0
$\begingroup$

I am quite confused about extracting features of the multichannel signals and wonder if anyone can help me out.

I want to get a feature set where feature is related to time, but I found the only way to do this was to cut the signals into different parts and extracted some features from each part, is there any way to extract something from the whole time series?

For example, there is a time series which contains 300 time epoch, and the feature set contains 5 features, features are time related, which means the last feature represents the last part of the signal.

Thanks in advance!

$\endgroup$
0
$\begingroup$

You can unroll your time series and use a LSTM or RNN. As for the test and train splitting of your time series, you could train your model using test data before a certain date and test with data after that date. It would help if you described your project as it would help me point you to the right tutorial.

$\endgroup$
  • $\begingroup$ Hello, I am now using RNN, but I don't want to directly input the signals because the results was not good , so I'd like to extract some features and use them as the input, that's why I need some time-related features as the RNN requires. And now I am confused about the feature selection...Thanks for your advice! $\endgroup$ – Ddj Apr 27 '18 at 4:04
0
$\begingroup$

The Zak transform or the discrete-time Fourier transform can accept a continuous input and divide it into it's composite parts based upon fixed or variable intervals. Desired or undesirable components, or gating, can be mathematically applied and a new (filtered or amplified) output derived.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.