AMX Digital

[waltr]

I have been also thinking Discrimination is highly desired. Lack of Disc in a PI is the biggest reason I mostly hunt with a TGLS.

Could the VooDoo type of Disc be added into a PI - where detector runs in pure PI mode until a target is detected then switches periodically into Disc mode for iron reject?

at 96Ksps (not 96sps)

Yep, that was a typo. Thanks for correction.

[ivconic]

Why do you say that? I think discrimination should be on the table as it can be critical in some places. Most of the places I hunt are littered with iron flecks from old rotted cans. To a PI without discrimination they sound just like a nugget. In Africa some of the places they hunt are littered with iron-age artifacts such as arrow heads. While you & I might think they are cool, the Africans don't want to dig a 1 meter hole to pull out a spear point.

Without discrimination, you probably want 2 channels per TX pulse width: one for target and one for ground. So for 2 pulse widths you need 4 channels. But when running even just 2 pulse widths there may be other samples you want to take. This is why I favor direct sampling initially, it removes constraints. If I were to start with analog demods I would automatically assume 8 channels, because I would want to do some R&D.

But how realistic is it possible to distinguish pieces of rusty iron from pieces of native gold?
In most cases these will be very small pieces of gold.
Such discrimination is very difficult even with VLF I/B, let alone PI detectors.
I watched on one of the "History...National.." channels a series about Australian native gold prospectors.
In the sequels.
Most use, of course, ML.
The vast majority do not use discrimination, either they turn it off or the detector itself does not have it at all.
But over time, more experienced searchers trained their ears to recognize subtle variations in audio response.
I have to admit that I was not able to recognize those variations.
Probably because I don't use the ML detector.
But the point is that in such conditions the discrimination that does not work perfectly correctly (with PI I do not know one that works); it is not even desirable that it exists.
But ok, the rest of your post explains the need for more channels on an ADC.
Which directly disclassifies all these ADCs I mentioned, with only 2 channels.
And not only that; it also disclassifies a good number of processors I had in mind.
Obviously this is a job for the very "heavy artillery"...
And it won't be possible to do it in the home version the way we've done it so far.
The only possible way is to design modules with stages, order the production of complete modules in China,
at JLPcb or PCBWay and later assemble and work with the modules at home.
About the same as you already did from those pictures you posted.
...
The eventual problem that I mentioned on another topic remains.
What if JLPcb and PCBWay can't source all the "delicate" material from the BOM?
Or they can from foreign suppliers, but at several times higher prices.

[Carl-NC]

All the elements of this project are not 100% known, there is some R&D to be done even after the architecture is defined and circuitry designed. This includes best methods of ground balance and possible methods of iron disc. So what you do is design for the R&D of those possibilities. Starting out with just a 2-channel design ties your hands immediately, and you are likely confined to a TDI-like detector.

In ML GPX detectors, iron disc works when using DD coils and looking at the TX turn-on reactive signal. In CCPI that signal is only a few microseconds wide and is accompanied by several hundred volts of flyback (or flyforward?), so using it is a real challenge. But there is also the possibility of looking at the decay in more detail. Just like a VRM response has a different decay curve than an eddy decay, so does an iron target. To look at this you need at least 3 samples. If you direct sample the decays, well, then you have a ton of information to use, and iron disc might not be so hard.

BTW, iron disc with nails is relatively easy because it's primarily a BH-curve response. Flat iron like bottle caps or the flakes of rotted cans are more difficult because they have substantial eddies. Which is why I want more information.

[ivconic]

MPU3914 gives 24 bits at 125KHz WHEN CLOCKED from 16MHz with integrated MTPX of 8 channels
CS5343 gives 24 bits at 96Ksps (not 96sps) with single input

The two ADC's come at about the same price on MOUSER.
These ADC's could only be used with an analog integrated sampling. They are not fast enough for the direct sampling technique.

