JP7304060B2 - Structural internal deformation characteristics detector - Google Patents

Description

In general, the present invention not only determines the presence or absence of internal deformation that exists inside a structure such as concrete (hereinafter referred to as a structure), but also determines the internal deformation characteristics such as its position, shape, range, type, and physical properties. Apparatus for detecting based on the relationship between a stimulus acting on a structure and the response thereto. More specifically, the present invention relates to an apparatus and method characterized by creating training data on the relationship between the internal deformation characteristics of a structure subjected to a specific stimulus and its response.

Non-destructive detection of characteristics such as presence/absence, position, range, type, etc. of internal deformation, which cannot be observed from the outside because it exists inside the structure and there is no means to easily grasp its physical characteristics. To do so, it is common practice to analyze the response obtained by artificially applying a physical stimulus external to the structure or under the stimulus of natural physical action. The simplest stimulus is sunlight heating, and more advanced stimuli include natural microtremors, blows, vibrations, artificial heating, ultrasonic waves, X-ray transmission, and the like.

For example, if the object is a relatively small-scale structure such as a human body or a compact object, the internal conditions can be accurately determined by measuring it in a well-equipped closed space using X-rays, ultrasonic waves, etc. Since it can be observed in detail, it is relatively easy to detect internal deformation characteristics. At the other end of the spectrum, however, when the object is a large-scale structure that constitutes social capital in a natural environment, it is usually not possible to directly observe the inside, and the surface of the structure cannot be observed in response to a relatively simple stimulus. It is practical to measure only the response at . Although humans can recognize the presence or absence of internal deformation just by looking at such surface response data, it is practically impossible to identify even the characteristics of internal deformation.

Therefore, there is a need for a means of classifying and identifying internal deformation characteristics from the obtained surface response through some logic. As one of such means, it is conceivable to obtain the internal deformation characteristics from the surface response by inverse analysis using a computer. It is necessary and not quite practical from a cost-effective point of view when the subject is large and mass.

Currently, deep learning (hereinafter referred to as DL) has been developed. It is conceivable to use such a DL as a means of relating surface response data of structures obtained to specific stimuli with characteristics of internal deformation. If this detection means is used, the internal deformation cannot be visualized in detail by X-rays or the like, and the internal deformation cannot be determined with specific numerical values by inverse analysis. From the point of view of effectiveness, it is thought that the detection of internal deformations in large structures of general public infrastructure can be improved, which can only be dealt with by visual inspection and hammering inspection. However, in order to use DL for recognizing and identifying internal deformations of such large infrastructure structures, it is necessary to solve the following problems.

Fig. 1 compares the DL for general image recognition and identification with the DL for detection of deformation characteristics inside the structure targeted by the present invention, using image recognition and identification as an example. is shown in the left half of the figure and the latter in the right half. As shown in the left half of FIG. 1, in general image recognition/identification, teacher data is usually composed of image data that allows humans to visually identify objects. In other words, objects to be recognized/identified contained in images of training data, such as people, animals, automobiles, structures, characters, etc., may have a wide variety of individual forms, but their morphological characteristics are the same. , and it is possible to directly acquire the features of the object visually. Moreover, such image data can be easily collected, stored, and processed in large quantities with high precision, thanks to advances in related technology. Therefore, even if part of the object is hidden in the image or the appearance varies due to different viewing angles, the morphological characteristics and appearance of each object and the object can be paired. Basic information or methods for associating with 1 are accumulated. Even if it is not directly visible in this way, it is possible to indirectly visualize it by means of an image obtained using X-rays or ultrasonic waves, for example, when targeting internal organs of the human body. In addition, since there is generally a large amount of anatomical information accumulated over many years in the background, the preparation of training data does not differ greatly from general image recognition/identification. Also, in the case of large-scale structures of social infrastructure, high precision is required for the recognition and identification of phenomena that appear on the surface, such as cracks that appear on the surface of concrete structures and rust that exists on the surface of steel structures. It is thought that this can be handled relatively easily by applying an image from a CCD camera or the like and general image authentication/identification technology.

In addition, in the case of DL that uses wave/vibration data such as voice, in general DL, it is possible to associate the wave/vibration data with the characteristics of the source of the data. The same can be said.

However, in the work of recognizing and identifying internal deformations from surface responses, such as in the case of large-scale structures that make up social infrastructure, human beings cannot see the response data, as mentioned in the background art section. Although it is possible to recognize the presence or absence of internal deformation only by the measurement, it is practically impossible to identify even the internal deformation characteristics. More specifically, as shown in the right half of FIG. 1, the stimulus applied to a large structure generally includes natural heating, impact, vibration, artificial heating, and ultrasonic waves. , wave distribution, displacement distribution, temperature distribution, etc., appearing at each point on the surface of the structure at a certain point in time or over time. It can only detect the local presence of outlier data related to internal deformations in relatively stable data in relatively large temporal or spatial regions without further speculation about the details of the internal deformations. is virtually impossible.

Therefore, when creating teacher data on the relationship between the internal deformation characteristics of a structure that received a specific stimulus and the response, the mere existence of response data is enough to create teacher data that associates the corresponding internal deformation characteristics with the response data. cannot be created.

In addition, even if it is not the training data of the target structure itself, if there is training data of a similar structure in the past, it is possible to make an analogy from it, but the internal deformation of the large-scale structure of social capital has become a problem and time is short. Therefore, there is almost no measured data relating the internal deformation characteristics of a structure to a response to a specific stimulus. Even if they do exist, the physical and spatial characteristics inside a large-scale structure built in the natural environment as described above are uneven due to the influence of the materials used, the construction method, and the construction conditions. In addition, if such a structure is subject to the effects of the natural environment over time and causes mechanical, chemical, etc. causes, such as destruction, damage, deformation, chemical change, corrosion, etc., to occur over time, Since the diversity of deformation is multiplied by that, it is considered almost impossible at present to divert between different structures in the first place. In the future, even if we can obtain a large amount of measured data on the response to some kind of stimulus in a large structure, in order to enrich the training data corresponding to such diversity, we will continue Since it is practically impossible to examine the internal deformation in detail by extracting the internal deformation locations from large structures that are sometimes inaccessible because they are buried in the ground, training data on the characteristics of internal deformation is still incomplete. As a countermeasure, it is conceivable to collect training data through experiments using test specimens that can clearly set internal deformation and stimulus characteristics, but due to cost considerations, the number of internal deformation characteristics that can be set is limited. Compared to the diversity of structures, it is nothing more than an insignificant thing, and even if it can be used as a supplement, it seems that in many cases it cannot be a fundamental solution.

In this way, in the application of DL to the recognition and identification of internal deformation of large structures, which is the target of the present invention, it is necessary to create training data such as in the recognition and identification of general images or wave / vibration data. is practically impossible, so it is necessary to devise a new method of creating teacher data.

Regarding the above problems, at present, it is realistically possible to physically relate the relationship between the response when a structure having various internal deformations receives a stimulus and the internal deformation relatively correctly. Since it is convenient to use analysis such as finite element analysis for example, it is conceivable to create part or all of the basic teaching data by analysis. That is, (a) the structural characteristics of the structure to be detected, (b) the characteristics of the internal deformation of the detection target, (c) the stimulus given to the structure, etc. are set as a physical model, and (d) the structure for the stimulus is set. analytically find the response of The analysis at this time is not the so-called inverse analysis of identifying the internal deformation characteristics from the response as described above, but the forward analysis of obtaining the response from the physical model. It has a track record of being widely used in research and design practice. Therefore, it is considered that the individual training data obtained in this way are almost guaranteed to have practical physical consistency except for special cases. In addition, if analysis technology progresses further in the future, it will not only be possible to improve the accuracy of analysis, but also increase the number of parameter types to broaden the range of their fluctuations, and bring the training data closer to reflecting the diversity of reality. It is possible.

