Hammering inspection system and program

Abstract

To provide a striking sound inspection system that can automatically and highly accurately inspect a state of a structure.SOLUTION: A striking sound inspection system 1 comprises: a teacher data learning unit that learns normality teacher data prepared based on a reference normal striking sound of a structure found out to be normal, and abnormality teacher data prepared based on a reference abnormal striking sound of a structure found out that abnormality occurs; an inspection data acquisition unit that acquires inspection data on the basis of an inspection striking sound of an inspection object structure; a probability acquisition unit that analyzes the inspection data on the basis of the normality teacher data and abnormality teacher data, and acquires a probability in which the inspection striking sound is a striking sound of the structure in a normal state; and a judgement unit that judges to be normal when the probability is more than a first threshold, judges to be abnormal when the probability is equal to or less than a second threshold, and judges to be cautious when the probability is greater than the second threshold and smaller than the first threshold.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a hammering inspection system and program for determining the condition of a structure based on the hammering sound generated by striking the structure. [Background technology]

[0002] Conventionally, when inspecting a structure, a skilled worker would strike the surface of the structure with a hammer or the like, and check for any abnormalities based on the resulting striking sound.

[0003] However, due to the declining birthrate and aging population in recent years, as well as the shrinking working population, there has been a shortage of such workers, and there are concerns that it will be difficult to secure the necessary personnel for testing and that the accuracy of testing will decline.

[0004] Therefore, an inspection device has been developed to assist in the analysis of impact sounds of structures (see Patent Document 1). This inspection device identifies and visualizes specific parts where the sound pressure level is high in the frequency distribution of impact sounds of the structure, and by using this inspection device, it becomes easier to detect abnormalities in the structure based on the impact sounds. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Laid-Open No. 2013-253947 Summary of the Invention [Problem to be solved by the invention]

[0006] However, although the inspection device according to the above-mentioned Patent Document 1 can visualize the frequency distribution of impact sounds and is useful for inspection, it is still the worker who judges whether there is an abnormality based on the visualized frequency distribution, so the situation remains that skilled workers are still required. Also, although Patent Document 1 mentions the automation of inspections, it does not consider at all how to specifically automate inspections.

[0007] The present invention has an object to solve such problems, and to provide a hammering inspection system and program that can automatically and highly accurately inspect the condition of a structure. [Means for solving the problem]

[0008] The present invention is characterized by comprising a teacher data learning unit that learns normal teacher data created based on standard normal impact sounds of structures that are known to be normal and abnormal teacher data created based on standard abnormal impact sounds of structures that are known to have abnormalities; an inspection data acquisition unit that acquires inspection data based on the inspection impact sounds of the structure under inspection; a probability acquisition unit that analyzes the inspection data based on the normal teacher data and the abnormal teacher data and acquires the probability that the inspection impact sound is an impact sound of a structure in a normal state; and a judgment unit that judges the probability to be normal if the probability is equal to or greater than a first threshold, judges the probability to be abnormal if the probability is equal to or less than a second threshold, and judges the probability to be abnormal if the probability is greater than the second threshold and less than the first threshold.

[0009] In the above-described hammering sound inspection system, it is preferable that the normal teacher data is obtained by applying a Hann window to the reference normal impact sound to create a reference normal spectrogram with time on the X axis and frequency on the Y axis, creating a reference normal matrix from the reference normal spectrogram, and normalizing each element of the reference normal matrix; the abnormal teacher data is obtained by applying a Hann window to the reference abnormal impact sound to create a reference abnormal spectrogram with time on the X axis and frequency on the Y axis, creating a reference abnormal matrix from the reference abnormal spectrogram, and normalizing each element of the reference abnormal matrix; and the inspection data is obtained by applying a Hann window to the inspection impact sound to create an inspection spectrogram with time on the X axis and frequency on the Y axis, creating a check matrix from the check spectrogram, and normalizing each element of the check matrix.

[0010] Furthermore, in the hammering sound inspection system, it is preferable that the teacher data learning unit learns by setting elements of the reference normal matrix corresponding to the reference normal impact sound for the normal teacher data to 1, and by setting elements of the reference abnormal matrix corresponding to the reference abnormal impact sound for the abnormal teacher data to 0.

