Amplifier Linearisation by neural network

[CopperTop]

Over in the Sound Quality vs. Measurements thread there have been discussions on amplifiers that measure well (THD, IMD etc.) but apparently sound bad, and there's an ever-present assumption that feedback is the culprit, particularly on transients. For many, the goal seems to be an analogue amplifier that's linear without feedback - but unless the load is a resistor, it's hard to see how that would work well.

I don't know enough to judge whether there really is a problem with feedback, but if there is, the only true way to feedback-less operation would seem to be something like the methods described in the papers below. Has anyone here ever considered anything along these lines?

AES E-Library Adaptive Predistortion Filter for Linearization of Digital PWM Power Amplifier using Neural Networks
http://etheses.bham.ac.uk/510/1/Xiao09PhD_A1b.pdf
http://www.ece.mtu.edu/faculty/ztian...amplifiers.pdf

The idea would be to take a simple nonlinear output stage and drive it with a pre-distorted signal to achieve a perfect output into the load. The pre-distortion would be achieved digitally, using something akin to a tapped delay line feeding an artificial neural network (ANN) that had been previously trained on that particular output stage and speaker load. Thus, the system would compensate for the basic nonlinearity of the output stage, and would also control the speaker cone with a high damping factor (involving 'memory' as described in the papers above) without using any feedback at all.

The system would have to be trained offline using test signals and monitoring the output into the load in the time domain, seeking to reduce the errors by adjusting the ANN weights. Maybe a combination of a lookup table and ANN might simplify the requirements of the ANN and the training..?

Intuitively the ANN and its delay line would have to be large enough to handle the 'longest resonance' of the speaker. Say 100ms = 4000 taps or so? Calculating an ANN that size in real time at 44k1 Hz sounds like a huge task, but modern Intel processors can reputedly perform tens of GFLOPS - although the biggest hurdle might be memory bandwidth. I don't really have a handle on whether this is orders of magnitude too much for a supercomputer, or quite comfortable for an 8 year old PC. Maybe the problem would be an ideal one for GPU-based systems, CUDA etc. Maybe there are cleverer ways to reduce the problem.

It might be helpful to reduce the number of dimensions that the ANN needs to cover by, say, regulating the temperatures of the output transistors by external means, or else the transistor termperature might have to be another input to the ANN, along with ambient temperature and any other factors that could affect the predictability of the system.

It all sounds extremely complicated, but the actual hardware would be extremely simple, and all the clever stuff would be done in software and during the training. I can also see that there is a dichotomy between the simplicity and purity of the idea of 'no feedback' and all this digital stuff.

[jez]

You may be on to something there but for me that just sounds like dangerous modern talk and intuitively wrong.....
Yours,
T.Rex.

[counter culture]

Interesting suggestion CopperTop. Predistortion is an accepted technique for improving RF PA fidelity. It's a way of achieving results that are impossible any other way at present.

It's not the only way to achieve feedback-less operation, however. All that is required is linear devices.

My intuition suggests that predistortion is unlikely to satisfy the anti-feedback lobby.

Audiophile ideas seem to travel as a set. Let me see if I can list them.

Feedback is bad.

Digital is bad.

Solid state is bad.

Capacitors are bad.

Cables are bad.

Engineers are bad.

Measurement is bad.

Objective testing is bad.

Since predistortion would appear to require all of measurement, digital and engineers and possibly even some objective testing, this would appear to rule it out.

Audiophilistinism is not about finding solutions, rather it is about rejecting solutions. It is about one-upmanship, snobbery, retrogression. No solution will ever be good enough, and the profound hope of every audiophile is that intrinsically linear amplifying devices will never become available, because that would shoot the fox. If they should, then true creativity will be required, because a reason for rejecting that solution will then have to be invented.

The unspeakable in pursuit of the inaudible.

Don't get me started.

[Bigun]

Learn to play the piano and throw all your gear out. Problem solved.

[CopperTop]

Originally Posted by counter culture

All that is required is linear devices.

Presumably feedback is used in order to simulate a linear device that doesn't actually exist..? That is, a device that can provide a programmed output voltage regardless of the load. And the load isn't simply a resistor, but a springy electromechanical thing with coils and magnets.

[janneman]

The idea is OK in principle but will be (too) hard to implement.

There's basic two ways to do it:

1) predistort to cancel known non-linearities in the amp.
The problem with that is that those non-linearities are basically not known to any required precision, and they also vary with time, signal, load etc.
The principle is used sometimes with speaker drivers because there the nonm-linearities ARE known to some extend and are also larger than in electronic devices, so even with partial cancellation you can make the system more linear.

2) Derive the difference between input and output signal (the non-linearity or error signal) and either
2a) modify the effective input signal to get an error less output;
or
2b) add it somehow to the output to make the total output error less.

2a is feedback and we have already seen that this is controversial on psychological and prejudice grounds.
counter_culture's post # 3 is rather accurate in this regard.

2b is feedforward, and the problem here is to add two signals at the amp output, which is very low impedance. It can (and has) been done but is quite hard and not necessarily an improvement.


BTW counter_culture you forgot one: "three legs good, eight legs bad" ;-)

jan didden

[janneman]

Originally Posted by CopperTop

Presumably feedback is used in order to simulate a linear device that doesn't actually exist..? [snip].

You probably mean 'emulate' instead of 'simulate', but no, it isn't.
It is simply a way to subtract (a sample of) the output error from the input signal so that the output becomes the input without the error.

jan didden

[juma]

Originally Posted by counter culture

....The unspeakable in pursuit of the inaudible....

Great analysis, true said and well written!

[DF96]

Yes. Can you imagine the response from people who are frightened of coupling capacitors? Granted, the ones who prefer to use a DC servo might be willing to embrace something even more complicated but most would be horrified.

You need to consider the aim of people who use predistortion to 'linearise' an RF amplifier: they just want to avoid interfering in the adjacent channel (spectral regrowth) while preserving enough of the signal quality in their own channel so that the receiver can understand it (sufficiently low bit error rate). This is nothing like fidelity in the audio sense. They are trying to get distortion down from the -20/30dB region to -50/60dB (if my memory serves me). We want to get from -40 down to -80 or more.

[tiefbassuebertr]

Originally Posted by counter culture

Audiophilistinism is not about finding solutions, rather it is about rejecting solutions.

I agree. One makes always that experience, if one make (as an outside technical person) a suggestion for improvement to the developement engineer from certainly company.