New kind of feedback or just re-inventing the wheel?

With normal negative feedback you compare the output with the input and apply the difference as a correction signal. Of course, as you reduce the distortion and nonlinearities you also reduce the ability to reduce them(!) because you haven't got as much error as you had before (good) so in proportion you lose the ability to reduce what error is left. Sort of a 1/x function.

What about this idea I thought of today? Say you have a *unity* gain buffer or maybe an emitter follower half of an output stage. Compare the input and output signals and subtract i/p from o/p to get an error signal. No big deal so far. Put a summing cct upstream of your buffer amp. Feed the error signal to this summer so that what comes out of it and feeds the amp is signal + error. (Error may be pos or neg so may boost or buck signal when added to it).

Here is the big deal part - the feedback does not try to correct the amp per se, the summer merely feeds the amp an "antidistorted" signal that the amp then proceeds to distort into a good signal!! What's more, because the distorting efect of the amp is still present, you are still dealing with error signals of the original size, not vanishingly small ones that have been made so by nfb. Also, nfb can only make distortion smaller, even a squillion times smaller, but never eliminate it. But what I propose might be able to eliminate it completely in theory at least.

Is this something new or just old stuff I've never come across?

GP.
 

gromanswe

Disabled Account
2002-07-08 8:42 am
North
www.google.com
The timing is the issue.

When correcting a signal
you take a sample of the input
and compare it to the output.
But the correctionprocess in it self takes time.
In meanwhile there is a new input and a new output.

To be able to correct signal completely
there has to be some sort of delay.

This delay is found in Portable CDplayers.
They have memory buffer for the input signal
and then can correct the signal.

That is why you cannot have ultimate bandwidth
in a coventional NFB device.
The currents are somewhat dealyed through
the transistors.
So if you delay the amplifier a bit
correction has time to take place.

Groman
to much time at keyboard
 
Hi Circlotron

What you describe is basically the idea of error-feedforward, applied to just a part of an amplifier.
The original error feedforward approach was to compare the output signal of a whole amplifier (which might by itself use error feedback) with the input signal and add the inverted error signal at the output.
But what sounds easy in theory can be tough to implement in practice, since the device that sums the signals has to be perfect (as well as the error-amplifier).

Regards

Charles
 

Nelson Pass

The one and only
Paid Member
2001-03-29 12:38 am
What you describe with a unity gain follower ends
up being negative feedback in the classic sense.

This scheme can be seen to be operating in my
dynamic bias patent #3,995,5228 and later work
by Hawksford and (separately) Cordell.

There are plenty of other possible variations.
 
phase_accurate said:
The original error feedforward approach was to compare the output signal of a whole amplifier (which might by itself use error feedback) with the input signal and add the inverted error signal at the output.

Addding the error signal to the output becomes difficult when a power output stage is the one you are trying to linearise!:) The method I am thinking of is to apply the error signal to the incoming signal before it has reached anything else.

The opamp cct I have attached is deceptively simple. Beware! The opamp stage has a x2 gain. The incoming signal is halved in amplitude by the top left 100k and the one to it's right. The x2 gain brings it back to normal size again at the opamp output. Also anything fed back to the opamp + input is also halved by the top two 100k's. This way the opamp functions as a non-inverting summing cct.

Notice that both the - input of the opamp and it's negative supply rail is referenced to the fet source, so any opamp output at all is viewed as an error signal. i.e. if the opamp swings above the source voltage this means the source voltage didn't follow it perfectly and so there is an error (distortion).

Now the good bit. This error voltage - the ugly difference between the gate drive and the source output voltage - is fed back to the opamp input and *added* to the input signal so it now can make the output go where it should, and in theory at least *completely* eliminating any distortion.

In practice? At first it oscillated badly until I put the 47pf in place. Then I capacitor coupled a small 8 ohm speaker and was able to get a 5v p/p clean sinewave out of the thing. On music it sounded nice and clean and clear, if not terribly loud. I think it would make a great headphone amplifier.

Yes, there is feedback to the + input of the opamp, and the schematic is correct. It wasn't simulated, it was built. (I can't yet drive the blasted software to simulate it - PSpice Student.

My next trick will be to see about incorporating it into the CDA.

GP.
 
Whoops! forgot the schematic

`
 

Attachments

  • error feedforward opamp and source follower.gif
    error feedforward opamp and source follower.gif
    6.5 KB · Views: 1,227

gromanswe

Disabled Account
2002-07-08 8:42 am
North
www.google.com
On FEEDBACK use/non-use

My views are posted elsewhere.
But on the whole,
it depends.
An Adaptor might need feedback.
Amplifier is a bad terminology to use
I think.
Or not.
It is for you to determine
what is best for any given adoptation
using any given materials/components.

Not to adjust to environment/situation
will not give best performance.

