Today between 19:25 UTC and 20:45 UTC I measured the CHARD_Y open-loop gain with two broadband noise injections (different shapes). The result is a pretty good measurement that will enable a better design of the CHARD_Y controller.
Right now: the controller has little gain at 2.6 Hz, and a very large (10 dB) gain peaking at 1 Hz.
The template is available in /ligo/home/gabriele.vajente/ASC/CHARD_Y_shaped_exc_2023_07_31b.xml
Here's a fit of the CHARD_Y plant (open loop gain divide by controller, including gain 60). zeros and poles in s-domain below (rad/s)
z = [-4.27268475+16.14989614j, -4.27268475-16.14989614j,
-0.84677811+11.7910939j , -0.84677811-11.7910939j ,
-0.15001959 +3.1577398j , -0.15001959 -3.1577398j ,
-0.14073526 +2.60978661j, -0.14073526 -2.60978661j,
-0.89432965 +1.06440904j, -0.89432965 -1.06440904j]
p = [-1.26172189+18.71400182j, -1.26172189-18.71400182j,
-0.42125296+16.04217932j, -0.42125296-16.04217932j,
-1.88894857+14.52668064j, -1.88894857-14.52668064j,
-0.17624191 +6.42823571j, hisproduces -0.17624191 -6.42823571j,
-0.08126954 +3.12803462j, -0.08126954 -3.12803462j,
-0.33177813 +2.74361234j, -0.33177813 -2.74361234j,
-1.01199783 +0.j , -0.20377166 +0.j ]
k = -2608.840762444897
Using this model too predict the closed-loop gain underestimates a bit the gain peaking observed at about 1 Hz. So I did a refinement of the fit to better model the little gain and phase wiggles at 1.4 Hz:
z = [-4.27268475+16.14989614j, -4.27268475-16.14989614j,
-0.84677811+11.7910939j , -0.84677811-11.7910939j ,
-0.15001959 +3.1577398j , -0.15001959 -3.1577398j ,
-0.14073526 +2.60978661j, -0.14073526 -2.60978661j,
-0.89432965 +1.06440904j, -0.89432965 -1.06440904j,
-0.55930335 +8.32753387j, -0.55930335 -8.32753387j,
-0.13681604 +6.53617672j, -0.13681604 -6.53617672j,
-0.75698522 +6.05299388j, -0.75698522 -6.05299388j]
p = [-1.26172189+18.71400182j, -1.26172189-18.71400182j,
-0.42125296+16.04217932j, -0.42125296-16.04217932j,
-1.88894857+14.52668064j, -1.88894857-14.52668064j,
-0.17624191 +6.42823571j, -0.17624191 -6.42823571j,
-0.08126954 +3.12803462j, -0.08126954 -3.12803462j,
-0.33177813 +2.74361234j, -0.33177813 -2.74361234j,
-1.01199783 +0.j , -0.20377166 +0.j ,
-0.53443931 +8.46054777j, -0.53443931 -8.46054777j,
-0.12110921 +6.58734716j, -0.12110921 -6.58734716j,
-0.59423727 +5.96287439j, -0.59423727 -5.96287439j]
k = -2557.795730924652)
This reproduces pretty well the closed loop transfer function. So we can be confident that it will model well the performance of a new controller design. We'll include the refined fit in the model
Here's a compariison of the fit from the previous measurement and from the new measurement from today. They are quite different, and that explains why the previous controller design failed
Based on the measured CHARD_Y plant, I designed a new controller that would give us more supression at 2.6 Hz, and avooid gain peaking at 1 Hz. This comes at the price of about 10 times large noise injection at 10 Hz and above. This might be too much at 10-15 Hz, but ok above 15-20 Hz. Hopefully once the new controller is engaged and we have a more stable CHARD_Y, we should be able to fine tune the A2L to gain back some noise coupling at frequencies below 20 Hz.
Comparison of the open-loop gain in the same three configurations
Comparison of the loop suppression in the same three configurations
The new controller can be engaged with the nominal gain of 60, and it is stable for an increased gain of 180. Even with a gain of 60, it removes completely the gain peaking at 1 Hz, and provides 3 dB of suppression at 2.6 Hz.
