ARTEMIS Offline Signal Identification
ฝัง
- เผยแพร่เมื่อ 23 ก.ย. 2024
- Here we check out Artemis, an application that runs on Windows, Linux, MacOS and Raspberry Pi. Artemis assist with signal identification without the requirement of the internet.
www.aresvalley...
If you would like to show your support for this channel, then please consider becoming a member by clicking here:
/ @techmindsofficial
Products used in my videos can be purchased through my Amazon storefront.
US ► www.amazon.com...
UK ► www.amazon.co....
Patreon ► / techminds
Twitter ► / techmindsyt
Facebook ► / techmindsyt
Instagram ► / tech.minds.official
Send $$ ► www.paypal.me/...
Merch ► teespring.com/...
Some links provided in my videos are afilliate links, whereby I earn a small amount of revenue from sale using my links. You will NEVER pay more that the advertised price and every penny raised from affiliate links is spent on purchasing new cool radio related gadgets to keep you guys entertained.
#TechMinds #SDR #HamRadio
Thank you so much for the awesome video! I’m really happy to see Artemis in action and to know that you enjoyed it. We are continuously working to improve and add new features (we are currently working on audio sample analysis, like the old Signal Analyzer), so your feedback is very valuable. Thanks again for your support!
I'm super excited to see where your project goes! All you guys are really missing now is a large dataset so that you can start training an AI to automate the identification for you.
Great works guys ! Please incoroporate automated signal identification in the near future 🙂
On my Windows 10 PC the audio output always goes to loudspeakers even when headphones are plugged in and work fine with eg. youtube audio from web browser. Can't seem to get the audio output from Artemis to headphones. On Ubuntu 22.04 the AF output to headphones works as expected.
For future improvements, it would be great to be able to navigate the signal list on the left with the cursor keys to be able to quickly go through signal types without having to move mouse/trackpad and then click. Thanks for a great software anyway !
@@fotografm Thanks for opening a GitHub issue. The feature has been implemented and will be available in the next release
Hoping onenday we can have an automated real-time classifier that can do the ID live
A feature would be great, where you coud upload a short audio file, which is converted into the spectrum diagram. This spectrum image could be compared to the existing examples and ordered by similarity. This was it could be much more easy to find the right encoding. 73
Hey Leroy, this is something we are (Pierpaolo right now) actively working on. We want to add a section just for the analysis of a personal audio file such as in the 'old but gold' Signal Analyzer by Sergey Makarov. We already have the plot with the Fourier transform (FFT), and autocorrelation will be implemented soon.
@@marcodallatiezza Hey, that is really great news! Thanks for your effort for the community. :)
The old but gold method is a close cousin to convolution (I.e autocorrelation), but you’d only ever be able to do the autocorrelation from generalized radio data, and I’m not sure audio data is enough. Not all audio data is complex, I think you need IQ data. I take it images of signals won’t work with this approach unless you extract the FFT from the bitmap matrix. You can do that, but convolution would be better. Images aren’t complex numbers. This is a non trivial problem, autocorrelation is highly analytical.
@@skunkwerx9674 well it's true that the quality of the result strongly depends on how the audio sample has been recorded (bandwidth, sample rate, SNR, ...) but it is totally feasible to get the autocorrelation with excellent results from a simple audio file. Check the project documentation: we analyzed a stanag and found the acf of the main frame as NATO decommissioned specifics. Convolution, cross-correlation and autocorrelation does not involve complex analysis or complex numbers as you can see from the theoretical introduction in the same section of Artemis documentation
@@marcodallatiezzaoh I completely agree it is feasible, I just think it’s important to consider the real world scenario of your users and their radios and the original sample set. I don’t know if you can tag the samples and determine how they were recorded originally in the known signal dataset to better match what the user gives.
Anyways it was just some food for thought, I’m also implementing a similar system but haven’t gone down the auto correlation route as it has been finicky to work with noisy and spurious data versus some other approaches.
I wish you all the best and I will be keenly watching your progress, good luck!
I've been using version 3 for a couple of months and all I say is that I'm really glad that version 4 is out making the frequency search dialog a hell of a lot more intuitive 👍
Thanks! We simplified many things and we are still working on it because there are still some minor graphical improvements to do. I can see is the green range label in the filter page: it seems still too small. Anyway, we always appreciate any feedback/suggestions or wanted new features (feel free to open an issue on the GitHub repo)
Saved, when I get my radios set up again I will definitely grab a copy. Well done to those with the slills to create this handy utility!
Nice project, it will be great to see how it develops. If it could somehow integrate with an SDR# or SDR++ plugin which used live audio listening (ai sound recognition perhaps) it would automatically identify the signal. I used Artemis v4 for a bit and found it fascinating, trawling manually through the signal samples was fun at first, but it became a chore which some level of automation could solve. I have seen that iphones are able to listen live for various sounds using ai and then issue a notification on detection such as sirens or alarms. Machine learning ai audio classification software is out there such as tensorflow and keras etc, but it is not easy to integrate and results vary, also creating a training model based on a changing and developing database of sounds could prove tricky for offline users. I would guess but using cloud processing to regularly update the sound database training model might be a good thing to look into. I look forward to seeing how this whole thing develops
I guess also image classification could be used to identify based on waterfall patterns. Might be easier than doing audio classification for now.
