• Home
  • Audio
  • Coding
  • Photo
  • About
  •        

     

    April 10th, 2019


    This project is quite remarkable as it does what every well-seasoned electronics engineer does – it guesses. It learns by guessing, and finds the most optimal (yet cheap) way of delivering required performance from any given circuit. A million times per second.
    So, what is it?
    - You set the target response of the circuit (by drawing real-world biquad filters on the chart)
    - You draw approximate circuit that might implement the desired target response (without component values, as they are nearly impossible to find out without guesswork)
    - … yada yada …
    - You get the component values that implement desired target response! Magic!

    What it does:
    - Simulates any given electronic circuit (passive, analog), connected to any supplied impedance (important)
    - Randomizes component values in the circuit, and learns the outcome of every randomization (some might call it Monte-Carlo). If you keep all of the solutions in a neat array along with component values we’ve got thru randomization you’d get…
    - Machine Learning! All you have to do is calculate the distance (evaluate similarity) between given target response and previously obtained circuit solutions and pick the best one. Ta-dams, Artificial Intelligence! Almost there!
    - The moment you get a good match from previous step, you could kind of improve it by slightly randomizing the values of components (like 10-20%?), and find a lil bit better solution thru several passes of randomization and pinpointing (10, 8, 5, 1% “fuzziness” for each cycle).
    - But the most important thing is… Magic!

    Update 9:
    Another BIG update. Now with UI that makes sense, as well as event engine that makes everything update in real time. MVVM, bindings and ooooh. You change input response? Everything that uses it recalculates and displays results. You change targets? No worries!


    Update 8:
    THIS ONE IS BIG.
    Really big. I've started to put all the previous parts together so i could
     - draw the circuit
     - define target response
     - click the button and see the magic
    well, magic happenned and it took like 30sec to solve the more-or-less complicated circuit i've defined, as i needed to ramp-up the resolution for these shallow notches. Uh-oh, FAILURE.
    Yet,to get that solution i had to go thru pretty much every possible solution of that circuit... Uhm... what if i store all the solutions and their charts in memory, and then do a simple look-up of "most appropriate solution" to the target response from the memory? (calculating distance between 2 charts is very efficient and simple)...
    Hella yes! Let's do it in real time, as it is exceptionally fast!
    BOOM, MAGIC: it calculates the circuit by real-time-changing target response in real time:




    Update 7:
    Now we're getting graphic. Or graphical or something... I desperately need a target response which is monotonously sloped from top to bottom. No possible biquads could cover this range of frequencies in such a smooth linear manner (linear on log/log scale, thus very loggish)... Untill we chain a bunch of biquads together (shelving filters)... SUCCESS:




    Update 6:
    Now we need a tool to draw target responses. With a mathematically possible curves, typically your's truly low-pass, high-pass approximizations as well as PEQs, shelves and everything.
    Should conform to LTI, as we are talking passive-analog schemes to be solved.
    So, IIR filters we go, and biquads we use!
    Lil bit of chaining of biquads, tying their parameters to the movable "control points" on the chart and fast recalculations of signal chains in the background as you move the mouse and we are set:





    Update 5:
    Live dragging of component values on the schematics and LIVE simulation of the resulting circuit!




    Update 4:
    Circuit mutations! Creates all possible interconnections and components for specific number of parts.
    Like if we decide on 3 parts, it produces all possible interconnections and configuration of component types.
    Might help us in the future for absolutely-everything-solver implementation...



    Update 3:
    Import of frequency response and impedance files.
     - With smoothing! (mathematically approximate, yet more-or-less similar to other software packages. Frequency-dependant window function of the filter on log scale, so it keeps it's vital property of being "n-octave", in 2 modes - classic and complex!).
     - Phase extraction from frequency response via Hilbert Transform.
     - Various tweaks as time/phase alignment




    Update 2:
    Now we need some UI for circuit editor. Like this fancy one. Yeah, CURVES! (and easier to code, damn those splines):






    Update 1:
    Comparision of target response Vs solution. Green is target, red is solution. They-are-perfectly-the-same.
    IT WORKS!




    Hooray it's alive and kicking!
    Top pic - simulation by LTSpice
    Mid pic - my solver iterates thru randomntess (depicted by several layers of blue lines) to get the solution!
    Bottom pic - circuit i am trying to calculate (it's simulation is in the top picture).
    Monte-Carlo solver looks kinda promising!
    AND MY SIMULATOR IS ACCURATE!!!







    By s3t © 2011