Floor slab deterioration detection method and floor slab deterioration detection device

Abstract

To simplify work for detecting deterioration while improving detection accuracy.SOLUTION: A floor slab deterioration detection method includes a learning step and an evaluation step. The learning step includes irradiating a set range including at least a part of a road surface of a road bridge with electromagnetic waves, generating learning data by associating three-dimensional reflection information obtained from the reflected waves with a correct answer value indicating the presence or absence of gravelling on an upper surface of the floor slab for each of a plurality of divided regions obtained by arbitrarily dividing the set range, and constructing a determination model that outputs the presence or absence of gravelling on the upper surface of the floor slab in response to the input of three-dimensional reflection information through machine learning on the basis of the learning data. The evaluation step includes determining the presence or absence of gravelling on the upper surface of the floor slab in an evaluation target region on the basis of the reflection information for evaluation obtained by irradiating the evaluation target region on the road surface of the road bridge with the electromagnetic waves and the determination model.SELECTED DRAWING: Figure 7

Description

[Technical field]

[0001] The present disclosure relates to a deck deterioration detection method and a deck deterioration detection device. [Background technology]

[0002] Patent Documents 1 to 3 disclose techniques for detecting deterioration in the deck of a bridge (road bridge). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2014-228349 A [Patent Document 2] JP 2016-188545 A [Patent Document 3] Patent Publication No. 2021-81352 Summary of the Invention [Problem to be solved by the invention]

[0004] The present disclosure provides a deck deterioration detection method and a deck deterioration detection device that can improve detection accuracy while simplifying the work required to detect deterioration. [Means for solving the problem]

[0005] [1] A deck deterioration detection method including a learning step and an evaluation step, wherein the learning step includes: irradiating electromagnetic waves to a set range including at least a part of the road surface of a road bridge A; for each of a plurality of divided areas obtained by arbitrarily dividing the set range, generating learning data DA by associating three-dimensional reflection information obtained from reflected waves accompanying the irradiation of the electromagnetic waves with a ground truth value indicating the presence or absence of gravel on the deck surface; and constructing a judgment model by machine learning based on the learning data DA, which outputs the presence or absence of gravel on the deck surface in response to input of the three-dimensional reflection information; and the evaluation step includes irradiating electromagnetic waves to an evaluation target area on the road surface of the road bridge A or a road bridge B different from the road bridge A, and determining the presence or absence of gravel on the deck surface in the evaluation target area based on evaluation reflection information corresponding to the three-dimensional reflection information and the judgment model.

[0006] [2] A deck deterioration detection method as described in [1] above, wherein the three-dimensional reflection information indicates the position in each of two horizontal directions, and the relationship between the arrival time of the reflected wave associated with the irradiation of electromagnetic waves and the intensity of the reflected wave.

[0007] [3] A deck deterioration detection method as described in [2] above, wherein in the three-dimensional reflection information, the maximum value of the arrival time of the reflected wave associated with the irradiation of the electromagnetic wave is set to the time at which the reflected wave arrives from a position that is below the highest rebar in the deck and above the underside of the deck.

[0008] [4] A deck deterioration detection method described in any one of [1] to [3] above, wherein the learning data DA includes, as the correct value, information indicating the result of determining whether the road surface in the set range has been excavated and turned into gravel for each of the multiple divided areas.

[0009] [5] The deck deterioration detection method according to any one of [1] to [4] above, further comprising: a preparation step of preparing, after construction of the judgment model, another learning data DC corresponding to the learning data DA from a road bridge C different from the road bridge A; a construction step of constructing, by machine learning, a CNN model which is a convolutional neural network model that outputs the presence or absence of gravel on the deck top surface in response to input of first converted information obtained by wavelet transforming the three-dimensional reflection information in the learning data DA, based on the learning data DA; a first acquisition step of inputting the first converted information into the CNN model and acquiring information on feature quantities from an intermediate layer of the CNN model; a second acquisition step of inputting second converted information obtained by wavelet transforming the three-dimensional reflection information in the learning data DC into the CNN model and acquiring information on feature quantities from an intermediate layer of the CNN model; and a display step of displaying, on a monitor, information on the feature quantities acquired in each of the first acquisition step and the second acquisition step.

[0010] [6] A deck deterioration detection device comprising: a measurement unit that irradiates electromagnetic waves onto a set range including at least a part of the road surface of a road bridge A; a learning data generation unit that generates learning data DA by matching three-dimensional reflection information obtained from the reflected waves accompanying the irradiation of the electromagnetic waves with a ground truth value indicating the presence or absence of gravel on the deck surface for each of a plurality of divided areas obtained by arbitrarily dividing the set range; a model construction unit that constructs by machine learning a judgment model based on the learning data DA, which outputs the presence or absence of gravel on the deck surface in response to input of the three-dimensional reflection information; and a judgment unit that judges the presence or absence of gravel on the deck surface in the evaluation target area based on evaluation reflection information corresponding to the three-dimensional reflection information obtained by irradiating electromagnetic waves onto an evaluation target area on the road surface of the road bridge A or a road bridge B different from the road bridge A, and the judgment model. Effect of the Invention

[0011] According to the present disclosure, there is provided a deck deterioration detection method and a deck deterioration detection device that can improve detection accuracy while simplifying the work required to detect deterioration. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a schematic diagram showing an example of a deck deterioration detection device. [Diagram 2] FIG. 2 is a block diagram illustrating an example of a functional configuration of the arithmetic device. [Diagram 3] FIG. 3 is a block diagram illustrating an example of a hardware configuration of the arithmetic device. [Figure 4] 4(a) is a flowchart showing an example of a series of processes in the learning phase, and FIG 4(b) is a flowchart showing an example of a series of processes in the evaluation phase. [Diagram 5] 5(a) and 5(b) are schematic diagrams for explaining the processing content in the learning phase. [Figure 6] Fig. 6(a) is a schematic diagram for explaining the processing content in the learning phase, and Fig. 6(b) is a graph for explaining the processing content in the learning phase. [Figure 7] FIG. 7 is a schematic diagram for explaining the processing content in the learning phase. [Figure 8] FIG. 8 is a diagram showing the results of evaluating the degree of influence on the output result of the judgment model. [Figure 9] FIG. 9 is a schematic diagram illustrating a method for acquiring a feature amount map. [Figure 10] 10(a) and 10(b) are diagrams showing an example of a feature map obtained without performing wavelet transformation. [Figure 11] FIG. 11 is a diagram showing an example of information after wavelet transformation. [Figure 12] 12(a) and 12(b) are diagrams showing an example of a feature amount map obtained when performing wavelet transformation. [Figure 13] FIG. 13 is a flowchart showing an example of a series of processes executed when performing an evaluation using a feature amount map. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, an embodiment will be described with reference to the drawings. In the description, the same elements or elements having the same functions are denoted by the same reference numerals, and duplicated description will be omitted.

[0014] [Deck deterioration detection device] FIG. 1 shows a schematic diagram of a deck deterioration detection device according to an embodiment. The detection device 1 (deck deterioration detection device) shown in FIG. 1 is a device that executes at least a part of the work for detecting deterioration of a deck in a road bridge 100. The road bridge 100 is a structure that creates a road that crosses a river, a valley, a strait, or the like, and is sometimes called a bridge. The road bridge 100 may be an elevated road (viaduct) that crosses a ground structure such as a road or runs parallel to a road. The road bridge 100 is constructed so that vehicles can run on its road surface 101. The road bridge 100 may be any of a straight bridge, a skew bridge, and a curved bridge.

