Summary:
Using the data I collected yesterday, the POP angular FF filters have been trained. The offline time-domain performance looks (unbelievably) good, online performance will be verified at the next available opportunity(see update).

Details:

The sequence of steps followed is the same as that done for the MCL FF filters. The trace that is missing from Attachment #1 is the measured online subtraction. Some rough notes:

• The "target" channels for the subtraction are the POP QPD PIT/YAW signals, normalized by the QPD sum. For the time that the PRMI was locked yesterday, the QPD readouts suggested that the beam was well centered on the QPD, but the POP QPD (OT-301) doesn't give me access to individual quadrant signals so I couldn't actually verify this.
• I used 64s impulse time on the FIR filter for training. Maybe this is too long, but anyways, the calculation only takes a few seconds even with 64^2 taps.
• I found that the Levinson matrix algorithm sometimes failed for this particular dataset. I didn't bother looking too much into why this is happening, the brute force matrix inversion took ~4 times longer but still was only ~5 seconds to calculate the optimal filter for 20 mins of training data sampled at 64 Hz.
• The actuator TF was measured with >0.9 coherence between 0.3 Hz - 10 Hz and fitted, and the fit was used for subsequent analysis. Fit is shown in Attachment #2.
• FIR to IIR fitting took considerable tweaking, but I think I got good enough fits, see Attachments #3, #4. In fact, there may be some benifit to making the shape smoother outside the subtraction band but I couldn't get IIRrational to cooperate. Need to confirm that this isn't re-injecting noise.

Update Apr 5 1145pm:

• Attachment #1 has now been updated to show the online performance. The comparison between the "test" and "validation" datasets aren't really apple-to-apple because they were collected at different times, but I think there's enough evidence here to say that the feedforward is helping.
• Attachment #5 shows that the POP DC (= PRC intracavity buildup) RMS has been stabilized by more than x2. This signal wasn't part of the training process, and I guess it's good that the intracavity power is more stable with the feedforward on. Median averaging was used for the spectral densities, there were still some abrupt glitches during the time this dataset was collected.
• _{The next step is to do the PRFPMI locking with all of these recently retuned feedforward loops engaged and see if that helps things}.