I'm using a sliding window to obtain features (like mean, varianve etc) of a labelled time series of sensor readings. The goal is to train a binary classifier (like linear regression or SVM) to detect the 'event'(green) labelled parts of the data.

enter image description here

The problem is that I don't know how to handle windows that contain some green and some blue labels. My current approach is to set a threshold at which a window is taken as 'event'. For example if more than 90% of the step labels are 'event', then the whole window is considered as 'event', otherwise it is 'no event'. In the plot:

  • window a) covers 100% step labels 'event' => considered as 'event'
  • window b) covers 60% 'event' which is below 90% threshold => considered as 'no event'
  • window c) covers 0% 'event' => considered as 'no event'

Does that make sense or is there a better way? Do you have any suggestions?

I have to admit that I'm new to this field and I don't even know the proper technical terms for this kind of problem. Thank you and best regards


  • $\begingroup$ You say you want to "detect events", but the events don't seem to really exist--you defined events arbitrarily. What do you really want to determine in your data & why? $\endgroup$ – gung Mar 6 '16 at 22:43
  • $\begingroup$ Isn't the solution as simple as saying anything > 0 is an event everything else is no event? $\endgroup$ – forecaster Mar 6 '16 at 22:57
  • $\begingroup$ @forecaster unfortunately it is not that easy: the events are looking like the data points in the green colored area, they have certain characterstics (in terms of variance, amplitude etc). there are also outliers like high but small peaks in the signal. so just taking a simple threshold does not work $\endgroup$ – CShor Mar 6 '16 at 23:22
  • $\begingroup$ @gung Sorry if my description was misleading: I want to detect events in a stream of sensor readings using a classifier. I know what these events look like (more or less like the data points above the green area) so I labelled recorded data manually. Then a classifier should be trained using data. After training the classifier should be able to read new data coming from the sensor and detect the event. $\endgroup$ – CShor Mar 6 '16 at 23:32
  • $\begingroup$ Can you post sample data? $\endgroup$ – forecaster Mar 6 '16 at 23:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.