Installation:

chmod +x install.sh
./install.sh

Running demo:

source .env/bin/activate
python3 demo.py

Assumptions:

• We know how the signal looks like when it's clean (it starts clean)
• The anomaly is additive/subtractive (increases/decreases the energy level)
• The anomaly is at least (1/n_frec_div)*sample_bandwith wide

Assumptions that can be compensated or fixed with better control mechanisms:

• The anomaly lasts at least 5nfft (it can be reduced to one nfft or less)
• The anomaly starts/stops suddently (not gradually)
• There can only be one simultaneous anomaly, and it can't change it's class

As shown in experiments.ipynb, the model has a high accuracy under these conditions.

Accuracy: 0.9997777777777778

Confusion Matrix:
[[1500    0    0]
 [   1 1499    0]
 [   0    0 1500]]

The dataset is structured like:

• dataset/Jamming/Clean
• dataset/Jamming/Narrowband
• dataset/Jamming/Wideband

Each folder has multiple files, each one is a single IQ recording saved as a np.complex64 numpy buffer.

The metadata file is located at dataset/Jamming/metadata.csv, and it should look like this:

FileName,SignalType,JammingStartTime,AveragePower_dB
000498000_000012000_000050000_1731003770598630_DVBT.data,Clean,-1,-28.867297172546387
000498000_000012000_000050000_1731084527072217_DVBTWidebandJamming.data,Wideband,18671,-28.867297172546387
000498000_000012000_000050000_1731084527088622_DVBTNarrowbandJamming.data,Narrowband,18671,-28.867297172546387
...

The config file is located at dataset/Jamming/config.csv, and it should look like this (although it is currently being ignored):

SampleRate,CenterFrequency,NFFT
12000000.0,498000000.0,1024

Credits

Authors:

• .
• .
• .

Made in collaboration with University of Santiago de Compostela.

Co-tutors:

• Francisco Javier Valera Sánchez
• Anxo Tato Arias