The Smart System Research Group of the Information Technology Research Institute (Director: Tomohiro Kudo) of the National Institute of Advanced Industrial Science and Technology (AIST; President: Ryoji Chubachi), and Nihon Water Solution Co., Ltd. (NWS; Representative Director: Jun-ichi Fukazawa) have developed a technology to narrow down the locations of suspected water leakage with high-accuracy in advance, using a sound anomaly analysis technology which learns judgments of skilled workers by machine learning.

Field tests have been conducted in two cities, likely to result in a reduction of the locations that must be inspected for water leakage by skilled workers to one-fifth, compared to the past. Narrowing down the inspection locations will lead to significant reduction in inspection costs, which will provide relief to local governments that need to reduce maintenance and management costs brought on by the decline in water rate revenue due to a decrease in population. In addition, providing this as a maintenance and management technology for social infrastructure is expected to contribute to the low-cost supply of safe drinking water in countries of Southeast Asia, now suffering from a water leakage rate of over 30 %.

Positioning of water leakage detection through learning-type sound anomaly analysis

The total laid length of water pipes in Japan is approximately 610,000 km. Of this, a large amount of the water pipes laid during the high-growth period will be reaching its statutory useful life (40 years) and will be up for replacement in the near future. On the other hand, income from water rates remains stagnant at 3 trillion yen per year, making it difficult, cost-wise, to replace all the aged pipes exceeding their useful life. However, once a water leak accident occurs, there are possibilities not only of primary damages such as the suspension of water supply, but also of secondary damages, such as traffic accidents caused by cave-in of roads, making leak prevention measures indispensable. Therefore, day-to-day maintenance, based on water leakage detection, to confirm and maintain the safety of the water pipes for long periods is required.

Furthermore, although water supply systems are being rapidly introduced in Southeast Asian countries, the water leakage rate is very high (30 % or higher) and is obstructing further economic development. A high water leakage rate not only brings about a steep increase in water rates and places a burden on the waterworks organization, but there is also the risk that contaminants may enter the drinking water, so water leakage countermeasures are very important from the standpoint of safe water supply as well.

However, currently, water leakage detection relies heavily on the work of skilled workers. There is a limit in the number of skilled workers in Japan. Future reduction in their numbers from aging is a concern. By contrast, Southeast Asia is still at the stage of developing human resources. Replacing part of the skills of the skilled worker with IT and narrowing down the locations requiring water leakage inspection by skilled workers is required.

AIST has been developing technologies to cope with the aging of social infrastructure, such as the technology to detect deterioration in roads and bridges. One of these is learning-type sound anomaly analysis technology, which has been applied to the hammering test to inspect for various defects in concrete structures non-destructively. Such machine learning technology is one of the technologies that are essential for utilizing various data and AIST has been proposing the adaptive-learning-type general-purpose recognition systems to achieve this since the 1980s. Until now, work has been conducted in anomaly detection technologies for still and moving images, and currently, targets of the work have been broadened to include acoustic data and time series sensor data.

Since 2012, AIST has contributed to solving the international water problem under the “Asian Strategy, Water Project,” and aimed to contribute to enhancement of water business competitiveness through collaboration among the government, the private sector, and research organizations. As a part of these activities, AIST has been conducting joint research for water leakage detection using a sound anomaly analysis technology together with NWS, which has been involved in the inspection of water pipe leaks for many years, amassing relevant expertise.

One of the advanced approaches to water leakage detection is technology to narrow down the water leakage locations (screening method) indicated in Fig. 1. The primary inspection is conducted in parallel with the water meter reading made every two months, by placing an acoustic leak checker in contact with the water meter to check for the presence of water leakage sounds. As the secondary inspection, a skilled worker goes to the site where water leakage is suspect and conducts a detailed water leakage inspection using a leak sound detection bar, etc. The track record of primary inspections up to this point indicates that the number of secondary inspections has been reduced to less than 10 % of all households. However, in many cases, ambient noise, etc. (pseudo leak sounds) resulted in false detections, necessitating many man-hours from skilled workers in secondary inspections.