However, as shown in the following embodiments, the number of parameters that make up the analysis model is large, and at present there is a practical limit to setting the values of each parameter in consideration of even realistic variations. do. Therefore, here, while aiming to improve efficiency and rationalization by reducing the number of analysis cases as much as possible, the following methods of creating training data that can deal with the diversity of actual target structures and internal deformations are proposed. I also suggest adding more.
(a) From the measurement data on the response obtained when the target structure was actually measured, extract the part that is considered to be highly related to the internal deformation that is assumed to be highly likely to exist, and measure them A portion of the data is added to the teacher data in relation to the internal deformation. As a result, it is possible to add data reflecting the relationship between the actual internal deformation occurring in the target structure and the response to the teacher data.
(b) For at least some of the internal deformation occurrence locations of the target structure corresponding to the part of the measurement data extracted in (a), conduct a more detailed investigation of internal deformation characteristics than other measurement locations. By doing so, the relationship between the characteristics of the actual internal deformation and its response is more accurately reflected in the training data, improving the reliability of the training data. It is conceivable to improve the accuracy of DL by using teacher data created in this way.

In addition, in large structures such as those that constitute social capital, measurable responses are often limited to responses on the surface of the structure. is not limited to surface responses.

It is a schematic diagram comparing the general DL and the DL for the structure targeted by the present invention. 1 is a schematic diagram showing the configuration of an apparatus according to an embodiment of the present invention; Fig. 2 is a schematic diagram schematically showing the growth of two internal deformations present inside the structure over time and the corresponding relationship between the respective surface responses appearing on the measurement surface; 3 is a schematic diagram showing the contents of data processing in each means constituting the apparatus shown in FIG. 2 and the flow of data between the means; FIG. 4 is a schematic representation of a flow chart of a method for detecting characteristics of internal deformations from responses, according to one embodiment of the present invention; FIG. An example of a crack analysis model assuming the presence of a single crack in the structure is shown. FIG. 7 is a perspective view showing a typical surface temperature distribution on the upper surface in the crack analysis model shown in FIG. 6; In the figure, (a) shows the case of front-side heating in which the upper surface is heated, and (b) shows the case of back-side heating in which the lower surface is heated. The temperature distributions on the vertical bisector of the crack opening are shown for three cases in which the inclination angle δ in the crack analysis model shown in FIG. 6 is 30°, 45° and 60°. In the figure, (a) shows the case of front-side heating in which the upper surface is heated, and (b) shows the case of back-side heating in which the lower surface is heated. An example of a float analysis model in which a float exists inside the structure is shown. FIG. 10 is a perspective view showing a typical surface temperature distribution on the upper surface in the float analysis model shown in FIG. 9; In the figure, (a) shows the case of front-side heating in which the upper surface is heated, and (b) shows the case of back-side heating in which the lower surface is heated. The temperature distribution on the center line of the float is shown for three cases where the depth e is 10 mm, 30 mm and 50 mm in the float analysis model shown in FIG. In the figure, (a) shows the case of front-side heating in which the upper surface is heated, and (b) shows the case of back-side heating in which the lower surface is heated. FIG. 7 shows the surface temperature distribution to which noise is added when the inclination angle δ=45° in the crack analysis model shown in FIG. In the figure, (a) shows the case of front-side heating in which the upper surface is heated, and (b) shows the case of back-side heating in which the lower surface is heated. FIG. 9 shows the surface temperature distribution to which noise is added when the depth e=30 mm in the float analysis model shown in FIG. In the figure, (a) shows the case of front-side heating in which the upper surface is heated, and (b) shows the case of back-side heating in which the lower surface is heated. The result of applying this detection method to the surface temperature distribution of the surface temperature of the crack model when δ = 45° and the front side is heated and does not include noise is shown. For reference, the crack analysis model is shown on the left. The results of applying this detection method to the surface temperature distribution of the crack model when δ = 45° and the back side is heated without noise are shown. The result of applying this detection method to the surface temperature distribution of the surface temperature of the crack model when δ = 45° and the front side is heated and includes noise is shown. For reference, the crack analysis model is shown on the left. The result of applying this detection method to the surface temperature distribution of the crack model when δ = 45° and the backside heating includes noise is shown. Fig. 3 shows the positive sine-related and positive cosine-related components of the detection results for front side heating for a multi-crack analytical model with four different tilt angles; An orthogonal crack analysis model is shown in which two cracks with δ = 45° are orthogonal. FIG. 19 shows positive sine wave-related components and positive cosine wave-related components of the detection results for front side heating in the case of the orthogonal crack analysis model shown in FIG. 19 . FIG. 9 shows the detection result when the floating model shown in FIG. 9 is heated on the front side and includes noise. FIG. 9 shows the detection results when the floating model shown in FIG. 9 is subjected to backside heating and noise is included. 1 shows a schematic diagram of a crack test specimen with simulated cracks; FIG. FIG. 23 shows the surface temperature distribution on the surface of the crack test specimen shown in FIG. 23 when the surface on which the simulated crack is provided is placed directly above and heated by solar radiation on a fine day. In the figure, (a) is an RGB image of the surface temperature distribution itself, and (b) is an image showing the result of normalization by the maximum absolute value of the result of dividing the average temperature from the surface temperature distribution. FIG. 24(a) shows detection results of the surface temperature distribution forming the RGB image. FIG. 24(b) shows the detection result after performing noise removal in advance on the surface temperature distribution forming the RGB image. It is an RGB image and a photograph showing the surface temperature distribution on the lower surface of the floor slab when the lower surface of the RC floor slab of the actual bridge is measured by an infrared imaging device on a sunny day in winter. FIG. 27 shows the result of applying this detection method to the surface temperature distribution of the upper region R1 surrounded by the rectangular solid line in the RGB image shown in FIG. FIG. 27 shows the result of applying this detection method to the surface temperature distribution of the intermediate area R2 surrounded by the rectangular solid line in the RGB image shown in FIG. FIG. 27 shows the result of applying this detection method to the surface temperature distribution of the lower region R3 surrounded by the rectangular solid line in the RGB image shown in FIG. FIG. 4 is an explanatory diagram of a method of detecting internal deformation by striking an end portion of a rod-shaped structure including internal deformation; FIG. 32 shows an example of acceleration data actually measured by an acceleration sensor at the end of an actual rod-shaped structure similar to the rod-shaped structure of FIG. 31 when the end is hit. FIG. 31 shows the results obtained by analysis of acceleration data when the upper end of the rod-shaped structure is hit, the vibration propagates through the rod-shaped structure and returns to the upper end again. (a) is the case without internal deformation, and (b) is the case with internal deformation in the central portion of the rod-like structure. FIG. 32 shows the result of applying noise removal by wavelet degeneracy using multi-resolution analysis with orthogonal wavelets to the actually measured acceleration data shown in FIG. FIG. 24 shows the approximate locations of cracks detected using the DL technique from the actually measured surface temperature distribution data of the specimen shown in FIG. FIG. 27 shows the approximate locations of cracks detected using the DL technique from the surface temperature distribution data actually measured on the lower surface of the RC floor slab shown in FIG.

1 Structural internal deformation detection device according to the present invention
1.1 Configuration of Apparatus FIG. 2 shows an example of a specific configuration of an apparatus 100 for detecting internal deformation existing on the surface or inside a structure from a response to a specific stimulus, for carrying out the present invention. The arrows in the figure generally represent the flow of data between each means.

2, apparatus 100 includes actual measurement means 101, structure analysis means 102, experiment means 103, actual measurement data storage means 104, teacher data storage means 105, teacher data extraction means 106, estimation means 107, a display means 108; The actual measurement means 101 is means for measuring at least the response value when the target structure receives some stimulus and responds. An acoustic sensor or the like that measures sound is conceivable. As the structural analysis means 102, for example, general finite element analysis software capable of executing heat transfer analysis, mechanical analysis, acoustic analysis, etc. can be considered. As the experimental means 103, for example, it is conceivable that an internal deformation of a known specification is set in a full-size or reduced-size simulated test body of a part or the whole of the target structure. As the actual measurement data storage means 104, a computer HDD, SDD, or the like that stores the actual measurement data acquired by the actual measurement means 101, the target structure specifications, the measurement conditions, and the like can be considered. The teacher data extracting means 106 extracts, from at least a part of the measured data stored in the measured data storage means 104, a part of the measured data evaluated to be strongly related to the response of the internal deformation to be detected. The details of this are described later. The teacher data storage means 105 is a computer that stores teacher data extracted from at least a part of the measured data via the teacher data extraction means 106, and teacher data created by the structure analysis means 102 and the experiment means 103. HDD, SDD, etc. can be considered. The estimating means 107 uses the teacher data stored in the teacher data storing means 105 for each of the sections forming at least a part of the measured data stored in the storing means 104 to estimate the internal deformation of the object to be detected. It is a means of recognizing and discerning the relevance of presence-based responses. As the display means 108, a display or the like for displaying the results obtained by the estimation means 107 can be considered.
1.2 Internal Deformation Detection Procedure Next, the contents of data processing in each means constituting the device shown in Fig. 2 and the flow of data between means are as follows, assuming a specific internal deformation and measurement method. will be described more specifically.

