[Gabriele, Elenna, Jamie]
We used the NonSENS linear subtraction code to fit filters for the HAM1 to CHARD_P feedforward. We are using all four L4C signals (RX, RY, X, Z) and we could obtain a significant improvement bewteen 10 and 30 Hz. See first plot.
Filters are loaded in the FF_X, FF_Z, FF_RX, FF_RY filter banks (see second plot).
Note that the filters obtained by training on older data (November 12) are different from those obtained with today's data. The old filter did not provide optimal subtraction.
We left the filter loaded, but the main FF switch off.
There are several technical details that we had to figure out:
- We train the subtraction with offline data: ASC-CHARD_P_IN1_DQ1 is saved at 256 Hz, while the test point used in the front end subtraction is generated at 2048 Hz. There is a downsampling filter T_down applied to the DQ channel. We measured it direclty online, see third plot. The main effect is an equivalent 6ms delay.
- The feedforward signals are originated in the SEI model (4 kHz sampling), and then sent via IPC to the ASC model (2 kHz sampling). This add a delay equal to one sample of the sender, so 1/4096 s
- We need to compensate for effects 1) and 2) in the NonSENS training. This is done by specifying the "shaping" filter parameter. We need to multiply the witness channels by the downsampling filter, so that both witness channels and target channels are filtered by T_down in the training, and filtered by nothing in the real time frontend. In this way the "ratio" is the same for offline and onlien data. We also need to add the delay in 2) to the offline witness channels, to reproduce what's there in the frontend
- NonSENS naturally train the filter in z-domain, so we had to write a small code snippet to convert the IIR filters to s-domain that can be uploaded to foton
The python code is attached as a jupyter notebook (renamed to .txt to attach it). You'll need to pull the latest version of nonsens from git.ligo.org. The code in the notebook should be enough to retrain when needed.
In general it's not possible to directly convert offline-trained noise subtraction or feedforward to the online case when there are loops involved. One would need to take care of the closed loop response of the loop acting on the target signal, for example by measuring it seprately and using it as part of the "shaping" filter in training.
In the CHRAD_P case, the goal was to achieve noise subtraction at frequencies above 10 Hz, where the loop effect is neglegible, due to the aggressive cutoff in the control flter.
I tested HAM1 FF for CHARD P in nominal low noise again and it works to reduce the noise in CHARD (see plot 1), however there is no appreciable difference in DARM when the FF is on or off. I did an on/off test to determine this: plot 2.
Interesting observation at LLO that the delay between models is not what was expected.