Cross-mapped NSW AIAF assessment for Predictive maintenance — water mains (UC-002)
ML model trained on SCADA + work-order history to predict pipe failure 30 days out. Pilot zone: Westbrook CBD network.
Predictive maintenance ML model for the Westbrook CBD water main network. Pilot zone covers 14.2km of mains across 9 suburbs. Out of scope: stormwater, sewer, recycled water networks (future phase).
Provisional: High (makes decisions about Council asset spend). Confirmed High after assessment — false positives may trigger unnecessary pipe replacement at $25k+ per incident; false negatives may miss imminent failures causing service outage.
SCADA telemetry (every 15 min), asset register (TechnologyOne EAM), work-order history (5 years), BoM weather feed. No resident PII. Retention: 5-year rolling window for training data; predictions retained 7 years per Records Act.
Azure ML pipeline; XGBoost classifier. Training set: 8,400 historical failure events. Test set: holdout final 12 months. Currently: F1 = 0.71, FP rate 11.2%, FN rate 8.4%. AIRC condition: FP rate must remain under 15%.
No personal information processed. PIA confirms PPIPA/HRIPA non-applicability. PIA artefact: EV-2026-0147.
Awaiting Records Officer review (queued 2026-05-25). Predictions and model decision logs must be classified as State records and retained per GA28 retention schedule.
Any predicted failure with recommended spend > $25k must be confirmed by a senior water engineer before works are scheduled. Predictions reviewed monthly by Water & Infra team. Manual override always available.
Azure ML (AU East), Python 3.11, XGBoost 2.1, MLflow tracking. Hosted in Council's own Azure tenant (no third-party SaaS). Data sovereignty: never leaves AU.
Recommend Approve with Conditions at G2. Conditions: (a) FP rate must remain < 15% reviewed monthly; (b) human-in-loop for spends > $25k; (c) 6-month formal review by AIRC.