In general, when assuming internal deformations to be detected, it should be noted that considering a wide variety of internal deformations from the beginning makes the problem more complicated than necessary. Therefore, it is important to start with the detection of internal deformations that have a small number of basic characteristics that are likely to occur in the target structure based on engineering experience, especially in the early stages of deterioration progression. Efficiency is often good for inspection level detection. Another point to note is that the internal deformation does not remain constant, but changes over time, especially as it expands in scale.

To explain these more specifically, for example, FIG. Schematically shows surface responses, such as surface temperature distribution, that appear corresponding to respective internal deformations on the measurement surface of the lower surface of the flat plate structure when sunlight is applied. The left side of FIG. 3(a) shows a cross-section of the structure at an early stage of deterioration progression, and corresponding to the presence of two spaced apart internal deformations, the surface responses are also independent of each other. On the other hand, the right side of FIG. 3(a) shows the stage where the deterioration has progressed. Their relationship is complicated. FIG. 3(b) is a plan view of the measurement surface of the flat plate structure and schematically shows surface responses appearing on the measurement surface corresponding to two internal deformations. The AA cross section shown on the left side of FIG. 3(b) corresponds to the left side of FIG. 3(a), and the BB cross section shown on the right side of FIG. 3(b) corresponds to the right side of FIG. 3(a). Thus, in the initial stage when the growth of internal deformation is not very advanced, the temperature distributions associated with each internal deformation are thought to be scattered almost independently. There is a high possibility that the characteristics such as the existence position of each internal deformation can be detected only by the relationship between and the response. However, as the growth of each internal deformation progresses, the interval between the response distributions related to the internal deformation becomes narrower, and when there appears a place where both are interfering with each other or overlap each other, multiple internal deformations occur. It seems necessary to consider the influence of the response of However, in detecting the level of inspection investigation at the initial stage of progress of deterioration, it is often sufficient to detect the state in which each internal deformation exists independently.

Therefore, in the following explanation, two internal deformations exist independently in one structural section, and the location where the internal deformation occurs at the initial stage when the growth of the internal deformation has not progressed much will be described. A procedure for segmentation and detection will be described below. In the detection of internal deformation, when the upper surface and/or the lower surface of this structural section is stimulated, the response may be measured at the upper surface and/or the lower surface of the structural section. Suppose a stimulus is applied and the response is measured on the lower surface. Also, the measured response may be a two-dimensional distribution and/or temporal variation at several points of the structure.

Assuming that the internal deformations of interest are two independent internal deformations 1 and 2, as shown on the left side of FIG. 3(a) and on the left side of FIG. 3(b), such For example, as shown in the upper left part of FIG. and 302 are subjected to model stimuli 401 and 402 (e.g., heating, impact, etc.) that model actual stimuli, and model responses 601 and 602 (e.g., temperature distribution, vibration, etc.) are obtained by the structural analysis means 102. It depends. In this case, the drawings show different types of internal deformation, but the same type of internal deformation may be used.

Data in which the model internal deformation 501 and the model response 601 and the model internal deformation 502 and the model response 602 are associated are stored in the teacher data storage means 105 together with analysis information such as analysis setting conditions. As described above, in the initial stage when the growth of internal deformation has not progressed much, each internal deformation is often scattered independently. A relatively simple and independent specification that takes into account typical or average characteristics may also be acceptable. However, if it is not enough to deal with the actual phenomenon, or if the growth of the internal deformation progresses further, the types of model internal deformation are increased sequentially and/or each model internal deformation is configured. It is conceivable to take a step-expansion approach to widen the range of parameter values.

On the other hand, as shown in the upper right part of FIG. , the actual response data is stored in the actual measurement data storage means 104 together with the measurement information such as the structural dimensions of the actual structure of interest.

Optionally, a range of responses considered to have a strong correlation with the responses related to the assumed internal deformation 1 and internal deformation 2 is selected from the actual response data stored in the actual measurement data storage means 104 as a teacher. The teacher data storage means 105 additionally stores the data extracted by the data extraction means 106 and associated with the internal deformation of the response data in these ranges. For example, when it is evaluated that there is a strong relationship between the actual response 1001 and the model response 601 of the model internal deformation 501, the actual internal deformation 901 that caused the actual response 1001 is similar to the model internal deformation 501. Considering that, the actual response 1001 is stored in the teaching data storage means 105 as part of the teaching data of the internal deformation 1 . If the relationship between the actual response 1002 and the model response 602 is the same, then the actual response 1002 is stored in the teaching data storage means 105 as part of the training data for the internal deformation 2 . By this operation, it is considered that the relationship between the actual internal deformation existing in the target real structure and the related actual response can be reflected in the training data. Optionally, when data corresponding to teacher data is created by the experiment means 103 assuming the real structure, real internal deformation, real stimulus, etc., this data is added to the teacher data storage means 105. Remember.

Training data created based on such structural analysis, actual measurements, and experiments cannot necessarily be obtained in a large amount and as easily as training data for general image recognition, so if you are not satisfied with both quality and quantity There is Therefore, data expansion methods used in DL (rotating, scaling, translating, shearing, blurring, adding noise, etc.) are used to increase the amount of these teacher data and diversify the teacher data. Sometimes it is useful to add gender.

Using the teacher data stored in the teacher data storage means 105 and the measured data and the like stored in the measured data storage means 104 obtained as described above, learning, evaluation, etc., are executed in the estimation means 107. By the operation of , the portion of the actual response considered to be related to the assumed internal deformation is recognized and guessed from the actual response, and displayed on the display means 108 in correspondence with the actual structure. At this time, it is conceivable to display information about the type of internal deformation that is highly likely to be associated with the range of the actual response, the possibility of its existence, and the like.

FIG. 5 shows an example of a flow chart describing the main steps out of the above steps, and each step will be specifically described based on this flow chart. In step 1 (S1), for example, the surface temperature of the actual structure whose characteristics such as the position, range, physical properties, and type of internal deformation are detected, is subjected to a specific actual stimulus such as heating, impact, vibration, and ultrasonic waves. etc. are measured as actual response data by actual measurement means. In step 2 (S2), the actual response data is stored in the actual measurement data storage means together with the measurement information. In step 3 (S3), one or more internal deformations to be detected that are considered to exist in the target structure are assumed based on engineering experience and judgment, and a model corresponding to the assumed internal deformations When a model stimulus is applied to a model structure including internal deformation, analysis is performed by the structural analysis means to obtain model response data. At step 4 (S4), the model response data is stored in the teacher data storage means together with the analysis information.

At this time, an optional step 10 (S10) is additionally performed to extract, from the actual response data, portions that can be evaluated as having a strong relationship with the response corresponding to the assumed internal deformation, and update them. may additionally be stored in the teacher data storage means as teacher data. In addition, an optional step 11 (S11) is additionally performed, and experimental data obtained from experiments assuming the target real structure, real internal deformation, real stimulus, etc. are used as further teacher data by the teacher data storage means. may additionally be stored in Furthermore, optional step 12 (S12) may be additionally performed to process the teacher data accumulated in the teacher data storage means by the data expansion method or other techniques, thereby improving the quality and quantity of the teacher data. good.

Next, in step 5 (S5), using the teacher data stored in the teacher data storage means and the actual response data stored in the measured data storage means obtained as described above, a method such as DL is used to Evaluate and estimate the characteristics such as the position, range, physical properties, type, etc. of internal deformation that exists in the structure.

Finally, in step 6 (S6), the characteristics of the detected internal deformation are displayed on display means such as a display. At this time, related information about the internal deformation, for example, the position in the actual structure where the internal deformation occurs, the type of internal deformation that is likely to be associated with the range of the actual response, the possibility of its existence, etc. information may be displayed.

