Method for detecting abnormality of support structure

Abstract

To perform high-accuracy inspection of road structures in service for abnormalities, by using an auto-encoder.SOLUTION: The method for detecting an abnormality of a support structure includes: a step of measuring a physical quantity by a sensor arranged on a road structure; and an abnormality diagnosis step for detecting an abnormality of the support structure on the basis of a feature amount calculated from a measurement result, the abnormality diagnosis step including: a first step of storing the physical quantity; a second step of calculating a normal-state feature amount in a normal state; a third step of performing learning of an auto-encoder such that an identical output is reconstructed at a decoder layer by inputting the normal-state feature amount to an encoder layer; a fourth step of calculating a monitoring feature amount during monitoring; and a fifth step of outputting reconstructed data by inputting the monitoring feature amount to the trained auto-encoder and detecting abnormalities of the support structure on the basis of an error between the input and the output.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a method for detecting abnormalities in a support structure, which uses an autoencoder to detect abnormalities due to deterioration in the support structure of a road structure in use. [Background technology]

[0002] For example, bridges on which vehicles pass or road information boards installed around roads on which vehicles pass gradually deteriorate due to aging caused by the passage of vehicles. Note that in this disclosure, structures such as bridges and road information boards that are in use will be referred to as road structures in use.

[0003] Many road structures currently in use were constructed in large quantities in response to the rapid expansion of road traffic demand during the period of high economic growth. Accidents caused by deterioration of road structures in use often cause severe damage, so it is extremely important to detect deterioration before they break down.

[0004] However, the main method for inspecting the deterioration of road structures in service is still to carry out periodic inspections by inspectors, either visually or using some kind of measuring instrument.

[0005] As mentioned above, it is important to prevent the occurrence of serious damage due to deterioration of the support structures used in road structures in use, and in recent years there has been a demand for methods that can detect abnormalities in road structures in use unmanned, without the need for inspectors, and at intervals shorter than regular inspections.

[0006] One possible method for unmanned detection of abnormalities in road structures in use is to apply an anomaly detection method using an autoencoder. A method for detecting abnormalities in mechanical devices using an autoencoder has been disclosed as prior art (see, for example, Non-Patent Document 1).

[0007] Specifically, in Non-Patent Document 1, a discrimination model of an autoencoder using a neural network is created for detecting abnormalities in a machine having a hydraulically driven piston mechanism, and the discrimination between normal and abnormal states of the machine is automated. As a result, the burden of visual inspection is reduced and the risk of overlooking an abnormality is reduced. [Prior art documents] [Non-patent literature]

[0008] [Non-Patent Document 1] Proceedings of the 35th Fuzzy Systems Symposium (FSS2019 Osaka University) FD1-3, Takaaki Mine et al., pp. 506-508 Summary of the Invention [Problem to be solved by the invention]

[0009] In Non-Patent Document 1, the sensor pattern acquired from the piston, which is the target of anomaly detection, is input directly to the autoencoder. However, with this type of input, it is difficult to accurately determine whether an abnormality has occurred in the rigidity of the road structure.

[0010] Furthermore, in Non-Patent Document 1, the average value of the maximum normal value and the minimum abnormal value is calculated using other data that was not used in learning, and a threshold value for determining an abnormality is set. In other words, in order to set a threshold value, the abnormality determination method in Non-Patent Document 1 requires measurement data under abnormal conditions in addition to the measurement data under normal conditions used in learning.

[0011] Therefore, in order to use an autoencoder to detect anomalies in road structures in use with high accuracy, it is important to consider the characteristics of the road structure being detected, and to establish inputs suitable for learning the autoencoder and anomaly detection methods.

[0012] The present disclosure has been made to solve the above-mentioned problems, and aims to provide a method for detecting abnormalities in support structures that can accurately identify whether a support structure included in a road structure in service is normal or abnormal using an autoencoder, in order to prevent serious accidents by inspecting the support structures for abnormalities. [Means for solving the problem]