[0011] Furthermore, in the above-mentioned hammering sound inspection system, it is preferable that the normal teacher data, the abnormal teacher data, and the inspection data are obtained by normalizing each element of the reference normal matrix, the reference abnormal matrix, and the inspection matrix, respectively, and then multiplying them by a weighting matrix corresponding to a predetermined frequency characteristic of the abnormal impact sound.

[0012] In addition, in the hammering inspection system, it is preferable that the weighting matrix is ​​a matrix in which the value of each element in the column corresponding to the predetermined frequency is 1, and the elements in the other columns are 1 or less.

[0013] In addition, in the hammering inspection system, it is preferable that the teacher data learning unit, the probability acquisition unit, and the determination unit are realized by a neural network.

[0014] In addition, in the hammering inspection system, it is preferable that the teacher data learning unit, the probability acquisition unit, and the determination unit are realized by a cloud service.

[0015] The present invention also provides a program for causing a computer to execute a hammering inspection system, the program comprising: a teacher data learning process for learning normal teacher data created based on standard normal hammering sounds of structures known to be normal and abnormal teacher data created based on standard abnormal hammering sounds of structures known to have an abnormality; an inspection data acquisition process for acquiring inspection data based on the inspection impact sound of the structure to be inspected; a probability acquisition process for analyzing the inspection data based on the normal teacher data and the abnormal teacher data to acquire the probability that the inspection impact sound is the standard normal impact sound; and a judgment process for determining the sound to be normal if the probability is equal to or greater than a first threshold, determining the sound to be abnormal if the probability is equal to or less than a second threshold, and determining that caution is required if the probability is greater than the second threshold and less than the first threshold. [Effects of the Invention]

[0016] According to the hammering inspection system and program of the present invention, the condition of a structure can be inspected automatically and with high accuracy. [Brief explanation of the drawings]

[0017] [Figure 1] FIG. 1 is a schematic diagram showing an example of how a hammering inspection system according to an embodiment of the present invention is used. [Figure 2] FIG. 2 is a functional block diagram of a system management server used in the hammering inspection system according to the embodiment of the present invention. [Figure 3] 1 is a flowchart showing a process for determining the state of a structure by the hammering inspection system according to the embodiment of the present invention. [Figure 4] 10 is a graph showing a waveform of an impact sound obtained by striking a structure. [Figure 5] This is a graph in which one wave of the impact sound graph in FIG. 4 is cut out so that the wavelength becomes maximum after a predetermined time has elapsed. [Figure 6] 6 is a graph showing other waves of the impact sound graph in FIG. 4 cut out under the same conditions as FIG. 5. [Figure 7] This is a spectrogram obtained from normal training data. [Figure 8] This is a spectrogram obtained from abnormal training data. [Figure 9] FIG. 1 is a schematic diagram showing a neural network used in a teacher data learning unit of a hammering inspection system according to an embodiment of the present invention. [Figure 10] 10 is a graph showing the normal probability and the abnormal probability calculated by a probability acquisition unit for an impact sound generated when a part of a structure in a normal state is struck. [Figure 11] 10 is a graph showing the normal probability and abnormal probability calculated by a probability acquisition unit for an impact sound generated when a part of a structure in which an abnormality has occurred is struck. [Figure 12] 10 is an example of a graph showing the normal probability and the abnormal probability calculated by a probability acquisition unit for an impact sound generated when an actual structure is struck. [Figure 13] 10 is another example of a graph showing the normal probability and the abnormal probability calculated by the probability acquisition unit for the impact sound generated when an actual structure is struck. DETAILED DESCRIPTION OF THE INVENTION

[0018] Hereinafter, an embodiment of a hammering inspection system according to the present invention will be described with reference to the accompanying drawings.

[0019] Fig. 1 is a schematic diagram showing a hammering inspection system 1 according to an embodiment of the present invention. As shown in Fig. 1, the hammering inspection system 1 includes a sound collection terminal 2 that collects impact sounds generated when a surface of a structure S made of concrete or the like is struck by an impact device 4, and a system management server 3 that inspects the condition of the structure based on the impact sound data provided from the sound collection terminal 2 via a network N.

