are the objective variables. The learning device 51 stores these teacher data in its own device or in an external storage means.

[0069] 2-2. Learning Fig. 12 is a flowchart illustrating a learning process of a machine learning model. The flow of Fig. 12 is started, for example, when a learning instruction is given by an administrator or a user of the damage degree prediction device 1. Alternatively, the flow of Fig. 12 may be started when a predetermined amount of teacher data is accumulated. The learning device 51 performs the learning process according to the flowchart of Fig. 12.

[0070] First, the learning device 51 acquires teacher data to be used for learning (step S301). In this example, the learning device 51 acquires teacher data prepared in an external device other than the damage level prediction device 1.

[0071] Next, the learning device 51 provides training data to the machine learning model to perform machine learning (step S302). This training data includes the explanatory variables and the objective variables described above.

[0072] Next, the learning device 51 stores the machine learning model in which the parameters of the intermediate layer 92 have been learned by machine learning, as a learned model in the storage means (step S303).

[0073] Then, the learning device 51 outputs the generated trained model M1 to the damage degree prediction device 1. The damage degree prediction device 1 stores the trained model M1 in the storage means 11. In this way, the damage degree prediction device 1 obtains the trained model M1.

[0074] The machine learning in the learning device 51 will be further explained below. First, the problems with the technology described in the above-mentioned Patent Document 1 (hereinafter referred to as the "prior art") will be described. Fig. 13 is a diagram for explaining the problems with the prior art. This prior art relates to a damage prediction device that learns the shape and deterioration degree judgment results of a pier and the damage results obtained by structural analysis as image information, and predicts the damage results from the deterioration degree judgment results of an unknown pier.

[0075] In the prior art damage prediction device, when making a prediction for an unknown pier shape (a pier shape that has never been learned), the pier shape may be erroneously recognized. Fig. 13(A) is a diagram showing the actual shape of the pier to be predicted (i.e., the correct shape), with damage information from structural analysis added to the overall shape of the pier. If such a pier shape is unknown to the trained model, inputting the diagram of the shape shown in Fig. 13(A) into the trained model and making a prediction may result in the erroneous output of a pier shape that is similar in terms of the number of beam rows, etc., from among the trained pier shapes.

[0076] For example, Figure 13(B) shows the prediction results output by the trained model. This is the result of outputting data for a pier with a different shape than Figure 13(A). Because the shape is incorrect, the damage prediction will also be incorrect.

[0077] Here, the reason why the prior art damage prediction device erroneously recognizes the shape of a pier will be explained. FIG. 14 is a diagram for explaining a problem with the prior art damage prediction device, and is a diagram for explaining erroneous recognition of the shape of a pier.

[0078] Fig. 14(A) is an example of a simplified display of the pier shape shown in Fig. 13(A). Since the focus is only on the shape, the beams are displayed in white and the other parts (piles and floorboards) are displayed in black. Fig. 14(B) shows each of the beams, piles, and floorboards in the pier shape of Fig. 14(A) as standardized square areas (pixels). The beams are displayed as white pixels and the other parts (piles and floorboards) are displayed as black pixels.

[0079] In the machine learning model, the length and width of the vector, which is the size of the image data, must be fixed. Each of the multiple piers has a different shape and size, and an image formed by square pixels of a standardized size as shown in FIG. 14(B) will have different length and width lengths depending on the pier. In machine learning, the size of the image data as shown in FIG. 14(B) is corrected and read so that it fits into a fixed size (a size in which the number of pixels in the vertical and horizontal directions is fixed to a preset number).

[0080] Fig. 14(C) shows an example of image data of a pier with a fixed size. An image corresponding to a pier, which is the maximum size, can be set as a fixed size as shown in Fig. 14(C). Fig. 14(D) shows an example of image data in which the size of an image of a pier smaller than the maximum size of the pier in Fig. 14(C) has been corrected to fit into a fixed size. As shown in Fig. 14(D), when fitted into a fixed frame indicating a fixed size, the area outside the pier is treated as image data with black pixels, similar to parts other than the beams (piles, floorboards, etc.).

[0081] Fig. 14(E) shows an example of image data in which the size of a pier smaller than the maximum size pier in Fig. 14(C) and larger than the pier in Fig. 14(D) has been corrected to fit into a fixed size. As in the case of Fig. 14(D), the area outside the range of the pier when fitted into the fixed frame is treated as image data with black pixels. As described above, image data of various piers such as those shown in Figures 14(C), (D), and (E) (image data formed by pixels as described above) are loaded into the learning device and subjected to machine learning.