[0013] The support structure abnormality detection method according to the present disclosure is a support structure abnormality detection method including a measurement step of measuring a physical quantity that serves as an index value for detecting an abnormality in the support structure by a sensor arranged on the support structure included in a road structure in service, and an abnormality diagnosis step of calculating a feature value that shows a change when an abnormality occurs in the rigidity of the support structure based on the measurement result by the sensor, and detecting an abnormality in the support structure based on the feature value, wherein the abnormality diagnosis step includes a first step of storing the physical quantity acquired from the sensor as a measurement result continuously or periodically as measurement information in a storage unit, a second step of calculating a normal state feature value as a feature value from a calculation processing result for the measurement information when the support structure is normal, and a third step of detecting the abnormality. The method comprises the following steps: a third step in which machine learning is performed on an autoencoder consisting of an encoder layer, a feature learning layer, and a decoder layer used for continuous monitoring, with the normal state features calculated in the second step as input to the encoder layer, so that the same output is reconstructed in the decoder layer; a fourth step in which monitoring state features are calculated as features from the results of calculations on the measurement information during daily monitoring; and a fifth step in which, during daily monitoring, the monitoring state features calculated in the fourth step for the autoencoder that has been trained in the third step are input to the encoder layer, the reconstructed data is output to the decoder layer, and anomalies in the support structure are detected based on the error between the input and output. [Effects of the Invention]

[0014] According to the present disclosure, a method for detecting abnormalities in support structures can be obtained that uses an autoencoder to accurately identify whether a support structure included in a road structure in service is normal or abnormal, in order to prevent serious accidents from occurring by inspecting the support structure for abnormalities. [Brief explanation of the drawings]

[0015] [Figure 1] 1 is a functional block diagram of an anomaly detection system that executes a support structure anomaly detection method according to a first embodiment of the present disclosure. [Figure 2] 4 is a flowchart illustrating a series of processes executed by a sensor and a controller as a method for detecting an abnormality in a support structure according to the first embodiment of the present disclosure. [Figure 3] FIG. 2 is an explanatory diagram showing a specific arrangement of sensors when an F-type information board is used as a road structure in use in the first embodiment of the present disclosure. [Figure 4] 1 is a structural diagram of an autoencoder that is applied when performing a support structure anomaly detection method according to a first embodiment of the present disclosure. FIG. [Figure 5] FIG. 2 is an explanatory diagram regarding a reconstruction error according to the first embodiment of the present disclosure. [Figure 6] FIG. 10 is a diagram showing a histogram of reconstruction errors according to the first embodiment of the present disclosure. [Figure 7] 1 is an explanatory diagram showing a specific arrangement of sensors when an I-shaped information board is used as a road structure in use in the first embodiment of the present disclosure. FIG. [Figure 8] 1 is an explanatory diagram showing a specific arrangement of sensors when a gate-shaped information board is used as a road structure in use in the first embodiment of the present disclosure. FIG. [Figure 9] FIG. 2 is an explanatory diagram showing a specific arrangement of sensors when a bridge is used as a road structure in service in the first embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0016] Hereinafter, preferred embodiments of the support structure anomaly detection method of the present disclosure will be described with reference to the drawings. The support structure anomaly detection method according to the present disclosure realizes a technique for detecting anomalies in a support structure with high accuracy using reconstruction errors from an autoencoder, and has the following two technical features. Feature 1: The results of physical quantity measurements taken by sensors are converted into feature quantities that show changes when an abnormality occurs in the rigidity of the support structure, and the converted feature quantities are used as input to the autoencoder. Feature 2: The threshold is set to a value that is statistically significant from the output of the trained autoencoder, eliminating the need for additional measurement data to set the threshold for determining anomalies.

[0017] In the following, a detailed explanation will be given using road information boards and bridges as specific examples of road structures in use.

[0018] Embodiment 1 1 is a functional block diagram of an anomaly detection system that executes a support structure anomaly detection method according to a first embodiment of the present disclosure. The anomaly detection system shown in FIG. 1 includes a sensor 10 and a controller 20.

[0019] The sensor 10 is placed on a support structure included in a road structure in service, and measures a physical quantity that serves as an index value for detecting abnormalities in the support structure. The number of sensors may be one or more, and Fig. 1 illustrates a case where one sensor 10 is used.

[0020] Examples of road structures in use include bridges on which vehicles pass, or road information boards installed around roads on which vehicles pass. Specific installation locations of the sensor 10 relative to bridges or road information boards will be described later with reference to Figures 3, 7 to 9.

[0021] The controller 20 performs an abnormality diagnosis to detect an abnormality due to deterioration of the support structure using the reconstruction error of the autoencoder based on the measurement results of the sensor 10. In order to perform such an abnormality diagnosis, the controller 20 includes an index value data collection unit 21, a feature calculation unit 22, a machine learning unit 23, and an abnormality determination processing unit 24.

