See for example here:
Example strategies could be e.g.
- only look after cap artefact filtering
- treat each step individually (ignoring ramps)
- with some sliding window approach, using a "sensible" time scale e.g. dt=5ms, look for cases where
|mean(I[t, t+dt]) - mean(I[t+2dt, t+3dt])| > some sensible threshold, based on noise estimate, in words: a little segment before the jump differs from a little segment after the jump- but the slope of the before and after segment is the same (differs less than some threshold)
Or
- only look after cap artefact filtering
- treat each step individually
- fit the signal with multiple exponentials, or splines, or anything that wouldn't fit a little jump
- use the quality of fit as a jump detector
See for example here:
Example strategies could be e.g.
|mean(I[t, t+dt]) - mean(I[t+2dt, t+3dt])| > some sensible threshold, based on noise estimate, in words: a little segment before the jump differs from a little segment after the jumpOr