Abnormal part identification device

Abstract

The present disclosure is an abnormal part identification device that uses frequency components of vibration data at a plurality of points in the longitudinal direction of an optical fiber to calculate an evaluation value representing the degree of abnormality for each of the plurality of points, and compares the evaluation values at the plurality of points to identify an abnormal part in the longitudinal direction of the optical fiber.

Description

Anomaly identification device

[0001] This disclosure relates to an apparatus for identifying abnormal locations using vibration data acquired using a distributed optical fiber vibration sensor (DAS).

[0002] Technologies are being explored for remotely monitoring and detecting anomalies in infrastructure facilities such as roads, water pipes, and gas pipes, as well as communication equipment such as fiber optic cables, which require daily checks for their integrity. For example, Non-Patent Document 1 describes a method for detecting anomalies in objects where IoT (Internet of Things) sensors are installed, by combining one-dimensional time-series data obtained from IoT sensors with deep learning.

[0003] A distributed optical fiber (DAS) has been proposed that enables remote monitoring by connecting optical fibers (see, for example, Non-Patent Document 2). With a DAS, vibrations can be measured at each point in the optical fiber, so if there is a change in vibration, that point can be identified as an anomaly. However, the inventors have discovered that there are anomalies that cannot be detected by observing only that point.

[0004] Chunyong Yin et al., “Anomaly Detection Based on Convolutional Recurrent Autoencoder for IoT Time Series”, IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 52, NO. 1, JANUARY 2022. Lasertec Corporation Technical Information “Optical Fiber Distributed Vibration Sensor (DAS)” https: / / lazoc.jp / technical / das / Mathworks Autoencoder https: / / jp.mathworks. com / discovery / autoencoder. html

[0005] DAS allows for the observation of the distribution of optical fibers. Therefore, this disclosure aims to enable the detection of anomalies that cannot be detected by observation at a single point by comparing vibration data at multiple points along the longitudinal direction of the optical fiber.

[0006] The abnormal location identification system of this disclosure comprises an optical fiber installed in an object to be inspected, a vibration measuring instrument for measuring the vibration of the optical fiber, and an abnormal location identification device that analyzes the vibration data measured by the vibration measuring instrument and identifies abnormal locations in the object to be inspected.

[0007] The abnormal location identification device of this disclosure performs the abnormal location identification method of this disclosure. The abnormal location identification method of this disclosure involves the abnormal location identification device calculating an evaluation value representing the degree of abnormality for each of the multiple locations using the frequency components of vibration data at multiple locations in the longitudinal direction of the optical fiber, and identifying the abnormal location in the longitudinal direction of the optical fiber by comparing the evaluation values ​​at the multiple locations.

[0008] The abnormal location identification device may create input data for the multiple locations using vibration data from the multiple locations, input the input data for the multiple locations into a pre-learned normal data model to calculate reconstructed data for the multiple locations, compare the reconstructed data and the input data for each location to calculate the reconstruction error for the multiple locations, and use the reconstruction error to calculate the evaluation value for the multiple locations.

[0009] By using the reconstruction error as the evaluation value, the probability distribution to which the evaluation value follows can be aligned along the longitudinal direction of the optical fiber. Therefore, the standard for the evaluation value can be aligned at each point along the longitudinal direction of the optical fiber.

[0010] In this disclosure, anomaly detection may be performed by combining machine learning and statistical processing, using input data that emphasizes the relationship of vibration characteristics in the distance direction. For example, the anomaly identification device may divide the vibration data of the optical fiber into sections in the longitudinal direction of the optical fiber, calculate the frequency spectrum for each section using the vibration data for each section, and normalize the spectral intensity values ​​in the distance direction for each section to create the input data.

[0011] By arranging the frequency spectra obtained from vibration data acquired by DAS according to distance and normalizing each divided section in the distance direction, the distance-direction relationship of the frequency characteristics can be emphasized and used as input data for machine learning. In this case, the resolution of the frequency spectrum may be reduced. This can improve noise immunity.

