How the ranking system supports field technicians

Visual simulation

From nightly sound measurements to a focused inspection route

Devices detect acoustic activity. The ranking system compares each signal with its own history. Technicians remain responsible for confirming the real-world cause.

Decision supportNot automatic leak confirmation
Illustrative example: device names, scores and outcomes below are fictional. The workflow uses the real project design and the model combination selected by the benchmark.

The complete operational loop

1SensorDevices listen443 network devices record nightly acoustic summaries.
→
2DataSignals arriveMinimal, Frequency and Spread are added to each device history.
→
3RankSystem ranksThree complementary signals produce one relative priority list.
→
4VisitTechnicians inspectSeven ranked candidates plus one random audit visit.
→
5ResultReality feeds backConfirmed leak or non-leak becomes evidence for later validation.

What changes for the technician?

Current alarm view

Many devices look equally urgent

D-104CriticalMinimal 531
D-219CriticalMinimal 518
D-087CriticalMinimal 502
D-311CriticalMinimal 487

The threshold sees the latest level. It does not show whether the value is unusual for that specific device or persistent over time.

+
Ranking-assisted view

A relative priority with context

1D-087Sharp change from own baseline94
2D-104Persistent rise over 7 days89
3D-311Models agree on elevated priority81
7D-219High level, but normal for this device56

Scores are relative priority points, not leak probabilities. The technician still verifies the pipe, valve, customer usage and surrounding noise.

Example inspection route

12alarm candidates
7highest ranked visits
1random audit visit
8total field checks
Visit order Device Why it is on the route Selection Technician records
1 D-087 Strong change from its own normal level Ranked Leak / no leak + observation
2 D-104 High and persistent acoustic signal Ranked Cause, location and confidence
3 D-311 Linear and tree models agree Ranked Repair required or false alarm
4–7 Four next devices Next highest combined priorities Ranked Same structured field result
8 D-402 Randomly selected outside the top seven Audit Checks what the ranking may miss

The audit visit is essential. Without it, the system only learns about devices it already preferred and cannot measure hidden misses.

What the three ranking components contribute

1 · Personal history

Device anomaly

Asks: “Is tonight unusual for this device?” This avoids treating every naturally noisy location in the same way.

2 · Pattern rules

Extra Trees

Finds combinations such as persistence, variability and frequency changes that may matter together.

3 · Bootstrap voting

Random Forest

Checks whether the same patterns remain useful across many resampled versions of the training data.

Device anomaly+Extra Trees+Random Forest→One inspection priority

What this can improve—and what it cannot

Expected operational value

  • Give technicians an ordered list instead of equally urgent alarms.
  • Use recent history, not only one threshold.
  • Record why a device was prioritized.
  • Build better evidence from every completed visit.

Current limits

  • Only 99 usable confirmed leaks trained the corrected benchmark.
  • The historical visits were already biased toward alarms.
  • The system cannot confirm a leak without field inspection.
  • Prospective campaigns are required before production automation.

Safe rollout

NowShadow mode

Generate rankings without changing technician decisions. Compare afterward.

NextAssisted pilot

Use seven ranked visits plus one audit. Monitor misses and false alerts.

LaterProduction support

Deploy only after new campaigns confirm stable improvement.

The ranking system recommends where to look first. The technician determines what is actually happening.