Is the idea to do a simple Sample during the coil/target decay?
It depends what you mean by 'simple sample'
The integrated sampling reads the output of each demodulator and optionally makes a supplementary digital integration. before using the net result in DSP.
e.g. An ADC giving 125Ksps (real data stream, possibly with DMA support) could be digitally integrated 40x to give a net result every 5msec.

sample at 500kHz during the decay to obtain the shape of the decay curve.
This is direct sampling, that supposes a DMA support and a completely different type of processing.
ADC's of 24 bits with that type of data rate or more are still difficult to find in stock at the usual reliable suppliers.

What/ When is the ADC taking samples?
This is something still to be discussed later but it is indeed an important question.

Everything is very interesting to me, but I don't really understand much of it.
I'm sorry if I rambled on about some nonsense, PI is not my field, I'm here as a guest who is just getting acquainted with this field.
I don't quite understand the "math" behind the need to sample at the highest possible sampling rate as 500ksps.
And that 96ksps is not enough to reproduce such a signal.
Let me try to do some math and you correct me if I am wrong.
96ksps = 96 000 samples per second.
Respecting Tony's proposal: 25kpps = 25 000 pulse per second = 25 000 decay curves.
96/25 = 3.84 samples per decay curve.
500/25 = 20 samples per decay curve.
I don't know why Tony insists on 25kpps?
If we reduce the pps to 5kpps, let's see...
96/5 = 19.2 samples per decay curve.
500/5 = 100 samples per decay curve.
How many samples do we realistically need for one decay curve?
Isn't it unnecessary to sample the whole decay?
I didn't follow the threads about it, but how much was the TC for the average gold nugget?
But I remember from the occasion that we used to talk about the need to analyze the decay curve only after it falls to a certain value.
Unless you want to reproduce the complete decay curve...
Then even 500ksps will not be enough for 25kpps.
I need someone to explain all this to me, I don't understand anything...

[ivconic]

All the elements of this project are not 100% known, there is some R&D to be done even after the architecture is defined and circuitry designed. This includes best methods of ground balance and possible methods of iron disc. So what you do is design for the R&D of those possibilities. Starting out with just a 2-channel design ties your hands immediately, and you are likely confined to a TDI-like detector.

In ML GPX detectors, iron disc works when using DD coils and looking at the TX turn-on reactive signal. In CCPI that signal is only a few microseconds wide and is accompanied by several hundred volts of flyback (or flyforward?), so using it is a real challenge. But there is also the possibility of looking at the decay in more detail. Just like a VRM response has a different decay curve than an eddy decay, so does an iron target. To look at this you need at least 3 samples. If you direct sample the decays, well, then you have a ton of information to use, and iron disc might not be so hard.

BTW, iron disc with nails is relatively easy because it's primarily a BH-curve response. Flat iron like bottle caps or the flakes of rotted cans are more difficult because they have substantial eddies. Which is why I want more information.

It looks like we were writing at the same time so you already answered some of the questions I wrote to Willy.
But there is still mathematics that needs to be clarified, that is, the question; how many samples do we really need per decay curve?
You:
"... If you direct sample the decays, well, then you have a ton of information to use, and iron disc might not be so hard..."

Exactly! Looks like my "math" isn't that bad. Because it shows that a high sampling rate on such pps will give a literally "ton" of data.
Ok, now let's see the further process that will happen.
A "ton" of data must be processed in real time (slightly slower in reality, but fast enough for the user not to notice).
What is the DSP in question? Are there any filters? What kind? How much useful information do we expect from a "ton" of samples... in a unit of time?
What processor will manage to do it all? At the same time, let's keep in mind that this is not the only task that the processor has to perform.
The more details we reveal here; it becomes clearer that the task is anything but easy and naive.
And that it will greatly narrow down the choice of processor as well as ADC.
This is all about Tony's 25kpps proposal.
Why 25kpps? Does it have to be like that?
What is the lowest pps at which we can still successfully detect the smallest gold nuggets, with all the additional notes we have already
established here about possible hardware?
Do we know the pps and sps at already existing machines? Say GPX5000, GPZ 7000?
(I really wasn't interested until now, so I didn't do any research.)
...
Instead of sampling the entire decay curve; is it possible to design a moving, adjustable "window" with fewer samples?
This would greatly simplify the work and reduce the load on the processor. But it also expands the choice of ADCs.
Sampling the entire decay is a "brute force" method that will yield at least 50% of unnecessary information.
Maybe more than that.
Can it be done in a more intelligent way, at the very beginning before DSP.

