0
$\begingroup$

I want to know if my impact-echo signals are proper with CNN or RNN regression model.

I got some simulated signal, as following shows.figure

In previous research, people mostly consider frequency or even time-frequency analysis. For deeper crack the signal response lower frequency here.

Singal input are their 1D signal, and want output present their depths. But my train simulated crack are all create integer (e.g. 8, 10, 12, 14, 16, 18, 20, 22, 24, 26), so I'm wondering if these signals in time domain using CNN like this or RNN like this model do regression task is reasonable. (Is my RNN model wrong or BIG concept misleading)

And I hope the model can learn (e.g. train 10cm and 12cm depth crack signal, when the model input is a 11cm signal it can be predict as 11).

Also if it is reasonable, which paper should I read or some paper related to enhance knowledge.

If the question is not clearly describe, please told me (I'm not native in English). Thank you guys.

Improve this question
$\endgroup$
5
  • $\begingroup$ I see you have many points (X axis) in each depth crack. Do you need to put entire signal to neural network or could you cut to small period like 1024 points and feed to neural network? $\endgroup$
    – Cloud Cho
    Commented Sep 21, 2023 at 21:41
  • $\begingroup$ @CloudCho Oh you are right! In frequency domain analysis, we need to take a longer time to achieve higher frequency resolution. However, on the flip side, a longer time duration can make the signal more susceptible to interference from boundaries or other reflections. This is one of the advantages that I think 1D signals can get good performance. Currently, I'm using 448 data points (μs) for each input signal. I'm wondering what I'm doing (make CNN or RNN doing these regression) is reasonable. $\endgroup$
    – hui30319
    Commented Sep 21, 2023 at 22:49
  • $\begingroup$ 448 point in X axis would be doable in CNN. By the way, how human detect the depth of crack visually from the graph? $\endgroup$
    – Cloud Cho
    Commented Sep 21, 2023 at 22:50
  • $\begingroup$ On-site, I think people use FFT to check frequency peak (at high frequency). And decide if there have a crack under impact point. $\endgroup$
    – hui30319
    Commented Sep 21, 2023 at 22:59
  • $\begingroup$ Then you may consider to use FFT plot instead of time series as your dataset. $\endgroup$
    – Cloud Cho
    Commented Sep 21, 2023 at 23:04

0

Reset to default

You must log in to answer this question.