Hacker News Reader

Comment by rollulus | Hacker News Reader

Are there any ML approaches to identifying signals? Since using a receiver that produces sound given a FM/SSB demodulation of whatever true modulation is used, or visually inspecting a waterfall certainly has limitations.

View on HN

My students Dianne, Pauli, Mikko and Juho tried [1] this. They wrote a blog [2] and a git repository [3]. It's not a ready to use product, but an exploration and learning journey to using machine learning for identifying signals.

Multiple steps are required to get to readable or audible output: identifying a signal, identifying modulation and demodulating.

[1] https://terokarvinen.com/2022/ai-sdr-analyze-radio-signals-w... [2] https://muikkurf.wordpress.com/ [3] https://github.com/kajami/SDR-project

Yes of course, but anybody doing anything non-trivial is selling it to the government and not writing about it online. Using the spectrum or waterfall as the input to an RFML system is very common and in my experience works pretty well. A lot of systems also train on I/Q data. Usually these systems are designed to be trained on signals that are relevant to a specific mission. So a product might look like both a platform for annotating and training a model as well as deploying it, usually as a plug-in into a separate expensive SIGINT platform. The biggest products I'm aware of aren't even advertised and last I heard cost in the six figures.

Rohde & Schwarz ran a contest on this idea a while ago. I can't find a link with a quick google, but I'm pretty sure I'm not making it up -- if anyone here knows more I'd love to know what came of that.

ML feels like an overkill for a single signal. If you want to process large, wideband scans, you could iterate over peaks and check things like: does phase shift create any interesting set of points, does PLL find consistent amplitude keying, does the frequency move around the centre. That covers 99% of what you're going to find in the wild. The first part (transform the signal) is definitely not great for ML, but the second part (does the result look like set of points / digital on/off key / voice) could be classified that way.

There are some existing projects like https://github.com/randaller/cnn-rtlsdr (that one tries to identify more specific tv signals though)

Automatic classifiers tend to look for power above the background noise and then AM demodulate the signal around it. That demodulated signal, or video as it's called, is centered around 0 Hz and can be matched against a database of spectrum masks for various modulations, baud rates and other parameters.

No neural nets required. Just good old regression.

So how would they fare against something below the noise floor, like FT8

FT8 isn't actually below the noise floor if you look at the bandwith the signal is detected in, rather than the 2500 Hz reference bandwidth.

https://tapr.org/pdf/DCC2018-KC5RUO-TheReal-FT8-JT65-JT9=SNR...

Has details, but basically you should add 26 dB to account for the difference between 2500 Hz and the 6.25 Hz bandwidth each FT8 tone is detected in.

You can use ML to do the visual inspection - https://www.macaulaylibrary.org/2021/06/22/behind-the-scenes... is for birdsong, but that's just a lower-frequency signal, right? :-)

There is a fair amount of published research on this.

code for a CNN-based approach here:

https://github.com/randaller/cnn-rtlsdr

Yes, there are companies specializing in this:

https://www.deepsig.ai/

Reply on news.ycombinator.com