Tagged: neural networks

Near Field Localization with Machine Learning and 7 Coherent RTL-SDRs

Thanks to Laakso Mikko and Risto Wichman researchers at the Department of Signal Processing and Acoustics in Aalto University, Finland for submitting news that their recent paper titled "Near-field localization using machine learning: an empirical study" is available on IEEE Xplore. (To access the paper you need an IEEE subscription, but we see no harm in letting individuals know that they can search for the DOI on sci-hub to get it for free).

The work described in the paper uses 7 RTL-SDR dongles with their clocks connected together. Combined with noise source calibration, this results in a coherent SDR. They then train a Deep Neural Network to perform near field localization using an antenna array. If you are interested, we have out own 5-channel coherent SDR called "KrakenSDR" which will soon be released for crowd funding. The abstract reads:

Estimation methods for passive near-field localization have been studied to an appreciable extent in signal processing research. Such localization methods find use in various applications, for instance in medical imaging. However, methods based on the standard near-field signal model can be inaccurate in real-world applications, due to deficiencies of the model itself and hardware imperfections. It is expected that deep neural network (DNN) based estimation methods trained on the nonideal sensor array signals could outperform the model-driven alternatives. In this work, a DNN based estimator is trained and validated on a set of real world measured data. The series of measurements was conducted with an inexpensive custom built multichannel software-defined radio (SDR) receiver, which makes the nonidealities more prominent. The results show that a DNN based localization estimator clearly outperforms the compared model-driven method.

The paper notes that the code used in the experiments is open source and available on GitHub.

If you're interested, we also posted about Laakso's previous work on beamforming with a phase coherent 21-channel RTL-SDR array back in February.

Examples of MUSIC pseudospectra. The units are [m] for range r on the vertical axis and degrees for θ on the horizontal axis. Red crosses mark the true location and black circles the NFLOPnet estimated location.

SDRA2021 Talks: Electrosense, Neural Network Signal Classification, gr-rpitx, Radio Astronomy and More

The 2021 Software Defined Radio Academy conference was held online this year on June 26/27 and the talks have been recently uploaded to YouTube. There are some interesting talks this year including a presentation on various SDR related topics including Electrosense, gr-rpitx, 21cm radio astronomy with low cost SDR hardware, and using deep learning neural networks for automatic signal identification. Our favorite talks and blurbs are collected below for easy access, and the full set of talks can be found on their YouTube channel.

Dr. Henning Paul: Building a flexible Multi-Antenna-capable SDR using open Source

The availability of Open Source software components enables the ambitious hardware hacker to design their own powerful SDR. This talk is the follow-up to the talk on Scientific SDR and recapitulates the steps towards the current design of a Homebrew SDR based on a Xilinx Zynq SoC using the Linux kernel and other Open Source components. Furthermore, one of its applications, receiving shortwave radio with antenna diversity is presented.

SDRA2021 - 04 - Dr. Henning Paul: Building a flexible Multi-Antenna-capable SDR using open Source

Jean-Michel Friedt: GNURadio compatible gen. purpose SDR emitter using RasPi4 PLL

GNU Radio, the Raspberry Pi single board computer and Digital Video Broadcast Terrestrial receivers make an awesome combination for educational purposes of Software Defined Radio. gr-rpitx aims at complementing these tools with emitting capabilities, combined with the flexibility of GNU Radio.

SDRA2021 - 08 - Jean-Michel Friedt: GNURadio compatible gen. purpose SDR emitter using RasPi4 PLL

Sreeraj Radjendran: Knowledge extraction from wireless spectrum data

In this half-hour talk, the need for large scale wireless spectrum monitoring will be discussed. A short introduction to a large scale wireless spectrum monitoring framework, Electrosense, will be given. Furthermore, how anomaly detection and signal classification can be performed using the collected data will also be discussed. Insights to the major problems with state-of-the-art machine learning models will also be discussed in this context.

SDRA2021 -11- Sreeraj Radjendran: Knowledge extraction from wireless spectrum data

Stefan Scholl, DC9ST: Classification of shortwave radio signals with deep learning

Automatic mode classification of radio signals in the HF band is a valueable tool for band monitoring, operation of rare transmission modes and future applications of cognitive radio. In recent years, machine learning has established as a general and very powerful approach to classification problems. The presentation first provides an introduction to neural networks and deep learning. Then neural nets are applied to the task of radio signal classification. The result is an experimental deep convolutional neural net (CNN), that can distinguish between 18 different transmission modes occurring in the HF band, such as AM, SSB, Morse, RTTY, Olivia, etc.

Additional Links: Stefan Scholl's post on this topic 

SDRA2021 -12- Stefan Scholl, DC9ST: Classification of shortwave radio signals with deep learning

Marcus Leech: Mapping the sky at 21cm: Gnuradio and Radio Astronomy

We show the results of a year-long sky survey at the 21cm hydrogen line, producing an intensity map of the sky covering a declination range from -35 to +75DEG. We discuss the software tools used, Gnu Radio signal flows, and the hardware aspects of the instrument.

SDRA2021 -14- Marcus Leech: Mapping the sky at 21cm: Gnuradio and Radio Astronomy

Deep Learning Neural Network Signal Identification Software for the RTL-SDR

Recently GitHub user randaller released a piece of software that utilizes the RTL-SDR and neural networks for RF signal identification. An artificial neural network is an machine learning technique that is based on approximate computational models of neurons in a brain. By training the neural network on various samples of signals it can learn them just like a human brain could. A neural network trained on signal classification can then be used by anyone to identify unknown signals. Randallers neural network software can learn either from raw IQ data, FFT processed samples, slightly demodulated data, or demodulated audio data. The tensorflow machine learning library is used as the base code, and the deep learning technique is used.

At the moment the software is only really proof of concept, and the currently trained model is only able to identify WFM, TV SECAM Carriers and TETRA. It should be possible to train the network further by providing your own samples too, but a good graphics card is required for this as the software makes use of GPU processing for training. The output of the software is a percentage which shows how confident the neural network is that it is identifying a signal correctly.

If you are interested, there is also a Reddit thread discussing this software here.

Artificial Neural Network being used to identify a WFM signal with an RTL-SDR
Artificial Neural Network being used to identify a WFM signal with an RTL-SDR