NSPECKT: A Non-Parametric Change Point Detector for Gravitational Wave Data

Data analysis algorithms used in searches for gravitational wave (GW) signals are commonly based on the assumption that certain statistical properties of the noise in the data are time-invariant (i.e., stationary noise). However, real data exhibits non-stationarity in various forms over widely varying timescales. The estimation of noise properties from a non-stationary segment can lead to a corrupted calibration of the noise model and can adversely affect the performance of GW signal detection algorithms. Therefore, measuring the noise properties - specifically, the Power Spectral Density (PSD) - with sufficient accuracy requires the demarcation of stationary segments in the data. We present NSPECKT (Non-stationarity in SPECtrogram with Kolmogorov-Smirnov Test), a non-parametric algorithm for segmenting data into stationary and non-stationary regions using spectrogram-based change point detection. We quantify the performance of NSPECKT on real data from the Laser Interferometric Gravitational-Wave Observatory (LIGO) detectors and find that it effectively rejects non-stationarity caused by both instrumental transients ("glitches") and strong GW signals across a wide range (~10 milliseconds to ~100 seconds) of timescales. The latter ensures that the GW signal is not rejected due to accidental inclusion in PSD estimation.