# Extracting features from a signal: a 21st century approach?

I have a discrete signal of random length (let's say between 500 and 800 in time). This signal has an important information which I want to collect and to put it together in a feature vector to train a classifier. The signal also comes in random phase and in random scale (it depends on the distance to the sensor).

So there are some constraints: the feature vector has to be invariant to scale and phase, it has to be a sparse representation of the information contained in the signal and it has to deal with the problem of different signal length.

I have followed a classic approach to solve this problem:

I calculate the Fourier coefficients of the signal, a[n] and b[n]. I take as feature vectors the most significant coefficients for my problem (I fix this parameter experimentally). To solve the problem of scale and phase invariance I consider only the normalized magnitude r[n] = s[n]/s where s[n] = sqrt(a[n]^2+b[n]^2).

This solution actually works quite well, but I am sure that there has to be something more "up-to-date" to solve that. I would really appreciate any alternative approaches to solve this problem or some references to read.

Thank you.

• If it's simple and it works well, it very well might be the newest method, because who could ask for more? If you're looking for more advanced and sexy algorithms, the answer to your question probably will depend on the exact problem you're solving. – MackTuesday Apr 17 '14 at 19:39
• Because any improvement, even only a 1% in the performance of the classification stage, is crucial for my application. Also, I am curious about different approaches for this problem. – gui Apr 18 '14 at 11:35