[0012] The anomaly location identification device may accumulate the evaluation value for each point along the longitudinal direction of the optical fiber and identify the point along the longitudinal direction of the optical fiber with the highest accumulated value. In this case, the anomaly location identification device may calculate the evaluation value at each of the multiple points for each point determined by the longitudinal direction of the optical fiber and the frequency.

[0013] Furthermore, the above disclosures can be combined as much as possible.

[0014] According to this disclosure, anomalies that cannot be detected by observation at a single location can be detected by comparing vibration data from multiple locations along the longitudinal direction of the optical fiber.

[0015] This shows an example of the system configuration of this disclosure. This is an example of an anomaly that cannot be detected by observing only one point. This shows an example of the average frequency spectrum before and after repair. This shows an example of the distance distribution of the average spectral intensity. This shows an example of an embodiment of the pre-processing in this disclosure. This is an explanatory diagram of the operation of the pre-processing unit. This shows an example of the model generation method of this embodiment. This is an explanatory diagram of the operation of the generation unit. This is an explanatory diagram of the operation of the generation unit. This shows an example of an embodiment of the anomaly location identification method of this disclosure. This is an explanatory diagram of the operation of the detection unit. This is an explanatory diagram of the operation of the detection unit. This shows an example of the model generation method of this embodiment. This is an explanatory diagram of the operation of the generation unit. This is an explanatory diagram of the operation of the detection unit.

[0016] Embodiments of this disclosure will be described in detail below with reference to the drawings. However, this disclosure is not limited to the embodiments shown below. These examples are illustrative, and this disclosure can be implemented in various modified and improved forms based on the knowledge of those skilled in the art. In this specification and in the drawings, components with the same reference numerals refer to the same components.

[0017] (System Configuration) Figure 1 shows an example of the system configuration of the present disclosure. The system of this embodiment includes an optical fiber 12 installed on the object to be inspected 11, a vibration measuring instrument 13 that measures vibration at each point along the longitudinal direction of the optical fiber 12, and an abnormality location identification device 14 that analyzes the vibration data measured by the vibration measuring instrument 13.

[0018] The optical fiber 12 receives vibrations at any point inside or outside the object to be inspected 11 that is laid within a range where vibrations transmitted to the object to be inspected 11 are similarly transmitted. The object to be inspected 11 is any object to which vibrations are transmitted to the optical fiber 12, and examples include infrastructure facilities such as roads, water pipes, and gas pipes, and communication equipment such as optical fiber cables. The object to be inspected 11 may also be the optical fiber itself that receives the vibrations.

[0019] The vibration measuring instrument 13 measures the vibrations received by the optical fiber 12 as a distribution along the longitudinal direction of the optical fiber 12. The vibration measuring instrument 13 is any measuring device to which DAS (Distributed Acoustic Sensing) technology can be applied. For example, a C-OTDR using a light pulse light source with a stable optical frequency and a light receiving circuit for optical heterodyne detection or optical homodyne detection can be exemplified.

[0020] (Summary of this disclosure) Here, we will explain using the example of the detachment of an overhead optical fiber cable as shown in Figure 2. Using DAS, vibrations transmitted to the cable were acquired for 3 weeks and 6 weeks before and after repair, respectively, and the average frequency spectrum of "only the locations where detachment occurred" was compared.

[0021] Figure 3 shows an example of the average frequency spectrum before and after repair. The solid line shows the measurement results for the three weeks before repair, and the dashed line shows the measurement results from immediately after repair to the three weeks after repair. It can be seen that the change in frequency characteristics before and after repair is small for the average frequency spectrum itself. For this reason, it is difficult to determine whether the overhead optical fiber cable has fallen out by comparing the frequency characteristics at each point along the longitudinal direction of the optical fiber 12.

[0022] On the other hand, DAS measurement data also allows observation of characteristic changes in the distance direction. Figure 4 shows an example of the distance distribution of average spectral intensity. The solid line shows the measurement results for three weeks before repair, the dashed line shows the measurement results from immediately after repair to three weeks after repair, and the dotted line shows the measurement results from three weeks after repair to six weeks after repair. All measurement results show examples where the average spectral intensity is the average value normalized in the distance direction of spectral intensity in the frequency band between 10 Hz and 15 Hz.

