AlertPipe: Real-Time Fault Detection in WDNs

• AlertPipe System is a real-time, edge-oriented framework for unsupervised detection and localization of faults in water distribution networks.
• It integrates LSTM-VAE based anomaly detection, dual drift detectors, and change-point algorithms to rapidly identify pipe blockages, leakages, and bursts.
• The framework is optimized for resource-constrained edge platforms, ensuring scalability and robust performance even in non-stationary, label-scarce conditions.

An AlertPipe System is a real-time, edge-oriented framework for unsupervised event detection and localization in water distribution networks (WDNs). It integrates distributed pressure sensing, machine learning–based anomaly detection, statistical change-point detection, and graph-based localization to provide timely alerts for pipe blockages, leakages, and bursts. Deployments target resource-constrained edge platforms and support robust operation under non-stationary, label-scarce conditions. Architectures, algorithms, and evaluation results are detailed in "Unsupervised Online Detection of Pipe Blockages and Leakages in Water Distribution Networks" (Li et al., 22 Aug 2025) and "Real-time Pipe Burst Localization in Water Distribution Networks Using Change Point Detection Algorithms" (Mzembegwa et al., 2024).

1. System Architecture and Data Flow

The AlertPipe architecture comprises a multi-layer, hierarchical design enabling modular, scalable fault detection and localization.

• Sensor layer: Pressure sensors are installed at selected network junctions, capturing nodal pressure $p_j(t)$ at uniform intervals (e.g., every 0.2 s or 30 min).
• Edge agent nodes: Each sensor hosts a compute-efficient agent consisting of an LSTM-Variational Autoencoder (LSTM-VAE) for anomaly scoring, two drift detectors (Kolmogorov–Smirnov-based and distance-based), and local buffers. Agents operate fully decentralized; no inter-node communication is required (Li et al., 22 Aug 2025).
• Data ingestion and buffering: For burst scenarios, pressure readings are synchronized across junctions by a message broker (e.g., MQTT, Kafka) to permit instantaneous, network-wide snapshots (Mzembegwa et al., 2024).
• Detection and localization: Anomaly and drift signals from each node are logged and may be optionally routed to a central graph-based burst localization module, using WDN topology (via TSNet/WNTR/NetworkX) for pipe segment attribution (Mzembegwa et al., 2024).

The real-time per-sensor workflow (for LSTM-VAE-based detection) is:

1. Sample pressure $x^t_i$.
2. Construct a sliding window sequence $\{x^{t-T+1}_i, \ldots, x^t_i\}$ (default $T=10$).
3. Compute reconstruction $\hat x^t_i$, latent encoding $z^t_i$, and per-instance loss $\ell^t$.
4. Raise anomaly flags if $\ell^t > \theta^t$ over a cumulative window $\Delta$.
5. Populate buffers for drift analysis; on alarm, trigger retraining/reset protocol (Li et al., 22 Aug 2025).

For burst localization, real-time change-point detection (CUSUM or Shewhart) is applied per junction; earliest alarms propagate to a graph-based inference engine that identifies likely burst pipes based on temporal and amplitude ranking (Mzembegwa et al., 2024).

2. Anomaly and Drift Detection Methodologies

The AlertPipe system distinguishes two principal fault classes:

• Pipe blockages: Treated as collective anomalies, characterized by abrupt shifts in pressure that appear as elevated model reconstruction losses.
• Background leakages: Modeled as concept drift, manifested as slow, recurring alterations in nominal behavior over time (Li et al., 22 Aug 2025).

LSTM-Variational Autoencoder:

• Encoder: Inputs a length-$T$ pressure sequence; two LSTM layers (8 units, then 2 units) output $x^t_i$0, $x^t_i$1.
• Latent space: $x^t_i$2.
• Decoder: Mirrors the encoder via repeat vector, two LSTM layers, and a time-distributed output for reconstruction.
• Objective: Maximizes ELBO:

$x^t_i$3

with per-instance loss $x^t_i$4 ($x^t_i$5 default).

• Optimization: Adam (lr = $x^t_i$6), dropout $x^t_i$7, epochs 100/500 (initial/retraining) (Li et al., 22 Aug 2025).

Dual Drift Detection:

1. Statistical: Kolmogorov–Smirnov (KS) two-sample test over sliding latent buffers (size $x^t_i$8), with thresholds $x^t_i$9, $\{x^{t-T+1}_i, \ldots, x^t_i\}$0.
2. Distance-based: Euclidean norm between “anomalous” latent buffers (size $\{x^{t-T+1}_i, \ldots, x^t_i\}$1), alarm when $\{x^{t-T+1}_i, \ldots, x^t_i\}$2 (offline calibrated).

Upon drift detection, agents retrain the LSTM-VAE using up to 500 new samples, reset thresholds, and clear buffers to adapt to new operating regimes (Li et al., 22 Aug 2025).

Change-Point Detection Algorithms:

• CUSUM: Tracks cumulative deviations; triggers when $\{x^{t-T+1}_i, \ldots, x^t_i\}$3 or $\{x^{t-T+1}_i, \ldots, x^t_i\}$4 exceeds threshold $\{x^{t-T+1}_i, \ldots, x^t_i\}$5, using drift parameter $\{x^{t-T+1}_i, \ldots, x^t_i\}$6 for noise control.
• Shewhart: Applies control chart on sliding window; flags burst when $\{x^{t-T+1}_i, \ldots, x^t_i\}$7 exceeds $\{x^{t-T+1}_i, \ldots, x^t_i\}$8 or $\{x^{t-T+1}_i, \ldots, x^t_i\}$9 (Mzembegwa et al., 2024).

