Sensor Accuracy Data

[Nofossil]

There's been some discussion about sensor accuracy, so I decided to get some data. While I don't have NIST-traceable calibrated temperature references, I did do an experiment with three different sensor types to see what accuracy I could get out of them over the temperature range from freezing to boiling. I didn't have any 1-wire sensors to test - I'll try to do some of them later.

I used a big long mercury (shhh - don't tell the EPA) lab thermometer as a temperature reference. My ice bath measured 33.5 °F and my boiling water bath measured 212 °F on the nose. I also used a room temperature midpoint at 67.9 degrees.

I used three different sensors: the MCP9700A and MCP9701A form Microchip, and the LM61 from National Semiconductor. All sensors were connected to my controller using standard 25' cables. With these sensors, the NFCS has about a tenth of a degree of 'noise' from reading to reading. The raw (uncorrected) errors were as follows:

Sensor      Error at 33.5    Error at 67.9    Error at 212
MCP9700A                1             -0.3            -0.5
MCP9701A              0.5             -0.7            -0.9
MCP9701A                1                0            -0.6
MCP9701A                1             -0.2            -1.2
LM61                    0             -0.2             2.9
LM61                  0.5             -0.3             1.1
LM61                   -1               -1             1.4
LM61                 -1.4             -1.5             1.8

It appears that the Microchip sensors are a bit more accurate, and it also appears that the fourth LM61 is a bit squirrely.
I developed a spreadsheet which uses this data to calculate calibration (gain and offset) values for each channel, which are then entered into the NFCS. With the calibration values in place, results were as follows:

Sensor        Error at 33.   Error at 67.9    Error at 212
MCP9700A              0.1             -1.0            -0.1
MCP9701A              0.1             -0.9               0
MCP9701A                0             -0.7               0
MCP9701A             -0.1             -0.8             0.2
LM61                    0             -0.7            -0.1
LM61                  0.1             -0.9               0
LM61                 -0.1             -0.5             0.1
LM61                    0             -0.7             0.1

The calibration algorithm essentially zeros out error at the endpoints. The error that we see is noise. All sensors exhibit a bit of nonlinearity, with a negative error at room temperature. Presumably they'd have a small positive error at temperatures below freezing and above boiling. This calibration approach yields pretty close to minimum error across the measurement range. If accuracy room temperature were more important, the offset value could be tweaked to yield zero error at room temperature. If you split the difference, you'd have a worst case error of half a degree across all 8 sensors across the range of freezing to boiling. The sensors also track each other within about +/- 0.25 °F . That's a good deal better than I expected. These are pretty inexpensive sensors, after all.

Here's a screenshot of the sensors at room temp after the tests.