[0023] This graph shows a significant difference in the area around the point of detachment (980m) before and after repair. Therefore, by comparing the distance distribution of the average spectral intensity of optical fiber 12 within the same frequency band, it is possible to determine if an overhead optical fiber cable has detached.

[0024] Thus, when viewed along the distance direction of the optical fiber 12, the system vibrates due to the same vibration source, and the specific band components at each point in the system are close in value. However, the amount of change when the relative magnitudes of these components change is sufficiently large compared to the magnitude of the entire sequence of specific band components in the system, making it possible to ensure sufficient dynamics of the change.

[0025] The amount of change represents the degree of abnormality. Therefore, this disclosure uses the frequency components of vibration data at multiple points along the longitudinal direction of the optical fiber 12 to calculate an evaluation value representing the degree of abnormality for each of the multiple points, and identifies the location of abnormality in the longitudinal direction of the optical fiber 12 by comparing the evaluation values ​​at the multiple points.

[0026] This example illustrates the detachment of an overhead optical fiber cable, but by using the average spectral intensity of an appropriate frequency band depending on the object being observed, it is possible to determine anomalies in the monitored object. Thus, this disclosure provides an anomaly location identification system that can identify anomalies that cannot be determined as an anomaly based on observation of only one location, as well as an anomaly affecting the entire system.

[0027] (First Embodiment) In the abnormal location identification system according to this embodiment, the abnormal location identification device 14 identifies the abnormal location using machine learning. Specifically, the abnormal location identification device 14 includes a generation unit 42 that generates a machine learning model, a detection unit 43 that identifies the abnormal location using the model, and a pre-processing unit 41 that generates input data for the generation unit 42 and the detection unit 43.

[0028] The preprocessing unit 41 creates input data by calculating the frequency spectrum obtained from the vibration data for each longitudinal section of the optical fiber 12. In this embodiment, the vibration data is divided into longitudinal sections of the optical fiber 12, the frequency spectrum for each section is calculated using the vibration data for each measurement point in each section, and the input data is created by normalizing the spectral intensity values ​​for each section.

[0029] In this embodiment, the detection unit 43 inputs the input data from the multiple locations into a pre-learned data model for normal operation to calculate reconstructed data for the multiple locations. The detection unit 43 then compares the reconstructed data and the input data for each location to calculate the reconstruction error for the multiple locations, and uses the reconstruction error to calculate the evaluation value for the multiple locations. The detection unit 43 then accumulates the evaluation value for each location along the longitudinal direction of the optical fiber 12 and identifies the location along the longitudinal direction of the optical fiber 12 with the highest accumulated value.

[0030] Machine learning can use any method capable of generating statistical distribution information, for example, deep learning using an autoencoder [Non-Patent Literature 3]. In this embodiment, an example is shown in which the generation unit 42 performs deep learning using an autoencoder. The autoencoder may be any of the following: stacked autoencoder, convolutional autoencoder, variational autoencoder, or conditional variational autoencoder.

[0031] (During learning) Figure 5 shows an example of an embodiment of the preprocessing described herein. The preprocessing in this embodiment is performed by executing steps S1 to S4 in order.

[0032] (Procedure S1) Understanding the positional relationship between the equipment to be inspected and the optical fiber. Investigate which distance in the longitudinal direction of the optical fiber 12 that receives vibrations at an arbitrary point corresponds to which position on the object to be inspected 11. If the object to be inspected 11 is the optical fiber 12 that receives vibrations, investigate the longitudinal distance of the optical fiber 12 and the geographical location of the optical fiber 12.

[0033] (Procedure S2) When it is determined that the object 11 to be inspected is in a normal state, the vibration transmitted to the optical fiber 12 is measured using the vibration measuring instrument 13. The measurement period can be any time for which a model can be created using machine learning.

