Summary:
The core functionality of the Matlab DARM model has now been replicated in Python. Attached are figures in a single PDF file showing the primary results. By eye, these look reasonable. A more detailed study comparing to the Matlab model is forthcoming. I also replicated a study made for the L1 detector by Joe B. (see LLO aLOG 29622). I propose to call this code pyDARM.
Details:
I have ported most of the core functionality of the DARM model from Matlab into Python. So far, I have done spot checks by eye to make sure that the results look sensible. I have not yet done a detailed study comparing to the Matlab version, but will do so soon.
I have produced plots showing:
1) DARM digital filters
2) Sensing function
3) Actuation function
4) Open loop gain
5) Frequency dependent actuation authority of each stage compared to inverse sensing
6) Ratio of each stage to the overall calibration
As can be observed, the scale of each figure appears reasonable (comparing with, e.g., G1700316), the OLG is stable, and with a UGF with the correct value (by eye). The contribution of each suspension stage is closely matching L1's results using the Matlab model (see LLO aLOG 29622).
What was done:
- Write python version of Matlab functions to parse Foton filter files and to compute IOP downsampling filters from RCG code coefficients
- Exported numerical values of zeros, poles, gain, delay from analog AA and AI models (these are objects in .mat files)
- Exported ASCII file of the frequency response for the suspension force-to-length for each stage. This is read in and used in the python DARM model, and so far can only be at specific frequency points
- DARM filter bank digital filters computed, sensing function and actuation functions are computed from parameters
- Intermediary data products can be accessed
- Code structure is "flatter", meaning less jumps between different functions/files. Hopefully this makes the code more accessible and readable.
Required python modules (so far): scipy (e.g., filters), numpy (e.g., arrays/array math), collections (for namedtuples), matplotlib (for plotting)
Quirks found along the way:
- You can't directly multiply or add filter objects together in Python (the + and * functions are not overloaded in Python). I had to code up my own version to 1) add filters by using polynomials from roots, computing a transfer function filter, and then converting to a zpk object, all using scipy built-in functions; and 2) multiply filters by appending zeros together, poles together, and multiplying gain.
- The scipy sos2zpk() function is not exactly like Matlab's version. I found an extra zero and pole when converting because Matlab removes any zeros in the 3rd and 6th positions of an sos section before computing a filter from the sos coefficients
To do list (short term):
- More detailed study comparing Python with Matlab models
- Read in a config file
- Make an L1 model to check for any differences
- See if there is a way to read Matlab .mat file and objects therein. This didn't look trivial when I tried at first, which is why the analog AA and AI models were exported as well as the frequency response of each suspension stage
- Address how to get the force-to-length transfer function for each stage for arbitrary frequencies
- Add computations for GDS / DCS pipelines
Longer term:
- Hook this into a pipeline from measurement to model to uncertainty estimate pipeline
The IMC is locked with a build up slightly lower than what we had before this vent. On April 26th we had 165 counts on MC2_TRANS_SUM with 2 Watts input, we now have 67 counts with 0.85 Watts of input power. Some of the things that we ran into:
Craig and I re-phased the length demodulator as his comment says, and flipped the sign. We then added 219 degrees to the IMC WFS demod phases. We found that the LSC power normalization was not up to date (probably a consequence of using the rotation stage medm rather than the guardian to set the power) the gain of the IMC WFS was very low. We also reset the dark offsets for the MC WFS. SDF screenshot attached.
Keita found that there was angular feedback to M2 on MC2, which I don't think is correct. This was set this way on Feb 17th and accepted in SDF, we set it back to M2 off and accepted that in SDF.
We set the mode cleaner to offline for a jitter measurement using the IMC WFS DC starting at 1:04 UTC. The PSL is in science mode and there is 8.9 Watts injected.
On a small detour, I made some edits to the state generator for the OFFLOAD_ALIGNMENT_MANY state, bug fixes from the guardian re-work. The motivation for this is that the offload MCWFS script doesn't turn off the loops so it misaligns the IMC which the slow MCWFS loops take a while to recover from. I had intended to write a guardian state that uses this generator, however, this generic offload state relies on having the alignment intergators in the suspension top masses, which is not the case for the MCWFS. We also would need to edit the state to deal with the IMC PZT. There were a few bugs to work out with the generic offload script, I added fast_ezca to the list of things that are imported by ISC_GEN_STATES, and removed two variables which were just renamed in the state ie:
def gen_OFFLOAD_ALIGNMENT_MANY(dof,tramp,optics):
class OFFLOAD_ALIGNMENT(GuardState):
request = False
redirect = False
tramp = ramptime
optics = opticList
became
def gen_OFFLOAD_ALIGNMENT_MANY(dof,ramptime,opticList):
class OFFLOAD_ALIGNMENT(GuardState):
request = False
redirect = False
Attached are data taken with the IMC unlocked with 8.5Watts input power, compared to the data attached to 39434 which was taken with the HPO and the IMC unlocked. Both are calibrated into beam diameters (normalized pit and yaw *sqrt(pi/8)). The 70W amplifier has lower jitter below 50Hz, but not above. This might be due to the high noise on the PSL table and periscope and the floor that Terra and Anamaria have pointed out (41707 and comments).
I did a quick summary page look at the HAM1 ground motion and PSL table motion from November jitter data timeframe and today. I also compared with LLO (used PSL table data from before install started, so August). Floor motion at LHO is slightly lower in the hundreds of Hz region now compared to Nov.
first plot: HAM1 ground, LHO on left, LLO on right
second plot: PSL table motion, LHO on left, LLO on right