[0082] FIG. 14(F) shows an example of image data in which the size of the image data of the pier, which is the target of damage prediction, has been corrected to fit a fixed size. The learning device and the damage prediction device only recognize the image of the loaded image data, and therefore cannot directly recognize the size of the pier. Therefore, when trying to predict the damage level of the pier in the image data of FIG. 14(F) using a trained model trained with the image data of FIG. 14(C), (D), and (E), it is possible that, among the images of piers that have been trained so far, the image of the pier in FIG. 14(E) is recognized as having a similar shape, and the damage level is predicted based on the shape of the pier in FIG. 14(E) and the result is output. It is believed that such misrecognition can be eliminated to some extent by loading and training images of piers of various shapes, but it is not always possible to correctly recognize the number of beams or piles, and there is a possibility that misrecognition may occur.

[0083] In this embodiment, during learning and prediction (described later), a large fixed frame (e.g., 1,000 x 1,000 pixels, etc.) is set, and the image of the pier shape (with the beams, piles, and floorboards each converted into standardized square pixels) is enlarged vertically and horizontally to fit into the fixed frame. By carrying out such processing, an image of the pier shape without margins (black areas other than the pier area in the figure) as present in the images of Figures 14(D) and (E) is read, and learning or prediction is performed, making it possible to output the correct pier shape.

[0084] Figure 15 is a diagram for explaining the process of enlarging the size of image data showing the shape (beam structure) of a pier to a set number of pixels. In each of Figures 15(A) and (B), the diagram on the left is an example of image data before enlargement processing, and the diagram on the right is an example of image data after enlargement processing.

[0085] The diagrams on the left side of Figures 15(A) and (B) show an example in which images S1 and S2 (images shown in pixels as in Figure 14(B)) of a pier shape are arranged within an image showing a fixed frame F of a set number of pixels. Image S2 is an image smaller than image S1, and shows the shape of a smaller pier.

[0086] In the images S1 and S2, the pixels corresponding to the beams among the pixels of each pixel display information on the deterioration degree of the beams that has been determined in advance. The deterioration degree may be displayed identifiably by letters such as a, b, c, and d as in the above-mentioned FIG. 8, but in the images S1 and S2 in FIG. 15(A) and (B), it is displayed identifiably by light and dark colors. In the case of color display, it may be displayed identifiably by different colors.

[0087] The images on the right side of Fig. 15(A)(B) show the state in which images S1 and S2 have been enlarged vertically and horizontally to the same size as fixed frame F. When such an image is simply enlarged, the image on the right side of Fig. 15(A)(B) after enlargement (enlarged image) may have a situation in which multiple structural elements exist within one pixel of the enlarged image. For this reason, in this embodiment, a process is performed to determine the pixel value after enlargement by nearest neighbor interpolation. Here, structural elements refer to, for example, beams, piles, decks, etc. that constitute a pier or bridge.

[0088] Fig. 16 is a diagram for explaining determination of pixel values ​​by nearest neighbor interpolation. Fig. 16(A) shows a pixel in which multiple structuring elements exist in an image enlarged by simple enlargement (for example, the image on the right side of Fig. 15(A)(B)). Assume that pixel P in Fig. 16(A) has two structuring elements: a beam represented by a dark area R1, and a pile represented by a light area R2. Since one pixel can represent only one structuring element, in such a case, the pixel value is determined to be one of the multiple structuring elements (two in the case of Fig. 16(A)).

[0089] In such a case, the method for determining the pixel value is to adopt the structuring element of the pixel closest to the pixel as the pixel value of the pixel. In other words, the structuring element that occupies the largest proportion within the pixel when simply enlarged is determined as the pixel value of the pixel. For example, in FIG. 16(A), of the two structuring elements, the proportion of the beam region R2 is larger, so pixel P is determined to be a beam (shown in a light color) as shown in FIG. 16(B). Even if there are three or more structuring elements, the structuring element with the largest proportion should be determined.

[0090] The pixel value may be determined in a different manner than the above, for example, based on the types of structuring elements present in the enlarged pixel. For example, for multiple structuring elements present within a pixel, the pixel value can be trained to be one of the structuring elements, and the accuracy of each of the trained models constructed when each structuring element is used can be checked. The structuring element with the highest accuracy can then be used as the pixel value of the pixel.

[0091] In the above-mentioned method, it becomes difficult to determine the pixel value when the proportions of the two structuring elements in FIG. 16(A) are the same, 0.5:0.5, for example. In order to solve this problem, the following method may be adopted.

