what is the best way of detecting larger steps and slopes in a signal? (as shown in the pictures, 1. to be detected signal, 2. good signal) I thought about some kind of sliding window analysis. Thanks for help.
This problem is heavily related to the notion of piece-wise constant or piece-wise linear approximations. Either to fit the experimental curve at best, or to find the optimal parameters may requires some more details, like: do you know the number of parts? Any idea about the noise? Some literature to start with:
Will you be doing batch processing (processing the whole waveform at once) or online processing («live» processing with finite delay)? Assuming the latter:
If you can characterize the step vs slope and SNR is sufficient, you might get the job done with some tuned highpass filter (probably linear phase FIR ie «sliding window»)
If the only cases considered is the exact two plots offered (with a random noise component), you might look into matched filtering for detection.
Kalman filter detection seems like a direction if you want less ad-hoc tuning.