Im fairly new into the world of audio systems, although Ive been passionate about it for a long time. I am an electronic engineering student and for my Thesis, I want to do something in the field of audio engineering. However Im unsure about the theory that is in my head and I would be very open to hearing some feedback about it from people who have more experience than me.
The idea is to try to reduce the crossover distortion which happens in Class B amplifiers. To achieve this I plan to simulate a bunch off different input signals of a Class B amplifier in Simulink. This simulations can then be used and labeled as the training set data. Systematically varying the parameters that affect crossover distortion, such as bias currents and voltages, should in theory create a diverse set of training examples.
With a working Machine learning model, the AI can tweak the biasing of the transistors in the amplifier. By adjusting the bias current, it can ensure that both transistors are partially conducting even when they're not in the active phase of the signal thus reducing the crossover distortion.
Let me know what you think of this idea, and if you think its feasible for a newbie to get into it.
The idea is to try to reduce the crossover distortion which happens in Class B amplifiers. To achieve this I plan to simulate a bunch off different input signals of a Class B amplifier in Simulink. This simulations can then be used and labeled as the training set data. Systematically varying the parameters that affect crossover distortion, such as bias currents and voltages, should in theory create a diverse set of training examples.
With a working Machine learning model, the AI can tweak the biasing of the transistors in the amplifier. By adjusting the bias current, it can ensure that both transistors are partially conducting even when they're not in the active phase of the signal thus reducing the crossover distortion.
Let me know what you think of this idea, and if you think its feasible for a newbie to get into it.
Non-switching OPS's ? Yes , played with those ...
-Analog .... with just diodes/semi's , the "off" class B semi is held at a idle current when not conducting.
There are many examples of this amongst Japanese OEM's.
-Digital , a fast opto is integrated into the bias circuit. Opto's also monitor Re at the output devices. ESP32/8266
then optimizes bias accordingly. I've done this in "dumb" mode to expand "first watt" amplifier mode. To do this
in real time would require a high bandwith Vbe and very high sampling beyond ESP realm.
-CFA , An amp with current feedback to it's inverting input is so fast as to nearly (fully) correct the crossover glitch.
Even in full class B , .01%. My preferred CFA is under-biased to 25-40mA (runs real cool) and still turns .0005% 20K (5PPM) THD.
You can see the "odd shaped" VAS waveform as you reduce bias down toward <20mA.
Most AB amps with typical slew .... THD rises considerably without ideal biasing.
Using something like "AI" to set the algorithm for the bias ? Class D would be the way to go. Some D's actually have firmware
to tweak the amp's response with the output as reference.
OS
-Analog .... with just diodes/semi's , the "off" class B semi is held at a idle current when not conducting.
There are many examples of this amongst Japanese OEM's.
-Digital , a fast opto is integrated into the bias circuit. Opto's also monitor Re at the output devices. ESP32/8266
then optimizes bias accordingly. I've done this in "dumb" mode to expand "first watt" amplifier mode. To do this
in real time would require a high bandwith Vbe and very high sampling beyond ESP realm.
-CFA , An amp with current feedback to it's inverting input is so fast as to nearly (fully) correct the crossover glitch.
Even in full class B , .01%. My preferred CFA is under-biased to 25-40mA (runs real cool) and still turns .0005% 20K (5PPM) THD.
You can see the "odd shaped" VAS waveform as you reduce bias down toward <20mA.
Most AB amps with typical slew .... THD rises considerably without ideal biasing.
Using something like "AI" to set the algorithm for the bias ? Class D would be the way to go. Some D's actually have firmware
to tweak the amp's response with the output as reference.
OS
You are really referring to class AB since the output transistors conduct for more than half the cycle.
Optimum biasing for class AB amplifiers has been thoroughly studied by many people including me. The problem is simple enough to be solved by conventional means.
For a machine-learning project, you need to pick a more complicated problem.
Ed
Optimum biasing for class AB amplifiers has been thoroughly studied by many people including me. The problem is simple enough to be solved by conventional means.
For a machine-learning project, you need to pick a more complicated problem.
Ed
... if it can perform some future telling: adjusting the required settings in advance of the actual signal to be processed.