[0092] For example, if two structural elements, a beam and a pile, are present in a single pixel with equal proportions, a trained model in which the pixel is trained as a beam and another trained model in which the pixel is trained as a pile are generated. It is confirmed which of the two trained models has higher accuracy, and if, for example, the trained model trained as a beam has higher accuracy, the pixel is determined to be a beam. Even if three or more structural elements exist, trained models for each structural element are generated, and the structural element of the trained model with the highest accuracy is determined to be the structural element of the pixel.

[0093] Furthermore, since the original image before enlargement represents one structural element by one pixel, by enlarging the image, the structural element represented by one pixel is represented by multiple pixels. Therefore, it is expected that different damage results will be output for each of the multiple pixels. This is because it is difficult for the learning device 51 or the damage prediction device 1 to recognize that the multiple pixels are one structural element. In such a case, for example, processing may be performed by outputting the average value of each pixel value of the multiple pixels as a representative value, or outputting the most severe damage result (result with the most severe damage) of the damage results of each of the multiple pixels as a representative value.

[0094] FIG. 17 is a diagram for explaining how to determine pixel values ​​when one structural element is represented by multiple pixels after the image is enlarged. FIG. 17(A) shows one structural element (e.g., a beam) in the enlarged image, and shows a state in which this one structural element is represented by four pixels P1, P2, P3, and P4. The pixels P1, P2, and P3, which are represented in a dark color, are more severely damaged than the pixel P4, which is represented in a light color. For example, the pixels P1, P2, and P3 are in a "yield" state, and the pixel P4 is in a "crack" state. In this case, since there are many "yield" pixels, when the four pixel values ​​are averaged, the average pixel value indicates "yield". Therefore, the representative value of the four pixels representing the structural element in FIG. 17(A) is set to "yield", and all four pixels are set to a "yield" state, which is represented in a dark color, as shown in FIG. 17(B).

[0095] FIG. 17(C) is another example from FIG. 17(A) showing one structural element (e.g., a beam) in the enlarged image, with pixels P1 and P4 showing the "crack" state and pixels P2 and P3 showing the "yield" state. In this example, the pixel value showing the most severe damage state is taken as the representative value for this structural element. In other words, since "yield" is more severe (has greater damage) than "crack", the pixel value of this structural element is taken as "yield", and by making all four pixels the "yield" state, which is represented in a dark color as shown in FIG. 17(D), the damage state of the structural element is determined to be "yield".

[0096] In the state of Figure 17(C), if an attempt is made to determine a representative pixel value using the method described in Figure 17(A), the average value cannot be determined because there are an equal number of pixels indicating "cracks" and pixels indicating "yield." However, using this method, it is possible to determine the average value as described above. The above methods are examples for determining the representative value, and the present invention is not limited to these methods.

[0097] Next, machine learning based on structural analysis in the learning device 51 will be described. As for machine learning techniques, learning can be performed using techniques such as CNN (convolutional neural network).

[0098] In this embodiment, a structure simulating an autoencoder is used as a machine learning algorithm for an image showing a pier shape. An autoencoder is a neural network model that learns to make output data the same as input data. In a model configured by overlapping fully connected layers from an input layer to an intermediate layer and an output layer, the number of dimensions of the intermediate layer is made smaller than the number of dimensions of the input and output, and learning is performed so that data that is the same as the input can be restored. This makes it possible to obtain feature amounts equivalent to those obtained by compressing the dimensions of the input image in the intermediate layer.

[0099] In this embodiment, a convolutional autoencoder is used that incorporates a convolutional layer and a pooling layer structure into a fully connected layer. Figure 18 is a diagram for explaining the processing of a convolutional autoencoder. A convolutional autoencoder is a neural network model that adds a convolutional layer that captures the features of the input data and a pooling layer that reduces the dependency on the features between the input layer and the output layer.

[0100] In Fig. 18, an image showing the degree of degradation is input to an input layer 91. The input image is filtered in a convolution layer 95 to extract image features. Then, a pooling layer 96 performs a predetermined calculation to reduce the size of the feature map. A feature vector according to an external force condition is added to the feature generated in the pooling layer 96.

[0101] The deconvolution layer 97 performs a restoration process to restore the size of the reduced image. The fully connected layer 98 is for consolidating the outputs from the deconvolution layer 97, and is similar to the intermediate layer 92 in Fig. 3 described above. Then, the image of the analysis result is output to the output layer 93. In this way, a damage prediction model is constructed. In this embodiment, a convolutional autoencoder is used, but the present invention is not limited to this, and other autoencoders may be used.

[0102] 2-3. Prediction 19 is a sequence chart illustrating a damage state prediction process by the damage level prediction device 1. The damage state prediction is performed in response to a request from a user.

