Long ago, Evan and Haocun measured DHARD P alog 27518.
They determined that the way that DTT estimates OLG TFs is not really sufficient for regions where the excitation is not much much greater than the overall noise in the loop.

At that time, they extracted the IN1/EXC and IN2/EXC manually from DTT, and estimated the OLG like (IN1/EXC)/(IN2/EXC) alog 27617.
They also started using swept sine measurements to avoid bias and to achieve high coherence.
However, swept sines at these frequencies are extremely slow, and the measurements are unwieldly (doesn't have the bandwidth you want, can cause locklosses by shaking hard at a sensitive frequency, forget to save the right A channels).


Instead, I've drawn up a low-excitation long-term measurement infrastructure to estimate unbiased OLGs.
The advantage of this code is it saves the EXC (and any other test points you want) directly to disk, so you can plot whatever binwidth and averages you want for the duration of the measurement.

I stress tested this code by injecting into IMC-MCL_EXC for an hour, which saves three channels (IN1, IN2, and EXC) at fs = 16384 Hz.  
It saved a 458M .hdf5 file successfully.  Not sure what the limitation on the data file size.


Results
For this lock on February 1, the CHARD Y gain had been doubled for stability reasons, to avoid 2.2 Hz ringups alog 67100.
The digital filters had also recently changed alog 67101.

From the 0.1 Hz bandwidth plot (PDF 2), we only have around -18 degrees of phase margin at 2.2 Hz. 
Increasing the gain by a factor of two gave us stability by pushing the UGF from 2.2 Hz to 2.0 Hz, where we have -27 degrees of phase.
I would not have claimed that -18 degrees was enough to cause instability, but the interferometer could not lock without the CHARD Y gain increase.

Additionally, it's possible to see why increasing it further would cause us instability at either 6.5 Hz or 1.4 Hz.

Bounce and roll mode notches are at 9.7 Hz and 13.7 and 14.0 Hz, respectively.

Future
CHARD Y is totally dominating the ASC contribution to the DARM noise in the region of 10-25 Hz.  
Having measurements will help us reduce the loop gain in the 10-25 Hz region to avoid sensor noise reinjection will benefit to our DARM spectrum.

Attached are the current CHARD Y regular and radiation-pressure compensation filters.
It appears there is rapid phase loss in the 10-25 Hz region where the EDH_LP takes over.

Evan notes that the regular and RPC filters add together with a different signs and the same magnitude, which kills our loop gain at 0.7 Hz (last two PDFs).


Excitation smithing
The main consideration for forging you excitation is 
1) achieve a modest amount of coherence at the frequencies you care about,
2) runs for a long enough period of time that we can resolve the OLG with our modest coherence with enough averages,
3) doesn't cause a lockloss.

I usually compare the excitation to the IN1_DQ error signal during full lock, 
and try to shape the excitation to mirror it, times the approximate loop gain to overcome the loop suppression.

You can forge your excitation via foton using e.g. 

quick_psd H1:ASC-CHARD_Y_IN1_DQ -t1 1359315483 -d 1000 -b 0.01 --excitation --excitation_fs 2048

then typing your foton string in the Calibration textbox for the Excitation channel, e.g.

cheby2("BandPass",4,100,0.002,19)gain(2e4)zpk([0.5],[0.01],1.0,"n")resgain(0.5,3,20)resgain(0.95,15,15)resgain(1.25,5,20)resgain(2.15,20,15)resgain(3.5,10,20)

It is critical to include the excitation channel sampling frequency, otherwise the injection strength will be misinterpreted by awg alog 67142.

Code
The code must be run in conda env labutils, and exists in 

/ligo/home/craig.cahillane/Git/IFO/general/scripts


The injection script is 

python chard_yaw_injection_caller.py


This script relies on the class 

autoBandlimitedNoiseSweep.py


I postprocessed this script using 

ipython -i postprocess_hdf5.py ~/Git/IFO/CHARD_Y/data/Injections/20230201/chard_yaw_injection_0p01_10p0_Hz_gps_start_1359339293.hdf5


The injection script cannot be run by others as of now, only because it saves data directly into my own directory.
I will work on putting something convenient in /ligo/gitcommon/ so others can makes use of this.