[0020] The sound collection terminal 2 includes a sound collection unit that collects sound, a transmission / reception unit that converts the collected sound into data in a predetermined format and transmits it to the system management server 3 via the network N, and receives predetermined data indicating the inspection results from the system management server 3 via the network N, and a display unit that displays the received inspection results. As such a sound collection terminal 2, in addition to a dedicated terminal, a commercially available smartphone, mobile phone, tablet terminal, etc. may also be used, and in this embodiment, an example in which a smartphone is used will be described.

[0021] The system management server 3 is configured as a cloud server and includes a CPU (Central processing unit) 31 that controls the entire system, a ROM (Read-only memory) 32 that stores control programs and the like that run on the CPU 31, a RAM (Random access memory) 33 for temporarily storing various data, and a memory unit 34 for storing data on impact sounds transmitted from the sound collection terminal 2 and various data generated by the impact sound inspection system 1.

[0022] 2 is a functional block diagram of the system management server 3 used in the hammering inspection system 1 according to the embodiment of the present invention. As shown in FIG. 2, the system management server 3 functions as a teacher data learning unit 311, an inspection data acquisition unit 312, a probability acquisition unit 313, and a determination unit 314 by the CPU 31 loading a control program stored in the ROM 32 into the RAM 33 and executing it.

[0023] The teacher data learning unit 311 learns normal teacher data created based on reference normal impact sounds of structures that are known to be normal, and abnormal teacher data created based on reference abnormal impact sounds of structures that are known to be abnormal. Details of learning the normal teacher data and abnormal teacher data will be described later.

[0024] The inspection data acquisition unit 312 acquires inspection data based on the inspection impact sound of the inspection target structure. Details of the acquisition of inspection data will be described later.

[0025] The probability acquisition unit 313 analyzes the test data based on the normal teacher data and the abnormal teacher data, and acquires the probability that the test impact sound is a reference normal impact sound. Details of acquiring the probability that the test impact sound is a reference normal impact sound will be described later.

[0026] The judgment unit 314 judges the test impact sound to be normal if the probability that it is a reference normal impact sound is equal to or greater than a first threshold, judges it to be abnormal if it is equal to or less than a second threshold, and judges it to require caution if it is greater than the second threshold and less than the first threshold.

[0027] The impact device 4 is configured to include a main body 41 that houses a power supply, a control unit, a drive unit, etc. (none of which are shown), and a hammer unit 42 that is movable forward and backward relative to the main body 41 and can impact a structure with a constant impact force. By automating the impact of a structure using the impact device 4, it is possible to reduce variations in the impact force during impact and perform inspection with high accuracy. However, the impact device 4 is not an essential component of the present invention, and the impact device 4 may be omitted and an operator may manually impact the structure S using a hammer.

[0028] Next, a detailed description will be given of the operation of the hammering inspection system 1 having the above-described configuration. Fig. 3 is a flowchart showing a process for determining the state of the structure S by the hammering inspection system 1 according to the embodiment of the present invention.

[0029] As shown in Fig. 3, first, the teacher data learning unit 311 learns the teacher data (step S1). Specifically, the teacher data learning unit 311 creates normal teacher data based on reference normal impact sounds, which are impact sounds of a structure that are known to be normal, and creates abnormal teacher data based on reference abnormal impact sounds, which are impact sounds of a structure that are known to be abnormal, and learns these. The teacher data learning unit 311 learns the teacher data in advance before actually inspecting the structure S. Details of this teacher data learning are described below.

[0030] Figure 4 is a graph showing the waveform of the impact sound obtained by hitting the structure S. If the impact sound generated by repeatedly hitting the structure S with a predetermined hitting force at a predetermined time interval is graphed with time (seconds) on the X axis and sound intensity (bits) on the Y axis, a graph showing a wave for each hit is obtained, as shown in Figure 4.

