My approach to the Kaggle LANL Earthquake competition
The competition involves predicting when the next seismic activity (earthquake) will occur in a laboratory experiment. The training data is a large sequence of acoustic values at a high resolution. Each acoustic value in the sequence is labeled with the number of seconds until the next earthquake. The goal is to train a model to accurately detect when the next earthquake occurs, given very small segments of acoustic values from independent test runs.
My approach implemented in this repo is a deep learning model composed of three parts:
- gru on features extracted from spike segments
- acoustic 1d conv
- spectogram 2d conv
For more details see the python notebook.