Data augmentation methods, programs, trained models, inspection methods, and data augmentation systems

Abstract

To provide a data augmentation method capable of generating multiple training data sets that are used to generate a pre-trained model to determine the presence or absence of defects in concrete by data expansion.SOLUTION: The data augmentation method includes an acquisition step and a generation step. In the acquisition step, multiple actual measured data are acquired. In the generation step, the data is augmented. Data augmentation processing is a series of processing of generating multiple extended data based on information about at least one of the multiple elements and the multiple actual measured data.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure generally relates to a data augmentation method, program, learned model, inspection method, and data augmentation system. More specifically, the present disclosure relates to a data augmentation method, program, and data augmentation system for generating augmented data as learning data. The present disclosure also relates to a learned model generated using the above-described augmented data and an inspection method using this learned model.

Background Art

[0002] The inspection system described in Patent Document 1 performs non-contact inspection on an inspection object. The inspection system includes a first measurement unit, a first acquisition unit, and a first generation unit. The first measurement unit measures a first vibration generated in the inspection object using a first laser and generates first vibration data. The first acquisition unit acquires first inspection object information including first data based on the first vibration data. The first generation unit refers to a first reference database and generates first inspection data for the first inspection object information.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In order to generate a learned model for determining the presence or absence of a defect in an inspection object using information such as the first inspection data in Patent Document 1 as learning data, it is necessary to prepare a large number of learning data including learning data corresponding to the defect. However, when the frequency of occurrence of a defect in the inspection object is low, the number of learning data tends to be insufficient.

[0005] Therefore, there is a technique to increase the number of training data points by generating new augmented data based on existing training data, known as data augmentation. However, in some technical fields, known data augmentation is not always effective. Specifically, in the field of determining the presence or absence of defects in concrete, known data augmentation is not always effective.

[0006] This disclosure aims to provide a data augmentation method, program, and system that can generate multiple augmented data sets used to generate a trained model for determining the presence or absence of defects in concrete, as well as a trained model generated using the multiple augmented data sets, and an inspection method using this trained model. [Means for solving the problem]

[0007] A data augmentation method according to one aspect of this disclosure generates a plurality of augmented data using a plurality of measured data. The plurality of measured data are data based on a plurality of detection signals output from a laser detection device. The laser detection device comprises a pulsed laser, a vibration detection laser, and a photodetector. The pulsed laser emits a first laser beam for vibrating the surface of concrete. The vibration detection laser emits a second laser beam for detecting vibrations of the surface of the concrete. The photodetector detects the second laser beam reflected from the surface of the concrete and converts it into a detection signal. The data augmentation method comprises an acquisition step and a generation step. In the acquisition step, the plurality of measured data are acquired. In the generation step, the plurality of augmented data are generated. In the generation step, data augmentation processing is performed. The data augmentation processing is performed on the speed of sound inside the concrete. 、 The intensity of the first laser light irradiated from the pulsed laser 、 The intensity of the second laser light emitted from the vibration detection laser 、 Sensitivity of the aforementioned photodetector 、 Noise generated by the laser detection device 、and the distance between the laser detection device and the concrete. Each element at least one of the following When each element is simulated to change within its range of variation, the obtained values ​​are... By changing the aforementioned multiple measured data Multiple data are requested, and the requested multiple data are obtained In addition to the aforementioned multiple measured data, the aforementioned multiple extended data as This is the process of generating something.

[0008] A program according to one aspect of this disclosure is a program that causes one or more processors of a computer system to execute the data augmentation method.

[0009] A trained model according to one aspect of this disclosure is generated by machine learning using the plurality of augmented data generated by the data augmentation method as training data. It is a pre-trained model. The aforementioned trained model is Computers, Based on the input data corresponding to the detection signal, the system outputs an inspection result regarding the presence or absence of defects in the concrete. Make it work .

[0010] An inspection method according to one aspect of the present disclosure includes an inference step. In the inference step, input data corresponding to the detection signal is input to a trained model to obtain the inspection result regarding the presence or absence of defects in the concrete.

[0011] A data augmentation system according to one aspect of this disclosure generates augmented data using a plurality of measured data. The plurality of measured data are data based on a plurality of detection signals output from a laser detection device. The laser detection device comprises a pulsed laser, a vibration detection laser, and a photodetector. The pulsed laser emits a first laser beam for vibrating the surface of concrete. The vibration detection laser emits a second laser beam for detecting vibrations of the surface of the concrete. The photodetector detects the second laser beam reflected from the surface of the concrete and converts it into a detection signal. The data augmentation system comprises an acquisition unit and a generation unit. The acquisition unit acquires the plurality of measured data. The generation unit generates the plurality of augmented data. The generation unit performs data augmentation processing. The data augmentation processing involves the speed of sound inside the concrete. 、 The intensity of the first laser light irradiated from the pulsed laser 、 The intensity of the second laser light emitted from the vibration detection laser 、 Sensitivity of the aforementioned photodetector 、 Noise generated by the laser detection device 、 and the distance between the laser detection device and the concrete. Each element at least one of the following When each element is simulated to change within its range of variation, the obtained values ​​are... By changing the aforementioned multiple measured data Multiple data are requested, and the requested multiple data are obtained In addition to the aforementioned multiple measured data, the aforementioned multiple extended data as This is the process of generating something. [Effects of the Invention]

[0012] This disclosure has the advantage that multiple augmented data sets used to generate a trained model for determining the presence or absence of defects in concrete can be generated through data augmentation. [Brief explanation of the drawing]

[0013] [Figure 1] Figure 1 is a block diagram of a data augmentation system and related components according to one embodiment. [Figure 2]FIG. 2 is a flowchart showing a data expansion method using the same data expansion system as described above. [Figure 3] FIGS. 3A and 3B are waveform diagrams of detection signals input to the same data expansion system as described above. [Figure 4] FIGS. 4A and 4B are diagrams showing measured data obtained in the same data expansion system as described above. [Figure 5] FIGS. 5A to 5F are diagrams showing expanded data obtained in the same data expansion system as described above. [Figure 6] FIGS. 6A to 6F are diagrams showing expanded data obtained in the same data expansion system as described above. [Figure 7] FIGS. 7A to 7H are distribution diagrams of a plurality of measured data obtained in the same data expansion system as described above. [Figure 8] FIGS. 8A to 8H are distribution diagrams of a plurality of measured data obtained in the same data expansion system as described above. [Figure 9] FIGS. 9A to 9D are distribution diagrams of a plurality of measured data obtained in the same data expansion system as described above.