[0031] FIG. 5 is a graph of one wave from the graph of impact sounds in FIG. 4, cut out so that the wavelength reaches a maximum after a predetermined time has elapsed. FIG. 6 is a graph of another wave from the graph of impact sounds in FIG. 4, cut out under the same conditions as FIG. 5. For each wave in the graph shown in FIG. 4, a predetermined time range is cut out for each peak, as shown in FIGS. 5 and 6. Each of these cut-out waves represents an impact sound corresponding to one hit. Note that in FIGS. 5 and 6, the X-axis is in seconds, and the Y-axis is in bits.

[0032] Then, for each of the extracted waves, a spectrogram is created by applying a Hann window with the X axis representing time and the Y axis representing frequency, and a matrix as shown in Equation 1 is obtained from the spectrogram. The matrix shown in Equation 1 is called a reference matrix. Note that, for convenience, the reference matrix is ​​assumed to be a 3-row, 3-column matrix here, but a matrix consisting of any number of rows and columns can be used when learning actual training data or acquiring test data, which will be described later.

[0033]

number

[0034] Next, the reference matrix is ​​normalized. First, the maximum value of the elements of the reference matrix M (M=max(a ij )) and divide all elements of the matrix in Equation 1 by their maximum value M to obtain the matrix shown in Equation 2.

[0035]

number

[0036] Here, each element of number 2 is between 0 and 1.

[0037] Next, each element of the matrix in Equation 2 is multiplied by 10,000, and then logarithmized to obtain the logarithm shown in Equation 3. The matrix shown in Equation 3 has matrix elements consisting of two values, 0 and 10,000.

[0038]

number

[0039] Then, the matrix of Equation 3 is divided by Log10000 to obtain the matrix shown in Equation 4. The matrix shown in Equation 4 has matrix elements consisting of two values, 0 and 1.

[0040]

number

[0041] The teacher data learning unit 311 acquires and normalizes the reference matrix described above for a spectrogram (reference normal spectrogram) obtained from an impact sound of a structure known to be normal (reference normal impact sound) and a spectrogram (reference abnormal spectrogram) obtained from an impact sound of a structure known to be abnormal (reference abnormal impact sound).The teacher data learning unit 311 then learns, as normal teacher data, a matrix obtained by normalizing the reference matrix (reference normal matrix) obtained from the reference normal spectrogram, and as abnormal teacher data, a matrix obtained by normalizing the reference matrix (reference abnormal matrix) obtained from the reference abnormal spectrogram.The teacher data learning unit 311 sets the elements of the reference normal matrix corresponding to the reference normal impact sound for the normal teacher data to 1, and sets the elements of the reference abnormal matrix corresponding to the reference abnormal impact sound for the abnormal teacher data to 0.

[0042] The normal teacher data and the abnormal teacher data are each in a matrix form that makes it easy to distinguish frequencies characteristic of a structure in a normal state from frequencies characteristic of a structure in an abnormal state. This can be easily understood by converting the normal teacher data and the abnormal teacher data into spectrograms, as shown in Figures 7 and 8. Figure 7 is a spectrogram obtained from the normal teacher data. Figure 8 is a spectrogram obtained from the abnormal teacher data. The spectrogram obtained from the abnormal teacher data shown in Figure 8 has a characteristic peak around 4000 Hz that is not seen in the spectrogram obtained from the normal teacher data shown in Figure 7. This frequency peak is the peak of the impact sound specific to a structure in an abnormal state, which is used in this embodiment. This frequency peak is merely an example and will vary depending on the type of structure and the type of abnormality.

[0043] 9 is a schematic diagram showing a neural network 3110 used in the teacher data learning unit 311 of the hammering inspection system according to an embodiment of the present invention. As shown in FIG. 9, in this embodiment, the neural network 3110 is composed of three layers: an input layer 3111, an intermediate layer 3112, and an output layer 3113. The teacher data learning unit 311 uses this neural network 3110 to perform learning by setting the elements of a reference normal matrix to 1 and the elements of a reference anomalous matrix to 0. Specifically, the input layer 3111 acquires the reference normal matrix and the reference anomalous matrix, the intermediate layer 3112 performs predetermined pre-calculation using the reference normal matrix and the reference anomalous matrix to produce a second-order (0, 1) output, and the output layer 3113 produces the second-order output.

