I have a human activity accelerometer dataset. Here, in one activity segment, for example, for activity 1, the resting position signals, and the signals for activity 1 are mixed and the whole segment is misclassified as 1. There are no labels for the resting positions. Now, how can I separate or segment this 2 type of signal without any labels? Can anybody give any way or algorithm? I am attaching the picture of the signal. I want to separate the marked jones. Thanks in advance for your help.
Hope these papers be helful for you
The second one is a survey paper which has valuble information... can be very important for your task
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0208692
https://www.google.com/url?sa=t&source=web&rct=j&url=https://www.mdpi.com/2227-9709/5/2/27/pdf&ved=2ahUKEwja8JD5tKPrAhWpRBUIHVLVCrsQFjAHegQIBxAB&usg=AOvVaw369f0xtLVAwYaY6fAkYdmC
https://www.sciencedirect.com/science/article/pii/S1532046416300594
The parts of the curve that you have highlighted are very variable. For each time T, calculate V(T) = max(y(t))-min(y(t)) for t in a suitable window [T-W,T]. Plot V(T) against T. You will see that with a suitable choice of threshold V0, V(T)>V) will pick out the times you are interested in. Large W will reduce the false positives, small W will reduce the false negatives, so you will have to try different values.
I would pose the problem as a statistical hypotheses test. For example, the 'null' hypothesis might be that the means (or variance) of all samples over two adjacent (possibly overlapping) segments are equal. Create a test statistic to test this null hypothesis. Boundaries that mark changes in behavior are declared wherever this test statistic exceeds some (large) threshold. Use sliding windows to evaluate the test statistic, realized efficiently as FIR filters, or IIR if non-uniform weights are used.
- Badges
- Science topic