We performed noise subtraction using DeepClean on the strain data on 2025/03/18.
The witness channels used are:
- K1:IMC-IMMT1_TRANS_QPDA2_DC_PIT_OUT_DQ
- K1:LSC-REFL_PDA3_RF45_I_ERR_DQ
The coherence between these channels to the strain channel from 1426341931 s ~ 1426346027 s (4096 s) around 128Hz is shown in Fig.1 and Fig.2.
We trained 3 models focusing on different frequency bands using the data from 1426341931 s ~ 1426346027 s and cleaned the data from 1426346027 s ~ 1426350123 s (4096 s). Here are the frequency bands of these models:
- model 1: [118 Hz, 135 Hz]
- model 2: [110 Hz, 156 Hz]
- model 3: [110 Hz, 135 Hz]
The cleaning results of model 1 are shown in Fig.3 ~ Fig.5 (the ASD of the strain channel, the coherence between the witness channels and strain channel before and after cleaning). The results of model 2 are shown in Fig.6 ~ Fig.8 and the results of model 3 are shown in Fig.9 ~Fig.11.
We can see that the performance of model 3 is the best. The coherence spectrum of K1:LSC-REFL_PDA3_RF45_I_ERR_DQ channel at the training time is very different from the coherence spectrum at the cleaning time, but we can still decrease the coherence between this channel and the cleaned strain channel.
Once the detector is in lock again, we will proceed the same cleaning procedure and see if we still need to subtract beam jitter noise around 128Hz using DeepClean in O4c.
