This thesis addresses the problem of denoising and imputing missing values in
time series and multidimensional seismic data by means of Hankel low-rank
methods. We present the theoretical framework of structured low-rank
approximation and study the algorithms HLR, MSSA, damped MSSA (dMSSA) and the
weighted (Q,R)-norm approach. A unified experimental protocol is developed,
using RMSE and SNR as performance measures, which enables a fair comparison of
the methods on synthetic 1D signals and 5D seismic data. Particular attention
is devoted to the 5D dMSSA algorithm, where we show that, for high noise
levels and low sampling ratios, it consistently attains higher SNR than
classical MSSA. On the real-world \emph{Australian wines} time series, all
considered methods successfully reconstruct removed months and confirm the
practical usefulness of Hankel low-rank techniques for signal reconstruction
and noise reduction in realistic settings.
|
