Physics-Informed Neural Networks to enhance leakage detection in drinking water distribution systems

Leakages in drinking water distribution systems (DWDSs) are caused by structural failures of piping infrastructure and result in unnecessary loss of water. Prompt and accurate leakage detection is paramount for water utilities as it is both of public interest to prevent ecologic hazards and property damage as well as company interest to minimize revenue losses, insurance claims, and customer dissatisfaction from interrupted water supply.

An essential prerequisite for leakage detection is data gathered from sensors installed throughout a DWDS. With a hypothetical full coverage of flow meters, the leakage detection problem becomes trivial because leakages can be identified by a mass-balance calculation within delimited district metered areas. However, this scenario is not financially and practically viable. In most other cases, leakage detection is enabled through data gathered by a limited number of pressure sensors distributed throughout the DWDS. These pressure data are utilized to identify pressure losses from leakages caused by higher wall friction due to the augmented flowrate. Most of the methods utilising pressure data rely on a well-calibrated hydraulic model of the DWDS and some form of calibrated water demand patterns to capture the difference between the legitimate water demand, due to water usage in normal conditions, and additional flows due to leakages. Water demand calibration, however, becomes especially challenging if irregular or non-periodic water demands that do not follow the usual diurnal patterns and, hence, cannot be extrapolated into the future, are present. This type of demand may describe, for instance, certain industrial water usages.

In an earlier work developed as part of the Battle of the Leakage Detection and Isolation Methods (BattLeDIM), an international competition on leakage detection and localization, we introduced LILA, a purely data-driven approach to leakage detection and localization based on pressure data, without the need for a calibrated hydraulic model or physical parameters or water demand. LILA employs a linear regression model of the pressure losses between different sensor locations to establish a baseline and raises an alarm if deviations from that baseline are detected. However, in the presence of irregular demands, if unknown, the establishment of a linear baseline is only possible to a very limited extent, resulting in high fault tolerances and extremely long detection times. On the other hand, we demonstrated that known irregular demands may be incorporated into the linear regression model as an additional regressor.

In this work, we present an approach to predict unknown industrial water demands in an implicit fashion employing a physics-informed neural network (PINN), thus, enhancing the detection capability of LILA. The PINN incorporates the physics of the DWDS in the form of a loss function that reflects a modification of the Bernoulli principle. The input to the model is the pressure data, while the output is directly fed to the linear leakage detection model, training the PINN in an implicit manner. Preliminary results show that the time to detection of an abrupt leak can be reduced by up to a factor of 20 using PINN in comparison to the original LILA.