[Willy Bayot]

I shall try to answer the question of IVCONIC.
I am sure that some readers will find approximations in the following description but I have intentionally simplified the real process to make it clearer.


Let's take a pulse period of 40µsec(25Kpps).
(Why 25Kpps? That is to improve the SNR, more samples per second = more samples per unit of swing. A swing at 1m/s moves the coil 1mm/msec.
If you want to catch a small nugget, you need at least one NET SAMPLE per 10mm or 10msec. At 25Kpps, we get 250 raw samples for 10mm, thus, we can apply a digital integration of 250x)

From that, we should remove the time of the pulse itself and a bit more to take into account the damping delay necessary to remove all the oscillations due to the flyback.
We are left with say, 30µsec of exponential signal decay.
A small gold nugget has a TC of 1 µsec. This means that the signal has decayed to 37% after 1µs, 13% after 2µs, 5% after 3µsec,1% after 5µs.
Using an ADC which really captures data at 500sps means that you get a sample every 2µsec.
Over the 30µsec of decay, you can get 15 samples. That is already very short to extract any info from the decay.(1% left of the signal after 5µs!!!)
Using an ADC which really captures data at 96Ksps means that you get a sample every 10.5µsec.
Over the 30µsec of decay, you can get 3 samples.
What can you do with that? The whole decay is completely embedded in the first sample. The only thing you could extract from that is the fact that there was a TARGET of any type.

Let's take a pulse period of 200µsec(5Kpps) (At 5Kpps, we get 50 raw samples for 10mm, thus, we can only apply a digital integration of 50x)
We have 190µsec of exponential signal decay.
Over the 190µsec of decay, you can catch 95 samples. That is more reasonable BUT the only important part of the decay is still located in the first 5µsec. This will not be enough to make a difference between a gold nugget and a piece of iron junk.

This was with DIRECT SAMPLING.

For INTEGRATED SAMPLING, the calculations are quite different.
What we capture are the results of the analog integration of WINDOW periods defined by analog switches.
Since the integrators keep a memory between consecutive pulse periods, the ADC captures can be made slower without loosing info.
Thus, an ADC at around 100Ksps could still be used.

[ivconic]

My understanding of all this is very weak and wrong in many points.
But I start from the thinking of audio sampling.
A good audio ADC must faithfully reproduce audio signals in the 0 to 22kHz range.
That in theory.
In practice, few of us can hear above 14kHz.
I personally don't hear anything above 8-9kz. Because I'm hearing impaired.
But things in the industry are not made according to my ear, but according to the standards that have been adopted.
So the audio ADC is capable of reproducing an analog signal in the range from 0 to 22kHz.
According to the Nyquist theorem, the smallest sampling rate in this case is 44kHz.
I'll repeat, at the risk of being boring... WM8738, here are some specs for that ADC.
- 44.1K sampling rate, 24bit AD sampling to achieve nonderucve high-fidelity digital audio transmission,
- frequency range: 20Hz-20kHz
...
No, I do not suggest using the WM8738 ADC again. It is already clear to me that it does not suit this task and why it does not suit this task.
But I worked with it, the audio I get on the other side, on the DAC (WM8501) is fantastic quality!
(not because I'm half-deaf and can't hear, but because it's a fact that can be measured).
The delay, latency, is insignificant, as I already mentioned somewhere; it is not noticeable even when playing the guitar fast.
In "real life" there is no delay, latency. In the process itself in the chip, of course, there are delays, latency.
But that is not the topic or the point here.
The point is the following: is there more "data" in the sampled audio signal with a frequency range of 20Hz to 20kHz or is there more data in the decay curve of the PI detector?
When you clarify this for me and answer; many other things will become much clearer to me too.

