JP2020085631A - Internal state estimation system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To determine the internal state of a three-dimensional structure which is capable of observing the appearance and inspecting the outer surface but difficult in observing and inspecting the inside without using image recognition using radiation, ultrasonic waves, electromagnetic waves, etc. Also, make it possible to estimate accurately without relying on the experience of a skilled inspector.
SOLUTION: Vibration adding means 1 for applying a hit to a hit point h, vibration measuring means 2a to 2d, structure model data output means 3 for creating structure models M1 to Mn, hit data and structure data K1 to Kn. Three-dimensional model analysis means 4 for receiving and performing vibration analysis to generate a large amount of learning data L1 to Ln, learning data L1 to Ln are given in advance and optimized, and amplitude data Sa to Sd and time data Ta to Td. An internal state estimation system including a neural network 5 for estimating the internal structure of a three-dimensional structure and an internal structure display means 6 by receiving an input of
[Selection diagram]

Description

INDUSTRIAL APPLICABILITY The present invention is capable of accurately estimating changes in the internal structure of a structure due to long-term use or deterioration over time, alteration of the materials constituting the structure, and differences between the structure at the time of design and the structure after construction. The present invention relates to an internal state estimation system for doing.

Conventionally, as a method for inspecting the internal state of a structure without destroying the structure, a method using radiation, ultrasonic waves, electromagnetic waves, etc. is known, and it is also possible to estimate the degree of deterioration by a tapping sound inspection. ing.
However, if the structure is large, it is difficult to see through with radiation, and if the structure has a complicated shape or if the materials that make up the structure are altered, inspection using ultrasonic waves or electromagnetic waves is not accurate. There is a problem that is dropped.
In addition, guessing by hammering sound requires an inspector with experience, so it is labor-intensive, costly, and takes time to train successors, and it is difficult to specify the degree of damage and the position of damage. There are also problems.

In Patent Document 1 (Japanese Unexamined Patent Application Publication No. 2018-41178), machine determination by a neural network is studied in image recognition of an underground radar using electromagnetic waves in the hundreds of MHz band used for detecting an abnormal place in the ground. (See especially paragraphs 0002 to 0003), and as a method for generating a large amount of learning data required for that at high speed, an analysis model is used to target a probe wave of a specified frequency. It has been described that the incidence of light and the reception of reflected waves from the target system are simulated, and that the simulation is executed by a high-speed arithmetic processing element (see particularly paragraph 0013).
However, this learning data generation method is applied to a method of detecting an abnormal place in the ground by image recognition of the ground radar.

In Patent Document 2 (Japanese Patent No. 3705357), a vibration unit 3 that vibrates the structure 2, a measurement unit 4 that measures a response due to the given vibration, and a vibration mode based on a result measured by the measurement unit 4 The soundness diagnostic device 1 includes an experimental mode analysis unit 5 to be obtained, and a deformation estimation unit 6 that estimates a state (deformation) of a change such as a damaged portion from a result of a vibration mode analysis. The theoretical value calculation unit 61 that analyzes the modeled structure 2 to calculate the theoretical value of the natural frequency of the vibration mode corresponding to each model and the theoretical value and the natural frequency calculated by the experimental mode analysis unit 5 The point of estimating the damage location from the most adaptable model having an evaluation unit 62 that evaluates the actual measurement value based on the adaptation condition is described (see particularly paragraphs 0025 and 0030).
Further, in Patent Document 3 (Japanese Patent No. 4069977), in the deformation estimation unit 6 of the soundness diagnostic device 1 of Patent Document 2, the vibration parameters of the structures in the vibration modes from the low order to the high order are set as follows. The present invention provides a soundness diagnostic system capable of detecting the influence of a minute deformation of a structure by evaluating the vibration parameter in comparison with the sound vibration parameter.
However, in the inventions described in Patent Documents 2 and 3, it is not easy to accurately estimate the theoretical deformation value and the actual measurement value to accurately estimate the deformed portion of the structure. It is necessary to improve it.

JP, 2018-41178, A Japanese Patent No. 3705357 Japanese Patent No. 4069977

The problem to be solved by the present invention is to change the internal structure of a three-dimensional structure that is capable of observing the appearance and inspecting the outer surface but is difficult to observe and inspect the interior, and the material forming the three-dimensional structure. In order to understand the deterioration of the three-dimensional structure, the internal state (thickness, material strength, etc.) of such a three-dimensional structure is not recognized by image recognition using radiation, ultrasonic waves, electromagnetic waves (including underground radar), etc. Another object of the present invention is to provide an internal state estimation system that enables accurate estimation without relying on the experience of a trained inspector.

