It is possible that the calibration lines somehow play a role.
/virgoData/PrCali/PermanentLines.cfg:
# nameOut - Amplitude - sinus Frequency - phi0 - offset - Unit - fType - sampleRate - rampTime
PR_SINEWAVE_CH Ca_NE_zMar_l1 2e-5 13.6 0. 0. "au" normal 10000 5

NE bottom-stage DSP:
COMB tyPSDm 9_4Hz.cmb no 3 lin_tyAA

Since the noise is clearly caused by some sort of up-conversion, some noise could be expected around the sum-frequency of 13.6+9.4 Hz. Note that zMar and ty are applied to the same two coils of the marionetta. It would be interesting to switch off these lines or move them in frequency at a moment the bumps around 23 Hz are visible.