According to the Navy Facilities Asset Data Store (iNFADS), over two-thirds of the Navy’s water lines are either approaching or have exceeded their expected useful life – with 67% of water lines over 40 years old, and 37% of water lines over the age of 65. iNFADS also indicates that 57% of the current water line inventory is either in poor or failing condition, which can lead to a high likelihood of water main breaks and non-surfacing leaks. The main objective of the project was to demonstrate a novel artificial intelligence (AI) leak identification technology for online leak monitoring in water distribution systems. The demonstration tests and evaluates the performance of the AI leak detection technology by assessing its ability to detect and monitor both breaks and non-surfacing leaks.

The core of the demonstrated technology is based on sensing and analysis of the acoustic waves generated by leaks leaving the pressurized pipe boundary using AI, probabilistic methods, and visualization tools hosted on a server. When a leak is present, the noise generated by the leak is transmitted through various media including the pipe wall, surrounding soil and water column. The overall distance that leak noise can travel from the source of the leak is ultimately material dependent. Crucially, the demonstrated technology listens for the acoustic leak noise directly within the water column, to maximize transmissibility while simultaneously minimizing variability due to different pipe materials and soil conditions. The sensing device was designed to be retrofittable to existing wet barrel hydrants typical in California. The demonstration was conducted in two phases. The first phase considered an offline, asynchronous implementation of the technology, where sensor data from each device was manually collected and uploaded to the server for processing. The second phase demonstrated a fully online implementation in which sensor data could be automatically transmitted to the secured server in real-time via cellular connectivity. In both phases, the data was analyzed using AI-based algorithms and probabilistic methods for robust leak detection and localization. Model outputs were made available through a bespoke dashboard and user interface.

First, leak detection models were successfully trained using the leak features extracted from the baseline data collected over the demonstration period using simulated leaks. In Phase I, the AI models were able to classify unseen leaks with an overall accuracy of approximately 94%, encompassing leaks of various flow rates and distances from the instrumented hydrants. In Phase II, time-frequency analysis was used to gain an additional understanding of the acoustic data. These insights were used to introduce improvements to the data processing pipeline, resulting in a significant improvement in accuracy beyond 94% in Phase I to detecting all leaks in Phase II. The location of the leak was successfully estimated using the maximum likelihood estimation method. Results showed that it is possible to locate the leaks to the correct pipe segment level, however, exact pinpointing was challenging due to uncertainty with respect to the pipe layout and pipe characteristics. Device clock synchronization between the individual devices also likely contributed to the overall localization errors. In addition to leak identification, the technology also showcased the ability to detect and capture short-lived, pressure transient events in the network and notify operators in a timely manner.

The deployment of the continuous monitoring solution for water distribution networks provides several key, practical benefits. First, the ability to detect leaks in challenging materials such as polyvinyl chloride (PVC) is critical given the fast-growing representation of PVC pipe throughout the Department of War. Additionally, the ability to continuously monitor leaks and provide early warning of leaks will allow operators to respond in a timely manner and prioritize maintenance actions as needed. Lastly, real-time pressure transient capture also helps to monitor vulnerable areas of the network that may be susceptible to sudden breakage. (Project Completion - 2024)