JP7604412B2 - Method and system for identifying leakage phenomena using machine learning - Google Patents

Description

The present invention relates to a method and system for identifying leakage phenomena that can measure AE waves using a single sensor and identify fluid leakage phenomena in pressurized pipelines using machine learning.

In recent years, many water leakage accidents due to aging have occurred in water transmission and distribution pipelines that use internal water pressure, such as those used for drinking water and industrial water, and methods to identify leakage phenomena are being developed. Specifically, many methods have been proposed to identify the presence or absence of leakage and its location by installing sensors at multiple points and estimating the location of the leakage from differences in the measured leakage waveforms. For example, Patent Document 1 discloses a method of estimating the location of a leakage by calculating the ratio of high-frequency components to low-frequency components.

Patent document 2 also discloses an invention for a water leak location estimation system that estimates the leak path and location by installing a single sensor at one location.

JP 2019-039858 A JP 2021-183936 A

However, the method disclosed in Patent Document 1 requires sensors installed at two locations to detect leak waves with the same degree of accuracy, and has the problem of a greater workload than when using a single sensor. Furthermore, even with the system disclosed in Patent Document 2, a method for identifying the scale of leaks has not been confirmed, and useful waveform parameters for identifying the presence or absence of a leak and its location have not been fully clarified.

In other words, in conventional technology, there is still no established method for accurately identifying the presence or absence of a leak, its location, and its size using a single sensor.

The present invention aims to provide a method and system for identifying leakage phenomena that can measure AE waves using a single sensor and identify fluid leakage phenomena in pressurized pipelines, including water transmission and distribution pipelines, using machine learning.

The present invention is a method for identifying leakage phenomena by machine learning, capable of identifying a fluid leakage phenomenon in a pressurized pipeline by AE waves measured by a single leakage sensor, and is characterized by having at least a measurement step of measuring AE waves by the single leakage sensor installed in the pressurized pipeline, a learning step of machine learning the measured AE waves, an extraction step of extracting AE parameters in the learning step, and a leakage phenomenon identification step of identifying a fluid leakage phenomenon in the pressurized pipeline based on the extracted AE parameters.

The configuration of the present invention makes it possible to identify each leakage phenomenon with high accuracy using a single leakage sensor by identifying AE parameters that are effective in identifying the presence or absence of leakage, the location of the leakage, and the size of the leakage, which have not been clearly identified until now.

4 is a flow chart showing a procedure for identifying a water leakage phenomenon in an embodiment of the present invention. FIG. 2 is a schematic plan view showing the arrangement of water leakage sensors in a model water distribution pipeline in an embodiment of the present invention. 1A and 1B are examples of waveforms of detected continuous AE waves, where (a) is a waveform when there is a leak in the model water distribution pipeline, and (b) is a waveform when there is no leak in the model water distribution pipeline. 2 shows waveforms of AE waves for explaining AE parameters that are explanatory variables in an embodiment of the present invention. 11 is a table showing an F value, which indicates the accuracy of a model, for the determination result of the presence or absence of water leakage in an embodiment of the present invention. This is a table showing the top three AE parameters with high importance in the decision tree and random forest that had the highest accuracy in determining whether or not there was a leak. 11 is a table showing an F value, which indicates the accuracy of a model, for the results of identifying a water leakage position in an embodiment of the present invention. This is a table showing the top three important AE parameters in the decision tree and random forest that had the highest accuracy in identifying the leak location. This is a table showing the AE parameters with the highest importance in the decision tree and random forest that had high accuracy in identifying the leak location. 13 is a table showing an F value, which indicates the accuracy of a model, for the results of identification of the scale of leakage in an embodiment of the present invention. This is a table showing the top three important AE parameters in the decision tree and random forest that had high accuracy in identifying the scale of leakage. 13 is a table showing the AE parameters with the highest importance in the decision tree and random forest that had high accuracy in identifying the scale of leakage. 13A is a graph showing the importance of explanatory variables in a decision tree obtained by verification on an actual structure, and FIG. 13B is a graph showing the importance of explanatory variables in a random forest. This is a photograph of the leakage state at a leak point in a water transmission pipeline, which is an actual structure. 1 is a schematic block diagram of a water leakage phenomenon identification system according to an embodiment of the present invention.

