Methods for training machine learning models, methods for estimating the state of structures, training devices, estimation devices, and programs

Abstract

This technology provides the ability to estimate the condition of a structure under inspection, even when measurement data for a healthy learning structure is unavailable. [Solution] A method for training a machine learning model includes: a selection step of identifying a reference frequency having predetermined characteristics in a learning spectral response function obtained by applying elastic waves having a wide frequency band to a learning structure; a normalization step of normalizing the learning spectral response function in the frequency axis direction using the reference frequency; and a training step of training a machine learning model that uses the learning spectral response function obtained by measuring the vibration of a structure under inspection as training data, and uses data used to estimate the state of the structure under inspection as output data.

Description

[Technical Field]

[0001] This invention relates to a technique for estimating the state of a structure. [Background technology]

[0002] Techniques for estimating the condition of structures have been proposed. For example, as one of the non-destructive testing methods for evaluating the soundness of structures containing concrete, the impact elastic wave method (see, for example, Non-Patent Document 1) and the local vibration test method that applies the impact elastic wave method (see, for example, Non-Patent Document 2) are known.

[0003] The impact elastic wave method estimates the thickness of a structural member, the presence of internal defects, the presence of back voids, and the distance to internal defects by analyzing inspection data obtained from applying elastic waves to the structure.

[0004] Furthermore, the local vibration test method uses an electrodynamic vibrator to locally apply elastic waves with a wide frequency band to a structure and measure the response acceleration of the structure. Then, the spectral response function is obtained from the obtained response acceleration using a fast Fourier transform, etc., and the state of the part under inspection can be estimated by focusing on the peak frequency (=resonance frequency) of the reference spectral response function obtained from a healthy part of the structure. In this case, the resonance frequency, which is a modal parameter of the structure, can be used to estimate the state of the part under inspection according to the difference between the resonance frequency in the spectral response function obtained from the part under inspection and the resonance frequency in the reference spectral response function. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Concrete Diagnostic Technology '15, p. 139. [Non-Patent Document 2] Journal of Japan Society of Civil Engineers, Series E2 (Materials and Concrete Structures), Vol. 67, No. 4, 522-534, 2011. [Overview of the project] [Problems that the invention aims to solve]

[0006] Incidentally, when attempting to perform non-destructive testing on a particular structure, it is extremely rare for a structure with the same structure as the one being tested to have a known spectral response function in a sound state. Therefore, as the resonant frequency in the spectral response function for reference, either (1) the resonant frequency obtained from a part of the structure being tested that is considered sound is used, or (2) the resonant frequency estimated from a rod-shaped test specimen placed with both ends free, which was poured with concrete under the same conditions as the structure being tested, is used.

[0007] However, estimating the state by focusing on these resonant frequencies has the drawback of not fully utilizing important information (other frequency ranges) in the spectral response function, as it does not take into account vibration characteristics other than the resonant frequency in the spectral response function.

[0008] One aspect of the present invention has been made in view of the above-mentioned problems, and its object is to provide a technology that can estimate the state of a structure even when sound data of a sound structure is not available in advance. [Means for solving the problem]

[0009] To solve the above problems, a method for training a machine learning model according to one aspect of the present invention includes: a identification step of identifying a reference frequency having predetermined characteristics in a learning spectral response function obtained by applying elastic waves having a wide frequency band to a learning structure; a normalization step of normalizing the learning spectral response function in the frequency axis direction using the reference frequency; and a training step of training a machine learning model that uses the learning spectral response function obtained in the normalization step as training data, takes an inspected spectral response function obtained by measuring the vibration of an inspected structure as input data, and outputs data used to estimate the state of the inspected structure as output data.

[0010] Furthermore, a method for estimating the state of a structure according to one aspect of the present invention includes: a identification step of identifying a reference frequency having predetermined characteristics in a spectral response function of a structure to be inspected obtained by applying elastic waves having a wide frequency band to the structure to be inspected; a normalization step of normalizing the spectral response function of the structure to be inspected in the frequency axis direction using the reference frequency; and an estimation step of estimating the state of the structure to be inspected by inputting the spectral response function of the structure to be inspected obtained in the normalization step into a machine learning model that takes the spectral response function of the structure to be inspected obtained by measuring the vibration of the structure to be inspected as input data and data used to estimate the state of the structure to be inspected as output data.

[0011] Furthermore, a training device according to one aspect of the present invention includes: an identification unit that identifies a reference frequency having predetermined characteristics in a learning spectral response function obtained by applying elastic waves having a wide frequency band to a learning structure; a normalization unit that normalizes the learning spectral response function in the frequency axis direction using the reference frequency; and a training unit that trains a machine learning model using the learning spectral response function normalized by the normalization unit as training data, taking the spectral response function of a target structure obtained by measuring the vibration of the target structure as input data, and output data used to estimate the state of the target structure as output data.

[0012] Furthermore, an estimation device according to one aspect of the present invention includes: an identification unit that identifies a reference frequency having predetermined characteristics in a spectral response function of a structure to be inspected obtained by applying elastic waves having a wide frequency band to the structure to be inspected; a normalization unit that normalizes the spectral response function of the structure to be inspected in the frequency axis direction using the reference frequency; and an estimation unit that estimates the state of the structure to be inspected by inputting the spectral response function of the structure to be inspected, which is obtained by measuring the vibration of the structure to be inspected, as input data and data used to estimate the state of the structure to be inspected as output data, into a machine learning model.

[0013] The training device according to each aspect of the present invention may be implemented by a computer. In this case, a control program for the training device that causes the computer to operate as each part (software element) included in the training device to implement the training device by the computer, and a computer-readable recording medium on which the program is recorded are also within the scope of the present invention.

[0014] Also, the estimation device according to each aspect of the present invention may be implemented by a computer. In this case, a control program for the estimation device that causes the computer to operate as each part (software element) included in the estimation device to implement the estimation device by the computer, and a computer-readable recording medium on which the program is recorded are also within the scope of the present invention.

Advantages of the Invention

[0015] According to one aspect of the present invention, the state of the inspection target structure can be estimated even when there is no measurement data of a sound learning structure.

Brief Description of the Drawings

[0016] [Figure 1] It is a diagram showing the configuration of the state estimation system according to the embodiment. [Figure 2] It is a diagram showing an example of a method for measuring the vibration of a structure according to the embodiment. [Figure 3] It is a diagram showing an example of a spectral response function according to the embodiment. [Figure 4] It is a diagram showing an example of a learning spectral response function of a sound learning structure according to the embodiment. [Figure 5] It is a diagram schematically showing the normalization of the learning spectral response function according to the embodiment. [Figure 6] It is a diagram showing an example of the relationship between the member thickness and attributes of a structure and the resonance frequency according to the embodiment. [Figure 7] It is a flowchart showing an example of the flow of a method for generating a machine learning model and an example of the flow of a state estimation method according to the embodiment. [Figure 8] This is a diagram illustrating the method for calculating the degree of abnormality according to the embodiment. [Figure 9] This figure shows an example of the update process according to the embodiment. [Figure 10] This diagram shows the test flow for Example 1. [Figure 11] This figure shows the test cycle for Example 1. [Figure 12] This figure shows a crack sketch (bottom view) of a damaged slab test specimen according to Example 1. [Figure 13] This figure shows the damage status (bottom surface) at the end of the test according to Example 1. [Figure 14] This table shows the internal damage evaluation cases based on the degree of abnormality in Example 1. [Figure 15] This is a contour map showing the degree of abnormality related to Example 1. [Figure 16] This diagram shows the correlation between Case 1 and Cases 2, 3, and 4 in Example 1. [Figure 17] This figure shows the contour map and cross-section of the abnormality level for Case 1 according to Example 1. [Figure 18] This figure shows the abnormal values ​​at measurement points (rows C and D) near the cutting line in Example 1. [Figure 19] This figure shows an example of the measured spectrum and healthy spectrum related to Example 1. [Figure 20] These are a plan view and a cross-sectional view showing the dimensions of the pier that was the subject of measurement in Example 2. [Figure 21] This figure shows the estimated degree of abnormality for each measurement point in Example 2. [Figure 22] This figure shows the spectral response function of the measurement point exhibiting the minimum anomaly and the spectral response function of the measurement point exhibiting the maximum anomaly according to Example 2. [Modes for carrying out the invention]

[0017] [Embodiment] <Overview of the State Estimation System> An overview of a state estimation system according to one embodiment of the present invention is described below. Conventional local vibration testing methods focus only on the change (difference) in the resonant frequency of modal parameters, and therefore, even if a spectral response function is obtained, information in frequency bands other than the resonant frequency is not fully utilized. In contrast, the state estimation system according to this embodiment estimates the state (degree of abnormality) of a structure by comparing spectral response functions (healthy spectrum vs. spectrum of the structure under inspection). Here, the healthy spectrum is obtained in advance from an arbitrary (suitable) structure, normalized (amplitude normalized, frequency standardized) to give it generality, and then customized for the structure under inspection by transfer learning of the healthy spectrum using a small amount of data of structures that appear healthy.