The invention according to claim 1 for solving the above-mentioned problems is an internal state estimation system for estimating an internal state of a three-dimensional structure,
Vibration adding means for applying a single-shot impact with a constant intensity from a predetermined direction to an impact point provided on the outer surface of the three-dimensional structure,
At a plurality of vibration measurement points provided on the outer surface of the three-dimensional structure, the amount of amplitude after the impact is applied by the vibration applying means and the transmission time from the moment when the impact is applied until the vibration reaches Then, vibration measuring means for outputting each amplitude data and time data or each natural frequency,
Structure model data output means for creating a large number of structure models each having a different internal state for the three-dimensional structure and outputting structure data relating to the large number of structure models,
Receiving the strength data of the impact added to the impact point and the structural data output from the structure model data output means, the numerical analysis means is used for each structural model to correspond to the plurality of vibration measurement points. 3D that performs vibration analysis at a point and outputs, as learning data, specific information that identifies the structure model that was the target of the vibration analysis and multiple vibration analysis data obtained by the vibration analysis that was performed on the structure model Model analysis means,
After the impact is applied to the impact point by the vibration adding means, the amplitude data and the time data or the natural frequency output from the vibration measuring means are input and inverse analysis is performed to output the internal state estimation information. Neural network,
In accordance with the internal state estimation information output from the neural network, an internal state display means for displaying the internal state of the three-dimensional structure,
The neural network is characterized in that learning data about the many structural models output from the three-dimensional model analyzing unit is given in advance and is optimized by self-learning.

According to a second aspect of the present invention for solving the above-mentioned problems, in the internal state estimating system according to the first aspect of the present invention, the vibration measuring unit frequency-converts the vibration waveform based on the measured amplitude amount and transmission time. Then, at least the power spectrum density of the acceleration, the amplitude amount and the phase difference about the frequency component having the maximum are output as amplitude data and time data, respectively, and the three-dimensional model analysis means is replaced with a plurality of vibration analysis data, and a plurality of The frequency analysis is performed on the vibration analysis data, and at least the amplitude amount and the phase difference with respect to the frequency component having the maximum power spectrum density of the acceleration are output as learning data of the amplitude data and the time data.

According to a third aspect of the invention for solving the above-mentioned problems, in the internal state estimating system according to the first aspect of the invention, the three-dimensional structure is a solid body, and the internal state is a thickness of the solid body. Alternatively, it is characterized by the material strength inside the solid body.

The invention according to claim 4 for solving the above-mentioned problems is the internal state estimating system according to claim 1, wherein the three-dimensional structure is a hollow body, and the internal state is a wall of the hollow body. It is characterized by being thick.

In the internal state estimating system according to the first aspect of the present invention, the amplitude data and the time data or the natural frequency output from the plurality of vibration measurement points after the impact is applied to the impact point by the vibration adding means is transmitted to the neural network. Inverse analysis is performed by inputting, internal state estimation information is output, and internal state display means for displaying the internal state of the three-dimensional structure is provided according to the output internal state estimation information. The network is preliminarily given learning data for a large number of structure models output from the three-dimensional model analysis means, and is optimized by self-learning. Accurately estimate the internal state of a structure that is difficult to observe or inspect without relying on image recognition using radiation, ultrasonic waves, electromagnetic waves, etc., and without relying on the experience of a trained inspector. be able to.

In the internal state estimation system of the invention according to claim 2, in addition to the above-described effect of the invention according to claim 1, the vibration measuring means frequency-converts the vibration waveform based on the measured amplitude amount and the transmission time, and at least The amplitude amount and the phase difference for the frequency component having the maximum power spectrum density of acceleration are output as amplitude data and time data, respectively, and the three-dimensional model analysis means frequency-converts the plurality of vibration analysis data, and at least the power spectrum of acceleration is obtained. Since the amplitude amount and phase difference for the frequency component having the maximum density are output as the learning data of the amplitude data and the time data, the amount of learning data can be reduced without lowering the estimation accuracy.

The internal state estimation system of the invention according to claims 3 and 4 has the effect of the invention according to claim 1 as well as the thickness of the solid body or the material strength inside the solid body according to claim 2. In No. 3, the wall thickness of the hollow body can be accurately estimated.

