Tuning PPE Alerts So People Still Act On Them in Month Six

Less safe, because attention is finite: a supervisor spending an hour a day dismissing alerts is not walking the site, and walking the site was working. The general economics of alert trust are covered separately. What follows is specific to PPE, where most of the volume can be removed without weakening the safety case at all.

1. Require dwell before alerting

Someone crossing a corner of the yard bare-headed on the way to the welfare block is not the hazard you are managing. Someone standing under a suspended load without a helmet for thirty seconds is. Requiring a person to remain in the zone for a meaningful interval removes most transient detections and barely touches the cases that matter, because genuine exposure involves spending time in the hazard.

2. Scope zones to actual hazards, not to the site

A blanket rule that a helmet is required everywhere generates alerts in the car park, the office walkway and the canteen queue. Monitoring where the hazard is — the lifting area, the workshop, the tank farm — cuts volume enormously and makes each alert self-evidently legitimate. It also improves credibility with the workforce, because the rule matches the risk rather than being applied indiscriminately.

3. Suppress repeats of the same person in the same place

One worker without a vest standing at a bench for twenty minutes should generate one open finding, not an alert every few seconds. Without de-duplication a single sustained situation floods the log and buries everything else. Grouping consecutive detections of the same condition in the same zone into one item, which closes when the condition clears, is usually the single largest reduction in volume.

4. Route to a supervisor, not to a person's record

Alerts should reach whoever can intervene in the next few minutes — the area supervisor — rather than accumulating against named individuals. This is partly a legal boundary, since identifying specific individuals from their face is biometric processing requiring a Ministry permit under the PDPL, and partly practical: intervention prevents harm, while a record against a name only allocates blame afterwards.

5. Separate real-time from reporting

Two different audiences need two different outputs. A supervisor needs an immediate alert for an active hazard. A manager needs a weekly compliance rate by area and shift, and does not need to see individual events at all. Systems that send everything to everyone produce an inbox that both audiences learn to ignore. This distinction also produces the trend data that makes the programme defensible, as the audit record depends on.

6. Fix the false positives before tightening anything

If the system flags workers who are in fact compliant — a vest worn open, a cloth under a hard hat, a glove hidden behind a tool — resolve that before adjusting thresholds. Alerts that the site knows are wrong are the fastest way to lose the workforce's cooperation, and cooperation is what the programme actually runs on. This is a detection and geometry issue rather than a tuning one, and it is discussed here.

The governance point that decides whether any of this survives

How the system is introduced matters more than how it is configured. A monitoring programme presented as a way to catch and fine people gets defeated by the people it monitors: cameras obstructed, monitored routes avoided, supervisors quietly declining to escalate. You lose the data and the safety improvement together.

The version that survives is stated honestly as measurement — we are finding out where compliance drops so we can fix the cause — and the causes are frequently not worker negligence. They are PPE stored too far from the work, a helmet unbearable in July, gloves that do not fit, a shortcut route nobody realised was in a hazard zone. A programme that fixes those earns cooperation. A programme that issues warnings gets obstruction.