[ivconic]

I shall try to answer the question of IVCONIC.
I am sure that some readers will find approximations in the following description but I have intentionally simplified the real process to make it clearer.


Let's take a pulse period of 40µsec(25Kpps).
(Why 25Kpps? That is to improve the SNR, more samples per second = more samples per unit of swing. A swing at 1m/s moves the coil 1mm/msec.
If you want to catch a small nugget, you need at least one NET SAMPLE per 10mm or 10msec. At 25Kpps, we get 250 raw samples for 10mm, thus, we can apply a digital integration of 250x)

From that, we should remove the time of the pulse itself and a bit more to take into account the damping delay necessary to remove all the oscillations due to the flyback.
We are left with say, 30µsec of exponential signal decay.
A small gold nugget has a TC of 1 µsec. This means that the signal has decayed to 37% after 1µs, 13% after 2µs, 5% after 3µsec,1% after 5µs.
Using an ADC which really captures data at 500sps means that you get a sample every 2µsec.
Over the 30µsec of decay, you can get 15 samples. That is already very short to extract any info from the decay.(1% left of the signal after 5µs!!!)
Using an ADC which really captures data at 96Ksps means that you get a sample every 10.5µsec.
Over the 30µsec of decay, you can get 3 samples.
What can you do with that? The whole decay is completely embedded in the first sample. The only thing you could extract from that is the fact that there was a TARGET of any type.

Let's take a pulse period of 200µsec(5Kpps) (At 5Kpps, we get 50 raw samples for 10mm, thus, we can only apply a digital integration of 50x)
We have 190µsec of exponential signal decay.
Over the 190µsec of decay, you can catch 95 samples. That is more reasonable BUT the only important part of the decay is still located in the first 5µsec. This will not be enough to make a difference between a gold nugget and a piece of iron junk.

This was with DIRECT SAMPLING.

For INTEGRATED SAMPLING, the calculations are quite different.
What we capture are the results of the analog integration of WINDOW periods defined by analog switches.
Since the integrators keep a memory between consecutive pulse periods, the ADC captures can be made slower without loosing info.
Thus, an ADC at around 100Ksps could still be used.

We wrote again at the same time so I didn't see your post. But THANK YOU SO MUCH for such a clear explanation! Now some things are becoming clearer to me.
One of the questions that arises from your explanation is; is it a problem if the coil swing is slower than stated?
If we call that detector a "slow scan" PI detector?
Do you agree that it will soften the otherwise harsh reality of facts that you so truly noted?
I hope we can agree that when searching for the tiniest nuggets of gold; you need to move the coil really slowly.

[ivconic]

I shall try to answer the question of IVCONIC.
I am sure that some readers will find approximations in the following description but I have intentionally simplified the real process to make it clearer.


Let's take a pulse period of 40µsec(25Kpps).
(Why 25Kpps? That is to improve the SNR, more samples per second = more samples per unit of swing. A swing at 1m/s moves the coil 1mm/msec.
If you want to catch a small nugget, you need at least one NET SAMPLE per 10mm or 10msec. At 25Kpps, we get 250 raw samples for 10mm, thus, we can apply a digital integration of 250x)

From that, we should remove the time of the pulse itself and a bit more to take into account the damping delay necessary to remove all the oscillations due to the flyback.
We are left with say, 30µsec of exponential signal decay.
A small gold nugget has a TC of 1 µsec. This means that the signal has decayed to 37% after 1µs, 13% after 2µs, 5% after 3µsec,1% after 5µs.
Using an ADC which really captures data at 500sps means that you get a sample every 2µsec.
Over the 30µsec of decay, you can get 15 samples. That is already very short to extract any info from the decay.(1% left of the signal after 5µs!!!)
Using an ADC which really captures data at 96Ksps means that you get a sample every 10.5µsec.
Over the 30µsec of decay, you can get 3 samples.
What can you do with that? The whole decay is completely embedded in the first sample. The only thing you could extract from that is the fact that there was a TARGET of any type.

