Machine Learning Method Reconstructs Flow Across Entire Sewer Networks Using Just 3-4% Sensor Coverage

As cities grapple with increasingly intense storms and aging infrastructure, monitoring urban drainage networks in real time remains a critical challenge. Full deployment of sensors across a network is often impractical due to cost, installation complexity, and maintenance burden—yet decision-makers need reliable data to manage flow and prevent flooding.

A research team has demonstrated that a machine learning approach called data-driven sparse sensing (DSS) can reconstruct flow conditions across an entire 77-node urban drainage network using just a handful of strategically placed sensors. The method hinges on learning the underlying patterns in how flow behaves across the network, then positioning monitors at nodes that provide maximum information value.

How the Approach Works#

The researchers tested their workflow using a validated simulation model of a sewer system, applying three key steps:

• Build a spatial basis of typical flow patterns using singular value decomposition (SVD)
• Select optimal sensor locations via pivoted QR decomposition—an approach that outperformed computationally expensive alternatives
• Train a decoder to reconstruct unmeasured node conditions from sparse measurements

Results Across Diverse Conditions#

Testing a 3-node sensor layout (representing just 4% of the network) across 225 different scenarios—combining 25 different model calibrations with 9 rainfall events—revealed strong performance. The median Nash-Sutcliffe efficiency (NSE), a standard metric for comparing predictions to observations, reached 0.791, with even the 10th percentile exceeding 0.700 across all tested scenarios.

Importantly, the pivoted QR-selected sensor placements matched the performance of reference methods (Greedy D-optimal and genetic algorithms) without requiring iterative computational searches, offering a faster alternative.

Robustness and Failure Modes#

The researchers stress-tested the framework by adding measurement noise and simulating sensor failures. The method proved insensitive to realistic noise levels but showed sensitivity to which sensors were lost—some node positions proved more critical than others. Nodes with larger upstream drainage areas, bigger adjacent pipe diameters, and higher active flow fractions contributed most to overall reconstruction accuracy, a relationship they quantified with strong statistical correlation (0.781).