[0034] (Procedure S3) The pre-processing unit 41 for calculating vibration patterns in a normal state divides the vibration data acquired in procedure S2 in the distance direction of the optical fiber 12 and generates vibration time-series data for each point in the longitudinal direction of the optical fiber 12. In this embodiment, the pre-processing unit 41 further divides the vibration data acquired in procedure S2 in the time direction. This results in the vibration time-series data D shown in Figure 6. ３１ As shown above, a graph is obtained that shows the amplitude over 30 seconds at intervals of 50 meters along the longitudinal direction of the optical fiber 12.

[0035] Here, the division into distance and time directions can be set to any values ​​that you want to use to identify anomalies; for example, 50m and 30 seconds can be used. In this embodiment, we show an example where the vibration data is divided in both the distance and time directions, but it is also possible to divide it in only one of the distance or time directions.

[0036] Next, the preprocessing unit 41 obtains frequency spectra from vibration time-series data for each point along the longitudinal direction of the optical fiber 12. For example, the preprocessing unit 41 performs a Fourier transform on the vibration time-series data to obtain frequency spectra D every 50 meters and every 30 seconds along the longitudinal direction of the optical fiber 12. ３２ You can obtain this.

[0037] Next, the preprocessing unit 41 analyzes the frequency spectrum D of the optical fiber 12 at 50m intervals in the longitudinal direction. ３２Reduce the resolution in the frequency direction. The reduction of the resolution in the frequency direction can be exemplified by calculating the average value for each size of, for example, 2 Hz. By reducing the resolution in this way, the noise tolerance can be improved.

[0038] Next, the preprocessing unit 41 normalizes the frequency spectrum intensity of the frequency spectrum D with reduced resolution to a value in the range of 0 to 1 in the distance direction. As a result, the vibration pattern data D in the normal state ３２ is obtained every 50 m and every 30 s in the longitudinal direction of the optical fiber 12. In the present disclosure, the vibration pattern data in the normal state is referred to as "reference data". ３３

[0039] (Procedure S4) Model Generation FIG. 7 shows an example of the model generation method of the present embodiment. The model generation method of the embodiment includes the generation unit 42 having procedures S411 to S416 in order. The generation unit 42 inputs the reference data in the normal state calculated in procedure S3 into the autoencoder to learn the normal state and generate a data model indicating the normal state (S411). Here, when the division in the distance direction is performed in procedure S3, a data model is created for each division section.

[0040] Next, as shown in FIG. 8, the generation unit 42 inputs a plurality of reference data D used for learning ３３ into the autoencoder 44 after learning the data model, and creates reconstruction data D ３３ for each of the respective reference data D ４１ (S412). As a result, the reconstruction data D ４１ is obtained for each reference data D ３３

[0041] Next, as shown in FIG. 8, the generation unit 42 calculates the square of the difference between the reference data D ３３ and the reconstruction data D ４１ for each reference data D ３３ (S413). As a result, the squared data D ４２ of the reconstruction error is obtained for each reference data D ３３

[0042] The reference data D ３３ and the reconstruction data D ４１ ​​​Since both are two-dimensional distributions of distance and frequency, the squared data D of the reconstruction error ４２ This is a two-dimensional distribution of distance and frequency. The generation unit 42 generates the squared data D of the reconstruction error at different times, as shown in Figure 9. ４２ From this, extract values ​​of the same distance and frequency, and obtain the anomaly degree sequence data D. ４３ Create (S414). Here, extraction only requires extracting as many data as there are data points that make up the data set. This creates a sequence of abnormality data D at each distance and frequency in the longitudinal direction of the optical fiber 12. ４３ This is created.

[0043] As shown in Figure 9, the generation unit 42 generates the distribution D of the series data of the anomaly degree. ４４ The distribution of the series data of the anomaly score is calculated (S415). For example, as shown in Figure 9, D ４４ When the distribution follows a log-normal distribution, the generation unit 42 takes the logarithm of the values ​​in the abnormality sequence data and calculates the probability distribution of the log-sequence data. This gives the probability distribution for normal conditions.

