VLF MD with digital signal processing : Bee-Buzz 1

[Skippy]

Quote Carl: "Pretty much all these designs use a 24bit ADC, so a 12-bit ADC may prove to be a little light for this task"

My initial thoughts, too. As an example, the Teknetics T2 / Fisher F75 uses a 16-bit ADC with "dithering" to get about 2 bits extra. That is after an analogue front-end, so it's converting DC - 100Hz signals. The "dithering" is pseudo-random noise injection ( generated by the micro ), I recall 3 bits are used ( ie. a random number from 0 to 7 ) and then an average over 8 samples is taken.
You really should be considering 14 bits + some "extra" [ pseudo-random noise injection / averaging etc ] as an absolute minimum.

If you are after ideas for generating a Target ID value from your phase etc data, see this post:

https://www.geotech1.com/forums/foru...654#post417654


The whole thread contains useful info, if you take a look.

[Atul Asthana]

Quote Carl: "Pretty much all these designs use a 24bit ADC, so a 12-bit ADC may prove to be a little light for this task"
The whole thread contains useful info, if you take a look.

thanks, noted. dithering could be incorporated, requires additional analog hardware.

however, I am already oversampling the signal, and to further improve snr and resolution, I can sample 8 consecutive cycles, average out each of the points (8 points) of each cycle and use this averaged out 15 bit data of 8 points to feed to goertzel algorithm.

this gives me an additional 18 dB improvement in SNR : means that the power of the noise is reduced by a factor of approximately 63 (10^(18/10)). In simpler terms, I can now detect signals that are 63 times weaker than with the 12-bit ADC, assuming the noise floor is the limiting factor.

Presently, I still want to keep the MD simple, only using the stm32 peripherals: the 12 bit adc.
will look at a 16 bit adc if this 12/15 bit doesnt work for this hobby machine.

fot the kind of tid mentioned, I need to study equinox 800 in all its essence, and I think, I can relegate that to some time later.

please opine.

[Carl-NC]

We'll sample the target signal around 512 times per second. For each sample, we'll use the Goertzel algorithm to get signal strength and phase, then smooth these values using digital filters. These smoothed values will be used to determine the target type (VDI) and generate output.

A typical loop rate in a digital metal detector is 200Hz. The higher the better until you run out of processing time. Personally, I'd start out conservatively at 200Hz. That is 5ms, and if the TX is 6kHz then you get an even 30 TX cycles in a single loop. It's not critical that the TX frequency is a whole multiple of the loop rate but it probably makes things easier. This allows you to average each of the 8 sample points 30 times for roughly a 15dB SNR improvement (not sure where you got 60dB from originally, or the 18dB above). Normally, the STM32 allows you to do background accumulation on the ADC for N samples but I think in this case you will have to do it manually because successive samples have to be binned differently. This should all be handled by DMA and a short interrupt handler so that other tasks can proceed. Which brings up the job of handling tasks. I would recommend running everything under FreeRTOS instead of trying to roll your own task handler.

Reduced stress on the H-bridge transistors.

I don't understand this. From post #1 it sounds like you want to drive the TX coil with a class D waveform, is that right?

with 100 mW transmit power,

I don't know what this means.

To simplify coil tuning, we'll implement a frequency counter mode on the STM32 using a Colpitts oscillator. This will measure the resonant frequency of each coil. Ideally, the transmit frequency will be automatically adjusted to match the measured resonant frequency (within a few Hz of 6 kHz to avoid affecting filter and impedance calculations).

The White's TRX has an H-bridge TX driver that can be run in open-loop mode (self-driven) or driven by a TX clock. When the TRX is first turned on, it runs the TX in open-loop mode and measures the natural TX frequency. Then the micro drives the H-bridge at that frequency.