[0044] 3, after the teacher data learning unit 311 has learned the teacher data, the test data acquisition unit 312 then acquires the test data (step S2). In this embodiment, as shown in FIG. 1, the test data is acquired by transmitting data of the impact sound (test impact sound) of the structure S to be inspected by the sound collection terminal 2 via the network N to the system management server 3, and the test data acquisition unit 312 generates test data based on the data of the impact sound.

[0045] Specifically, the test data is obtained by test data acquisition unit 312 applying a Hann window to the test impact sound to create a test spectrogram with time on the X axis and frequency on the Y axis, creating a check matrix from the test spectrogram, and normalizing each element of the check matrix. The acquisition of the test spectrogram and check matrix from the test impact sound by test data acquisition unit 312 and the normalization of the check matrix are performed by using a neural network similar to neural network 3110 and executing a series of processes similar to those performed by teacher data learning unit 311 described above to acquire a reference normal spectrogram (reference abnormal spectrogram) and a reference normal matrix (reference abnormal matrix) from a reference normal impact sound (reference abnormal matrix), and normalizing the reference normal matrix (reference abnormal matrix).

[0046] Next, the probability acquisition unit 313 acquires the probability that the inspection impact sound is an impact sound of the structure S in a normal state based on the acquired inspection data, normal teacher data, and abnormal teacher data (step S3).

[0047] Fig. 10 is a graph showing the normal probability and abnormal probability calculated by the probability acquisition unit 313 for the impact sound generated when a structure S in a normal state is struck. Fig. 11 is a graph showing the normal probability and abnormal probability calculated by the probability acquisition unit for the impact sound generated when a structure S in an abnormal state is struck. The input matrix data is processed by neural network processing that references threshold data in the intermediate layer 3112 that has learned based on training data, and the probability of normality or abnormality is determined.

[0048] In the example of structure S in a normal state shown in Figure 10, when a structure known to be in a normal state is hit at 16 locations, the probability that any of the structures is in a normal state is calculated to be approximately 1, and the probability that any of the structures is in an abnormal state is calculated to be approximately 0.

[0049] In the example of structure S, which is known to be in an abnormal state, shown in Figure 11, if structure S, which is known to be in an abnormal state, is struck at 15 locations, the probability that any of the structures is normal is calculated to be less than 0.3, and the probability that any of the structures is in an abnormal state is calculated to be 0.7 or more.

[0050] In this way, it can be seen that the probability calculation by the probability acquisition unit 313 can accurately reflect the difference in probability between a structure S in a normal state and a structure S in which an abnormality has occurred.

[0051] A case where this probability calculation by the probability acquisition unit 313 is performed for an actual structure S will be described using Figures 12 and 13. Figure 12 is an example of a graph showing the normal probability and abnormal probability calculated by the probability acquisition unit 313 for an impact sound generated when an actual structure S is struck. Figure 13 is another example of a graph showing the normal probability and abnormal probability calculated by the probability acquisition unit 313 for an impact sound generated when an actual structure S is struck.

[0052] In the example shown in Figure 12, impacts were applied to 17 locations on structure S, and the probability that the structure was in a normal state for each impact was 0.7 or higher, and the probability that the structure was abnormal was less than 0.3.

[0053] On the other hand, in the example shown in Figure 13, impacts were performed at 18 locations on another structure S, and the results showed that there was variation in the probability that each location was in a normal state and the probability that an abnormality had occurred, and it was inferred that there were locations on structure S where an abnormality had occurred or where attention was required.

[0054] Returning to FIG. 3, next, the determination unit 314 determines the state of the structure S based on the probability calculated by the probability acquisition unit 313 in this manner (step S4).

[0055] Specifically, the determination unit 314 compares the probability calculated by the probability acquisition unit 313 with a first threshold value and a second threshold value that is lower than the first threshold value, thereby determining the state of the structure S. In this embodiment, the determination is performed using the first threshold value of 0.7 and the second threshold value of 0.4, but these threshold values ​​can be set to any numerical values ​​within a range greater than 0 and less than 1.0.