Also also, I figure one would need a fairly fast pc to run ai classification sdr plugins, so perhaps using a USB ai dongle such as intel neural compute stick will improve performance and offload some of the processing task. I know phones are including ai chips, and apple arm silicon have neural chips, but most sdr users are using whatever computer they have to hand. So perhaps the path might be a phone app which you hold close to the computer which is running SDR software and it listens and identifies the signal?
This is actually a pretty solid idea Ben - could make for an interesting machine learning project
Done
This was really nice to see - thanks for spending time showing off open source and GPL software like this.
Artemis is a pretty app - the development team should be super proud of themselves
Cheers
Excellent ! Many thanks for this review. As I told you before, I really like all your videos. David.
Honestly, a little bit dissapointed, I thought by now someone would have created an app that can automatically identify it from the recording, much like the "Merlin Bird ID" app does for birdsong!
We started talking about it at least 5 years ago: there should still be a discussion on github for automatic signal recognition, and several experiments have been done, not only by me. If it was a matter of creating a Shazam-like routine it would have already been done, but unfortunately it doesn't work that way. Let us proceed in order: a machine learning/neural network approach would not be complex if there were not the problem of dataframe completeness. The model needs a large amount of signals (tens of thousands, at least) to be trained. This would not be a major concern because the various encodings involving different signals can be created synthetically, then adding noise and artificial distortions to make it more like a real signal during RX. However, this allows us to have a spectrum that ranges only over civilian, non-proprietary and non-military signals, about 150 out of 500 if we are optimistic. More classical methods, such as similarity recognition of audio tracks (such as Mel-frequency cepstral coefficients) could be effective,albeit marginally, if applied on the audio samples we have collected: for the same signal encoding, the content of the signal alone can affect the similarity index and thus the effectiveness of the method. Listening to a signal that has the same encoding but different content is not like listening to the same song but with a lot of noise as shazam does.We are now working on an audio sample analysis module, but if we want to do things right about self-recognition of signals, we need to accurately assess the limitations of the methods we are dealing with. Unfortunately, it is not a trivial matter of training a neural network
We started talking about it at least 5 years ago: there should still be a discussion on github for automatic signal recognition, and several experiments have been done, not only by me. If it was a matter of creating a Shazam-like routine it would have already been done, but unfortunately it doesn't work that way. Let us proceed in order: a machine learning/neural network approach would not be complex if there were not the problem of dataframe completeness. The model needs a large amount of signals (tens of thousands, at least) to be trained. This would not be a major concern because the various encodings involving different signals can be created synthetically, then adding noise and artificial distortions to make it more like a real signal during RX. However, this allows us to have a spectrum that ranges only over civilian, non-proprietary and non-military signals, about 150 out of 500 if we are optimistic. More classical methods, such as similarity recognition of audio tracks (such as Mel-frequency cepstral coefficients) could be effective,albeit marginally, if applied on the audio samples we have collected: for the same signal encoding, the content of the signal alone can affect the similarity index and thus the effectiveness of the method. Listening to a signal that has the same encoding but different content is not like listening to the same song but with a lot of noise as shazam does.We are now working on an audio sample analysis module, but if we want to do things right about self-recognition of signals, we need to accurately assess the limitations of the methods we are dealing with. Unfortunately, it is not a trivial matter of training a neural network ...
We started talking about it at least 5 years ago: there should still be a discussion on github for automatic signal recognition, and several experiments have been done, not only by me. If it was a matter of creating a Shazam-like routine it would have already been done, but unfortunately it doesn't work that way. Let us proceed in order: a machine learning/neural network approach would not be complex if there were not the problem of dataframe completeness. The model needs a large amount of signals (tens of thousands, at least) to be trained. This would not be a major concern because the various encodings involving different signals can be created synthetically, then adding noise and artificial distortions to make it more like a real signal during RX. However, this allows us to have a spectrum that ranges only over civilian, non-proprietary and non-military signals, about 150 out of 500 if we are optimistic. More classical methods, such as similarity recognition of audio tracks (such as Mel-frequency cepstral coefficients) could be effective,albeit marginally, if applied on the audio samples we have collected: for the same signal encoding, the content of the signal alone can affect the similarity index and thus the effectiveness of the method. Listening to a signal that has the same encoding but different content is not like listening to the same song but with a lot of noise as shazam does.We are now working on an audio sample analysis module, but if we want to do things right about self-recognition of signals, we need to accurately assess the limitations of the methods we are dealing with. Unfortunately, it is not a trivial matter of training a neural network ...