[Atul Asthana]

I agree that many improvements are possible. Its really great that we are all pooling our knowledge and expertise on this project :

Averaging and SNR Improvement:

You're correct about the SNR improvement from averaging. Averaging N samples improves the SNR by 10log₁₀(N) dB. Averaging 30 cycles (as you calculated) gives an SNR improvement of 10log₁₀(30) ≈ 14.8 dB, which is close to 15 dB. The 60 dB and 18 dB figures I mentioned earlier were in the context of different oversampling scenarios (averaging 8 samples per cycle to get from 12 bit to 15 bit resolution), not averaging multiple cycles. My apologies for the confusion.

Loop Rate :

With a loop rate of 512 slots per second (approximately 2 ms per slot), using 8 samples within each slot provides good resolution and the ability to detect fast-moving targets. If necessary, the number of samples per slot can be increased to 16 for further resolution improvement and number of slots can be reduced to 256.

The ADC handling :

Since every transmitted wave is being oversampled and each sample is being integrated over 8 cycles, using dma may make it easier. An interrupt driven, software managed process will probably be a simpler approach. Both approaches can be tried out.

RTOS :

Though, I did consider this, but I am not very sure about the overheads. I may have to reduce number of cycles and samples to fit everything in the slot.

Tuning Tx Rx :

Tuning tx rx tanks (actually varying the tx frequency) on every switch on is a good idea, however, the slight amount of detuning of the tanks is acceptable. Retuning every use cycle should be considered if the tanks are likely to detune by a large amount.

in our case, the Q is not very high, and slight amount of detuning, will not make much of a difference.

• Q of TX Tank (Q_TX): Approximately 10
• Q of RX Tank (Q_RX): Approximately 5.3

that implies : a 1% drop in power for frequency shift by 30 hz tx and 57 hz rx.
for a hobby grade design, this drift can be tolerated.

Driving tx coil

The PWM signal driving the H-bridge generates the sine wave in the transmit LC tank circuit. Adjusting this PWM waveform allows for control of the transmit coil's power output.

100 mw transmitted power refers to the power with which we illuminate the target. its used as a reference just to calculate the likely voltage generated on the receive coil.

Primary objective

The primary goal remains to maximize the usage of the capabilities of the low-cost, readily available, and widely used STM32F103C8T6 to create a "somewhat" commercial-grade VLF metal detector, while maintaining hardware and software simplicity.

please comment to correct errors/misunderstandings and improve (and simplify) the design.

[Repwoc]

I am interested in this project because I have a similar project in the making, also based on a STM32 microcontroller.

The system revolves around the STM32F103C8T6 microcontroller running at 72 MHz.

Why have you decided on this chip? It is pretty low spec compared with other STM32 devices. I would say you would be much better off choosing one based on a Cortex M4 or M7 processor.rather than the Cortex M3 in the blue pill. The M3 doesn't have a math co-processor and the blue pill ADC is quite slow. Even the STM32F401CCU6 black pill has M4 processor, single precision floating point unit, much more memory (flash and SRAM), a much quicker ADC (2.4 x faster than the blue pill ADC, but still only 12 bits) and, significantly when you want to do DSP, dedicated DSP instructions using the CMSIS-DSP library. On the flip side, the black pill has only the one ADC whereas the blue pill has two and the black pill costs a little more. The STM32F411CCU6 is similar but has twice as much flash + SRAM and runs at 100 MHz. The STM32 Cortex M7 microcontrollers include a double-precision floating point unit and some have 16-bit ADCs.

I was reading that its possible to apply goertzel algorithm on just one single cycle and get good enough results,

Could you tell us where you read this? It looks more complex than a single-cycle instruction.
https://www.st.com/resource/en/desig...lectronics.pdf .

[Atul Asthana]

I am interested in this project because I have a similar project in the making, also based on a STM32 microcontroller.



