Induction Balance Stuff - Single/Multi Frequency Response, GB, Disc, Measurements, Ideas, Fun, etc.

[Aziz]

Hi all,

after using the Complex FIR-bandpass filter, with very narrow and sharp bandpass (60 dB Kaiser) for the Franklin LC oscillator, a lot of noise is gone.
The sensitivity has got a big boost.
The Franklin LC oscillator is one of the stablest LC oscillators you can get. It even outperforms most crystal oscillators regarding the stability.
Frequency change (phase noise) is less than 0.1 Hz (´+/- 11 ppm). The frequency drift is dependent on temperature changes and power supply voltage only.

Wow, what a nice LC oscillator!

We should focus on digital filtering to get most out of the system.

[moodz]

CC blows most other offerings out of the water .. it bases its answers on planning and mathematical certaintly .. not only developing simple blocks of code which is ho-hum nowadays. It runs the complete project.

All I say is something like "Build code to make ADC1 sample at its highest sample rate, run an IQ demod and display the results on a python app connected to the USB bus of this WEACT STM32H750 board"

... and it just does it .. no errors. Builds the code, finds and runs the compilers, flashes the hardware ( if its connected ), then launches the programs. Its like 6 weeks of research and coding all done in 10 minutes.

Not only that it writes proper DMA driven code with correct buffering and offers to add features like FFT and DSP functions. ... all that time I used to waste trying to get DMA to work properly

No GUIs ... all done from the CLI.

[Aziz]

Hi Paul,

thanks very much.
Aziz

[moodz]

...also I should mention. I had lots of chats with Claude about different aspects of metal detecting ( PI, VLF etc ) ... it uses the chat history with you to "personalise" the focus and context of the response.

For instance ... you can ask it to create a simulation for ltspice and it will with high complexity ( the layout might need arranging ). I wanted to sim a particular part.. CC reads the data sheet for that part and creates the ltspice symbol and lib for that part. You might have to tweak some behaviours .. but you dont do it manually .. just ask CC to fix it... simple.

[Aziz]

Hi Paul,

I have tried the web-based CC. But I have to test the C-Code generation yet. The output is really impressive.
Thanks for your help.
Aziz

[Aziz]

Hi all,

the free Claude Code time limit arrived quickly!
I may pause until 15 pm local time.
Anyway.

The "damped ringing PI" project seems to be possible soon.
I expect a lot of sensitivity boost in IB-Coil configuration (TX, RX). In radically reduction of noise using the Complex FIR bandpass filter. Note, that we have both TX as reference and RX as signal, which give us a lot of info about ground and target response. Frequency (Phase) and Envelope analysis in the ringing period included. Even adding an Hilbert-Spectrum to the Hilbert-Transform.

Instead of wasting TX pulse power (we don't want to burn it in a damping resistor), we let the power ringing in the TX coil until the losses eat the complete transmitted energy. Best done in high frequency ringing. In my case up to 48 kHz as the upper frequency regions is quite noisy on my G6 sound card.
BTW, we don't need much TX power. 500 mA TX-coil current should even be enough. Short TX pulses. Flyback voltages up to 200 V (depending on the mosfet).

Aziz

[moodz]

Aziz I sent you guest pass to CC in messages... click on the link.


Key Details About Claude Guest Passes:

• • Duration: Each pass grants 7 days of Claude Pro access.
  • Access: They include full access to the Claude Code tool.
  • Eligibility: Recipients must not have a previous subscription history on any Claude product.
  • Activation: Users need to enter payment information to start the trial, but are not charged during the 7-day period.

[Aziz]

Aziz I sent you guest pass to CC in messages... click on the link.


Key Details About Claude Guest Passes:

• • Duration: Each pass grants 7 days of Claude Pro access.
  • Access: They include full access to the Claude Code tool.
  • Eligibility: Recipients must not have a previous subscription history on any Claude product.
  • Activation: Users need to enter payment information to start the trial, but are not charged during the 7-day period.

Thanks Paul,

But I face with problems, problems....
7 days for free, but I have to make a contract and pay then. I don't have a credit card too...
It is easier to manage more free e-mail accounts. And login with different e-mail accounts.
Or simply waiting a couple of hours..

Anyway, thank you for your support.
Aziz

[KingJL]

Aziz, I also have been doing some 'chatting' with AI... I have been exploring the hilbert transformer implementations. Have you explored using cascaded biquad filter hilbert transform vs the FIR hilbert transform? Much cheaper as far as cpu usage/memory budget is concerned and the results (error/drift) are more than acceptable. Attached is the report generated by the 'chat' discussion.

Efficient Hilbert Transform Implementation.docx

In my case, the reason I was exploring the cascade biquad approach was that I plan to use the TLV320ADC5120 audio adc which is a 24 bit 768kHz capable IC. The TLV320ADC5120 has programmable biquad filters available on chip that can be configured to provide the hilbert transform with the I signal in one channel and the Q signal in the other channel, thus offloading the hilbert transform from the cpu/mcu (STM32H743) altogether.