[0018] Normalizing the spectral response function involves both amplitude axis normalization (0 to 1) and frequency axis normalization. Each axis (amplitude, frequency) offers different advantages. Amplitude axis normalization scales the amplitude spectrum intensity to 0 to 1, eliminating variations due to the contact condition of the concrete surface (roughness), variations by the measurer (how the exciter and accelerometer are pressed), and individual differences in measuring instruments. On the other hand, frequency axis normalization scales concrete structures with different member thicknesses or stiffnesses, such as beams and floor slabs, using a reference frequency (e.g., resonant frequency). This eliminates the need to learn sound data from various concrete structures, thus increasing versatility.

[0019] The "elastic waves with a wide frequency range" input to the vibrator can be white noise or sweep waves. White noise is a noise wave created by randomly combining waves with a wide range of frequencies and the same energy, and can be input for a short time (about 1 second). On the other hand, sweep is an input wave with a relatively long duration (several tens of seconds) created by sequentially combining waves from low frequency to high frequency in the time direction. In this embodiment, as an example, white noise is used for thin members such as floor slabs, and sweep is used for thick members such as beams to ensure that sufficient vibration energy is transmitted (white noise does not transmit vibration energy).

[0020] <Configuration of the state estimation system> One embodiment of the present invention will be described below. Figure 1 is a diagram showing the configuration of the state estimation system 100 according to this embodiment. The state estimation system 100 is a system for estimating the state of a structure. The structure is constructed from, for example, hydrated solidified materials such as concrete and mortar, inorganic solidified materials such as geopolymers, hydrates such as cement, wood, or a combination thereof. More specifically, it is a reinforced concrete (RC) structure or a steel reinforced concrete (SRC) structure. Examples of structures include, but are not limited to, road decks, bridge foundations, tunnels, and port facilities such as piers.

[0021] The condition of a structure can include, for example, the presence or absence of abnormalities or the degree of such abnormalities. Examples of structural abnormalities include deterioration (aging) or damage due to external forces (live loads such as earthquakes and wind). Examples of structural deterioration include wear, fatigue, salt damage, freeze-thaw damage, carbonation, ASR, and fire. More specifically, this can include spalling of the concrete cover due to rebar corrosion, or further deterioration leading to subsidence. By detecting structural abnormalities before deterioration progresses, preventive maintenance measures can be taken.

[0022] As shown in Figure 1, the state estimation system 100 comprises an information processing device 1, a sensor 2, and a vibrator 3. The information processing device 1 is a device that estimates the state of a structure based on the vibration of the structure measured by the sensor 2, and is a general-purpose computer as an example. The information processing device 1 is an example of the training device and estimation device according to this disclosure. The sensor 2 is an acceleration sensor and outputs sensing data representing acceleration. The vibrator 3 is a vibrator that excites vibrations of the structure.

[0023] (Method for measuring vibrations) Figure 2 shows an example of a method for measuring the vibration of a structure using a vibrator 3. In the example in Figure 2, the vibrator 3 is placed on the surface of the structure 4, and the vibrator 3 generates, for example, white noise to excite local vibrations of the structure 4. During excitation, the sensor 2 is brought into contact with the surface of the structure 4 to measure the response acceleration. Vibration measurements are taken at multiple locations on the surface 4a of the structure 4 (for example, at 300 mm intervals, etc.), and data is collected.

[0024] Here, we will explain the difference in vibration measurement results between the case where there is no internal damage to structure 4 and the case where there is, referring to Figure 2. When there is no internal damage (cracks, etc.) to structure 4, most of the elastic waves generated by the exciter 3 are reflected at the bottom surface 4b. On the other hand, when there is internal damage (cracks, etc.) to structure 4, among the elastic waves generated by the exciter 3, high-frequency waves are reflected at the crack surface 4c, while low-frequency waves are diffracted through the crack. Thus, by utilizing the existence of elastic waves that are reflected at the internal crack surface 4c (high-frequency waves) and elastic waves that are diffracted through the crack (low-frequency waves), the state of structure 4 can be evaluated by taking the Fast Fourier Transform of these response accelerations and capturing the change in the resonant frequency, which is the peak frequency, as the difference in spectral response function (a response function that shows the response to vibration for each frequency). This evaluation method is also called the local vibration test method, as mentioned above. However, the local vibration test method assumes that the spectral response function in a sound state is known. Here, "sound condition" refers to a state in which no abnormalities such as cracks have occurred in structure 4.

[0025] If the spectral response function of a healthy structure is unknown, it is possible to assess internal damage based on changes in the resonant frequency (the specific frequency at which an object vibrates naturally). Figure 3 shows an example of the spectral response function of structure 4. In Figure 3, spectral response function w11 is the spectral response function of structure 4 in a healthy state, and spectral response function w12 is the spectral response function of damaged structure 4. In spectral response functions w11 and w12, the horizontal axis represents frequency (Hz), and the vertical axis represents the amplitude spectrum. The resonant frequency of spectral response function w12 of the damaged structure is lower than the resonant frequency of spectral response function w11 in a healthy state.

[0026] When evaluating internal damage based on changes in resonant frequency, the resonant frequency in a healthy state is calculated, for example, by assuming an apparent speed of sound c using the following theoretical formula (1).

number

[0027] In equation (1), E d ρ is the dynamic elastic modulus, ρ is the density, c is the apparent speed of sound, and L is the member thickness. However, in evaluation methods based on changes in the resonant frequency, vibration characteristics other than the resonant frequency are not considered, so important information in the spectral response function cannot be fully utilized. In contrast, the state estimation system 100 according to this embodiment has the features described below, which enable estimation of the state of structure 4 even for structures 4 whose spectral response function in a healthy state is unknown.

[0028] <Configuration of the information processing device> As shown in Figure 1, the information processing device 1 comprises a control unit 10, a storage unit 20, a communication unit 30, an input unit 40, a first output unit 50, and a second output unit 60.

[0029] (Communications Department) The communication unit 30 communicates with external devices of the information processing device 1 via a communication line. The specific configuration of the communication line is not limited to this exemplary embodiment, but examples of communication lines include wireless LAN (Local Area Network), wired LAN, WAN (Wide Area Network), public telephone network, mobile data communication network, or a combination thereof. The communication unit 30 transmits data supplied from the control unit 10 to other devices and supplies data received from other devices to the control unit 10. Note that the information processing device 1 may be configured without a communication unit 30.

[0030] (Input section) The input unit 40 is configured to receive input to the information processing device 1, and may include, for example, an input device such as a keyboard, mouse, touch panel, camera, or microphone. The input unit 40 may also be configured to receive data from the input device via an interface such as USB (Universal Serial Bus). In this embodiment, sensing data from the sensor 2 is input to the input unit 40. However, the sensor 2 may be connected to the communication unit 30 instead of the input unit 40.

[0031] (First output section, second output section) The first output unit 50 and the second output unit 60 are configured to output from the information processing device 1, and may include, for example, output devices such as a display, printer, touch panel, or speaker. The first output unit 50 and the second output unit 60 may also be configured to output data to the output device via an interface such as USB.

[0032] (Storage part) The memory unit 20 stores various types of data that the control unit 10 references. In particular, the memory unit 20 stores instructions for the computer program executed by the control unit 10. Examples of data stored in the memory unit 20 include training data 201, measurement data 203, and the machine learning model 202.

[0033] (Training data / Measurement data) Training data 201 is the data used to train the machine learning model 202. Training data 201 includes, as an example, a learning spectral response function obtained from vibration measurement results of a healthy learning structure. Measurement data 203 represents the vibration measurement results of the learning structure.

[0034] (Machine learning model) The machine learning model 202 is a machine learning model used by the first estimation unit 131 and the second estimation unit 151, described later, to estimate the state of the structure under inspection. Here, when we say that the machine learning model 202 is stored in the storage unit 20, we mean that the parameters defining the machine learning model 202 are stored in the storage unit 20.

[0035] (Specific example of machine learning model 202) Machine learning model 202 is an autoencoder as an example. An autoencoder performs dimensionality reduction (feature extraction) by reducing the number of nodes in the hidden layer, and then reconstructs data close to the original before dimensionality reduction by performing the reverse process. The input layer and output layer have the same number of nodes. In this embodiment, only healthy learning spectral response functions are used as training data, and the autoencoder is trained so that the same learning spectral response function given to the input layer is reconstructed in the output layer. The trained autoencoder created in this way has only learned measurement data of healthy learning structures, so it can reconstruct healthy data but not abnormal data. Taking advantage of this property, separately acquired test data is input to the trained autoencoder, and data with small reconstruction errors between the input and output layers are identified as healthy, while data with large reconstruction errors are identified as abnormal.

