Neural network input/output problem, signal estimation

Hello everyone.

I'm new in NN stuff and I've got a problem.

I want to project a neural network which can estimate amplitudes and frequencies of periodic signal with additional noise. This signal is a sum of 2-4 sine waves (+noise). I tried to solve it in several ways but nothing works. Maybe can you give me any advice, what should I give to input, and what kind of array/matrix I need to put on target output? What kind of network should it be? I tried to put on output every single A*sin(2*pi*f*t) but on training data - that works, on testing data - definitely no.

I will be so grateful for every advice! Hope you can help me! :) Thank you.

Answer 1

[Greg Heath]

Neural Nets are not appropriate for that type of task. See

 help fft
 doc fft

Hope this helps.

Thank you for formally accepting my answer

Greg

[Michal Taraaram]

But theoretically if I have real amplitudes (A1, A2, A3) and frequencies (f1, f2, f3) from signal without noise, is there a possibility to project neural net that on input have a matrix with few rows containing few samples of signal with noise, and on target there will be some kind of array containing those amplitudes and frequencies?

Answer 2

[Staffan]

Hello Michal,

Could this solve your issue?

http://se.mathworks.com/matlabcentral/answers/121579-curve-fitting-to-a-sinusoidal-function

I just used the code in the accepted answer to fit a sine function. Since the original data (simplenar_dataset, may be used e.g. when working with narnet) had a rather sparse sampling rate I added a third vector to the figure in which I decreased the between the points with two orders of magnitude. If you would like to do the same, just replace the plot command with this line:

plot(x,y,'b',  xp,fit(s,xp), 'r', min(xp):0.01:max(xp),fit(s,min(xp):0.01:max(xp)),'g')

Regards

Staffan