search

radioML / dataset

Open RadioML Synthetic Benchmark Dataset
244 stars 140 forks source link

Weird analog modulation data #19

Open liuzhejun opened 5 years ago

liuzhejun commented 5 years ago 
    Hi,Tim, I downloaded the dataset from the DeepSig Inc website, I found that most of the analog modulation data in the dataset is very strange, their IQ graphics are like this:

iq, I am curious, is this normal?

zhw0211 commented 5 years ago

I want to know the website where you downloaded the dataset , could you please tell me?

liuzhejun commented 5 years ago

I want to know the website where you downloaded the dataset , could you please tell me?

https://www.deepsig.io/datasets

rutrilla commented 5 years ago

Hi there!

I've also noticed this behavior. I've been checking the code and I think that in the case of AM-SSB modulation, the audio signal is being killed before being transmitted, so probably what we're seeing are just the channel effects.

These are the AM-DSB and AM-SSB transmitters:

amdsb amssb

If we compare both of them, one of the main differences, and what I think is the source of the problem, is that AM-DSB operates with complex numbers while AM-SSB uses floats. In both cases, the audio signal is multiplied by a 0 Hz sine signal source. When the output of this signal source is complex, it results in a unitary real part and a 0 imaginary part. However, when the output is float, it only results in a 0 output. Thus, when multiplying this output with the audio signal in the float domain, the audio signal is killed.

I made this flowgraph to show you what I'm saying:

grc_diagram

This is the output of the flowgraph:

AM_modulations

I think this could be the source of the problem. I've checked previous commits, and I think this problem is present since the beginning. So I'm afraid that the AM-SSB modulation could be wrong in the dataset, compromising the classification results obtained with it to a certain extent.

Do my findings make sense to you or is there anything that I may have not understood properly? Please, check it and update us with your conclusions.

I look forward to hearing from you.

Regards,

Ramiro Utrilla

liuzhejun commented 5 years ago

Hi there!

I've also noticed this behavior. I've been checking the code and I think that in the case of AM-SSB modulation, the audio signal is being killed before being transmitted, so probably what we're seeing are just the channel effects.

These are the AM-DSB and AM-SSB transmitters:

amdsb amssb

If we compare both of them, one of the main differences, and what I think is the source of the problem, is that AM-DSB operates with complex numbers while AM-SSB uses floats. In both cases, the audio signal is multiplied by a 0 Hz sine signal source. When the output of this signal source is complex, it results in a unitary real part and a 0 imaginary part. However, when the output is float, it only results in a 0 output. Thus, when multiplying this output with the audio signal in the float domain, the audio signal is killed.

I made this flowgraph to show you what I'm saying:

grc_diagram

This is the output of the flowgraph:

AM_modulations

I think this could be the source of the problem. I've checked previous commits, and I think this problem is present since the beginning. So I'm afraid that the AM-SSB modulation could be wrong in the dataset, compromising the classification results obtained with it to a certain extent.

Do my findings make sense to you or is there anything that I may have not understood properly? Please, check it and update us with your conclusions.

I look forward to hearing from you.

Regards,

Ramiro Utrilla

Hello, utrilla. Thank you for the explanation! This is a reasonable explanation of the problems I have encountered before. So have you successfully generated normal AM-SSB modulated data?

rutrilla commented 5 years ago

Hi liuzhejun,

Those are my current source-sink pairs for AM-DSB and AM-SSB modulations. I've tested them with a real over-the-air transmission, and it seems they work, you can hear the audio on the sink side. Moreover, the spectrum shape also makes sense to me. However, I don't have much experience in modulations and GNURadio, so I'd like the people from RadioML to confirm the situation and the proper solutions. I think it's important to keep people working with the same procedures for signal generation.

I look forward to hearing from you.

Have a nice day!

Ramiro

GRC_RadioML_AMDSB_AMSSB.zip

liuzhejun commented 5 years ago

Hi liuzhejun,

Those are my current source-sink pairs for AM-DSB and AM-SSB modulations. I've tested them with a real over-the-air transmission, and it seems they work, you can hear the audio on the sink side. Moreover, the spectrum shape also makes sense to me. However, I don't have much experience in modulations and GNURadio, so I'd like the people from RadioML to confirm the situation and the proper solutions. I think it's important to keep people working with the same procedures for signal generation.

I look forward to hearing from you.

Have a nice day!

Ramiro

GRC_RadioML_AMDSB_AMSSB.zip

Hi, utrilla, thank you for your sharing! I'm very sorry for didn't respond to you in time because of work. And I agree with you that they need to confirm and correct this error. But it seems that they have not updated on this project for a long time. I sent an email to o'shea before, and I didn't get a reply. I am no longer involved in this project. In fact, my major is software engineering, so the knowledge of radio is very difficult for me, my work about this project has been replaced by others. But I still hope that the official staff can give a reply, I will pay attention to the latest developments. Good luck to you!

lintongtong123 commented 4 years ago

Hi, I want to know the carrier frequency of all modulation modes, but I can not find that in the code. Do you know?

kevinchez commented 4 years ago

Hi liuzhejun,

Those are my current source-sink pairs for AM-DSB and AM-SSB modulations. I've tested them with a real over-the-air transmission, and it seems they work, you can hear the audio on the sink side. Moreover, the spectrum shape also makes sense to me. However, I don't have much experience in modulations and GNURadio, so I'd like the people from RadioML to confirm the situation and the proper solutions. I think it's important to keep people working with the same procedures for signal generation.

I look forward to hearing from you.

Have a nice day!

Ramiro

GRC_RadioML_AMDSB_AMSSB.zip

Hi @rutrilla ! I'm not familiar with GNU Radio. So how to modified the code to generate the correct data?

rutrilla commented 4 years ago

Hi @kevinchez,

My source files are attached to the message you've quoted. If you open these files, GNU Radio will generate the corresponding python files and you can compare them with the code of the AM-DSB and AM-SSB transmitters.

amdsb amssb

However, if you just want to regenerate the dataset, I think that newer version, RML2016.10b, fixed the problem with the AM-SSB modulation, so you don't have to deal with the code.

Have a nice day!

kevinchez commented 4 years ago

Hi @kevinchez,

My source files are attached to the message you've quoted. If you open these files, GNU Radio will generate the corresponding python files and you can compare them with the code of the AM-DSB and AM-SSB transmitters.

amdsb amssb

However, if you just want to regenerate the dataset, I think that newer version, RML2016.10b, fixed the problem with the AM-SSB modulation, so you don't have to deal with the code.

Have a nice day!

Hi @rutrilla ! Thanks for your help!