diff --git a/ml_research/Diagram_ML.svg b/ml_research/Diagram_ML.svg new file mode 100644 index 0000000..f96baa4 --- /dev/null +++ b/ml_research/Diagram_ML.svg @@ -0,0 +1,4 @@ + + + +
Device produces a scan of N samples in Frequency F and loops over the entire Frequency band

Device cannot produce I/Q data
Device produces a sc...
Datasets contain I/Q data or Spectrograms (that can be built out of I/Q data)
Datasets contain I/Q...
SAME FORMAT











SAME FORMAT...
Generation of pseudo spectrogram with frequency bins and time length similar to train dataset
Generation of pseudo...
Spectrograms are cut so that the shown frequencies and time length is congruent with device data
Spectrograms are cut...
Train data
Convert to TF-Lite
(Model: MobileNetV3?)
Train data...
Prediction on device using TF-Lite library and saved model weights
Prediction on device...Text is not SVG - cannot display \ No newline at end of file diff --git a/ml_research/datasets.md b/ml_research/datasets.md index 60d95f0..ebf796a 100644 --- a/ml_research/datasets.md +++ b/ml_research/datasets.md @@ -1,3 +1,5 @@ +## Overview of ML Dataets + | dataset name | data content | data format | frequencies / Sampling rate | other infos | source paper/website | source data | open questions | |----------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------| | Noisy Drone RF Signal Classification (Glüge et al.) | Drones: DJI, FutabaT14, FutabaT7, Graupner, Taranis, Turnigy | I/Q Data as well as generated Spectrograms available | The device can scan a range of 6Ghz. But the newest plot is very weird. +/- 7 Mhz, can this really be? "non-overlapping signal vectors of length of 1048576 samples, which corresponds to approx. 74.9ms at 14MHz" | mixed with either Labnoise (50%) or Gaussian noise (50%). The noise class was created by mixing Labnoise and Gaussian noise in all possible combinations. Several levels of SNR were used over the entire dataset. | https://www.scitepress.org/Link.aspx?doi=10.5220/0012176800003595 https://github.com/sgluege/Noisy-Drone-RF-Signal-Classification https://github.com/sgluege/Noisy-Drone-RF-Signal-Classification-v2/tree/main | https://www.kaggle.com/datasets/sgluege/noisy-drone-rf-signal-classification-v2/data https://www.kaggle.com/datasets/sgluege/noisy-drone-rf-signal-classification | Not sure about the frequencies. | @@ -5,3 +7,7 @@ | DroneRF (Allahham et al.) | Drones: Bepop; AR; Phantom | "the dataset contains only time series data, and not the complex IQ signals" (From Glüge and not from the authors) - but what does this exactly mean? | capture the whole 2.4GHz bandwidth, we have used 2 RF receivers. Each RF receiver has a maximum instantaneous bandwidth of 40 MHz, so both receivers must be operating simultaneously to at least capture a technology spectrum such as WiFi (i.e. 80 MHz). Recorded using universal soft- ware radio peripheral (USRP) software-defined radio (SDR) transceivers. Signals that could be considered noise in the 2.4 GHz band (Bluetooth, Wi-Fi) were not recorded. | modes, including off, on and connected, hovering, flying, and video recording. | https://www.sciencedirect.com/science/article/pii/S2352340919306675?ref=pdf_download&fr=RR-2&rr=8bf5de727fa35d7f | https://data.mendeley.com/datasets/f4c2b4n755/1 | Apparently time versus db? Not very clear what is in the data | | Radio-Frequency Control and Video Signal Recordings of Drones (Vuorenmaa et al.) | Drones: DJI Inspire 2 (2.44 and 5.8 GHz), DJI Matrice 100 (2.44 GHz), DJI Matrice 210 (2.44 and 5.8 GHz), DJI Mavic Mini (2.44 GHz), DJI Mavic Pro (2.44 GHz), DJI Phantom 4 (2.44 GHz), DJI Phantom 4 Pro Plus (2.44 and 5.8 GHz), Parrot Disco (2.44 GHz), Parrot Mambo (2.44 GHz), Yuneec Typhoon H (2.44 and 5.7 GHz) | I/Q Data (And Video data) | TODO: not clear yet | | https://zenodo.org/records/4264467 | | | | Spectrogram Dataset (Wicht et al.) | Wi-Fi and Bluetooth signals (NOT DRONES) | Sepctrograms | | | https://www.mdpi.com/2306-5729/7/12/168 | | | + + +## Current Proposal for ML +![Diagram ML](./Diagram_ML.svg)