I have finally fitted the UIM dynamics that we measured in the run-up to my previous aLog. The ZPK file I have produced will be propagated through the tagged suspension models as an update of the UIM dynamics that had been coded in there previously. This will result in the generation of a number of extra tagged models to aid in making the needed foton filters and for analysis in pyDARM.
The fit is done on data generated after the response model of the UIM is removed from the initial data, hence the data can be directly interpreted as the difference from the current model.
The fit (2019-11-20_H1SUSETMX_L1_AllMeas_vs_Fit.pdf) is far from perfect after 200 Hz, and is largely representative from that point on. Above 500 Hz, there is reason the suspect that the TF continues to go up (i.e. we are not sitting on a 1/f^6 slope), but obtaining data there is difficult. I have this fit roll off to a flat level, such that I do not create enormous discrepancies at higher frequencies. The idea is that I can just slap this on top of the current mode without doing anything revolutionary.
Scripts to produce have been placed:
/ligo/svncommon/CalSVN/aligocalibration/trunk/Runs/O3/H1/Scripts/FullIFOActuationTFs
process_actuationmeas_UIMHFDynamics_LF_20191120.py (Low frequency part only)
process_actuationmeas_UIMHFDynamics_FULL_20191120.py (Assembling full ZPK filter through a number of rounds of fitting and some hand adding of zeroes)
Outputted files are placed in:
/ligo/svncommon/CalSVN/aligocalibration/trunk/Runs/O3/H1/Results/FullIFOActuationTFs
where they are named with appended date 2019-12-20_*
The useful file being: (values from here will go into matlab for making a new model)
2019-11-20_H1SUSETMX_L1_ZPK_Final.mat
In doing this I learnt much on how to wield iirrational:
- Results of complicated fits can vary wildly depending on initial conditions, including time when you run it (yes: running a fit one day, can give a slightly different result another day, if the fit you are fitting something to a very high order)
- Fitting high-Q features requires editing the source code of the master copy and reinstalling it
- in /v2/data2filter.py: Change "Q_rank_cutoff" to much smaller values as needed (it appears in a number of places where its used in different steps). values as low as 0.01 can force the fitting algorithm to try to fit just about any feature (whether it is real or not).
- I manipulate the uncertainty of some of my results to force iirrational to fit certain data with more emphasis than other (by reporting different SNR values to iirrational) - though I had mostly mixed results, I think its a useful trick.
- Although I desired the effect of the option "relative_degree = 0" (in combination with "never_unstable_zeros = True" and "never_unstable_poles = True"), ( this forces the end result to not have any 1/f^n features outside of the band where I had data), doing so created frightening mis-fitting in the 200-500 Hz region where the algorithm tries to step up in magnitude response and refuses to fit the data there (even with the trick in the previous bullet point). This is why I hand to had fit in my last step.
Going forward I'll be making new tagged models and propagating my change through to tagged models and to a new foton filter for the L1 stage in the H1CALCS filterbank.