Let's take a pulse period of 200µsec(5Kpps) (At 5Kpps, we get 50 raw samples for 10mm, thus, we can only apply a digital integration of 50x)
We have 190µsec of exponential signal decay.
Over the 190µsec of decay, you can catch 95 samples. That is more reasonable BUT the only important part of the decay is still located in the first 5µsec. This will not be enough to make a difference between a gold nugget and a piece of iron junk.

This was with DIRECT SAMPLING.

For INTEGRATED SAMPLING, the calculations are quite different.
What we capture are the results of the analog integration of WINDOW periods defined by analog switches.
Since the integrators keep a memory between consecutive pulse periods, the ADC captures can be made slower without loosing info.
Thus, an ADC at around 100Ksps could still be used.

I had to read this over and over again. And I started to understand little by little (hopefully).
And of course you are right! It's my big mistake that I even looked for any analogy between audio sampling and this task we have here.
One has nothing to do with the other! Hahaha!

But I am all the more worried. Because you're right.
We actually need a ADC speed of a few msps here in order to make a significant step in the development of PI detectors.
Which one is it? How to find it? Are there any and at what price? With of course more channels and at least with a resolution of 16 bits if not 24 bits.
Things became clearer to me.
Willy, thanks for your patience!
The way you explained things is good because as a layman I am starting to see the "big picture" of the whole story.
And it is clear that no significant progress will be made with this project if both ADC and processor are not with high-end features.
This is very far from the domain of the ordinary hobbyist.
And it is clear why direct sampling is no longer an option here.
I suggest that we further narrow down the story to integrated sampling and further develop the method that will give the best results.
...
I have always ignored long topics with dozens of pages, dealing with PI technology.
I have always wondered; what do these characters have so much to write there?
PI is a simple and stupid device. Why is there so much to write about it?
Now it becomes clear to me.
And now it's clear to me why Carl is so "dazzled" by PI technology in the long term.
This is really a big challenge!

[Carl-NC]

Ivica, you're asking all the right questions. I'm a Big Fan of illustrations so let's dive in (again, all this belongs in the Concept discussion, but we'll go with the flow). Here is a multi-pulse TX using 2 rates:


I used 25kHz because Tony suggested it, and 5kHz because it gives me a familiar 100us TX width. If you transmit only 25kHz then you are stuck with trying to squeeze a GB sample in that 20us window which is tough. Plus, your GB has a dreaded target hole. In this example I used (5) pulses of 25k per (1) pulse of 5k; in reality we might instead use 4:1 or 8:1 just to make scaling in DSP a little easier. You can also change the ratio to implement variable frequency weighting.

Let's suppose we use analog demods. Here is my first guess as to what to try:



Boy, I love being able to just <paste> these images! So I have 4 RX channels: (1) for 25kHz and (3) for 5kHz. Maybe we could squeeze in another sample for the 25kHz pulse but I doubt it would do any good. GB is achieved with a combination of RX25, RX5a, and probably RX5b. Discrimination might be achievable with RX5a,RX5b,and RX5c.

The clocks above are demod clocks. The demods are integrating the results so the ADC does not need to run this fast. You can sample the demods every 400us (2.5kHz) or even slower if you like. Let's say we sample at 2.5kHz. The ADC creates a DataReady pulse which triggers a micro interrupt. The micro reads the ADC and DMA's the result. So far this takes nearly zero overhead. But you don't want to run a processing loop at 400us unless you are using a really fast micro. So instead, you pile up the samples (either in a circular buffer or by just summing the data) and say, every 5ms, you send all the data to the DSP task. (Most metal detectors run a processing loop of 5ms.) A processing loop of 5ms on a micro running at 32MHz gives you 160,000 single-cycle instructions to work with.

So now you have a hard example of an analog demod solution: 4 demod channels, a 16-18 bit 4-channel simultaneous sampling ADC running at 2.5kHz, and not-too-aggressive micro.