1 is a block diagram of an internal state estimation system of Example 1. FIG. The figure which shows the kind of pattern in which the thickness of the outer side of the pipe object which is an analysis has decreased one side. 3 is a flowchart showing an algorithm of the internal state estimation system of Example 1. FIG. FIG. 5 is a diagram showing an iron plate to be analyzed in Example 2. The figure which shows the vibration analysis data in each vibration measurement point immediately after adding the hit|damage H of fixed intensity|strength to the hit|damage point h using the structure model without damage in Example 2. The figure which shows the vibration analysis data in each vibration measurement point immediately after adding the hit|damage H of fixed intensity|strength to the hit|damage point h using the structure model with damage in Example 2. The figure which shows the structure model in which the plate thickness of the shaded part is reducing 15 mm. The figure which shows the difference of the vibration analysis data which used the structure model without damage and the structure model with damage. The graph which shows the relationship between the frequency obtained by Fourier-transforming vibration analysis data, and the power spectrum density of acceleration. The graph which shows the relationship between the frequency and the phase difference obtained by Fourier-transforming the vibration analysis data. 5 is a flowchart showing an algorithm of the internal state estimation system of Example 2. FIG. The table which shows the estimation result of the damage degree and damage position obtained by carrying out the reverse analysis of the verification data.

Hereinafter, embodiments of the present invention will be described with reference to examples.

FIG. 1 is a block diagram of the internal state estimation system according to the first embodiment.
As shown in FIG. 1, the internal state estimation system according to the first embodiment includes a bent structure C installed between a floor surface and a wall surface (specifically, the center line has an arc shape, and This is for estimating the internal structure (thickness of the pipe) of the hollow pipe whose cross section is orthogonal to the circle, and comprises the following means.
The outer surface of the structure C can be visually observed, and the inner surface of the structure C cannot be visually observed, but the material of the pipe body and the shape of the inner wall at the time of installation (shown by a dotted line in FIG. 1). Part) is known.
(1) A vibration applying means 1 for applying an impact to an impact point h provided on the outer surface of the structure C.
(2) The amplitude amount after the impact H is added to the impact point h, which is installed at the vibration measurement point a provided on the outer surface of the structure C, and the transmission time from when the impact is applied until the vibration reaches And a vibration measuring means 2a for measuring amplitude and outputting amplitude data Sa and time data Ta.
(3) The amplitude amount after the impact H is added to the impact point h, which is installed at the vibration measurement point b provided on the outer surface of the structure C, and the transmission time from when the impact is applied until the vibration reaches And a vibration measuring means 2b for measuring amplitude and outputting amplitude data Sb and time data Tb.
(4) The amplitude amount after the impact H is added to the impact point h, which is installed at the vibration measurement point c provided on the outer surface of the structure C, and the transmission time from the moment when the impact is applied until the vibration reaches And a vibration measuring means 2c for measuring amplitude and outputting amplitude data Sc and time data Tc.
(5) The amplitude amount after the impact H is added to the impact point h, which is installed at the displacement measurement point d provided on the outer surface of the structure C, and the transmission time from when the impact is applied until the vibration reaches And a vibration measuring means 2d for measuring amplitude and outputting amplitude data Sd and time data Td.

(6) For the structure C, structure model data output means 3 that creates structure models M1 to Mn having different internal structures and outputs structure data K1 to Kn regarding each structure model.
(7) Receiving the striking data regarding the striking H added to the striking point h and the structure data K1 to Kn output from the structure model data output means 3, the numerical analysis means is used for each of the structure models M1 to Mn. Vibration analysis is performed at points corresponding to the vibration measurement points a to d, and for the structure body model M1, the vibration analysis data Aa1 obtained by the vibration analysis at the points corresponding to the structural data K1 and the vibration measurement points a to d. Ad1 is output as the learning data L1, and similarly, for the structure model Mn, similarly, vibration analysis data Aan to Adn (amplitude data and amplitude data obtained from the vibration analysis at the points corresponding to the structural data Kn and the vibration measurement points a to d). Three-dimensional model analysis means 4 for outputting (analysis data corresponding to time data) as learning data Ln.
Since the outer shape of the structure models M1 to Mn is the same as that of the structure C, the positions of the points corresponding to the vibration measurement points a to d in the structure models M1 to Mn are the vibration measurement points a to d. And all match.
(8) After the impact H having a constant intensity is added to the impact point h, an inverse analysis is performed by receiving the amplitude data Sa to Sd and the time data Ta to Td output from the vibration measuring means 2a to 2d. A neural network 5 that outputs internal structure estimation information.
(9) Internal structure display means 6 for displaying the internal structure of the structure C in accordance with the internal structure estimation information output from the neural network 5.