[0044] Then, the generation unit 42 generates the distribution D of the series data of the anomaly degree. ４４ Using this, the average value of the degree of abnormality in a normal state D ４５ and standard deviation D ４６ The distribution D of the series data of the anomaly degree is calculated and stored in the detection unit 43. The generation unit 42 also calculates and stores the distribution D of the series data of the anomaly degree. ４４ Convert the logarithm to a real number and use an arbitrary percentage as the threshold (threshold percentage D). ４７ The threshold value is calculated and stored in the detection unit 43 (S416). This threshold value can be any value set based on how far it deviates from the normal distribution to determine if it is abnormal. For example, in the case of the probability distribution shown in Figure 9, the value D, which is located at 90% in that probability distribution, is used. ４８ The 90th percentile is calculated, and this is converted to a numerical value to obtain the threshold percentile D of the anomaly series data. ４７ It can be set to that.

[0045] (During monitoring) Figure 10 shows an example of an embodiment of the abnormal location identification method of this disclosure. The abnormal location identification method of this embodiment is performed by executing steps S5 to S10 in order.

[0046] (Procedure S5) While monitoring the object 11 to be inspected, the vibrations transmitted to the optical fiber 12 are measured using the vibration measuring instrument 13.

[0047] (Procedure S6) The pre-processing unit 41 for calculating vibration patterns in the monitoring state generates vibration pattern data from the vibration data acquired in procedure S5 using the same procedure as in procedure S3. Specifically, the pre-processing unit 41 divides the vibration data acquired in procedure S5 into 50m in the distance direction and 30s in the time direction, and generates vibration time series data D ３１ The vibration time series data corresponding to this is generated. At this time, the division is performed at the same time interval and distance as the division in step S3. The vibration data obtained in this way is called "detection target data".

[0048] Then, the preprocessing unit 41 processes the frequency spectrum D ３２ Convert to the corresponding frequency spectrum and reference data D ３３ This generates vibration mode data corresponding to the above. This generates the vibration mode data D shown in Figure 11. ５１ However, this can be obtained every 50m and every 30s along the longitudinal direction of the optical fiber 12.

[0049] (Procedure S7) The reconstruction detection unit 43 using the trained data model, as shown in Figure 11, uses the vibration pattern data D calculated in procedure S6. ５１ Then, the data model saved in step S4 is input to the trained autoencoder 44, and the reconstructed data D for each input data is input. ７１ This creates the reconstructed data D. ７１ Vibration pattern data D ５１ Obtained each time.

[0050] Next, the detection unit 43, in the same manner as in step S4, generates vibration phase data D ５１ and reconstructed data D ７１ The difference in reconstruction errors is used for vibration mode data D. ５１ This is calculated for each. This gives the reconstruction error at each distance in the longitudinal direction of the optical fiber 12. In this embodiment, the square of this reconstruction error is used for the vibration phase data D ５１ This is calculated for each step. As a result, the squared reconstruction error data D is obtained, as shown in Figure 11. ７２Vibration pattern data D ５１ You can obtain it each time.

[0051] If a data model is created for each divided section in step S4, the reconstruction error calculation process is performed for each divided section using the data for the corresponding divided section from the vibration pattern data under the monitoring state calculated in step S6 as input.

[0052] (Procedure S8) The cumulative anomaly series creation detection unit 43 generates the squared reconstruction error data D ７２ Based on this, anomalies are determined at each distance and frequency in the longitudinal direction of the optical fiber 12. When the state differs from the normal state due to an anomaly, the data model learned in the normal state cannot be reconstructed properly, and the squared reconstruction error data D, which indicates the degree of anomaly, is generated. ７２ Therefore, the detection unit 43 determines the threshold percentage D calculated in step S4. ４７ When the value is greater than this, it can be determined to be an outlier that deviates significantly from the normal distribution, and thus can be judged as abnormal.

[0053] For example, the detection unit 43 uses the squared reconstruction error data D calculated in step S7. ７２ Of these, the values ​​are compared at the same frequency and distance as the threshold percentage value calculated in step S4. The detection unit 43 sets all values ​​that do not exceed the threshold percentage value to a reference value of 0.