What remains as a problem here is a specific extraction method for extracting a portion that is strongly related to the response corresponding to the assumed internal deformation from the actual response of the measured data, which constitutes the teacher data extraction means 106. is necessary. As noted above, the response of a structure with internal deformations to any stimulus contains internal deformations in a steady-state response with relatively little variation, reflecting a mostly healthy general part. A reflected, non-stationary response portion with relatively large fluctuations appears locally. In conventional signal processing techniques, there is a scanning method using a filter that is effective for extracting non-stationary responses among such stationary responses. For example, by applying the scanning method using this filter to the measured data of the structural response, we can extract non-stationary response portions that are similar to those considered to be related to the assumed internal deformation, and use this as the training data. can be added to Specific details regarding this will be described later.

In the above description, the internal deformation is assumed to be internal deformation 1 and internal deformation 2. However, by changing the filter characteristics variously, a new characteristic response is selected from responses that do not correspond to internal deformation 1 and internal deformation 2. If a part of the response can be extracted, that part can be used as the internal deformation 3 . In this case, even if the characteristics of the internal deformation 3 cannot be specifically determined by a clear specification such as setting in the analysis model, the teacher data will be used as a new characteristic internal deformation of the target structure. It is thought that the accuracy of classification and identification can be improved step by step by adding to

In this way, even if the present invention starts from training data based on relatively small amount of analysis data, it can reflect the various characteristics of internal deformation in the actual structure by utilizing the measured data of the actual structure, expanding the data, etc. It is possible to multiply the teacher data that has been used.
2 An example of how to create training data
2.1 Setting the Overall Situation Next, in order to specifically explain the method of creating training data, an example is assumed in which internal deformation 1 and internal deformation 2 exist in the structural section shown in Fig. 3(a). more specifically, select cracks and floats, which are typical internal deformations that exist in concrete slab structures, stimulate the side where cracks and floats exist and the opposite side of the slab structure by heating, and Consider the case where the surface temperature of the side on which cracks and floats are present, corresponding to heating (hereinafter referred to as front side heating and back side heating), is measured as a response.

Here, a crack is a thin, planar space that exists locally within a structure, and a part of the space intersects the structure surface. Suppose that there is a thin, planar space that exists locally in , and that space does not intersect the structural surface.
2.2 Creation of training data by analysis
2.2.1 Surface temperature distribution related to cracks Figure 6 shows a crack analysis model in which cracks exist in the structure. When conducting a steady-state heat transfer analysis for this model, for example, a uniform material with a thermal conductivity of κ (1.6 W/(m K)) is used. A crack opening with a width of 0.1 mm and a length of c (150 mm) is located near the center of the upper surface of a flat plate of height h (100 mm). It is necessary to set many conditions, such as a crack penetrating to a depth e (50 mm) from the upper surface of the flat plate. Furthermore, the heat transfer coefficient of the heat transfer boundary on the lower or upper surface is 6 W/m ² K, the ambient temperature T is 22 °, the four side surfaces are adiabatic boundary surfaces, and the heat flux φ flowing from the upper or lower surface is 100 W / m It is necessary to further set the thermal conditions such as ² . FIG. 7 shows a perspective view of a typical surface temperature distribution on the upper surface in each case of front-side heating or back-side heating in which the upper surface is used as the observation surface and the upper surface or the lower surface is heated under such conditions. (a) and 7(b). Various model cases can be considered by varying the parameters that make up these conditions. For example, the temperature distribution on the vertical bisector of the crack opening when the front side heating or the back side heating is performed for three cases where the inclination angle δ is 30°, 45°, and 60°. Figures 8a and 8b. When the shapes of both the front side heating and the back side heating are turned upside down, they have almost similar temperature distributions. Also, the shape of these temperature distributions roughly resembles a sine wave with the origin at the crack opening. The shape of this partial surface temperature distribution is thought to be one of the clues to detect cracking locations.
2.2.2 Surface temperature distribution related to float Next, when there is a float in the flat concrete structure, the effect of the float on the surface temperature distribution was obtained by analysis. The analysis model is shown in FIG. In this model, the depth e ( 30 mm), there is a square space with side length c (150 mm) and side length d (1500 mm) with a thickness of 0.1 mm. FIGS. 10(a) and 10(b) show typical surface temperature distributions over the entire upper surface when the heating conditions are the same as for the crack model, and when the front side heating and the back side heating are performed. Again, different model cases can be considered by varying these parameters. For example, FIGS. 11(a) and 11(b) show the temperature distribution on the center line of the float for three cases where the depth e is 10 mm, 30 mm and 50 mm. As for the float, the shapes of both the front side heating and the back side heating have almost similar temperature distributions when they are turned upside down. Further, the shape of these temperature distributions roughly resembles a sine wave with the center of the float as the origin. This shape is considered to be one of the clues for detecting the location where the floating occurs. 11(a) and 11(b) show different temperature distributions, but the temperature distribution in the same range on the upper surface of the analysis model is shown by 250×250 pixels in the former and 125×125 pixels in the latter. It can be seen that a small difference in the number of pixels has almost no effect on the roughness of the image.
2.2.3 Diversification of teaching data As described above, the analysis model can be used to create teaching data on surface temperature distributions related to cracks and floats, which are internal deformations of flat concrete structures. Furthermore, if necessary, with respect to the many parameters set above, for example, changing the individual dimensions, changing the thermal conditions of the concrete, changing the heating conditions, intersecting cracks, connecting cracks and floats. The training data may be reinforced by increasing the diversity of the model through operations such as changing the flaky space from a single space to a combination of multiple spaces, taking into account unsteady states, etc. . Furthermore, it is also possible to incorporate the effects of reinforcing bars, aggregates, etc. existing inside a typical concrete structure into the analysis model in order to better match the conditions of the actual structure.

In addition to changing the configuration, parameters, etc. of the analysis model as described above, the data expansion method used in the DL method, that is, rotation, scaling, translation, shearing, blurring, noise addition, as described above etc. may be applied to the surface temperature distribution.
2.3 Method of creating teacher data from measured surface temperature distribution
2.3.1 Analysis method for extracting the training data part As a filter for detecting local temperature change parts related to internal deformation included in the surface temperature distribution of the structure, for example, There is a filter based on two-dimensional wavelet transform that is said to be effective in detecting local changes in non-stationary signals, and a method of using this filter will be described below.

Assuming that the coordinate axis is x in the horizontal direction and y in the vertical direction, the basic equation of the two-dimensional wavelet transform is as follows.

ここに、t(x,y)は座標(x,y)における表面温度に関連する関数であって、温度分布だけでなく温度分布を反映した明度画像データ又はカラー画像を構成する刺激値でも代用できるが、以下、簡単のためこれらを含めて温度分布と呼ぶ。ψ(x,y)はマザーウェーブレットと呼ばれる関数であり、局在化した関数である。(x₀,y₀)はψ(x,y)のシフトパラメータであり、ａはスケーリングパラメータである。

Here, t(x, y) is a function related to the surface temperature at the coordinates (x, y), and not only the temperature distribution but also the brightness image data reflecting the temperature distribution or the stimulus value constituting the color image can be substituted. However, hereinafter, for the sake of simplification, these are also referred to as the temperature distribution. ψ(x,y) is a function called a mother wavelet and is a localized function. (x ₀ ,y ₀ ) is the shift parameter of ψ(x,y) and a is the scaling parameter.

Next, we need to choose a suitable mother wavelet ψ(x,y) to characterize the temperature distribution near the internal deformation of the concrete structure. Such a temperature distribution is characterized by changes in directionality, gradient, magnitude, etc., depending on the location. Therefore, the present invention is effective in detecting, for example, the local directionality and periodicity of signals. A two-dimensional Gabor wavelet with a Gabor function that
2.3.2 Characteristics of two-dimensional Gabor wavelet suitable for detection Let us redefine the mother wavelet ψ(x,y) as a two-dimensional Gabor wavelet (hereafter abbreviated as Gabor wavelet) shown in Equation (2). This Gabor wavelet is obtained by applying a two-dimensional Gaussian window to the complex oscillation ^ei(2πfx+φ) traveling in the x-axis direction. By applying a two-dimensional Gaussian window, the complex oscillation ^ei(2πfx+φ) becomes localized around the origin, reflecting the local periodicity and directionality of the temperature distribution in its vicinity. You get results.

ここに、fは中心周波数、σ_x及びσ_yはｘ軸方向及びｙ軸方向の窓幅、並びにφは位相を表す。

where f is the center frequency, σ _x and σ _y are the window widths in the x and y directions, and φ is the phase.

By using such a Gabor wavelet, the Gabor wavelet transform Ψ(x,y) of the temperature distribution t(x,y) is given by the following equation.

