AICompetition

This repository contains all the developed source code and the documentation produced for the RF signal classification project. Our implementation focuses on signals operating in the main ISM frequency band (2.40–2.48 GHz), including technologies such as WLAN, Bluetooth, and ZigBee, as well as a non-standard signal based on the SmartBAN specification. The goal of the project is to classify these overlapping signals within the increasingly congested ISM spectrum, with particular attention to future applications in Body Area Networks (BANs).

HOW TO INSTALL

To access all the materials, follow these steps:

• Copy the repository link: https://github.com/nico-net/AICompetition
• Open your terminal
• Run the command:

  git clone https://github.com/nico-net/AICompetition
  

HOW TO USE

Pipeline

To generate a synthethic dataset, launch mainCreatingImages.mlx from subfolder createDataset/ adjusting the parameters as preferred. Once a dataset is ready, trainNetwork.mlx can be launched from trainNetworkFolder/ , making sure that correct path and imagesize are set as parameters. The script includes:

• partitioning of aforementioned dataset for training, testing and validation;
• training of the chosen Neural Network with the given parameters
• performance evaluation and save of the resulting CNN

Use mainEvaluationMetricsSynthetic.m inside evaluationMetricsSynthetic/ to get a more precise evaluation of the CNN. A net has to be loaded as a variable.

To test the net on real captures, launch rxPluto.m (with a loaded net variable) from visualizationHelpers/ . Keep in mind that this was tested with the setup from the Report, which included external PLUTO synchronization. Two PLUTOs with synchronized clock are required, as results can't be guaranteed with only two indipendent PLUTO devices.

The following is a link to a Google Drive directory where one can find and download the dataset we used for our tests

https://drive.google.com/drive/folders/1NW-zfu33dwwFuJRUL_rlkAoi--zqWsN5?usp=drive_link

Here are the Networks we trained The subfolders refer to the size of the images used for training. att_noisy refers to net trained with attention gates and a dataset composed of both clear and interference images. fullnoise refers to nets trained with attention gates and with the dataset we uploaded.

https://drive.google.com/drive/folders/1qrnkdLk9yRq-4vhVCIGekWyOpfaIZuEB?usp=sharing

Folder Structure — AICompetition

Inside the folder AICompetition in your local environment, you will find the following directories:

---

createDataset

• Helper functions for signal simulation
• Signal mixer code used to combine signals within the considered frequency band
• Interference generator code used to introduce interferences
• mainCreatingImages.mlx: user interface code for dataset creation
• creatingTrainingImages.mlx: detailed script explaining all adopted and developed functionalities

---

evaluationMetricsSynthetic

• Semantic segmentation model evaluation code across different SNR levels
• Use mainEvaluationMetricsSynthetic.m to run the evaluation

---

plutoTxSmartBANAndZigBeeSDR

• Helper function for ZigBee: generates clean waveform for SDR transmission
• Helper function for SmartBAN: generates clean waveform for SDR transmission
• Generator code that uses the helper functions to create and transmit ZigBee or SmartBAN signals via PlutoSDR
• mainZigBeeSmartBANTx.mlx: compact and user-friendly interface script

---

predictionTime

• .mat files with prediction time values used to compute mean and variance
• These values are referenced in Table IV of the documentation

---

pruningQuantizationFolder

• pruningAndQuantization.m: end-to-end pruning and quantization pipeline for a ResNet-50 + attention-gated U-Net
• projectionFunction.m: compresses a neural network using neuron-PCA and performs fine-tuning

---

trainNetworkFolder

• trainNetworkDef.m: source code for training neural networks
• trainNetwork.mlx: compact and user-friendly interface for model training

---

visualizationFunctionsForSyntheticImages

• helperSpecSenseDisplayResults.m: visualization functions to show predicted and true labels
• helperVisualizationPredictedAndTrueLabels.m: loads samples, segments them with a trained model, and compares true vs. predicted labels
• mainVisualizationPredictionsSynthetic.mlx: user-friendly interface for visualizing predictions

---

visualizationHelpers

• plotOverLayed.m: overlays a colormapped matrix onto an image
• plotResults.m: visualizes net results by comparing predicted and true labels on an ISM band spectrogram
• reAllignLabels.m: reassigns mask labels to simplify colormapping
• rxPluto.m: performs synchronized IQ acquisition using two ADALM-Pluto SDRs and processes signals for AI-based classification and spectrogram visualization

VIDEO

You can see a detailed explaination of the code and our equipment in our YouTube video at this link: https://youtu.be/Wmrsy2saCLQ

LICENSE

This project is licensed under the BSD 2-Clause License.
See the LICENSE file for more details.