Too much Input layer in NN classifier

Help
#1

Project ID: 361954
Hello, I have a question: Why does my training model have an Input layer with 39 features, when each sample consists of 30 records? Please let me know.

#2

Hi @Mati81

Can you share the impulse configuration page, I guess there is some overlap configured here in your windowing.

Best

Eoin

#3

image
image1880×706 176 KB

Please let me know how can I change it.

Best regards,
Mati

#4

Addtionally I found arcticle about each parameters Impulse design | Edge Impulse Documentation, this article say that you should have more training windows, and features if you reducing window increase. It is true only for quantity training windows but for features I have const 39 features despite I changed window increase. Please let me know why it works in this way.

image
image1311×842 150 KB

Best regards,
Mati

#5

Hi @Mati81

For the spectral analysis block using FFTs the features extracted are based on the following:

  • There are two types of features generated per axis/channel: statistical features (RMS, skewness, kurtosis) and spectral features (maximum value from FFT frames for each bin).
  • The total number of features depends on the filter and FFT parameters.
  • Calculation for the number of features generated for a specific set of parameters is detailed in the docs here: Spectral features | Edge Impulse Documentation

Best

Eoin