This alog has been edited.

I think that we have understood one of reason that we have had trouble with the stability of our DARM loop over the last few months, which I've tried to explain in a dcc note: T1900148.  The summary is that we are using length to angle decoupling, and because our spot positions are far off center we also are using gains of about 5 in the angle to length decoupling on the ETX PUM (a DARM actuator) to avoid having the angular drive show up in length.  Since the output of the length to angle decoupling doesn't go through the angle to length decoupling filter, the output of that filter couples to DARM through the spot mis centering, which has been large enough to cause instabilities and calibration problems.  

• There is a systematic error in the calibration model of the PUM actuator, 47979, which is not seen when the measurement is made bypassing the drivealign matrix (driving at test)
• The DARM loop has been unstable at 8-9 Hz with the L2P and L2Y decoupling filters off, 47824  This seems likely to be a crossover instability similar to what we might have seen when the PUM gain was lowered to 15 over the weekend: PUM gain lower model

There have been several problems with the stability of the DARM loop over the last few months that might be explained by this, including the 4.2 Hz instability, which was initially solved by adding a boost to the PUM lock filter (47164) but has come back in the last few weeks, There have also been several attempts to change the DARM loop in ways that should have been stable according to the model but weren't, this effect is probably contributing to that difficulty.

This morning Niko and Jenne have been struggling with 4.2 Hz locklosses again, so we took a guess that turning off the L2P and L2Y might help fix the problems at 4.2 Hz, and that turning off the boost in the PUM stage would give us more phase margin for the crossover around 8 Hz and might help us avoid the locklosses we've had with the L2A off.  So far it seems that this is working.  If we can run like this it would make the calibration more accurate and make understanding our DARM loop easier. 

The two most obvious ways to get rid of this parallel DARM actuation path are to set the A2L gains to zero, and turning off the L2A decoupling.  We don't want to set the A2L gains to zero because we do not want the spot to be centered on the optic (power recycling gain, point absorbers). We would expect that turning off the L2A decoupling will increase the size of the ASC control signals.  We have never had accurate measurements for the length to angle decoupling above 2.5 Hz, these filters are rolled off at around 10 Hz. 

Some other conclusions from thinking about these cross couplings more: (updated)

• Making the L2A filters more accurate to higher frequencies doesn't get rid of the cross coupling to length. Rolling the filters off faster (even if that makes them less accurate) would help.  
• If the output of the length to angle filters are summed into the angle drives at the input to the drivealign matrix, this would reduce the unintended length actuation at frequencies where the angle to length decoupling is accurate.  This becomes more important the further our spots are from the center of the opitc. 
• We could copy the L2L filter into the L2A banks, so that it can just be updated if the L2L filter is changed with out revisiting the decoupling.  (Not doing this might be the reason we have had difficulty any time that we have tried to make changes in the L2L filters).
• If we want to do frequency dependent angle to length decoupling, and continue to be able to move the spots around, we could use two parallel filters in the place where P2L and Y2L are currently.  (some modeling might be helpful to decide how important this is).

All models were using the OBSERVE SDF reference except for h1seiproc and h1iopoaf1. These are new models and their SDF needed some TLC.

For both:

I moved the safe.snap out of target into userapps and put them into SVN, I then symlinked them to the target area.

I created OBSERVE.snap as a symlink to the userapps safe.snap for now. If h1seiproc needs a separate safe and OBSERVE snap, we can create the new file and re-link in the future.

Executive summary: H1 CW hardware injection recovery is looking good after the most recent DARM sign flip last week. L1 recovery is also good, except for an understood timing offset in the injection actuation.

Figures 1 and 2 are summary plots for the 18 CW injections (ranging from 12 Hz to 2991 Hz) from Jessica Leviton's monitoring infrastructure. Only the (admittedly optimistic) injection at 12 Hz shows significant discrepancies between injected and recovered parameters. Note that we have a deliberate mix of "loud" and quiet injections. In this context, loud means clearly visible in one day of data using coherent integration.

For this last period of ER14 (March 22 onward), the monitoring is found at this link (looks at low-noise data: H1:DMT-GRD_ISC_LOCK_NOMINAL): 

https://ldas-jobs.ligo-wa.caltech.edu/~jessica.leviton/Hardware_Injections_end_ER14.html

but an O3 version using observing-mode data as of April 1 will be set up next week.