Hereinafter, each means will be described in detail.
(Vibration adding means 1)
It is a means that can apply a hit in any direction from the hit point h, and is usually installed on a wall or a column near the structure C.
The hitting point h can be set at any position of the structure C, but in order to improve the estimation accuracy, it is better to set it at the center of the structure C or near the center of gravity. Therefore, the striking point h in FIG. 1 is set at the outer center portion of the cross section passing through the central axis of the structure C, and the striking direction is toward the center of the central axis.

(Vibration measuring means 2a to 2d)
The vibration measuring means 2a to 2d can be installed at arbitrary four points (a to d) of the structure C, and the amplitude which is the amplitude amount measured at each point after the hit H is added to the hit point h. It outputs data Sa to Sd and time data Ta to Td, which is the time (phase difference) from the time when the impact is applied until the vibration reaches.
In addition, in FIG. 1, the vibration measuring means 2b is installed at the same position as the striking point h (a position very close to it), and the vibration measuring means 2a and 2c are located at the striking point h outside the cross section passing through the central axis of the structure C. It is installed on the upper side and the lower side, and the vibration measuring means 2d is installed on the inner center portion in the same cross section.

(Structure model data output means 3)
The structure model data output means 3 uses, as the structure models M1 to Mn having different internal structures, the shape of the inner wall at the time of installation (the portion shown by the dotted line in FIG. 1, which can usually be identified from the design drawing). As the above, various patterns are created in consideration of abrasion due to the fluid flowing inside the structure C and increase in thickness due to adhered substances.
For example, a total of 17600 patterns of the following (a) to (f) can be mentioned.
(A) 200 patterns in which the thickness of the tubular body is uniformly reduced from the reference thickness, and the thickness reduction amount is increased by 0.1 mm.
(B) In the pattern in which the decrease in the inner thickness of the tubular body is (1-x) times the decrease in the outer thickness, the decrease in the outer thickness is increased by 0.1 mm. 200 patterns (however, x is 0.1, 0.2,..., 1 in 10 ways, so 2000 patterns in total).
(C) The thickness of the outer side of the tubular body is reduced, and the other part has the same pattern as the above (a) and (b) (the outer center part is reduced as shown in FIG. 2(1)). 2), a pattern in which the outer upper part is depleted as shown in FIG. 2(2), and a pattern in which the outer lower part is depleted as shown in FIG. 2(3)). 200 patterns + 2000 patterns.
(D) 200 patterns in which the thickness of the tubular body is uniformly increased from the reference thickness, and the increment of the thickness is increased by 0.1 mm.
(E) A pattern in which the increase in the inner thickness of the tubular body is (1+x) times the increase in the outer thickness, and 200 patterns in which the increase in the outer thickness is increased by 0.1 mm (However, since there are 10 patterns of x, 0.1, 0.2,..., 1, there are 2000 patterns in total).
(F) The thickness of the outer side of the tubular body is unevenly increased, and the other portions have the same pattern as the above (d) and (e), and 200 patterns+2000 patterns for each unevenly increased portion.
In the first embodiment, the pipe body is completely fixed to the floor surface and the wall surface, the outer diameter of the pipe is 14 cm, the inner diameter is 8 cm, and the structural data K1 to Kn regarding each structural body model is output.

(Three-dimensional model analysis means 4)
In the three-dimensional model analysis means 4 of the first embodiment, an analysis means by the finite element method (hereinafter referred to as "FEM analysis means") is used as the numerical analysis means.
According to the FEM analysis means, when a three-dimensional model is cut into a large number of tetrahedral meshes and the conditions are defined and then analyzed, when impact or vibration is applied to a predetermined position, the three-dimensional model within the three-dimensional model is analyzed. It is possible to calculate how any point (vertex of any tetrahedron) fluctuates (oscillates).
Therefore, when the structure model data output unit 3 creates the structure models M1 to Mn having different internal structures and outputs the structure data K1 to Kn related to each structure model, the hitting point for each structure model. Since the vibration analysis data of the points corresponding to the measurement points a to d when the impact H of constant strength is added to h is obtained, a large amount of learning data L1 to Ln can be prepared in a short time.

