Seminar by Wojciech Żuławiński

Application of robust statistics for periodicity detection in non-Gaussian signals

Many local damage detection procedures are based on the periodicity detection methods. There are various approaches that use for this purpose the autocorrelation (ACF) measure and its standard estimator called sample ACF. However, they may fail if the considered signal of interest is disturbed by strongly impulsive non-Gaussian noise, as the sample ACF is not resistant to significantly outlying values. Hence, for such signals, we propose to use robust (i.e. much less sensitive to outliers) ACF estimators for the periodicity detection methods in the time-frequency (spectrogram-based autocorrelation maps) and frequency-frequency domains (spectral coherence maps). We also propose robust modifications of other periodicity detection methods, namely the coherent and incoherent statistics, where the Fourier transform is calculated using the M-regression method. The presented methodologies are applied to the simulated and real datasets.