Missing Link between pulse and continuous waveform detectors ?

[moodz]

DPS5000 performance

... nearly 90db down on the receive but the lockin still pulls it in for analysis.
The pulse makes barely one pixel on the T display ( top half ) but enough phase and amplitude identify the target in the F display. ( bottom half )


moodz

[moodz]

dps5000 performance

... and no lockin results in messing up the phase / amplitude like this ...
descrimination is stuffed .... this is what happens on so called multifrequency MDS.


moodz

[simonbaker]

... and no lockin results in messing up the phase / amplitude like this ...
descrimination is stuffed .... this is what happens on so called multifrequency MDS.
[ATTACH]9587[/ATTACH]

moodz

Hi moodz:

Seems to me this is all very interesting stuff and the approach I'd want to take if I still had a brain... something good will come from it I have a feeling... I'll try to follow along best I can... keep us updated...

Cheers,

-SB

[moodz]

Hi moodz:

Seems to me this is all very interesting stuff and the approach I'd want to take if I still had a brain... something good will come from it I have a feeling... I'll try to follow along best I can... keep us updated...

Cheers,

-SB

Thanks Simon ... I am still looking at the maths behind all this ... here is another plot showing the phase relationship of the sinc pulse with increasing frequency, The top plot or T ( time domain) shows the rx pulse. The bottom plot or F ( frequency / phase ) shows the phase content plot of this pulse time in radians .... phase of course swings both negative and positive ... mathematically the phase variation should begin at a point ( ie phase = 0 ) as DC has no phase. You can see the nice linear ramp of the phase swings in the frequency content of the pulse from DC to 48 kHz. Different targets will effect those thousands of phase swing points in a very particular way depending on the nature of the target. In conjunction with the amplitude points descrimination is able to resolve target types with much more resolution. What is really nice about all this is that the processing is all just mathematics .. cant add any noise.



moodz

[moodz]

DPS5000 performance

... and the normalised phase ramp. ( absolute phase )



moodz

[Aziz]

Hi moodz,

very interesting work man. This is, what also could be done with laptop and usb sound card and some external hardware. The laptop would have a much higher number processing power. It also would reduce the post processing "digital noise" if using the double precision numbers (8 byte IEEE double numbers).

I am not surprised about the lock-in amp possibilities. It even allows signal measurements just few dB above the noise level. But DD coils are then very susceptible to ground effects and inductive coupling changes.

Looking forward to see more ...

Regards,
Aziz

[moodz]

DPS5000 functional

Thanks Aziz ... below is a functional block of the DPS5000 ... one of the tricky bits I have been pondering is how to process up to thousands of data points in a single sample ( each data value is a complex variable ie Amplitude and Phase ) and each can be processed as an elementary metal detector. This equates to thousands of parallel descrimination points.
The question is how to produce one output signal that indicates a good target or not.

After some preliminary testing of code on the back end ( where the Phase and Amplitude is pumped out of the processing engine ) I believe I have found an answer .....

A neural network .... using the same principle as the meatware on your shoulders ... a neural network can not only process thousands of parallel data inputs at the same time but it can also adapt ( or learn ) .... This would mean that from the thousands of amplitude / phase inputs to the NN there would be say several outputs ranging from ( for example ) dig, dont dig, hotrock, ferrite, dont know, run away etc.

Additionally the user could 'train' the detection system to recognise new 'unknown' targets if they are worth digging ( like coins ) or avoiding (like landmines )

Historically neural networks have very high performance for pattern recognition and the ability to 'lift' a wanted pattern from high levels of interferance ( eg ground clutter ). Most NNs can achieve hit ratios above 90% in this type of application .. ( ie pattern recognition ).



moodz

[bbsailor]

All,

You might enjoy reading this.

Pay attention to detection characteristics at very early, mid and late delay times.

http://uxoinfo.com/blogcfc/client/enclosures/htd.pdf

bbsailor

[moodz]

All,

You might enjoy reading this.

Pay attention to detection characteristics at very early, mid and late delay times.

http://uxoinfo.com/blogcfc/client/enclosures/htd.pdf

bbsailor

Thanks very much for that reference BB ..... so there must be some merit in using a neural network / fuzzy logic to descrim targets

I will have a closer look at this paper. Thanks very much for the link !

Regards,

moodz.

[Aziz]

Hi moodz,

I have a vast knowledge about neural networks (I hope, I did not forget too much).

Some hints on NN:

- a three or more layer back-propagation NN are convenient (for non-linear problems). A two layer is only capable for linear dependent problems.