[0015] The road bridge 100 includes a deck 104 and asphalt 102. The deck 104 is a foundation of the road bridge 100 and is made of concrete. The deck 104 may be made of concrete and a plurality of reinforcing bars 108. The deck 104 may be a reinforced concrete (RC) deck or a prestressed concrete (PC) deck. The asphalt 102 is a portion formed by paving the upper surface of the deck 104 with asphalt. The asphalt 102 may be referred to as asphalt pavement or simply pavement. The upper surface of the asphalt 102 corresponds to the road surface 101 of the road bridge 100. A waterproof layer may be provided between the asphalt 102 and the deck 104 to prevent water from flowing into or penetrating the deck 104, or the asphalt 102 may be directly paved on the deck 104.

[0016] In general, it has been believed that the deck 104 deteriorates (damages) from the underside due to fatigue caused by repeated wheel loads from vehicles. However, in recent years, there have been cases where deterioration has occurred from the top surface of the deck 104, and a representative example of this is gravelization (soil and sand) on the top surface of the deck 104. Gravelization on the top surface of the deck 104 (hereinafter simply referred to as the "top surface of the deck") can occur due to water intrusion, fatigue, frost damage, alkali-aggregate reaction, salt damage, or a combination of these factors. The detection device 1 detects the occurrence of gravelization on the top surface of the deck as deterioration of the deck 104. The detection device 1 includes a measurement device 10 and a calculation device 20.

[0017] The measuring device 10 (measurement unit) is a device capable of irradiating electromagnetic waves to a road surface 101 of a road bridge 100 and receiving reflected waves accompanying the irradiation of the electromagnetic waves. The measuring device 10 irradiates electromagnetic waves from above to the road surface 101. The measuring device 10 may irradiate electromagnetic waves in any frequency range as long as it is capable of receiving reflected waves from inside the road bridge 100, but for example, the frequency of the irradiated electromagnetic waves is 200 MHz to 3 GHz. The measuring device 10 may be capable of irradiating electromagnetic waves to one point on the road surface 101, or may be capable of irradiating electromagnetic waves to multiple points on the road surface 101 at approximately the same timing. The measuring device 10 may be mounted on a vehicle 190 for measurement. In one example, the measuring device 10 includes one or more pairs of transmitting antennas and receiving antennas.

[0018] The reflected waves received by the receiving antenna included in the measuring device 10 include waves reflected by the road surface 101 (top surface of the asphalt 102), waves reflected multiple times by the road surface 101, waves reflected by the top surface of the deck 104, waves reflected multiple times by the top surface of the deck 104, and waves reflected by the reinforcing bars 108 located at the topmost position in the deck 104. Therefore, the strength of the reflected waves may change depending on the top surface of the deck 104, the portion of the deck 104 between the top surface and the reinforcing bars 108, and the state of the reinforcing bars 108. The measuring device 10 outputs a signal corresponding to the reflected waves (for example, a signal indicating a time change in signal strength) to the computing device 20. Hereinafter, the signal corresponding to the reflected waves may also be simply referred to as a "reflected wave".

[0019] The arithmetic device 20 is a computer that performs calculations to determine the presence or absence of gravel on the upper surface of the deck using the reflected waves obtained from the measurement device 10. The arithmetic device 20 may be mounted on the vehicle 190 together with the measurement device 10. The arithmetic device 20 is configured to execute a learning process and an evaluation process. In the learning process, the arithmetic device 20 constructs a judgment model for determining the presence or absence of gravel on the upper surface of the deck by machine learning. In the evaluation process, the arithmetic device 20 uses the judgment model constructed in the learning process to determine the presence or absence of gravel on the upper surface of the deck for the road bridge 100 to be evaluated.

[0020] Machine learning is a method of autonomously finding laws or rules by repeatedly learning based on given information. The specific method of machine learning is not limited. In one example, the arithmetic device 20 executes machine learning using a machine learning model, which is a computational model including a neural network. A neural network is an information processing model that mimics the mechanism of the human brain and nervous system. The type of neural network used in the arithmetic device 20 is not limited, and for example, a convolutional neural network (CNN), a recurrent neural network (RNN), or a long short-term memory (LSTM) may be used. The arithmetic device 20 may execute machine learning using a machine learning model, which is a computational model including a transformer, instead of a neural network.

[0021] The arithmetic device 20 can train the machine learning model by repeating learning, and construct this machine learning model as a trained model. The stage of performing this training and acquisition corresponds to the learning phase. The arithmetic device 20 can also output the result of determining whether or not gravel has formed on the upper surface of the deck of the road bridge 100 to be evaluated, by processing an input vector (i.e., input data) using the trained model. The stage of performing this output corresponds to the evaluation phase or operation phase. The trained model constructed by the arithmetic device 20 is portable between computer systems. The arithmetic device 20 executes both the learning phase and the evaluation phase.

[0022] The road bridge 100 to be evaluated in the evaluation phase may be a road bridge different from the learning road bridge 100 used in the learning phase. For ease of explanation, the learning road bridge 100 used in the learning process executed in the learning phase will be referred to as the "road bridge 100A," and the road bridge 100 to be evaluated in the evaluation process executed in the evaluation phase will be referred to as the "road bridge 100B."

[0023] The detection device 1 includes an input / output device 22 connected to the arithmetic device 20. The input / output device 22 is a device for inputting information indicating an input from an operator to the arithmetic device 20 and outputting information from the arithmetic device 20 to the operator. The input / output device 22 may include a keyboard, an operation panel, or a mouse as an input device, and may include a monitor 24 (e.g., a liquid crystal display) as an output device. The input / output device 22 may be a touch panel in which an input device and an output device are integrated. The arithmetic device 20 and the input / output device 22 may be integrated.

[0024] 2 shows an example of the functional configuration of the arithmetic device 20. The arithmetic device 20 has, as functional components (hereinafter referred to as "functional blocks"), for example, a measurement information acquisition unit 32, a correct answer information acquisition unit 34, a learning data generation unit 36, a model construction unit 38, a model holding unit 42, a determination unit 44, and a result output unit 46. The processes executed by these functional blocks correspond to the processes executed by the arithmetic device 20.

[0025] The measurement information acquiring unit 32 acquires, from the measuring device 10, a reflected wave (a signal indicating a reflected wave) resulting from the irradiation of an electromagnetic wave to the road bridge 100. In the learning phase, the measurement information acquiring unit 32 acquires a reflected wave when the measuring device 10 irradiates an electromagnetic wave to a set range SA including at least a part of the road surface 101 of the road bridge 100A (road bridge A). The measurement device 10 irradiates the electromagnetic wave to the set range SA while moving the measuring device 10 by, for example, having an operator operate the vehicle 190. A plurality of measurement points may be set in advance in the set range SA, and the measurement information acquiring unit 32 may acquire a reflected wave and information indicating the position of the measurement point from the measuring device 10 for each measurement point in the set range SA. In the evaluation phase, the measurement information acquiring unit 32 acquires a reflected wave when the measuring device 10 irradiates an electromagnetic wave to an area to be evaluated on the road surface 101 of the road bridge 100B (road bridge B).

[0026] The correct information acquiring unit 34 acquires a correct value indicating the presence or absence of gravel for each of a plurality of divided regions obtained by arbitrarily dividing the set range SA. The correct value is, for example, binary data indicating whether each divided region included in the plurality of divided regions is actually evaluated as having gravel or not having gravel. The correct value can also be said to be a result of classifying into two categories: actually having gravel and not having gravel. The correct information acquiring unit 34 may acquire a correct value for each divided region based on input information from a worker or the like. The plurality of divided regions are set so that each divided region includes a plurality of measurement points arranged in a matrix in two different directions.

