README for collaborative work between L. Stephey, M. Marchevsky, and M. Mustafa at Lawrence Berkeley National Lab, 2019-2020.
The goal of this effort is to use unsupervised learning techniques to better understand acoustic events during the superconducting magnet training process.
About 20 sample spectrograms in .csv format. Kindly provided by M. Marchevsky.
Some exploratory work into k-means clustering with data contained in summary files (not the raw spectrograms themselves). Summary files kindly provided by M. Marchevsky.
Jupyter notebooks designed to build and train a 2D convolutional autoencoder to learn features in our unlabeled spectrograms. Several scripts to prepare the data, several scripts to build and train the network, and several scripts to plot and analyze the encoded data produced by the trained network.
To reproduce the workflow presented in a poster at the 2020 ai4science workshop, you can use the notebooks in the following order:
process_march_data.ipynb
data_prep.ipynb
process_post_quench_data.ipynb
post_quench_data_prep.ipynb
build_conv2d_autoencoder.ipynb
train_conv2d_autoencoder.ipynb
save_conv2d_autoencoder.ipynb
plot_conv2d_pca.ipynb
plot_conv2d_spectrograms.ipynb
These scripts and notebooks will not work "out of the box" since filepaths are hardcoded for NERSC's corigpu system and file locations may change, but hopefully the general ideas and techniques we used are clear and may be helpful to others.