3. Localization and Inference Algorithms

For pipe burst events, localization proceeds in several ranking and back-tracing steps following CPD-based alarms:

1. Alarm ranking: Ordered by earliest detection time $T=10$0 and largest pressure drop $T=10$1.
2. Neighborhood inference: For pairs of adjacent alarm nodes, identify connecting pipe as burst candidate.
3. Predecessor pruning: If a single upstream neighbor exists, attribute burst to corresponding segment.
4. Fallbacks: Use maximum pressure drop or lowest average pressure among predecessors (Mzembegwa et al., 2024).

Graph construction exploits WNTR/TSNet exports for directed network topology; traversal and neighbor queries use NetworkX primitives.

In non-burst contexts, system agents log and timestamp each anomaly or drift flag with the current loss, $T=10$2-value, and $T=10$3 statistic. Heartbeat reports aggregate buffer states and model health (Li et al., 22 Aug 2025).

4. Implementation, Memory, and Computational Properties

All detection operations are designed for real-time, edge deployment:

• Memory: Five fixed-size circular buffers (ref_N, mov_N: 200; ref_AN, mov_AN: 50; mov_warn: 1000) consume $T=10$4k floats ($T=10$510-20 kB).
• Model footprint: LSTM-VAE weights per agent occupy $T=10$610 kB.
• Per-step compute: LSTM-VAE inference (1 ms per sample on ARM Cortex-A53), KS-test over 200 points (2D), one distance calculation (50x2).
• Retraining: 200 ms/epoch × 500 epochs $T=10$7 100 s on edge GPU; infrequent due to drift rarity. Can be offloaded if necessary (Li et al., 22 Aug 2025).
• Data pipeline: For TSNet-simulated burst scenarios, pressure readings are streamed and synchronized per time step; CUSUM/Shewhart statistics maintained per node (Mzembegwa et al., 2024).

Scalability is preserved because each sensor acts independently; adding sensors scales linearly without global orchestrator dependencies. Isolation mode supports resilience under network partitioning (Li et al., 22 Aug 2025).

5. Evaluation and Comparative Results

AlertPipe-style systems demonstrate strong empirical performance using prequential G-mean and localization accuracy as primary metrics.

• Blockage/Leakage Detection (Li et al., 22 Aug 2025):

| Benchmark | Steady G-mean | Post-drift Drop | Recovery | False Positive Rate | |-----------|---------------|-----------------|----------|--------------------| | Hanoi | ≈0.95 | ≈0.6 | >0.9/100 steps | <5% | | ZJ | ≈0.75 | similar | >0.9/200 steps | ≈8% |

Downstream node (pipe 7, Hanoi): 991/1000 true positives, 174/1000 false positives; upstream/unaffected nodes: TP<200, FP<100.

Comparative baselines (iForest++, LOF++, VAE4AS) cap at G-mean 0.7 (Hanoi) and 0.6 (ZJ). LOF++ and VAE4AS demonstrate slow or unstable drift adaptation.

Ablation: Removing distance-based drift detection (DD2) increased drift detection latency by ≈25%; vanilla VAE (no LSTM) degraded blockage detection and reduced G-mean by ≈15%.

• Burst Localization (Mzembegwa et al., 2024):

| Capture Interval | CUSUM Accuracy | Shewhart Accuracy | |------------------|---------------|-------------------| | 0.2 s | 92% | 96% | | 2 s | 72% | 72% | | 5 s | 52% | 32% | | 10 s | 64% | 60% |

Average: CUSUM 70%, Shewhart 65%. CUSUM outperforms Shewhart at coarse sampling; Shewhart only excels at the highest sensor rates.

This suggests fine sampling intervals are critical for Shewhart optimality, while CUSUM offers better robustness at practical reporting rates.

6. Deployment, Operational Tuning, and Limitations

AlertPipe system deployments require calibration of window sizes ($T=10$8, $T=10$9), drift/anomaly thresholds ($\hat x^t_i$0, $\hat x^t_i$1, $\hat x^t_i$2), and retraining frequencies according to network dynamics and operational constraints. Recommendations include:

• Drift detector windowing: Increase for slower leakage evolution.
• Statistical thresholding: Initialize via 5–10% significance analysis on historical nominal data.
• Retraining cadence: Limit to once per day if possible; adopt incremental retraining for frequent drift (Li et al., 22 Aug 2025).

Deployment challenges:

• Sparse sensor placement may yield detection “blind spots”; placement optimization may be required (Mzembegwa et al., 2024).
• Data latency or missing data violates synchrony assumptions; buffering and time-outs become essential.
• Accurate burst localization presumes high-fidelity network hydraulic models and absence of background leaks; model uncertainties can degrade performance.
• Discriminating bursts from other transients (e.g., valve or pump events) may need external data correlation.

7. Conclusion and Outlook

The AlertPipe System synergizes unsupervised learning (LSTM-VAE), dual drift detection, and rapid CPD-based localization to deliver robust, adaptive detection of pipe blockages, leakages, and bursts in WDNs. The lightweight, edge-compatible implementation, bounded memory, and empirical superiority over classical baselines render it an effective paradigm for modern smart water infrastructure monitoring (Li et al., 22 Aug 2025, Mzembegwa et al., 2024). Scalability and resilience to communication loss or topological changes are inherent in its fully decentralized design. Further improvements in model-based sourcing of anomalies, context-aware sensor placement, and the integration of physics-informed or multimodal data are promising avenues for future research.