I'm trying to classify states with a 1D-CNN-Network structure. Therefore I need to analyze the incoming vibrational data in a certain window. Two examples of such signals are shown below.

I'm continuously streaming the vibrational data. It should be possible to detect a trigger to determine the starting point of the process. This ensures that only interesting states are analyzed and classified. My first approach for such an trigger detection was a rolling window with an average value. If this average value exceeds a certain threshold value, the trigger is set. But this window must be relatively large, because some signals stand still for a longer time (see second picture).

Each process have a different duration between 45s-85s. Therefore it would be perfect to get a trigger for the starting and for the end point of the process. The signal value for stall can be seen at the beginning and at the end of each graph. A single peak (caused from a collision) should not be detected as a trigger.

This signal is equal to the first class. First Image

And here is a different signal for a second class. Second Signal

Which other algorithms would be useful for this application?

  • $\begingroup$ Care to share some data? I would calculate a moving variance, apply a "debouncing" to remove events smaller than the smallest expected event. Then identify start/stop event candidates and eliminate those that doesn't respect the 45sec-85sec constraints. You should get something quite close to the solution. A tuning petiod will involve you identifying false triggers and understanding why they happens. $\endgroup$ – Pier-Yves Lessard Nov 5 '19 at 1:29
  • $\begingroup$ Thank's your for your ideas. I tried your approach in a simplified way, but the results are not robust enough for me. I now use an external sensor as trigger for a better robustness. $\endgroup$ – DaniK Nov 7 '19 at 6:54

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.