[0036] However, the machine learning model 202 is not limited to an autoencoder. For example, the machine learning model 202 may be a machine learning model that outputs classification results generated by supervised learning. In this case, the training data used to train the machine learning model 202 is, for example, a set of training spectral response functions and correct labels (e.g., abnormal / normal) obtained from measurement results of vibrations of a healthy training structure.

[0037] (Input and output of a machine learning model) The input data to machine learning model 202 is the spectral response function of the structure under inspection, obtained by measuring the vibration of the structure under inspection. The output data from machine learning model 202 is used to estimate the state of the structure under inspection. If machine learning model 202 is an autoencoder, its output is the reconstructed spectral response function of the structure under inspection. If machine learning model 202 is a classification model, its output, as an example, is data showing the classification result of the state of the structure under inspection.

[0038] (Control Unit) The control unit 10 comprises a data processing unit 11, a first training phase execution unit 12, a first estimation phase execution unit 13, a second training phase execution unit 14, and a second estimation phase execution unit 15. Each part of the control unit 10 is realized by the control unit 10 reading and executing instructions of a computer program stored in the storage unit 20. The data processing unit 11 comprises an acquisition unit 111, a specification unit 112, and a standardization unit 113. The first training phase execution unit 12 comprises a training unit 121. The first estimation phase execution unit 13 comprises a first estimation unit 131. The second training phase execution unit 14 comprises an extraction unit 141 and a retraining unit 142. The second estimation phase execution unit 15 comprises a second estimation unit 151.

[0039] (Acquisition Department) The acquisition unit 111 generates a learning spectral response function for a healthy learning structure based on the response acceleration measured by the sensor 2. More specifically, as an example, the acquisition unit 111 extracts the response acceleration measured by the sensor 2 at a predetermined extraction time, calculates multiple amplitude spectra (for example, 30 amplitude spectra) by performing a fast Fourier transform, and calculates a learning spectral response function by averaging these multiple amplitude spectra. However, the method by which the acquisition unit 111 generates the learning spectral response function is not limited to the example described above, and a learning spectral response function representing the vibration of the learning structure may be generated by other methods. In the following description, the learning spectral response function for a healthy learning structure is also referred to as "healthy data".

[0040] Furthermore, the acquisition unit 111 may acquire the learning spectral response function input to the input unit 40, or it may receive the learning spectral response function from another device connected via the communication unit 30. Alternatively, the acquisition unit 111 may acquire the learning spectral response function by reading it from a storage location specified by the user of the information processing device 1 (which may be a storage device within the information processing device 1 or a storage device outside the information processing device 1).

[0041] Figure 4 shows an example of a learning spectral response function for a sound learning structure. In Figure 4, the learning spectral response function w1 is the spectral response function of the slab, and the learning spectral response function w2 is the spectral response function of the beam. The horizontal axis represents frequency (Hz), and the vertical axis represents the amplitude spectrum. As shown in Figure 4, the spectral characteristics of a sound beam and slab are significantly different.

[0042] (Specific part) The identification unit 112 identifies a reference frequency having predetermined characteristics in the learning spectral response function (a learning spectral response function obtained by applying elastic waves with a wide frequency band to a healthy learning structure) generated by the acquisition unit 111. The reference frequency is the frequency used by the normalization unit 113, described later, to normalize the learning spectral response function in the frequency axis direction. The reference frequency is set based on existing data and the characteristics of the structure. The reference frequency is, for example, the resonance frequency of the structure, but is not limited to this. The reference frequency may be, for example, the resonance frequency of similar healthy data, or a theoretical resonance frequency, or any arbitrary frequency may be set. Furthermore, the reference frequency may be a frequency identified using the resonance frequency of the structure.

[0043] (Standardization Department) The normalization unit 113 normalizes the learning spectral response function acquired by the acquisition unit 111 in the amplitude direction (vertical axis direction in Figure 3) using its maximum amplitude, and also normalizes it in the frequency direction (horizontal axis direction in Figure 3) using a reference frequency. Here, normalization includes both normalization and standardization. Normalization means normalizing to a number between 0 and 1. Standardization means determining a reference and showing it by its variation. As an example, the normalization unit 113 standardizes the learning spectral response function in the frequency direction by dividing the frequency of the learning spectral response function by the reference frequency. However, the method of standardizing the learning spectral response function is not limited to this, and the normalization unit 113 may standardize the learning spectral response function by other methods.

[0044] Figure 5 schematically illustrates the normalization of the learning spectral response function. As shown in Figure 5, by normalizing the learning spectral response function in the frequency axis direction (scaling to a reference frequency), the learning spectral response function becomes independent of a specific thickness L or a specific stiffness K (the reference frequency matches). Therefore, by using healthy data obtained from any learning structure in conjunction, a general-purpose machine learning model 202 can be constructed. This makes it possible to estimate the state of the structure under inspection even when there is no healthy learning spectral response function for a similar learning structure.

[0045] (Training Department) The training unit 121 trains the machine learning model 202 using a standardized learning spectral response function as training data. The machine learning method used for the machine learning model 202 is not limited; for example, decision tree-based, linear regression, or neural network methods may be used, or two or more of these methods may be used. Examples of decision tree-based methods include LightGBM (Light Gradient Boosting Machine), Random Forest, and XGBoost. Examples of linear regression methods include Bayesian regression, Support Vector Regression, Ridge Regression, Lasso Regression, and ElasticNet. Examples of neural networks include deep learning.

[0046] (1st estimation part) The first estimation unit 131 acquires the spectral response function of the structure under inspection. The spectral response function acquired by the first estimation unit 131 is, for example, a spectral response function obtained by applying elastic waves with a wide frequency band to the structure under inspection. For example, the first estimation unit 131 extracts the response acceleration measured by the sensor 2 at a predetermined extraction time, calculates multiple amplitude spectra (for example, 30 amplitude spectra) by performing a fast Fourier transform, and calculates the spectral response function of the structure under inspection by averaging these multiple amplitude spectra.

[0047] Furthermore, the first estimation unit 131 may acquire the spectral response function to be tested input to the input unit 40, or it may receive the spectral response function to be tested from another device connected via the communication unit 30. Alternatively, the first estimation unit 131 may acquire the spectral response function to be tested by reading it from a storage location specified by the user of the information processing device 1 (which may be a storage device within the information processing device 1 or a storage device outside the information processing device 1).

[0048] When vibrations are measured at multiple locations within the structure under inspection, a spectral response function for the structure under inspection is generated from the measurement results at each location.

[0049] Furthermore, the first estimation unit 131 identifies a reference frequency having predetermined characteristics in the acquired spectral response function of the structure under inspection. The reference frequency identified by the first estimation unit 131 is, for example, the resonance frequency of a healthy structure corresponding to the structure under inspection. Here, a healthy structure corresponding to the structure under inspection is, for example, the structure in the state before an abnormality occurred in the structure under inspection.

[0050] The first estimation unit 131, as an example, identifies the reference frequency of the structure under inspection. More specifically, the first estimation unit 131, as an example, adds density ρ, member thickness L, and dynamic elastic modulus E to equation (1) above. d The resonant frequency f0 obtained by substituting is identified as the reference frequency. Alternatively, the first estimation unit 131 may identify the reference frequency by referring to a pre-constructed database. In this case, the first estimation unit 131 may, for example, refer to a database in which the stiffness, shape, and / or density of the structure are associated with the reference frequency to identify the reference frequency of the structure under inspection. In other words, the first estimation unit 131 can also identify the reference frequency using the stiffness, shape, and density of the structure under inspection.

[0051] Figure 6 shows an example of the relationship between the thickness and compressive strength of a structural member and its resonant frequency. In Figure 6, the horizontal axis represents the member thickness (mm), and the vertical axis represents the resonant frequency (Hz). In the example in Figure 6, the density is 2340 kg / m³. 3 So, the typical concrete strength is (18-50 N / mm²). 2 This figure shows the change in resonance frequency calculated using equation (1) with the stiffness estimated from ). As shown in Figure 6, the compressive strength and member thickness of typical concrete also have a significant effect on the resonance frequency.

[0052] The higher the stiffness, the faster the elastic waves propagate through the material; conversely, the lower the stiffness, the slower the elastic waves propagate. In the case of typical concrete, stiffness can be estimated from the compressive strength, with higher strength corresponding to higher stiffness. Within the elastic range, the stiffness of concrete is defined by the elastic modulus. Furthermore, the resonant frequency is the frequency at which an object vibrates intrinsically, and as shown in equation (1), it is determined by the object's stiffness, shape, and density.