[0022] 2 is a flowchart illustrating a series of processes executed by the sensor 10 and the controller 20 as a method for detecting an abnormality in a support structure according to the first embodiment of the present disclosure. The function of each component of the controller 20 will also be described with reference to the series of processes illustrated in the flowchart of FIG.

[0023] Step S201 corresponds to a measurement step executed by the sensor 10. Steps S202 to S206 correspond to abnormality diagnosis steps executed by the components of the controller 20, and are roughly divided into five steps, namely, first to fifth steps.

[0024] In the measurement step of step S201, the sensor 10 arranged on the support structure measures a physical quantity that is an index value for detecting an abnormality in the support structure. A specific example of the sensor 10 is an acceleration sensor, which is installed on the support structure to measure acceleration data corresponding to vibrations of the support structure as a physical quantity.

[0025] 3 is an explanatory diagram showing a specific arrangement of sensors 10 when an F-type information board is used as a road structure in use in the first embodiment of the present disclosure. The F-type information board 200 shown in FIG. 3 has a support structure including an information board 201 and a support pillar 202.

[0026] In this type of support structure, the following three points P31 to P33 shown in FIG. 3 are important positions that correspond to the poles of the vibration resonance frequency or the parts where stress is concentrated. P31: Position of the tip of the support 202 P32: The base position of the support 202 P33: Position of the tip of the information board 201 that is the pole in the in-plane direction (the direction in which the display surface of the information board 201 does not shake, i.e., the direction across the road)

[0027] Next, in the first step of the abnormality diagnosis step of step S202, the index value data collection unit 21 in the controller 20 continuously or periodically acquires physical quantities as measurement results from the sensor 10 in sequence, and stores the measurement information as time-series information in a memory unit (not shown).

[0028] Next, in the second step of the abnormality diagnosis step of step S203, the feature amount calculation unit 22 calculates a normal state feature amount as a feature amount from the measurement information collected when the support structure is normal.

[0029] Here, the feature amount in the first embodiment is a value that changes when an abnormality occurs in the rigidity of the support structure, and is a value converted from the measurement result by the sensor 10. Specifically, the feature amount calculation unit 22 can convert the measurement result by the sensor 10 into a spectrogram as a feature amount by performing a short-time Fourier transform.

[0030] Furthermore, the feature amount calculation unit 22 can convert the measurement results from the sensor 10 into a scalogram as a feature amount by performing a wavelet transform.

[0031] Furthermore, the feature amount calculation unit 22 can perform FFT (Fast Fourier Transform) analysis based on the measurement results from the sensor 10, thereby converting the results into a frequency spectrum as a feature amount.

[0032] In the support structure included in the road structure in use, which is the target of anomaly detection in the present embodiment 1, if an abnormality occurs in the rigidity, a change appears in the frequency characteristics. The spectrogram, scalogram, or frequency spectrum, which are feature quantities obtained by converting the physical quantities measured by the sensor 10, are feature quantities that represent the frequency characteristics, and in particular, the spectrogram and scalogram are feature quantities that also include the time change of the frequency characteristics.

[0033] Therefore, the spectrogram, scalogram, or frequency spectrum are values ​​that change when an abnormality occurs in the rigidity of the support structure, and by using these values ​​as input to the autoencoder, it is possible to evaluate with high accuracy whether an abnormality has occurred in the rigidity of the road structure.

[0034] Next, in the third step of the abnormality diagnosis step of step S204, the machine learning unit 23 performs machine learning on an autoencoder consisting of an encoder layer, a feature learning layer, and a decoder layer used for abnormality monitoring, using the normal state features calculated in the previous second step as input to the encoder layer so that the same output is reconstructed in the decoder layer.

[0035] A general backpropagation algorithm is applied to the learning of the autoencoder. Note that the specific structure of the autoencoder applied when executing the support structure anomaly detection method according to the first embodiment of the present disclosure will be described in detail with reference to FIG. 4 after the description of the flowchart in FIG. 2 is finished.

[0036] Next, in the fourth step of the abnormality diagnosis step of step S205, the feature amount calculation unit 22 calculates a monitoring feature amount as a feature amount from the result of arithmetic processing on the measurement information during daily monitoring.