[0054] The detection unit 43 detects squared data D that exceeds the threshold percentage value. ７２ As shown in Figure 11, the evaluation value D is determined for each point defined by the corresponding frequency and distance. ８１ The calculated evaluation value D is then calculated. ８１ The value is converted to a real number. The detection unit 43 then evaluates the value D after conversion to a real number. ８１ This is averaged over a predetermined time interval. This gives the average value D of the evaluation. ８２ However, it can be obtained at each point defined by frequency and distance.

[0055] Rating D ８１ For example, it can be calculated using the following: D ８１ = (log(D ７２ )-D ４５ ) / D４６ However, D ４５ and D ４６ These are the mean and standard deviation at the corresponding frequency and distance, which were saved in step S4.

[0056] Next, as shown in Figure 12, the detection unit 43 measures the average value D of the evaluation value. ８２ The data is arranged in order of distance and frequency. This results in the anomaly sequence data D along the entire length of the optical fiber 12. ８３ You can obtain this.

[0057] In this embodiment, the average value D ８２ Using the abnormality series data D ８３ An example of obtaining the reconstructed data D obtained in step S7 has been shown, but this disclosure is about the reconstructed data D obtained in step S7. ７１ It is possible to use any statistical value obtained using this method. For example, the mean, sum, or median can be used as the statistical value.

[0058] (Procedure S9) The abnormality detection unit 43, as shown in Figure 12, collects the abnormality sequence data D along the entire length of the optical fiber 12. ８３ This is accumulated for each distance along the entire length of the optical fiber 12. The detection unit 43 calculates the cumulative anomaly score D, which is the cumulative value of the anomaly score sequence data obtained in this way. ９１ This is used to detect abnormal areas.

[0059] For example, the detection unit 43 detects the average value D of each point at each distance of the optical fiber 12. ８２ Select the top 5 points from among them, and calculate the average value D of the selected points. ８２ By summing them up, the cumulative abnormality level D ９１ The detection unit 43 then calculates the cumulative anomaly score D, as shown in Figure 12. ９１ A threshold T set arbitrarily by ９１ Determine the distance that exceeds the threshold T. ９１ The distance exceeding this limit, i.e., the point along the longitudinal direction of the optical fiber 12, is determined to be an abnormal location.

[0060] Note: Cumulative abnormality level D ９１ The selected mean value D ８２It is possible to use any method capable of evaluating the degree of abnormality at each distance of the optical fiber 12, not limited to the sum of the values. For example, the detection unit 43 uses a selected average value D ８２ Furthermore, statistical values ​​such as the average are calculated and this is used to determine the cumulative abnormality D ９１ That is also acceptable.

[0061] (Procedure S10) The abnormal location estimation detection unit 43 estimates where the abnormality is occurring on the object to be inspected 11, using the distance to the abnormal location determined in procedure S9 and the relationship between the distance of the optical fiber 12 and the position of the object to be inspected 11, which were investigated in advance in procedure S1.

[0062] If the reconstruction error for each segment is calculated in step S9, the cumulative anomaly score D ９１ The calculation process is performed for each divided section, and the cumulative anomaly score D for each divided section is calculated. ９１ The calculation results are arranged in descending order of distance and stored as a cumulative anomaly score series.

[0063] (Second Embodiment) In the first embodiment, in steps S4 and S8, the distribution D of the series data of the degree of abnormality ４４ While it is assumed that the distribution follows a log-normal distribution, this disclosure is not limited to this. For example, the distribution of the series data of anomaly scores D ４４ is χ ２ Distribution is also acceptable.

[0064] (Procedure S4) Model generation Figure 13 shows an example of the model generation method of this embodiment. The model generation method of this embodiment has steps S421 to S427 in order. Steps S421 and S422 are the same as steps S411 and S412. Specifically, the generation unit 42 generates the reconstructed data D as shown in Figure 14. ４１ Reference data D ３３ Generate one for each.

