Automated Structure Detection in Non-Stationary Time Series Using Multi-Statistic Joint Analysis

We present a multi-statistic joint analysis method for automated detection of structured signals and state transitions in non-stationary time series. The method employs four independent metrics—coefficient of variation (CV), sample entropy, envelope skewness, and recurrence rate—to perform adaptive frequency-band scanning and output five classification labels: signal, white_noise, nonstationary, segmented, and signal_low. We validate the method on synthetic signals, sunspot records, and public LIGO O1/O4c1 data. Results show that the method automatically detects segmentation points (error < 1 s) and distinguishes detector noise states across observation periods. The method requires no pre-defined model and is suitable for automated screening of weak, non-stationary, single-channel time series.