[0103] 19, a user instructs the damage prediction device 1 to predict a damage level. Specifically, the user inputs input data for a beam structure of a structure to be subjected to damage prediction. The input means 21 receives, as input data, image data representing the deterioration level of each beam in the beam structure, resized to a set number of pixels, and external force conditions acting on a pier or bridge including the beam structure to be subjected to damage prediction (step S401).

[0104] In one example, the damage prediction device 1 itself provides a user interface for inputting the inspection results of the structure. The user refers to the judgment results of the deterioration degree of each beam of the structure, creates image data (hereinafter referred to as first image data) showing the deterioration degree of each beam as shown in FIG.

[0105] The damage degree prediction device 1 changes the input first image data (for example, image data as shown in FIG. 8) to image data (hereinafter referred to as second image data) whose size is enlarged to a set number of pixels as shown in FIGS. 15(B) and (D) in the same manner as described in the learning process by the learning device 51. Then, the second image data is input to the input means 21.

[0106] The process of changing the size from the first image data to the second image data may be performed within the damage degree prediction device 1 before inputting the data to the input means 21, or the first image data may be input to the input means 21 and then changed to the second image data within the input means 21. Furthermore, the process of changing the size from the first image data to the second image data may be performed by a device other than the damage degree prediction device 1, and the second image data may be input to the input means 21 of the damage degree prediction device 1.

[0107] When generating second image data by enlarging the first image data, a situation may arise in which multiple structural elements (beams, piles, floorboards, etc.) exist within one pixel of the enlarged second image data, as in the example described in Fig. 16. In such a case, as in the case of machine learning described above, the structural element that occupies the largest proportion within that pixel may be determined as the structural element of that pixel. Furthermore, in order to deal with the case where there are structuring elements that occupy a similar proportion of a pixel, the following method may be adopted.

[0108] Of the multiple structural elements present in a pixel, the pixel is provisionally determined as one of the structural elements, and a damage degree prediction is performed. A damage degree prediction is performed in the same manner for each of the multiple structural elements. Then, the structural element that results in the greatest damage among the prediction results is determined as the structural element for that pixel. For example, if two structural elements, a beam and a pile, exist in a single pixel and occupy an equal proportion of them, the prediction result of the damage degree prediction in which the pixel is assumed to be a beam is compared with the prediction result of the damage degree prediction in which the pixel is assumed to be a pile, and the structural element that results in the greatest damage (the most severe damage result) for that pixel is determined as the structural element for that pixel. Even if three or more structural elements exist, the structural element that results in the most severe damage result for that pixel among the prediction results when each structural element is used is determined as the structural element for that pixel.

[0109] Furthermore, when the second image data is generated by enlarging the first image data, a structural element represented by one pixel will be represented by multiple pixels in the enlarged second image data, similar to the example described in FIG. 17. For example, a pixel corresponding to one beam in the first image data will be represented by multiple pixels in the second image data. In such a case, the multiple pixels may each have a different damage result (pixel value) depending on the result of damage prediction. Since the multiple pixels are originally one beam, they need to be output as one damage result.

[0110] Therefore, the average value of each of the pixels can be set as the representative value for the pixels. For example, in the case of Fig. 17(A), the damage result of the damage prediction, called "yield" shown in dark color, can be set as the pixel values ​​of four pixels as shown in Fig. 17(B).

[0111] Alternatively, the most severe damage result (result with the greatest damage) among the damage results of each of the multiple pixels may be used as the representative value. In the above-mentioned method of determining the representative value of pixel values ​​by the average value, for example, as shown in Figure 17(C), if there are an equal number of pixels indicating "cracks" and pixels indicating "yield", the average value cannot be determined, but this method makes it possible to determine the average value. The above methods are examples for determining the representative value, and the present invention is not limited to these methods.

[0112] Next, the input means 21 generates a request to output damage state data indicating the damage state of the beam structure of the damage degree prediction target (step S402). This request includes the input data input, acquired, or specified in step S401. Next, the input means 21 requests the damage state data of the beam structure of the damage degree prediction target from the storage means 11 (step S403). Note that in this example, it can be said that the input means 21 essentially inputs the received input data to the trained model M1.

[0113] Next, the storage means 11 receives a request to output damage state data of the beam structure to be predicted (step S404). The storage means 11 inputs the input data included in the received request to the trained model M1 (step S405). The trained model M1 outputs an output result, that is, damage state data of the beam structure to be predicted according to the input data. The damage state data is data in the same format as the map of damage distribution mapped on the image of the beam structure used as the teacher data, as exemplified in FIG. 11. Next, the storage means 11 acquires the damage state data, which is the output result output from the trained model M1 (step S406). The storage means 11 outputs the acquired output result to the output means 31 (step S407). The output means 31 inputs the damage state data, which is the output result, and outputs information indicating the predicted damage state according to the input output result (step S408). The output means 31 may output the information output from the trained model M1 to the user (the device that sent the request) as is, or may output the information output from the trained model M1 after performing some processing on it.

