TAP sequence recognition with accelometer

I’m having trouble with the recognition of different tap sequence (i.e. tap, double tap, tap).

I’m trying to teach my model with Spectrogram and Spectral Analysis processing blocks and Neural Network (Keras) learning block. I’m not sure which parameters should I use. Can someone help plz?

Which filter should I use to filter out ringing after a tap. Main goal of the model is to differentiate between multiple tap sequencies based on the time between peaks while disregarding the amplitude of peaks.

Link to project: https://studio.edgeimpulse.com/studio/25257/

Thank you =)

Hey @Smartylocky looking at your project it seems like you’ve made some improvements already, some tips:

  1. If the amplitude / frequency of the peak does not matter you can actually reduce the frequency bands to 4 or something.
  2. Your training / test set is pretty unbalanced right now which makes it hard to validate how well the model really works. You can rebalance the dataset from the Dashboard (at the bottom of the dashboard).
  3. Spectrogram block looks like a good choice here.

With 1 & 2 this yields 100% validation & 94.44% accuracy on the test set.