[L. D'Onofrio, G.Vajente]

Using Band Limited RMS [1][2], we tracked the variation of the detector noise during O3 data.
 
There are evidences that part of the variation in the BLRMS time series in the frequency band 1120-1400 Hz is linked to channels that belong to PSL subsystem.  
Our model does not explain all the variation of the BLRMS segment.
 
To find correlations between the variation of noise and the auxiliary channels (we chose 10 subsystems, around 80 thousand of channels) we used LASSO (for more details, see [3].
To quantify the “importance” of channels selected by LASSO we defined the ε score that is the normalized residual computed for single channel (0< ε <1; ε ∼ 1 means no correlation).
 
We analyzed different BLRMS segments in the last four months of O3 data.
We obtained the best reconstruction for the 22-hour BLRMS segment starting from 1249000652 GPS time (Aug 05, 2019 00:37:14 UTC). This segment is splitted in two equal parts; the first part is the training set for LASSO and the second part, the test set. In Figure 1, the top plot shows the predicted values (in orange) using the 4 channels selected by LASSO while the bottom plot, the residuals for the test set.  
 
The selected channels are: 

• H1:PSL-FSS_PC_MON_OUTPUT (ε score : 0.42)
• H1:PSL-FSS_PC_MON_OUTPUT (standard deviation minute trend) (ε score: 0.58)
• H1:PSL-FSS_PC_PP (ε score: 0.77)
• H1:PSL-PMC_OSCILLATION_MON (ε score: 0.86)

 
In the shown BLRMS segment the residuals are close to zero and the reconstruction is good. This does not happens for all the analyzed segments. Further analysis is required.