[0114] Fig. 20 is a diagram illustrating an example of information indicating a predicted damage state output from the output means 31. In this example, the damage level (i.e., the degree of damage) of each beam is divided into a plurality of categories (i.e., damage categories), and each category is visually represented (for example, in a different color). The categories used here are the same as those illustrated in Fig. 11, and for example, crack damage is represented in yellow, yield damage in red, and ultimate damage in black. In addition, colors may be used to indicate the deterioration level displayed in Fig. 5 or the soundness displayed in Fig. 6.

[0115] The information output from the output means 31 also includes the area of ​​damage or the proportion of the area of ​​damage. Here, the proportion of the area of ​​damage refers to the proportion of the number of damaged beams out of the total number of beams in the beam structure of interest (or the proportion of damaged and undamaged beams), or the proportion of the area of ​​the beams having a given damage out of the total area of ​​the beams in the beam structure of interest. In the latter case, for example, if the total area of ​​the beams in the beam structure of interest is 5,000 m, 2 In the case where the total area of ​​the beams with at least some crack damage is 3,000 m 2 When this occurs, the crack area ratio is 60%. The damage area ratio may refer to the ratio of the area of ​​the beams of the beam structure in question to the area of ​​the entire pier superstructure.

[0116] The format of the information indicating the predicted damage state is not limited to the example in FIG. 20. The information indicating the damage state may include at least one of the following information: the range of damage, the degree of damage, and the damage ratio. For example, the output means 31 may not distinguish between damage categories, and may represent the presence or absence of damage as a binary value (i.e., damaged / not damaged). Alternatively, the output means 31 may not output the damage state in a map, but may output only the damage ratio.

[0117] As described above, according to the damage prediction device 1 of the present embodiment, even if image data representing the deterioration degree of the beam is given as input data, an appropriate output result can be output, so there is no need to convert the deterioration state of the beam into numerical data for input, and input processing can be performed quickly. Therefore, prediction results can be obtained quickly.

[0118] In addition, according to the damage prediction device 1, by storing a trained model generated using structural analysis, at the time of prediction, the damage state of the beam can be predicted (or evaluated) using the trained model from the deterioration degree judgment results (a to d) and the position information of the beam without performing structural analysis, and prediction results can be obtained quickly.

[0119] In addition, an image enlargement algorithm using nearest neighbor interpolation makes it possible to place an image of any pier shape within a fixed frame without any margins, so that damage results can be output using an image of the correct pier shape.

[0120] In addition, since the deterioration degree information of the beam can be captured as image data, it is not necessary to add conditions for the spatial position information of the beam. However, there are cases where accuracy is improved by reading the position information of the beam.

[0121] Furthermore, when attempting to obtain and grasp the damage state of multiple piers, the damage prediction device 1 can significantly reduce the labor and time required. For example, after evaluating the remaining strength of multiple piers using the damage prediction device 1, it is possible to utilize the damage prediction device 1 as a primary screening function by carrying out a more detailed structural analysis only on those with significant damage to grasp the specific damaged areas.

[0122] In addition, the damage prediction device 1 can obtain outputs according to various external forces (crane work load, berthing force, traction force, level 1 earthquake motion, level 2 earthquake motion, etc.). Furthermore, the damage level prediction device 1 can predict the damage state not only for piers but also for bridges by using, for example, road bridge soundness assessment categories.

[0123] Furthermore, the damage degree prediction device 1 can solve the problems of the technique (prior art) described in Patent Document 1, as explained in FIG. 13, and does not erroneously recognize the shape of the pier during prediction.

[0124] 3. Variations The present invention is not limited to the above-described embodiment, and various modifications are possible. Some modifications are described below. Two or more of the features described in the following modifications may be used in combination.

[0125] 3-1. Explanatory variables and input data In the above-described embodiment, the explanatory variables used when generating the trained model M1 were the external force conditions acting on a pier or bridge including beams, data related to the beam structure of the beams, and numerical data of the beams in the beam structure, but these data do not necessarily have to be used as explanatory variables.

[0126] In addition, in the above-described embodiment, when predicting the level of damage, the external force conditions acting on the pier or bridge including the beam structure to be predicted are used as input data, but the external force conditions do not necessarily have to be used as input data.

