[Shu-Wei Yeh, Chia-Jui Chou]
We performed noise subtraction using DeepClean on strain data from April 24.
Coherence between the witness channels and the strain channel was evaluated using data from the silent run state on April 24 (1429559610) over a 1024-second interval. Witness channels were selected based on coherence values greater than 0.4 (Figures 1 and 2). The selected witness channels are listed below.
- K1:LSC-REFL_PDA1_RF45_I_IN1_DQ
- K1:LSC-REFL_PDA1_RF45_I_OUT_DQ
- K1:LSC-REFL_PDA3_RF45_I_ERR_DQ
- K1:LSC-REFL_PDA3_RF45_I_IN1_DQ
- K1:LSC-REFL_PDA3_RF45_I_OUT_DQ
- K1:LSC-REFL_PDA3_RF45_Q_IN1_DQ
- K1:LSC-REFL_PDA3_RF45_Q_OUT_DQ
DeepClean was applied at 116 Hz, and the improvement in inspiral range was evaluated. The training segment started at 1429559610 with a duration of 4096 seconds, and the cleaning segment began at 1429563706, also with a duration of 4096 seconds.
We prepared two models, each using different band-pass filter configurations, to test the noise subtraction performance:
-
Figure 3 shows that the model trained with a 114–119 Hz band-pass filter achieved an inspiral range improvement of 0.232%.
-
Figure 4 shows that the model trained with a 113–121 Hz band-pass filter achieved an even better improvement of 0.358%.
We observed that the second model provided the best performance, likely due to a clearer peak in the coherence plot.
We will continue testing coherence and DeepClean performance with the upcoming data.