(Neural network 5)
As shown in FIG. 1, the neural network 5 includes an input layer 5I, an output layer 5O, learning data accumulating means 5L, learning data inputting means 5T for inputting a large amount of learning data L1 to Ln to the learning data accumulating means 5L, and not shown. Built with multiple middle layers, etc.
Further, the intermediate layer is configured by combining various devised layers such as a convolutional layer, a pooling layer, and a fully-bonded layer, if necessary.
Then, the neural network 5 is optimized by accumulating a large amount of learning data L1 to Ln in the learning data accumulating means 5L and allowing the learning data to self-learn. Then, an estimation accuracy of 95% or more can be obtained by using parameters optimized by a sufficient number (10,000 to several hundreds of thousands depending on the complexity of the target shape) of learning data.
After that, when the amplitude data Sa to Sd and the time data Ta to Td output from the vibration measuring means 2a to 2d are input to the input layer 5I, the neural network 5 performs an inverse analysis, and the internal structure in the actual pipe. The internal structure estimation information close to is output with high accuracy.

(Internal structure display means 6)
The internal structure display means 6 displays the internal structure of the structure C in accordance with the internal structure estimation information output from the neural network 5. In a normal display mode, a cross section passing through the central axis of the structure C is displayed. Display status.
It is also possible to select a mode in which the state of the cross section perpendicular to the central axis of the structure C is displayed at a plurality of locations.

FIG. 3 is a flow chart showing an algorithm of the internal state estimation system of the first embodiment. A large number of learning data L1 to Ln are generated by the following procedures (T1) to (T4) and stored in the neural network 5. After self-learning, the internal structure of the structure C is estimated in steps (1) to (6), and the estimated internal structure is displayed.
(T1) A large number of structure models (M1 to Mn) are created based on the current state of the structure C, design drawings, and the like.
(T2) Structure data K1 to Kn of each structure model is created.
(T3) Vibration analysis is performed by the FEM analysis means based on the impact data regarding the impact H added to the impact point h and the respective structural data K1 to Kn, and the vibration analysis data (amplitude) of the points corresponding to the vibration measurement points a to d. Output analysis data corresponding to data and time data).
(T4) A large number of learning data L1 to Ln are generated and stored in the neural network 5.
(1) The striking point h of the structure C is applied with a striking H having a constant strength from a predetermined direction.
(2) The amplitude data Sa to Sd and the time data Ta to Td after the impact H is added at the vibration measurement points a to d are measured.
(3) Send the amplitude data Sa to Sd and the time data Ta to Td to the neural network 5.
(4) The neural network 5 performs an inverse analysis based on the input amplitude data Sa to Sd and the time data Ta to Td.
(5) The internal structure estimation information is output by the inverse analysis and sent to the internal structure display means 6.
(6) The internal structure display means 6 displays the internal structure based on the internal structure estimation information.

FIG. 4 is a diagram showing an iron plate to be analyzed by the internal state estimation system of the second embodiment.
This iron plate has a width of 1000 mm, a height of 600 mm, and a thickness of 30 mm, the lower end of FIG. 4 is completely restrained, and one side is in contact with the standing wall and cannot be seen.
The hitting point h is set at the upper left corner, and a single hitting H is applied from above with a constant strength (100N).
Further, the vibration measurement points were 11 points as indicated by circles in FIG.
In addition, in Example 2, since the analysis target is an iron plate having a thickness of 30 mm, the strength of the impact H was set to 100 N, but the strength is appropriately changed depending on the size and material of the analysis target.