Below, with reference to the drawings, an embodiment of the leakage phenomenon identification method and identification system using machine learning according to the present invention will be described, taking as an example a method for identifying a leakage phenomenon in a water transmission/distribution pipeline.

The method and system for identifying leakage phenomena in water transmission/distribution pipelines of this embodiment uses machine learning to learn acoustic emission waves (hereinafter simply referred to as "AE waves") in water transmission/distribution pipelines, making it possible to identify the presence or absence of leakage in a water transmission/distribution pipeline, as well as the location and magnitude of the leakage, using a single leakage sensor 1.

In other words, by identifying AE parameters that are effective in identifying the presence or absence of a leak, the location of the leak, and the scale of the leak, which have not been clearly defined until now, it is possible to identify each leak phenomenon with high accuracy using a single leak sensor 1.

(Leakage phenomenon identification flow)
The flow of identifying a water leakage phenomenon in this embodiment is shown in Fig. 1. To briefly explain the flow, in step S100, AE waves are measured in a water transmission/distribution pipeline using a water leakage sensor 1 that is composed of a known elastic wave measuring device and a resonance type sensor.

In the next step S101, it is determined whether AE waves have been detected, and if no AE waves have been detected, the measurement conditions are reset in step S102. Resetting the measurement conditions involves adjusting the frequency band, preamplifier, and amplitude detection threshold shown in FIG. 4. If it is determined in step S101 that AE waves have been detected, a class label is assigned in step S103.

Next, in step S104, the machine learning model of this embodiment determines whether or not there is a leak. If it is determined that there is a leak, AE parameters that are effective for that determination are extracted in step S105.

Next, in step S106, the machine learning model of this embodiment determines whether the leak location has been identified. If it is determined that the leak location has been identified, AE parameters that are effective for this determination are extracted in step S107. If it is determined that the leak location cannot be identified, the measurement conditions are reset in step S102, as described above.

Next, in step S108, the machine learning model of this embodiment determines whether or not the scale of the leak has been identified. If it is determined that the scale of the leak has been identified, AE parameters that are effective for this determination are extracted in step S109. If it is determined that the scale of the leak cannot be identified, the measurement conditions are reset in step S102, as described above.

(Survey and verification on a model water distribution pipeline)
In this embodiment, in order to identify various water leakage phenomena and to verify effective AE parameters, a model water distribution pipeline P as shown in FIG. 2 was installed and verification was performed. A stand 2.9 m above ground was installed on the upstream side of the model water distribution pipeline P, and a water tank (not shown) was installed on the stand. In addition, a terminal valve (not shown) was installed on the downstream side, and it was configured to be able to switch on and off the discharge. That is, the model water distribution pipeline P is assumed to be a relatively low-pressure water distribution pipeline branching off from a high-pressure water supply pipeline. In addition, the arrangement relationship between each water leakage sensor 1 and the water leakage hole H is as shown in FIG. 2. When measuring the AE wave in the model water distribution pipeline P, the AE wave was measured for the model water distribution pipeline P with the water leakage hole H and the undamaged model water distribution pipeline P without the water leakage hole H. For reference, the model water distribution pipeline P in this embodiment uses a VP pipe with an inner diameter of 100.8 mm.

The measurement cases are as shown in Table 1 below. Case A is when the end valve (not shown) of the model water distribution pipeline P is blocked and water is only drained from the leak hole H. Case B is when the end valve was opened to 11.3° and water was drained from both the leak hole H and the end valve. Case C is when the end valve was opened to 11.3° in an undamaged model pipeline P without a leak hole H and water was drained only from the end valve.

Furthermore, each water leakage sensor 1 is positioned relative to the water leakage hole H as shown in FIG. 2, with 150 kHz resonant type sensors (R15α) attached to "ch1", "ch4", "ch5", and "ch8", 60 kHz resonant type sensors (R6α) attached to "ch2" and "ch6", and 30 kHz resonant type sensors (R3α) attached to "ch3" and "ch7".

Each of the water leakage sensors 1 is attached to the bottom of the pipe of the model water distribution pipeline P using a high vacuum sealing oil compound, and is further fixed to the pipe by wrapping a band around it.

