Waiting for new ideas about the origin of PR trips, a software mitigation of the problem has been implemented yesterday. The method is based on a few systematic characteristics of the trips:

1) the trip is anticipated by a fast glitch on the accelerometers;

2) the IP displacement along Z, which is the cause of unlock, is proportional to the IP rotation TY. This is a clean signal where one can read quite early the direction of the trip;

3) the trip can be well reproduced by adding a step to the accelerometers in a certain combination. The accordance is very good: very likely the trip is produced by a disturbance which adds a constant to the three signals. In principle, a mechanical disturbance (a tilt) can produce such an a effect on an accelerometer, but a common tilt on the structure where the three accelerometer are placed would produce basically a translation of IP. The large component of rotation seems definitely associated to an electrical disturbance on the sensors.

The fast glitch can be used as a trigger for producing an artificial step to be sent to the accelerometers. The TY signal can be used to know, in less than 1 second, the sign of the step needed to generate an opposite trip. In order to do that, the trigger is used to generate a time window (fig 1), during which TY is compared to a positive and a negative threshold. When a threshold is crossed, the appropriate step is sent. The step generates a trip only in Z combination, the only one which actually needs to be compensated. TY is not compensated, because interacting with the witness could produce a second dangerous crossing of the threshold.

This night two trips occurred: a large one (fig 2) and a small one (fig 3). Comparing the first to a similar past trip, one can see that the amplitude of the Z component has been reduced enough to avoid the unlock. Also the small one did not produce an unlock, but the figure shows that the compensation produced a dangerous trip in the opposite direction (over-compensation). The amplitude of the step has been defined in order to stay about in the middle between large and small trips, but we cannot exclude that a small harmless trip could be compensated so much to produce an unlock.

On this purpose, an improvement of the feed-forward can be studied, because the delay between the fast glitch and the TY crossing of a threshold contains the information to define a good amplitude of the step. Maybe it can be implemented after acquiring some statistics on the compensated trips.

A bug in the feed-forward code has been found and fixed. It was preventing the application of the step in some unlucky situation: that is the reason why the unlock has not been prevented during the latest PR trip.

Figure 1 and 2. I have looked at two example of the PR F0 accelerometer glitches in the raw_full data that are sampled at 10kHz. The glitch time scale is ~1ms.

The glitch is also visible in the COIL channels as pointed out by Didier this morning. But this seems rather like a consequences, as it happens a few ms later and has a longer time scale by a factor few. What are these COIL channels? Are these corrections of a feedback loop, and if yes what is its bandwidth?

The glitch is also visible in the V1 vertical accelerometer, but it has a different shape. It looks like a step. It is not present in the V2 vertical accelerometer. Is there something that is common between the H1, H2, H3 and V1 accelerometer, but different for the V2 accelerometer? For example is there a cable connector common to H1, H2, H3 and V1, but different for V2?

Figure 3 shows that the horizontal feedback of the accelerometer looks like a step response, so it might be that the horizontal accelerometer error signal doesn't look like a step, as the step is removed by the feedback loop, while for the vertical accelerometer it is not removed by the feedback loop keeping the proof mass centered inside the accelerometer.

This morning LVDT modulation frequency of the horizontal accelerometers has been changed (from 49.5 to 48.3 kHz). The demodulation phases has been adapted, the sensing elements have been re-centered, the TFs re-measured and the controllers re-estimated. Horizontal inertial damping gains has been re-measured as well and changed accordingly.

So far we had just one fast gitch on the three PR ACC, not too different from the usual ones, but it was not followed by any trip. Let's wait for the next.

After the modification of PR ACC H modulation frequency, we had 3 fast glitches - the usual rate observed recently - but no one triggered a trip. This is a good news: it seems that the problem has been fixed. To be noticed that the trips disappeared also from the vertical loop, but PR ACC V modulation frequency was not touched.