Let's say we want to use direct sampling on the above TX waveform. I will take a guess and say I want 4 samples for every 20us pulse. This now sets my ADC sample rate at 200kHz, and it needs to be 24 bits. You will be hard-pressed to grab data that fast in the same micro as the one doing the DSP. So maybe you need a data capture micro that simply sums up the data and eventually feeds it to the DSP processor at the 5ms loop rate. Or maybe you need a really fast M7 micro. Or maybe I only get data at 100kHz or, worse, 50kHz. Note that you probably don't need 20 samples per 5kHz pulse, but it's way easier to run the ADC at a constant rate and just toss unwanted samples.

Before I hit the Post button, notice that the TX waveform has a positive 100us pulse every 400us. The TDI/Goldscan has a 100us pulse every 320us. So just the response from the single 100us positive pulse should be somewhat comparable to the TDI. But we also get a second (negative) 100us pulse, and 10 (!) 20us pulses in the same time frame. So this thing should hands-down beat a TDI on small gold.

[ivconic]

Ivica, you're asking all the right questions. I'm a Big Fan of illustrations so let's dive in (again, all this belongs in the Concept discussion, but we'll go with the flow). Here is a multi-pulse TX using 2 rates:


I used 25kHz because Tony suggested it, and 5kHz because it gives me a familiar 100us TX width. If you transmit only 25kHz then you are stuck with trying to squeeze a GB sample in that 20us window which is tough. Plus, your GB has a dreaded target hole. In this example I used (5) pulses of 25k per (1) pulse of 5k; in reality we might instead use 4:1 or 8:1 just to make scaling in DSP a little easier. You can also change the ratio to implement variable frequency weighting.

Let's suppose we use analog demods. Here is my first guess as to what to try:



Boy, I love being able to just <paste> these images! So I have 4 RX channels: (1) for 25kHz and (3) for 5kHz. Maybe we could squeeze in another sample for the 25kHz pulse but I doubt it would do any good. GB is achieved with a combination of RX25, RX5a, and probably RX5b. Discrimination might be achievable with RX5a,RX5b,and RX5c.

The clocks above are demod clocks. The demods are integrating the results so the ADC does not need to run this fast. You can sample the demods every 400us (2.5kHz) or even slower if you like. Let's say we sample at 2.5kHz. The ADC creates a DataReady pulse which triggers a micro interrupt. The micro reads the ADC and DMA's the result. So far this takes nearly zero overhead. But you don't want to run a processing loop at 400us unless you are using a really fast micro. So instead, you pile up the samples (either in a circular buffer or by just summing the data) and say, every 5ms, you send all the data to the DSP task. (Most metal detectors run a processing loop of 5ms.) A processing loop of 5ms on a micro running at 32MHz gives you 160,000 single-cycle instructions to work with.

So now you have a hard example of an analog demod solution: 4 demod channels, a 16-18 bit 4-channel simultaneous sampling ADC running at 2.5kHz, and not-too-aggressive micro.

Let's say we want to use direct sampling on the above TX waveform. I will take a guess and say I want 4 samples for every 20us pulse. This now sets my ADC sample rate at 200kHz, and it needs to be 24 bits. You will be hard-pressed to grab data that fast in the same micro as the one doing the DSP. So maybe you need a data capture micro that simply sums up the data and eventually feeds it to the DSP processor at the 5ms loop rate. Or maybe you need a really fast M7 micro. Or maybe I only get data at 100kHz or, worse, 50kHz. Note that you probably don't need 20 samples per 5kHz pulse, but it's way easier to run the ADC at a constant rate and just toss unwanted samples.

Before I hit the Post button, notice that the TX waveform has a positive 100us pulse every 400us. The TDI/Goldscan has a 100us pulse every 320us. So just the response from the single 100us positive pulse should be somewhat comparable to the TDI. But we also get a second (negative) 100us pulse, and 10 (!) 20us pulses in the same time frame. So this thing should hands-down beat a TDI on small gold.

That graph clearly demonstrates the need for a multi-channel ADC.
Analog demod solution looks so better choice now.
Especially after the daunting demands of direct sampling.
Thanks to Your and Willy's goodwill; in a couple of posts I learned something that I have not been able to understand for years.


It will be an interesting "math" to take those samples later. An "block" algorithm for further processing should be drawn/written. But in "folk words". To make it easier to understand.