[0053] Furthermore, the first estimation unit 131 normalizes the acquired spectral response function under test in the amplitude direction (vertical axis direction in Figure 3) using its maximum amplitude, and normalizes the spectral response function under test in the frequency axis direction (horizontal axis direction in Figure 3) using the reference frequency identified by the first estimation unit 131. The process by which the first estimation unit 131 normalizes the spectral response function under test is the same as the process performed by the normalization unit 113 described above, and a detailed explanation is omitted here. In the example in Figure 1, the normalization unit 113 and the first estimation unit 131 are shown as different components, but the normalization unit 113 and the first estimation unit 131 may be a single component, or the normalization process may be performed using a common library.

[0054] Furthermore, the first estimation unit 131 estimates the state of the structure under inspection by inputting the normalized spectral response function of the structure under inspection into the machine learning model 202. If the machine learning model 202 is an autoencoder, the first estimation unit 131 estimates the state of the structure under inspection according to the normalized spectral response function of the structure under inspection and the spectral response function of the structure under inspection output from the autoencoder. If the machine learning model 202 is a classification model generated by supervised learning, the first estimation unit 131 estimates the state of the structure under inspection based on the classification result obtained by inputting the spectral response function of the structure under inspection into the machine learning model 202.

[0055] (Extraction part) The extraction unit 141 extracts from among the multiple spectral response functions of the subject to be inspected output by the first estimation unit 131 those whose estimation results by the first estimation unit 131 satisfy predetermined conditions. Here, predetermined conditions are, for example, that the rank when the abnormality degree calculated for each spectral response function of the subject to be inspected is sorted in ascending order is higher than a predetermined rank (e.g., top 10%). Alternatively, if the majority of the structure to be inspected is judged to be in a sound state, it is also effective to train the autoencoder using only the measurement data (spectral response functions of the subject to be inspected) obtained from the structure to be inspected, self-evaluate the data used for training with the trained model, sort it in descending order of abnormality degree, and extract data that is higher than the predetermined rank. In other words, the extraction unit 141 extracts the measurement results of locations that are likely to be sound (no abnormalities have occurred) from among the vibration measurement results at each of the multiple locations of the structure to be inspected.

[0056] (Retraining Department) The retraining unit 142 updates the machine learning model 202 trained by the training unit 121 by machine learning using the spectral response function of the object under inspection extracted by the extraction unit 141 as training data. As an example, the retraining unit 142 updates the machine learning model 202 by transfer learning using the spectral response function of the object under inspection extracted by the extraction unit 141 as training data. In this case, in other words, the machine learning model 202 can also be said to be a model updated by transfer learning using training data in which the spectral response function of the object under inspection, representing multiple measurement results that appear to be healthy for the structure under inspection, is normalized in the amplitude axis direction and normalized in the frequency axis direction.

[0057] However, the method by which the retraining unit 142 updates the machine learning model 202 is not limited to the examples described above. The retraining unit 142 may update the machine learning model 202 by fine-tuning using the spectral response function of the subject under test extracted by the extraction unit 141, for example.

[0058] (Second estimation part) The second estimation unit 151 estimates the state of the structure under inspection using a machine learning model 202. The processing performed by the second estimation unit 151 is the same as that performed by the first estimation unit 131, and a detailed explanation is omitted here.

[0059] <An example of the process for generating and estimating the state of a machine learning model> Figure 7 is a flowchart showing an example of the process for generating (training) a machine learning model 202 performed by the state estimation system 100, and an example of the process for estimating the state. In the example in Figure 7, the processing flow is explained when the machine learning model 202 is an autoencoder.

[0060] (Obtaining the spectral response function for training) In step S11, the acquisition unit 111 acquires a learning spectral response function obtained from the measurement results of vibrations of a healthy learning structure. The learning spectral response function is acquired by the following method as an example. First, the administrator of the state estimation system 100 brings the vibrator 3 and sensor 2 into contact with the surface of the learning structure, which is the test object, and generates white noise, for example, using the vibrator 3 to excite local vibrations of the learning structure, which is the test object. The acquisition unit 111 acquires the response acceleration measured by sensor 2 during excitation. The response acceleration is acquired, for example, at a sampling frequency of 50 kHz.

[0061] Furthermore, for example, in step S12, the acquisition unit 111 may augment the training data 201. As an example, the acquisition unit 111 may adjust one learning spectral response function f to accommodate slight changes in the learning spectral response function due to differences in how the vibrator 3 is pressed against the test specimen during vibration measurement. i to f i The data is copied to -s(-3≦s≦+3) and shifted along the frequency axis, resulting in a 7x data augmentation. The acquisition unit 111 may set to zero any frequency components that would result in data loss due to the shift. The acquisition unit 111 may also be configured not to perform the processing in step S12.

[0062] (Normalization of the spectral response function for training) In step S12 (an example of a specific step, an example of a normalization step), the specific unit 112 identifies a reference frequency having predetermined characteristics in the learning spectral response function (a spectral response function obtained by applying elastic waves with a wide frequency band to a learning structure) generated by the acquisition unit 111, and the normalization unit 113 normalizes the learning spectral response function in the amplitude direction and normalizes the learning spectral response function in the frequency axis direction using the reference frequency. By normalizing the amplitude axis direction in the learning spectral response function, variations due to the contact state between the vibrator 3 and sensor 2 and the surface of the learning structure, as well as variations due to the measurer, can be reduced, and by normalizing in the frequency axis direction using the reference frequency, the spectral characteristics become independent of a specific thickness L or a specific stiffness K. Therefore, sound data (learning spectral response function) obtained from different measurers, surface conditions, and different types of learning structures (member thickness, stiffness, etc.) can be jointly used as training data.

[0063] (Machine Learning) In step S13 (an example of a training step), the training unit 121 trains the machine learning model 202 using a normalized learning spectral response function as training data.

[0064] (Acquisition of the spectral response function under examination) In step S21, the first estimation unit 131 obtains the spectral response function of the structure under inspection obtained by measuring the vibration of the structure under inspection. The spectral response function of the structure under inspection is obtained by the following method as an example. First, the administrator of the state estimation system 100 brings the vibrator 3 and the sensor 2 into contact with the surface of the structure under inspection, and the vibrator 3 generates, for example, white noise to excite local vibrations of the structure under inspection. The first estimation unit 131 obtains the response acceleration measured by the sensor 2 during the excitation. The response acceleration is obtained, for example, at a sampling frequency of 50 kHz. For example, the first estimation unit 131 extracts the waveform at a predetermined extraction time (for example, 0.3072 seconds). For example, the first estimation unit 131 performs a fast Fourier transform 30 times consecutively on the extracted waveform and generates the spectral response function of the structure under inspection by averaging the 30 amplitude spectra.

[0065] (standardization) In step S22 (an example of the inspection target normalization step), the first estimation unit 131 normalizes the generated inspection target spectral response function. More specifically, the first estimation unit 131 first identifies a reference frequency having predetermined characteristics. The first estimation unit 131 also normalizes the acquired inspection target spectral response function in the amplitude direction and normalizes the inspection target spectral response function in the frequency axis direction using the identified reference frequency. Here, the first estimation unit 131 normalizes the inspection target spectral response functions measured at multiple locations on a single inspection target structure in the amplitude direction and the frequency axis direction, respectively.

[0066] (Estimation of state) In steps S23 to S24 (an example of the estimation step), the first estimation unit 131 estimates the state of the structure under inspection based on the output data output from the machine learning model 202 by inputting the normalized spectral response function of the structure under inspection into the machine learning model 202. Here, if the machine learning model 202 is an autoencoder, the first estimation unit 131 estimates the state of the structure under inspection according to the spectral response function of the structure under inspection input to the autoencoder and the reconstructed spectral response function output from the autoencoder. Here, the first estimation unit 131 estimates the above state by inputting the spectral response functions of multiple locations measured on a single structure under inspection into the machine learning model 202. Steps S23 to S24 will be explained in order below.

[0067] First, the first estimation unit 131 inputs the normalized spectral response function of the object under test to the autoencoder and obtains the reconstructed spectral response function output from the autoencoder. For the sake of explanation, the normalized spectral response function of the object under test input to the autoencoder will be referred to as the "measured spectrum," and the reconstructed spectral response function, which is reconstructed in a normalized state and output from the autoencoder, will be referred to as the "healthy spectrum."

[0068] Furthermore, the first estimation unit 131 performs smoothing and noise reduction processing on both the measured spectrum and the healthy spectrum. As an example, the first estimation unit 131 performs smoothing by calculating a moving average with a predetermined parameter width. Because the spectral response function has a steep multimodal nature, even a slight difference between the peak frequency of the spectral response function under inspection input to the autoencoder and the peak frequency of the reconstructed spectral response function results in a large reconstruction error. Therefore, the peaks of the spectral response function are smoothed by calculating a moving average.

