The objective of this project is to demonstrate an Artificial Intelligence (AI) leak detection technology for real-time drinking water distribution system leak monitoring and leak localization at Department of Defense (DoD) installations.
This project responds to the need of providing a specific solution to accurately and cost-effectively identify leaks in water distribution systems. The technology detects sound within pressurized water lines using highly sensitive hydrophones placed at strategic locations within a piping network. Hydrophones are inserted into hydrants for monitoring, thereby eliminating the need to create new access points. In this system, changes in the acoustic field created by leaks can be obfuscated within the acoustic field variability due to operational factors, such as pumps, flows, etc., which causes the baselines to change over time, making it difficult for many conventional detection algorithms to detect. In order to overcome this, this technology employs robust data-driven machine learning, a branch of AI techniques, to detect and localize leaks and pressure transients (using pressure sensor, which is also present). The main difference between this technology and others in the market are:
The system can be configured for both wet and dry barrel hydrants. Monitoring data is acquired, stored, and transmitted using a custom-designed data-acquisition system to enable signal processing at the sensor locations to optimize data transfer rates and extend battery life. Analog sensor data is converted to digital data using a 24-bit Analog-to-Digital Conversion (ADC) board. The data processing, storage and communication modules include a micro-processor development board (Teensy 3.6), a flash memory storage module, a global positioning system chip receiver with an antenna, and a cellular modem. The ADC unit, as well as battery and data storage and communication modules are located inside a dry housing mounted on the hydrant at street level. Data storage and communication can be customized to meet specific DoD installation’s needs and cyber-security requirements.
The leak detection algorithms pre-process the noisy or corrupted sensor data with signal filtering and decomposition techniques, followed by calculating and modeling features using both non-parametric and parametric AI family of models (such as clustering). Once adequate amount of data has been acquired, anomalous events can be identified as those that cannot adequately be described by the models. This machine learning approach is ideal for long-term real-time monitoring and has significant advantages over conventional leak detection approaches which rely simply on the signal structures and correlation.
Benefits of this project are mainly water savings resulting from avoidance of pipe leaks at the early stages. Small leaks, if not detected and mitigated, can lead to larger leaks, which could cause severe property damages, traffic disruption, and cross contamination to drinking water supplies. An added benefit of this technology is the detection of water hammers that enables a more robust operation of the distribution system. Detecting water pressure transients can help identifying vulnerable areas to avoid unnecessary capital expenditures for repair or replacement of equipment due to failure caused by local pressure surges. Based on the project team’s interaction with both DoD utility personnel as well as its counterpart municipalities, all have expressed interest in this multi-sensor monitoring system due to multiple practical operational benefits in addition to its leak detection and localization capabilities.