As for the measurement conditions, the preamplifier amplification value is 40 db, and the detection threshold, described below, is 30 db. The measurement interval for AE waves is 1 MHz. To confirm the effects of water pressure, pressure sensors are placed at the upstream and downstream parts of the model water distribution pipeline P, and the water pressures are measured at each part as a synchronization signal in addition to the measurement of AE waves. The measurement conditions for AE waves are as shown in Table 2.

By applying machine learning to the AE waves detected by the water leakage sensor 1 in the model water distribution pipeline P, it is possible to identify each water leakage phenomenon and extract effective AE parameters. For reference, FIG. 3 shows an example of the waveform of a continuous detected AE wave, where (a) shows the waveform when there is a water leakage hole H in the model water distribution pipeline P, and (b) shows the waveform when there is no water leakage in the model water distribution pipeline P.

(Analysis methods in machine learning)
Next, a method of analysis will be described regarding identification of the presence or absence of a water leak and extraction of effective AE parameters by machine learning. Fig. 4 shows the waveform of an AE wave for explaining the AE parameters, which are explanatory variables.

As shown in the figure, in this embodiment, the AE parameters extracted from the measured AE waves are "AMP", which is the maximum amplitude value, "RISE", which is the rise time, "DURATION", which is the duration, and "COUNT", which is the count value of the intersection with a predetermined detection threshold, as well as "A-FRQ", which is the average frequency, "C-FRQ", which is the centroid frequency, "RMS", which is the effective voltage, and "P-FRQ", which is the peak frequency.

Next, the class labels set for "determining the presence or absence of a leak," "identifying the location of the leak," and "identifying the scale of the leak (leak hole diameter)" are shown in Tables 3 to 5.

The measurement data obtained by measuring AE waves was divided into training data and evaluation data to construct an algorithm and verify the evaluation accuracy of the algorithm. For "determining the presence or absence of a leak," the training data was 70, and the evaluation data was 50. For "identifying the location of the leak" and "identifying the scale of the leak (leak hole diameter)," there were two cases of training data, 40 and 200, and the evaluation data was 200, and the effect of data size on the evaluation accuracy was examined.

When constructing the algorithm, we use "decision trees," "random forests," and "support vector machines (hereafter referred to as "linear SVMs")," which are types of supervised learning.

A machine learning model was constructed in the manner described above, and the accuracy of identifying various water leakage phenomena using this machine learning model and the identification of effective AE parameters are described below.

(1) Determining whether there is a leak
Fig. 5 shows the F value, which indicates the accuracy of the model, for the results of determining the presence or absence of leakage using machine learning, based on measurement data measured in a model water distribution pipeline P with no leakage when discharging (valve opening of the terminal valve: 11.3°) and a model water distribution pipeline P with a leakage hole H. As shown in the figure, when comparing the accuracy of each algorithm, it can be seen that the decision tree and random forest provide higher accuracy than the linear SVM.

When comparing the accuracy for each diameter of the leak hole H, it can be seen that the larger the diameter of the leak hole H, the higher the discrimination accuracy (dashed line A in the figure), and when the diameter of the leak hole H is 3 mm or more, the F value is 0.9 or more, and it can be seen that the presence or absence of a leak can be accurately determined. Furthermore, regardless of the resonant frequency of the leak sensor 1, the discrimination accuracy of "ch5", "ch6", and "ch7", which are close to the leak hole H, tends to be higher than that of other leak sensors 1 (solid line B in the figure).

Figure 6 also shows the top three AE parameters with high importance in discriminating between decision trees and random forests, which had high discrimination accuracy. As shown in the figure, it is clear that the parameters with high importance common to both algorithms are "RMS", which is the effective voltage, "P-FRQ", which is the peak frequency, and "C-FRQ", which is the center of gravity frequency.

Since the majority of the AE parameters with the highest importance in the decision tree are "RMS," which is the effective voltage, it can be seen that a certain level of discrimination accuracy can be obtained with just one AE parameter.

In addition, the important AE parameters differ depending on the diameter of the leak hole H. When the diameter of the leak hole H is 1 mm, the peak frequency "P-FRQ" is more important, and when the diameter of the leak hole H is 5 mm, the effective voltage "RMS" is more important.