[0027] The learning data generating unit 36 ​​generates learning data by associating, for each of the divided regions, three-dimensional reflection information acquired from the reflected waves resulting from the irradiation of the electromagnetic waves with the correct answer value indicating the presence or absence of gravel on the upper surface of the deck. Hereinafter, the three-dimensional reflection information is referred to as "reflection information RD," and the learning data generated based on the road bridge 100A is written as "learning data DA." In the learning data DA, the reflection information RD is associated with the correct answer value for each of the divided regions.

[0028] The model construction unit 38 constructs a determination model M by machine learning based on the learning data DA, which outputs the presence or absence of gravel on the upper surface of the deck in response to the input of the reflection information RD. The model construction unit 38 constructs the determination model M using, for example, a convolutional neural network (CNN). The determination model M may be constructed so as to output either binary data indicating whether it is estimated that gravel has been formed or that it is estimated that it is not estimated that gravel has been formed. The model holding unit 42 holds the determination model M constructed by the model construction unit 38.

[0029] The judgment unit 44 judges the presence or absence of gravel on the deck surface in the evaluation target area based on the evaluation reflection information corresponding to the reflection information RD obtained by irradiating an electromagnetic wave to an evaluation target area on the road surface 101 of the road bridge 100 to be evaluated (for example, the road bridge 100B) and the judgment model M held in the model holding unit 42. The evaluation target area is set to the same size as the divided area set in the learning phase, for example. The judgment unit 44 judges the presence or absence of gravel by acquiring the output from the judgment model M when the evaluation reflection information is input to the judgment model M.

[0030] The result output unit 46 outputs the determination result by the determination unit 44 to the input / output device 22. The result output unit 46 displays, for example, information indicating the presence or absence of gravel, which is the determination result by the determination unit 44, on the monitor 24 included in the input / output device 22.

[0031] 3 shows an example of a hardware configuration of the arithmetic device 20. The arithmetic device 20 has a circuit 60. The circuit 60 includes a processor 62, a memory 64, a storage 66, an input / output port 68, and a timer 69. The storage 66 is configured with one or more non-volatile memory devices such as a flash memory or a hard disk. The storage 66 stores a program that causes the arithmetic device 20 to execute the learning process and the evaluation process. For example, the storage 66 stores a program that causes the arithmetic device 20 to configure the above-mentioned functional blocks.

[0032] The memory 64 is composed of one or more volatile memory devices such as a random access memory. The memory 64 temporarily stores a program loaded from the storage 66. The processor 62 is composed of one or more arithmetic devices such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The processor 62 executes the program loaded in the memory 64, thereby configuring each of the above-mentioned functional blocks in the arithmetic device 20. The results of the arithmetic operations by the processor 62 are temporarily stored in the memory 64.

[0033] The input / output port 68 inputs and outputs electrical signals between the measuring device 10 and the input / output device 22, etc., according to instructions from the processor 62. The timer 69 measures the elapsed time, for example, by counting reference pulses at a fixed period. Note that the circuit 60 is not necessarily limited to one that configures each function by a program. For example, at least some of the functions of the circuit 60 may be configured by a dedicated logic circuit or an ASIC (Application Specific Integrated Circuit) that integrates such a logic circuit.

[0034] [Deck deterioration detection method] Next, as an example of a deck deterioration detection method, a detection process executed using the detection device 1 will be described. This detection process (deck deterioration detection method) includes, for example, a learning process executed using a learning road bridge 100A, and an evaluation process for determining the presence or absence of gravel on the deck top surface of a road bridge 100B to be evaluated.

[0035] FIG. 4(a) is a flowchart showing a series of processes executed in the learning process (learning phase). In the learning process, step S11 is first executed by an operator and the computing device 20. In step S11, for example, an operator uses the measuring device 10 to irradiate electromagnetic waves to a set range SA including at least a part of the road surface 101 of the road bridge 100A. FIG. 5(a) shows a schematic plan view of the road bridge 100A and roads 150 connected to both ends of the road bridge 100A. The road bridge 100A is formed to extend in one horizontal direction, for example. In FIG. 5(a), "110a" and "110b" each represent the boundary between the road bridge 100A (road surface 101 of the road bridge 100A) and the road 150.

[0036] The set range SA is arbitrarily set in advance by, for example, an operator. The set range SA may be set to include the entire road surface 101 of the road bridge 100A, or may be set to include a large part (e.g., 80% to 90%) of the entire road surface 101. The set range SA may be set for each section (e.g., for each lane or for each vehicle travel direction) on the road surface 101 of the road bridge 100A. The set range SA may be a rectangular range, or may be a range in which at least a part of the outer edge is inclined or curved in accordance with the shape of the road surface of the road bridge 100A. In the rectangular set range SA, a pair of edges may be set to follow the extension direction of the road bridge 100A. Hereinafter, the extension direction of the road bridge 100A is referred to as the "X-axis direction", and a horizontal direction perpendicular to the X-axis direction is referred to as the "Y-axis direction".

[0037] In step S11, within the set range SA, electromagnetic waves are irradiated to each of a plurality of measurement points that are different in at least one of the positions in the X-axis direction and the Y-axis direction. For example, the plurality of measurement points are set to be arranged in a matrix in the X-axis direction and the Y-axis direction in a plan view. In FIG. 5(a), the number of measurement points in the X-axis direction is indicated by "Xn", and the number of measurement points in the Y-axis direction is indicated by "Yn". In this case, in the set range SA, electromagnetic waves are irradiated to each of a total of (Xn x Yn) measurement points. The interval between adjacent measurement points in the X-axis direction and the interval between adjacent measurement points in the Y-axis direction may be the same. The interval between adjacent measurement points in the X-axis or Y-axis direction may be 40 mm or more, 50 mm or more, or 60 mm or more, or 100 mm or less, 90 mm or less, or 80 mm or less.

[0038] Focusing on one measurement point, the measurement device 10 receives a reflected wave at that measurement point as a result of irradiating the measurement point with an electromagnetic wave. The measurement information acquisition unit 32 of the calculation device 20 acquires a signal corresponding to the reflected wave received by the measurement device 10. The measurement information acquisition unit 32 acquires, for example, a signal indicating the change in intensity of the reflected wave over time (see FIG. 6(b)). The intensity of the reflected wave may be represented by an amplitude (real number) or a complex number.

[0039] Next, step S12 is executed by the calculation device 20. In step S12, for example, the measurement information acquisition unit 32 generates reflection information RD acquired from a reflected wave accompanying irradiation of an electromagnetic wave for each of a plurality of divided regions dr obtained by arbitrarily dividing the set range SA. That is, the measurement information acquisition unit 32 generates reflection information RD for each divided region dr included in the plurality of divided regions dr. FIG. 5(b) shows a schematic relationship between the set range SA and the plurality of divided regions dr.

[0040] As shown in FIG. 5(b), a plurality of divided regions dr may be set so as to be two-dimensionally arranged in M ​​rows and N columns in a plan view. Each of M and N is an integer of 2 or more. M divided regions dr are arranged in the Y-axis direction, and N divided regions dr are arranged in the X-axis direction. In the example shown in FIG. 5(b), the set range SA is divided into (M×N) divided regions dr. The areas of the plurality of divided regions dr may be the same. Each divided region dr may be a square.

[0041] Each divided region dr includes multiple measurement points in the X-axis direction and multiple measurement points in the Y-axis direction. If the number of measurement points in one divided region dr in the X-axis direction is represented as "xn" and the number of measurement points in the Y-axis direction is represented as "yn", then each divided region dr includes (xn x yn) measurement points. xn and yn are the same. Each of xn and yn may be 10 or more, 15 or more, or 20 or more, or 50 or less, 45 or less, or 40 or less.