しかし、式（２）のウェーブレットは、ｘｙ平面上で一方向にのみ進行する波であるのに対して、変状に起因する温度変化の方向は任意であるので、温度分布t(x,y)の各点での温度変化の任意方向の特性を検出するためには、各点の周りでウェーブレットを回転させる必要がある。そこで、ψ(x,y)を各点で反時計回りに角度θ_r回転させて得られるψ_r(x,y,θ_r)を次式で定義し、解析においては必要な範囲にわたってこれを回転させる。

However, while the wavelet of equation (2) is a wave that travels in only one direction on the xy plane, the direction of temperature change due to deformation is arbitrary, so the temperature distribution t(x,y ), we need to rotate the wavelet around each point to detect the directional characteristics of the temperature change at each point. Therefore, ψ _r (x, y, θ _r ), which is obtained by rotating ψ(x, y) counterclockwise by an angle θ _r at each point, is defined by the following equation. rotate.

ここで、ひび割れ又は浮きに関連した表面温度分布の形状を検出するのに適している式（２）のガボールウェーブレットの構成について考察する。ひび割れに関連した表面温度分布は、図４～６に関して述べたように、開口部の直線に垂直な方向において、開口部を境界として表面温度が急変し、最高温点を含む峰部と最低温点を含む谷部とが開口部の両側に存在するので、開口部に原点を有する正弦波に類似していると考えてもよいと思われる。従って、複素振動e^i(2πfx+φ)のうちの正弦波sin(2πfx)によって、ひび割れに関連した温度分布を検出できる可能性があると考えられる。

We now consider the construction of the Gabor wavelet of equation (2) suitable for detecting the shape of surface temperature distributions associated with cracks or floats. As described with reference to FIGS. 4 to 6, the surface temperature distribution related to cracking is such that, in the direction perpendicular to the straight line of the opening, the surface temperature changes suddenly with the opening as the boundary, and the peak containing the highest temperature point and the lowest temperature point. Since troughs containing points exist on both sides of the aperture, it can be thought of as analogous to a sine wave with its origin at the aperture. Therefore, it is thought that there is a possibility that the sine wave sin(2πfx) of the complex vibration ^ei(2πfx+φ) can detect the crack-related temperature distribution.

On the other hand, the surface temperature distribution associated with the float, as described with respect to FIGS. It can be thought of as analogous to the characteristics of a cosine wave with its origin at the temperature point. Therefore, it is conceivable that the cosine wave cos(2πfx) of the complex oscillation ^ei(2πfx+φ) can detect the surface temperature distribution related to float.

Furthermore, in order to analyze the characteristics of the surface temperature distribution related to internal deformation in more detail, the surface temperature distribution of the detection target range was measured by measuring the surface temperature of a healthy portion that was hardly or not affected by internal deformation (reference It is conceivable to obtain a temperature distribution consisting of positive and negative temperatures expressed as a difference between the temperature and the temperature, and to apply the wavelet transform to this temperature distribution. The sine wave component and cosine wave component thus obtained can evaluate the characteristics of the internal deformation from various angles depending not only on the magnitude of the value but also on whether the value is positive or negative. be able to. For example, in the cases shown in FIGS. 5 to 6 and 9 to 10, the surface temperature of the sound portion is about 31° C. when the front side is heated and about 29° C. when the back side is heated. Moreover, if it is not easy to set the surface temperature of the sound portion, the average temperature of the surface temperature distribution in the detection target range may be used as the reference temperature.
2.3.3 Procedure for detecting temperature distribution related to internal deformation Using the Gabor wavelet ψ(x,y) as described above, the surface temperature distribution t(x,y) is affected by the presence of internal deformation. The procedure for detecting the location of the temperature distribution subjected to , that is, the location of the characteristic temperature distribution related to the internal deformation is as follows.

For example, when detecting a temperature distribution location related to a deformation similar to a crack, the temperature distribution that is highly related to the sinusoidal wave sin(2πfx) of the complex vibration ^ei(2πfx+φ) in Equation (2) you will find the spot. Therefore, in order to give width to the characteristics of the sine wave sin(2πfx), a set of values of frequency f and window widths σ _x , σ _y within an appropriate range are sequentially selected, and the Gabor wavelet ψ(x ,y). angles spaced counterclockwise

The Gabor wavelet expansion coefficients Ψ _ri of the surface temperature distribution t(x, y) of the target region are obtained from ψ _r (x, y, θ _r ) rotated by . Next _{, the} value θ _r0 of the angle θ _ri that maximizes the sum of the squares of the absolute values of the expansion coefficients Ψ _{ri is} obtained. be the direction of travel of the sine wave of interest for .

Next, by arranging the absolute values of the Gabor wavelet expansion series of the entire target region in descending order, leaving only the expansion coefficients with the highest p% (expansion coefficient ratio) and correcting the other expansion coefficients to 0, as a result, a minute temperature change can be detected. Relevant values are removed, thus emphasizing the larger of the temperature changes that show the characteristic trend of the deformation. When the distribution of the squared value of the corrected expansion coefficient is displayed in the figure, it shows the locations where the surface temperature distribution related to the internal deformation exists. Even in this way, the influence of noise is large, and if the method of the present invention alone cannot sufficiently perform detection, it is effective to apply the noise removal method to the surface temperature distribution of the target region in advance. One example of the denoising method is by wavelet shrinkage using multiresolution analysis with orthogonal wavelets.

Also, when detecting a temperature distribution occurrence location related to a deformation similar to floating, the cosine wave cos(2πfx) in the complex vibration ^ei(2πfx+φ) in Equation (2) is highly relevant. To find the temperature distribution point, the procedure is the same as above.

In the above description, the complex oscillation ^ei(2πfx+φ) of the filter is composed of a sine wave or cosine wave of a single frequency. good too. Furthermore, the filter may be expanded to any function. As a result, if a new characteristic surface response portion can be extracted, that internal deformation may be used as a new internal deformation even if the characteristics of the internal deformation related to that portion cannot be clearly specified. .
2.3.4 Verification of detectability for analysis data
(1) Target surface temperature data and noise to consider The effectiveness of the detection method for detecting the location of the surface temperature distribution was verified.

A point to be considered here is that noise caused by various factors is added to the actual surface temperature distribution as shown in FIGS. The temperature distribution characteristic of internal deformation may be greatly disturbed by the influence of noise. Therefore, when examining the detectability, it is necessary to show that the surface temperature distribution to which noise is added can be sufficiently detected.

Considering the surface temperature distribution noise, the noise characteristics are local non-uniformity of thermal properties due to concrete being composed of composite materials such as cement, fine aggregate, and coarse aggregate, irregularities and stains on the concrete surface, Gaussian white noise with an average of 0 based on the standard deviation σ of the surface temperature distribution of concrete with internal deformation, although it cannot be defined unconditionally because it is affected by uneven heating conditions and environmental conditions. given as

For example, in the crack analysis model shown in FIG. 6, when δ=45°, the standard deviation σ of the temperature distribution on the observed surface during front side heating and back side heating is 0.157. Further, in the float analysis model shown in FIG. 9, the standard deviation σ of the temperature distribution on the observed surface during front side heating and back side heating when d=30 mm is 0.261. For these crack analysis model and float analysis model, the temperature distribution when Gaussian white noise with standard deviation σ _w of twice the standard deviation σ of each is added to the surface temperature distribution as noise is shown in Fig. 12 and It is shown in FIG. Since the surface temperature of the location where the internal deformation exists is higher or lower than that of the surrounding area, it may be possible to visually recognize the existence if the color is displayed. It becomes difficult to see temperature changes.
(2) Verification of the detectability of cracks Regarding the surface temperature distribution obtained by the crack analysis model in the case of front side heating and back side heating, when they do not contain noise and when the above-mentioned Gaussian white noise is added Verification was carried out for the surface temperature distribution.

Of the surface temperatures for the crack model shown in Fig. 6, the results of applying this detection method to the surface temperature distribution when δ = 45°, the front side heating and the back side heating, and no noise are included are shown. 14 (a) to (d) and FIG. 15 (a) to (d). In the case of front-side heating, FIG. 14(a) shows ^the positive part (hereinafter referred to as the positive sine wave-related component ), and FIG. 14(b) is the negative part (hereinafter referred to as negative sine wave-related component) of the sine wave-related part, and FIG. 14(c) is the cosine wave-related component FIG. 14(d) is a positive part (hereinafter referred to as a positive cosine wave related component) of the part related to the cosine wave (hereinafter referred to as a negative cosine wave related component). component). In this study, the main issue is the presence or absence of the sine wave-related component and the cosine wave-related component within the very narrow range of the expansion coefficient ratio p = 4%, so the value of each component itself has no particular meaning. .