[0127] 3-2. Damage prediction device using a probabilistic model In the above-described embodiment, the damage prediction device may include a prediction means for predicting a future deterioration level or soundness using a predetermined probability model.

[0128] Fig. 21 is a block diagram illustrating a functional configuration of a damage degree prediction device according to a modified example, and shows a damage degree prediction device 1A having a configuration in which a prediction means is added to the block diagram illustrated in Fig. 1. In Fig. 21, the storage means 11, the input means 21, the output means 31, and the learning device 51 are the same as those illustrated in Fig. 1.

[0129] The prediction means 41 has a probability model M2, and predicts the deterioration degree of the beam at a future time point according to the probability model M2. The probability model M2 is a model that estimates a future state transition based on a probability based on a past state transition, and an example of the model is a Markov chain model. The prediction is performed by combining the trained model M1 and the probability model M2.

[0130] The input means 21 inputs input data similar to that of the damage prediction device 1 in Fig. 1, and also inputs data for the current time and the future time to be predicted input by the user. The prediction means 41 accepts the data input by the input means 21, predicts the deterioration level of the beam to be subjected to damage prediction at the future time using the probabilistic model M2, and outputs deterioration level data corresponding to the predicted future deterioration level to the trained model M1 in the storage means 11. The trained model M1 outputs damage state data relating to the damage state of the beam structure to be subjected to damage prediction, which is output data, according to the deterioration level data from the prediction means 41 and the input data input by the storage means 11 from the input means 21. The damage state data is output from the output means 31.

[0131] 3-3. Explanatory variables The explanatory variables used in the teacher data in learning and the input data in prediction are not limited to those exemplified in the embodiment. The explanatory variables may include at least one of the following (a) to (f). (a) The number of rows of beams perpendicular to the normal of the pier. (a) Pile characteristic value β [m -1 ]. The characteristic value β is defined, for example, by the following formula (1) (quoted from "Japan Port and Harbour Association: Technical Standards for Port Facilities and Commentary (Volume 2), pp. 704-705, 2018") or formula (2) (quoted from "Japan Road Association: Specifications for Highway Bridges and Commentary, IV, Substructure Edition, pp. 259-260, 2017"):

number

number

[0132] Regarding (a), even if the total number of beams is the same, the magnitude of the influence of external forces is considered to change depending on the number of rows of beams perpendicular to the normal line of the pier, so it is meaningful to use this as an explanatory variable. Regarding (b), the magnitude of the influence of external forces is considered to change depending on the characteristic value of the pile (e.g., 1 / β), so it is meaningful to use this as an explanatory variable. Regarding (c), in a beam structure, the behavior of a beam located at the end against external forces may differ from that of beams at positions other than the end, so adding information on whether the beam is located at the end or not to the explanatory variable may improve the prediction accuracy. Regarding (d), there are two possible modes for the joint of a beam located at the end, a single joint and a double joint. Since the behavior of a single joint and a double joint against external forces may differ between the single joint and the double joint, when the beam is located at the end, it may improve the prediction accuracy by adding information on whether the beam is a single joint or a double joint to the explanatory variable. Note that a double joint refers to a joint structure in which both ends of the beam are supported by piles, and a single joint refers to a joint structure in which only one end of the beam is supported by a pile.

[0133] Regarding (e), FIG. 22 is a diagram illustrating a dimensionless distance indicating the position of the beam. This diagram shows the distance in the x direction. The distance in this example is a numerical value obtained by dividing the distance from the reference position (upper left end) in the standardized beam structure by the maximum value (13 in this example), multiplying it by 10, and rounding up the decimal part. By using the dimensionless distance in this way, a unified distance can be used for beam structures of various sizes. Note that the specific calculation method of the dimensionless distance is not limited to this, and it is optional whether to use an integer and how many digits to use (2 digits in the example of the figure).

[0134] Regarding (f), since behavior in response to external forces may differ depending not only on the beam structure but also on the pile projection length or pile length, prediction accuracy may be improved by adding information about the piles supporting the beam structure to the explanatory variables.

[0135] When adding the projection length or length of the pile to the explanatory variables, in addition to using real values, the maximum value ratio may be used, as in "3-4. Beam length" described later. Two-class or three-class classification may also be considered. For example, in the case of two-class classification, if the pile is below the average projection length (or length) of the piles in the pier, the value may be 0, and if it exceeds the average value, the value may be 1.

