Skip to content

AquaSentinel: Next-Generation AI System Integrating Sensor Networks for Urban Underground Water Pipeline Anomaly Detection via Collaborative MoE-LLM Agent Architecture

🕒 Published (v1): 2025-11-19 20:53 UTC · Source: Arxiv · Venue: AAAI 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AquaSentinel is a physics-informed AI framework for real-time anomaly detection in urban underground water pipelines that achieves network-wide monitoring using only 20–30% sensor coverage. It combines sparse sensor placement, physics-based virtual sensor augmentation, a Mixture-of-Experts (MoE) ensemble of spatiotemporal GNNs, dual-threshold anomaly detection (RTCA), and an LLM agent for generating actionable maintenance reports. On 110 simulated leakage scenarios, the system achieves 100% detection accuracy with 90.91% of anomalies detected within 10 timesteps (100 minutes).

Problem

Dense sensor deployments for underground pipeline monitoring are economically prohibitive for legacy infrastructure, while sparse deployments sacrifice detection coverage. Existing methods face a trilemma: cost, robustness, and interpretability cannot be simultaneously satisfied. No prior work achieves network-wide anomaly detection from sparse sensors by tightly integrating physical conservation laws with spatiotemporal learning, and pure data-driven or physics-only approaches each fail independently.

Method

AquaSentinel operates as a four-stage pipeline:

  1. Strategic sparse sensor placement: Nodes are selected by a composite score combining betweenness centrality, hydraulic significance (flow volume × pressure differential), and failure risk. Sensors cover 20–30% of nodes, deployed non-invasively on manhole covers.

  2. Physics-based state augmentation: States at unmonitored nodes are inferred by solving a constrained optimization (eq. 5) that enforces mass conservation at junctions (eq. 6) and Hazen-Williams head-loss equations (eq. 7), creating "virtual sensors" network-wide.

  3. MoE spatiotemporal prediction: Six expert GNNs (CaST, GMAN, ST-SSL, STG-MAMBA, STGCN, HydroNet) are dynamically weighted by exponentially smoothed recent loss to forecast normal hydraulic states. STGCN and HydroNet dominate due to lowest MAE/RMSE on the campus dataset.

  4. RTCA dual-threshold detection: For each node, both instantaneous error (e^RT) and cumulative error over window W (e^C) must exceed adaptive thresholds (based on exponential moving-average statistics with k₁=2.5, k₂=3.0) for T consecutive timesteps before an anomaly is confirmed. Localization traces upstream to find the root source node (eq. 15). An LLM then converts detection outputs into severity-classified, prioritized maintenance reports using structured prompting (D, C, H, T inputs).

Key Contributions

  • Demonstrates that sparse sensor deployment (20–30% coverage) with physics-based propagation matches dense-deployment detection performance.
  • Proposes the RTCA algorithm: dual-threshold (instantaneous + cumulative) adaptive anomaly detection that jointly suppresses false positives and detects slow-developing leaks.
  • Introduces causal flow-based localization that identifies leak source nodes and affected pipe segments from indirect observations.
  • Integrates a domain-specific LLM agent that translates algorithmic outputs into severity-classified, actionable maintenance reports with prioritization scores.
  • Provides a real-world sewer dataset (TAMUCC campus) and a 110-scenario PCSWMM-simulated benchmark for evaluation.

Results

  • 100% detection accuracy across all 110 simulated leakage cases (22 conduits × 5 leakage types).
  • 90.91% of anomalies detected within 10 timesteps (100 minutes).
  • Detection delay by scenario:
  • Constant >25% leakage: avg. 3 timesteps, 100% within 10 steps
  • Constant 15–25%: avg. 5 timesteps, 100% within 10 steps
  • Constant 5–15%: avg. 8 timesteps, 100% within 10 steps
  • Constant <5%: avg. 12 timesteps, 81.8% within 10 steps
  • Dynamic (0%→35%): avg. 15 timesteps, 73.6% within 10 steps
  • Best MoE expert (HydroNet): depth MAE=0.0085 ft, flow MAE=0.0038 cfs on real campus data.
  • No quantitative comparison against external baseline systems is reported; evaluation is ablative across leakage scenario types.

Limitations

  • No real-world leak validation: All 110 leakage scenarios are synthetically generated via PCSWMM simulation; no confirmed real leaks occurred during the deployment period.
  • Thresholds (k₁, k₂, W, T) were tuned on simulated data and may require recalibration for real municipal deployments with different pipe materials, flow regimes, and noise profiles.
  • Evaluation is on a single small campus network (23 nodes); city-scale scalability is unvalidated.
  • No comparison against prior anomaly detection baselines (e.g., CUSUM, isolation-forest, other DL approaches) is provided.
  • The LLM reporting component is qualitatively demonstrated but not quantitatively evaluated (e.g., no accuracy, hallucination rate, or user study reported).

Relevance to Agentic AI / LLM Agents

AquaSentinel exemplifies the emerging pattern of LLM agents embedded in closed-loop physical monitoring systems: the LLM is not the core detector but serves as an interpretive and decision-support layer that translates numerical outputs (anomaly flags, confidence scores, node IDs) into natural-language maintenance instructions with severity triage and prioritization. This is a concrete instance of domain-specific LLM agents operating at the interface of sensor data and human field operators, a pattern increasingly studied in CPS and infrastructure domains. For researchers tracking agentic AI, this paper is notable for its structured prompting architecture (D, C, H, T inputs) and severity-to-action mapping, which operationalize "tool-use + report generation" in a safety-critical physical domain rather than software or web environments.