[0056] In a specific determination, first, the probability is compared with a first threshold value (step S5). If the probability is equal to or greater than the first threshold value (step S5: Yes), the determination unit 314 determines that the structure S is "normal."

[0057] If the probability is lower than the first threshold (step S5: No), the probability is then compared with the second threshold (step S6). If the probability is not equal to or lower than the second threshold (step S6: No), the determination unit 314 determines that there is a possibility that an abnormality has occurred in the structure S and that caution is required, and determines that "Caution" is required.

[0058] On the other hand, if the probability is equal to or less than the second threshold value (step S6: Yes), the determining unit 314 determines that an abnormality has occurred in the structure S and determines that the structure S is "abnormal."

[0059] Then, data indicating the judgment result of the judgment unit 314, "normal," "caution," or "abnormal," is stored in the memory unit 34 so that it can be referenced by external devices, and is also transmitted from the system management server 3 to the sound collection terminal 2 via the network N, and the judgment result is displayed on the display unit of the sound collection terminal 2. This allows the worker at the work site to instantly know the condition of the struck part of the structure S.

[0060] In this way, according to the hammering inspection system 1 of the present invention, condition inspection of the structure S by hammering, which has conventionally been carried out based on the intuition and experience of skilled workers, can be carried out automatically and with high accuracy and without variation in judgment.

[0061] In addition, hammering sound data and judgment data stored in the system management server 3, which is a cloud server, can be shared in real time with the site, office, etc. Furthermore, since it is possible to compare past hammering sound data and judgment data stored in the system management server 3, it is possible to observe the deterioration of the structure S over time.

[0062] The present invention is not limited to the above-described embodiment, and various modifications can be adopted.

[0063] For example, when the above-described teacher data learning unit 311 and the test data acquiring unit 312 acquire the normal teacher data, the abnormal teacher data, and the test data, respectively, the elements of the reference normal matrix, the reference abnormal matrix, and the test matrix may be normalized and then multiplied by a weighting matrix corresponding to a predetermined frequency characteristic of the reference abnormal impact sound. For example, as in the above-described embodiment, if a characteristic peak appears around 4000 Hz in the spectrogram acquired from the teacher data, the elements may be multiplied by a weighting matrix in which each element in the column corresponding to 4000 Hz is 1 and each element in the other columns is 1 or less, and the value gradually decreases as the frequency moves away from the frequency, as shown in the following equation 5.

[0064]

number

[0065] Here, the matrix obtained by normalizing each element of the reference normal matrix, the reference abnormal matrix, and the inspection matrix is ​​defined as D=(d ij ), and the weighting matrix is ​​W=(w ij ), then the weighted matrix E=(w ij *d ij )

[0066] In this way, by multiplying a weighting matrix that emphasizes peaks found in abnormal structure S by a matrix obtained by normalizing each element of the reference normal matrix, reference abnormal matrix, and inspection matrix, normal teacher data, abnormal teacher data, and inspection data can be obtained that allow for a clearer determination of abnormalities in structure S.

[0067] Furthermore, in the above-described embodiment, the system management server 3 is equipped with the inspection data acquisition unit 312, but the present invention is not limited to this. The sound collection terminal 2 may be equipped with the inspection data acquisition unit 312, and the state of the structure S may be determined by transmitting the inspection data obtained by the sound collection terminal 2 to the system management server 3.

[0068] Furthermore, in the above-described embodiment, the sound collection terminal 2 and the system management server 3 are configured as separate units, but the present invention is not limited to this. For example, the system management server 3 may be eliminated, and the sound collection terminal 2 may incorporate the above-described functions of the system management server 3 (teacher data learning unit 311, test data acquisition unit 312, probability acquisition unit 313, and judgment unit 314).

