When you find out at the next calibration check.
A thermocouple in a furnace is checked periodically and trusted between calibrations. Everything produced during that interval depends on the assumption that its calibration has not changed.
What fails
Base-metal thermocouples in furnace service drift. Thermal cycling, contamination from the atmosphere and metallurgical change in the wire all shift the relationship between the voltage the junction produces and the temperature it is at. None of it is dramatic and none of it announces itself in the reading.
The commercial consequence is specific to this sector. A heat treatment run is qualified against a periodic system accuracy check, so a sensor that has moved since the last check puts every part processed in the interval into question. The cost is not the sensor. It is the batch, and the paperwork that says the batch was made at a temperature nobody can now confirm.
Why recalibration is not the answer
You can recalibrate a thermocouple, and the industry does. The problem is that a periodic check tells you what the sensor is doing on the day of the check, and infers the rest. Tighten the interval and you buy confidence with downtime. Widen it and you carry more product on an assumption. There is no interval at which the underlying question, has this drifted since we last looked, actually goes away.
What a noise measurement changes, and what it does not
Our thermometer has no calibration relationship in the sensor to lose. We determine temperature from the Nyquist relation on every measurement, using a resistance taken from the same block of data as the noise, so a probe that has aged, oxidised or changed composition does not, by that alone, invalidate the temperature result.
The honest comparison is worth making in full, because it does not go our way throughout. When we measured ours against a thermocouple at 1,000 °C, the thermocouple was the quieter instrument: 0.47 °C of short-term scatter against 1.08 °C for ours. On a single reading, taken today, the conventional sensor wins.
What it cannot do is hold that performance. Scatter is random and falls with longer averaging. Drift is not random and does not. So the trade is short-term scatter now against a reading whose meaning does not walk away as the sensor ages, and whether that is a good trade depends entirely on which of the two is costing you money.
| Bath temperature | Measured scatter | Rice’s equation floor |
|---|---|---|
| 24.4 °C | 0.39 °C | 0.12 °C |
| 800 °C | 0.91 °C | 0.44 °C |
| 1,000 °C | 1.08 °C | 0.53 °C |
Standard deviations from the first high-temperature system, in a block bath, with a 1 MHz electronics bandwidth and a 5.872 s sampling interval. The effective bandwidth is lower than 1 MHz because of RC filtering in the probe, which is most of the gap between the measured figures and the statistical limit.
Where this has got to
High temperature is the most active part of the programme. The figures above are from our first high-temperature system, and we now run the work inside ThermoSI, where work package 2 is taking practical Johnson noise thermometry to 1,200 °C with a target uncertainty below 3 °C. That target is an uncertainty and not a scatter figure, so it is not the same quantity as the table, and it covers the systematic terms as well as the noise.
The part that needs application work is the probe. Sensor elements that stay resistive, stable and connectable at 1,200 °C are a materials problem before they are a measurement problem, and they are what we are developing in that work package alongside the electronics.
- Measured scatter at 1,000 °C, first high-temperature system
- 1.08 °C at 1,000 °C
- At 1,000 °C a thermocouple shows less short-term scatter than the noise thermometer, at 0.47 °C against 1.08 °C
- 0.47 °C thermocouple · 1.08 °C JNT
- ThermoSI work package 2 is taking practical Johnson noise thermometry to 1,200 °C, with a target uncertainty below 3 °C
- <3 °C at 1,200 °C
If this is your measurement
The thermometer is a working second-generation prototype. Where it goes next will be shaped by real applications. If calibration drift is limiting a measurement you are responsible for, we would like to hear from you.