BEST MODE FOR CARRYING OUT THE INVENTION

[0014] (Embodiment) Hereinafter, a data expansion method, a program, a learned model 4, an inspection method, and a data expansion system 1 according to an embodiment will be described with reference to the drawings. However, the following embodiment is only one of various embodiments of the present disclosure. The following embodiment can be variously modified according to design and the like as long as the object of the present disclosure can be achieved. Also, each figure described in the following embodiment is a schematic figure, and the ratio of the size and thickness of each component in the figure does not necessarily reflect the actual dimensional ratio.

[0015] (Overview) FIG. 1 shows the configuration of the data expansion system 1 of the present embodiment. The data expansion system 1 is used together with a laser detection device 2 and a learning device 3.

[0016] The laser detection device 2 emits a second laser beam L2, detects the second laser beam L2 reflected by the concrete C1, and outputs a detection signal D1 indicating the detection result. More specifically, the detection signal D1 is a waveform signal output from the photodetector 23 of the laser detection device 2. The data augmentation system 1 generates actual measurement data based on the detection signal D1. The actual measurement data is, for example, data generated by performing a time-frequency analysis on the detection signal D1. The data augmentation system 1 generates augmented data based on the actual measurement data. Furthermore, the data augmentation system 1 outputs multiple training data T1, which includes multiple actual measurement data and multiple augmented data, to the learner 3. The learner 3 performs machine learning using the multiple training data T1 and generates a trained model 4. The trained model 4 takes data based on the output signal D2 output from the photodetector 23 as input and outputs a result determining whether or not there are defects in the concrete C1.

[0017] In this context, the output signal D2 that forms the basis of the data input to the trained model 4 is a signal newly detected for determining the presence or absence of defects in concrete C1, separate from the detection signal D1 detected for generating the training data T1. However, the output signal D2 is detected in the same manner as the detection signal D1 detected for generating the training data T1.

[0018] The data augmentation system 1 is particularly suitable for processing detection signals D1 for box culvert concrete or unreinforced concrete. Concrete C1 constitutes, for example, the inner walls of tunnels, bridges, or buildings. Defects in concrete C1 include, for example, voids or cracks present within the concrete C1.

[0019] The pre-trained model 4 referred to here may include, for example, a model that uses machine learning. The pre-trained model 4 can be implemented, for example, by implementing the pre-trained model on an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or FPGA (Field-Programmable Gate Array).

[0020] The data augmentation system 1 of this embodiment generates multiple augmented data using (based on) multiple measured data. The multiple measured data are data based on multiple detection signals D1 output from the laser detection device 2. The laser detection device 2 includes a pulsed laser 21, a vibration detection laser 22, and a photodetector 23. The pulsed laser 21 emits a first laser beam L1 to vibrate the surface of the concrete C1. The vibration detection laser 22 emits a second laser beam L2 to detect vibrations on the surface of the concrete C1. The photodetector 23 detects the second laser beam L2 reflected from the surface of the concrete C1 and converts it into a detection signal D1. The data augmentation system 1 includes an acquisition unit 132 and a generation unit 133. The acquisition unit 132 acquires multiple measured data. The generation unit 133 generates multiple augmented data. The generation unit 133 performs data augmentation processing. The data augmentation process generates multiple augmented data based on at least one of the following: information on the variation in sound velocity within the concrete C1, information on the variation in intensity of the first laser beam L1 emitted from the pulsed laser 21, information on the variation in intensity of the second laser beam L2 emitted from the vibration detection laser 22, information on the variation in sensitivity of the photodetector 23, information on the variation in noise generated in the laser detection device 2, and information on the distance between the laser detection device 2 and the concrete C1, along with multiple measured data.

[0021] According to this embodiment, the data augmentation system 1 can generate multiple augmentation data sets. Based on the multiple training data sets T1 which include the multiple augmentation data sets, the learner 3 performs machine learning to generate a trained model 4 suitable for determining the presence or absence of defects in the concrete C1.

[0022] Furthermore, functions similar to those of the data augmentation system 1 can be realized by a data augmentation method. In the data augmentation method of this embodiment, multiple augmented data are generated using (based on) multiple measured data. The multiple measured data are data based on multiple detection signals D1 output from the laser detection device 2. The laser detection device 2 includes a pulsed laser 21, a vibration detection laser 22, and a photodetector 23. The pulsed laser 21 emits a first laser beam L1 to vibrate the surface of the concrete C1. The vibration detection laser 22 emits a second laser beam L2 to detect vibrations on the surface of the concrete C1. The photodetector 23 detects the second laser beam L2 reflected from the surface of the concrete C1 and converts it into a detection signal D1. The data augmentation method has an acquisition step and a generation step. In the acquisition step, multiple measured data are acquired. In the generation step, multiple augmented data are generated. In the generation step, data augmentation processing is performed. The data augmentation process generates multiple augmented data based on at least one of the following: information on the variation in sound velocity within the concrete C1, information on the variation in intensity of the first laser beam L1 emitted from the pulsed laser 21, information on the variation in intensity of the second laser beam L2 emitted from the vibration detection laser 22, information on the variation in sensitivity of the photodetector 23, information on the variation in noise generated in the laser detection device 2, and information on the distance between the laser detection device 2 and the concrete C1, along with multiple measured data.

[0023] Furthermore, the data augmentation method can be implemented in a program. The program in this embodiment is a program that causes one or more processors of a computer system to execute the data augmentation method. The program may be recorded on a non-temporary recording medium that is readable by the computer system.

[0024] Furthermore, some of the processing in the data augmentation method may be performed by a human. For example, in the process of generating augmented data by changing the parameters that constitute the measured data, a human may determine the amount of change in some of the parameters. In this embodiment, unless otherwise specified, all processing in the data augmentation method will be described as being performed by a data augmentation system 1, which mainly consists of a computer system.

[0025] The trained model 4 in this embodiment is generated by machine learning using multiple augmented data generated by the data augmentation method as training data T1. The trained model 4 outputs an inspection result regarding the presence or absence of defects in concrete C1 based on input data corresponding to the detection signal D1 (data based on the output signal D2).

[0026] The inspection method of this embodiment includes an inference step. In the inference step, input data corresponding to the detection signal D1 (data based on the output signal D2) is input to the trained model 4 to obtain an inspection result regarding the presence or absence of defects in the concrete C1.

[0027] Figure 2 schematically shows the overall flow of the data augmentation method of this embodiment and the process of performing machine learning using multiple training data T1, which includes multiple augmented data generated by the data augmentation method. However, the flowchart in Figure 2 is merely an example, and the order of processing may be changed as appropriate, or processing may be added or omitted as appropriate.

[0028] First, the data augmentation system 1 acquires the detection signal D1 from the laser detection device 2 (step ST1). Next, the data augmentation system 1 extracts components below the upper frequency limit from the detection signal D1 (step ST2). This removes high-frequency components that are not useful in machine learning from the detection signal D1. Furthermore, the data augmentation system 1 generates actual measurement data based on the detection signal D1 from which the high-frequency components were removed in step ST2 (step ST3). In other words, the data augmentation system 1 converts the detection signal D1 into actual measurement data in a format suitable for machine learning.

