I applied DeepClean to clean KAGRA strain data from 1432093271 s to 1432004295 s (03:40:53, May 24, 2025 UTC ~ 03:57:57, May 24, 2025 UTC, 1024 seconds in total). The DeepClean models are trained using the witness channels and strain channel from 1432094295 s to 1432096343 s (03:57:57, May 24, 2025 UTC ~ 04:32:05, May 24, 2025 UTC, 2048 s in total).

The frequency bands cleaned are: [298Hz, 302Hz], [322Hz, 343Hz], [348Hz, 380Hz], [418Hz, 422Hz], [440Hz, 450Hz], [462Hz, 472Hz], [533Hz, 542Hz]. The witness channels used for each frequency bands can be found here: JGW-T2516710.

The inspiral range of the strain data was calculated using 84 seconds of data starts from 1432093271 s with fftlength=8s, overlap=4s, fmin=10Hz. and we can see the inspiral range got improved by 0.476% after cleaning.

We are working on the cleaning using Independent Compoment Analysis (ICA) method and also trying to improve DeepClean by introducing coherence ratio as part of the loss function.