Introduction

This app uses convolutional neural networks to classify seismogram data into a positive and negative class. Two competing objectives dictate the model design: (1) reducing the number of false positives and (2) doing early on detection. Hidden Markov Models are considered as an alternative.

Data Processing

• Fixed sizes of arrays
• Created conditionals to ensure size of time array is consistent

Feature Selection

• Used TSFresh to extract features
  • x, y and z abs_energy
  • x, y and z mean
  • x, y and z median
  • x, y and z sum_values
  • x, y and z maximum
• For future work, I could consider using a 2DFFT spectrum as a feature

Model Selection

TBD

Prediction and Results

TBD

Future Work

• Finish getting CNN to work
• Get confusion matrix and accuracy number
• Re-do trainig and validation with smaller windows of data. Check minimum window for high accuracy.
• Do classification for every 4s window, check how early we can detect an earthquake.
• Re-do training with some information about how much more often 'noise' files exist than 'earthquake' file, instead of assuming they are equally likely events
• Make unit test cases
• Get requirements.txt and make Travis integration tests