Analyzing vibration using machine learning

Question

I would like a few suggestions on an idea that I have -

I am trying to make a musical instrument (percussion), whilst just having a PVC disc. I am hitting the disc in a variety of styles (in order to produce a variety of sounds correspondingly), just like the way the actual percussion instrument is hit. I am converting the mechanical vibrations on the PVC disc to an electrical signal using a transducer, performing an FFT analysis of the different strokes, and trying to identify the stroke which is hit. Using this technique, I could get an accuracy of only 80 percent. I would like it to be extremely accurate ( more than 95 percent recognition). I was using only frequency as the parameter used to distinguish the sounds.

Now, I am thinking that if I could use other parameters too in order to identify the stroke, I might be able to get the required accuracy. I am thinking of resorting to Machine Learning for this. I am kind of new to this and would like to know what I might need to know before I proceed with this idea.

Any help would be greatly appreciated.

Answer 1

If you want to use machine learning for such a project, you can use vibrations data directly, and treat the problem as a regular audio classification problem.

A simple approach would be to use a Neural Network with Convolutions. This would take care of features extraction for you. And maybe follow these by dense layers at the end.

Given that, it would be easier to make suggestions if you posted samples of your data.

Edit: Also, keep in mind that machine learning usually requires large datasets so if you are collecting data all by yourself in the way you describe, you might not have enough sample to run a good model. In such a case when number of samples is limited one can use transfer learning - using a pre-trained model - but I am not aware of any such pre-trained models for wave data.