0
$\begingroup$

I am rather new to the world of signal processing, and am struggling to understand a fundamental concept: How are filters actually implemented?

I have read a significant portion of this online book, and scrounged the internet, finding snippets of useful information here and there, but I cannot quite tie it all together.

In brief, I have a vibration profile (~200,000 samples) in the time domain, which I would like to analyze using a particular frequency weighting. This involves a four-step filter, using a (Butterworth) high-pass and low-pass, followed by an a-v transition and an upward step. The analog equations are shown below.

Filter Req's

Yet I am still struggling on how to implement them. I have seen mentions of the bilinear transform, used in programs in MATLAB/Python, but such implementations seem to omit the 's' or 'p' variable from the filters shown, typically creating the numerator 'B' and the denominator 'A' from the coefficients of the analog filters.

Other sources suggest taking the Fourier of the filter equations and the signal, multiplying, then applying the inverse Fourier to bring it back into the time domain. Yet I cannot discern how to take the Fourier of an equation.

My same problem occurs with using convolution in the time domain, as I have not understood how to convolve an equation yet.

I am sure I have come near the answer, and I know this is a basic principle, but I cannot seem to comprehend it, and my Googling is off the mark. Any help on this would be appreciated. General advice is also welcome, as I have a lot to learn.

** If important, I am currently working in Python, using the scipy.signal package, with access to MATLAB for testing and evaluation purposes.


Currently I am looking at substituting j2πf for p, then simply running the data through the equations and taking the inverse Fourier to bring it back into the time domain.

Alternatively, my eyes have just noticed what was staring me in the face: the second equation given for each filter is in terms of 'f', which would suggest I am able to simply run the frequency-domain data through this equation before taking the inverse Fourier to recover the time-domain. If this is a correct realization, this question can be closed (or I can 'answer' my own, now self-evident, question).

$\endgroup$
  • 1
    $\begingroup$ It's very unusual to go and start with an analog filter idea (butterworth in your case) and then transform it to the digital domain, just to achieve a certain frequency response – why not directly design the filter digitally? $\endgroup$ – Marcus Müller Aug 15 '17 at 21:45
  • 1
    $\begingroup$ The analog design given was specified by a standard that I am attempting to follow. It also gives the weighting function in third-octaves as decibels, but if it is feasible, I would like to use the (presumably) more accurate method $\endgroup$ – Pyrotrain Aug 15 '17 at 23:13
0
$\begingroup$

There are a few different ways to proceed. They involve doing some algebra but it is more an issue of tedium. The notation of your analog filters is a little non standard, essentially your $p$ is $s$ in most books.

Traditionally, phase wasn't considered important in hearing so standards tended to be specified by magnitude, so there is some ambiguity with respect to the digital implementation's phase response. You have 2 basic simple choices, a digital IIR filter or FIR filter. FIR is likely to be easier but easy and better are typically not the same. FIR can be direct or FFT based. IIR has a number of choices such as biquads or direct. Floating point math has its own easy/better issues.

One path way is to focus on the terms inside the magnitude brackets, i.e. the complex analog transfer functions. The other path is to focus on the analog magnitude.

So, given the analog transfer functions, you can either use a table of Laplace transforms and deduce the impulse response of each component you listed, or use the bilinear transform to express each component that is in $s$ in terms of $z$.

If you derive the impulse responses, you can sample points of the impulse response and provided a generally exponential decay, use those samples as coefficients of a FIR filter of sufficient length. The FIR filters are then applied sequentially. One nice property of ideal LTI filters, the order that you apply the filters are ideally interchangeable.

If you choose the bilinear transform and do the algebra to derive the discrete time numerators and denominators, you need to be aware that the bilinear transform squeezes the response as it approaches the Nyquist frequency. One needs to plot the discrete time response to asses how the corresponding analog response is distorted. If some modest tweaking results in an acceptable response then, the numerators and denominators need to be factored so that they are compatible with the IIR implementation, such as a cascade of biquads.

If you go with the magnitudes, the task becomes a direct sampling from the analog frequency domain to the discrete time periodic frequency domain. One can choose a linear phase response, and appropriately impose symmetry and define the frequency domain FIR filter. This might not sound like an analog implementation.

The essential challenge is mapping a frequency response in $-\infty < \Omega < \infty$ to $-\pi \le \omega \le \pi$

There are other approaches like state variables, so this is not the entire story.

$\endgroup$
-1
$\begingroup$

you already have the basic knowledge about analog filter, then get a book of "digital signal processing", check the chapter about "z transform","transfer function", "digital filter design"

$\endgroup$

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.