Metrics, placement validation, and provenance for a defensible demonstration.
IMERG_2024_08_11
Network at rest. Four proposed sensor locations are online.
Top-1 localisation
78.9%
Held-out rainfall groups
Top-3 localisation
100.0%
Candidate ranking
Macro F1
0.729
Grouped test split
Severity MAE
0.139
Optional regressor
The persisted summary compares equal-budget placements on the same hydraulic state matrix.
Lower is better. The scale is logarithmic because random placements span a wide error range.
How well four readings describe the rest of the modelled state.
Physical readings
4
Proposed sensor locations
Estimated locations
34
Reduced-order reconstruction
RMSE
—
Current analysed scenario
NSE
—
Current analysed scenario
RMSE and NSE describe one scenario at a time. Run analysis on the command centre to fill them in.
The classifier is evaluated by rainfall event, so test weather windows are absent from training.
Held-out events
IMERG_2023_07_09, IMERG_2024_08_11, IMERG_2024_09_12
Every layer carries an explicit category. Controlled faults are not field incident records.
What the categories mean
Drain geometry and rainfall are observed source inputs. Hydraulic response is physics-derived from EPA SWMM. Sensor placement, reconstruction, and fault probabilities are computed outputs. Controlled obstruction scenarios are simulations, not observed municipal blockage records.
What this prototype does not claim.
The SWMM network is not fully calibrated against municipal field sensors. Missing elevations, catchments, and some engineering dimensions are tagged prototype assumptions. Restriction is modelled as effective cross-section reduction, not solids or debris transport. The classifier covers ten sensitivity-ranked prototype conduits and is not trained on observed blockage incidents.
OSDosed · Smart India Hackathon 2026
GMDA · IMERG · SWMM · SVD/QR