[Tinkerer]

My understanding of all this is very weak and wrong in many points.
But I start from the thinking of audio sampling.
A good audio ADC must faithfully reproduce audio signals in the 0 to 22kHz range.
That in theory.
In practice, few of us can hear above 14kHz.
I personally don't hear anything above 8-9kz. Because I'm hearing impaired.
But things in the industry are not made according to my ear, but according to the standards that have been adopted.
So the audio ADC is capable of reproducing an analog signal in the range from 0 to 22kHz.
According to the Nyquist theorem, the smallest sampling rate in this case is 44kHz.
I'll repeat, at the risk of being boring... WM8738, here are some specs for that ADC.
- 44.1K sampling rate, 24bit AD sampling to achieve nonderucve high-fidelity digital audio transmission,
- frequency range: 20Hz-20kHz
...
No, I do not suggest using the WM8738 ADC again. It is already clear to me that it does not suit this task and why it does not suit this task.
But I worked with it, the audio I get on the other side, on the DAC (WM8501) is fantastic quality!
(not because I'm half-deaf and can't hear, but because it's a fact that can be measured).
The delay, latency, is insignificant, as I already mentioned somewhere; it is not noticeable even when playing the guitar fast.
In "real life" there is no delay, latency. In the process itself in the chip, of course, there are delays, latency.
But that is not the topic or the point here.
The point is the following: is there more "data" in the sampled audio signal with a frequency range of 20Hz to 20kHz or is there more data in the decay curve of the PI detector?
When you clarify this for me and answer; many other things will become much clearer to me too.

The coil swing speed:
The only coil swing speed standard I could find is the one used by the military mine detectors. To be able the compare detectors from many manufacturers around the world, they set that standard of 1m/s.
For nugget detectors, we can talk about fast and slow sweep speed. So what is slow? What is fast? We could say slow is less than the standard 1m/s and fast is more than 1m/s. For comparison, a rattlesnake strikes at a speed of about 3m/s.
The smaller the coil, the faster the sweep.
Large diameter coils are swept slower because they are more cumbersome and heavier and to integrate more target samples.
Often you see people sweeping very slow. Why? Because with a slow sweep you get the integration of more samples.
A much simplified example:
If a small nugget gives you a response of 10uV, integrating 10 samples gives you about 100uV. Integrating 500 samples gives you about 5mV.
If you have a noise floor level of 3mV, you might just be able to hear the 5mV target.

When people talk about a "slow" PI, they mean that it has difficulties integrating enough samples, for example if it runs on only 1000 pulses per second. So the sweep has to be slow to increase sensitivity.
If the delay of the target indication is too long, the coil is already past the target when you hear the sound. This is also called a slow detector.

[Olly]

For direct sampling perhaps an FPGA would provide the required horsepower, especially if coupled with a 24-bit, 2.5Ms/Sec ADC such as the AD7760. This is a combination I’ve successfully used in the past for very low noise current profile measurements.

[Willy Bayot]

An illustration of the challenges we have to solve with PI detection.

Those are REAL decays which have already been amplified a lot, normalized and filtered.

Do not get me wrong, they are not the shapes of the decaying signals, they are the difference between the ground signal and the target signal.

As Carl has rightly said, there are a lot of info in the decays as you can guess by just looking at them. You can see the various TC (GOLD = shortest TC).

BUT:

1. You have first to capture them (all or useful parts of them) with as little noise as possible and as much magnitude as possible (Note that the end of the decays are mostly lost in the noise)
2. You have to decipher them using DSP and extract the useful info
3. You must have the time to do the DSP
4. You need to define how to report the results back to the user

[Willy Bayot]

For direct sampling perhaps an FPGA would provide the required horsepower, especially if coupled with a 24-bit, 2.5Ms/Sec ADC such as the AD7760. This is a combination I’ve successfully used in the past for very low noise current profile measurements.

Yes, available, rather costly but you only need one with direct sampling
BUT
data interface : 16 parallel pins!!!

Would need to interface with a dedicated CPU (FPGA is out of question here) external to the DSP processor.