Why have you decided on this chip? It is pretty low spec compared with other STM32 devices. I would say you would be much better off choosing one based on a Cortex M4 or M7 processor.rather than the Cortex M3 in the blue pill. The M3 doesn't have a math co-processor and the blue pill ADC is quite slow. Even the STM32F401CCU6 black pill has M4 processor, single precision floating point unit, much more memory (flash and SRAM), a much quicker ADC (2.4 x faster than the blue pill ADC, but still only 12 bits) and, significantly when you want to do DSP, dedicated DSP instructions using the CMSIS-DSP library. On the flip side, the black pill has only the one ADC whereas the blue pill has two and the black pill costs a little more. The STM32F411CCU6 is similar but has twice as much flash + SRAM and runs at 100 MHz. The STM32 Cortex M7 microcontrollers include a double-precision floating point unit and some have 16-bit ADCs.



Could you tell us where you read this? It looks more complex than a single-cycle instruction.
https://www.st.com/resource/en/desig...lectronics.pdf .

oh, its single cycle of sine wave, not single instruction cycle.
goertzel algorithm takes quite a few instruction cycles.

Processor :
bluepill is low priced, easily available in India unlike other variants and will do the needfull in this case, by not aiming to include all possible functionalities.
Alternatively other low priced boards could be looked at based on availability and price.

Other stm32 boards are high priced and slightly difficult to procure, making it a hard task for a hobbyist.

[Repwoc]

You can't buy here?

https://www.aliexpress.us/item/1005001456186625.html

[Atul Asthana]

You can't buy here?

https://www.aliexpress.us/item/1005001456186625.html

not in India
most of the chinese business sites are banned in India.
and so are most of the imports.

[Repwoc]

Wow. That must really hinder commercial and industrial innovation.

[Carl-NC]

The 60 dB and 18 dB figures I mentioned earlier were in the context of different oversampling scenarios (averaging 8 samples per cycle to get from 12 bit to 15 bit resolution), not averaging multiple cycles.

You can't average the 8 samples in the TX cycle together, they have to be binned separately. You can only average each one over multiple cycles.

Since every transmitted wave is being oversampled and each sample is being integrated over 8 cycles, using dma may make it easier. An interrupt driven, software managed process will probably be a simpler approach. Both approaches can be tried out.

DMA will make it faster. Since you have to bin 8 separate sample points you will have to retrieve those samples and bin them manually. An interrupt procedure is about the only way that makes sense. You sure don't want this in a polling loop.

RTOS : Though, I did consider this, but I am not very sure about the overheads. I may have to reduce number of cycles and samples to fit everything in the slot.

Whether you use RTOS or roll your own task manager, there is overhead. RTOS is nice in that is has truly autonomous tasks and uses time slicing with priority levels, which makes task management simple.

100 mw transmitted power refers to the power with which we illuminate the target. its used as a reference just to calculate the likely voltage generated on the receive coil.

Metal detectors don't transmit power, just a magnetic field. The TX coil is a reactive element so if the resistance is zero, then there is zero power loss no matter how big the signal is. The TX energy that went into the coil gets completely recycled via the tank capacitor. So it makes no sense to say you want 100mW of TX power.

[Atul Asthana]

yes, you are right,
The entire process will have to be interrupt driven, since like you say, all the sampled points have to be binned separately.

in writing my own task manager, I will know the overhead and will be able to synchronise everything. with rtos, I dont know the load and the synchronicity.


From what I know,
Unless a certain amount of power is transmitted, no current will be induced in the target or its surrounding medium.

In fact, in an md, most of the transmitted power is wasted, being absorbed by the surrounding medium. This is why detecting deeper targets is challenging—most of the transmitted energy and the target-generated response is dissipated in the medium.

The surrounding medium is typically resistive, with minimal phase shift—this constitutes the ambient response. When modeling the target and its surrounding medium, it can be represented as a combination of an inductor, capacitor, and resistor. For this combination to affect the receiving coil, it must generate energy, which originates from the magnetic field created by the transmission.

