Anomaly Detection in Sewer Systems Using Gaussian Process Regression

Authors: Mohsen Rezaee, Peter Melville-Shreeve, Hussein Rappel Year: 2025 Tags: gaussian-process-regression, anomaly-detection, sewer-systems, blockage-detection, probabilistic-forecasting, urban-drainage

TL;DR

A GPR model is trained on hydraulic simulator output to probabilistically forecast manhole water depth in a combined sewer; deviations outside a 95% credible interval sustained for ≥3 consecutive hours trigger a blockage alarm. The approach is demonstrated on a single simulated blockage scenario, providing an alarm with roughly 8 hours of lead time before manhole overflow.

First pass — the five C's

Category. Research prototype / proof-of-concept (conference short paper).

Context. Urban hydroinformatics / smart sewer monitoring. Builds on: Rezaee et al. (2025) preprint — same sewer network and emulator; Rosin et al. (2022) — ANN + statistical process control for near-real-time blockage detection; Glyn-Davies & Girolami (2022) — GP-based anomaly detection in streaming data; Williams & Rasmussen (2006) — foundational GP textbook.

Correctness. Load-bearing assumptions: (1) the hydraulic simulator faithfully represents real sensor observations so that training on simulator output transfers to field deployment; (2) blockages manifest as sustained, directional depth deviations detectable within the 95% credible interval; (3) a fixed 3-hour persistence threshold is operationally meaningful. Assumption (1) is never validated against real sensor data in this paper.

Contributions. - Frames sewer anomaly detection as a credible-interval exceedance problem under full probabilistic GPR output. - Shows directional signatures differ upstream (depth rise) and downstream (depth drop) of a blockage, enabling localization. - Establishes a 3-hour consecutive-exceedance trigger that provides lead time before overflow in the tested scenario. - Notes explicitly that gradual blockages embedded in training data constitute a systematic miss-detection failure mode.

Clarity. Writing is concise and accessible; however, quantitative model performance metrics are withheld ("model checking metrics indicate a satisfactory fit"), weakening reproducibility for a methods paper.

Second pass — content

Main thrust: GPR is fitted on one month of simulator-generated data (15-minute intervals, inputs: time and rainfall depth) to produce 95% credible intervals for manhole water depth; sustained exceedance triggers a blockage alarm roughly 8 hours before overflow occurs.

Supporting evidence: - Training set: 1 month of data, 15-minute resolution (≈2,880 points); no separate validation set size reported. - Blockage inserted between nodes J2 (upstream) and J3 (downstream); downstream depth falls below the 95% credible interval, upstream depth rises above it. - Alarm triggered after 3 consecutive hours of exceedance. - Manhole overflow occurs approximately 11 hours after blockage initiation, leaving ≈8 hours of operational lead time. - No precision, recall, F1, or RMSE values are reported anywhere in the paper.

Figures & tables: Two figures (Figures 1 and 2) show time-series plots of predicted mean (red line) and observed values (blue bullets) at downstream and upstream manholes respectively. The 95% credible interval appears as shading (described as "shaded red"). Axis units and labels are referenced in captions but not described in detail in the text; statistical significance is not reported; no error bars on predictions beyond the credible interval itself; no tables of metrics.

Follow-up references: - Rezaee et al. (2025) preprint — the extended version of this work with full methodology and likely more results. - Rosin et al. (2022) — nearest competing method (ANN + SPC for blockage detection) that should serve as a direct baseline. - Glyn-Davies & Girolami (2022) — GP-based anomaly detection framework directly relevant to the probabilistic approach used here. - Williams & Rasmussen (2006) — essential background for GPR kernel selection and inference.

Third pass — critique

Implicit assumptions: - Simulator-to-reality transfer: the GPR is trained entirely on PySWMM/SWMM hydraulic model output, not real sensor data. If the simulator does not capture real sensor noise, drift, or diurnal patterns accurately, the credible intervals may be miscalibrated in deployment — this assumption is untested and would break the anomaly detection logic. - Stationarity of the GP prior: the kernel choice and its hyperparameters are not described; non-stationarity in real sewer flow (storm events, seasonal infiltration) could inflate false positive rates. - The 3-hour threshold is treated as a tunable parameter but no sensitivity analysis is provided; its optimality is asserted, not demonstrated.

Missing context or citations: - No comparison to the Rosin et al. (2022) ANN + SPC method, which is cited but not benchmarked against. - No engagement with threshold-free or scoring-based anomaly detection approaches (e.g., isolation forests, autoencoders) that are common in this domain. - Real-world sensor noise, missing data, and data drift are acknowledged in passing but no citation or method addresses them. - Only a single blockage type and location tested; pump failure and structural collapse (mentioned in the introduction as target anomalies) are not evaluated.

Possible experimental / analytical issues: - Single synthetic scenario: one blockage event in one network; no cross-validation, no multiple blockage severities, no partial/gradual blockage experiments beyond acknowledging they are a limitation. - No quantitative metrics reported (precision, recall, false positive rate, detection latency distribution); "satisfactory fit" is unverifiable. - Training and test data are both from the same simulator run; there is no out-of-distribution test. - Kernel specification entirely omitted — readers cannot assess or reproduce the GPR fit. - The claim that GPR "performs well with limited data" is invoked as motivation but 2,880 training points is not limited for a scalar regression task; this framing is unsupported.

Ideas for future work: - Validate on real sensor data from operational sewer networks, including sensor faults and missing observations, to test the simulator-to-reality gap. - Systematic sensitivity analysis of the persistence threshold (3 hours) across blockage severities and network topologies to derive principled threshold selection. - Benchmark against Rosin et al. (2022) and at least one non-probabilistic baseline (e.g., LSTM, ARIMA) using standard detection metrics on a shared test set. - Incorporate physics-informed kernel design (e.g., encoding hydraulic residence time or pipe capacity constraints) to improve robustness to gradual blockages — consistent with the stated future direction.