[0065] Next, the generation unit 42 generates the reconstructed data D ４１ and reference data D ３３ The difference is calculated using the reference data D. ３３ Calculate for each (S423). This results in the reconstruction error D for the same frequency and the same distance. ５２ Reference data D ３３ Obtained each time.

[0066] Next, the generation unit 42 calculates the average value D ５２ and the standard deviation D ５５ of the reconstruction error D for each same frequency and same distance (S424). Then, the generation unit 42 subtracts the average value D ５６ from the reconstruction error D ５２ and divides by the standard deviation D ５５ to convert it into a value of the standard normal distribution (S425). As a result, the series data of the standard normal distribution of the reconstruction error D ５６ is calculated for each same frequency and same distance. ５２

[0067] Next, the generation unit 42 squares each value of the series data of the standard normal distribution of the reconstruction error D ５２ and calculates a threshold percentage point D ５７ (for example, 94%) assuming that the series data follows a chi-square distribution (S426).

[0068] Next, the generation unit 42 stores the average value D ５２ of the reconstruction error D ５５ , the standard deviation D ５６ , and the threshold percentage point D ５７ in the chi-square distribution for each distance and frequency in the longitudinal direction of the optical fiber 12 in the detection unit 43 (S427).

[0069] (During monitoring) Procedures S5 and S6 are the same as those in the first embodiment. Regarding procedure S7, in this embodiment, as shown in FIG. 15, the detection unit 43 creates reconstruction data D ５１ obtained for each vibration mode data D ７１ . Then, the detection unit 43 calculates the difference between the reconstruction data D ７１ and the vibration mode data D ５１ for each vibration mode data D ５１ in the same manner as in procedure S423. As a result, the reconstruction error D ７４ at each distance in the longitudinal direction of the optical fiber 12 is obtained.

[0070] The detection unit 43 determines an evaluation value D ７４ for each value of the reconstruction error D ８４ ​The following is used to calculate the value. In this embodiment, for example, the following can be used to calculate it: D ８４ = (D ７４ -D ５５ ) / D ５６ However, D ５５ and D ５６ These are the mean and standard deviation at the corresponding frequency and distance, which were saved in step S4.

[0071] Rating D ８４ By calculating the reconstruction error D, ７４ The value can be converted to a standard normal distribution value. The detection unit 43 evaluates the evaluation value D ８４ The value obtained by squaring this value is the threshold percentage D calculated in step S4. ５７ The values ​​are compared with those of the same frequency and distance. The detection unit 43 is the threshold percentage D ５７ All values ​​that do not exceed the specified value are set to the baseline of 0.

[0072] The detection unit 43 uses the threshold percentage D stored in step S4. ５７ For values ​​exceeding this limit, the evaluation value D is calculated for each frequency and distance. ８４ The average of the squared values ​​is calculated. This results in the anomaly series data D shown in Figure 12. ８３ Similar to the reconstruction error D ７４ An anomaly series data is obtained, represented by the mean of the squares of the values.

[0073] For the abnormality determination procedure S9 using the abnormality degree series data, and the subsequent abnormality location estimation procedure S10, the abnormality degree series data D in the first embodiment is used. ８３ Abnormality series data D ８５ You can simply rephrase it as follows.

[0074] (Other Embodiments) In the embodiments described above, the preprocessing unit 41 divided the vibration data every 50 m and every 30 s in the longitudinal direction of the optical fiber 12. However, the division here can be any value suitable for modeling such as the autoencoder 44, and the division may be in the distance direction only, or in the time direction only.

[0075] Furthermore, the vibration pattern data input to the autoencoder 44 in the above-described embodiment may be image data obtained by converting the spectral intensity of the vibration pattern data into pixel intensity.

[0076] Furthermore, the abnormal location identification device 14 can also be implemented using a computer and a program, and the program can be recorded on a recording medium or provided via a network.

[0077] As described above, the preprocessing unit 41 divides the vibration data acquired by the vibration measuring instrument 13 into multiple sections and performs a process to calculate the frequency characteristics of the vibration data for each measurement point in each section. This creates input data for deep learning.

