Machine Learning Project 2 : 'Characterization and automatic differentiation between minor and major disruptions'

This is part of the Project 2 of the CS433-Machine Learning course given at the EPFL Fall 2023. Note: It is important to mention that we received the project data approximately 17 days following the official start date, which significantly reduced the amount of time available to us.

Team NUCLEAR-FUSION :

• Kaan Uçar
• Elias Nicolas Naha
• Riccardo Carpineto

Project Description

The aim of the project is to predict minor and major disruptions based on personal features. To perform this classification task we need to perfom the following :

• Preprocessing physical data from the Tokamak sensors
• Implement Deep Learning and Advanced machine learning models
• Hyperparameter tuning and joint-model pipelining

We then select the most performing one. The performance is measured by the F1 score which is a good perfomance value for imbalanced data. In our case the data is unbalanced over the plasma events : 899 time-series windows of 20 datapoints; 642 for no-event label, 207 for minor events and 50 for major events. The best run we achieved a weighted 0.9684 F1 score using ResNet Architecture Classifier to classify theses windows.

Structure of Repository :

• figure folder : contains plot from grid searches for hyperparameter tuning from RF and SVM, as well as the learning curve from SVM

• data folder : contains all the data needed to run the code, the raw data from the experiments as well as the windows centered on the events given

• models : contains models and saved weights from the ResNet and NN models to avoid to train the model again

• code folder: contains all the executable files to run the code

  • GMM.ipynb : executable file that reproduces our implementation of a Gaussian Mixture Model (GMM), an unsupervised probabilistic model

  • NN_general.ipynb : executable file implementing the neural network for the basic Neural Network for classification

  • NN_window.ipynb : executable file that helps forming the windows thanks to the use of a tuned Neural Network, with label 0s and 1s

  • RNN.ipynb : executable file implementing a Recurrent Neural Network used to classify the different labels window by window

  • ResNet.ipynb : executable file implementing a Residual Neural Network used to classify the different labels window by window, leading the best results from the others models

  • SVM.ipynb : executable file implementing SVM

  • RF.ipynb : executable file implementing Random Forest

  • scaler.pkl : normalizer used in the ResNet to pre-process the data

  • window_creation.ipynb : executable file creating the windowed data from time series data. This data is saved in the data folder.

  • pipeline.ipynb : file containing the pipeline showing the process for predicting labels for a raw experiment, implementing in a first part the 'NN_window' to form the differents windows, and then apply the ResNet to predict the labels from theses.

Instructions to run :

Python version : 3.11.5

Before running the files you can use "pip install -r requirements.txt" in your python environnement to install the required librairies for this project.

To try out a model you can run any .ipynb model file with the presence of the data folder and scaler.pkl file. The models used for the joint-model are 'NN_window.ipynb' to form windows and then ResNet.ipynb to predict their labels. You can run them separately to observe their accuracy, and then run pipeline.ipynb to see how the pipeline works in details.