I am currently working on developing a strategy to measure a time-varying signal and would appreciate some input. In my signal acquisition process, the acquired signal $a(t)$ can be thought of as the convolution of the "true" signal, $y(t)$, with a kernel, $k(\tau)$.

\begin{equation} a(t) = \int y(\tau) k(t-\tau)\, \mathrm{d}\tau \end{equation}

The kernel I'm investigating exhibits the rise and fall-off characteristics of a cosine function, where the rise-time and fall-off time are determined by specific system parameters. However, I have the flexibility to introduce an arbitrary delay after the rise-period. This delay I can vary on a time-scale much shorter than the rise time. During this delay, the kernel "just continues" to aquire signal. Furthermore, I can sweep the kernel over the signal.

Signal situation

Currently, I'm considering two naive approaches. The first approach involves sweeping a minimal kernel over the signal, while the second approach entails increasing the waiting time. The challenge with the first approach is that the achievable time-resolution is largely limited by the slower rise and fall-off periods of the kernel. On the other hand, the second approach requires numerical differentiation in signal processing, which can be problematic.

Intuitively, I believe there could be benefits if we combine both approaches. However, I'm unsure how to effectively perform data analysis in such a scenario. I have the feeling that this problem must have been solved elsewhere.

I hope I have provided enough context, and I look forward to hearing your suggestions and insights.

Thank you!

  • $\begingroup$ I don't see anyone commenting. This was meant to be a comment, but I don't think you have anything to lose if you see this here instead? (my account too low in pts to post an actual comment) 1. >the acquired signal a(t) can be thought of as the convolution of the "true" signal, y(t), with a kernel, k(τ) Are you sure you mean convolution, not per-sample multiplication? Because a common thing to do ("problem that has been solved elsewhere", as you put it) is to take overlapping windows (your kernels; the step size would be something larger than 1 sample (so not sliding) but less than 1/2 the len $\endgroup$
    – user25849
    Jun 7, 2023 at 4:21
  • $\begingroup$ I don't follow the problem. Your "time-varying signal" is just "signal", and you're convolving it with something. If you're trying to recover $y(t)$ from $a(t)$, why not deconvolve? Or are $y, k$ both unknown? $\endgroup$ Jun 8, 2023 at 9:29
  • $\begingroup$ What is your goal? $\endgroup$
    – Royi
    Jun 9, 2023 at 7:51
  • $\begingroup$ Yes, I mean really convolution. The thing with the deconvolution is that has very unfavorable properties on noisy data. $\endgroup$
    – P. Egli
    Jun 12, 2023 at 14:00


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.