[0069] Also, since most of the frequency components of the plurality of parameters are noise, the first estimator 131 subtracts a certain threshold value and sets the components that become negative values to zero to perform noise removal. As an example, the threshold value may be set to a height of 0.2 when the maximum value of the spectral response function before performing the moving average is set to 1.

[0070] Also, the first estimator 131 normalizes each of the measured spectrum and the healthy spectrum in the amplitude direction. Although these spectral data are already normalized, the amplitude of the spectral response function changes due to the smoothing process and the noise cutting process, so normalization is performed again to maintain the amplitude normality.

[0071] Also, the first estimator 131 performs an evaluation regarding the state of the inspection target structure based on the measured spectrum and the healthy spectrum. As an example of the evaluation, it is a value indicating the degree of health (abnormality degree) of the inspection target structure. More specifically, as an example, the first estimator 131 calculates at least any one of the following abnormality degrees Ansco, Ansco2, and Ansco3 as an evaluation of the inspection target structure.

[0072] FIG. 8 is a diagram for explaining a method of calculating the abnormality degrees Ansco, Ansco2, and Ansco3. In FIG. 8, the reconstructed spectral response function w22 is a healthy spectrum output from the autoencoder. The inspection target spectral response function w21 is a measured spectrum obtained by measuring the vibration of the inspection target structure. In the diagram for explaining this method of calculating the abnormality degree, the amplitude axis uses a normalized one, but the frequency axis direction may be normalized or not normalized (the abnormality degree results are the same). In the figure, the horizontal axis indicates the frequency (Hz), and the vertical axis indicates the amplitude spectrum. The frequency fs is the center of gravity of the reconstructed spectral response function w22. The amplitude Am i is the amplitude spectrum of the frequency f of the inspection target spectral response function w21. The amplitude As i is the amplitude spectrum of the frequency f in the reconstructed spectral response function w22. j is the amplitude spectrum of the frequency f in the reconstructed spectral response function w22. j is the amplitude spectrum of the frequency f in the reconstructed spectral response function w22.

[0073] (Abnormality Ansco) The first estimation unit 131 calculates the anomaly score Ansco as an example using the following formula (2). In formula (2), As j Each frequency f of the amplitude-normalized measurement spectrum j This is the amplitude, Am i Each frequency f of the amplitude-normalized healthy spectrum i This is the amplitude. In this example, the anomaly score Ansco is the sum of the absolute values ​​of the amplitude differences between the measured spectrum and the healthy spectrum.

number

[0074] (Abnormal degree Ansco2, Ansco3) The above-mentioned anomaly score Ansco may not reflect the extent of the damage. This is because the method using the anomaly score Ansco only considers the amplitude difference between the measured spectrum and the healthy spectrum. For example, even if there is a large decrease in the resonant frequency, the anomaly score may be calculated to be small. Therefore, the extent of the damage may be evaluated using anomaly scores Ansco2 and Ansco3, which are based on the area moment of the spectral response function. As an example, the first estimation unit 131 may calculate anomaly scores Ansco2 and Ansco3 using the following equation (3). In equation (3), fs is the centroid of the healthy spectrum (see Figure 8), and the amplitude Am i The frequency f of the spectral response function w21 under examination. i This is the amplitude. Amplitude As j The frequency f in the reconstructed spectral response function w22 is j This is the amplitude.

number

[0075] In this case, to put it another way, the first estimation unit 131 can also calculate the degree of abnormality regarding the state of the structure under inspection using the area moment Mm of the measured spectrum and the area moment Ms of the healthy spectrum with respect to the centroid fs of the healthy spectrum. More specifically, the first estimation unit 131 can also calculate the degree of abnormality regarding the state of the structure under inspection as the value obtained by dividing the area moment Mm of the measured spectrum by the area moment Ms of the healthy spectrum (Mm / Ms) with respect to the centroid fs of the healthy spectrum. Alternatively, the first estimation unit 131 can also calculate the degree of abnormality regarding the state of the structure under inspection as the square root of the value obtained by dividing the area moment Mm of the measured spectrum by the area moment Ms of the healthy spectrum (Mm / Ms) with respect to the centroid fs of the healthy spectrum.

[0076] (Extraction of relatively healthy data) In step S31 (an example of an extraction step), the extraction unit 141 extracts relatively healthy data. For example, the extraction unit 141 extracts from among multiple measurement spectra those whose estimation results using the machine learning model 202 satisfy predetermined conditions. More specifically, as an example, the extraction unit 141 extracts from among the multiple measurement spectra used to calculate the anomaly score those whose rank is higher than a predetermined rank when the calculated anomaly scores are sorted in ascending order.

[0077] The retraining unit 142 performs an update process (update step) to update the machine learning model 202 using machine learning with the measurement spectrum extracted by the extraction unit 141 as training data. As an example, the retraining unit 142 updates the final layer of the machine learning model 202 using transfer learning with the measurement spectrum extracted by the extraction unit 141 as training data.

[0078] As described above, by normalizing the spectral response function in the frequency axis direction, the condition of the structure under inspection can be evaluated even without sound data for similar structures. On the other hand, if sound data for the same type of structure as the structure under inspection is available, the accuracy of the machine learning model 202, which was trained on a different training structure, can be improved by using transfer learning.

[0079] Figure 9 shows an example of the update process performed by the retraining unit 142. In the example in Figure 9, the retraining unit 142 retrains the final layer 2021 of the machine learning model 202, which was trained in steps S11 to S13 of Figure 7, using data representing the measurement results of the structure under inspection or data of a small amount of healthy structures of the same type as the structure under inspection. The main advantage of retraining is that if a small amount of healthy data of the structure under inspection is available, the pre-trained machine learning model 202 can be customized for the structure under inspection. Hereafter, the machine learning model 202 before the retraining unit 142 performs the update process will also be referred to as the "general-purpose model".

[0080] For example, not all parts of a structure under inspection are damaged; often, sound parts remain. Therefore, by acquiring even a small amount of information on the spectral response function of the healthy state of the structure under inspection and customizing the general-purpose machine learning model 202 constructed in steps S11 to S13 of Figure 7, estimation accuracy can be improved. By incorporating actual measured values ​​of the structure under inspection, the type of concrete, boundary conditions, and the influence of the vibrator are incorporated into the model, resulting in improved prediction accuracy of the spectral response function in a healthy state. Note that if no spectral response function of a healthy state of the structure under inspection can be obtained, the retraining unit 142 does not perform the retraining process.

[0081] The second estimation unit 151 estimates the state of the structure under inspection using the machine learning model 202 generated by retraining. More specifically, the second estimation unit 151 performs the processes shown in steps S21 to S24 of Figure 7. The content of the processes performed by the second estimation unit 151 is the same as that performed by the first estimation unit 131, and a detailed explanation is omitted here.

[0082] <Effects of the Embodiment> As mentioned above, the spectral response function of structure 4 varies depending on the member thickness and stiffness, and for example, the spectral characteristics of a floor slab and a beam differ significantly, as shown in Figure 4. Therefore, if the spectral response function is used as training data without normalizing it along the frequency axis, there is a problem that the estimation accuracy will be low depending on the type of structure. For example, a model trained using training data for a slab cannot appropriately estimate the state of a beam. Furthermore, there are differences in stiffness and density depending on the type of cement (e.g., ordinary cement or blast furnace cement), and considering that the measured vibration characteristics (spectral response function) also depend on boundary conditions (structural shape, fixing conditions, load conditions) and the characteristics of the exciter 3, the number of required data cases becomes enormous, making it practically difficult to comprehensively obtain sound data.

[0083] In contrast, according to this embodiment, the spectral response function is normalized in the amplitude axis direction using a reference frequency, and the machine learning model 202 is trained using the spectral response function normalized in the frequency axis direction. As a result, the spectral characteristics no longer depend on a specific member thickness or specific stiffness. Therefore, sound data obtained from different measurers, surface conditions, and different types of structures 4 (member thickness, stiffness, etc.) can be jointly used as training data. In other words, according to this embodiment, even when there is no sound data or the sound data is insufficient, a machine learning model 202 capable of more accurately estimating the state of various types of structures 4 can be constructed.

[0084] Furthermore, according to this embodiment, a general-purpose model is trained based only on an arbitrary healthy spectral response function, and the model is updated by transfer learning or the like to match the target structure (concrete stiffness, member thickness, boundary conditions, etc.). This makes it possible to estimate the state of various structures. In addition, according to this embodiment, in order to efficiently utilize important information of the spectral response function as an anomaly score, an index can be provided that effectively utilizes frequency band information other than the resonant frequency.

[0085] [Example 1] An example of state estimation using the state estimation system 100 according to the above embodiment will be described. In this example, the damage to a structure will be evaluated (output), assuming practical use.

[0086] In this example, data from a wheel load test is used to evaluate the damage to the road PC slab. The wheel load test is a test that evaluates the fatigue resistance of a road PC slab by applying a wheel load multiple times (for example, 250kN more than 100,000 times).