[0136] 3-4. Beam length The damage prediction device 1 may use the length of the beam as an explanatory variable, but the length used as the explanatory variable may be a real value, or a value obtained by normalizing the real value with a reference value (for example, the maximum length of the beam in the beam structure) as a maximum value ratio. Alternatively, the damage prediction device 1 may treat the length of the beam as a classified class rather than as a numerical value. As an example, in the case of two-class classification, "zero" is used as the length for beams having a length equal to or less than the average length of all beams included in the beam structure, and "1" is used as the length for beams having a length exceeding the average length. In addition, classification may be performed for each of the x direction and y direction of the beam. This applies not only to beams but also to the above-mentioned piles.

[0137] 3-5.Multi-layer structure When the target beam structure is a bridge beam structure and has a deck structure with multiple layers (n layers), it can also be the target of prediction by the damage prediction device 1. An example of a deck structure with multiple layers is a double deck structure of a road bridge and a railway bridge, or a double deck structure of two road bridges (for example, an up line and a down line). A deck structure with multiple layers has multiple superstructures (for example, road surfaces) in the vertical direction. In this example, the distance to other beams is considered in each layer as in the case of a single layer, but the vertical distance is not considered. However, if the correlation between multiple layers is not considered, there is a problem that the higher the layer of the beam structure, the more likely it is to shake during an earthquake, but learning or prediction will be made without distinguishing between them. Therefore, in this example, a deck structure with multiple layers is considered by including information indicating which layer the beam structure is and information indicating the distance between adjacent layers in the deterioration degree information.

[0138] That is, in the case of a beam structure of a layered bridge, the teacher data may include, as explanatory variables, information indicating which layer the beam structure is on and information indicating the distance between adjacent layers, and the input means may input information indicating which layer the beam structure is on in the beam structure and information indicating the distance between adjacent layers. Also, the length of the entire pier, the length of the pier between each layer, or if the foundation structure of the target bridge is a pile foundation, information such as the pile projection length or pile length may be included in the explanatory variable (F).

[0139] FIG. 23 is a diagram illustrating an example of identification information given to each beam in a multi-layer beam structure. In the figure, each layer is labeled as "kth layer," and this information is also used as an explanatory variable. Note that it is preferable that the machine learning model for the multi-layer structure is trained using training data (for a multi-layer bridge) different from that of the machine learning model for the single-layer structure.

[0140] 3-6. Irregular beam structure FIG. 24 is a diagram illustrating an irregular beam structure. The beam structure to be examined is not limited to a regular lattice (square) structure, and may be irregular as shown in this figure. As a method for dealing with such an irregular beam structure, it is possible to provide dummy pile heads between long beams, divide the beams, and set them in a regular lattice. However, in the above-mentioned embodiment, it is not necessary to provide dummy pile heads or beams to form a regular lattice, and it is sufficient to read the information as it is in the acquired image data. Note that in the above-mentioned embodiment, it is not necessary to provide dummy pile heads or beams to form a regular lattice, but it is also possible to provide dummy pile heads or beams to form a regular lattice.

[0141] 3-7.Other When the target beam structure has a multi-layer beam structure, the beam structure to be predicted is not limited to a single layer. The learning and prediction processes may be applied to multiple layers.

[0142] The order of processing in the preparation of training data, learning, and prediction is not limited to that exemplified in Fig. 12 and Fig. 19. For example, the explanatory variables may be input in any order.

[0143] The hardware configuration of the damage prediction device 1 is not limited to that exemplified in the embodiment. The damage prediction device 1 may have any hardware configuration as long as the required functions can be implemented. For example, a part of the functions of the damage prediction device 1, for example, the storage means 11 and a part or all of the trained model M1, may be implemented in an external device such as a server device. In this case, a computer device having an input means 21 and an output means 31 cooperates with this external device and functions as the damage prediction device 1 according to the embodiment as a whole. This server device may be a real server or a virtual server.

[0144] Software elements such as programs and trained models may be provided in a state recorded on a computer-readable recording medium such as a CD-ROM (Compact Disc Read only memory), or may be downloaded via a communication network such as the Internet. [Explanation of symbols]

[0145] 1, 1A...damage prediction device, 11...memory means, 21...input means, 31...output means, 41...prediction means, 51...learning device, 90...machine learning model, 91...input layer, 92...intermediate layer, 93...output layer, 95...convolution layer, 96...pooling layer, 97...deconvolution layer, 98...fully connected layer, 101...CPU, 102...memory, 103...storage, 104...input device, 105...display device, M1...trained model, M2...probabilistic model, F...fixed frame, S, S1, S2...image of pier shape, R1, R2...area, P, P1, P2, P3, P4...pixels.

