Unfortunately, it looks like the SDS011 might not be suitable for longer-term comparisons.
My assumption at the beginning of the experiment was that I could calibrate the sensors by adjusting for measurement differences; by running the machines all in one place, I would get a single number that I could add (or subtract) from each machine to bring it into alignment with the others.
This might not be true.
When, for instance, compare Pi1 to Pi0, and subtract the adjustment number for Pi1, the results start off fine–but toward the end, they dip into the negative! Air pollution, obviously, can’t be negative. The trouble is that adjustment number. It is larger than the measured air pollution, so subtracting it leads to a number less than 0.
I’m pretty sure that this means that Pi1’s measurements are drifting over time–and that they’re drifting downward.
To find out, I subtracted Pi1’s measurements (near the tracks) from Pi0’s (far from the tracks). And, indeed, the difference between them grows, if slowly, over time. This implies that they are drifting out of alignment. Pi1’s measurements are drifting slowly downward. (Because of double negatives, (Pi0–[–Pi1]), the line has a positive slope.)
The results for Pi2 are similar, but less pronounced. Pi2 is drifting upward, but slower than Pi1 is drifting.
Of course, we must assume that Pi0 is also drifting.
The question now is, can I use a m+bx equation to adjust each Pi instead of a flat m? Of course, this won’t work for this experiment, since I can’t determine the m+bx from the data I have, since they’ve already been placed. But, perhaps in the future, I can, instead of using a fixed number, use an equation.