[0037] Next, in the fifth step of the abnormality diagnosis step of step S206, the abnormality determination processor 24 performs abnormality detection of the support structure using the autoencoder trained in step 3. Specifically, in the fifth step, the abnormality determination processor 24 inputs the monitoring feature calculated in step 4 during daily monitoring to the encoder layer of the trained autoencoder.

[0038] The abnormality determination processing unit 24 then outputs the data reconstructed by the decoder layer of the autoencoder, and performs abnormality detection of the support structure based on the error between the input and the output.

[0039] 4 is a structural diagram of an autoencoder applied when executing the support structure anomaly detection method according to the first embodiment of the present disclosure. The autoencoder 100 according to the first embodiment is configured with a total of seven layers, for example, three encoder layers 101, one feature learning layer 102, and three decoder layers 103.

[0040] The measurement frequency is 100 Hz. The spectrogram of the acceleration signal for 5 seconds is converted into an image with a frequency resolution of 64 pixels and a time resolution of 5 pixels, and the results are shown as the input and output of the autoencoder.

[0041] The anomaly detection method according to the first embodiment is based on the assumption that it is not possible to reconstruct feature data with a different tendency from that used in the learning of the autoencoder, that is, feature data obtained by converting data measured by a sensor when a change in the condition of a road structure in service occurs.

[0042] The learning process is as follows: Step 1 and Step 2. Step 1: Convert the physical quantity measured by the sensor 10 into a feature quantity that shows a change when an abnormality occurs in the rigidity of the support structure. For example, acceleration is measured as the physical quantity and converted into a spectrogram that corresponds to a feature quantity that also includes time changes in frequency characteristics. Step 2: The spectrogram is input to the encoder layer 101, and training is performed to minimize the reconstruction error of the training data.

[0043] By performing learning in steps 1 and 2, the autoencoder 100 after learning becomes a model that can correctly and accurately reproduce data only for features converted based on measurement results under normal conditions.

[0044] Therefore, the abnormality determination processing unit 24 can determine whether the support structure to be diagnosed is abnormal or normal from the reconstruction error when the converted feature amount during monitoring is input to the model created by the machine learning unit 23. The sum of squares loss of the reconstruction error used for the determination is shown in the following equation (1). Note that, as an example of the learning rate, -4 The number of epochs was set to 100.

[0045]

number

[0046] 5A and 5B are explanatory diagrams relating to the reconstruction error in the first embodiment of the present disclosure. Fig. 5A shows the transition state of the reconstruction error by the autoencoder 100 in a normal state in which no fluctuation is observed in the fixed state of the support structure. Fig. 5B shows the transition state of the reconstruction error by the autoencoder 100 in an abnormal state in which a fluctuation in the dominant frequency of the support structure occurs from time T1 in the figure.

[0047] In Figures 5(A) and 5(B), the horizontal axis represents the elapsed time over approximately one year, and the vertical axis represents the reconstruction error. In Figures 5(A) and 5(B), the solid line represents the reconstruction error, and the dotted line represents the 10-day average of the reconstruction error.

[0048] As shown in Fig. 5(A), under normal conditions, a temporary increase in the reconstruction error is observed, but the 10-day average trend is stable. On the other hand, as shown in Fig. 5(B), after time T1 when the abnormal condition occurred, the reconstruction error clearly shows an upward trend.

[0049] Fig. 6 is a diagram showing a histogram of reconstruction errors according to the first embodiment of the present disclosure. In Fig. 6, the horizontal axis represents reconstruction errors and the vertical axis represents frequency, showing the frequency of reconstruction errors before and after the occurrence of an abnormality in Fig. 5(B) above.

[0050] Comparing the distribution 601 before the occurrence of the abnormality with the distribution 602 after the occurrence of the abnormality, the distribution shapes are different, and the average value of the distribution 602 after the occurrence of the abnormality is higher than the average value of the distribution 601 before the occurrence of the abnormality, making it quantitatively clear that the reconstruction error increases after time T1.

[0051] Therefore, the anomaly determination processing unit 24 can stably detect anomalies related to the support structure over the long term, and can detect anomalies by appropriately setting the threshold for the reconstruction error in accordance with the required reliability based on the distribution of the reconstruction error during learning, and by monitoring the state in which the reconstruction error exceeds the threshold over time.

[0052] Regarding the setting of statistical thresholds, it is conceivable that changes in distribution averages can be detected based on interval estimation of the population mean, and changes in distribution shape can be detected based on interval estimation of the population variance.