[0042] As described above, the time change in the intensity of the reflected wave is obtained for each measurement point. Here, the time axis in the time change of the reflected wave is called the "Z axis". Although the Z axis is the time axis that represents the arrival time of the reflected wave accompanying the irradiation of the electromagnetic wave, if the direct wave and the multiply reflected wave are not taken into consideration, the arrival time of the reflected wave changes depending on the depth position of the point where the electromagnetic wave is reflected, so the Z axis can also be said to be the depth direction.

[0043] The upper part of FIG. 6(a) shows the distribution of the intensity of the reflected wave at a certain time in the entire set range SA, and the intensity is indicated by the darkness of the color. The bridge from which the intensity of the reflected wave illustrated in FIG. 6(a) was obtained is a skew bridge, and the set range SA is set approximately along a parallelogram. The lower part of FIG. 6(a) shows the reflection information RD corresponding to one divided area dr. The reflection information RD shows, for example, the relationship between the positions in the X-axis direction and the Y-axis direction (two horizontal directions), the arrival time of the reflected wave caused by the irradiation of the electromagnetic wave, and the intensity of the reflected wave. The positions in the X-axis direction and the Y-axis direction in the reflection information RD may be expressed based on an origin set at the same position (corresponding position) among multiple divided areas dr. In the reflection information RD, the intensity of the reflected wave is obtained for each three-dimensional coordinate on a three-dimensional graph consisting of the position in the X-axis direction, the position in the Y-axis direction, and the arrival time of the reflected wave.

[0044] FIG. 6(b) shows the change over time of the intensity of the reflected wave at one measurement point (for example, a measurement point whose coordinates on the X-axis and Y-axis are 1,1) among the multiple measurement points included in the reflection information RD. The reflection information RD includes data showing the change over time of the intensity of the reflected wave for each two-dimensional coordinate consisting of the X-axis coordinate and the Y-axis coordinate. The Z-axis (horizontal axis) of the graph shown in FIG. 6(b) represents the number of samplings, and the sampling period is 0.34 ns, and data in the range from 0 ns to about 30 ns is shown. The reflection information RD includes, for example, data from 15 to 20 samplings to 65 to 80 samplings, and does not include data in the remaining range. "zn" in FIG. 6(a) and FIG. 6(b) represents the number of data (sampling number) on the Z-axis corresponding to the arrival time of the reflected wave. zn may be about 50 to 60.

[0045] The start point of the range on the Z axis included in the reflection information RD is set, for example, after the time when the direct wave arrives from the transmitting antenna to the receiving antenna of the measuring device 10. The maximum value of the range on the Z axis included in the reflection information RD means the maximum value of the arrival time of the reflected wave in the reflection information RD. The above maximum value may be set in advance by an operator. In one example, the above maximum value is set to the time when the reflected wave arrives from a position that is below the reinforcing bar 108 located at the topmost position in the deck 104 and above the lower surface of the deck 104. In the reflection information RD described above, (xn×yn×zn) voxels are set, and the intensity of the reflected wave is obtained for each voxel.

[0046] Next, step S13 is executed by the arithmetic device 20. In step S13, for example, the correct answer information acquisition unit 34 acquires a correct answer value indicating the presence or absence of gravel on the deck surface for each of the multiple divided regions dr. FIG. 7 shows a schematic diagram visually representing the relationship between the correct answer information CD including the correct answer value for each divided region dr and the reflection information RD. In the correct answer information CD, for example, either "graveled" or "not graveled" is indicated by binary data for each divided region dr. The correct answer information acquisition unit 34 may acquire the correct answer information CD based on input information by a worker via the input / output device 22.

[0047] The correct answer information CD may be information indicating the result of determining whether the road surface 101 of the road bridge 100 has been excavated and whether it has become gravel. In this case, after the asphalt 102 forming the road surface 101 is actually removed, the presence or absence of gravel may be determined by visual inspection by a worker, and the determination result may be input to the arithmetic device 20. The presence or absence of gravel may be evaluated by visual inspection by a worker at multiple points in one divided area dr, and the result may be input to the arithmetic device 20. In this case, when the ratio of the gravel-covered area in the divided area dr to the divided area dr is equal to or greater than a predetermined threshold value (for example, 20% to 40%), the correct answer information acquisition unit 34 may set the correct answer value in the divided area dr to "graveled". When the ratio is smaller than the predetermined threshold value, the correct answer information acquisition unit 34 may set the correct answer value in the divided area dr to "no gravel". As described above, by comparing the percentage of areas evaluated as having turned to gravel with a specified threshold value to determine the correct value, damage in small areas can be matched to the actual state of road surface repairs that are not subject to repair, ensuring stable estimation accuracy.

[0048] Next, step S14 is executed by the arithmetic device 20. In step S14, for example, the learning data generating unit 36 ​​generates learning data DA by associating the reflection information RD with a correct answer value indicating the presence or absence of gravel on the deck top surface for each of the multiple divided regions dr as shown in Fig. 7. The learning data DA includes, for each divided region dr, the reflection information RD obtained in that divided region dr and the correct answer value in that divided region dr as one data set.

[0049] Next, step S15 is executed by the arithmetic device 20. In step S15, for example, the model construction unit 38 constructs a judgment model M by machine learning based on the learning data DA generated in step S14, which outputs the presence or absence of gravel on the deck surface in response to the input of reflection information RD. The judgment model M can also be said to be a model that outputs a result of estimating (or classifying) the presence or absence of gravel in a certain divided area in response to the input of reflection information RD obtained in that divided area. The model construction unit 38 may construct the judgment model M so as to output either information indicating the result of estimating (or classifying) that gravel has been formed or information indicating the result of estimating (or classifying) that gravel has not been formed in response to the input of reflection information RD.

[0050] The model construction unit 38 may execute machine learning using a convolutional neural network based on the learning data DA to construct the determination model M. The convolutional neural network (determination model M) has an input layer, multiple intermediate layers, and an output layer. The model construction unit 38 constructs the judgment model M, which is a trained model, by repeating learning and determining weights in an intermediate layer of the model so that the accuracy rate of the output result by the judgment model M is high. The judgment model M constructed by the model construction unit 38 may be stored in the model holding unit 42.

[0051] 4(b) is a flowchart showing a series of processes executed in the evaluation process (evaluation phase). In the evaluation process, first, step S21 is executed by a worker and the computing device 20. In step S21, for example, a worker uses the measuring device 10 to irradiate electromagnetic waves to an irradiation target range including at least a part of the road surface 101 of a road bridge 100B to be evaluated that is different from the road bridge 100A. The irradiation target range may be set by the worker.

[0052] The irradiation target range may be set so as to be divided into a plurality of divided regions dr1 corresponding to the plurality of divided regions dr in the learning process. The size of the divided region dr1 is the same as the size of the divided region dr. The irradiation of the electromagnetic wave in step S21 is performed with the same settings as step S11 in the learning process, except for the road bridge to be irradiated and the range to be irradiated. For example, the frequency of the electromagnetic wave, the output of the electromagnetic wave, and the interval between adjacent measurement points in the X-axis direction or the Y-axis direction are the same between the learning process and the evaluation process. As in step S11, the measurement information acquisition unit 32 acquires a signal (time change in the intensity of the reflected wave) corresponding to the reflected wave received by the measurement device 10 for each measurement point.

