O.I Traore, L. Pantera, N. Favretto-Cristini, P. Cristini,
S. Viguier-Pla, P. Vieu (2017)
Time series analysis by singular spectrum methods: Application to
the
processing of acoustic emission signals.
Anglo French Physical Acoustic Conference, Marseille, 23-25 janvier 2017
- Keywords
- singular spectrum analysis; acoustic
emission; structural changes detection; signal denoising
Abstract :
Several methods using the Fourier transform or the time-frequency
decomposition can be used to
analyze and process acoustic emission (AE) signals. Most of them are
based on specific assumptions
on the stochastic behavior of the source signal or the noise. For
example, in a denoising problem,
a method like the spectral subtraction assumes that the noise is a
stationary random process. In
general, since acoustic emission data are closely linked to the
experimental protocol that allows
their recording, their stochastic behavior also depends on this
experimental protocol. Therefore,
it is relevant to test processing methods that make no assumption on
the stochastic behavior of
the noise or the source signal and are potentially efficient for any
signal pattern.
The Singular Spectrum Analysis (SSA) method has received increasing
attention since the early
nineties. Recently it has been successfully applied to various topics,
for instance in geophysics and
economics. Unlike most methods for time series analysis, SSA needs no
statistical assumption on
signal or noise, while performing analysis and investigating the
properties of the algorithm. By
using a decomposition of the signal into the sum of a small number of
independent and interpretable
components, SSA allows to perform various tasks such as extraction of
specific components from
a complex signal (noise, trend, seasonality ...), detection of
structural changes and missing values
imputation. To our best knowledge, very few works have explored the
ability of the SSA to analyze
and denoise AE signals. This is the main objective of this work.
Several tools based on the classical SSA are tested. The results
obtained with simulated data
match with those obtained with real data from nuclear safety
experiments. In both cases, analysis
of the heterogeneity matrix (H-matrix) leads to the identification of
the main components of the
corrupted signal, even for low signal-to-noise ratio. The H-matrix
also allows the estimation of the
correlation (in terms of heterogeneity) between these components. For
signal denoising purposes,
the SSA leads to an excellent estimation of the source signal when the
separability between the
noise and the source signal is such that the weighted correlation
tends to zero. However, denoising
becomes more difficult with increasing weighted correlation. We
observe waveform distortions of
the source signal and some artifacts are created when the noise and
the source signal share some
frequency range.
Résumé détaillé