From the above, it became clear that the presence or absence of water leakage can be determined with high accuracy by using a decision tree or random forest. In addition, it became clear that the AE parameters that are effective for determination are mainly "RMS", which is the effective voltage, and "P-FRQ", which is the peak frequency.

(2) Identification of leak location We examined whether it was possible to identify the leak location in two conditions: "Case B" with discharge (valve opening of terminal valve: 11.3°) and "Case A" without discharge. Figure 7 shows the F value, which is the model accuracy, for the results of leak location identification by machine learning for each of the above conditions. Comparing the identification accuracy for each algorithm, it can be seen that decision trees and random forests provide higher accuracy than linear SVM, just as in determining the presence or absence of leaks.

In addition, when comparing identification accuracy by data size, although there is a tendency for data size 200 to provide higher identification accuracy, it is clear that by using a 30 kHz resonant sensor (R3α), high identification accuracy with an F value of 0.9 or more can be obtained regardless of data size (solid line A in the figure). Furthermore, it was confirmed that there is a tendency for identification accuracy to be high in the "no discharge" case, and in addition, it was shown that when the diameter of the water leakage hole H is 1 mm, the identification accuracy decreases with the 60 kHz and 150 kHz resonant sensors.

Figure 8 shows the top three important AE parameters in the decision tree and random forest that showed high identification accuracy. As shown in the figure, the important parameters common to both algorithms are "RMS" which is the effective voltage, "P-FRQ" which is the peak frequency, "C-FRQ" which is the centroid frequency, and "COUNT" which is the count value of the intersection with a predetermined threshold.

The important AE parameters differ depending on whether or not there is discharge. As shown in Figure 9, in the case of "no discharge," the important AE parameter is "RMS," which is the effective voltage, and in the case of "discharge," the important parameter is "P-FRQ," which is the peak frequency.

The above results demonstrate that the use of a 30 kHz resonant sensor (R3α) makes it possible to identify the location of a leak with high accuracy. In addition, it was shown that "RMS", "P-FRQ", "C-FRQ", and "COUNT" are effective AE parameters.

(3) Identification of the Leakage Scale We examined whether the diameter of the leakage hole H could be identified in two conditions: "Case B" with discharge (valve opening of the terminal valve: 11.3°) and "Case A" without discharge. When comparing the identification accuracy of each algorithm, as shown in Figure 10, the decision tree and random forest achieved higher identification accuracy than the linear SVM, as in the above-mentioned cases (1) and (2).

Furthermore, when comparing the identification accuracy by data size, a data size of 200 provided higher identification accuracy. When comparing the identification accuracy by the presence or absence of discharge, it was shown that "no discharge" provided higher accuracy. In particular, when the data size was 200 and "no discharge" was provided, the identification accuracy of the decision tree and random forest was extremely high, with an F-value of 0.94 or higher for all resonant sensors. As shown by the solid line A in Figure 10, the identification accuracy of "ch5", "ch6", and "ch7" was higher than that of the other resonant sensors, so even when "discharge" is provided, the size of leak hole H can be identified if the resonant sensor is close to leak hole H.

Figure 11 shows the top three important AE parameters in the decision tree and random forest that showed high identification accuracy. The important parameters common to both algorithms are "RMS", which is the effective voltage, "P-FRQ", which is the peak frequency, and "COUNT", which is the count value of the intersection with a specified threshold.

As shown in Figure 12, the AE parameter "RMS" was shown to be an important indicator regardless of whether or not release was performed, while the importance of "COUNT" was also high when "release was performed" and "P-FRQ" was also high when "no release was performed."

From the above, it became clear that the scale of the leak could be identified with high accuracy, and that the AE parameters that were effective in this case were "RMS," "P-FRQ," and "COUNT."

(Verification on actual structures)
The machine learning model described above was used to verify the accuracy of identifying water leakage phenomena in a cast iron water pipeline (φ250 mm) that is actually buried and in common use. In the water pipeline to be verified, a water leakage hole H with a diameter of 3 mm was confirmed at a position 3.0 m from the measurement point of the water leakage sensor 1. In addition, the internal water pressure of the water pipeline was measured under water-stop conditions from a very high pressure state of 0.8 MPa (maximum value: water head 78 m) to 0.01 MPa (water head 1 m) during the measurement period. Note that FIG. 14 shows a photograph of a situation in which water is leaking vigorously from a water leakage hole H of a water pipeline, which is an actual structure, and the AE waves associated with such a jet flow will be measured.