[0029] Next, the data augmentation system 1 generates augmented data based on the measured data (step ST4). Furthermore, the data augmentation system 1 outputs the training data T1 (measured data and augmented data) to the learner 3 (step ST5).

[0030] If the number of training data T1 input to the learning device 3 is below the threshold (step ST6: No), steps ST1 to ST5 are repeated, and new measured data and augmented data are generated.

[0031] If the number of training data T1 input to the learner 3 is greater than the threshold (Step ST6: Yes), the learner 3 performs machine learning based on the multiple training data T1 and generates a trained model 4 (Step ST7). The learner 3 outputs the generated trained model 4 (Step ST8).

[0032] (detail) (1) Judgment System The configuration of this embodiment will be described in more detail below.

[0033] As shown in Figure 1, the determination system 100 includes a data augmentation system 1, a laser detection device 2, a learner 3, and a trained model 4.

[0034] (2) Laser detection device The laser detection device 2 includes, for example, a pulsed laser 21, a vibration detection laser 22, a photodetector 23, and an optical system 24.

[0035] The optical system 24 includes, for example, a mirror group 26, a half mirror 25a, a reflective mirror 25b, a half mirror 25c, and a half mirror 25d, as shown in Figure 1. The optical system 24 may also include lenses.

[0036] The mirror group 26 comprises mirrors 26b, 26a, and 26c. Mirror 26a is a dichroic mirror. The mirror group 26 is configured to adjust the irradiation position of the first laser beam L1 and the second laser beam L2.

[0037] The pulsed laser 21 emits a first laser beam L1. The first laser beam L1 is, for example, a shock wave excitation laser beam used to vibrate the surface of the concrete C1, which is the object to be inspected. The first laser beam L1 is a pulsed laser beam. As the pulsed laser 21, for example, an Nd:YAG laser (fundamental wave) that outputs a pulsed laser beam with a pulse width of 10 ns can be used, but is not limited to this.

[0038] The vibration detection laser 22 emits a second laser beam L2. The second laser beam L2 is used to detect vibrations on the surface of the concrete C1, which is the object to be inspected. In the laser detection device 2, the reflected light of the second laser beam L2 from the concrete C1 is used as the probe beam L21. The vibration detection laser 22 includes, for example, a CW laser. As the vibration detection laser 22, for example, an Nd:YAG laser (2nd harmonic) that continuously outputs laser light with a wavelength of 532 nm can be used, but is not limited to this.

[0039] The optical system 24 is configured to form a first optical path, a second optical path, a third optical path, and a fourth optical path. The first optical path guides the first laser beam L1 emitted from the pulsed laser 21 to the surface of the concrete C1. The second optical path guides the second laser beam L2 emitted from the vibration detection laser 22 to the surface of the concrete C1. The third optical path guides the probe light L21, which is the reflected light of the second laser beam L2 irradiated onto the surface of the concrete C1, to the photodetector 23. The fourth optical path guides a portion of the second laser beam L2 emitted from the vibration detection laser 22 as reference light L22 to the photodetector 23.

[0040] The first optical path is formed by mirror 26a and mirror 26b.

[0041] The second optical path is formed by mirror 26c and mirror 26b.

[0042] The third optical path is formed by mirror 26b, mirror 26c, half mirror 25a, and reflection mirror 25b.

[0043] The fourth optical path is formed by half mirror 25c and half mirror 25d.

[0044] In FIG. 1, the first laser beam L1 is schematically shown by a solid line, and the second laser beam L2, the probe beam L21, and the reference beam L22 are schematically shown by a dashed line.

[0045] The first laser beam L1 emitted from the pulse laser 21 is reflected by the mirrors 26a and 26b and irradiates the surface of the concrete C1.

[0046] The second laser beam L2 emitted from the vibration detection laser 22 passes through the half mirrors 25c and 25a, is reflected by the mirror 26c, passes through the mirror 26a, and then is reflected by the mirror 26b and irradiates the surface of the concrete C1.

[0047] The probe beam L21, which is the reflected light of the second laser beam L2 irradiated on the surface of the concrete C1, is reflected by the mirror 26b, passes through the mirror 26a, and then is reflected by the mirror 26c toward the half mirror 25a. Thereafter, the probe beam L21 is reflected by the half mirror 25a and the reflection mirror 25b, passes through the half mirror 25d, and irradiates the photodetector 23.

[0048] A part of the second laser beam L2 emitted from the vibration detection laser 22 is reflected by the half mirrors 25c and 25d and irradiates the photodetector 23 as the reference beam L22.

[0049] The photodetector 23 detects the probe beam L21 and the reference beam L22 formed by the second laser beam L2. More specifically, the photodetector 23 detects the interference fringes between the probe beam L21 and the reference beam L22. The photodetector 23 outputs a waveform representing the temporal change in the detection intensity of the interference fringes as a detection signal D1.

[0050] Figure 3A shows an example of the detection signal D1 obtained when the first laser beam L1 and the second laser beam L2 are irradiated onto concrete C1 containing defects. Figure 3B shows an example of the detection signal D1 obtained when the first laser beam L1 and the second laser beam L2 are irradiated onto concrete C1 without defects.

[0051] Figures 3A and 3B each show the components (intensity) of all frequencies in a single diagram. When a defect is present, the intensity of the detected signal D1 increases at certain frequencies compared to when the defect is absent. The components for each frequency are shown in Figures 4A and 4B, which will be discussed later.

[0052] (3) Data expansion system (3.1) Configuration As shown in Figure 1, the data expansion system 1 comprises a communication unit 11, a storage unit 12, and a processing unit 13.

[0053] The communication unit 11 includes a communication interface device. The data expansion system 1 can communicate with the laser detection device 2 and the learner 3 via the communication interface device. In this disclosure, "communication possible" means that signals can be sent and received directly or indirectly via a network or repeater, etc., by an appropriate communication method such as wired communication or wireless communication.

[0054] The storage unit 12 is a storage device composed of a hard disk drive (HDD) or a solid-state drive (SSD), etc. The storage unit 12 stores information. For example, the storage unit 12 stores a detection signal D1 and a plurality of learning data T1. The storage unit 12 also stores, for example, a (computer) program executed by the processing unit 13.

[0055] The processing unit 13 includes a computer system having one or more processors and memory. At least some of the functions of the processing unit 13 are realized when the processor of the computer system executes a program recorded in the memory of the computer system. The program may be recorded in memory, provided via a telecommunication line such as the Internet, or provided on a non-temporary recording medium such as a memory card.

