Artificial Neural Network - ANN and detectors

[WM6]

I would like use members experience for this idea.

Why complicated?

You can easily solve your whole problem in so called Raikov Effect way.

[Qiaozhi]

Why complicated?

You can easily solve your whole problem in so called Raikov Effect way.

Now all you need to do is find a way to hypnotize your metal detector.

[WM6]

Now all you need to do is find a way to hypnotize your metal detector.

Or yourself. And then you even do not need detector. Your neural system can clearly see gold from chair.

Raikov effect can help to realize every wild imagination and wishes. No more limits to treasure hunters.

[moodz]

Why complicated?

You can easily solve your whole problem in so called Raikov Effect way.

So if I take a cheapo ebay detector and paint major manufacturer logo on it ... it will perform better due to Raikov Effect :-)

[moodz]

The operative word is "if". Nobody is going to bell the cat, because anyone who understands the problem understands that belling that particular cat has no return on investment. The cat will eat the mouse alive.

But suppose I'm wrong, and someone does bell that cat? They've got a metal detector they've invested ten years of their life in, the underlying technology is obsolete, and it's been trained "in a particular ground location" that the trainer cleaned out as a necessary part of the training process. Completely useless even to the guy who built the thing. Thus does the cat eat the mouse even if the mouse bells the cat.

https://www.researchgate.net/publica...neural_network

[vbeeeks]

Does it not look like another advanced patent template in the making

[TLH858]

I think it's a GREAT idea! Once trained, neural networks are great at pattern recognition even with corrupted or missing data. They can generalize. Of course training will be the big challenge. I can conceive of a detector that stores the last four sweeps (2x E/W and 2x N/S) and a menu of possible targets and depths. For coin hunters there are only a finite number of targets and depths that we are interested in like penny, nickel, dime, etc and lead bullet, brass shell casing, tinfoil, etc.

You might think that training would take forever, but what if 1,000 (or 10,000) people owned the same detector and shared files? It wouldn't take long before it started to perform well, and every year it would get better.

Rick

[Skippy]

Quote:"For coin hunters there are only a finite number of targets and depths that we are interested in like penny, nickel, dime"

I'm not interested in nickels and dimes, we don't have them in my country. We have a massive selection of coins from 250 BC up to present day. And corrosion can make two examples of the same coin behave differently. Even you US guys have this problem - example: your Indian-Head cents are an alloy that corrodes badly, changing the characteristics of the coin, depending on what the dirt/soil they are buried in, how long they've been buried, etc. And don't your zinc-cored cents corrode badly, too? That would make them hard to ID ( for the purposes of not digging them)

However ... there is possibly some merit to a machine that 'stores the last 4 sweeps' as the extra information may make it possible to give a better ID.
I did propose (on Dankowski's forum) that a machine that recorded the previous few seconds of sweep could be useful. As well as simply being able to 'repeat' the sweep on command, it could repeat it at a slower rate with appropriate increase in intelligibility, and it could repeat it with more detailed 'analysis', due to being able to use pre-target and post-target data. And also present multiple opinions, eg 4 different readouts on screen, based on 4 different ID strategies.

[TLH858]

Interesting! Thanks.
Rick

[Koala]

for coin hunters there are only a finite number of targets and depths
Rick

With an infinite amount of orientations, corrosion and ground conditions and so on.

[Qiaozhi]

"..... the only 100% discriminator is a shovel."
-- White's Electronics Metal Detector XLT Engineering Report.

[TLH858]

Understood, but don't underestimate a neural networks ability at pattern recognition. There are neural networks today performing task of pattern recognition that humans can't. In fact people train networks to do things and they don't even know how the network does what it does. Most deep learning today is software running on large machines, but if we are going to build realistic embedded systems we will need to minimize size, cost, and power consumption. I suspect that we will soon see a lot more neural network chips (hardware) coming on the market.
Rick

[sled]

Quote:"For coin hunters there are only a finite number of targets and depths that we are interested in like penny, nickel, dime"

However ... there is possibly some merit to a machine that 'stores the last 4 sweeps' as the extra information may make it possible to give a better ID.
I did propose (on Dankowski's forum) that a machine that recorded the previous few seconds of sweep could be useful. As well as simply being able to 'repeat' the sweep on command, it could repeat it at a slower rate with appropriate increase in intelligibility, and it could repeat it with more detailed 'analysis', due to being able to use pre-target and post-target data. And also present multiple opinions, eg 4 different readouts on screen, based on 4 different ID strategies.

That's the job of a RNN / LSTM (long-short term memory) network, it can a learn and make decision based on previous inputs.

@TLH858 The problem you will face is data, getting clean user labelled data is tedious. For a start, I'd suggest to hook a logger onto an existing detector (bought or built) and monitor the demodulated channels + disc and ground balance setting and target output then try to learn the transfer function

f(demod_x, demod_y, disc, balance) = target (beep / no beep)

Go out metal detecting, logging data. Then train a network, implement it, connect it to your detector the same way as with the logger and verify it works (beeps) as expected. This would already be a big accomplishment. Next step would be to learn the demodulation also by using the raw signal after the preamp i.e f(rx_signal, disc, balance) = target

PS: Make your life easy and use a single board computer (like RaspberryPI) or connect directly to your laptop or stream to an AWS cluster over LTE with your phone. Whatever is the easiest and fastest to prove your idea works.

[TLH858]

Great suggestions. Thanks. I'm already using an Arduino as a data logger for some other projects.
Rick

[WM6]

With an infinite amount of orientations, corrosion and ground conditions and so on.

Bitter truth in the world of metal detectors. On other side, this is why such forum exist. Neverendingstory .....