|
Identifier |
000423989 |
|
h |
Title |
Acoustic signal characterization using hidden markov models with applications in acoustical oceanography |
Alternative Title |
Στατιστικός χαρακτηρισμός ακουστικών σημάτων με χρήση κρυπτο-μαρκοβιανών μοντέλων και εφαρμογές στην ακουστική οκεανογραφία |
Author
|
Σμαραγδάκης, Κωνσταντίνος
|
Thesis advisor
|
Ταρουδάκης, Μιχαήλ
|
Reviewer
|
Μακράκης, Γεώργιος
Τσόγκα, Χρυσούλα
Dosso, Stan
Χαρμανδάρης, ευάγγελος
Σκαρσούλης, Εμμανουήλ
Τζαγκαράκης, Γεώργιος
|
Abstract |
The scope of this PhD thesis is to develop a new probabilistic characterization scheme for
acoustic signals recorded in the marine environment, with applications in acoustical oceanog-
raphy. We will refer to the proposed scheme as Probabilistic Signal Characterization Scheme
(PSCS). The schemes aims at the definition of a set of observables (signal features) that could
characterize a signal to a unique way. To this end, a signal is decomposed into several levels
using the stationary wavelet packet transform. This decomposition provides a time-frequency
analysis of the characteristics of the signal. The stationary wavelet packet coefficients of the
various levels are then modeled by a single left-to-right Hidden Markov Model (HMM) with
Gaussian emission distributions. The concept behind the decision of using a sequential mod-
eling of the signal’s extracted coefficients, was the fact that a signal after propagation through
a dispersive medium such as water column in the marine environment, exhibits evolving
time-frequency characteristics. The association of a signal with a representative HMM is per-
formed by means of the Expectation-Maximization (EM) algorithm. Eventually the signal is
characterized by a set of parameters which describe the HMM.
The proposed signal characterization methods has been applied in inverse problems of acous-
tical oceanography. In particular, problems associated with the retrieval of the marine environ-
mental parameters using measured features of the acoustic field due to a sound source have
been considered. These problems being in nature non-linear are solved with optimization pro-
cedures requiring comparison of the characteristic of the measured acoustic signal with same
of replica signals. In this work the Kullback-Leibler divergence is employed as the similarity
measure of two signals, comparing their corresponding HMMs. To validate the performance
of the proposed characterization scheme, the thesis presents few characteristic test cases in
which simulated and real data have been considered. The measured signals are characterized
by means of the proposed PSCS method and the model parameters of the marine environment
have been estimated by employing a Genetic Algorithm (GA) over three sets of population of
candidate model parameters. The GA leads to distributions of the model parameters of the
vii
final population using Gaussian Mixture Model (GMM). This representation provides the so-
lutions of the inverse problems in the form of the maximum of the marginal densities and a
qualitative indication of the confidence intervals of the recoverable parameters. The results
are compared with those obtained using the Statistical Signal Characterization Scheme (SSCS)
proposed by Taroudakis et al. In addition, the results corresponding to the experimental data
are compared to various approaches from the literature. The applications presented here con-
firmed the reliability and efficiency of the method when applied with typical signals used in
acoustical oceanography.
|
Language |
English |
Subject |
Machine learning |
|
Probabilistic models |
|
Ακουστική διάδοση |
|
Μηχανική μάθηση |
|
Πιθανοθεωρητικά μοντέλα |
Issue date |
2019-07-26 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Applied Mathematics--Doctoral theses
|
|
Type of Work--Doctoral theses
|
Permanent Link |
https://elocus.lib.uoc.gr//dlib/4/4/d/metadata-dlib-1564131596-73724-5318.tkl
|
Views |
1037 |