[0056] The processing unit 13 includes a signal processing unit 131, an acquisition unit 132, a generation unit 133, and a data division unit 134. Note that these merely represent the functions implemented by the processing unit 13 and do not necessarily represent an actual physical configuration.

[0057] The signal processing unit 131 extracts components below the upper frequency limit from the detected signal D1 and discards components above the upper frequency limit. The method for determining the upper frequency limit will be described later.

[0058] The signal processing unit 131 generates measured data based on the detected signal D1 (more specifically, the components of the detected signal D1 below the upper frequency limit). More specifically, the signal processing unit 131 generates one measured data from one detected signal D1. For example, the signal processing unit 131 generates measured data by performing a time-frequency analysis on the detected signal D1.

[0059] By performing a time-frequency analysis on the detection signal D1 in Figure 3A, the measured data shown in Figure 4A is obtained. Similarly, by performing a time-frequency analysis on the detection signal D1 in Figure 3B, the measured data shown in Figure 4B is obtained. In Figures 4A and 4B, points that are closer to white in color have higher intensity. In Figures 4A and 4B, the data for the period in which output corresponding to the presence or absence of defects in concrete C1 is obtained is enclosed in a rectangular white frame. The period outside the white frame, i.e., the period from 0 seconds to approximately 0.01 seconds, is a silent period in which the reflected light of the second laser beam L2 is not detected.

[0060] The acquisition unit 132 acquires the measured data from the signal processing unit 131 and provides it to the generation unit 133.

[0061] The generation unit 133 generates extended data based on measured data. More specifically, the generation unit 133 generates one or more extended data from one measured data. In this embodiment, it will be explained that the generation unit 133 generates two or more extended data from one measured data.

[0062] The data splitting unit 134 divides the multiple augmented data generated by the data augmentation process into a training data set, a validation data set, and a test data set. Further details will be described later.

[0063] (3.2) Data augmentation process There are multiple factors that determine the waveform of the detected signal D1. One of these factors is, for example, the intensity of the first laser beam L1.

[0064] When a certain element changes by a certain amount, the extent to which the waveform of the detection signal D1 changes, and consequently the extent to which the measured data based on the detection signal D1 changes, can be known (predicted) empirically based on experimental results. The memory unit 12 stores the correspondence between the changes in each element and the changes in the measured data.

[0065] Furthermore, the extent to which each element can realistically change (range of change or variability) can be determined empirically based on experimental results. Alternatively, the range of change for each element may be determined as appropriate by the user. The memory unit 12 stores the range of change for each element.

[0066] The generation unit 133 randomly determines a change in at least one of the multiple elements within the range of change stored in the storage unit 12, and determines the change in the measured data provided by the acquisition unit 132 based on the above correspondence stored in the storage unit 12. This allows the generation unit 133 to obtain measured data when at least one element changes randomly. The measured data obtained in this way is, in other words, the extended data generated by the generation unit 133. That is, the generation unit 133 obtains measured data when at least one element changes randomly as extended data. The generation unit 133 can, for example, use random numbers (pseudo-random numbers) to simulate a situation where at least one element changes randomly.

[0067] Next, Figures 5A to 6F will be explained. Here, components below the upper frequency limit were extracted from the detection signal D1 obtained by irradiating concrete C1 with the first laser beam L1 and the second laser beam L2, and multiple such detection signals D1 were prepared. Furthermore, multiple measurement data were obtained based on the multiple detection signals D1, and the multiple measurement data were classified into "defective measurement data" and "normal measurement data". Defective measurement data is measurement data based on the detection signal D1 obtained when concrete C1 with defects is irradiated with the first laser beam L1 and the second laser beam L2. Normal measurement data is measurement data based on the detection signal D1 obtained when concrete C1 without defects is irradiated with the first laser beam L1 and the second laser beam L2. Based on these multiple measurement data, multiple extended data were obtained. The extended data obtained in this way is shown in Figures 5A to 6F.

[0068] Figures 5A to 5C show extended data obtained based on the same actual defect measurement data (S1). Figures 5D to 5F show extended data obtained based on the same actual defect measurement data (S2). The actual defect measurement data (S2) is different from the actual defect measurement data (S1).

[0069] Figures 6A to 6C show extended data obtained based on the same normal measurement data (S3). Figures 6D to 6F show extended data obtained based on the same normal measurement data (S4). Normal measurement data (S4) is different from normal measurement data (S3).

[0070] (3.3) Multiple elements The elements that determine the waveform and measured data of the detected signal D1 may include the following elements. The generation unit 133 obtains the measured data as extended data when at least one of the following elements changes randomly. That is, the generation unit 133 obtains the extended data based on information about at least one of the following elements.

[0071] One element is the speed of sound inside the concrete C1. The speed of sound inside the concrete C1 changes depending on the presence or absence of defects in the concrete C1, as well as conditions such as the composition of the concrete C1, its water content, and its temperature. If defects are present, the strength of the concrete C1 decreases, and the speed of sound also decreases. Furthermore, if defects are present, the measured data also changes compared to the case where no defects are present. The range of change (variation) in the speed of sound is predetermined to some extent and can be determined experimentally. Similarly, the change in measured data accompanying the change in the speed of sound can also be determined experimentally. Therefore, using the speed of sound as one element, the generation unit 133 can obtain measured data when the speed of sound changes as extended data. In other words, the generation unit 133 can obtain extended data based on information regarding the variation in the speed of sound.

[0072] Another element is the intensity of the first laser beam L1 emitted from the pulsed laser 21. When the intensity of the first laser beam L1 changes, the measured data also changes. The range of change (variation) in the intensity of the first laser beam L1 is fixed to some extent, for example, about 2%. Furthermore, the change in the measured data associated with a change in the intensity of the first laser beam L1 can be determined experimentally. Therefore, using the intensity of the first laser beam L1 as one element, the generation unit 133 can obtain the measured data when the intensity of the first laser beam L1 changes as extended data.

[0073] Another element is the intensity of the second laser beam L2 emitted from the vibration detection laser 22. Generally, a higher-precision laser is used as the vibration detection laser 22 compared to the pulsed laser 21, and the variation in the intensity of the second laser beam L2 is, for example, about 1%. Similar to the case where the intensity of the first laser beam L1 is one element, the generation unit 133 can use the intensity of the second laser beam L2 as one element to obtain measured data as extended data when the intensity of the second laser beam L2 changes.

[0074] Another factor is the sensitivity of the photodetector 23. When the sensitivity of the photodetector 23 changes, the measured data also changes. The photodetector 23 includes a light-receiving element, and the output current of the light-receiving element generates noise equivalent to the square root of the current value. Therefore, the range of change (variation) in the sensitivity of the photodetector 23 is equivalent to the range of the square root of the current value. Furthermore, the change in the measured data accompanying the change in the sensitivity of the photodetector 23 can be determined experimentally. Thus, using the sensitivity of the photodetector 23 as one factor, the generation unit 133 can obtain the measured data when the sensitivity of the photodetector 23 changes as extended data.