[0053] The abnormality determination processing unit 24 can also calculate the sum of squared residuals regarding the error between the input and output of the autoencoder 100, determine a threshold value through statistical testing of the sum of squared residuals over a specified period, and perform abnormality detection of the support structure.

[0054] In addition, the abnormality determination processing unit 24 can calculate the sum of squared residuals regarding the error between the input and output of the autoencoder 100, estimate the interval of the population from the distribution of the sum of squared residuals over a specified period, and perform abnormality detection of the support structure by threshold processing based on the interval estimation.

[0055] Finally, a specific arrangement of the sensor 10 will be further explained with reference to FIGS. 7 is an explanatory diagram showing a specific arrangement of sensors 10 when an I-shaped information board is used as a road structure in use in the first embodiment of the present disclosure. I-shaped information board 210 shown in FIG. 4 has a support structure including information board 211 and support posts 212.

[0056] In this type of support structure, P71 to P74 shown in Fig. 7 are examples of important positions that correspond to the poles of the vibration resonance frequency or the parts where stress is concentrated, and therefore, sensors 10 are attached to these positions.

[0057] 8 is an explanatory diagram showing a specific arrangement of sensors 10 when a gate-shaped information board is used as an in-service road structure in the first embodiment of the present disclosure. Gate-shaped information board 220 shown in Fig. 8 has a support structure including information board 221 and support posts 222. Support post 222 is divided into horizontal support post 222H that is parallel to information board 221 and vertical support posts 222V that are provided on both sides of horizontal support post 222H.

[0058] In this type of support structure, the base of the support, the shoulder, the center of the horizontal support, and the quartiles of the support 222 correspond to the poles of the vibration resonance frequency or the parts where stress is concentrated. Therefore, Figure 8 shows an example in which sensors 10 are arranged at equal intervals at nine locations: P81, P82(1), P82(2), P83(1), P83(2), P84(1), P84(2), P85(1), and P85(2).

[0059] 9 is an explanatory diagram showing a specific arrangement of sensors 10 when a bridge is used as a road structure in service in the first embodiment of the present disclosure. Bridge 230 shown in FIG. 9(A) has a support structure including main girder 231 and deck 232. Bridge 240 shown in FIG. 9(B) has a support structure including main girder 241, deck 242, and pier 243 provided at the center of main girder 241 in the longitudinal direction.

[0060] In the support structure of the type shown in Figure 9(A), the center, ends, and quartiles of the main girder 231 supporting the deck slab 232 correspond to the poles of the vibration resonance frequency or the areas where stress is concentrated. Therefore, Figure 9(A) shows an example in which sensors 10 are placed at five locations: P91, P92(1), P92(2), P93(1), and P93(2).

[0061] In addition, in the support structure of the type shown in Figure 9(B), the center and end portions of the left side of the pier 243 of the main girder 241 supporting the deck slab 242 and the right side of the pier 243 of the main girder 241 correspond to the poles of the vibration resonance frequency or the areas where stress is concentrated. Therefore, Figure 9(B) shows an example in which sensors 10 are placed at six locations: P94(1) to P94(4), P95(1), and P95(2).

[0062] The physical quantities measured by the sensor 10 depend on the environmental temperature, and therefore the feature quantities converted based on the measurement results of the physical quantities are also affected by the environmental temperature.

[0063] Therefore, in order to prevent the accuracy of abnormality diagnosis from being deteriorated due to the environmental temperature, it is possible to measure the environmental temperature at the time when the physical quantity is measured as temperature information, store data correlating the physical quantity and the temperature information in a memory unit as measurement information, and take the environmental temperature into consideration when considering the converted feature quantity as well.

[0064] Specifically, when collecting data, the index value data collector 21 acquires temperature information as well as physical quantities, and stores data associating the physical quantities with the temperature information in the storage unit as measurement information.

[0065] Furthermore, when calculating the normal state feature amount, the feature amount calculation unit 22 can calculate the normal state feature amount for each temperature for each environmental temperature by referring to the temperature information included in the measurement information.

[0066] Furthermore, the machine learning unit 23 can calculate a weighting coefficient for learning according to the environmental temperature by performing machine learning for each environmental temperature using the normal state feature amounts for each temperature.

[0067] Furthermore, when calculating the monitoring feature, the feature calculation unit 22 can identify the learning weighting coefficient according to the corresponding environmental temperature by referring to the temperature information included in the measurement information.

