Growing populations, climate change, and deteriorating water supply infrastructure are exerting unsustainable demands on urban water resources worldwide. A significant portion of water is lost between treatment and delivery, which can largely be attributed to leakage and bursts in distribution systems. Detection of large bursts is relatively straightforward, while detection of smaller bursts and leaks is more difficult. A reliable, highly sensitive leak monitoring system must have a high robustness to background noise due to uncontrolled operating, weather, and environmental changes associated with real world water supply systems. And, importantly, these systems should also have the ability to operate autonomously or semi-autonomously for long periods of time without day-to-day supervision or tuning.
Leak detection approaches can be broadly classified as model-based or data-driven. Model-based techniques are limited when dealing with complex large-scale systems, hence data-driven approaches have become appealing alternatives. This study focused on long-term monitoring and detection systems where hydraulic data (e.g. pressure, flow, acceleration) are continuously or semi-continuously monitored in a data-driven framework. While it is clear from the literature that events associated with large changes in the hydraulic conditions can be detected using monitored pressure data, what remains unclear is whether it is possible to discern background leaks embedded in high background noise. In such cases, hydroacoustic signatures are more suitable to detect the frequency changes generated from the leaks within the noisy environment. This paper explores the application of singular spectrum analysis (SSA) in extracting leak components from noisy measurements.
Methodology
An experimental test bed was constructed to simulate a small portion of a typical full-scale water distribution system in North America (Figure 1). The test bed consisted of a series of grey scale 80 PVC pipes with 15.24 cm inner diameter, two tees, a fire hydrant, and one service connection valve. The total length of the pipe system was approximately 15 m. Leaks were simulated by opening valves at four locations, each of which resulted in approximately 17 L/min flows when fully opened. The system was pressurized by a direct connection to the water main flow from the City of Waterloo, Ontario, Canada to an average of 345 kPa. Two hydrophones located in the network were used to measure the acoustic characteristics. The simulated leaks created different acoustic signatures at various locations due to the system configuration, thus resulting in a rich set of test cases to evaluate the proposed SSA approach. In addition to simulating leaks, larger flows simulating operational flows running concurrently and non-concurrently with the leaks were used to develop and validate methods that are sensitive to background leak signatures buried in operation-induced acoustic noise.
Figure 1: Laboratory setup and hydrant cut-out
Outcomes
For data processing, the experimental hydro acoustic signals were partitioned into consecutive non- overlapping segments. Features such as entropy, effective value and spectral peak were computed for each segment and the Bhattacharya distance was computed between the histograms of leak and non-leak feature values in order to quantify the histograms’ similarity. SSA, a powerful signal processing technique, was then applied where a group of elementary components was assumed to be carrying the signatures related to the leak, while the remaining components formed the background variation insensitive to the leak.
The capability of descriptive features to discriminate between leak and non-leak data was enhanced through the application of machine learning to the data processed using SSA. To automate this process, a powerful classifier, the one-class support vector machine (OCSVM), was applied to features computed from leak-free SSA-processed data. Under the assumption that the set of elementary components can be divided into two disjoint subsets of sensitive and insensitive components, an OCSVM can be trained for each elementary component and separate models can be obtained. Over time, poor models will be recognized by their randomly varying predictions, while efficient models based on the highly sensitive components are likely to make consistent predictions. Training and continuously tuning models can be costly. Model inefficiency can be reduced, however, using SSA pre-processing, or the separation of elementary signals and their noise components.
Receiver operating characteristic (ROC) graphs were developed to show the probability of models correctly predicting leaks against the probability of false alarms. The area under ROC curves (AUC) is an indicator for the accuracy of leak detection (Figure 2). When the valve is closed and the network is relatively quiet, the largest SSA component results in an AUC of 0.90 (dashed line in Figure 2(a)), meaning a high detection accuracy, while the remaining components led to AUC values around 0.5, which mean they perform almost random predictions. The models are also tested against leaks that occur when the valve was open (Figure 2(b)). It showed that many components are expected to perform poorly (AUC < 0.6) while a few components (2, 9, 21) resulted in good predictions. Of note were the components 3, 4, 5, 6, 7, 8 (marked lines in 2(b)) which led to a high AUC of > 0.8, while 4, 5, 6 even have AUC > 0.85. Once a leak occurs, those latter components (4, 5, 6) gave deterministic and constant alarms leading to a robust decision making process.
Figure 2 shows that the proposed algorithm based on SSA pre-processing of raw measurements has many advantages. The approach is fully data-driven and only requires the availability of sensor data. It has the potential to detect small leaks, whose acoustic signatures are hidden in the background noise, primarily due to the decomposition of the raw signals into elementary components. Ultimately, the approach is flexible in that it can be applied in both supervised as well as semi-supervised leak detection settings.
Figure 2: Evaluation of leak detection accuracy using the AUC of an OCSVM model: (a) valve closed, (b) valve open.
Conclusions
The results from this study show significant promise in detecting leaks coupled to strong background noise. The selection of the proper components which carry the leak signature is supported by analyzing the singular spectrum distributions of leak and leak-free data. In practice, data corresponding to all possible leak scenarios are usually unavailable, therefore a semi-supervised approach was proposed within the framework of SSA and OCSVM. Instead of pre-selecting the sensitive components, ensemble modelling is proposed where a model is trained on each SSA component. Results show that pre-processing data with SSA is able to provide sufficiently reliable leak predictions.
The authors propose a multi-sensor system which includes multiple modalities, including acceleration, acoustic pressure, temperature, and pressure which can monitor a range of hydraulic conditions, including acoustic signatures produced by background leaks. Data from these sensors can be pre-processed using a range of sophisticated signal processing techniques and fed into machine learning algorithms to detect even small leak induced hydraulic changes, specifically acoustic pressure. Results from laboratory and ongoing field studies gathered in highly noisy and non-stationary acoustic environments which are typical of uncontrolled real-world operating water distribution systems, have demonstrated that this technology holds significant promise for autonomous and semi-autonomous implementation at the field scale.
Cody, R., Harmouche, J., Narasimhan, S. (2018). Leak detection in water distribution pipes using singular spectrum analysis, Urban Water Journal, 15:7, 636- 644, DOI: 10.1080/1573062X.2018.1532016
Contact: Sriram Narasimhan, Department of Civil and Environmental Engineering
For more information about WaterResearch, contact Julie Grant.