[0078] The generation unit 42 then generates a model using deep learning based on the reference data created by the preprocessing unit 41. The generation unit 42 further generates statistical distribution information of the calculated anomaly score based on the output data obtained by inputting the reference data into the model.

[0079] The detection unit 43 detects abnormal locations in the data using the abnormality level calculated based on the output data obtained by inputting the detection target data created by the preprocessing unit 41 into the model, and the statistical distribution information of the abnormality level obtained by inputting the reference data into the model, and identifies the abnormal locations by comparing them with the distance information in the longitudinal direction of the optical fiber cable.

[0080] Here, the input data for deep learning is a distance spectrogram obtained by arranging the spectra in the distance direction. That is, the preprocessing unit 41 of this disclosure may obtain frequency spectra from time-series data for each distance in the longitudinal direction of the optical fiber cable from vibration data for each measurement point in the section, create a series of frequency spectra for each distance arranged in the distance direction, and use the series as reference data and detection target data.

[0081] In this embodiment, the preprocessing unit 41 reduces the resolution of the distance spectrogram and normalizes the spectral intensity in the distance direction. That is, the preprocessing unit 41 of this disclosure may perform the following processing: divide the frequency spectrum for each distance into arbitrary bands, calculate the average value for each divided band and arrange them in order of frequency to create a resolution-reduced frequency spectrum; create data of a group of frequency spectra by arranging the resolution-reduced frequency spectra for each distance in the longitudinal direction of the optical fiber cable; and create data by converting the spectral intensity values ​​of the group of frequency spectra to values ​​in the range of 0 to 1 for each series in the distance direction using an arbitrary method.

[0082] In this embodiment, the generation unit 42 may identify distances with high anomaly scores using statistical values ​​of anomaly scores. That is, the generation unit 41 may input the reference data into the generated model and, based on the output data obtained, calculate a statistical distribution of possible anomaly scores for each distance and frequency. Here, the statistical distribution information of anomaly scores may be the parameters of the probability distribution of possible anomaly scores for each distance and frequency, and the values ​​of arbitrary percent points of the probability components.

[0083] Furthermore, the detection unit 43 may convert the degree of anomaly calculated based on the output data obtained by inputting the data to be detected into the model into a value that follows the same probability distribution based on the parameters of the probability distribution calculated by the generation unit, retain only the degree of anomaly that exceeds the percentage point calculated by the generation unit, convert the rest into a standard value, and create an anomaly series data by calculating the statistical value of the converted degree of anomaly, and identify the location or range of anomalies from the anomaly series data.

[0084] 11: Object to be inspected 12: Optical fiber 13: Vibration measuring instrument 14: Anomaly location identification device 41: Pre-processing unit 42: Generation unit 43: Detection unit 44: Autoencoder

Claims

1. An anomaly identification device that uses the frequency components of vibration data at multiple points along the longitudinal direction of an optical fiber to calculate an evaluation value representing the degree of anomaly for each of the multiple points, and identifies an anomaly in the longitudinal direction of the optical fiber by comparing the evaluation values ​​at the multiple points.

2. An abnormal location identification device according to claim 1, comprising: creating input data for the multiple locations using vibration data from the multiple locations; calculating reconstructed data for the multiple locations by inputting the input data for the multiple locations into a pre-trained normal data model; calculating the reconstruction error for the multiple locations by comparing the reconstructed data and the input data for the multiple locations for each location; and calculating the evaluation value for the multiple locations using the reconstruction error for the multiple locations.

3. The abnormal location identification device according to claim 2, comprising: dividing the vibration data of the optical fiber into longitudinal sections of the optical fiber; calculating the frequency spectrum for each section using the vibration data for each section; and creating the input data by normalizing the spectral intensity values ​​in the distance direction for each section.

4. A method for identifying abnormal locations in the longitudinal direction of an optical fiber, comprising: using the frequency components of vibration data at multiple points along the longitudinal direction of the optical fiber, calculating an evaluation value representing the degree of abnormality for each of the multiple points, and comparing the evaluation values ​​at the multiple points to identify abnormal locations in the longitudinal direction of the optical fiber.

Images