In this study, at locations where water leakage was suspect in the primary inspection by the leak checker, leak sounds were discriminated from pseudo leak sounds using a sound anomaly analysis technology, in order to drastically reduce the targets of secondary inspection by skilled workers (Fig. 2).

Figure 1 : Screening method using water leak checker An advanced water leak detection method currently being introduced, but further reduction in man-hours is sought.

Figure 2 : Water leakage detection method targeted by this study Introduce sound anomaly analysis technology to improve the screening method

There are two major features of the sound anomaly analysis technology used in this study. The first is providing the computer with case records of judgment made by skilled workers to let the computer automatically learn the optimum rules to detect sound anomaly, rather than having the humans determine what the sound anomaly is. With conventional sound anomaly analysis technology in which humans have determined what the sound anomaly is in advance, the computer is unable to detect an unexpected strange sound as the sound anomaly. However, with the developed technology, by having the computer learn the range of sounds under normal conditions, it can detect any sound outside the range as sound anomaly. In addition, adding further cases of judgment by skilled workers will make the rules for detection of sound anomaly more accurate.

The second feature is the method to calculate the feature value of the sound when providing judgment cases to the computer. Generally, in sound anomaly analysis, a compact and useful information for judgment, known as the feature value is computed from the waveform of the target of analysis. In conventional sound anomaly analysis technologies, results of Fourier analysis (frequency analysis) were generally used as the feature value. However, this proved to be insufficient for the detection of water leakage sounds. With the developed sound anomaly analysis technology, in addition to frequency analysis, the variation in the time axis was added, to facilitate learning by the computer.

Verification of the developed sound anomaly analysis technology was conducted using inspection data for 77,789 households in two regional cities (Fig. 3). Water meters were inspected using the leak checker and 7,081 households were flagged for suspected water leaks. The secondary inspection conducted through visits by skilled workers for water leakage sounds using a leak sound detection bar revealed that, contrary to the results of the primary inspection, sounds could not be heard in 4,663 cases, while they were heard in 2,418 cases. When 198 cases were selected randomly from the 2,418 cases and used as subjects of this experiment, a total of 28 cases of real water leaks were confirmed. The developed analysis technology judged that out of the 198 cases, 160 cases were not water leaks, and only 38 cases were judged to be water leaks. In other words, of the 198 cases suspected in the primary inspection, 160 cases could be removed from the group of targets of secondary inspection, enabling a reduction of the workload of skilled workers to almost one fifth (=38/198).

Figure 3 : Detailed verification results

Figure 4 : Breakdown of judgment results using the developed technology

However, as shown in Fig. 4, of the 160 cases judged not to be water leaks, there were 5 cases that were actual water leaks. These constitute misjudgments (missed detection) but the water leakage sound in these cases was small, leading to the assumption that it was the initial stage of a leak and was at a level at which even the skilled worker may have overlooked them (minute leak location). Should the leak from these minute leak locations increase, the water leak sound will increase and the developed technology would correctly judge it to be a water leak.

On the other hand, of the 38 cases judged by this technology to be water leaks, there were 15 cases of misjudgment (over-detection). However, considering that leak judgment using a leak checker had over-detected 170 cases (= 198-28), it can be said that the number of over-detected households had been drastically reduced to 8 % (=15/170).

With the developed sound anomaly analysis technology, the number of locations inspected by skilled workers could be reduced to one-fifth compared to the conventional technology, enabling a significant reduction in inspection costs. This technology is considered to be especially effective for local governments facing population decline, where water rate revenue is declining but the government must maintain and manage the aging water pipes under a limited budget.

In the future, the sound anomaly analysis technology will be further improved to reduce misjudgments. In addition, NWS is aiming to commercialize and introduce a sensor device using the developed technology during 2015. Furthermore, there are plans to deploy this sound anomaly analysis technology to the Southeast Asian countries for utilization in water leakage detection.