The new filter worked better than I expected in a live data situation. This combines two ideas:
1. Sample the data as fast as possible and periodically solve at a slower rate a more regular linear regression as of the current timestamp with all measurements since the last time that happened.
2. Combine this result with an exponentially weighted moving average. I set an alpha of 0.34 based on messing around in a spreadsheet.
This turns out to be a very nice improvement.
Follow
Obviously the various parameters need to be adjustable as this sort of thing tends to need to be tunable to the specific roasting system and measurement hardware in use.