[0069] Furthermore, in the above-described embodiment, the hammering test is performed by transmitting the hammering data collected by the sound collecting terminal 2 to the system management server 3 in real time, but the present invention is not limited to this. The hammering test may be performed by transmitting hammering data that has been collected and saved in advance using the sound collecting terminal 2 or the like to the system management server 3 at a desired timing. In this case, the hammering data may be transmitted to the system management server 3 not only from the sound collecting terminal 2 but also from, for example, a PC (Personal Computer) installed in an office. The determination result by the system management server 3 may also be transmitted not only to the sound collecting terminal 2 but also to other external devices. [Explanation of symbols]

[0070] 1: Hammering inspection system 2: Sound collection device 3: System management server 4: Impact device 31:CPU 32:ROM 33:RAM 34: Storage part 311: Teacher Data Learning Department 312: Inspection data acquisition unit 313: Probability acquisition section 314: Judgment section S: Structure N: Network

Claims

1. a teacher data learning unit that learns normal teacher data created based on a reference normal impact sound of a structure that is known to be normal and abnormal teacher data created based on a reference abnormal impact sound of a structure that is known to be abnormal; an inspection data acquisition unit that acquires inspection data based on the inspection impact sound of the inspection target structure; a probability acquisition unit that analyzes the inspection data based on the normal teacher data and the abnormal teacher data and acquires a probability that the inspection impact sound is an impact sound of a structure in a normal state; a determination unit that determines the probability to be normal if the probability is equal to or greater than a first threshold, determines the probability to be abnormal if the probability is equal to or less than a second threshold, and determines that caution is required if the probability is greater than the second threshold and smaller than the first threshold; A hammering inspection system comprising:

2. the normal teacher data is obtained by applying a Hann window to the reference normal impact sound to create a reference normal spectrogram with time on the X axis and frequency on the Y axis, creating a reference normal matrix from the reference normal spectrogram, and normalizing each element of the reference normal matrix; the abnormal teacher data is obtained by applying a Hann window to the reference abnormal impact sound to create a reference abnormal spectrogram with time as the X axis and frequency as the Y axis, creating a reference abnormal matrix from the reference abnormal spectrogram, and normalizing each element of the reference abnormal matrix; 2. The hammering inspection system according to claim 1, wherein the inspection data is obtained by applying a Hann window to the inspection impact sound to create an inspection spectrogram in which time is represented on the X axis and frequency is represented on the Y axis, creating a check matrix from the inspection spectrogram, and normalizing each element of the check matrix.

3. 3. The hammering sound inspection system according to claim 2, wherein the teacher data learning unit performs learning by setting an element of the reference normal matrix corresponding to the reference normal impact sound for the normal teacher data to 1, and setting an element of the reference abnormal matrix corresponding to the reference abnormal impact sound for the abnormal teacher data to 0.

4. 4. The hammering sound inspection system according to claim 2 or 3, wherein the normal teacher data, the abnormal teacher data, and the inspection data are obtained by normalizing each element of the reference normal matrix, the reference abnormal matrix, and the inspection matrix, respectively, and then multiplying the elements by a weighting matrix corresponding to a predetermined frequency characteristic of the reference abnormal impact sound.

5. 5. The hammering inspection system according to claim 4, wherein the weighting matrix is ​​a matrix in which each element in a column corresponding to the predetermined frequency has a value of 1 and elements in other columns have values ​​of 1 or less.

6. 6. The hammering inspection system according to claim 1, wherein the teacher data learning unit, the probability acquisition unit, and the determination unit are realized by neural networks.

7. The hammering inspection system according to claim 1 , wherein the teacher data learning unit, the probability acquisition unit, and the determination unit are implemented by a cloud service.

8. A program for causing a computer to execute a hammering inspection system, a teacher data learning process for learning normal teacher data created based on a reference normal impact sound of a structure that is known to be normal and abnormal teacher data created based on a reference abnormal impact sound of a structure that is known to be abnormal; an inspection data acquisition step of acquiring inspection data based on the inspection impact sound of the inspection target structure; a probability acquisition step of analyzing the test data based on the normal teacher data and the abnormal teacher data to acquire a probability that the test impact sound is the reference normal impact sound; a determination step of determining that the probability is normal if the probability is equal to or greater than a first threshold, determining that the probability is abnormal if the probability is equal to or less than a second threshold, and determining that caution is required if the probability is greater than the second threshold and less than the first threshold; A program comprising:

Images