[0087] Figure 10 shows the test flow, and Figure 11 shows the test cycle. Figure 12 shows a crack sketch (bottom surface) of a damaged slab specimen, and Figure 13 shows the damage status (bottom surface) at the end of the test (end of STEP 2). The subjects evaluated were the damaged slab specimens shown in Figures 12 and 13, which were subjected to the test flow shown in Figure 10 and the test cycle shown in Figure 11.

[0088] Figure 14 is a table showing evaluation cases for internal damage based on the degree of abnormality. Local vibration measurements were performed on the deck slab test specimen at the end of the test, and internal damage was evaluated in four scenarios. Case 1 in Figure 14 is an ideal scenario in which a large amount of sound data of the inspected structure can be obtained. This is an extremely rare scenario in practice, but because a dedicated model of the inspected structure can be constructed, it is a case in which the evaluation of internal damage is highly accurate (i.e., this is the correct value).

[0089] Cases 2-4 are scenarios that closely resemble real-world situations, assuming that a large amount of sound data for the structure to be inspected cannot be obtained in advance. A general-purpose model is trained using sound data obtained from arbitrary structures, and only in Case 4 is transfer learning performed using a small amount of sound data for the structure to be inspected.

[0090] Figure 15 is a contour plot of the degree of abnormality. In Figure 15, contour plots C1 to C4 are the contour plots of the degree of abnormality for cases 1 to 4 in Figure 14, respectively. Figure 16 is a diagram showing the correlation between case 1 and cases 2, 3, and 4.

[0091] Case 2 is a case where frequency normalization is not performed, so the general-purpose model trained on arbitrary structures estimates a high degree of anomaly at all measurement points. As shown in Figure 16, no correlation is observed with Case 1. This suggests that a general-purpose model trained on healthy data of structures with different thicknesses and stiffnesses cannot be directly applied.

[0092] In Case 3, frequency normalization is performed, so the trained model produces results that do not depend on the thickness and stiffness of the structure being inspected. As shown in Figure 15 for Case 3, the anomaly distribution is relatively similar to that of Case 1, but the anomaly values ​​in Case 3 are slightly higher than in Case 1. In this case, although a correlation is observed as shown in Figure 16, it is located below the 1:1 line (i.e., the anomaly values ​​are higher than in Case 1). This is because the specific characteristics of the structure being inspected (type of concrete, boundary conditions, etc.) are not reflected in the general-purpose model.

[0093] Case 4 utilizes a small amount of data from seemingly healthy regions of the test specimen at the end of the test and applies transfer learning (Case 4 in Figure 15). As shown in Figure 16, the anomaly score shows a strong correlation with Case 1 and is located near the 1:1 line. These results show that Case 4 is very similar to the ground truth value in Case 1, and that the general-purpose model can be customized by introducing normalization and transfer learning.

[0094] Figure 17 shows the contour map of the abnormality level for Case 1 and the cross-section of the test specimen. The test specimen was cut in the direction of the bridge axis for direct observation of internal damage, and the cross-section is shown in the lower part of Figure 17. Figure 18 shows the abnormal values ​​at measurement points near the cutting line (rows C and D, straddling the cutting line). Cracks that have progressed into the interior of the deck slab were observed, and the numbers on the cross-section indicate the maximum crack width measured with a crack gauge.

[0095] As shown in Figure 18, Panel A showed a higher degree of abnormality (compared to Panel B), which correlates with the larger number of internal cracks observed on the cross-section of Panel A (see Figure 17). From these observations, it can be confirmed that the PC slab has not only surface cracks but also numerous internal cracks. In particular, Panel A showed a relatively high degree of abnormality, which corresponds to the fact that many cracks occurred on its cross-section. While there is no significant difference in the damage between Panel A and Panel B based solely on the observed bottom surface crack distribution, it has been shown that internal damage can be appropriately evaluated by the state estimation method according to this embodiment.

[0096] Figure 19 shows examples of the spectral response function under test and the reconstructed spectral response function in Case 3 and Case 4 of Figure 14. The upper part of Figure 19 shows an example of the spectral response function under test and the reconstructed spectral response function at a healthy measurement site in Case 3, and the lower part of Figure 19 shows an example of the spectral response function under test and the reconstructed spectral response function at a healthy measurement site in Case 4. Since it is a healthy measurement site, the spectral response function under test and the reconstructed spectral response function must match. Comparing Case 3 and Case 4, although the peaks are aligned in both Case 3 and Case 4, the shapes of the spectral response function under test and the reconstructed spectral response function are more similar in Case 4 than in Case 3.

[0097] [Example 2] An example of measurement on an actual structure is shown below. Figure 20 is a plan view and a cross-sectional view showing the condition of the pier that was measured. The measurement was performed from above on the pier's beam (beam height 1.4m). Using the general-purpose model described in Example 1 (training data: arbitrary healthy area measurements, 360 locations), a small amount of healthy data from the seemingly healthy areas of this pier (locations with low anomaly scores estimated by the general-purpose model) was used, and transfer learning was applied.

[0098] Figure 21 shows the estimated degree of anomaly for each measurement point. Figure 22 shows the spectral response function of measurement point 8 (left), which shows the minimum degree of anomaly, and the spectral response function of measurement point 14 (right), which shows the maximum degree of anomaly. As shown in Figure 22, measurement point 14, which has a high degree of anomaly, shows a difference in the shape of the estimated healthy spectrum and the measured spectrum, suggesting internal damage to the beam.

[0099] [Examples of implementation using software] The functions of the information processing device 1 (hereinafter referred to as "the device") are programs that cause the device to function as a computer, and these programs can be realized by programs that cause each control block of the device (especially each part included in the control unit 10) to function as a computer.

[0100] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., memory) as hardware for executing the program. By executing the program using this control device and storage device, the functions described in each of the embodiments are realized.

[0101] The above program may be recorded on one or more computer-readable recording media, not temporary ones. These recording media may or may not be provided by the above device. In the latter case, the program may be supplied to the above device via any wired or wireless transmission medium.

[0102] Furthermore, some or all of the functions of each of the above control blocks can also be realized by logic circuits. For example, an integrated circuit in which logic circuits functioning as each of the above control blocks are formed is also included in the scope of the present invention. In addition, it is also possible to realize the functions of each of the above control blocks by, for example, a quantum computer.

[0103] Furthermore, each process described in the above embodiments may be performed by AI (Artificial Intelligence). In this case, the AI ​​may operate on the control device described above, or it may operate on other devices (for example, an edge computer or a cloud server).

[0104] The present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

[0105] 〔summary〕 A method for training a machine learning model according to Embodiment 1 of the present invention includes: a identification step of identifying a reference frequency having predetermined characteristics in a learning spectral response function obtained by applying elastic waves having a wide frequency band to a learning structure; a normalization step of normalizing the learning spectral response function in the frequency axis direction using the reference frequency; and a training step of training a machine learning model that uses the learning spectral response function obtained in the normalization step as training data, takes an inspected spectral response function obtained by measuring the vibration of an inspected structure as input data, and outputs data used to estimate the state of the inspected structure as output data.

[0106] According to the above embodiment, a machine learning model is trained using a learning spectral response function with a wide frequency band, as well as a reference frequency (e.g., resonant frequency) having predetermined characteristics. Therefore, by estimating the state of the structure under inspection using a machine learning model trained using this training method, vibration characteristics other than the reference frequency can also be considered. Thus, even if measurement data of a healthy learning structure is unavailable, the state of the structure under inspection can be estimated. Furthermore, as shown in Figure 19, frequency band information other than the resonant frequency can also be effectively utilized.

[0107] Furthermore, the training method according to aspect 2 of the present invention is characterized in that, in the training method described in aspect 1, in the normalization step, the learning spectral response function is normalized in the frequency axis direction using the reference frequency, and the learning spectral response function is normalized in the amplitude axis direction. According to the above aspect, by normalizing the amplitude of the learning spectral response function, it is possible to reduce the contact state between the vibrator 3 and the sensor 2 and the surface of the learning structure, as well as variations due to the measurer.

[0108] Furthermore, the training method according to aspect 3 of the present invention is characterized in that, in the training method described in aspect 1 or 2, the machine learning model is an autoencoder. According to the above aspect, the autoencoder is trained using a learning spectral response function having a wide frequency band, as well as a reference frequency (e.g., resonant frequency) having predetermined characteristics. Therefore, by estimating the state of the structure under inspection using the autoencoder trained using this training method, vibration characteristics other than the reference frequency can also be considered, so that the state of the structure under inspection can be estimated even if there is no measurement data of a healthy learning structure.