[0075] Furthermore, the information regarding the variation in the sensitivity of the photodetector 23, as information regarding one element, may also be information regarding the variation in sensitivity for each frequency. In other words, the generation unit 133 may obtain extended data based on the information regarding the variation in sensitivity for each frequency.

[0076] Another factor is the noise generated by the laser detection device 2. This factor is represented, for example, by the signal-to-noise ratio (S / N ratio) of the second laser beam L2. The S / N ratio of the second laser beam L2 changes depending on the inspection environment. For example, if the inspection environment is inside a tunnel, the noise due to reflection will be greater compared to when it is outside the tunnel. Therefore, the S / N ratio of the second laser beam L2 changes depending on whether the inspection environment is inside or outside the tunnel. Also, for example, the S / N ratio of the second laser beam L2 changes depending on the distance between the laser detection device 2 and the concrete C1. Furthermore, the reflectivity of the concrete C1 changes depending on the moisture content and the degree of surface contamination, so the S / N ratio of the second laser beam L2 reflected by the concrete C1 changes. In addition, the S / N ratio of the second laser beam L2 reflected by the concrete C1 changes depending on the irradiation angle of the second laser beam L2. The range of change (variation) of the S / N ratio can be determined based on the inspection environment, etc. Furthermore, the changes in measured data associated with changes in the signal-to-noise ratio (S / N ratio) can be determined experimentally. Therefore, using the S / N ratio (noise) as one element, the generation unit 133 can obtain the measured data when the S / N ratio (noise) changes as extended data.

[0077] Another factor is the distance (irradiation distance) between the laser detection device 2 and the concrete C1. Note that information regarding this factor may overlap with information regarding the variation in the signal-to-noise ratio of the second laser beam L2. As the irradiation distance increases, the second laser beam L2 attenuates, and therefore the measured data also changes. Generally, the intensity of the second laser beam L2 is proportional to the -2 power of the irradiation distance. Furthermore, as the irradiation distance increases, the focused beam diameter of the second laser beam L2 increases, and the measured data changes. If w is the focused beam diameter, D is the beam diameter before focusing, f is the focal length of the lens in the optical system 24 (see Figure 1), and λ is the beam wavelength, then the relationship w = fλ / D holds. Note that the irradiation distance changes depending on the irradiation angle of the second laser beam L2. The range of change (variation) in the irradiation distance can be determined based on the inspection environment, etc. Furthermore, the change in measured data associated with changes in irradiation distance can be determined experimentally. Therefore, using the irradiation distance as one element, the generation unit 133 can obtain measured data as extended data when the irradiation distance changes.

[0078] Information regarding the range of change for each element (for example, the variation in the sensitivity of the photodetector 23) and the changes in measured data associated with the changes in the elements is pre-stored in the storage unit 12. This information may be acquired, for example, by the communication unit 11 of the data expansion system 1, or it may be input by user operation on the input interface provided by the data expansion system 1.

[0079] (3.4) Labels Each measured data point is labeled as either defective or not defective. The labels may be assigned, for example, by the user. Specifically, information indicating the labels may be acquired by the communication unit 11 of the data augmentation system 1, or it may be input by the user through an operation on the input interface provided by the data augmentation system 1.

[0080] Furthermore, as explained in the next section "(3.5) Upper Frequency Limit," multiple measured data are classified into defective measured data and normal measured data using a classification algorithm. Based on this result, the data augmentation system 1 may label the defective measured data as defective and the normal measured data as not defective.

[0081] Extended data generated from measured data labeled as defective will be labeled as defective. Extended data generated from measured data labeled as free of defects will be labeled as free of defects.

[0082] (3.5) Upper limit frequency The signal processing unit 131 extracts components below the upper frequency limit from the detected signal D1 and discards components above the upper frequency limit. As a result, the measured data generated based on the detected signal D1, and the extended data generated based on the measured data, will contain components below the upper frequency limit and not components above the upper frequency limit. In this way, the signal processing unit 131 performs an extraction step to extract components below the upper frequency limit from each of the multiple measured data. In the data extension process, the generation unit 133 generates multiple extended data based on the components below the upper frequency limit from each of the multiple measured data.

[0083] Figures 7A to 9D show the results of verifying whether the classification algorithm could classify multiple measured data generated from multiple detection signals D1 when the upper frequency limit was varied in various ways. From these results, it is preferable that the upper frequency limit be between 40kHz and 60kHz. That is, by setting the upper frequency limit to between 40kHz and 60kHz and extracting components below the upper frequency limit from the detection signal D1, the training data T1 (measured data and augmented data) becomes data suitable for machine learning.

[0084] Figures 7A to 9D will be explained in detail below.

[0085] Figures 7A to 9D show the distribution of numerous measured data, including defective and normal measured data, after classifying them using a classification algorithm. By reducing the dimensionality of the multidimensional measured data, which includes intensity information for each frequency, Figures 7A to 9D represent the measured data as two-dimensional data consisting of features represented on the vertical axis and features represented on the horizontal axis. White circles indicate defective measured data, and black circles indicate normal measured data.

[0086] Figures 7A-7D, 8A-8D, and 9A-9B each show the results of classifying a large number of observed data using PCA (Principal Component Analysis) as the classification algorithm through unsupervised learning. Figures 7E-7H, 8E-8H, and 9C-9D each show the results of classifying a large number of observed data using t-SNE (t-distributed Stochastic Neighbor Embedding) as the classification algorithm through unsupervised learning.

[0087] Figures 7A and 7E show the results when the upper frequency limit is set to 10 kHz. Figures 7B and 7F show the results when the upper frequency limit is set to 20.161 kHz. Figures 7C and 7G show the results when the upper frequency limit is set to 29.762 kHz. Figures 7D and 7H show the results when the upper frequency limit is set to 40.323 kHz.

[0088] Figures 8A and 8E show the results when the upper frequency limit is set to 44.643 kHz. Figures 8B and 8F show the results when the upper frequency limit is set to 50 kHz. Figures 8C and 8G show the results when the upper frequency limit is set to 54.348 kHz. Figures 8D and 8H show the results when the upper frequency limit is set to 59.524 kHz.

[0089] Figures 9A and 9C show the results when the upper frequency limit is set to 69.444 kHz. Figures 9B and 9D show the results when the upper frequency limit is set to 78.125 kHz.

[0090] When the upper frequency limit is 10kHz, 20.161kHz, and 29.762kHz, defective and normal measured data are not sufficiently separated. When the upper frequency limit is 40.323kHz and 44.643kHz, defective and normal measured data are somewhat separated. When the upper frequency limit is 50kHz, defective and normal measured data are clearly separated. When the upper frequency limit is 54.348kHz and 59.524kHz, defective and normal measured data are somewhat separated. When the upper frequency limit is 69.444kHz and 78.125kHz, defective and normal measured data are not sufficiently separated.

