Which may be good techniques to face this abstract problem?
You have a data stream of a continuous signal, as one from a physical sensor. That signal has real (discretized) values, no attribute; addictional features (e.g., power, auto-correlation, entropy) might be extracted. You can assign one label from a finite set to a window of the signal. Let this label be a training label. You have to choise start and end points of the window as well as the window label.
The task is to classify next windows online, as just as the signal is received.
I am asking for an incremental algorithm, in the sense that it should increase its detection performance given more training labels. But it must be able to classify even after only one training label.
If the problem turns out to be too hard because of windows-boundaries detection, let's say you can fix their size at a small constant. Thus the algorithm classifies little slices of the signal and then it merges adjacent ones with same labels. If you use that simplified approach, please justify why it is reasonable.