If succesfull, I would turn that thing into a moneymaking machine instead!
They are already doing this with class D. Better to mix psychology/marketing to figure out how to scam you fellow citizen instead.
A "smarter" speaker would be the way. Primitive coil/cone does 1% distortion. Some amps correct for back EMF (primitive magnetic.
Correction at the amp is 1% of the problem.
Wouldn't it be funny if someone came up with a little 90% efficient speaker with .01% distortion and nobody could stand anything but class A anymore..
I wish you well and highly suggest you read Douglas Self's book on power amplifiers. A virtual treatise on crossover issues.
Underbiased solid state has a particularly nasty characteristic to it. 5% speaker distortion isn’t as bad as 0.5% from a crossover notch.
Nobody is coming up with a small 90% efficient speaker, BTW. If anything, speakers will get less efficient over time. We already know how to make 50% efficient speakers. I have some. They are just HUGE.
I know, I just get annoyed sometimes with the fact that we've gone from incandescent lamps to LEDs that are starting to approach complete conversion efficiency while we piddle around all day with incredible amplifiers that drive speakers which throw away more than 99% of the output no matter how you make it.
Even with inefficient loudspeakers, you only need a watt or two for home music listening at levels that won't make you go deaf.
Some interesting work done in the following thread: https://www.diyaudio.com/community/threads/power-amp-output-stage-measurements-shootout.374367/
Indeed and what it shows is the rather complex handling of the crossing phase by the relay-run of the complementairy devices. Can a by the op proposed machine take care of this? Sine waves, triangles (even harmonics) and square waves (uneven harmonics) are predictable, but not the 1813 cannons.
Seems like it would be helpful if we knew if the thesis is for an undergraduate or graduate degree, what sort or level of learning institution is involved (e.g is it Harvard, MIT, online certificate program, etc.?), what the OP future goals are, a job in industry, further education, self-employment in a field for the love of it, etc. To some extent a thesis project might be chosen to help prepare for some future goal in life.
For one example of a possibly more demanding project, investigating AB crossover biasing could be taken on at a somewhat complex level, with digital delays, pre-analysis of music files, development of a practical dynamic/programmable biasing system, etc. Or, maybe this is supposed to be more of a quick, easy project for an undergrad CS major focusing in machine intelligence?
OTOH, if this were to be for a serious dissertation that resolves a number of controversial problems in audio once and for all, then that would be yet another level to consider.
For one example of a possibly more demanding project, investigating AB crossover biasing could be taken on at a somewhat complex level, with digital delays, pre-analysis of music files, development of a practical dynamic/programmable biasing system, etc. Or, maybe this is supposed to be more of a quick, easy project for an undergrad CS major focusing in machine intelligence?
OTOH, if this were to be for a serious dissertation that resolves a number of controversial problems in audio once and for all, then that would be yet another level to consider.
Last edited:
Well, maybe let's see if we hear from the OP again? No sense going too far with ideas if he's losing interest already.
Thank you guys for all your replies. I have been doing some reading about amplifiers and audio systems lately, and oh boy is it a complex field. Im still figuring out what amplifier to use, as someone recommended the Class D on this thread. However just building a similiar amplifier seems to be a big challange in itself, and would broaden the scope of my project alot. Im still looking for a way to do this with the class B or AB amplifiers.
My thesis is for undergraduate level, and Im majoring in Information Engineering. Im not too concerned about having figured out every part of the problem and solving it. My aim is to try this idea, laying some ground work and maybe in the future it can be picked up by other researchers.
My thesis is for undergraduate level, and Im majoring in Information Engineering. Im not too concerned about having figured out every part of the problem and solving it. My aim is to try this idea, laying some ground work and maybe in the future it can be picked up by other researchers.
Of the classes you mentioned, only Class AB has some output device bias current when operated in the class A region. Class B in theory have no bias current through the output devices, as only one device is passing current at any given moment. Class D is another matter, as it doesn't have something like a class A operating region (at least in the sense that the output devices are operated as switches, rather than as linear amplifiers). Therefore it depends what you mean by 'bias' in that case? Deadband timing, inductor-less modulation scheme timing, or what exactly?