[0053] Next, step S22 is executed by the arithmetic device 20. In step S22, for example, the judgment unit 44 generates, for each divided region dr1 of the irradiation target range, evaluation reflection information RD (evaluation reflection information) obtained by irradiating the divided region dr1 with electromagnetic waves based on the measurement information obtained in step S21. The evaluation reflection information RD is information corresponding to the reflection information RD obtained in the learning process, and is generated with the same settings as the reflection information RD obtained in the learning process. However, while in the learning process, correct answer information on the presence or absence of gravel is associated with the reflection information RD, the evaluation reflection information RD is information in which the presence or absence of gravel is unknown.

[0054] Next, step S23 is executed by the arithmetic device 20. In step S23, for example, the determination unit 44 determines the presence or absence of gravel in one divided region dr1 (region to be evaluated) based on the reflection information RD for evaluation obtained in the divided region dr1 and the determination model M held in the model holding unit 42. In one example, the determination unit 44 inputs the reflection information RD for evaluation obtained in the divided region dr1 to the determination model M and obtains an output from the determination model M to determine the presence or absence of gravel.

[0055] Next, step S24 is executed by the arithmetic device 20. In step S24, for example, the judgment unit 44 judges whether the judgment of the presence or absence of gravel has been performed in all of the divided regions dr1 set in the irradiation target region. In step S24, if it is judged that the judgment of the presence or absence of gravel has not been performed in all of the divided regions dr1 (step S24: NO), the processing executed by the arithmetic device 20 returns to step S23.

[0056] On the other hand, if it is determined in step S24 that the presence or absence of gravel has been determined for all of the divided regions dr1 (step S24: YES), the processing executed by the arithmetic device 20 proceeds to step S25. In step S25, for example, the result output unit 46 outputs information indicating the determination result in step S23 that has been repeatedly executed to the input / output device 22. In one example, the result output unit 46 displays information indicating the presence or absence of gravel for each divided region dr1 in the irradiation target range on the monitor 24 of the input / output device 22. By executing the above series of processing shown in FIG. 4(b), the occurrence of gravel on the deck top surface of the road bridge 100B can be detected (estimated) without actually excavating the road surface 101.

[0057] The series of processes shown in Fig. 4(a) and Fig. 4(b) are merely examples and may be modified as appropriate. In the series of processes, one step and another step may be executed in parallel or in a different order. In the series of processes, a different step may be additionally executed.

[0058] The learning process and the evaluation process were performed using different road bridges, but after obtaining learning data DA using a portion of the road surface of the road bridge 100A (set range SA) and constructing a judgment model M, the presence or absence of gravel can be determined using the judgment model M in areas of the road surface of the road bridge 100A other than the set range SA.

[0059] [Verification example] Next, a verification example will be described in which a judgment model M was constructed using an actual road bridge 100 and the model was evaluated. Two road bridges 100 were used for the verification. For convenience of explanation, one of the two road bridges 100 is referred to as "Bridge A1" and the other is referred to as "Bridge A2". A machine learning model using a convolutional neural network was constructed as the judgment model M. Specifically, a judgment model M having 52 convolutional layers was constructed with reference to a Residual Network (ResNet). In the reflection information RD, xn, which represents the number of data in the X-axis direction, was set to 28, yn, which represents the number of data in the Y-axis direction, was set to 28, and zn, which represents the number of data in the Z-axis direction, was set to 53.

[0060] The reflection information RD and the correct answer value linked to the reflection information RD were treated as one data set, and a part (most part) of the entire data set was used as learning data, and the remaining part was used as verification data. The reflection information RD of the verification data, whose correct answer value is known, was input to the judgment model M, and the judgment result was compared with the correct answer value. In addition, the judgment result of the judgment model M was compared with the correct answer value for each of the following CASE1 to 3, and the judgment accuracy of the judgment model M was confirmed. CASE1: A decision model M is constructed from the learning data for bridge A1, and the model is evaluated using the validation data for bridge A1. CASE2: A decision model M is constructed from the learning data for bridge A2, and the model is evaluated using the validation data for bridge A2. CASE 3: A judgment model M is constructed from the learning data for bridges A1 and A2, and the model is evaluated using the validation data for bridges A1 and A2.

[0061] In addition to the reflection information RD as it is, the training data set includes information obtained by inverting the reflection information RD on the X-axis, Y-axis, or Z-axis. For the training data for bridge A1, 9764 data sets classified as not graveled and 8200 data sets classified as graveled were prepared. For the validation data for bridge A1, 176 data sets classified as not graveled and 192 data sets classified as graveled were prepared. For the training data for bridge A2, 11877 data sets classified as not graveled and 4595 data sets classified as graveled were prepared. For the validation data for bridge A2, 225 data sets classified as not graveled and 75 data sets classified as graveled were prepared.

[0062] The results of verification using verification data for each of CASE 1 to 3 are shown in Table 1 below. [Table 1]

[0063] The accuracy rate, precision rate, and recall rate in Table 1 are calculated by the following formulas (1) to (3), respectively. Correct answer rate=(TP+TN) / (TP+FP+TN+FN) ···(1) Precision rate=TP / (TP+FP) ···(2) Recall rate=TP / (TP+FN) ···(3) In the formulas (1) to (3), the meanings of TP (True Positive), FP (False Positive), TN (True Negative), and FN (False Negative) are as follows: TP: The number of datasets that the decision model M judged to have gravel compared to the correct answer. FP: The number of datasets in which the decision model M judged the data to be gravelly when the correct answer was no gravelly. TN: The number of datasets in which the decision model M judged the sample to be "no gravel" when the correct answer was "no gravel" FN: The number of datasets in which the decision model M judged the data to be not gravelly when the correct answer was that the data was gravelly.

[0064] In Table 1, AUC means the area between the ROC curve, whose vertical axis is TPR (True positive rate) and horizontal axis is FPR (False positive rate), and the horizontal axis. TPR and FPR are calculated by the following formulas (4) and (5). TPR = TP / (TP + FN) (4) FPR = FP / (TP + FN) (5)

[0065] From Table 1, it can be seen that when learning and model judgment were performed using a single bridge, the AUC was greater than 0.95, and high judgment accuracy was obtained (the presence or absence of gravel was appropriately classified).In addition, when learning and model judgment were performed using two bridges, the AUC was greater than 0.90, and it can be seen that in this case too, high judgment accuracy was obtained.

[0066] FIG. 8 shows the result of evaluating which part of the reflection information RD the judgment model M focused on to judge the presence or absence of gravel. Specifically, FIG. 8 shows the result of performing Grad-CAM (Class Activation Mapping) to visualize which part in the Z-axis direction the judgment model M focused on to judge the presence of gravel. The top row of FIG. 8 shows the correct answer information for the bridge A1, where each lattice-like cell is a divided area dr, and the area with a circle (white circle) in the upper left corner of the cell is the area where the correct answer value is judged to be the presence of gravel. Then, by executing Grad-CAM, the degree of influence on the judgment result at three points with different Z-axis values ​​(near the road surface, the upper surface of the deck, and the upper reinforcing bar) is visualized. The upper reinforcing bar is the reinforcing bar 108 located at the top of the deck 104.

[0067] In the three pieces of information visualizing the degree of influence arranged under the correct answer information, the degree of influence of the area determined to have gravel is represented by the darkness of the color. The darker the color, the greater the influence on the judgment model M's judgment that gravel has occurred. In FIG. 8, in the same divided area dr indicated by the symbol "P", the correct answer value is that gravel has occurred, and the output of the judgment model M also judges that gravel has occurred. When focusing on the area indicated by "P", it can be seen that the influence of the reflected information RD from the upper surface of the deck and the upper reinforcing bars is greater than that near the road surface. It is considered that the influence of gravel on the upper surface of the deck is strongly expressed in the reflected waves from the upper surface of the deck or the upper reinforcing bars. Therefore, from the results shown in FIG. 8, it can be seen that the judgment model M bases its judgment on the reflected information RD at the location that is strongly influenced by gravel on the upper surface of the deck.