[Aziz]

Hi KingJL,

Aziz, I also have been doing some 'chatting' with AI... I have been exploring the hilbert transformer implementations. Have you explored using cascaded biquad filter hilbert transform vs the FIR hilbert transform? Much cheaper as far as cpu usage/memory budget is concerned and the results (error/drift) are more than acceptable. Attached is the report generated by the 'chat' discussion.

[ATTACH]n445174[/ATTACH]

In my case, the reason I was exploring the cascade biquad approach was that I plan to use the TLV320ADC5120 audio adc which is a 24 bit 768kHz capable IC. The TLV320ADC5120 has programmable biquad filters available on chip that can be configured to provide the hilbert transform with the I signal in one channel and the Q signal in the other channel, thus offloading the hilbert transform from the cpu/mcu (STM32H743) altogether.

no, I didn't. I don't know anything about the "cascaded biquad filter hilbert transform". I have to ask my AI mate yet.
Well, we (I definitely) should use the complex hilbert transformer version as a FIR-bandpass. This has exceptionally good decoding characteristics even it will cost more processing power and memory usage. The achievable sensitivity and low noise decoding is really really exceptionally great. The standard Hilbert-Transform is not enough.

Just made the complex FIR-bandpass with Claude Code AI. I am not impressed at all yet. Google AI gave me a much more "stable" version, after forcing him up to 10 times to correct himself. Claude Code made an automatic tap size filter, which has bad characteristics but uses less CPU power. I'll try another AI query with given tap size (filter size).

BTW, we need really big CPU horse power for the multi-dimensional multi-spectral data processing. This will include FFT spectrums on Envelope(t) and d Phi(t)/dt outputs. For both TX and RX channels.
Aziz

[Aziz]

Hi all,

I have finally built and tested the Claude Code (CC) AI for the complex FIR-Bandpass Filter. It is running. As good as Google AI code. (Google AI is here better).
This is very strange: CC must search the gain of 0 dB response in the passband and adapt the coefficients.

Attached is a complete zip-file (CC output) for your entertainment.
Aziz

[Aziz]

Hi all,

I have now 2 Google AI and 1 CC solutions for the Complex FIR-Bandpass. All behave same.
I can detect frequency changes of 0.001 Hz of my Franklin LC-Oscillator now. This is insane. The Complex FIR-Bandpass is the most powerful tool ever.
I will get rid of the standard Hilbert-Transform. The bandpass works only for the Q-channel.

[moodz]

Hi all,

I have now 2 Google AI and 1 CC solutions for the Complex FIR-Bandpass. All behave same.
I can detect frequency changes of 0.001 Hz of my Franklin LC-Oscillator now. This is insane. The Complex FIR-Bandpass is the most powerful tool ever.
I will get rid of the standard Hilbert-Transform. The bandpass works only for the Q-channel.

Aziz to even be in the ballpark ... you need a few orders more of sensitivity.
The method I use results in a frequency sensitivity of 16 uHz or 0.000016 hz at the phase detector. ( 100 nV of signal to noise )
If the target is 2.5 diameters away from the coil .. thats a very small change in the far field to get back into the detector.

PS the secret to using CC is not to tell it what to do ..... tell it the problem you are trying to solve.

[moodz]

This is nothing special ... a basic crystal oscillator on a CPU board such as I am using will have 20 ppm spec over a wide temperature range.
Basic crystal oscillators can hold a stability of ±20ppm over a wide temperature range (e.g., -40 to +85°C). So 16 uHz is not a big thing.
The CPU is dividing the 25 Mhz clock down to 5 - 100 khz .. so spec is even tighter.
The main problem for the Franklin oscillator is that it might have stability as far as LC oscillators go .. but is far away from the stability of even $1 crystal oscillators.

[Aziz]

Hi Paul,

PS the secret to using CC is not to tell it what to do ..... tell it the problem you are trying to solve.

Yes, I made the CC a very complex query. But it still solved well. I have to learn to make right queries to the AI yet. I'm a bloody beginner. I'll try another queries soon.

Aziz to even be in the ballpark ... you need a few orders more of sensitivity.
The method I use results in a frequency sensitivity of 16 uHz or 0.000016 hz at the phase detector. ( 100 nV of signal to noise )
If the target is 2.5 diameters away from the coil .. thats a very small change in the far field to get back into the detector.

100nV is around -200 dB! That's a long way to get there.
I don't know yet, whether I can get into the same region. But the "damped ringing PI" is a good candidate for. The complex demodulators reduce a lot of noise and are translating the signals into the low frequency time-domain signals without additional noise. And I don't need any front-end amplifiers.

Cheers,
Aziz