[0068] Then, by using the identified learning weighting coefficients, the abnormality determination processing unit 24 can detect abnormalities in the support structure using a trained autoencoder weighted according to the ambient temperature, which makes it possible to suppress the influence of the ambient temperature and to prevent deterioration in the accuracy of abnormality diagnosis.

[0069] When the characteristic amount calculation unit 22 calculates the normal characteristic amount for each temperature, it is also possible to adopt the following steps 1 to 3. Step 1: The index value data collecting unit 21 sets in advance the range of the environmental temperature when the sensor 10 measures the physical quantity as a plurality of environmental temperature ranges equally divided at predetermined temperature intervals.

[0070] Step 2: When collecting data on physical quantities, the index value data collection unit 21 refers to the temperature information, classifies the physical quantities associated with the temperature information as data for the corresponding environmental temperature range among multiple environmental temperature ranges, and stores the data in the memory unit as measurement information by temperature.

[0071] Step 3: When calculating the normal state feature amount, the feature amount calculation unit 22 calculates the normal state feature amount by temperature based on the measurement information by temperature classified as data in the environmental temperature range.

[0072] In this way, by classifying the collected physical quantities into a plurality of environmental temperature ranges equally divided at predetermined temperature intervals, it is possible to efficiently calculate normal state feature quantities for each temperature.

[0073] As described above, according to the first embodiment, physical quantities that serve as indicators for diagnosing deterioration of the support structure are collected as measurement information, and abnormalities in the support structure are detected using an autoencoder that inputs the feature quantities obtained by converting the physical quantities.

[0074] In particular, the method for diagnosing deterioration of a support structure according to the first embodiment can achieve the following effects. Effect 1: The sensor measurement results are converted into feature quantities that show changes when an abnormality occurs in the rigidity of the support structure, and the converted feature quantities are input to the autoencoder. Specifically, the time-series information of acceleration data measured as a physical quantity is converted into feature quantities that also include time changes in frequency characteristics, such as spectrograms, scalograms, or frequency spectra, and these are input to the autoencoder.

[0075] As a result, it is possible to diagnose deterioration of the support structure using physical quantities that change when an abnormality occurs in the rigidity of the support structure, and it is possible to realize a method for detecting abnormalities in the support structure that can accurately distinguish between normal and abnormal conditions using an autoencoder.

[0076] Effect 2: A threshold for detecting anomalies can be set as a statistically significant value based on the transition state of the reconstruction error, which is the output of a trained autoencoder. Specifically, an appropriate threshold can be set by statistically testing the sum of squares of the residuals between the input and output. Therefore, additional measurement data is not required to set a threshold for determining anomalies.

[0077] With these effects 1 and 2, it is possible to easily detect abnormalities in support structures by retrofitting sensors to existing road structures in service. Furthermore, although autoencoders require calculations such as learning, they are a nonlinear method and excel at stably monitoring minute and complex changes in the support structures that are the subject of diagnosis over the long term.

[0078] Therefore, by using an autoencoder that realizes Effects 1 and 2, it is possible to obtain a method for detecting anomalies in a support structure that can prevent serious accidents and perform anomaly detection with high accuracy. [Explanation of symbols]

[0079] 10 Sensor, 20 Controller, 21 Index value data collection unit, 22 Feature calculation unit, 23 Machine learning unit, 24 Anomaly detection processing unit, 100 Autoencoder, 101 Encoder layer, 102 Feature learning layer, 103 Decoder layer, 200 F-type information board, 201 Information board, 202 Pillar, 210 I-type information board, 211 Information board, 212 Pillar, 220 Gate-type information board, 221 Information board, 222 Pillar, 222V Vertical pillar, 222H Horizontal pillar, 230 Bridge, 231 Main girder, 232 Deck slab, 240 Bridge, 241 Main girder, 242 Deck slab, 243 Pier.