The position where the positive sine wave-related component shown in FIG. 14(a) exists almost coincides with the position on the side where the surface temperature at the crack opening is higher than the reference temperature, and FIG. 14(b) shows The position where the indicated negative sine wave-related component exists almost coincides with the position on the side where the surface temperature at the crack opening is lower than the reference temperature. Therefore, it can be seen that the position of the crack opening can be detected by positive and negative sinusoidal related components.

14(c), the positive cosine-wave-related component shown in FIG. The temperature gradually decreases from the maximum temperature point at , which is almost the same as the range where the temperature is higher than the reference temperature. Further, for the negative cosine wave function component shown in FIG. 14(d), the temperature rises gently from the lowest temperature point on the side where the crack surface exists, as shown in the graph of FIG. 8(b). It almost matches the range where the temperature is lower than the temperature. Therefore, it can be seen that the side on which the crack surface exists can be detected by the positive and negative cosine wave related components.

In this way, from the sine wave-related components and cosine wave-related components of FIGS. can be detected.

On the other hand, in the case of back side heating, as in the case of front side heating, in FIGS. 15(a) to 15(d), the positions of crack openings are indicated by positive and negative sine wave-related components. In the case of backside heating, as shown in FIG. 8(b), there is a maximum temperature point on the side where the crack surface does not exist, and a range where the temperature gradually decreases from that point, and the temperature is higher than the reference temperature. On the opposite side of the opening, there is a range where the temperature is lower than the reference temperature, where the temperature gradually rises from the lowest temperature point. The former temperature range is detected by the positive cosine related component shown in FIG. 15(c), and the latter temperature range is detected by the negative cosine related component of FIG. 15(d). Therefore, even in the case of back surface heating, the information about the crack from the sine wave-related components and cosine wave-related components in FIGS. existing side and can be detected.

Next, FIG. 16(a) to (d) and FIG. 17( a) to (d). In the case of front side heating, even if noise is included, the effect of noise is almost completely removed, and the noise is as clear as the case without noise shown in FIGS. The position of the crack opening can be determined from the positive and negative sine wave related components in (b), and the side on which the crack face exists from the positive and negative cosine wave related components in FIGS. 16(c) and 16(d). On the other hand, in the case of backside heating, as shown in Figs. 17(a) to 17(d), the temperature change on the observation surface of the observation target is smaller than in the case of front side heating, so the effect of noise cannot be completely eliminated. Although the detected image is somewhat distorted, it is quite possible to determine the position of the crack opening and the side on which the crack face exists.

In the previous studies, one image included cracks with one crack inclination angle (δ = 45°). Consider the case where cracks are included. In addition to the three crack models with cracks with inclination angles δ of 30°, 45° and 60° shown in FIG. Consider a crack model with an overall size of 1000mm x 1000mm x 100mm, with two pieces connected vertically and horizontally. FIG. 18 shows the positive sine wave-related component and the positive cosine wave-related component among the results of applying this detection method to the surface temperature distribution obtained by heating this crack model under the same conditions as the front side heating conditions. .

In FIG. 18, as before, the sine wave-related component indicates the position of the crack opening, and the cosine wave-related component indicates the side where the crack surface exists. The position of the crack opening and the side on which the crack surface is present are detected as signals of a wider range as the inclination angle δ is smaller. In other words, the range displayed by each component is wide in the order of δ = 30° and δ = 45°. , the signal for both the position of the crack opening and the side where the crack face exists does not appear.

As can be seen from FIG. 8, the reason for this is that the smaller the inclination angle δ, the higher the temperature at the highest temperature point, so that the sine wave-related component appears more strongly in both front-side heating and back-side heating. In this case, since the crack depth is the same, the smaller the inclination angle δ, the larger the area of the crack surface extending from the crack opening to the inside of the structure. It is thought that it is from Thus, it can be seen that the difference in crack inclination angle can be detected from the difference in surface temperature distribution characteristics.

Furthermore, as a more complicated condition, as shown in FIG. 19, an orthogonal crack where two cracks perpendicular to the δ = 45° crack shown in FIG. Think model. 20(a) to 20(d) show the results of applying this detection method to the surface temperature distribution when noise is included in the case of front side heating. Also in this case, it can be seen that the positions of two orthogonal crack openings can be detected from the positive and negative sine wave-related components in FIGS. 20(a) and 20(b). Also, it is possible to determine the side on which two crack planes exist from the positive and negative cosine wave-related components in FIGS. 20(c) and 20(d).
(3) Verification of the possibility of detecting the float The detection results when the front side heating and the back side heating for the float model shown in FIG. 9 are included and noise is included are shown in FIGS. a) to (d). For the floating model, the positive and negative sinusoidal related components of FIGS. 21(a) and 21(b) and FIGS. 22(a) and 22(b) do not appear. The reason for this is that the surface temperature does not suddenly change across the crack opening as in the case of the crack model. A positive cosine wave-related component appears because a dome-shaped portion having a higher temperature than the reference temperature is generated on the observation surface due to the surface heating. On the other hand, backside heating causes a dome-shaped portion of the observation surface that is cooler than the reference temperature, so that a negative cosine-related component appears.

Thus, by comprehensively considering the heating surface and how the sine wave related component and the positive and negative cosine wave related components appear, the presence of the float can be detected and the surface temperature distribution associated with the float can thus be determined. can be specified as teacher data.
2.4 Verification of crack detection by experimental data
2.4.1 Outline of Experiments After verification by analytical data, the detection performance of the method of the present invention was verified by experiments for cracks exhibiting a relatively more complex surface temperature distribution than floats.

Fig. 23 shows a concrete specimen of 450 mm × 600 mm × 200 mm (a1 = 200 mm, a2 = 250 mm, a3 = 150 mm, b1 = 100 mm, b2 = 250 mm, b3 = 100 mm), with an inclination angle δ of 30 to the surface. 1 mm thick (350 mm wide, 200 mm and 115 mm long) acrylic plates simulating the space of the crack are inserted to be 60° and 60°.

The thermal conductivity of acrylic plate and air is about 0.21W/(m・K) and about 0.025W/(m・K) respectively. It is believed to have comparable thermal insulation performance.

FIG. 24(a) is a surface temperature image of the front side when the front side with the simulated cracks is placed directly above and heated by solar radiation on a fine day. The surface temperature distribution at this time had a maximum temperature of 19.4°C, a minimum temperature of 13.8°C, an average temperature of 16.3°C, and a standard deviation of 1.1°C. FIG. 24(b) displays the result of dividing the mean temperature from the surface temperature distribution normalized by their maximum absolute value. From these images, the position of the crack opening with an inclination angle δ of 30° can be visually recognized, but the position of the crack opening with an inclination angle δ of 60° cannot be visually recognized. As shown in the previous section, the larger the inclination angle δ, the more difficult it is for the temperature distribution characteristic of cracks to occur. In addition, due to the positional relationship between the two acrylic plates, the amount of heat to be stored at the point where the highest temperature point appears for the crack with an inclination angle δ of 60° It is thought that one of the reasons is that it was further promoted because it flowed to
2.4.2 Detection Results Figures 25(a) to 25(d) show the detection results for the image data shown in Figure 24(b). , (c) the positive cosine-related component, and (d) the negative cosine-wave component.

The positive and negative sinusoidal related components clearly indicate the location of the crack opening with an inclination angle δ of 30°. However, cracks with an inclination angle δ of 60° could not be detected at all. For the reason described above, it is considered that the characteristic temperature distribution targeted by the present detection method did not occur at the location of the crack with the inclination angle δ of 60°.

The positive and negative cosine wave-related components, which are located slightly away from the crack opening with an inclination angle δ of 30°, correctly capture temperature changes on the side where the crack surface exists and on the opposite side. Therefore, by superimposing the cosine wave-related component on the sine wave-related component, not only does it become useful complementary information for detecting the crack occurrence location, but the surface temperature changes relatively gently even by itself. It is thought that it may be possible to detect areas of higher temperature than the surrounding area, which is not affected by deformation.