[0091] Thus, when the upper frequency limit is approximately 40-60 kHz, the measured data for defects and the measured data for normal conditions are relatively well separated. Therefore, it is preferable that the upper frequency limit be between 40 kHz and 60 kHz.

[0092] (3.6) Data partitioning The data partitioning unit 134 of the data augmentation system 1 executes a data partitioning step. The data partitioning step is the process of dividing (classifying) multiple augmented data into a training data group, a validation data group, and a test data group.

[0093] The data splitting step includes a first data splitting step and a second data splitting step. The second data splitting step is the process of splitting (classifying) multiple augmentation data into a training data set, a validation data set, and a test data set, using a different splitting (classification) method than that of the first data splitting step.

[0094] In the following, data belonging to the training dataset will be referred to as training data, data belonging to the validation dataset will be referred to as validation data, and data belonging to the test dataset will be referred to as test data. That is, in both the first and second data splitting steps, the training data T1 is classified into training data, validation data, and test data. This is the classification used in the holdout method of machine learning.

[0095] The training data is used in the so-called learning phase of machine learning, and a trained model 4 is built based on the training data. Subsequently, the hyperparameters of the trained model 4 are tuned based on validation data. Finally, the performance of the trained model 4 is evaluated based on test data.

[0096] When focusing on a single set of augmented data, the classification in the first data splitting step may differ from the classification in the second data splitting step. For example, augmented data classified as training data in the first data splitting step may be classified as validation data or test data in the second data splitting step.

[0097] In the first data splitting step, three or more augmented data sets generated based on one actual data set are selected: augmented data belonging to the training data set (training data), augmented data belonging to the validation data set (validation data), and augmented data belonging to the test data set (test data).

[0098] In the second data splitting step, the augmented data generated based on the first set of measured data is assigned to the training data set. Furthermore, in the second data splitting step, the augmented data generated based on the second set of measured data is assigned to the validation data set. Finally, in the second data splitting step, the augmented data generated based on the third set of measured data is assigned to the test data set.

[0099] Table 1 below shows an example of splitting (classifying) multiple extended data sets using the first data splitting step. Table 2 below shows an example of splitting (classifying) multiple extended data sets using the second data splitting step.

[0100] [Table 1]

[0101] [Table 2]

[0102] Tables 1 and 2 distinguish between multiple measured data points and multiple extended data points by assigning them numbers. Data points with the same number are the same data.

[0103] The label "nd" (no defect) means that the measured data located in the same row is normal measured data, and that each extended data located in the same row is extended data generated based on that normal measured data.

[0104] The label "d" (defect) indicates that the measured data located in the same row is defective measured data, and that each extended data located in the same row is extended data generated based on that defective measured data.

[0105] As shown in [Table 1] and [Table 2], each of the training data set, validation data set, and test data set includes augmented data generated based on normal measured data and augmented data generated based on defective measured data.

[0106] In the example shown in [Table 1], focusing on any single measured data point, four of the eight augmented data points generated based on that data point belong to the training data group, two belong to the validation data group, and two belong to the test data group. Furthermore, multiple measured data points are appropriately divided into training, validation, and test data sets.

[0107] In the example shown in [Table 2], the multiple measured data are classified into three groups: the first measured data (numbers 1-4, 9-12) belonging to the training data group (corresponding); the second measured data (numbers 5-6, 13-14) belonging to the validation data group (corresponding); and the third measured data (numbers 7-8, 15-16) belonging to the test data group (corresponding). The augmented data generated based on each measured data is classified (belongs to) the group corresponding to the original measured data.

[0108] Note that Tables 1 and 2 are merely examples, and the number of training data, validation data, and test data are not particularly limited.

[0109] In the second data splitting step shown in [Table 2], the training data, validation data, and test data are each generated based on different actual measurement data. Therefore, compared to the first data splitting step shown in [Table 1], it is expected that a training dataset closer to the training dataset obtained from actual measurements (a training dataset consisting only of actual measurement data) is obtained.

[0110] The data augmentation system 1 outputs to the learner 3 information regarding the first data partition determined in the first data partitioning step and information regarding the second data partition determined in the second data partitioning step. The information regarding the first data partitioning and the information regarding the second data partitioning are information that divides multiple (m) measured data and multiple (m × n) augmented data into training data, validation data, and test data, respectively.

[0111] In the following, the multiple measured data and multiple augmented data sets that have been divided into training data, validation data, and test data by the first data splitting step will be referred to as the training dataset after the first data splitting. Similarly, in the following, the multiple measured data and multiple augmented data sets that have been divided into training data, validation data, and test data by the second data splitting step will be referred to as the training dataset after the second data splitting.

[0112] The learner 3 generates a first trained model as trained model 4 based on the training dataset that has undergone the first data partitioning. The learner 3 also generates a second trained model as trained model 4 based on the training dataset that has undergone the second data partitioning. Depending on the type of defect and conditions such as the composition of concrete C1, the suitability of the first or second trained model for the determination will vary. The user should select the trained model that is more suitable for the determination between the first and second trained models.

[0113] Furthermore, the user may, for example, determine that neither the first nor the second trained model is suitable for determining the presence or absence of defects if the desired accuracy cannot be obtained in at least one of the first or second trained models. In such cases, the user can simply re-select the models and generate new first and second trained models.

[0114] If trained model 4 is the first trained model, in the inference phase, a first determination result regarding the presence or absence of defects in the concrete C1 being measured is obtained by inputting data based on the output signal D2 into the first trained model. The data based on the output signal D2 is generated in the same way as the process by which measured data is generated based on the detection signal D1. More specifically, the data based on the output signal D2 is generated by performing time-frequency analysis and removing components below the upper frequency limit on the output signal D2.

[0115] Alternatively, if trained model 4 is the second trained model, in the inference phase, a second determination result regarding the presence or absence of defects in the concrete C1 being measured is obtained by inputting data based on the output signal D2 into the second trained model. Alternatively, both the first and second determination results may be obtained based on a single output signal D2. In other words, if trained model 4 is both the first and second trained models, both the first and second determination results can be obtained.

[0116] Furthermore, the first trained model and the second trained model may each utilize different learning methods. For example, the first trained model may be a model using a neural network, while the second trained model may be a model trained on a different model than a neural network.

[0117] As the pre-trained model 4, a model using a neural network is preferable, for example.

[0118] (Modified version of the embodiment) The following are examples of modifications of the embodiment. These modifications may be implemented by combining them as appropriate. Hereafter, the embodiments described above will be referred to as the "basic example".