Also shown below (Figures 3 and 4) are summary plots for L1 data for a longer time baseline (March 2 onward), where there is, again, good agreement between intended and recovered parameters, except for a phase discrepancy that grows linearly with frequency that is consistent with a fixed time delay in actuation of two 61-usec clock cycles, which is expected for the O2-era actuation function being used there for CW injections. The timing discrepancy is not seen for H1 because Evan Goetz recently updated the actuation function used in H1 injections. 

The L1 monitoring infrastructure can be found here: (again, this location will change as of next week)

https://ldas-jobs.ligo-la.caltech.edu/~jessica.leviton/Hardware_Injections_ER14.html

Attachments:

Figure 1 - Summary of amplitude and phase agreement for injections:
graph 1 - Injected (red crosses) and recovered (blue dots with errors) injection amplitudes vs frequency
graph 2 - Fractional amplitude agreement (injected - true) / true with red line at zero showing the ideal
graph 3 - Phase offset between recovered and injected with red line at zero showing the ideal

Figure 2 - More detail on phase agreement:
graph 1 - Same as graph 3 above, but with linear frequency scale
graph 2 - Same information with phase discrepancy converted to a time offset 
graph 3 - Same as graph 2 but zoomed in vertically 

Figure 3 - Same as Figure 1, but for L1

Figure 4 - Same as Figure 2, but for L1

Kissel is running a broad band calibration sweep right now, expected to run for roughly an hour.

Out of OBS for some other settings anyways...

After an sdf change to H1:CAL-INJ_TINJ_OUTCOME at 13:15 took us out of Observing, we are unmonitoring the attached channels.

The following DIAG_MAIN warning:

"SERVO_BOARDS: LSC-REFL_SERVO_COMEXCEN is ON"

is the result of the changes made by commissioners in order to monitor the TCS adjustments with "additional lines" that will be turned off before O3.

These injections are scheduled to begin at 05:00:00 PDT 29 March 2018, the first injection should occur at 05:00:10 PDT 29 March 2018, and the entire excitation should last only 91 sec.

The expected GPS times of all these injections can be found in GraceDb via: https://gracedb.ligo.org/search/?query=H1+Test+HardwareInjection+gpstime%3A+1237896018+..+1237896120+HWINJREQ&query_type=E&results_format=S

These injections are scheduled to begin at 02:00:00 PDT 29 March 2018, the first injection should occur at 02:00:10 PDT 29 March 2018, and the entire excitation should last only 91 sec.

The expected GPS times of all these injections can be found in GraceDb via: https://gracedb.ligo.org/search/?query=H1+Test+HardwareInjection+gpstime%3A+1237885218+..+1237885320+HWINJREQ&query_type=E&results_format=S

These injections are scheduled to begin at 23:00:00 PDT 28 March 2018, the first injection should occur at 23:00:10 PDT 28 March 2018, and the entire excitation should last only 91 sec.

The expected GPS times of all these injections can be found in GraceDb via: https://gracedb.ligo.org/search/?query=H1+Test+HardwareInjection+gpstime%3A+1237874418+..+1237874510+HWINJREQ&query_type=E&results_format=S

I have started another, somewhat slower IMC VCO sweep, in the frequency region around 78.8 MHz.  Andy's plot in alog 47970 indicates that this is our best candidate for a whistle-free region. 

I started moving the offset slider at about 19:16:50 UTC to get my starting place where I wanted it, a little low of 78.8 MHz.  Then at about 19:23:35 UTC I started the script. 

Picking upper and lower limits of the clean region of Andy's plot by eye, I'm sweeping the VCO between 78.8 MHz (-3.37 V) and 78.95 MHz (-1.48 V). 

As with last night, this is a very slow change, and today we should be out of the whistle region for almost all of the sweep time, so we're leaving the IFO in Observing (with Keita's permission).

J. Kissel

SDF is still having problems comparing / checking very small or very large EPICs records. This is usually exposed in the calibration group's "epics records" i.e. DARM loop model parameters at a reference time used to compute time-dependent correction factors, which are typically very small or very large -- in the 10^-17 levels or 10^+18 levels (see attached). It's unclear *when* exactly the claimed differences happen (maybe when the .snap comparison file is changed from safe to OBSERVE? Or maybe because when you *accept* the diffs, they're written at a lower precision than that comparison is made?), but it's something we're having to constantly reconcile before every observation stretch these days. 

A series of DetChar safety injections is scheduled to go into LHO at 05:00:00 PDT tomorrow morning (28 March 2018). The injection file starts at 05:00:00 PDT, lasts for 91 seconds, and the first actual injection should go in at 05:00:10 PDT.