We started talking about it at least 5 years ago: there should still be a discussion on github for automatic signal recognition, and several experiments have been done, not only by me. If it was a matter of creating a Shazam-like routine it would have already been done, but unfortunately it doesn't work that way. Let us proceed in order: a machine learning/neural network approach would not be complex if there were not the problem of dataframe completeness. The model needs a large amount of signals (tens of thousands, at least) to be trained. This would not be a major concern because the various encodings involving different signals can be created synthetically, then adding noise and artificial distortions to make it more like a real signal during RX. However, this allows us to have a spectrum that ranges only over civilian, non-proprietary and non-military signals, about 150 out of 500 if we are optimistic. More classical methods, such as similarity recognition of audio tracks (such as Mel-frequency cepstral coefficients) could be effective,albeit marginally, if applied on the audio samples we have collected: for the same signal encoding, the content of the signal alone can affect the similarity index and thus the effectiveness of the method. Listening to a signal that has the same encoding but different content is not like listening to the same song but with a lot of noise as shazam does.We are now working on an audio sample analysis module, but if we want to do things right about self-recognition of signals, we need to accurately assess the limitations of the methods we are dealing with. Unfortunately, it is not a trivial matter of training a neural network ...
We started talking about it at least 5 years ago: there should still be a discussion on github for automatic signal recognition, and several experiments have been done, not only by me. If it was a matter of creating a Shazam-like routine it would have already been done, but unfortunately it doesn't work that way. Let us proceed in order: a machine learning/neural network approach would not be complex if there were not the problem of dataframe completeness. The model needs a large amount of signals (tens of thousands, at least) to be trained. This would not be a major concern because the various encodings involving different signals can be created synthetically, then adding noise and artificial distortions to make it more like a real signal during RX. However, this allows us to have a spectrum that ranges only over civilian, non-proprietary and non-military signals, about 150 out of 500 if we are optimistic. More classical methods, such as similarity recognition of audio tracks (such as Mel-frequency cepstral coefficients) could be effective,albeit marginally, if applied on the audio samples we have collected: for the same signal encoding, the content of the signal alone can affect the similarity index and thus the effectiveness of the method. Listening to a signal that has the same encoding but different content is not like listening to the same song but with a lot of noise as shazam does.We are now working on an audio sample analysis module, but if we want to do things right about self-recognition of signals, we need to accurately assess the limitations of the methods we are dealing with. Unfortunately, it is not a trivial matter of training a neural network
If only folk would use the TXIDs... 😙
Thanks for the heads up Mat. I'll give it a go, looks straight forward enough. :-)
Hi Teck Minds.What SDR software are you using at the start Of video .
Expertsdr3
What would really be useful would be a plugin that works with sdr programs that automatically identifies signals... I've been hoping this would be accomplished for several years as an outgrowth of "cognitive radio". I'm not a fan of AI but this is one instance where it could be very useful.
Yes, a good use case for AI- many signals to 'train' on...and much more constantly being generated- and with a active group of 'enthusiasts', could easily set up a training platform > Show fft/ waterfall with sound along with description and have input of correct or not correct. Someone with the knack for such development could easily gamify this and make it into a fun game rewarding the user for quantity and quality of input...
Making a LLM for this however im not so good with C++ lol
needs an option to decode and ID signal type
What is the name of your sdr reciever software
is there any (preferentially open source) software to automatically recognize known signals, by feeding a recorded sample? sounds like an easy task for AI..
great video. You have answered many question that I had about these signals and how to understand them. Thank You, VE7QJ
Glad it was helpful!
show parabéns pelo canal 73´s.
TNX
if only android
👍👍
🚭
And As Always, I Can`t Do Anything Because I Have WIndows 7 And I Can`t Upgrade Because Im To Poor. Please, Someone, if You Could, Buy Me A Computer. I Know Its Bold Thing To Do And Ask, But Hey, Its Internett, More Details, Im Disabled And I have Had Two Openm Heart Surgeries And I need Third One That I Know I Cant Afford And So I Want To Make Most of My Time And Im Listening To Russians And I record Them And Publicly Upload The Recordings Into My Other TH-cam Channels And A Lot of The Russians Are Being Identified And Banned From Travel To Any Countries. More Safe World. But Thats Only Single Thing I Do. :) I Would lOVE TTo Do Much More, But I can`t because I Cant Get Any Software Because Of My Compputer, I Have Good Graphics Card, I Only Need Good Motherboard With Ram And Stuff. :) I Know Noone Will help me, But Maybe There Are Chance That Someone Can Or Want. Thank You For Even reading This Comment :)
I Know Miracles Hapen, But They Have Never Hapened To Me.
I Would Do Soo Much With This Software, OMG! IW ould Do SOO MUCH, Its Crazy !
I had Windows 7 for more than a decade now. I use it as my main work computer, coded great pieces of software in it. Created a company out of it and now I'm in the top 5% earners of the country I live in.
For sure Windows 7 is not your problem. Look somewhere else and find the real culprit!
Load Linux on your Win 7 box. It'll run great, and costs nothing.
oh woe is me..... at least you have a pc...that can run windows 7 which more than enough to run a decent Linux setup... Oh, wait...many applications, very powerful, very useful...they've been made on Windows 7. Don't be a loser.
Thanks for this video! vy73 from Hamburg/Germany, Dietmar, DL4HAO :)
Thanks for watching!