[0119] The learner 3 may generate the trained model 4 through supervised learning, or it may generate the trained model 4 through unsupervised learning.

[0120] In the basic example, the trained model 4 generated by the learner 3 outputs a result indicating whether or not there are defects in concrete C1. However, the trained model 4 may also output information about the type of defect if there is a defect, in addition to the result indicating whether or not there are defects in concrete C1.

[0121] The trained model 4 may take the output signal D2 as input data corresponding to the detection signal D1 and output a result of determining whether or not there are defects in the concrete C1.

[0122] The learner 3 may update the trained model 4 by performing additional training (retraining).

[0123] In the basic example, the signal processing unit 131, which generates measured data based on the detection signal D1, is part of the data expansion system 1. However, the signal processing unit 131 may be an external component of the data expansion system 1.

[0124] The photodetector 23 is not limited to an instrument that detects interference fringes, but may be, for example, a laser Doppler vibrometer.

[0125] The entity executing the data augmentation system 1 or data augmentation method in this disclosure includes a computer system. The computer system mainly consists of a processor and memory as hardware. At least part of the functions of the entity executing the data augmentation system 1 or data augmentation method in this disclosure are realized by the processor executing a program recorded in the computer system's memory. The program may be pre-recorded in the computer system's memory, provided via a telecommunications line, or provided on a non-temporary recording medium such as a memory card, optical disk, or hard disk drive that is readable by the computer system. The processor of the computer system consists of one or more electronic circuits including semiconductor integrated circuits (ICs) or large-scale integrated circuits (LSIs). The integrated circuits referred to here, such as ICs or LSIs, are named differently depending on the degree of integration and include integrated circuits called system LSIs, VLSIs (Very Large Scale Integration), or ULSIs (Ultra Large Scale Integration). Furthermore, FPGAs (Field-Programmable Gate Arrays) that are programmed after the manufacture of LSIs, or logic devices that allow for the reconfiguration of junction relationships or circuit compartments within LSIs, can also be used as processors. Multiple electronic circuits may be integrated onto a single chip or distributed across multiple chips. Multiple chips may be integrated onto a single device or distributed across multiple devices. The computer system referred to here includes a microcontroller having one or more processors and one or more memories. Therefore, the microcontroller also consists of one or more electronic circuits, including semiconductor integrated circuits or large-scale integrated circuits.

[0126] Furthermore, it is not essential for the data augmentation system 1 to have multiple functions integrated into a single device; the multiple components of the data augmentation system 1 may be distributed across multiple devices. Moreover, at least some of the functions of the data augmentation system 1, for example, the functions of the data partitioning unit 134, may be implemented by a server or cloud (cloud computing), etc.

[0127] Conversely, in the embodiment, multiple functions that are distributed across multiple devices may be consolidated into a single device. For example, multiple functions that are distributed across the data augmentation system 1 and the learner 3 may be consolidated into a single device.

[0128] (summary) Based on the embodiments described above, the following aspects are disclosed.

[0129] In the data augmentation method according to the first embodiment, multiple augmented data is generated using multiple measured data. The multiple measured data are data based on multiple detection signals (D1) output from a laser detection device (2). The laser detection device (2) includes a pulsed laser (21), a vibration detection laser (22), and a photodetector (23). The pulsed laser (21) emits a first laser beam (L1) to vibrate the surface of the concrete (C1). The vibration detection laser (22) emits a second laser beam (L2) to detect vibrations on the surface of the concrete (C1). The photodetector (23) detects the second laser beam (L2) reflected from the surface of the concrete (C1) and converts it into a detection signal (D1). The data augmentation method includes an acquisition step and a generation step. In the acquisition step, multiple measured data is acquired. In the generation step, multiple augmented data is generated. In the generation step, data augmentation processing is performed. The data augmentation process generates multiple augmented data based on at least one of the following: information on variations in the speed of sound inside the concrete (C1), information on variations in the intensity of the first laser beam (L1) emitted from the pulsed laser (21), information on variations in the intensity of the second laser beam (L2) emitted from the vibration detection laser (22), information on variations in the sensitivity of the photodetector (23), information on variations in noise generated in the laser detection device (2), and information on the distance between the laser detection device (2) and the concrete (C1), along with multiple measured data.

[0130] According to the above configuration, multiple augmented data sets can be generated. Then, the learner (3) performs machine learning based on multiple training data sets (T1) containing multiple augmented data sets, thereby generating a trained model (4) suitable for determining the presence or absence of defects in concrete (C1).

[0131] Furthermore, the data augmentation method according to the second embodiment further includes an extraction step in the first embodiment. In the extraction step, components below the upper frequency limit are extracted from each of the multiple measured data. In the data augmentation process, multiple augmented data are generated based on the components below the upper frequency limit of each of the multiple measured data.

[0132] According to the above configuration, a more suitable pre-trained model (4) can be generated to determine whether or not there are defects in the concrete (C1).

[0133] Furthermore, in the data expansion method according to the third embodiment, the upper frequency limit is 40 kHz or more and 60 kHz or less, according to the second embodiment.

[0134] According to the above configuration, a more suitable pre-trained model (4) can be generated to determine whether or not there are defects in the concrete (C1).

[0135] Furthermore, the data augmentation method according to the fourth embodiment further includes a data splitting step in any one of the first to third embodiments. In the data splitting step, the multiple augmented data generated in the data augmentation process are divided into a training data group, a validation data group, and a test data group. The data splitting step includes a first data splitting step and a second data splitting step. In the first data splitting step, from three or more augmented data generated based on one measured data, augmented data belonging to the training data group, augmented data belonging to the validation data group, and augmented data belonging to the test data group are selected. In the second data splitting step, augmented data generated based on the first measured data from among multiple measured data is assigned to the training data group. In addition, in the second data splitting step, augmented data generated based on the second measured data from among multiple measured data is assigned to the validation data group. Furthermore, in the second data splitting step, augmented data generated based on the third measured data from among multiple measured data is assigned to the test data group.

[0136] With the above configuration, the learner (3) can generate a first trained model generated based on multiple augmented data divided by the first data division step, and a second trained model generated based on multiple augmented data divided by the second data division step. Therefore, the user can select the trained model that is more suitable for decision-making from the first trained model and the second trained model.

[0137] Configurations other than those in the first embodiment are not essential to the data augmentation method and can be omitted as appropriate.

[0138] Furthermore, the program relating to the fifth embodiment is a program that causes one or more processors of a computer system to execute a data extension method relating to any one of the first to fourth embodiments.

[0139] According to the above configuration, multiple augmented data sets can be generated. Then, the learner (3) performs machine learning based on multiple training data sets (T1) containing multiple augmented data sets, thereby generating a trained model (4) suitable for determining the presence or absence of defects in concrete (C1).