As can be seen from FIGS. 25(c) to (d), fine point-like components are present. This noise is caused by unevenness of local thermal properties, unevenness and dirt on the concrete surface, etc., due to the fact that the concrete is composed of composite materials such as cement, fine aggregate, and coarse aggregate, as described above. This is thought to be due to the influence of We investigated the effect of preprocessing to remove such noise components. The denoising method used is by wavelet shrinkage using multi-resolution analysis with orthogonal wavelets as described above. For the image data shown in FIGS. 23a to 23b, degeneration using only the top α=1% when arranging the absolute values of the Daubechies (N=10) wavelet level j=5 scaling coefficients and wavelet expansion integers in descending order FIGS. 26(a) to 26(d) show the results of performing the detection process again with the expansion coefficient ratio p=4% after performing the noise removal in advance. Small temperature variations are eliminated, and the presence of crack openings with an inclination angle δ of 30° is more clearly detectable.

It is assumed that the temperature distribution of the actual structure may contain more noise components than this specimen. In such cases, the noise removal method shown here is considered to be effective.
2.5 Examination of applicability to actual measurement data of actual structures
2.5.1 Measured data in actual structures As an example of application to surface temperature distribution in actual structures, application to flat concrete structures generally called floor slabs will be explained. Such structures are widely used in bridges or elevated structures of roads or railways, and are exposed to severe deterioration factors such as traffic load and rainwater, so their deterioration is a particular problem. Therefore, application of the present invention is one of desirable applications. The stimulation in this case is equally applicable to artificial heating, but here the simplest and least costly method is to heat the top side of the structure directly or indirectly under the influence of sunlight. Consider the case where

In this case, as internal deformations to be detected, cracks and floats that are typically seen in floor slabs of concrete structures are assumed. Consider detecting the location of the surface temperature distribution.

The thermal conditions in which such a structure is placed are substantially similar to those shown in the schematic diagram of FIG. When this is compared with the analysis model described above, the observation surface and the heating surface are different in structural aspects as shown in FIG. 6(b) for cracks and FIG. 9(b) for floats This corresponds to the case where the heating condition of the back surface heating works.
2.5.2 Detection results Next, from the surface temperature distribution obtained when the target structure was actually measured, a study was conducted to extract a range that is highly related to the surface temperature distribution corresponding to the cracks and floats assumed in the analysis. Show the results.

FIG. 27 is a photograph of the situation when the underside of the RC floor slab of a bridge was measured with an infrared imaging device on a sunny day in winter, and an RGB image showing the surface temperature distribution on the underside of the floor slab. 28 to 30 show the results of applying this detection method to the surface temperature distributions of regions R1 to R3 surrounded by rectangular solid lines in the RGB image, respectively. As before, in each figure, (a) positive sine-related component, (b) negative sine-related component, (c) positive cosine-related component, and (d) negative cosine-related component Ingredients are indicated.

Under this structural condition and the heating condition of back side heating, as shown in the examination of the analysis data, the range related to the float in this case becomes lower temperature than the surroundings, so only negative cosine wave related components appear, and cracks occur. The range associated with , appears as a set of positive and negative sinusoidal related components, with positive sinusoidal related components appearing incidentally around the crack. In this way, a range of surface temperature distribution that is highly related to the internal deformation of the object is extracted from the actual measurement data, and the surface temperature distribution in that range is used as teaching data related to cracks and floats. can be done.
3 Creation of teacher data when response is oscillatory data
3.1 Description of the Overall Situation Under Consideration In one embodiment according to the present invention, internal deformation and vibrational data are obtained when the stimulus given to a structure with internal deformation is an impact and the resulting response is vibrational data. A method for creating teacher data about the relationship between is described below.

As an example of the target structure and internal deformation, it is assumed that there is internal deformation in a part of the middle portion of the rod-shaped structure as shown in FIG. By hitting the end of such a rod-shaped structure and observing the vibration that propagates through the structure and returns to the end, it is possible to determine whether internal deformation has occurred, and to what position and to what extent. In some cases, the number of internal deformations is detected. This is because the shape of the vibration and the presence of the internal deformation are related because part of the vibration propagating through the rod-like structure is reflected and part of the vibration is transmitted at the location where the internal deformation exists.

For example, if this rod-shaped structure is in the ground, the details such as the length of the structure are unknown, and the relationship between the given vibration data and the internal deformation cannot be directly related. It is difficult to create training data systematically. Vibration data can be acceleration data detected by an acceleration sensor installed near the edge of the structure, or acoustic data detected by an acoustic sensor installed in the air around the edge where the vibration propagated in the air. . FIG. 32 shows an example of acceleration data measured by an acceleration sensor on a real structure.

An example of a rod-like structure shown here is generated by a stimulus such as a blow to the surface of the structure when the plate structure shown in FIGS. It can also be applied to the creation of teacher data when vibration, sound, etc. are responses.
3.2 Creation of teaching data by analysis When the rod-shaped structure shown in Fig. 31 does not contain internal deformation and has internal deformation in the center of the rod-shaped structure, when the upper end of the rod-shaped structure is hit, FIG. 33(a) and FIG. 33(b) show the results obtained by analysis of the acceleration data in the wave motion data in which the wave motion propagates through the rod-like structure and returns to the upper end. When the internal deformation shown in FIG. 33(a) is not included, the wave propagating in the rod-like structure periodically repeats reflection at the tip and upper end of the rod-like structure and gradually attenuates. On the other hand, when an internal deformation exists in the middle part of the rod-shaped structure, as shown in FIG. becomes very complex. Even if there is only one internal deformation, the shape is complicated. Therefore, when the number of internal deformations increases, humans can identify the position of the internal deformation only by observing the shape of the wave. is difficult.
3.3 Diversification of teaching data It is possible to create teaching data of acceleration data at the end of a rod-shaped structure with internal deformation by setting an analysis model as described above and applying a blow to the end by analysis. . Furthermore, if necessary, many parameters set for the rod-shaped structure shown in the figure, such as changing individual dimensions, changing the strength conditions of the constituent materials, adding restraint conditions from the surroundings, and the position of internal deformation The training data may be reinforced by increasing the diversity of the model structure through operations such as changing the range, number, and the like. In addition, the effects of reinforcing bars, aggregates, etc. present inside typical concrete structures can be incorporated into the analytical model to better match the conditions of the actual structure.

In the case of this rod-shaped structure, the measured data obtained from the actual structure often contains a lot of noise, so it may be desirable to remove the noise to clarify the shape of the wave motion. As an example of the denoising method applied to the actually measured data shown in FIG. 32, wavelet reduction using multi-resolution analysis with orthogonal wavelets can be considered. FIG. 34 shows the result of applying this noise removal to the measured data of FIG.

In addition, not only the measured data itself is used as the target vibrational data, but also the measured data decomposed by Fourier transform, orthogonal wavelet multi-resolution analysis, etc., can be used as the target for internal deformation. It may be possible to consider the impact in more detail.

In addition, as described above, in addition to changing the configuration of the model structure used for analysis, parameters, etc., the response data obtained by analysis can be used in the data expansion method used in the DL method, that is, rotation, scaling, translation, Shearing, blurring, noise addition, etc. may be applied.
4 Example of crack generation position detection using DL
4.1 Detection Procedure The detection procedure mainly follows the flow chart shown in FIG. Optional steps S10 and S11 are excluded, however, since only teacher data obtained by analysis is used.

According to step 1, the surface temperature distribution in the target structural section for detecting internal deformation occurrence locations is measured, and according to step 2, this is stored as surface temperature distribution data.

Next, according to step 3, one or more internal deformations to be detected existing in the target structural section are assumed based on engineering experience and judgment. Only cracks present in the Then, the surface temperature distribution at the crack generation location was obtained by analysis using the crack analysis model shown in FIG. As described in 2.2.1, the parameters that make up the analysis model are diverse, and the numerical values that each parameter takes can be set freely within a certain range. However, here, as the most basic case, one crack analysis model having a crack with a crack inclination angle δ of 45° was targeted using the values of each parameter shown in 2.2.1.