In its simplified form, The system can be understood as a transformer with a very lossy core, the transmission (Tx) coil acts as the primary winding, the medium and target function as the core, and the receiving (Rx) coil serves as the secondary winding. In this analogy, the Tx coil is the source, and the Rx coil is the load, where in all of the transformer principles apply.

[ actually knowing the effect of interaction of the generated alternating mag field (AMF) with the soil + target is of value, forming the essence of the generic theory of detection.
The modeling of an alternating magnetic field effect on a buried metal piece involves analyzing electromagnetic induction and eddy currents within the metal. The AMF induces currents in the metal, causing energy dissipation as heat and generating secondary magnetic fields. Soil conductivity and permeability influence the field's penetration and attenuation. Simultaneously, the AMF generation coil experiences energy loss due to resistive heating and coupling inefficiencies with the buried object. The model requires solving Maxwell's equations, accounting for soil and metal properties, coil geometry, and frequency. ]

however, this is a moot point. the fact is that the target produces a response, and thats what is of interest

[Marchel]

Hi

I am also working on a metal detector with STM32, initially I used STM32H7A3 processor in my project which has 16bit ADC but then I upgraded to STM32L476 which has faster ADC and I didn't notice much difference between 12 and 16bit ADC. In my metal detector I do 16 samples and the resampling is 64 times.

https://www.st.com/en/microcontrolle...2h7a3-7b3.html

https://www.youtube.com/watch?v=nUejeLYKmfw&t=729s

[ivconic]

I have a few blue pills in the pile, they have been unused for a long time.
The technology at that moment when I got them; developed too fast, there was no time, I immediately came to a couple of ESP32 modules and immediately after that to a couple of Nucleo 144.
So I don't know where I will go first and what I will do first!
How to choose the right development platform?
None of the processors I have, despite their great features, have a good enough ADC.
And that is a stumbling block when I think about using such a processor for these purposes.
However, I will always welcome any initiative based on one of the processors I already have.
Because they are sitting in a pile unused anyway.

[Atul Asthana]

the ADC debate :

I have a feeling that the audio codecs are being used because they are mass produced, cheap, very well documented and are available a plenty, ond not because this grade of vlf md needs such a high resolution and such a large dynamic range.

the dynamic range of the response from the target, and thats what I was trying to calculate with 100 mW of power, remains not better than 1 micro volt to 4 mV (for a very large target, physically visible on the surface) as response in the receiving coil, this is around 70 dB, thats 12 bits. Even if you were to consider an error of 4, that makes it 14 bits.

Actually, careful design and alignment of the coils + amp in a commercial detector, still has a noise floor of around 1 microvolt or above. provided the opamp is very low noise and adds minimum of its own noise, the ground is very quite and temperatures are low + there are no electrical power lines nearby.

This low noise floor is crucial for detecting weak target signals, enabling greater detection depth, smaller target detection, and improved target discrimination. Factors influencing noise include detector design, operating frequency, ground conditions, environmental EMI, and coil design. Minimizing noise is a key design consideration for maximizing detector sensitivity and performance.

So, there isnt much to gain by increasing the dynamic range of the digitisation, when the signal itself has, at best, a dynamic range of 12-14 bits above the noise floor.

Kindly opine.

[Carl-NC]

The first digitizing detectors placed the ADC after the demods, so that it sampled a baseband signal which had additional gain besides just the preamp. All of the designs I'm aware of used an ADC with an ENOB of 14-16 bits. When you direct sample, you need an ADC with even higher precision because you no longer have the demod and extra gain stage. You can do it with a 12b ADC but you will either lose faint targets or you will overload on both strong and moderate targets. If you want good depth, then you use extra preamp gain and overload on more targets. But that also means you will overload on bad ground. I don't know of a single modern direct sampling design that is not using a 24b ADC.

Personally, I think it's a good idea to start with the micro's 12b ADC. It's much easier to work with. But don't expect it to do much more than demonstrate feasibility. It's probably not gonna make a good field hunter.