[0140] Furthermore, the trained model (4) according to the sixth embodiment is generated by machine learning using multiple augmented data generated by the data augmentation method according to any one of the first to fourth embodiments as training data (T1). The trained model (4) outputs an inspection result regarding the presence or absence of defects in the concrete (C1) based on input data corresponding to the detection signal (D1) (data based on the output signal (D2)).

[0141] According to the above configuration, a trained model (4) suitable for determining the presence or absence of defects in concrete (C1) can be generated.

[0142] Furthermore, the inspection method relating to the seventh aspect includes an inference step. In the inference step, input data corresponding to the detection signal (D1) (data based on the output signal (D2)) is input to the trained model (4) to obtain an inspection result regarding the presence or absence of defects in the concrete (C1).

[0143] According to the above configuration, it is possible to determine whether or not there are defects in the concrete (C1).

[0144] Furthermore, the data augmentation system (1) according to the eighth embodiment generates multiple augmented data using multiple measured data. The multiple measured data are data based on multiple detection signals (D1) output from the laser detection device (2). The laser detection device (2) includes a pulsed laser (21), a vibration detection laser (22), and a photodetector (23). The pulsed laser (21) emits a first laser beam (L1) to vibrate the surface of the concrete (C1). The vibration detection laser (22) emits a second laser beam (L2) to detect vibrations on the surface of the concrete (C1). The photodetector (23) detects the second laser beam (L2) reflected from the surface of the concrete (C1) and converts it into a detection signal (D1). The data augmentation system (1) includes an acquisition unit (132) and a generation unit (133). The acquisition unit (132) acquires multiple measured data. The generation unit (133) generates multiple extended data. The generation unit (133) performs data extension processing. The data extension processing is a process that generates multiple extended data based on at least one of the following: information on the variation in the speed of sound inside the concrete (C1), information on the variation in the intensity of the first laser beam (L1) irradiated from the pulsed laser (21), information on the variation in the intensity of the second laser beam (L2) irradiated from the vibration detection laser (22), information on the variation in the sensitivity of the photodetector (23), information on the variation in noise generated in the laser detection device (2), and information on the distance between the laser detection device (2) and the concrete (C1), along with multiple measured data.

[0145] According to the above configuration, multiple augmented data sets can be generated. Then, the learner (3) performs machine learning based on multiple training data sets (T1) containing multiple augmented data sets, thereby generating a trained model (4) suitable for determining the presence or absence of defects in concrete (C1).

[0146] Not limited to the above embodiments, various configurations (including modifications) of the data augmentation system (1) according to the embodiment can be realized by a data augmentation method, a (computer) program, or a non-temporary recording medium on which the program is recorded. [Explanation of Symbols]

[0147] 1. Data Expansion System 2. Laser detection device 4. Pre-trained models 21 pulsed lasers 22. Laser for vibration detection 23 Photodetector 132 Acquisition Department 133 Generation part C1 Concrete D1 detection signal D2 output signal L1 First laser beam L2 Second laser beam T1 training data

Claims

1. A data augmentation method that generates multiple augmented data using multiple measured data based on multiple detection signals output from a laser detection device, The laser detection device is A pulsed laser that emits a first laser beam to vibrate the surface of the concrete, A vibration detection laser that emits a second laser beam for detecting vibrations of the surface of the concrete, The system includes a photodetector that detects the second laser light reflected from the surface of the concrete and converts it into a detection signal, The aforementioned data expansion method is: The acquisition step involves acquiring the aforementioned multiple measured data, The process includes a generation step of generating the aforementioned multiple extended data, In the above generation step, The process involves performing data augmentation, which involves obtaining multiple data obtained by changing the acquired multiple measured data when at least one of the following elements is simulated to be changed within the range of change for each element: the speed of sound inside the concrete, the intensity of the first laser beam emitted from the pulsed laser, the intensity of the second laser beam emitted from the vibration detection laser, the sensitivity of the photodetector, the noise generated in the laser detection device, and the distance between the laser detection device and the concrete, and generating the obtained multiple data as multiple augmented data separately from the acquired multiple measured data. Data expansion method.

2. The method further includes an extraction step of extracting components below the upper frequency limit from each of the aforementioned plurality of measured data, In the data augmentation process, the plurality of augmented data are generated based on the components of each of the plurality of measured data that are below the upper limit frequency. The data expansion method according to claim 1.

3. The aforementioned upper frequency limit is 40 kHz or more and 60 kHz or less. The data expansion method according to claim 2.

4. The process further includes a data splitting step of dividing the multiple augmented data generated in the aforementioned data augmentation process into a training data set, a validation data set, and a test data set. The aforementioned data partitioning step includes a first data partitioning step and a second data partitioning step, In the first data splitting step, from three or more extended data sets generated based on one measured data set, extended data belonging to the training data set, extended data belonging to the verification data set, and extended data belonging to the test data set are selected. In the second data partitioning step, The extended data generated based on the first measured data from the aforementioned plurality of measured data is assigned to the training data group. The extended data generated based on the second measurement data from the aforementioned plurality of measurement data is assigned to the verification data group. The extended data generated based on the third measurement data from the aforementioned plurality of measurement data is assigned to the test data group. The data expansion method according to claim 1.

5. To cause one or more processors of a computer system to execute the data expansion method described in any one of claims 1 to 4, program.

6. A trained model generated by machine learning using the plurality of augmented data generated by the data augmentation method described in any one of claims 1 to 4 as training data, Computers, The system is configured to output inspection results regarding the presence or absence of defects in the concrete, based on input data corresponding to the detection signal. A pre-trained model.

7. The invention provides a pre-trained model according to claim 6, comprising an inference step of inputting input data corresponding to the detection signal to obtain the inspection result regarding the presence or absence of defects in the concrete, Testing method.

8. A data augmentation system that generates multiple augmented data using multiple measured data based on multiple detection signals output from a laser detection device, The laser detection device is A pulsed laser that emits a first laser beam to vibrate the surface of the concrete, A vibration detection laser that emits a second laser beam for detecting vibrations of the surface of the concrete, The system includes a photodetector that detects the second laser light reflected from the surface of the concrete and converts it into a detection signal, The aforementioned data expansion system is The acquisition unit acquires the aforementioned multiple measured data, The system comprises a generation unit that generates the aforementioned plurality of extended data, The generating unit is The process involves performing data augmentation, which involves obtaining multiple data obtained by changing the acquired multiple measured data when at least one of the following elements is simulated to be changed within the range of change for each element: the speed of sound inside the concrete, the intensity of the first laser beam emitted from the pulsed laser, the intensity of the second laser beam emitted from the vibration detection laser, the sensitivity of the photodetector, the noise generated in the laser detection device, and the distance between the laser detection device and the concrete, and generating the obtained multiple data as multiple augmented data separately from the acquired multiple measured data. Data expansion system.

Images