[0109] Furthermore, the training method according to aspect 4 of the present invention is characterized in that, in the training method described in any one of aspects 1 to 3, the reference frequency is the resonant frequency of the learning structure. According to the above aspect, by normalizing the learning spectral response function in the frequency axis direction using the resonant frequency, the soundness of the structure to be inspected can be estimated even when there is no measurement data for a sound learning structure.

[0110] Furthermore, the training method according to aspect 5 of the present invention further includes, in the training method described in any one of aspects 1 to 4, an inspection target normalization step in which the inspection target spectral response function representing the result of measuring the vibration of the inspection target structure is normalized in the frequency axis direction, and an estimation step in which the state of the inspection target structure is estimated based on output data output from the machine learning model by inputting the inspection target spectral response function obtained in the inspection target normalization step into the machine learning model. According to the above aspect, even if there is no measurement data of a healthy training structure, the state of the inspection target structure can be estimated.

[0111] Furthermore, in the training method according to embodiment 6 of the present invention, the machine learning model is an autoencoder, and in the estimation step, the state of the structure to be inspected is estimated according to the spectral response function of the structure to be inspected and the reconstructed spectral response function output from the autoencoder.

[0112] According to the above embodiment, even if measurement data for a healthy learning structure is unavailable, the state of the learning inspection target structure can be estimated by using the output of the autoencoder.

[0113] Furthermore, the training method according to embodiment 7 of the present invention further includes, in the training method described in embodiment 5 or 6, the following steps: in the inspection target standardization step, the inspection target spectral response functions representing the measurement results of vibrations at each of the multiple locations of the inspection target structure are standardized in the frequency axis direction; in the estimation step, the state is estimated for each of the multiple locations by inputting each of the inspection target spectral response functions obtained in the inspection target standardization step into the machine learning model; an extraction step is taken to extract from the multiple inspection target spectral response functions obtained in the inspection target standardization step those whose estimation results in the estimation step satisfy predetermined conditions; and an update step is taken to update the machine learning model trained in the training step by machine learning using the inspection target spectral response functions extracted in the extraction step as training data.

[0114] According to the above-described embodiment, the estimation accuracy of the machine learning model can be improved by updating the machine learning model using a portion of the estimation results of the state of the structure under inspection.

[0115] Furthermore, in the training method according to aspect 8 of the present invention, in the training method described in aspect 7, the machine learning model is updated in the update step by transfer learning using the spectral response function of the subject to be tested obtained in the standardization step of the subject to be tested, which was extracted in the extraction step, as training data.

[0116] According to the above embodiment, the estimation accuracy of the machine learning model can be improved by using a portion of the estimation results of the state of the inspected structure to transfer learning the machine learning model.

[0117] Furthermore, the training method according to aspect 9 of the present invention, in the training method described in any one of aspects 5 to 8, in the estimation step, calculates the degree of abnormality regarding the state of the structure to be inspected based on the output data obtained by inputting the spectral response function of the structure to be inspected obtained in the inspection target standardization step into the machine learning model. According to the above aspect, even if there is no measurement data of a healthy training structure, the degree of abnormality regarding the state of the structure to be inspected can be calculated. In addition, by calculating the degree of abnormality using the above-described equations (2) and (3), frequency band information other than the resonant frequency can also be effectively utilized.

[0118] Furthermore, the training method according to embodiment 10 of the present invention, in the training method described in embodiment 7 or 8, in the estimation step, calculates the degree of abnormality regarding the state of the structure to be inspected based on the output data obtained by inputting the spectral response function of the structure to be inspected obtained in the inspection target standardization step into the machine learning model, and in the extraction step, extracts from among the plurality of spectral response functions of the structure to be inspected that have a rank higher than a predetermined rank when the calculated degree of abnormality is sorted in ascending order.

[0119] According to the above embodiment, the estimation accuracy of the machine learning model can be improved by sorting the degree of abnormality among the estimated state results at multiple locations of the structure under inspection from smallest to largest, and updating the machine learning model using those with a higher rank than a predetermined rank.

[0120] Furthermore, the training method according to aspect 11 of the present invention is characterized in that, in the training method described in aspect 9 or 10, the abnormality level is a value indicating the degree of soundness of the structure to be inspected. According to the above aspect, even if there is no measurement data for a sound training structure, the degree of soundness of the structure to be inspected can be estimated.

[0121] Furthermore, in the training method according to embodiment 12 of the present invention, in the training method described in any one of embodiments 5 to 11, the reference frequency used for normalizing the spectral response function of the object under inspection is the resonant frequency of a sound structure corresponding to the object under inspection. According to the above embodiment, the state of the object under inspection can be estimated with greater accuracy.

[0122] Furthermore, in the training method according to aspect 13 of the present invention, in the training method described in aspect 12, in the inspection target standardization step, a reference frequency to be used for standardizing the inspection target spectral response function is identified using the rigidity, shape, and density of the inspection target structure, and the inspection target spectral response function is standardized using the identified reference frequency.

[0123] According to the above embodiment, the condition of the structure under inspection can be estimated with greater accuracy by specifying a reference frequency using at least one of the rigidity, shape, and density of the structure under inspection.

[0124] Furthermore, in the training method according to aspect 14 of the present invention, in the training method described in aspect 6, in the estimation step, the degree of abnormality regarding the state of the structure to be inspected is calculated using the area moment of the spectral response function of the object to be inspected obtained in the inspection object normalization step with respect to the centroid of the reconstructed spectral response function and the area moment of the reconstructed spectral response function. According to the above aspect, the state of the structure to be inspected can be estimated with greater accuracy.

[0125] Furthermore, in the training method according to aspect 15 of the present invention, in the training method described in aspect 14, in the estimation step, the area moment of the inspected structure spectral response function obtained in the inspected structure normalization step with respect to the centroid of the reconstructed spectral response function is divided by the area moment of the reconstructed spectral response function, or the square root of that value is calculated as the degree of abnormality regarding the state of the inspected structure. According to the above aspect, the state of the inspected structure can be estimated with greater accuracy.

[0126] Furthermore, the method for estimating a structure according to embodiment 15 of the present invention includes: a identification step of identifying a reference frequency having predetermined characteristics in a spectral response function of a structure to be inspected obtained by applying elastic waves having a wide frequency band to the structure to be inspected; a normalization step of normalizing the spectral response function of the structure to be inspected in the frequency axis direction using the reference frequency; and an estimation step of estimating the state of the structure to be inspected by inputting the spectral response function of the structure to be inspected obtained in the normalization step into a machine learning model that takes the spectral response function of the structure to be inspected obtained by measuring the vibration of the structure to be inspected as input data and data used to estimate the state of the structure to be inspected as output data.

[0127] According to the above embodiment, vibration characteristics other than the reference frequency can also be considered, so even if measurement data for a sound learning structure is unavailable, the condition of the structure under inspection can be estimated.

[0128] Furthermore, the estimation method according to aspect 17 of the present invention is the estimation method described in aspect 16, wherein the machine learning model is an autoencoder, and in the estimation step, the state of the structure to be inspected is estimated according to the spectral response function of the structure to be inspected obtained in the normalization step and the spectral response function of the structure to be inspected output from the autoencoder.

[0129] With the above configuration, the state of the structure under inspection can be estimated with greater accuracy by inputting the spectral response function of the structure under inspection, normalized using a reference frequency, into an autoencoder and estimating the state of the structure under inspection in accordance with the output obtained and the input to the autoencoder.

[0130] Furthermore, the estimation method according to aspect 18 of the present invention is the estimation method described in aspect 16 or 17, wherein the machine learning model is a model updated by transfer learning using training data in which the spectral response function of the inspected object, which represents the measurement results of vibrations at multiple locations of the inspected object, is normalized in the frequency axis direction.

[0131] With the above configuration, the estimation accuracy of the machine learning model can be improved by updating the machine learning model through transfer learning using training data normalized in the frequency axis direction of the spectral response function of the inspected structure, which represents the measurement results of vibrations at multiple locations of the inspected structure.

[0132] Furthermore, the training apparatus according to embodiment 19 of the present invention comprises: an identification unit that identifies a reference frequency having predetermined characteristics in a learning spectral response function obtained by applying elastic waves having a wide frequency band to a learning structure; a normalization unit that normalizes the learning spectral response function in the frequency axis direction using the reference frequency; and a training unit that trains a machine learning model using the learning spectral response function normalized by the normalization unit as training data, taking an inspection target spectral response function obtained by measuring the vibration of the inspection target structure as input data, and output data used to estimate the state of the inspection target structure.

[0133] According to the above embodiment, a machine learning model is trained using a learning spectral response function with a wide frequency band, as well as a reference frequency (e.g., resonant frequency) having predetermined characteristics. Therefore, by estimating the state of the structure under inspection using a machine learning model trained using this training method, vibration characteristics other than the reference frequency can also be considered. Thus, even if measurement data of a healthy learning structure is unavailable, the state of the structure under inspection can be estimated.

