Skip to content
allanninal.dev/build
Collaborate

Part 4 of 7 · Cold chain monitor series ~5 min read

How a broken sensor is caught

The worst failure in temperature monitoring is not a sensor that stops. It is a sensor that keeps reporting a plausible, reassuring number that has nothing to do with the temperature.

Key takeaways

  • Real temperature is noisy. A perfectly flat line is a broken sensor.
  • Drift is caught by comparing a unit against its own recent behaviour, not against a standard.
  • A second sensor in one unit is cheap and settles most arguments.
  • Calibration is scheduled, recorded, and the certificate is stored with the readings.
  • Battery and signal quality are readings too, and they warn before the failure.

The flat line

How a stuck temperature sensor looks compared with a working oneA horizontal row of five boxes. Working sensor: minus eighteen point two, minus eighteen point four, minus eighteen point one. Noise is normal, from compressor cycles. Suspicious: minus eighteen point zero, repeated exactly. Eleven hours identical, to the decimal. It is stuck, not stable. A note says every threshold rule in the world says this fridge is fine.THE MOST DANGEROUS READINGWorking sensor-18.2, -18.4, -18.1Noise is normalcompressor cyclesSuspicious-18.0, -18.0, -18.011 hours identicalto the decimalIt is stucknot stableEvery threshold rule in the world says this fridge is fine.
Fig 1. How a stuck sensor presents. It passes every range check, produces a beautiful chart, and is reporting nothing at all.
  • Compute
  • Security & identity
  • Management
  • Analytics

The check is straightforward once you know to look for it: if the variance across a window is essentially zero, something is wrong. A working refrigeration unit cycles, and the sensor sees the cycling.

The threshold needs a little care because some sensors report to whole degrees, in which case identical consecutive readings are normal. The rule is really about the absence of any variation over a period long enough that a compressor must have cycled several times.

Drift, with nothing to compare against

How sensor drift is detected without a calibration referenceA vertical chain of five steps entered by a box labelled Is this sensor honest, with no reference available. Step one compares it to its own past, same unit and same season. Step two asks whether the baseline has shifted by a degree over months; if not it exits to Probably fine, keep watching. Step three asks whether anything changed, such as a new door seal or a new load pattern; if so it exits to Explained, annotate it. Step four asks whether a second sensor agrees, drawing on a side box explaining that two sensors settle arguments; if it agrees, it exits to The unit changed, not the sensor. Step five suspects drift and recommends calibration or replacement. A note says without a second sensor the last two boxes are a guess, and second sensors are inexpensive.AWS ACCOUNTIs this sensor honest?no reference availableCompare to its own pastsame unit, same seasonBaseline shifted?a degree over monthsProbably finekeep watchingnoDid anything change?new door seal, new loadExplainedannotate ityesSecond sensor agrees?if there is oneWhy twoit settles argumentsThe unit changednot the sensoryesSuspect driftcalibrate or replaceWithout a second sensor, the last two boxes are a guess. They are inexpensive.
Fig 2. How drift is investigated when there is no reference thermometer. The second sensor turns the final step from an inference into a measurement.
  • Compute
  • App integration
  • Machine learning
  • Security & identity
  • Management
  • Analytics

Two sensors per critical unit

The cost of a second sensor is trivial next to the cost of one ambiguous incident, and it converts an entire class of unanswerable question into an answerable one. When two sensors in one unit disagree by more than a degree, that is a finding in itself and it is worth alarming on.

They should be placed differently — one near the door, one at the back — which also produces genuinely useful information about how evenly the unit is performing. A walk-in that is three degrees warmer at the front has a problem that a single well-placed sensor will never reveal.

Calibration

What the schedule looks like

  • Annually at minimum, and more often where the stock justifies it.
  • Recorded in the same system as the readings, with the date, the reference used, and the offset found.
  • The certificate is stored alongside, because the question is always asked later.
  • A due date that is passed is an alarm, not a note. An uncalibrated sensor is an unverified record.
  • The offset is never applied retrospectively to old readings. It is recorded, and the readings stand as they were taken.
  • Ice point checks are cheap, take ten minutes, and catch gross errors between calibrations.

The fifth line is the one that surprises people. When calibration reveals a sensor was reading half a degree low, the temptation is to correct the historical record. The correct behaviour is to record the finding and leave the readings alone, because the record is of what was observed and the offset is a separate, later fact.

Battery and signal are readings too

A wireless sensor reports its battery level and signal strength alongside the temperature, and both degrade before they fail. A battery warning three weeks in advance is a maintenance task; the same battery failing at two in the morning is a gap in the record and a breach.

Treating those as first-class readings, with their own thresholds and their own place in the daily check, converts most sensor failures from incidents into scheduled work.

Next: who decides what happens to the stock.

All posts