Reducing coherent filtering artefacts in time-domain operational modal analysis

Signals collected for modal analysis are often filtered in order to reduce sensor noise, out-of-band oscillations, or for dealing with closely-spaced modes. This filtering introduces filtering artefacts into the data due to the non-ideal filter response and is well understood for impulse excitation. However, for ambient vibration data filter- ing artefacts become superimposed, preventing their visual identification, and will corrupt the correlation function of the data used in time-domain operational modal analysis (OMA) techniques such as covariance-driven stochastic subspace identifi- cation, and the random decrement technique. This corruption leads to inaccurate, misleading, biased or spurious frequency and damping estimates, with the inaccu- racy increasing for systems with higher damping, or lower signal-to-noise ratios. Counter-intuitively, the error in damping estimates is as large for modes with natural frequencies far from the filter cutoff frequency as for modes which are close to the cutoff frequency.
In this paper, an alternative to filtering for time-domain OMA, trimming of the cor- relation of noise from unfiltered correlation functions, is introduced and tested using 10,000 numerically generated ambient vibration data sets. It has been shown that this technique reduces the mean absolute error in the frequency estimates by over 200% and the mean absolute error in the damping estimates by over 400%. Additionally, a new technique which incorporates fitting of filtering artefacts as part of the modal analysis is introduced for where filtering of ambient vibration data is unavoidable, and is demonstrated using real-world acceleration data collected from a two-way spanning concrete slab subject to footfall excitation.