Two weeks after go-live the duty room is silent all night, or phones flood. The failures look opposite; the cause is often the same: thresholds copied from a demo, never tuned to floor noise—or an alert fires with no owner, no ticket and no close rule. Split “did the signal arrive” from “should we wake someone.”
- No alert: first ask whether the point is collecting. Offline with no cache goes silent: Sites that drop offline: how collection should be designed. Then ask whether the threshold is so high the floor never reaches it, or a maintain-silent was left on.
- Too many alerts: look for hold time and hysteresis. A spike that opens a ticket treats a door, a wash and a weld as incidents.
- Pop-ups with no loop: people are numb. Turn alerts into assignable tickets: Fewer screens to watch: how alerts become assignable tickets and After exceptions open tickets, missed handling drops.
A threshold is not a frozen number. It is a rule with a window
The same thermometer is not one line for a day-shift open door and a sealed night. Thresholds by shift and process beat one plant-wide 60°C. Keep an observe mode in trial: record fires, do not wake the lead, tighten after a week of data. Devices online with people still staring is often trust killed by false alerts: Devices are online. Why the floor still stares at screens. Cameras that only show a picture also look like “no alert”: Cameras without work orders are not software.
A wrong placement looks like haunting no matter how you tune software. A probe on an exhaust, a CT backwards, will be stably “wrong.” Bring floor photos; do not only edit config. Test notify paths too: SMS signature, WeCom, voice—right rules into junk still look like no alert. Who can change a threshold is a right; a night shift that raises the high will hide a real heat rise from days. Changing a threshold should itself audit.
When DaXi Technology accepts a hardware-software project, we look at false-alert rate, miss samples and close latency—not how many red lamps the wall shows. Close must take a reason. “OK” only brings the same false alert next week and you cannot tune. Send two weeks of “should have fired / would not stop” examples to Hardware-Software Integration, —easier to locate than “the system is wrong.”
Open observe mode first, then tighten thresholds
Trial starts in observe: record fires, do not wake the lead, tighten after a week. All-hands SMS on day one punches trust in three days. Day-shift open doors and sealed nights are not one line; thresholds by shift and process beat one plant-wide 60°C.