So in signal adaptors
So in Nature

gro

or else Nature WILL Revenge
sometimes I have to walk the other way, without feedback
 
Here's another setup

Here it is as applied to a complementary symmetry setup. Note the zero volt point of the opamp supply rails rides up and down on the loudspeaker output as per the other cct.. This will definitely have to be a new project for me. The only thing about it, is that it looks so conventional it belies just exactly what goes on in the cct. The output at Q11 & Q14 emitters because of normal nonlinearity, does not quite follow Q12 & Q13 bases. The amount the driver bases differ from the output emitters, this same amount is also fed back by R9 and added to the input signal so that *exactly the right amount* is fed to the driver bases to make the output emitters do the right thing.

With all due modesty, ;) I think this is the best thing since sliced bread!

GP.
 

Attachments

  • eff to comp symm.gif
    eff to comp symm.gif
    12.9 KB · Views: 1,172

mirlo

Member
Paid Member
2002-07-12 9:55 am
San Diego
Hmmm.

Two things:

1) You have to make damn sure the negative feedback is greater than the positive feedback, or it will latch one way or the other. Since the resistive dividers are nominally equal and the emitter followers have a gain ever so slightly less than one, this could be a problem.

2) It can reduce the errors but not eliminate them. To see this imagine that the amp you are trying to correct has a bad second order nonlinearity such that if the input is X the output is X^2.

Then to get the output to be X, ie what you want, the input has to be the square root of X. But if your circuit is working to eliminate the errors, the input to the difference amp is some linear combination of the input X and the perfect output X. Can't happen.

On the other hand, playing with how NFB works can have advantages. You might find that you can make an amp faster and remain stable. by some such machination.

On a similar vein, I think there is an interesting app note on the Analog Devices website about improving opamp circuit phase response by putting a replica opamp in the NFB loop.

I am probably rambling incoherently now -- time for me to go pretend to sleep so i can get up and pretend to work tomorrow.

Zzzzzzzzzzzzzz ...........
 
Got you tricked, huh?

Hi mirlo, Thanks for your interest. :)

As for the points you raise:
1) would be true if the opamp output drifts positive for example, and the left end of R4 was tied to earth. The drift would be divided by 2 at the R6 / R4 junction and amplified x2 by the opamp. Big risk.

BUT... as the opamp output drifts positive, the amp output stage starts going positive too and this pulls the entire opamp "upward" which is the same as pulling the left side of R4 downward. R4 LHS goes effectively downward by *almost* the same amount that R6 RHS goes upward. So instead of the drift being halved and then amplified x2, it is reduced to nearly zero (the inverse of the emitter follower gain) and then amplified x2.

2) You say " the input to the difference amp is some linear combination of the input X and the perfect output X." Not quite. The input to the difference amp + input is a linear combination (actually the arithmetic sum) of input Y and (base drive sqrt X - amp output X). Amp output X gets subtracted from the input by the aforementioned lifting up of the opamp effectively putting a negative signal on R4 LHS. Tricky little jigger, isn't it?

So it isn't really negative feedback as such, at least the way I see it anyway. There is no 1/X funtion as per normal NFB. If you have say 1 volt error in the output stage, that 1 volt is fed back as a *positive difference feedback* to the input to make the input 1v bigger. Unlike NFB, the reduction in distortion does not reduce our ability to reduce the remaining distortion a la 1/X.

In a nutshell: We derive the error correction signal from the *constant* input / output difference *inside* the loop, not like NFB does, from the *diminishing* input / output difference *outside* the loop.

Having only a gain of x2, it remains to be seen what stability etc is like. Maybe someone could simulate it for me and let us know?

GP.

PS. I am awfully pleased with this little cct. :cool: :cool: :cool: :) :) :)
 
New kind of feedback etc

Circlotron,

Yes a neat little jigger!

I did some research on these types of circuits, even did an AES preprint on it in '98 I think it was. Dont have the reference here, will post it tomorrow.

There is a host of literature on the combinations of pos and neg feedback, dating back to the Radiotron Designers Handbook by Langford-Smith, now reissued on CDROM. A guy called Miller (yes, son of...) produced a couple of 1000 tube power amps based on this principle in the 50ies. Basic problems are:

- As mentioned above, these circuits can be very unstable. As soon as the neg feedback gets lower (for instance when the output approaches the rails) the pos feedback may dominate and the thing latches up;

- If you do complex analysis you will see that the distortion reduction only works for either a specific very small freq range, or only for DC and VLF;

- Although the distortion of the power stage is reduced, the distortion of the corrective stage (the opamp) is multiplied by the same factor. That may or may not be a problem in this specific case, of course.

One way to look at these circuits is that with the balancing pos feedback you create an amp stage that has infinite gain. From there on, applying neg feedback of course reduces the TDH to zero.

Cheers, Jan Didden
 
New feedback etc

Circlotron,

If you're serious about this, take a look at:

- Power amplifier output stage design incorporating error feedback correction with current dumping enhancement, M J Hawksford, AES preprint # 1993 (B-4);

- Novel feedback topology obviates the need for high loop gain, Johannes M Didden, AES preprint # 4597 (N-2);

- A mosfet power amplifier with error correction, Robert R Cordell, AES preprint # 1931 (D-9)


This is serious stuff, but you need to take the next step if you want to grow into this.
Contact me off-line if you have no access to these papers.

Cheers, Jan Didden
 
Fun with transformers...

Righteo then. Here's a version that is conceptually a bit easier to follow. Though it is quite simple it should be completely practical. I haven't made it up yet to see just what it goes like.

The 2 diodes, 2 caps and 2 resistors at the opamp output are for transistor biasing and can be ignored for now. The opamp is a unity gain buffer. Ideally the speaker signal should follow the opamp output perfectly but in real life we all know it doesn't. That's what makes DIY audio so much fun :) Any difference between the two will be impressed on the transformer primary, and so it will appear on the secondary in series with the input so as to add or subtract from the input. This forces the speaker output to follow the input properly.

For example, the speaker output is 10v, but the base drive from the opamp has to be (for the sake of round numbers) 11v to make that happen. This 1 volt difference finds itself applied to the transformer primary, dotted end positive in this case. That 1v appears on the secondary also, so the opamp + input will be 1v higher than the input signal. So the input signal only needs to be 10v to make the output to the speaker 10v! :) That's just what we want.

Seeing the transformer only has to swing a volt or so, it can be quite small. Best idea I think too is to reference the opamp zero volt point to the emitters and run floating rails for it, although not absolutely necessary for a small amp with limited swing.

Well, that's it. Now to go home and try it out. YeeHah!

GP.
 

Attachments

  • comp symm ff with trans.gif
    comp symm ff with trans.gif
    7.4 KB · Views: 758
New kind etc

OK, I can see where you are going. But I beg to differ: this is conceptually not different from the previous diagram with the opamp with the 100k resistors around it.

That one was conceptually a gain of 2 amp with 50% pos feedback, coming to an amp block with infinite gain. This is then wrapped in a neg feedback loop from the output, reducing the output stage THD to (theoretically) zero.

In the latest diagram you have a unity gain opamp with 100% pos feedback, again coming to an amp block with infinite gain, etc.

This will work fine, provided you make sure the wrapped-around neg feedback always dominate. That is difficult, because as the amp swings from zero to max output, or when the amp is faced with high frequency signals, the excess gain necessary for the feedback to work varies quite a lot, and all of a sudden it latches up for no apparent reason. I fried quite a few woofers with a similar amp. There are other caveat as discussed in my earlier post. Also, with the reduction in output stage THD, you get an increase in the opamp THD. In the end I found it didn't really provide a lot of advantages, other than a stimulating intellectual journey (which was worth it).

But it IS a fascinating concept, and maybe you find a better way to implement it than I did. I will follow this thread very interestedly!

Cheers, Jan Didden
 
It worked, but...

Stability was awful till I put the opamp zero volt point to ground instead of to the output emitters. Also, I was so lazy I just strapped the bases together and drove then straight from the opamp to see if it would flatten the xover distorion. Eventually I ended putting a pot across the trans secondary to adjust the amount of feedback and then it worked quite well till I got past a certain point and then it gradually slid into oscillation at about 1MHz. By this time it was 12:30 AM so I quit for the night. But it worked, so that's the main thing.

GP.

P.S. I actually did mean it was conceptually the same as the first cct, just that this one might be easier to follow, that's all. :)
 
I suppose there is such a thing as negative and positive feedforward, just as there is in feedback. Can anyone tell me the significance of the two (neg and pos?)

*******************

BTW, I just thought of a good example of FF and FB working together. In a modern car engine, the mass airflow sensor measures the mass (i.e. quantity of molecules) of air entering the engine and tells the computer, which then looks up a table to inject the corresponding amount of fuel. That is *feedforward*.

The exhaust manifold has an oxygen sensor screwed into it that sends a signal to the computer regarding the residual oxygen content in the exhaust gases so the computer can trim the air fuel mixture (normally 14.7:1 by weight) to a constant figure. This is the figure that makes the catalytic converter work best. That is *feedback*.

GP.
 
Calling Jan Didden...

Jan, I was re-reading the other day those papers you sent me, and thinking further about the opamp approach for this kind of cct, it appears that the positive feedback path merely nullifies the negative feedback path to any extent you want (or more!). Even though the opamp may only have an aparent gain of 2 because of the negative feedback resistors and therefore hopefully nice and stable when the loop is closed around the next device e.g. emitter follower, the positive feedback path makes the + input of the opamp move the same direction as the - input causing the output to have to swing much further to satisfy the feedback loop. I can't see the difference between that and simply raising the value of the nfb resistor and having no pfb. The end result appears to be the same in practice. What do you think?