Water inside the water transmission pipeline was drained only from leak hole H, and it took 33 hours for the water to be drained from a full-water state to leak hole H. AE wave measurements were taken at a pumping station near leak hole H and near leak hole H, at 15-minute intervals. Measurements were taken for 30 seconds each, and the measurement conditions were a detection threshold of 20 db and an amplification value by the preamplifier of 60 db. As AE wave measurements were taken at the pumping station, the internal water pressure of the water transmission pipeline was evaluated simultaneously with the AE wave measurements using a water pressure gauge installed in the nearby pump equipment. Note that in this verification, a wideband sensor was used as the water leak sensor 1.

Next, the analysis method used in the verification will be explained. Eight AE parameters were extracted as explanatory variables: "AMP", which is the maximum amplitude value; "RISE", which is the rise time; "DURATION", which is the duration; and "COUNT", which is the count value of the intersection with a predetermined detection threshold. In addition, the following parameters were extracted: "A-FRQ", which is the average frequency; "C-FRQ", which is the centroid frequency; "RMS", which is the effective voltage; and "P-FRQ", which is the peak frequency.

Furthermore, to examine the differences in AE waves due to the water head, the class labels were set as shown in Table 6 below, and the acquired AE wave measurement data was divided into training data and evaluation data to construct an algorithm and verify the evaluation accuracy of the algorithm. The data size was set to 75 for training data and 58 for evaluation data, and similar to the model pipeline P described above, linear SVM, decision tree, and random forest were used to construct algorithms and compare the evaluation accuracy.

Table 7 shows the verification results for a water supply pipeline of an actual structure. Regarding the evaluation accuracy for each algorithm, a high evaluation accuracy of 0.98 or more was obtained for all indices for decision trees and random forests. From this, it can be seen that the method for identifying leakage phenomena in water supply and distribution pipelines using machine learning according to this embodiment is effective not only for low-pressure water supply pipelines such as the model water supply pipeline P described above, but also for high-pressure water supply pipelines. It can also be seen that leakage phenomena can be identified with high accuracy even in actual structures that are significantly affected by noise.

Figure 13(a) shows the importance of explanatory variables in the decision tree obtained in this verification, and Figure 13(b) shows the importance of explanatory variables in the random forest. As shown in the figure, it can be seen that "COUNT" is the most important AE parameter in both algorithms. This means that by using "COUNT" as an index, it is possible to identify leakage phenomena for each head difference with high accuracy.

(System Configuration)
The method for identifying a water leakage phenomenon by machine learning according to this embodiment has been described above, but it is needless to say that this series of processes can be systematized. For example, Fig. 15 shows a schematic block diagram of a water leakage phenomenon identification system 100 according to this embodiment. The water leakage phenomenon identification system 100 has a single water leakage sensor 1 and a preamplifier 2 that are installed in a water transmission/distribution pipeline, and the measurement data measured by the water leakage sensor 1 is input to a PC or application server that serves as an analysis means via a communication means (wired or wireless).

The PC or application server includes at least a measurement data analysis unit 101 that analyzes the measurement data measured by the water leakage sensor 1, a measurement condition setting unit 102 that sets the measurement conditions as shown in Table 2 above, a parameter extraction unit 103 that extracts AE parameters for identifying the water leakage phenomenon, a machine learning unit 104 that constructs a predetermined algorithm by supervised learning, a water leakage information output unit 106 that outputs identification information such as the presence or absence of a water leakage, the location of the leakage, and the scale of the leakage, and a measurement facility information input unit 105 that inputs facility information such as the pipe diameter and water pressure of the water distribution pipeline to be measured.

By executing the above-mentioned method for identifying a water leakage phenomenon using the device configuration or program configuration described above, it becomes possible to identify the water leakage phenomenon with high accuracy.