[0134] Furthermore, the estimation device according to embodiment 20 of the present invention includes: an identification unit that identifies a reference frequency having predetermined characteristics in a spectral response function of a structure to be inspected obtained by applying elastic waves having a wide frequency band to the structure to be inspected; a normalization unit that normalizes the spectral response function of the structure to be inspected in the frequency axis direction using the reference frequency; and an estimation unit that estimates the state of the structure to be inspected by inputting the spectral response function of the structure to be inspected, which is obtained by measuring the vibration of the structure to be inspected, as input data and data used to estimate the state of the structure to be inspected as output data, into a machine learning model.

[0135] According to the above embodiment, vibration characteristics other than the reference frequency can also be considered, so even if measurement data for a sound learning structure is unavailable, the condition of the structure under inspection can be estimated.

[0136] The program according to aspect 21 of the present invention is a program for causing a computer to function as a training device as described in aspect 19, and is a program for causing the computer to function as the specified unit, the standardization unit, and the training unit.

[0137] The program according to aspect 22 of the present invention is a program for causing a computer to function as an estimation device as described in aspect 20, and is a program for causing the computer to function as the identification unit, the standardization unit, and the estimation unit. [Explanation of symbols]

[0138] 1. Information Processing Device 2 sensors 3. Vibrator 4 Structures 10 Control Unit 11 Data Processing Unit 12. First Training Phase Execution Unit 13. First Estimated Phase Execution Unit 14. Second Training Phase Execution Unit 15. Second Estimated Phase Execution Unit 20 Memory section 30 Communications Department 40 Input section 50 First Output Section 60 Second Output Section 100 State Estimation Systems 111 Acquisition Department 112 Specific section 113 Standardization Section 121 Training Department 131 1st estimation part 141 Extraction part 142 Retraining Department 151 Second estimation part

Claims

1. A selection step involves identifying a reference frequency having predetermined characteristics in a learning spectral response function obtained by applying elastic waves with a wide frequency band to a learning structure, A normalization step in which the learning spectral response function is normalized in the frequency axis direction using the reference frequency, A training step involves training a machine learning model that uses the learning spectral response function obtained in the normalization step as training data, the spectral response function of the structure under inspection obtained by measuring the vibration of the structure under inspection as input data, and the data used to estimate the state of the structure under inspection as output data. A method for training machine learning models, including the following.

2. In the normalization step, the learning spectral response function is normalized in the frequency axis direction using the reference frequency, and the learning spectral response function is normalized in the amplitude axis direction. A method for training a machine learning model according to claim 1.

3. The aforementioned machine learning model is an autoencoder. A method for training a machine learning model according to claim 1 or 2.

4. The reference frequency is the resonant frequency of the learning structure. A method for training a machine learning model according to claim 1 or 2.

5. A standardization step for the inspected structure, which normalizes the spectral response function of the inspected structure, which represents the result of measuring the vibration of the inspected structure, in the frequency axis direction, The method further includes an estimation step of estimating the state of the structure to be inspected based on output data output from a machine learning model, by inputting the spectral response function of the structure to be inspected obtained in the standardization step of the structure to be inspected into the machine learning model. A method for training a machine learning model according to claim 1 or 2.

6. The aforementioned machine learning model is an autoencoder. In the estimation step described above, the state of the structure under inspection is estimated according to the spectral response function of the structure under inspection and the reconstructed spectral response function output from the autoencoder. A method for training a machine learning model according to claim 5.

7. In the standardization step of the object to be inspected, the spectral response function of the object to be inspected, which represents the measurement results of the vibrations at each of the multiple locations of the object to be inspected, is standardized in the frequency axis direction. In the estimation step, the state is estimated for each of the multiple locations by inputting each of the spectral response functions of the target under test obtained in the standardization step of the target under test into the machine learning model. An extraction step in which, from among a plurality of spectral response functions of the subject to be tested obtained in the standardization step of the subject to be tested, those that satisfy predetermined conditions in the estimation step, An update step, which updates the machine learning model trained in the training step by machine learning using the spectral response function of the subject to be examined extracted in the extraction step as training data, A method for training a machine learning model according to claim 5, further comprising:

8. In the update step, the machine learning model is updated by transfer learning using the spectral response function of the subject to be tested obtained in the normalization step of the subject to be tested, which was extracted in the extraction step, as training data. A method for training a machine learning model according to claim 7.

9. In the estimation step, the degree of abnormality regarding the state of the structure under inspection is calculated based on the output data obtained by inputting the spectral response function of the structure under inspection obtained in the inspection target standardization step into the machine learning model. A method for training a machine learning model according to claim 5.

10. In the estimation step, the degree of abnormality regarding the state of the structure under inspection is calculated based on the output data obtained by inputting the spectral response function of the structure under inspection obtained in the inspection target standardization step into the machine learning model. In the extraction step, from among the multiple spectral response functions of the subject to be examined used to calculate the abnormality score, those whose rank is higher than a predetermined rank when the calculated abnormality scores are sorted in ascending order are extracted. A method for training a machine learning model according to claim 7.

11. The aforementioned degree of abnormality is a value that indicates the degree of soundness of the structure being inspected. A method for training a machine learning model according to claim 9.

12. The reference frequency used for normalizing the spectral response function of the object under inspection is the resonant frequency of a sound structure corresponding to the object under inspection. A method for training a machine learning model according to claim 5.

13. In the standardization step of the object to be inspected, the reference frequency used for standardizing the spectral response function of the object to be inspected is determined using the rigidity, shape, and density of the structure to be inspected, and the spectral response function of the object to be inspected is standardized using the determined reference frequency. A method for training a machine learning model according to claim 12.

14. In the estimation step, the degree of abnormality regarding the state of the structure under inspection is calculated using the area moment of the spectral response function of the object under inspection obtained in the inspection object normalization step with respect to the centroid of the reconstructed spectral response function and the area moment of the reconstructed spectral response function. A method for training a machine learning model according to claim 6.

15. In the estimation step, the area moment of the spectral response function of the object under inspection obtained in the inspection object normalization step with respect to the centroid of the reconstructed spectral response function is divided by the area moment of the reconstructed spectral response function, or the square root of that value is calculated as the degree of abnormality regarding the state of the structure under inspection. A method for training a machine learning model according to claim 14.

16. A selection step involves identifying a reference frequency having predetermined characteristics in the spectral response function of the object to be inspected, obtained by applying elastic waves with a wide frequency band to the object to be inspected, and A normalization step in which the spectral response function under test is normalized in the frequency axis direction using the aforementioned reference frequency, An estimation step is performed by inputting the spectral response function of the structure to be inspected, obtained by measuring the vibration of the structure to be inspected, into a machine learning model, which takes the spectral response function of the structure to be inspected obtained by measuring the vibration of the structure to be inspected as input data, and the data used to estimate the state of the structure to be inspected as output data, into the machine learning model, thereby estimating the state of the structure to be inspected. A method for estimating the state of a structure, including [specific elements].

17. The aforementioned machine learning model is an autoencoder. In the estimation step, the state of the structure to be inspected is estimated according to the spectral response function of the structure to be inspected obtained in the normalization step and the spectral response function of the structure to be inspected output from the autoencoder. A method for estimating the state of a structure according to claim 16.

18. The aforementioned machine learning model is a model updated by transfer learning using training data in which the spectral response function of the inspected object, which represents the measurement results of vibrations at multiple locations of the inspected object, is normalized in the frequency axis direction. A method for estimating the state of a structure according to claim 16 or 17.

19. In a learning spectral response function obtained by applying elastic waves with a wide frequency band to a learning structure, a specification unit identifies a reference frequency having predetermined characteristics, A normalization unit that normalizes the learning spectral response function in the frequency axis direction using the reference frequency, A training unit trains a machine learning model that uses the learning spectral response function standardized by the standardization unit as training data, takes the spectral response function of the structure under inspection obtained by measuring the vibration of the structure under inspection as input data, and outputs data used to estimate the state of the structure under inspection as output data. A training device equipped with the following features.

20. A specific unit identifies a reference frequency having predetermined characteristics in the spectral response function of the object to be inspected, obtained by applying elastic waves with a wide frequency band to the object to be inspected, A normalization unit that normalizes the spectral response function under inspection in the frequency axis direction using the aforementioned reference frequency, An estimation unit estimates the state of a structure under inspection by inputting the spectral response function of the structure under inspection, obtained by measuring the vibration of the structure under inspection, as input data, and data used to estimate the state of the structure under inspection as output data, into a machine learning model; An estimation device equipped with the following features.

21. A program for causing a computer to function as a training device according to claim 19, the program for causing the computer to function as the specific unit, the standardization unit, and the training unit.

22. A program for causing a computer to function as an estimation device according to claim 20, the program for causing the computer to function as the identifying unit, the standardizing unit, and the estimation unit.

Images