- reduce the input parameters (decimation is ok). The more input parameters, the more training sample sets are necessary (very difficult to get and manage thousands of sample sets).

- try to reduce the number of hidden neurons (in the inner layers). This will make the generalisation more possible and avoid over-training.

- log sigmoid activation function is appropriate (see http://en.wikibooks.org/wiki/Artific...tion_Functions )

- the input parameters need to be scaled (-1 .. +1 or better -0.9 .. +0.9)

- every class output classification needs one output neuron (one for "precious", one for "junk", one for "mine!", one for "run!", ..)

- the output parameter could be 0 for less likely, 1 for high likely output, 0.5 could be used as a threshold (see log sigmoid activation function)

If you have some specific questions, you are welcome to ask me.

Aziz

[moodz]

Hi moodz,

I have a vast knowledge about neural networks (I hope, I did not forget too much).

Some hints on NN:

- a three or more layer back-propagation NN are convenient (for non-linear problems). A two layer is only capable for linear dependent problems.

- reduce the input parameters (decimation is ok). The more input parameters, the more training sample sets are necessary (very difficult to get and manage thousands of sample sets).

- try to reduce the number of hidden neurons (in the inner layers). This will make the generalisation more possible and avoid over-training.

- log sigmoid activation function is appropriate (see http://en.wikibooks.org/wiki/Artific...tion_Functions )

- the input parameters need to be scaled (-1 .. +1 or better -0.9 .. +0.9)

- every class output classification needs one output neuron (one for "precious", one for "junk", one for "mine!", one for "run!", ..)

- the output parameter could be 0 for less likely, 1 for high likely output, 0.5 could be used as a threshold (see log sigmoid activation function)

If you have some specific questions, you are welcome to ask me.

Aziz

Hi Aziz ... yeh I am going to use a 3 layer backprop network with a sigmoid transfer model. As you say it will be a no brainer to get it up and running. The training is going to be a little more difficult however I have added an extra input node called attached to a toggle switch which can have three positions called 1. BAD 2. UNKNOWN 3. GOOD
The data will also have a Z dimension ... the average of the last N samples.
So there will be Phase Frequency ZPhase ZFrequency. This acts as the autotune sample.
I intend to present the detector with known targets with/without clutter and 'reward' the detector by pressing GOOD for good target. 'Punish' the detector for detecting bad targets as good by pressing BAD. I might have to automate the process somewhat.

moodz

[moodz]

DPS5000 real world

Just in case you think the DPS hardware dont exist !!

Sinc Rx waveform .....

Note that the width of the main pulse is about 50 uS ... not so different from a PI



Now some math on the scope does the FFT ( purple trace ) and we have a spectral "pulse" that is 20 Khz wide ( from DC to 20 Khz ). Noise floor is better than 40db down ( 10 db / div ).



moodz

[moodz]

DPS5000 real world

... take 2

.. I should learn to drive this scope better

The previous spectral shot hardly does justice to the spectral pulse performance achievable with the a Sinc(x) pulse.

You can clearly see the 'brickwall' rolloff in spectral energy ( pink trace ) at 24 Khz Try and do that with an analogue filter .... The sample frequency of the ADC is 48 Khz in this case. Passband ripple less than 1 db.

Cant show you the phase performance ... the scope computer says no. ( cough )



moodz

[Aziz]

Hi moodz,

you have a nice scope (enviable). You did a good job. Thanks for sharing your findings.

I am wondering, whether it would be possible using just the sound-card output and input (without external hardware allowing a high s/n ratio). I can assist you with the laptop experiments and the frequency response of targets.

I also have a Hi-Fi amplifier (20 W, up to 20 kHz), which should allow more current to the coil. Fortunatelly, there is almost no change necessary in my experimental software (sinc burst only must be implemented yet).

Lets have a look into this matter. Looking forward to confirm your finds.

Aziz

[simonbaker]

... take 2

.. I should learn to drive this scope better

The previous spectral shot hardly does justice to the spectral pulse performance achievable with the a Sinc(x) pulse.

You can clearly see the 'brickwall' rolloff in spectral energy ( pink trace ) at 24 Khz Try and do that with an analogue filter .... The sample frequency of the ADC is 48 Khz in this case. Passband ripple less than 1 db.

Cant show you the phase performance ... the scope computer says no. ( cough )

[ATTACH]9631[/ATTACH]

moodz

This is wild cool stuff... just watching and enjoying...

the scope computer says no.

Little Britain fan????

-SB