Next, according to the optional step 12, the surface temperature distribution at the crack generation location obtained from the one crack analysis model is processed. Here, in order to generalize the obtained surface temperature distribution, it is normalized so that the minimum temperature is 0 and the maximum temperature is 1, and the normalized temperature distribution is subjected to -90 Rotation in the range of ° to +90 °, enlargement in the range of 1.0 times to 1.5 times, and aspect ratio change in the range of 1:1.5 to 1.5:1 By diversification, a total of 50 training data were created.

Next, in step 5, the actually measured surface temperature distribution data of the target was read out, and DL detection was performed using the teacher data.

Finally, according to step 6, the locations where cracks are assumed to be detected are shown superimposed on, for example, the original surface temperature distribution image.
4.2 Examination with experimental data
4.2.1 Target experimental data
The surface temperature distribution data shown in FIG.
4.1.2 Detection Result The DL method was applied to the above surface temperature distribution, and detection was performed using locations where there was a local surface temperature distribution related to cracks, which was the source of the training data. In FIG. 35, the approximate position of such a location is indicated by a square frame superimposed on the image of the surface temperature distribution. Similar to the results shown in FIGS. 25 and 26, this result clearly detects the position of the opening of the crack with an inclination angle δ of 30°. , indicates that only the upper right part of the figure can be detected.
4.2 Examination using measured data of actual structures
4.2.1 Target measured data The surface temperature distribution data actually measured on the lower surface of the RC floor slab shown in Fig. 27 was targeted.
4.2.2 Detection Results The DL method was applied to the above surface temperature distribution, and detection was performed using locations where there was a local surface temperature distribution related to cracks, which was the source of the training data. In FIG. 36, the approximate position of such a location is indicated by a square frame superimposed on the image of the surface temperature distribution.

These two examples show the results of executing DL using training data created from a single basic crack analysis model. change the depth or depth, increase the number of cracks, make the relationship between multiple cracks parallel, intersecting, etc., and/or add a float, which is a representative internal deformation, to the internal deformation of the object. It is thought that this can contribute to the improvement of detection accuracy.

REFERENCE SIGNS LIST 100 structural internal deformation detection device 101 actual measurement means 102 structural analysis means 103 experimental means 104 actual measurement data storage means 105 teacher data storage means 106 teacher data extraction means 107 estimation means 108 display means

Claims (18)

A structural internal deformation detection device that detects the characteristics of real internal deformation that exists on the surface or inside the actual structure and is invisible from the outside by deep learning,
measurement means for measuring an actual response of the real structure when a real stimulus is applied to the real structure and acquiring the actual response as actual response data;
measured data storage means for storing the actual response data acquired by the measured means;
A model structure is set based on the actual structure, and an arbitrary model internal deformation composed of basic parameters for the actual internal deformation is selected from a small number of models that are likely to occur in the actual structure. creating the model structure further incorporating the model internal deformation set to have basic properties and sequentially varying the values of some of the basic parameters; and based on the real stimuli structural analysis means for generating model response data by sequentially structurally analyzing model responses of the model structure when the set model stimulus is applied to the model structure;
teacher data storage means for storing the model response data as teacher data used in the deep learning;
learning the teacher data stored in the teacher data storage means by the deep learning, and using the results of the learning and the actual response data stored in the measured data storage means, characteristics of the real internal deformation a guessing means for guessing
display means for displaying the estimated characteristics of the real internal deformation on a screen;
A structural internal deformation detection device comprising: further comprising teacher data extracting means for extracting, from the actual response data stored in the measured data storage means, a part of the actual response data considered to have a strong correlation with the model response; 2. The structural internal deformation detection device according to claim 1, further comprising, as said teacher data, a part of said actual response data extracted by said teacher data extracting means and considered to have a strong correlation with said model response. . further comprising experimental means for obtaining experimental data on response when a stimulus set based on the real stimulus is applied to the test body simulating the real structure and the real internal deformation, wherein the teacher data storage means 3. The structural internal deformation detection device according to claim 1, further comprising said experimental data obtained by said experimental means as said teacher data. 4. The structural internal deformation according to any one of claims 1 to 3, wherein a data augmentation method is applied to said teacher data stored in said teacher data storage means to increase the quantity and variety of said teacher data. detection device. 5. The structural internal deformation detection device according to claim 1, wherein the characteristics of the internal deformation include the position, range, physical properties, or type of the internal deformation. 6. The structural internal deformation detection device according to claim 1, wherein the actual stimulus includes any one of heating, impact, vibration, and ultrasonic waves. When the actual structure is a concrete plate structure and the model internal deformation is a thin planar crack, the basic parameters constituting the model internal deformation are the width and length of the crack opening, the crack 7. The structural internal deformation detection device according to any one of claims 1 to 6, including an inclination angle with respect to the upper surface of the flat plate and a penetration depth of the crack from the upper surface of the flat plate. When the actual structure is a concrete plate structure and the model internal deformation is a thin planar float, the basic parameters that constitute the model internal deformation are the vertical length, the horizontal length and the thickness of the float. 7. The structural internal deformation detection device according to any one of claims 1 to 6, including the height and the depth from the upper surface of the flat plate of the float. When the actual structure is a rod-shaped structure, the actual stimulus is a blow, and the actual response data is vibrational data, the noise contained in the vibrational data is wavelet using multi-resolution analysis by orthogonal wavelets. 6. The structural internal deformation detection device according to claim 1, wherein the removal is performed by degeneracy. A structural internal deformation detection method for detecting, by deep learning, characteristics of real internal deformation that exists on the surface or inside the actual structure and cannot be visually recognized from the outside,
an actual measurement step of measuring an actual response of the real structure when a real stimulus is applied to the real structure and acquiring the actual response as actual response data;
an actual measurement data storage step of storing the actual response data acquired by the actual measurement step;
A model structure is set based on the actual structure, and an arbitrary model internal deformation composed of basic parameters for the actual internal deformation is selected from a small number of models that are likely to occur in the actual structure. creating the model structure further incorporating the model internal deformation set to have basic properties and sequentially varying the values of some of the basic parameters; and based on the real stimuli a structural analysis step of sequentially analyzing a model response of the model structure when the set model stimulus is applied to the model structure to create model response data;
a teacher data storage step of storing the model response data as teacher data used in the deep learning;
learning the teacher data stored in the teacher data storing step by the deep learning, and using the result of the learning and the actual response data stored in the measured data storing step, characteristics of the real internal deformation a guessing step of guessing
a display step of displaying the estimated characteristics of the actual internal deformation on a screen;
A structural internal deformation detection method comprising: further comprising a teacher data extraction step of extracting a portion of the actual response data that is considered to have a strong correlation with the model response from the actual response data stored in the actual measurement data storage step; 11. The structural internal deformation detection method according to claim 10, further comprising a portion of said actual response data extracted by said teacher data extraction step and considered to have a strong correlation with said model response as said teacher data. . further comprising an experiment step for obtaining experimental data on response when a stimulus set based on the real stimulus is applied to the test body simulating the real structure and the real internal deformation; 12. The structural internal deformation detection method according to claim 10, further comprising said experimental data obtained by said experimental step as said teacher data. 13. The structural internal deformation according to any one of claims 10 to 12, wherein a data augmentation method is applied to said teacher data stored in said teacher data storage step to increase the quantity and diversity of said teacher data. Detection method. The internal structural deformation detection method according to any one of claims 10 to 13, wherein the characteristics of the internal deformation include the position, range, physical properties, or type of the internal deformation. The structural internal deformation detection method according to any one of claims 10 to 14, wherein the actual stimulus includes any one of heating, impact, vibration, and ultrasonic waves. When the actual structure is a concrete plate structure and the model internal deformation is a thin planar crack, the basic parameters constituting the model internal deformation are the width and length of the crack opening, the crack 16. The structural internal deformation detection method according to any one of claims 10 to 15, including an inclination angle with respect to the upper surface of the flat plate and a penetration depth of the crack from the upper surface of the flat plate. When the actual structure is a concrete plate structure and the model internal deformation is a thin planar float, the basic parameters that constitute the model internal deformation are the vertical length, the horizontal length and the thickness of the float. 16. The structural internal deformation detection method according to any one of claims 10 to 15, wherein the height and the depth from the upper surface of the flat plate of the float are included. When the actual structure is a rod-shaped structure, the actual stimulus is a blow, and the actual response data is vibrational data, the noise contained in the vibrational data is wavelet using multi-resolution analysis by orthogonal wavelets. 16. The structural internal deformation detection method according to any one of claims 11 to 15, wherein removal is performed by degeneracy.

Images