In addition, in an actual structure currently in use that has been investigated using the water leakage phenomenon identification system 100, information such as the position of the actually confirmed water leakage hole H (distance from the water leakage sensor 1), the diameter of the water leakage hole H, and the water pressure is input to the machine learning unit 104 as supervised data each time, which makes it possible to gradually improve the accuracy of identifying water leakage phenomena.

Other Embodiments
The above describes an embodiment of the leakage phenomenon identification method and identification system using machine learning of the present invention, taking as an example a method for identifying a leakage phenomenon in a water transmission/distribution pipeline. However, the present invention is not necessarily limited to the configuration described above, and various modifications such as those described below are possible.

For example, the leakage phenomenon identification method and identification system of the present invention using machine learning is not necessarily limited to water supply and distribution pipelines, but can be applied to pressurized pipelines that transport various fluids, such as fuel pipes, gas pipes, and steam pipes. By taking into account the characteristics of the fluid and installing a single leakage sensor in the pressurized pipeline, it is possible to identify the leakage phenomenon of the fluid (presence or absence of leakage, leakage location, leakage scale (diameter of the leakage hole), etc.) with high accuracy.

Although the embodiments of the present invention have been described above with reference to the drawings, the specific configuration is not limited to these embodiments. The scope of the present invention is indicated by the claims, not by the description of the above embodiments, and further includes all modifications that are equivalent in meaning to and within the scope of the claims. In addition, the specific materials, dimensions, shapes, etc. described in the above examples may be modified within the scope of the present invention.

1. Water leakage sensor (leakage sensor)
2 Preamplifier 100 Leakage phenomenon identification system 101 Measurement data analysis unit 102 Measurement condition setting unit 103 Parameter extraction unit 104 Machine learning unit 105 Measurement facility information input unit 106 Leakage information output unit

Claims (4)

A method for identifying leakage phenomena by machine learning, capable of identifying a fluid leakage phenomenon in a pressurized pipeline by an AE wave measured by a single leakage sensor, comprising:
a measuring step of measuring an AE wave by the single leak sensor installed in the pressurized pipeline;
A learning step of machine learning the measured AE waves;
A leakage phenomenon identification step of identifying the presence or absence of a leakage phenomenon, a leakage position, and a leakage scale (diameter of a leakage hole) by a machine learning model constructed based on the learning step,
In the leakage phenomenon identifying step,
Extraction of at least one of RMS and P-FRQ, which are AE parameters effective for determining the presence or absence of the leakage phenomenon;
Extracting at least one of RMS, P-FRQ, and C-FRQ, which are AE parameters effective for determining the leak location;
Identifying a fluid leakage phenomenon in the pressurized pipeline based on extraction of at least one of RMS, P-FRQ, and COUNT, which are AE parameters effective for determining the leakage magnitude.
A method for identifying leakage phenomena using machine learning, comprising: The method for identifying a leakage phenomenon by machine learning according to claim 1 , wherein in the learning step, an algorithm is a decision tree or a random forest . The method for identifying leakage phenomena using machine learning according to claim 1 or 2 , wherein the pressurized pipeline is a water transmission/distribution pipeline, and the fluid is water . A machine learning leak phenomenon identification system capable of identifying a fluid leak phenomenon in a pressurized pipeline, comprising:
a single leak sensor installed in the pressurized pipeline;
a measurement data analysis unit that analyzes measurement data measured by the single leak sensor;
a measurement condition setting unit that sets measurement conditions for the single leak sensor;
A parameter extraction unit extracts AE parameters for identifying a leakage phenomenon in the pressurized pipeline;
A machine learning unit that constructs a predetermined algorithm through supervised learning;
A leakage information output unit that outputs identification information of a leakage phenomenon in the pressurized pipeline;
The measuring facility information input unit inputs facility information of the pressurized pipeline to be measured.
The parameter extraction unit
Extraction of at least one of RMS and P-FRQ, which are AE parameters effective for determining the presence or absence of a leakage phenomenon;
Extraction of at least one of RMS, P-FRQ, and C-FRQ, which are AE parameters effective for determining a leak location;
Extraction of at least one of RMS, P-FRQ, and COUNT, which are AE parameters effective for determining the leak size.
A leakage phenomenon identification system using machine learning.

Images