[0068] [Evaluation of road bridge conditions] As a result of various studies by the present inventor, it was found that it is possible to evaluate the conditions of road bridges to be collected in the future as learning data by constructing a machine learning model using information obtained by wavelet transforming the reflection information RD as input and obtaining a feature map from the model. Examples of road bridge conditions include bridge type, amount of antifreeze used, amount of snowfall, temperature, waterproofing specifications, cross slope, longitudinal slope, drainage facilities, and pavement conditions. Below, the steps executed to evaluate the conditions of road bridges after the learning step and evaluation step in the deck deterioration determination method will be described, including the process of the study.

[0069] FIG. 9 shows a schematic diagram of how a feature map FM (information on features) is obtained from an intermediate layer of a machine learning model using a convolutional neural network. For example, the feature map FM is obtained by the following procedure. For each data set included in a plurality of data sets (a plurality of data sets in which the reflection information RD and the correct answer value are respectively associated) in which the correct answer value is known, the reflection information RD is input to the judgment model M, and high-dimensional (e.g., 512-dimensional) data output from the intermediate layer of the judgment model M is obtained. In one example, high-dimensional data output from the last (most output-side) convolutional layer of the intermediate layers of the judgment model M is obtained.

[0070] After acquiring the high-dimensional data, the high-dimensional data is compressed into two-dimensional or three-dimensional data by a dimension compression algorithm such as t-SNE (Stochastic Neighbor Embedding). The feature map FM shown in FIG. 9 is obtained by compressing the high-dimensional data output from the intermediate layer of the determination model M into two-dimensional data and plotting it on a two-dimensional graph for each data set. In the feature map FM, each data set (plot) may be classified into any one of TP (True Positive), FP (False Positive), TN (True Negative), and FN (False Negative).

[0071] FIG. 10(a) shows a feature map FM1. The feature map FM1 is obtained by constructing a judgment model M (for convenience of explanation, referred to as "judgment model M1") using the bridge A1, and then inputting the reflection information RD of each data set of the learning data obtained at the bridge A1 into the judgment model M1. In the feature map FM1, each plot corresponding to each divided region dr is classified into the following four categories. In the following classifications, "0-0%" indicates that there are no areas in the divided region dr that are evaluated as having gravel, and "30-100%" indicates that the proportion of areas in the divided region dr that are evaluated as having gravel is 30% or more. · train_Bridge_A1_rate=0-0%_FP: Dataset in which the correct answer was no gravel, but the judgment result of the judgment model M1 was gravel. · train_Bridge_A1_rate=0-0%_TN: Dataset in which the correct answer was no gravel, whereas the judgment result of the judgment model M1 was no gravel. · train_Bridge_A1_rate=30-100%_FN: Dataset where the correct answer was that there was gravel, but the judgment result of the judgment model M1 was that there was no gravel. · train_Bridge_A1_rate=30-100%_TP: Dataset in which the correct answer was that there was gravel, whereas the judgment result of the judgment model M1 was that there was gravel.

[0072] FIG. 10(b) shows the feature map FM2. The feature map FM2 is obtained by inputting the reflection information RD of each data set of the learning data obtained at the bridge A2 into the judgment model M1 constructed using the bridge A1. In the feature map FM2, each plot corresponding to each divided region dr is classified into the following six categories. In the following classification, "0-0%" indicates that there is no area evaluated as having gravel in the divided region dr, "0-30%" indicates that the proportion of areas evaluated as having gravel in the divided region dr is more than 0% and less than 30%. Also, "30-100%" indicates that the proportion of areas evaluated as having gravel in the divided region dr is 30% or more. Bridge_A2_rate=0-0%_FP: Dataset in which the correct answer was no gravel, whereas the judgment result of the judgment model M1 was gravel. Bridge_A2_rate=0-0%_TN: Dataset in which the correct answer was no gravel, whereas the judgment result of the judgment model M1 was no gravel. Bridge_A2_rate=0-30%_FP: Dataset in which the correct answer was no gravel, whereas the judgment result of the judgment model M1 was gravel. Bridge_A2_rate=0-30%_TN: Dataset in which the correct answer was no gravel, whereas the judgment result of the judgment model M1 was no gravel. Bridge_A2_rate=30-100%_FN: Dataset in which the correct answer was that there was gravel, whereas the judgment result of the judgment model M1 was that there was no gravel. Bridge_A2_rate=30-100%_TP: Dataset in which the correct answer was that gravel had formed, whereas the judgment result of the judgment model M1 was that gravel had formed.

[0073] When feature map FM1 and feature map FM2 were evaluated by overlapping them, they were plotted in different positions between bridge A1 and bridge A2, and no trend was found depending on the presence or absence of gravel. Therefore, it is presumed that the judgment model M1 constructed using the learning data for bridge A1 has learned the characteristics of bridge A1 itself that are different from the characteristics depending on the presence or absence of gravel, and overlearning has occurred.

[0074] Similar verification was performed by replacing the input information to the machine learning model with information obtained by performing wavelet transform on the reflection information RD (hereinafter referred to as "transformed information"). FIG. 11 shows information obtained by performing wavelet transform on data showing the relationship between the arrival time (Z axis) of the reflected wave and the intensity of the reflected wave in one two-dimensional coordinate on the XY plane consisting of the X axis and the Y axis. That is, when performing wavelet transform on the reflection information RD, data showing the time change of the intensity of the reflected wave is wavelet transformed for each two-dimensional coordinate on the XY plane. When performing wavelet transform, for example, Ricker Wavelet is used as the mother wavelet. As for the frequency domain (frequency axis), in one example, the range is set to 0.2 GHz to 3.0 GHz, and the range is divided into 64. As a result, in one two-dimensional coordinate in the transformed information, data showing the time change of the intensity of the reflected wave is obtained for each of the 64 divided frequency ranges.

[0075] When the converted information is used as input information, a judgment model using a convolutional neural network is constructed so as to output the presence or absence of gravel in response to the input of the converted information. Hereinafter, the judgment model that uses the converted information as input is referred to as a "CNN model" to distinguish it from the judgment models M and M1 when the input information is the reflection information RD.

[0076] FIG. 12(a) shows the feature map FM3. The feature map FM3 is obtained by constructing a CNN model using the bridge A1, and then inputting converted information (first converted information) obtained by converting the reflection information RD of each data set of the learning data obtained at the bridge A1 into the CNN model. When obtaining the feature map FM3 from the CNN model, high-dimensional data is compressed into two-dimensional data, as in the case of obtaining the feature map FM1 from the intermediate layer of the determination model M1. In the feature map FM3, each plot is classified into four categories under the same conditions as in the feature map FM1.

[0077] FIG. 12(b) shows the feature map FM4. The feature map FM4 is obtained by inputting converted information (second converted information) obtained by converting the reflection information RD of each data set of the learning data obtained at bridge A2 into the CNN model constructed using bridge A1. When obtaining the feature map FM4 from the CNN model, high-dimensional data is compressed into two-dimensional data, similar to the case of obtaining the feature map FM2 from the intermediate layer of the judgment model M1. In the feature map FM4, each plot is classified into six categories under the same conditions as in the feature map FM2.