In creating a large number of structural models, the damage position and the degree of damage (amount of decrease in plate thickness) on one side of the iron plate were changed. It is assumed that there are four ways of reducing the plate thickness: 10 mm, 15 mm, 20 mm and 25 mm.
In order to identify the damage position, a mesh of 10 mm in length and width is created, and the size of the damage is 10 mm×30 mm, 10 mm×50 mm, 10 mm×70 mm, 10 mm×90 mm, 30 mm×10 mm, 30 mm×30 mm, 30 mm×50 mm, 30 mm x 70 mm, 30 mm x 90 mm, 50 mm x 10 mm, 50 mm x 30 mm, 50 mm x 50 mm, 50 mm x 70 mm, 50 mm x 90 mm, 70 mm x 10 mm, 70 mm x 30 mm, 70 mm x 50 mm, 70 mm x 70 mm, 70 mm x 90 mm, 90 mm x There are 24 types of 10 mm, 90 mm×30 mm, 90 mm×50 mm, 90 mm×70 mm, and 90 mm×90 mm.
All the structural models are obtained by shifting the positions of these 24 kinds of damages, but the position of the damage of 10 mm in length is 60×(98+96+94+92)=22800, and the position of damage of 30 mm in length is 58. ×(100+98+96+94+92)=27840 ways, 50mm length damage can take 56×(100+98+96+94+92)=26880 ways, 70mm length damage can get 54×(100+98+96+94+92)=25920 ways, 90mm length damage The number of positions to be obtained is 52×(100+98+96+94+92)=24960, and four reductions are assumed for each position. Since there is one model without damage, the number of all structure models is 513601.
Then, as in the first embodiment, the structural data K1 to Kn of each structural body model is created, and the impact point h is added with the impact H having a constant strength. The vibration analysis data was obtained.

FIGS. 5 and 6 are diagrams showing vibration analysis data at each vibration measurement point immediately after the impact H of constant strength is applied to the impact point h using such a structure model, and FIG. FIG. 6 shows vibration analysis data for a non-existent structure model, and FIG. 6 shows vibration analysis data for a structure model in which the mesh thickness of the meshed portion is reduced by 15 mm as shown in FIG. ..
8A and 8B are enlarged views of a portion surrounded by a thick line in FIGS. 5 and 6, and a horizontally elongated view is a structure model without damage and damage. A diagram showing vibration analysis data at each vibration measurement point after a certain amount of impact H is applied to the impact point h using the structure model, and the diagram shown in the upper right is surrounded by a thick line in a horizontally long diagram. FIG.
As can be seen by comparing the diagrams shown in the upper left and upper right of FIGS. 8(a) and 8(b), the vibration analysis data of the damaged structure model has a different waveform from the vibration analysis data of the undamaged structure model. Therefore, it is possible to estimate the damage position by the difference, but since the amount of data becomes too large when raw waveform data is used, it was decided to estimate using the value after Fourier transform.
9 is a graph showing the relationship between the frequency obtained by Fourier transforming certain vibration analysis data and the power spectrum density of acceleration, and FIG. 10 is the frequency obtained by Fourier transforming the same vibration analysis data. It is a graph which shows the relationship between and a phase.
As can be seen from these graphs, since the power spectrum density of acceleration has a very large difference depending on the frequency, only the nine frequency components with the largest values are the vibration measurement points at the same positions as the amplitude amount and the impact point. (The vibration measurement point (1) in FIG. 10) is used to calculate the phase difference from the vibration and obtain the amplitude data and the time data.

FIG. 11 is a flow chart showing an algorithm of the internal state estimation system of the second embodiment, in which a large number of learning data L1 to Ln are generated by the following procedures (T1′) to (T5′) and stored in the neural network 5. Then, after self-learning, the internal structure of the iron plate is estimated in steps (1′) to (7′), and the estimated internal structure is displayed.
(T1′) A large number of structure models (M1 to Mn) are created based on the current state of the iron plate, design drawings, and the like.
(T2′) Structure data K1 to Kn of each structure model is created.
(T3′) The FEM analysis means performs vibration analysis based on the impact data regarding the impact H added to the impact point h and the respective structural data K1 to Kn, and determines the points corresponding to the vibration measurement points (1) to (11). Vibration analysis data (analysis data corresponding to amplitude data and time data) is output.
(T4') The vibration analysis data for each structure model is frequency-converted by the fast Fourier transform means (FFT), and the amplitude amount and phase difference are output only for the nine frequency components from the one having the largest power spectrum density of acceleration. To do.
(T5′) A large number of learning data L1 to Ln (amplitude data and time data, reduction amount of plate thickness, damage size) based on the output amplitude amount and phase difference and internal structure data of each structure model. And a damage position), and accumulates them in the neural network 5.
(1′) A hit H having a constant strength is applied to the hit point h of the structure C from a predetermined direction.
(2') The vibration after the hit H is applied is measured at the vibration measurement points (1) to (11) ((1) and (11) indicate the numbers in circles). The amplitude data Sa to Sd and the time data Ta to Td are measured.
(3') The measured vibration data is frequency-converted by a fast Fourier transform means (FFT), and amplitude data and time data are output for each frequency having a large power spectrum density of acceleration.
(4') The output amplitude data S1 to S11 and time data T1 to T11 are sent to the neural network 5.
(5') The neural network 5 performs inverse analysis based on the input amplitude data S1 to S11 and time data T1 to T11.
(6') The internal structure estimation information is output by the inverse analysis and sent to the internal structure display means 6.
(7') The internal structure display means 6 displays the internal structure based on the internal structure estimation information.
Of the vibration analysis data (amplitude amount and phase difference after frequency conversion) about 11 vibration measurement points output in the procedure (T4′), 90% of the vibration analysis data is used for learning, and the remaining 1 The vibration analysis data was used for verification.

