Alert Signal-to-Noise Ratio (ASNR)

GMI-WT-010Water TreatmentIntegration-Required

Definition

Proportion of investigated sensor alerts confirmed as real events requiring action, measuring whether the monitoring system produces signal operators can trust or noise they learn to ignore.

The Measurement Gap

A mid-size water utility invests $4.8 million in IoT sensors across its distribution system and treatment plants: pressure transducers, flow meters, water quality analyzers, acoustic leak detectors, level sensors. The system generates over 2,300 alerts per month. Operators investigate when they can, dismiss most, and silently ignore the rest.

After 18 months, a post-mortem on a missed contamination event reveals that the relevant sensor triggered an alert 47 minutes before customer complaints began. The alert was real. It looked identical to the 2,100 other alerts that month that were not real. The operator on duty had already dismissed 14 alerts that shift: all false alarms. Alert number 15 was the one that mattered, and it died in the queue.

This is not an operator failure. This is a system design failure that no existing metric captures. Water utilities track total alerts generated, alerts acknowledged, and response time to acknowledged alerts. None of these metrics answer the question that actually determines whether the IoT investment produces safety value: what fraction of alerts correspond to real events that require action?

Without ASNR, a utility cannot distinguish between a sensor network producing 80% actionable alerts (operators trust and respond quickly) and one producing 8% actionable alerts (operators have rationally learned to ignore the system). Both networks generate the same total alert volume. Both show similar acknowledgment rates. Only ASNR reveals that one is a safety system and the other is a noise generator that has trained operators to ignore real events.

Formula

Layer 1: Plain LanguageFor every alert the sensor system generates, track what happens next. Was it confirmed as a real event requiring action (true positive)? Was it investigated and found to be sensor noise or a non-event (false positive)? Was it never investigated at all? Divide confirmed real events by total alerts that were investigated.

Layer 2: Mathematical Notation

ASNR = TP / (TP + FP)

Complement metric:Alert_Miss_Rate = FN / (TP + FN)

Where:

  • TP = alerts confirmed as real events requiring action in period (count; source: alert disposition tracking)
  • FP = alerts investigated and classified as false alarm or non-actionable in period (count; source: alert disposition tracking)
  • FN = real events that occurred without generating an alert (count; source: incident reports where no prior alert existed)Note: Alerts that were never investigated are excluded from ASNR but tracked separately as "uninvestigated." A high uninvestigated count is its own problem (see ART, GMI-WT-011).

Proposed formula, pending validation pilots.

At a Glance

Unit
Ratio (0 to 1)
Trend Direction
Higher
Reporting Frequency
Monthly
Category
Prediction-to-Action Conversion
← SCADA Anomaly Attribution Accuracy (SAAA)Anomaly Resolution Throughput (ART) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment