The Tayloe Mixer: A Low Noise Solution

[Carl-NC]

True, but again you do not require any amplification prior to the mixers, and as a bonus you don't have to mess with balancing the opamps prior to mixing. Double balanced action is achieved with centre tap on Rx coil.

The "full-wave" demod method also doesn't require amplification prior to the mixers, it's just that all detectors do it that way. It also could easily be implemented with a CT coil.

Point is, it appears that VLFs have been using Tayloe's method for many years.

[Carl-NC]

I assume Carls drawing clocks at 4x the same as Dans

Actually, no... VLF full-wave demods use overlapping IQ clocking, which means that the clocks run at the TX frequency, not a multiple. Drawings below illustrate clocking for full-wave vs. Tayloe, and the resulting signal gating (without integration) that they give.

So that is a difference between traditional full-wave demod and Tayloe, although there is no reason why full-wave demods could not be clocked identically to Tayloe. It would be interesting to see if either method (overlapping vs non-overlapping clocks) has an advantage.

- Carl

[simonbaker]

Actually, no... VLF full-wave demods use overlapping IQ clocking, which means that the clocks run at the TX frequency, not a multiple. Drawings below illustrate clocking for full-wave vs. Tayloe, and the resulting signal gating (without integration) that they give.

So that is a difference between traditional full-wave demod and Tayloe, although there is no reason why full-wave demods could not be clocked identically to Tayloe. It would be interesting to see if either method (overlapping vs non-overlapping clocks) has an advantage.

- Carl

The Tayloe circuit Carl compared in this example is not "full wave", so I would presume it would be at a disadvantage S/N-wise.

To make the full-wave Tayloe it seems you need to add some complexity, one way is the "double balanced" design Golfnut attached which also added the full-wave feature as well.

I haven't had a chance to see what is the simplest way to make a full-wave Tayloe.

For MD work, Carl's circuit seems most direct way to get I/Q info, although generating the sync pulses requires some hidden complexity (delay circuit? How is it done typically?) if you want to use the same freq as the TX. (maybe integrate or take derivitive of I signal to get a zero-crossing at quadrature???).

The appeal of the Tayloe circuit is it's flexibility in the way it divides the cycle into four parts which can be combined in various ways -- kind of a mathy thing... disadvantage for some MDs is inability to clock off the TX oscillator.

-SB

[Carl-NC]

For MD work, Carl's circuit seems most direct way to get I/Q info, although generating the sync pulses requires some hidden complexity (delay circuit? How is it done typically?)

Traditionally the clocks are generated from the TX signal through comparators, using simple RC delays. An RC-CR filter makes a handy 90-degree phase shifter.

Newer designs let the processor generate the clocks; it's pretty easy to use the PWMs to create autonomous clocking.

[golfnut]

This is good stuff, I must apologise as my limited experience of vlf dets is IGSL and that is my default benchmark.
Potential improvements I read about are relative to IGSL.
So I was looking at Dans work vs a simple 4066 type. The potential improvements - as yet unproven would be a little like the attached.
This is what I was eluding to but not very clearly. In this very simple model, a lineup with a preamp versus a non preamp and sampling integrator mixer.

What I dont know is the real NF of the IGSL preamp - anyone - so I can plug the number in..

This model shows that having no preamp and a low loss mix can be better than an amp and a 4066 type mix.

"So that is a difference between traditional full-wave demod and Tayloe, although there is no reason why full-wave demods could not be clocked identically to Tayloe. It would be interesting to see if either method (overlapping vs non-overlapping clocks) has an advantage."

Dont know, cant get my head around it - Overlapping clocks,, is this like sampling twice?
Two bites of the cherry to fill up the integrator cap with bites of signal?

I guess all of this is like having an AtoD with 1 sample per 1/2 cycle or 2 samples per 1/2 cycle . In time we may not use a mix/det at all and sample the Return and the Tx directly and do the math in dsp.

Thanks for the effort here everyone it is appreciated

Steve

[golfnut]

late again

If nobody voices any objection - I would lke Dan to Join in with this thread and get some input from him. Ive seen his mail addr here and there. He may pick up some tips from current scemes too.

S

[simonbaker]

Dont know, cant get my head around it - Overlapping clocks,, is this like sampling twice?
Two bites of the cherry to fill up the integrator cap with bites of signal?

I think it's more like using a single sample for two separate calculations (similar to the guts of a FFT calc), because the same sample belongs to two different sums (overlapping sums).

I guess all of this is like having an AtoD with 1 sample per 1/2 cycle or 2 samples per 1/2 cycle . In time we may not use a mix/det at all and sample the Return and the Tx directly and do the math in dsp.

Direct sampling would make it very easy to reuse some samples for the different overlapping sums, and excellent way to try different ways to process the data. However, it is hard to beat the way a capacitor integrator averages out the noise.

-SB

[golfnut]

I guess were into integrating the area under the curve. I say, when you sample affects how much signal you get.

I look at Carls sampled area resembling a cosine - and this may average to zero as it has both polarities, so may get no net gain there.

Or depending on target phase shift you could sample on a low part of the target return sinus. Not ideal


I was thinking that the time of the sample is key, to integrate the most signal energy.

Knowns,
where the Tx signal phase is in time.
where the target phases are of the signals were hoping for

You can see where this is going..

If we used a sampling clock from a state machine with predefined target phase shifts.
Say only 6 clocking predefined phase shifts to cover the main phases like nickel, silver, etc

n deg
n+1 deg
.
.
n+x deg
jmp back to beginning of list.. start over

We know as we fly round the various clocking/sampling phases in the list which one we were on when we got the largest hit - good stab at ID there.

More importantly, the main point is that you will clock/sample at the correct part of the return (of a specific target metal/phase shift) to record the peak signal from that phase - having sampled it on or near its Peak value.
(Avoiding sampling all the same way - where for some target phases its amplitude could be sampled low - due to sampling the at the wrong phase/time)


So there is, a set of predefined clocking phases to sample at the expected amplitude peak of known target phases.

Im normally 4 to 20 years behind the times with clevers, Carl am I barking up the wrong tree again?

Steve

[simonbaker]

I guess were into integrating the area under the curve. I say, when you sample affects how much signal you get.

I look at Carls sampled area resembling a cosine - and this may average to zero as it has both polarities, so may get no net gain there.

Or depending on target phase shift you could sample on a low part of the target return sinus. Not ideal


I was thinking that the time of the sample is key, to integrate the most signal energy.

Knowns,
where the Tx signal phase is in time.
where the target phases are of the signals were hoping for

You can see where this is going..

If we used a sampling clock from a state machine with predefined target phase shifts.
Say only 6 clocking predefined phase shifts to cover the main phases like nickel, silver, etc

n deg
n+1 deg
.
.
n+x deg
jmp back to beginning of list.. start over

We know as we fly round the various clocking/sampling phases in the list which one we were on when we got the largest hit - good stab at ID there.

More importantly, the main point is that you will clock/sample at the correct part of the return (of a specific target metal/phase shift) to record the peak signal from that phase - having sampled it on or near its Peak value.
(Avoiding sampling all the same way - where for some target phases its amplitude could be sampled low - due to sampling the at the wrong phase/time)


So there is, a set of predefined clocking phases to sample at the expected amplitude peak of known target phases.

Im normally 4 to 20 years behind the times with clevers, Carl am I barking up the wrong tree again?

Steve

I think I gather what you are saying -- it is similar to saying that with our TGSL, you can set the DISC pot for maximum response to nickel, or aluminum, or whatever by aligning the sync pulse smack dab over the peak of the signal corresponding to nickel, etc. However, when you do that, you are not discriminating for that metal. To discriminate everything below nickel, for example, you set the DISC pot so nickel just barely registers - at that setting, you may be optimal for copper, or some other metal.

Now when you noticed that one of Carl's quadrature integrations averaged to zero, that is OK, because you are integrating the same signal in both "channels" (I and Q), and then doing some math to figure out the amplitude and phase of the signal. Both integrations will never be zero -- one will always have a healthy response (or both will have a fair response). The fact that one averages to zero just means the phase of the signal is exactly the same as the other channel sync pulse.

To me, your idea of "sampling at the peak" (or rather centering your integration at the peak) is a darn good one for making a "gold hunter" machine, where you don't even worry about discrimination, you just want to optimize the sensitivity to gold. Generally, iron is far enough away in phase that it will probably have a very low response with that sample strategy, and in most gold areas there isn't much trash to worry about.

However, I still believe that it is probably best to always sample the whole waveform to get the best S/N -- even the off-peak samples should contribute statistically to improve the detected signal. But you can still use your idea of centering the integration window on your metal of choice, so long as you don't need to discriminate other metals than perhaps iron.

If you have limited processing power and can only take close samples in bursts, then that might also be a reason to "sample at the peak" -- but again you are losing your normal discrimination control I believe. You will discriminate, but the cut-off will be quite a ways from your metal of choice.

-SB

[Davor]

Peak is not as "pointy" as the null so there you have it. There is a way to have both maximum reading, and exact phase, and it involves Costas loop filter. To make it tick you just need I and Q outputs ... something that you already have with Tayloe

Micros are not my thing, but I guess creating a virtual Costas loop would be a right thing to do, so a bit of effort could fare very well. All other algorithms are just funny in comparison.

How it works? Imagine a circle, and two perpendicular axes I and Q. Now imagine a vector Z at any odd angle, and a new pair of perpendicular axes I' and Q' that would rotate in a way that I' is aligned with Z and Q' is perpendicular to it, so that you have both full amplitude and a precise phase reading. The thing that does it very well is called "Costas loop" and is usually designed using a PLL ... a modest challenge for virtualisation. Or not.

Anyway, it is feasible using a PLL, but I'm afraid it will perform only down to certain low level signals. Hence, my high expectations are still with DSP-like implementation. Even PLL should beat a semi-deaf TGSL.

[simonbaker]

Peak is not as "pointy" as the null so there you have it. There is a way to have both maximum reading, and exact phase, and it involves Costas loop filter. To make it tick you just need I and Q outputs ... something that you already have with Tayloe

Micros are not my thing, but I guess creating a virtual Costas loop would be a right thing to do, so a bit of effort could fare very well. All other algorithms are just funny in comparison.

How it works? Imagine a circle, and two perpendicular axes I and Q. Now imagine a vector Z at any odd angle, and a new pair of perpendicular axes I' and Q' that would rotate in a way that I' is aligned with Z and Q' is perpendicular to it, so that you have both full amplitude and a precise phase reading. The thing that does it very well is called "Costas loop" and is usually designed using a PLL ... a modest challenge for virtualisation. Or not.

Anyway, it is feasible using a PLL, but I'm afraid it will perform only down to certain low level signals. Hence, my high expectations are still with DSP-like implementation. Even PLL should beat a semi-deaf TGSL.

How would you use a PLL? What signal are you locking onto? How would you do discrimination?

Also, isn't the TGSL Synchronous Detector basically the same as the simple Tayloe in terms of S/N? They both are "half-wave" integrators rather than full-wave. The "double tayloe" design seems to be full-wave, I can see that has some advantage, but Carl's full-wave SD example should equate to that.

-SB

[Davor]

How would you use a PLL? What signal are you locking onto? How would you do discrimination?

Please see about the Costas loop on Wikipedia. In essence Costas loop rotates the IQ plane in a way that I' is in phase with the Rx signal, and all discrimination is simply derived as a phi against the original I. A PLL is used as a phase shifter, and the beauty of it is that it remains at the angle even with signal discontinuities - thanks to the PLL LPF. The main purpose of Costas loop is demodulation of various phase shifted modulations, where demodulated information is derived directly from Q'. For MD purposes it would simply be used for Phi readout.

Clever, eh?

[simonbaker]

Please see about the Costas loop on Wikipedia. In essence Costas loop rotates the IQ plane in a way that I' is in phase with the Rx signal, and all discrimination is simply derived as a phi against the original I. A PLL is used as a phase shifter, and the beauty of it is that it remains at the angle even with signal discontinuities - thanks to the PLL LPF. The main purpose of Costas loop is demodulation of various phase shifted modulations, where demodulated information is derived directly from Q'. For MD purposes it would simply be used for Phi readout.

Clever, eh?

The phase of the RX signal (and even the delta phi of the RX signal) is not at all the same as the phase of the target signal we are trying to detect, so we're not really interested in measuring the phase of the RX signal, at least for discrimination purposes. I don't know the subtleties of the Costas loop, but we want a detector that helps us detect (or manipulate) the phase of the target signal, not the RX signal which is a combination of the null signal and the smaller target signal. Just wondering how that is accomplished with those alternate methods.

-SB

[Carl-NC]

The phase of the RX signal (and even the delta phi of the RX signal) is not at all the same as the phase of the target signal we are trying to detect, so we're not really interested in measuring the phase of the RX signal, at least for discrimination purposes. I don't know the subtleties of the Costas loop, but we want a detector that helps us detect (or manipulate) the phase of the target signal, not the RX signal which is a combination of the null signal and the smaller target signal. Just wondering how that is accomplished with those alternate methods.

-SB

True, the RX phase includes null residual, ground, and target. We normally run the demod outputs through filters to take out the null residual and ground, assuming the target is a transient delta. I've never worked with a Costas loop, but it looks like it might be an interesting way to combine the demod & filter functions, whereby the VCO loop filter removes long-term null & ground phase but not transient target phase.

[Davor]

Yup. A three course meal for free.