With lots of help from JamieR, GabrieleV, MaddieW, and AaronV, I think I've maybe gotten a better handle on applying appropriate phase advance / delays to the NonSENS training, for online cleaning that creates the _CLEAN channel.
A caveat is that we haven't had a nicely thermalized IFO this afternoon to try the cleaning on, but I'm hopeful that since I trained everything on a thermalized IFO from last night, that once the IFO thermalizes the subtraction will be good.
For tonight, I am letting the subtraction subtract, so I've set the NOISE_CLEAN guardian to send signal out to the GDS calibration pipeline. I think I've accepted everything in safe and observe sdfs, so that Tony will have no problem going to Observing later.
A caveat is that some of the training needs more time, including this LSC training result attached, showing that I'm expecting I'll be injecting noise below 25 Hz, but cleaning things up above 25 Hz. This is noise injection something that requires improvement over the next few days, and is a work in progress.
Today I updated the Jitter cleaning, with a slightly different time-stamping delay. But, if I just ask it to subract the Jitter noise, it injects noise. Not good. I've accepted for tonight a -1 in the gain on the (updated, but overall the same as yesterday) Jitter noise estimate.
Attached is measured data showing that the subtraction is working (again, with Jitter having an unexplained minus sign, all others having the expected plus sign). The colors on the various contributions to the Noise_Estimate are roughly meant to align with the colors on the nice summary page that Derek updated recently.
The summary page will also be helpful for showing how the cleaning efficacy changes with time (it's not being updated throughout the lock stretches). I still don't think I've got this phase delay thing quite right, so still some room for improvement, but it's at least doing something.
Notes for self:
Jitter was trained with 239e-6 sec hand-added, others trained with 408e-6 s advance hand-added. Need to update other traces, since 239e-6 sec is the value that is used in the calibration pipeline. The 408 number was me misreading things. None of these are trained using the nonsens_firfilt and re-time-stamped with the nonsens_firfilt_delay.