[0078] When feature map FM3 and feature map FM4 were evaluated by overlaying them, compared to when feature map FM1 and feature map FM2 were evaluated by overlaying them, there was a tendency for them to be plotted in similar positions between bridge A1 and bridge A2, and the plot positions tended to differ depending on whether gravel was present or not. In addition, when observing the data group that was erroneously determined to have gravel (data group indicated by FP), the correct answer was plotted in the cluster of the data group that did not have gravel. From the above, it is presumed that when converted information is used as input information for the machine learning model, learning is performed with a greater focus on the characteristics of the presence or absence of gravel.

[0079] In consideration of the above verification, the deck deterioration detection method according to the present disclosure may include a preparation step, a construction step, a first acquisition step, a second acquisition step, and a display step, in addition to the learning step and the evaluation step. The preparation step is a step of preparing another learning data (hereinafter referred to as "learning data DC") corresponding to the learning data DA from a road bridge 100 (hereinafter referred to as "road bridge 100C") different from the road bridge 100A after the determination model M is constructed. In the learning data DC, too, reflection information RD is associated with a correct answer value indicating the presence or absence of gravel for each of a plurality of divided areas dr obtained by arbitrarily dividing a set range including at least a part of the road surface 101 of the road bridge 100C (road bridge C).

[0080] The construction process is a process of constructing a CNN model, which is a convolutional neural network model that outputs the presence or absence of gravel on the deck surface in response to input of first converted information obtained by wavelet transforming reflection information RD in the learning data DA, based on the learning data DA, by machine learning. The CNN model is constructed in the same manner as, for example, constructing the judgment model M, although there are differences in the information used for input.

[0081] The first acquisition step is a step of inputting the first converted information on the road bridge 100A to the CNN model and obtaining information on the feature amount (for example, the above-mentioned feature amount map FM) from the intermediate layer of the CNN model. The second acquisition step is a step of inputting the second converted information obtained by wavelet transforming the reflection information RD in the learning data DC on the road bridge 100C to the CNN model and obtaining information on the feature amount (for example, the above-mentioned feature amount map FM) from the intermediate layer of the CNN model. The computing device 20 may have, as a functional block, a feature amount acquisition unit 48 that executes the first acquisition step and the second acquisition step (see FIG. 2). The display step is a step of displaying the information on the feature amount obtained in each of the first acquisition step and the second acquisition step on the monitor 24.

[0082] 13 shows a series of processes executed to evaluate the condition of a road bridge when updating the judgment model M after constructing the judgment model M using the road bridge 100A. This series of processes is executed using the road bridge 100C. First, step S51 is executed by the calculation device 20. In step S51, for example, the learning data generation unit 36 ​​generates (prepares) learning data DC from a reflected wave obtained by irradiating an electromagnetic wave from the measurement device 10 to the road surface 101 of the road bridge 100C. The learning data generation unit 36 ​​may generate the learning data DC in a manner similar to that for generating the learning data DA.

[0083] Next, steps S52 and S53 are executed by the arithmetic unit 20. In step S52, for example, the feature acquisition unit 48 performs wavelet transform on the reflection information RD in the learning data DA to acquire first converted information, and performs wavelet transform on the reflection information RD in the learning data DC to acquire second converted information. In step S53, for example, the model construction unit 38 constructs a CNN model by machine learning based on the learning data DA, which outputs the presence or absence of gravel on the deck surface in response to the input of the first converted information. The constructed CNN model may be stored by the model holding unit 42.

[0084] Next, steps S54 and S55 are executed by the arithmetic device 20. In step S54, for example, the feature acquisition unit 48 acquires high-dimensional data output from the intermediate layer of the CNN model when the first conversion information is input to the CNN model constructed in step S53. In addition, the feature acquisition unit 48 acquires high-dimensional data output from the intermediate layer of the CNN model when the second conversion information is input to the CNN model constructed in step S53. In step S55, for example, the feature acquisition unit 48 performs dimensional compression on each of the two types of high-dimensional data obtained in step S54 using a dimensional compression algorithm such as t-SNE. As a result, a feature map FM as shown in FIG. 12(a) and FIG. 12(b) is obtained.

[0085] Next, step S56 is executed by the arithmetic device 20. In step S56, for example, the result output unit 46 displays the two types of feature amount maps FM for the road bridge 100A and the road bridge 100C obtained in step S55 on the monitor 24 of the input / output device 22. This allows the worker to understand the two types of feature amount maps FM and perform evaluation using those feature amount maps FM.

[0086] In one example, when the worker observes two types of feature amount maps FM superimposed on each other and finds a non-overlapping data group (a data group plotted away from other clusters), the worker can consider the conditions of the road bridge for which learning data should be collected in the future by investigating the characteristics of the divided area dr from which the data group was obtained. In one example, when the worker observes two types of feature amount maps FM superimposed on each other and finds that they are almost overlapping (similar), the worker can evaluate that the difference in the road bridge conditions between the road bridge 100A and the road bridge 100C has already been learned or does not affect the determination of the presence or absence of gravel.

[0087] In one example among the various examples described above, at least a part of the matters described in the other examples may be combined.

[0088] [Summary of this disclosure] The deck deterioration detection method described above includes a learning step and an evaluation step. The learning step includes irradiating an electromagnetic wave to a set range SA including at least a part of the road surface 101 of the road bridge 100A, and for each of a plurality of divided areas dr obtained by arbitrarily dividing the set range SA, generating learning data DA by associating three-dimensional reflection information RD obtained from the reflected wave accompanying the irradiation of the electromagnetic wave with a correct answer value indicating the presence or absence of gravel on the deck surface, and constructing a judgment model M by machine learning based on the learning data DA, which outputs the presence or absence of gravel on the deck surface in response to the input of the reflection information RD. The evaluation step includes judging the presence or absence of gravel on the deck surface in the evaluation target area based on the evaluation reflection information RD obtained by irradiating an electromagnetic wave to an evaluation target area on the road surface 101 of the road bridge 100A or a road bridge 100B different from the road bridge 100A, and the judgment model M.

[0089] Unlike the above-mentioned deck deterioration detection method, a method is conceivable in which a worker looks at information obtained from the reflection of electromagnetic waves inside the deck or information obtained by exploring the inside using infrared thermography to diagnose the deterioration of the deck top surface. In this method, the worker needs to be skilled to make the diagnosis, and it takes time to perform the diagnosis. In particular, when using information obtained from the reflection of electromagnetic waves, it is necessary to estimate the abnormal part while generating and comparing the reflection information in multiple horizontal and vertical cross sections from the reflection information configured in three dimensions. In this way, it tends to be difficult to process a large amount of information manually and accurately judge the deterioration by the judgment of the worker. In contrast, in the above-mentioned deck deterioration detection method, the set range SA where the electromagnetic waves are irradiated is divided into multiple divided areas dr, learning data is prepared, and a judgment model M is constructed by machine learning using information on the reflected waves in one divided area dr. Then, the judgment model M is used to judge whether or not gravel has been formed in the evaluation target area. This eliminates the need for judgment by the worker, and simplifies the processing of data by the worker. In addition, by using the three-dimensional reflection information RD as input information, the device can determine the presence or absence of gravel using the same method as when a worker performs the diagnosis. Furthermore, by dividing the set range SA into multiple divided regions dr, various learning data sets can be obtained from a single road bridge. From the above, it is possible to simplify the work of detecting deterioration while improving detection accuracy.

[0090] The reflection information RD may indicate the relationship between the position in each of two horizontal directions, the arrival time of the reflected wave accompanying the irradiation of the electromagnetic wave, and the intensity of the reflected wave. It is also possible to determine the presence or absence of gravel by analyzing the time change in the intensity of the reflected wave at one measurement point on the road surface 101. In contrast, in the above method, the presence or absence of gravel is determined using a set of time changes in the intensity of the reflected wave over a certain range, rather than using local information on the road surface 101. This makes it possible to improve the detection accuracy.