FIG. 12 is a table showing damage degree and damage position estimation results obtained by inputting 20 pieces of vibration analysis data for verification (20 sets of amplitude data and time data) into a neural network and performing inverse analysis. is there.
The numerical values described in the analysis value and the predicted value are numerical values given in advance to the combinations of the reduction amount of plate thickness, the size of damage and the damage position.
Comparing the analysis value and the estimated value, 19 out of 20 are correct and 1 is incorrect, but 1 of the incorrect answers has a different amount of damage but a different amount of reduction in plate thickness. It was confirmed that the damage positions were the same and that the extent and position of the damage could be estimated.

Modification examples of the embodiment will be listed.
(1) The internal state estimating system according to the first embodiment is for estimating the internal structure of the hollow pipe (thinning or thickening of the pipe), but not limited to thinning or thickening of the pipe. It may be used to estimate the material strength of. For this reason, in this specification, there are some places where “internal state” or “internal state...” is described instead of “internal structure” or “internal structure... ”.
For example, in the case where the inside of the pipe is deteriorated, a large number of structure models in which the strength of the region at a predetermined depth from the inner surface of the pipe body is small may be created and a large amount of learning data may be generated.
(2) The internal state estimation system of the first embodiment is for estimating the internal structure of the hollow pipe (thinning or thickening of the pipe), but it is not limited to the hollow pipe, and a hollow portion is provided inside. It is also possible to estimate the wall thickness of the hollow body by applying it to a three-dimensional structure (hollow body) having.

(3) The internal state estimation system of the second embodiment was intended to estimate the reduction (damage) of the thickness of the iron plate, but it is not limited to the reduction or damage of the thickness of the iron plate, and is composed of various materials. It is also possible to estimate the reduction in thickness, damage, and material strength of the solid body by applying it to the solid three-dimensional structure (solid body).
In such a case, for example, in the case where the reinforcing bars are included inside the solid body, the structure model in which the strength of some or all of the reinforcing bars is low, the structure model in which some of the reinforcing bars are not installed, and A large number of learning data may be generated by creating a structure model that combines them. In addition, in the case where the material around the reinforcing bar deteriorates, a large number of structure models in which the material strength around the reinforcing bar is small in addition to the strength of the reinforcing bar are created, and a large number of learning data may be generated.
(4) In the internal state estimation system of the first and second embodiments, the FEM analysis means is used for generating the learning data and the verification data, but the numerical analysis means based on the difference method, the boundary element method, etc. is not limited to the FEM analysis means. May be used.

(5) In the internal state estimation system of the second embodiment, with respect to the data after frequency conversion, only the nine frequency components from the one having the largest power spectrum density of acceleration, the vibration measurement point at the same position as the amplitude amount and the impact point. Although the phase difference (arrival delay time) with the vibration obtained in step 1 was calculated, the number of frequency components for calculating the phase difference is not limited to nine, and may be six or three. The amplitude amount and the phase difference may be calculated for the maximum frequency component and used as learning data.
(6) In the internal state estimation system of the second embodiment, a large number of structural models in which 24 kinds of damages exist in one place of the iron plate are created, but many damages exist in two or more places of the iron plate. The structure model may be created to obtain the vibration analysis data for each structure model.
However, if there are multiple damages, a huge number of combinations of structure models will be created, which will impose a large load on the generation of learning data. You may get it.
For example, even if it is assumed that the iron plate having a width of 1000 mm, a height of 600 mm and a thickness of 30 mm used in Example 2 has damages of 24 sizes and four kinds of depths at two places, the width is 1000 mm and the height is 600 mm. If you create a structure model by dividing a 200 mm square area into 15 areas and creating 24 types of damages in the center of each area, you can reduce the number of all structure models. The data generation load can be reduced.
In the case of this example, the number of all structure models is 24×4×24×4×15×14/2=967680, which is less than double that of 513601 in the second embodiment.

