[M. Wade, A. Viets]

I made a new GDS filters file to be installed tomorrow during maintenance.  The main change is an improvement in the high-pass filters and a reduction in latency.  The actuation filters were reduced in length from 6.0s (between 2 separate filters) to 3.5s (all one filter), which in theory should reduce the latency by 2.0s.  The following changes were made to the design of the high-pass filters:

• In the frequency domain, the filters roll off to zero faster.  They previously rolled off to zero from 8 Hz to DC, and now they roll off to zero from 9 Hz to 4.5 Hz, as they did in O2.  Unlike in O2, the filters are windowed in the time domain with a Kaiser window, rather than a Tukey window.  The choice of window makes all the (many orders of magnitude of) difference.
• The inverse sensing path is filtered with an additional high-pass filter at 2048 Hz.  This allows us to apply a longer high-pass filter without unmanageable computational cost.  The model filter is now 1.0s long, and the additional high-pass filter is 2.5s long, making the total filtering length equal in temporal duration to that of the actuation path.  How can this be done without losing information above 1kHz? 
  1. Spectrally invert the high-pass filter to make a low-pass filter with a gain of -1.
  2. Tee the inverse sensing path and downsample one copy.
  3. Apply the low-pass filter to the downsampled data.
  4. Upsample the low-passed data to 16384 Hz.
  5. Add the result, which only contains the content of h(t) below 9 Hz (times -1), to the original h(t).
  Then the model filter is applied at the full sample rate of 16384 Hz.  The model still includes a high-pass filter, in order to attenuate low-frequency noise as much as possible.

The filters file is in calibration SVN revision 6577 at this location:
trunk/Runs/ER14/GDSFilters/H1GDS_1234630818.npz

It was made using the script
trunk/Runs/ER14/H1/Scripts/TDfilters/H1_run_td_filters_1234630818.sh

The parameters file used was from the 2019-01-18 model:
trunk/Runs/O3/H1/params/modelparams_H1_20190118.py

The suggested configuration files for running this on the production and testing DMTs are:
trunk/Runs/ER14/GDSFilters/H1GDS_1234630818.ini
trunk/Runs/ER14/GDSFilters/H1GDS_1234630818_TEST.ini

The attached plots are:
1-2) Bode plots of the filters
3) ASD spectrum comparisons showing GDS-CALIB_STRAIN and GDS-CALIB_STRAIN_CLEAN.  Data was from 2019-01-30.
4-5) Comparisons of the frequency-domain model for the control corrections to the actual effect of the filters on the data.  Data was from 2019-01-30.
6-7) Comparisons of the frequency-domain model for the residual corrections to the actual effect of the filters on the data.  Data was from 2019-01-30.
8-9) Comparisons of the frequency-domain model for the response function and GDS-CALIB_STRAIN / DARM_ERR.  Data was from 2019-01-30.  Note the significant discrepancies.  One possible contributor is the fact that the front-end may not have been using this reference model during the time this data was taken from.  It is likely there are problems with the GDS correction filters as well.
10) A plot of the latency of the pipeline when reading from shared memory on ldas-pcdev1.ligo-wa.caltech.edu for ~4 minutes of data taken earlier today.  This does not include the 1.0s length of the frames.