State Switching to Maintain Interferometer Lock
Goal of project: To learn at what times or situations we should switch into more robust interferometer controls, so we don't lose lock, and maintain the thermal state of the interferometer. This may mean temporarily noisier controls, but will reduce thermal settling time since we won't have to re-lock, so will overall increase science mode time.
Finding Patterns
- Masha's neural network classification will be round 1 of this.
- Larry suggests looking into Kari's SPR classification method.
- Jan thinks reinforcement learning could be a useful method to classify reasons for lockloss.
- Determine minimum set of sensors needed to tell you about a pattern (either lockloss or state-switching trigger).
- For seismic, use directional information as well as frequency bands - can watch source (ex. truck) move.
Implementation Tests
- Install on MC, or Arms at 40m.
- We can create a model to do a little testing to make sure algorithms work, but must actually test on the real system to get useful info out.
- Test switching states automatically to prevent lockloss.
- Allow false alarms - we'd rather keep lock to maintain thermal state than be sure an event really happened.
- If an event happened with 75% or 80% confidence, switch into robust mode.
- Eventually test switching based on other environmental sensors, not just seismic.
- Use real-time data from USGS for teleseismic events - Jan is working on getting this for us.
- Round 1, pre-define various robust control states. See, ex. Brett's thesis.
- Eventually let the machine learn what a robust state should look like, design robust controller.