[0091] In the reflection information RD, the maximum arrival time of the reflected wave associated with the irradiation of the electromagnetic wave may be set to the time at which the reflected wave arrives from a position below the reinforcing bar 108 located at the top in the deck 104 and above the bottom surface of the deck. When gravelization occurs on the top surface of the deck, the condition of the reinforcing bar 108 located at the top in the deck 104 may also be different from normal, so the detection accuracy may be improved by using information associated with the reflected wave from the reinforcing bar 108 to determine whether gravelization occurs. On the other hand, it is considered that the reflected wave from below the reinforcing bar 108 does not change much depending on whether gravelization occurs. From the above, the above method makes it possible to achieve both improved detection accuracy and reduced data processing load.

[0092] The learning data DA may include, as a correct answer value, information indicating the result of determining whether the road surface 101 in the set range SA has been excavated and turned into gravel for each of the multiple divided regions dr. In this case, since the road surface 101 has actually been excavated, a more accurate correct answer value can be obtained. Then, by using the learning data including such a correct answer value, it is possible to improve the judgment accuracy of the judgment model M.

[0093] The above deck deterioration detection method may further include, after constructing the judgment model M, a preparation step of preparing another learning data DC corresponding to the learning data DA from a road bridge 100C different from the road bridge 100A; a construction step of constructing a CNN model, which is a convolutional neural network model that outputs the presence or absence of gravel on the deck top surface in response to input of first converted information obtained by wavelet transforming the reflection information RD in the learning data DA, by machine learning based on the learning data DA; a first acquisition step of inputting the first converted information into the CNN model and obtaining information related to the feature quantity from an intermediate layer of the CNN model; a second acquisition step of inputting second converted information obtained by wavelet transforming the reflection information RD in the learning data DC into the CNN model and obtaining information related to the feature quantity from an intermediate layer of the CNN model; and a display step of displaying the information related to the feature quantity obtained in each of the first acquisition step and the second acquisition step on a monitor 24.

[0094] It is not easy to actually excavate the road surface of a road bridge and investigate whether it has become gravel. Therefore, when a judgment model M is constructed using a certain road bridge and then learning data is obtained from other road bridges to update the judgment model M in order to increase the versatility of the model, it is important to consider in advance what conditions of the road bridge to use to obtain the learning data. As described above, it has been found that the CNN model, which uses information obtained by wavelet transforming the reflection information RD as input, may be learning by focusing more on the characteristics of gravel. Therefore, by displaying information on the feature amounts obtained in the first acquisition step and the second acquisition step by the above method on the monitor 24, it is possible to evaluate the conditions of road bridges to be collected in the future as learning data.

[0095] The detection device 1 described above comprises a measuring device 10 that irradiates electromagnetic waves onto a set range SA including at least a part of the road surface 101 of the road bridge 100A, a learning data generation unit 36 ​​that generates learning data DA by associating three-dimensional reflection information RD obtained from the reflected waves accompanying the irradiation of the electromagnetic waves with a ground truth value indicating the presence or absence of gravel on the top surface of the deck for each of a plurality of divided areas dr obtained by arbitrarily dividing the set range SA, a model construction unit 38 that constructs by machine learning a judgment model M that outputs the presence or absence of gravel on the top surface of the deck in response to input of reflection information RD based on the learning data DA, and a judgment unit 44 that judges the presence or absence of gravel on the top surface of the deck in the above-mentioned evaluation target area based on the judgment model M and the evaluation reflection information RD obtained by irradiating electromagnetic waves onto an evaluation target area on the road surface 101 of the road bridge 100A or a road bridge 100B different from the road bridge 100A. Like the deck deterioration detection method described above, this detection device 1 is capable of improving detection accuracy while simplifying the work required to detect deterioration. [Explanation of symbols]

[0096] 1...detection device, 10...measuring device, 20...computing device, 22...input / output device, 24...monitor, 34...correct answer information acquisition unit, 36...learning data generation unit, 38...model construction unit, 44...judgment unit, 100...road bridge, 101...road surface, 104...deck slab, M...judgment model, SA...setting range, dr...divided area, DA...learning data, RD...reflection information.

Claims

1. A deck deterioration detection method including a learning process and an evaluation process, The learning step includes: Irradiating electromagnetic waves to a set range including at least a part of a road surface of a road bridge A; For each of a plurality of divided regions obtained by arbitrarily dividing the set range, three-dimensional reflection information obtained from the reflected waves accompanying the irradiation of the electromagnetic waves is associated with a correct answer value indicating the presence or absence of gravel on the upper surface of the deck, thereby generating learning data DA; and constructing a determination model by machine learning based on the learning data DA, which outputs the presence or absence of gravel on the deck surface in response to the input of the three-dimensional reflection information; The evaluation step includes: A deck deterioration detection method including determining whether or not gravel has formed on the deck surface in an evaluation target area on the road surface of the road bridge A or a road bridge B different from the road bridge A, based on evaluation reflection information corresponding to the three-dimensional reflection information obtained by irradiating electromagnetic waves onto the evaluation target area on the road surface of the road bridge A, or a road bridge B different from the road bridge A, and the judgment model.

2. The deck deterioration detection method described in claim 1, wherein the three-dimensional reflection information indicates the position in each of two horizontal directions, and the relationship between the arrival time of the reflected wave associated with the irradiation of electromagnetic waves and the intensity of the reflected wave.

3. The deck deterioration detection method described in claim 2, wherein in the three-dimensional reflection information, the maximum value of the arrival time of the reflected wave associated with the irradiation of the electromagnetic wave is set to the time at which the reflected wave arrives from a position below the highest rebar in the deck and above the underside of the deck.

4. The deck deterioration detection method described in claim 1, wherein the learning data DA includes, as the correct value, information indicating the result of determining whether or not the road surface in the set range has been excavated and turned into gravel for each of the multiple divided areas.

5. a preparation step of preparing, after the construction of the judgment model, another learning data DC corresponding to the learning data DA from a road bridge C different from the road bridge A; A construction process of constructing a CNN model, which is a convolutional neural network model that outputs the presence or absence of gravel on the deck surface in response to input of first converted information obtained by wavelet transforming the three-dimensional reflection information in the learning data DA based on the learning data DA, by machine learning; A first acquisition step of inputting the first converted information into the CNN model to obtain information about features from an intermediate layer of the CNN model; A second acquisition step of inputting second converted information obtained by performing a wavelet transform on the three-dimensional reflection information in the learning data DC to the CNN model and acquiring information related to features from an intermediate layer of the CNN model; A deck deterioration detection method described in any one of claims 1 to 4, further comprising a display process for displaying information regarding the features obtained in each of the first acquisition process and the second acquisition process on a monitor.

6. A measurement unit that irradiates electromagnetic waves to a set range including at least a part of the road surface of the road bridge A; a learning data generating unit that generates learning data DA by associating three-dimensional reflection information obtained from the reflected waves caused by the irradiation of electromagnetic waves with a correct answer value indicating the presence or absence of gravel on the upper surface of the deck for each of a plurality of divided regions obtained by arbitrarily dividing the set range; A model construction unit that constructs a determination model by machine learning based on the learning data DA, which outputs the presence or absence of gravel on the deck surface in response to the input of the three-dimensional reflection information; A deck deterioration detection device comprising: a judgment unit that judges whether or not the deck surface in the evaluation target area has become gravel based on evaluation reflection information corresponding to the three-dimensional reflection information obtained by irradiating electromagnetic waves onto an evaluation target area on the road surface of the road bridge A or a road bridge B different from the road bridge A, and the judgment model.

Images