Claims

1. a storage means for storing a trained model generated by machine learning using training data in which image data showing the deterioration degree of each beam in a beam structure of a pier or bridge is arranged according to the position of the beam is used as an explanatory variable, and the damage state of the beam is used as an objective variable; An input means for inputting second image data obtained by changing the size of first image data showing a beam structure of a pier or bridge that is a target for predicting a damage level to a set number of pixels into the trained model; an output means for outputting damage state data corresponding to the damage state of the beam structure that is the damage degree prediction target obtained in response to an input by the input means; A damage prediction device comprising:

2. The trained model is generated by machine learning using image data in which an image showing the deterioration degree of each beam is arranged according to the position of the beam, and the image data is resized to have the set number of pixels as an explanatory variable. The damage prediction device according to claim 1 .

3. The image data in which the image showing the deterioration degree of each beam is arranged according to the position of the beam and the beam shown in the image based on the first image data each correspond to one pixel. The damage prediction device according to claim 2 .

4. The input means changes the first image data by enlarging the first image data so that the second image data has a set number of pixels. The damage prediction device according to any one of claims 1 to 3.

5. When the first image data is enlarged to a determined size, if one pixel corresponds to a plurality of structural elements in the beam structure, the input means changes the second image data to have a set number of pixels by setting the structural elements determined based on the proportion of the one pixel as pixel values. The damage prediction device according to claim 4 .

6. When the first image data is enlarged to a determined size, if one pixel corresponds to a plurality of structural elements in the beam structure, the input means changes the second image data to have a set number of pixels by setting a structural element determined based on the type of each structural element included in the one pixel as a pixel value. The damage prediction device according to claim 4 .

7. When the number of pixels of the beam shown in the image based on the first image data is different from the number of pixels of the beam shown in the image based on the second image data, the output means determines a representative value of a plurality of pixel values ​​constituting the beam shown in the image based on the second image data as a damage state of the beam. The damage prediction device according to any one of claims 1 to 3.

8. The representative value is an average value of a plurality of pixel values ​​constituting a beam shown in an image based on the second image data. The damage prediction device according to claim 7.

9. The representative value is a pixel value that is the most damaged pixel value among a plurality of pixel values ​​that constitute a beam shown in an image based on the second image data. The damage prediction device according to claim 7.

10. The trained model is generated by an algorithm using an autoencoder or convolutional autoencoder that compresses input image data, retains features, and then performs restoration processing. The damage prediction device according to any one of claims 1 to 3.

11. The trained model is generated by adding the external force conditions acting on the pier or bridge including the beam, data on the beam structure of the beam, and numerical data of the beam in the beam structure as explanatory variables, The input means inputs external force conditions acting on a pier or bridge including a beam structure that is the target of damage prediction, and the output means outputs damage state data for the external force conditions. The damage prediction device according to any one of claims 1 to 3.

12. The numerical data of the beam includes at least one of the dimensions of the beam, the presence or absence of piles at both ends of the beam, the length or protruding length of the piles supporting the beam, and characteristic values ​​of the piles. The damage prediction device according to claim 11.

13. The data on the beam structure includes the number of rows of beams constituting the pier or bridge. The damage prediction device according to claim 11.

14. The external force conditions include earthquakes, dynamic loads, and static loads. The damage prediction device according to claim 11.

15. The damage state data is data representing at least one of the degree of damage to the beam, the area of ​​the damage, or the area ratio of the damage. The damage prediction device according to any one of claims 1 to 3.

16. The output means outputs, as the damage state data, a damage degree distribution based on the damage classification in the beam structure, by mapping it on a diagram showing the beam structure. The damage prediction device according to any one of claims 1 to 3.

17. The beam structure is an n-layer bridge beam structure, The teaching data includes, as the explanatory variables, information indicating which layer the beam structure is in and information indicating a distance between the beam structure and other adjacent layers, The input means inputs information indicating which layer the beam structure is in the beam structure and information indicating a distance between the beam structure and other adjacent layers. The damage prediction device according to any one of claims 1 to 3.

18. A damage prediction method implemented by using the damage prediction device according to any one of claims 1 to 3.

19. A method for generating a trained model by machine learning using teacher data in which image data showing the deterioration degree of each beam in a beam structure of a pier or bridge is arranged according to the position of the beam is used as an explanatory variable, and the damage state of the beam structure is used as a target variable, A generation method in which image data in which an image representing the deterioration degree of each beam is arranged according to the position of the beam, and image data resized to have a set number of pixels is further used as an explanatory variable.

20. The trained model is generated by compressing image data in which an image showing the deterioration degree of each beam is arranged according to the position of the beam, and performing restoration processing after retaining the feature amount by an algorithm using an autoencoder or a convolutional autoencoder.

20. The method of claim 18.

21. A trained model generated by the generation method according to claim 19 or 20.

Images