RadioPulseDenoising

This repository contains scripts for the construction and training of convolutional neural networks (CNNs) designed for classifying and denoising radio pulses from cosmic-ray air showers. The CNNs were developed and trained using TensorFlow and Keras, so these libraries are required to run the provided scripts.

The repository is organized as follows:

• models/: Contains the pre-trained networks.

• data/: Contains waveform datasets used for testing and demonstration.

  • sample_set/: A small sample of waveforms that can be used to see how the data is prepared for training the neural networks. This is not sufficient for actual model training — a larger dataset should be prepared in the same format as demonstrated here.
  • validation_data/: A validation dataset for evaluating the performance of the provided pre-trained models. This data is intended exclusively for testing the already trained models found in the "models" directory.

• Notebooks: Provide essential workflows for data preprocessing, network training, and model application.

Notebooks Overview

1. DataReshapingForNetworks

This notebook preprocesses waveform data for network training. It reshapes waveforms from two channels into single samples and supports:

• Preparing datasets for the Denoiser, requiring both pure signal traces and signal-plus-noise traces.
• Preparing datasets for the Classifier, requiring signal-plus-noise traces and noise-only traces.
• Splitting the data into training and test sets.

2. NetworksArchAndTraining

This notebook defines the architecture of the CNNs (both Classifier and Denoiser) and handles their training.

3. Apply_Networks

This notebook evaluates the pre-trained networks located in the /models directory using the validation dataset.

Reference Proceedings:

The following are the proceedings that were written for this work:

1- Application of Machine Learning to Identify Radio Pulses of Air Showers at the South Pole. Link: https://pos.sissa.it/470/034/

2- Search for Cosmic-Ray Events Using Radio Signals and CNNs in Data from the IceTop Enhancement Prototype Station. Link: https://pos.sissa.it/444/291/

3- Deep Learning for the Classification and Recovery of Cosmic-Ray Radio Signals against Background Measured at the South Pole. Link: https://pos.sissa.it/424/012/

4 -Classification and Denoising of Cosmic-Ray Radio Signals using Deep Learning. Link: https://pos.sissa.it/395/417/pdf

Acknowledgments:

We want to acknowledge the Tunka-Rex people, Kostunin, Dmitriy (dmitriy.kostunin@desy.de and github @kostunin) and Dmitry Shipilov (justforprince@gmail.com and github @Dofollow), for sharing their CNN code with us, which became the basis for this work, their code can be found here: https://gitlab.iap.kit.edu/tunkarex/denoiser