Doing a similar fit to the BRMS around 50 Hz yields a time constant of around 1 hour, see fig 2. Assuming that this is some thermal effect driven by the beam-power, we can set t=0 at the moment the dark-fringe is reached and obtain an amplitude of around 7 fringes. If this would be caused by some sort of etalon, it would correspond to a change of optical path length of 3.5 micron.
We suspected that the noise is is caused by some object on the detection bench, whose amount of scattered light might be modulated by a spurious etalon. For an etalon of a piece of glass, the change in optical path-length is given by DeltaOPD = L * dn/dT * DeltaT = n_fringes * lambda / 2. Using a typical dn/dT of 1e-5, we are thus searching for an object for which L * DeltaT ~= 0.4 during a typical lock. This would require 4 degrees for an object of 10 cm (this excludes the main mirrors, which also have a much longer time constants) or 40 degrees for an object of 1 cm (excluding e.g. the prisms of the OMC). The only objects that might be heated enough are some components in the injection system, but it is hard to explain how they introduce the noise. An alternative explanation might be the interference between two pieces of optics separated by some long piece of aluminum, like the suspended bench itself that might be heated by scattered light. Other ideas welcome ...
Apart from reducing the light on the dark fringe and aligning the bench as good as possible (which does not kill the bumps totally), the problem might be solved by reducing the amount of seismic noise. The 100 Hz line might be hard to remove (unless it is a power-supply attached directly to the tower, to be checked). We should spend some time to find the source of the 130 Hz line and try to cure it.
It is possible that we can reduce this vibration by detaching the fans from the pump body, similarly to what done for other pumps, but NOT yet done for the detection pump. This isolation operation should be not difficult, to check if it can be managed during next Tuesday mantenance break.