Claims

1. a measuring step of measuring a physical quantity that serves as an index value for detecting an abnormality in a support structure included in a road structure in service by a sensor arranged on the support structure; an abnormality diagnosis step of calculating a feature amount that shows a change when an abnormality occurs in the rigidity of the support structure based on the measurement result by the sensor, and detecting an abnormality in the support structure based on the feature amount; A method for detecting an abnormality in a support structure, comprising: The abnormality diagnosis step includes: a first step of storing the physical quantity acquired continuously or periodically as the measurement result from the sensor in a storage unit as measurement information; a second step of calculating a normal state feature quantity as the feature quantity from a calculation processing result of the measurement information when the support structure is normal; a third step of performing machine learning on an autoencoder used for anomaly monitoring, which is composed of an encoder layer, a feature learning layer, and a decoder layer, such that the normal feature calculated in the second step is input to the encoder layer and the same output is reconstructed in the decoder layer; a fourth step of calculating monitoring feature quantities as the feature quantities from calculation processing results for the measurement information during daily monitoring; a fifth step of, during the daily monitoring, using the monitoring feature calculated in the fourth step for the autoencoder trained in the third step as an input to the encoder layer, using reconstructed data in the decoder layer as an output, and detecting an abnormality in the support structure based on an error between the input and the output; A method for detecting an abnormality in a support structure, comprising:

2. the measuring step measures the physical quantity and further measures temperature information that is an environmental temperature of the support structure; The abnormality diagnosis step includes: In the first step, the temperature information is acquired together with the physical quantity, and data in which the physical quantity and the temperature information are associated with each other is stored in the storage unit as the measurement information; In the second step, when calculating the normal state characteristic amount, the normal state characteristic amount for each temperature is calculated for each environmental temperature by referring to the temperature information included in the measurement information, and the normal state characteristic amount for each temperature is stored in the storage unit; In the third step, machine learning is performed for each environmental temperature using the normal feature amounts for each temperature, thereby calculating a learning weight coefficient according to the environmental temperature; In the fourth step, by referring to the temperature information included in the measurement information, a weighting coefficient for the learning corresponding to the calculated environmental temperature corresponding to the monitoring feature is identified; In the fifth step, by using the weighting coefficients of the learning identified in the fourth step, anomaly detection of the support structure is performed using the trained autoencoder weighted according to the environmental temperature. The method for detecting an abnormality in a support structure according to claim 1 .

3. The abnormality diagnosis step includes: a plurality of environmental temperature ranges are set in advance by equally dividing the range of environmental temperature at predetermined temperature intervals when the physical quantity is measured by the sensor; In the first step, the temperature information is acquired together with the physical quantity, the physical quantity associated with the temperature information is classified as data of a corresponding environmental temperature range among the plurality of environmental temperature ranges, and the classified data is stored in the storage unit as measurement information by temperature; In the second step, when the normal state characteristic amount is calculated, the normal state characteristic amount for each temperature is calculated based on the measurement information for each temperature. The method for detecting an abnormality in a support structure according to claim 2.

4. The abnormality diagnosis step includes: In the fifth step, a residual sum of squares of the error is calculated based on the error between the input and the output, and an abnormality of the support structure is detected by a statistical test of the residual sum of squares for a predetermined period. The method for detecting an abnormality in a support structure according to claim 3.

5. The abnormality diagnosis step includes: In the fifth step, a residual sum of squares of the error is calculated based on the error between the input and the output, an interval of a population is estimated from a distribution of the residual sum of squares over a predetermined period, and an anomaly detection of the support structure is performed by threshold processing based on the interval estimation. The method for detecting an abnormality in a support structure according to claim 3.

6. The abnormality diagnosis step calculates a spectrogram as the feature quantity by performing a short-time Fourier transform based on the measurement result by the sensor, and detects an abnormality in the support structure based on the spectrogram. The method for detecting an abnormality in a support structure according to any one of claims 1 to 5.

7. The abnormality diagnosis step calculates a scalogram as the feature quantity by performing a wavelet transform based on the measurement result by the sensor, and detects an abnormality in the support structure based on the scalogram. The method for detecting an abnormality in a support structure according to any one of claims 1 to 5.

8. The abnormality diagnosis step calculates a frequency spectrum as the feature quantity by performing an FFT analysis based on the measurement result by the sensor, and detects an abnormality in the support structure based on the frequency spectrum. The method for detecting an abnormality in a support structure according to any one of claims 1 to 5.

9. The sensor is configured as a plurality of sensors, and is installed so as to include a portion of the support structure where stress is concentrated or a position corresponding to a pole of the resonance frequency of the support structure, and is installed at equal intervals from one another. The method for detecting an abnormality in a support structure according to any one of claims 1 to 5.

10. The sensor is an acceleration sensor. The method for detecting an abnormality in a support structure according to any one of claims 1 to 5.

Images