(7) In the internal state estimation system of Examples 1 and 2, at a plurality of vibration measurement points, the amount of amplitude after the hit H is added to the hit point h and the vibration from the time when the hit is applied until the vibration reaches. The transmission time was measured, and the internal structures of the structure C and the iron plate were estimated based on the amplitude data and the time data (transmission time or phase difference), but instead of the amplitude data and the time data, at a plurality of vibration measurement points. The natural frequencies of the structure C and the iron plate may be calculated from the measured amplitude amount and the transmission time, and the internal structures of the structure C and the iron plate may be estimated based on the natural frequencies.
The natural frequency is calculated by frequency-converting the vibration data obtained at each vibration measurement point and averaging the frequency intervals at which the power spectrum density of a large acceleration appears.
For example, in the case of FIG. 9, the natural frequency is around 910 Hz.
In addition, as a matter of course, when estimating the internal structure by the natural frequency, the natural frequency of each structure model is obtained by using the numerical analysis means such as the FEM analysis means, a large number of learning data are generated, and the neural network is preliminarily prepared. It is necessary to store the data in the network 5 and optimize it.

DESCRIPTION OF SYMBOLS 1 vibration addition means 2a-2d vibration measurement means 3 structure model data output means 4 three-dimensional model analysis means 5 neural network 5I input layer 5L learning data storage means 5O output layer 5T learning data input means 6 internal structure display means a-d Vibration measurement points Aa1 to Ad1, Aan to Adn Vibration analysis data C Structure h Hitting point H Hitting K1 to Kn Structural data L1 to Ln Learning data M1 to Mn Structure model Sa to Sd, S1 to S11 Amplitude data Ta to Td, T1 to T11 time data

Claims (4)

An internal state estimation system for estimating the internal state of a three-dimensional structure,
Vibration adding means for applying a single-shot impact with a constant intensity from a predetermined direction to an impact point provided on the outer surface of the three-dimensional structure,
At a plurality of vibration measurement points provided on the outer surface of the three-dimensional structure, the amount of amplitude after the impact is applied by the vibration applying means and the transmission time from the moment when the impact is applied until the vibration reaches Then, vibration measuring means for outputting each amplitude data and time data or each natural frequency,
Structure model data output means for creating a large number of structure models each having a different internal state for the three-dimensional structure and outputting structure data relating to the large number of structure models,
At the point corresponding to the plurality of vibration measurement points using the numerical analysis means for each structure model, receiving the impact data added to the impact point and the structure data output from the structure model data output means. A three-dimensional model analysis that performs vibration analysis and outputs, as learning data, specific information that identifies the structure model that was the target of the vibration analysis and multiple vibration analysis data obtained by the vibration analysis performed on the structure model. Means and
After the impact is applied to the impact point by the vibration adding means, the amplitude data and the time data or the natural frequency output from the vibration measuring means are input and inverse analysis is performed to output the internal state estimation information. Neural network,
In accordance with the internal state estimation information output from the neural network, an internal state display means for displaying the internal state of the three-dimensional structure,
The internal state estimation system is characterized in that the neural network is preliminarily provided with learning data for the large number of structure models output from the three-dimensional model analyzing means, and is optimized by self-learning. The vibration measuring means frequency-converts the vibration waveform based on the measured amplitude amount and transmission time, and at least the amplitude amount and the phase difference for the frequency component with the maximum power spectrum density of acceleration as amplitude data and time data, respectively. Output,
The three-dimensional model analysis means, instead of the plurality of vibration analysis data, frequency-converts the plurality of vibration analysis data, at least the amplitude amount and phase difference about the frequency component having the maximum power spectrum density of acceleration, the amplitude data and The internal state estimation system according to claim 1, wherein the internal state estimation system outputs the learning data as time data. The internal state estimating system according to claim 1 or 2, wherein the three-dimensional structure is a solid body, and the internal state is a thickness of the solid body or a material strength inside the solid body. The internal state estimation system according to claim 1 or 2, wherein the three-dimensional structure is a hollow body, and the internal state is a wall thickness of the hollow body.