Estimation and Classification of NonStationary Processes: Applications in TimeFrequency Analysis
Research output: Thesis › Doctoral Thesis (compilation)
Abstract
This thesis deals with estimation and classification problems of nonstationary processes in a few special cases.
In paper A and paper D we make strong assumptions about the observed signal, where a specific model is assumed and the parameters of the model are estimated.
In Paper B, Paper C, and Paper E more general assumptions about the structure of the observed processes are made, and the methods in these papers may be applied to a wider range of parameter estimation and classification scenarios.
All papers handle nonstationary signals where the spectral power distribution may change with respect to time. Here, we are interested in finding timefrequency representations (TFR) of the signal which can depict how the frequencies and corresponding amplitudes change.
In Paper A, we consider the estimation of the shape parameter detailing time and frequency translated Gaussian bell functions.
The algorithm is based on the scaled reassigned spectrogram, where the spectrogram is calculated using a unit norm Gaussian window.
The spectrogram is then reassigned using a large set of candidate scaling factors.
For the correct scaling factor, with regards to the shape parameter, the reassigned spectrogram of a Gaussian function will be perfectly localized into one single point.
In Paper B, we expand on the concept in Paper A, and allow it to be applied to any twice differentiable transient function in any dimension.
Given that the matched window function is used when calculating the spectrogram, we prove that all energy is reassigned to one single point in the timefrequency domain if scaled reassignment is applied.
Given a parametric model of an observed signal, one may tune the parameter(s) to minimize the entropy of the matched reassigned spectrogram.
We also present a classification scheme, where one may apply multiple different parametric models and evaluate which one of the models that best fit the data.
In Paper C, we consider the problem of estimating the spectral content of signals where the spectrum is assumed to have a smooth structure.
By dividing the spectral representation into a coarse grid and assuming that the spectrum within each segment may be well approximated as linear, a smooth version of the Fourier transform is derived.
Using this, we minimize the least squares norm of the difference between the sample covariance matrix of an observed signal and any covariance matrix belonging to a piecewise linear spectrum.
Additionally, we allow for adding constraints that make the solution obey common assumptions of spectral representations.
We apply the algorithm to stationary signals in one and two dimensions, as well as to onedimensional nonstationary processes.
In Paper D we consider the problem of estimating the parameters of a multicomponent chirp signal, where a harmonic structure may be imposed.
The algorithm is based on a group sparsity with sparse groups framework where a large dictionary of candidate parameters is constructed.
An optimization scheme is formulated such as to find harmonic groups of chirps that also punish the number of harmonics within each group.
Additionally, we form a nonlinear least squares step to avoid the bias which is introduced by the spacing of the dictionary.
In Paper E we propose that the WignerVille distribution should be used as input to convolutional neural networks, as opposed to the often used spectrogram.
As the spectrogram may be expressed as a convolution between a kernel function and the WignerVille distribution, we argue that the kernel function should not be chosen manually.
Instead, said convolutional kernel should be optimized together with the rest of the kernels that make up the neural network.
In paper A and paper D we make strong assumptions about the observed signal, where a specific model is assumed and the parameters of the model are estimated.
In Paper B, Paper C, and Paper E more general assumptions about the structure of the observed processes are made, and the methods in these papers may be applied to a wider range of parameter estimation and classification scenarios.
All papers handle nonstationary signals where the spectral power distribution may change with respect to time. Here, we are interested in finding timefrequency representations (TFR) of the signal which can depict how the frequencies and corresponding amplitudes change.
In Paper A, we consider the estimation of the shape parameter detailing time and frequency translated Gaussian bell functions.
The algorithm is based on the scaled reassigned spectrogram, where the spectrogram is calculated using a unit norm Gaussian window.
The spectrogram is then reassigned using a large set of candidate scaling factors.
For the correct scaling factor, with regards to the shape parameter, the reassigned spectrogram of a Gaussian function will be perfectly localized into one single point.
In Paper B, we expand on the concept in Paper A, and allow it to be applied to any twice differentiable transient function in any dimension.
Given that the matched window function is used when calculating the spectrogram, we prove that all energy is reassigned to one single point in the timefrequency domain if scaled reassignment is applied.
Given a parametric model of an observed signal, one may tune the parameter(s) to minimize the entropy of the matched reassigned spectrogram.
We also present a classification scheme, where one may apply multiple different parametric models and evaluate which one of the models that best fit the data.
In Paper C, we consider the problem of estimating the spectral content of signals where the spectrum is assumed to have a smooth structure.
By dividing the spectral representation into a coarse grid and assuming that the spectrum within each segment may be well approximated as linear, a smooth version of the Fourier transform is derived.
Using this, we minimize the least squares norm of the difference between the sample covariance matrix of an observed signal and any covariance matrix belonging to a piecewise linear spectrum.
Additionally, we allow for adding constraints that make the solution obey common assumptions of spectral representations.
We apply the algorithm to stationary signals in one and two dimensions, as well as to onedimensional nonstationary processes.
In Paper D we consider the problem of estimating the parameters of a multicomponent chirp signal, where a harmonic structure may be imposed.
The algorithm is based on a group sparsity with sparse groups framework where a large dictionary of candidate parameters is constructed.
An optimization scheme is formulated such as to find harmonic groups of chirps that also punish the number of harmonics within each group.
Additionally, we form a nonlinear least squares step to avoid the bias which is introduced by the spacing of the dictionary.
In Paper E we propose that the WignerVille distribution should be used as input to convolutional neural networks, as opposed to the often used spectrogram.
As the spectrogram may be expressed as a convolution between a kernel function and the WignerVille distribution, we argue that the kernel function should not be chosen manually.
Instead, said convolutional kernel should be optimized together with the rest of the kernels that make up the neural network.
Details
Authors  

Organisations  
Research areas and keywords  Subject classification (UKÄ) – MANDATORY
Keywords

Original language  English 

Qualification  Doctor 
Supervisors/Assistant supervisor 

Award date  2019 Jun 14 
Place of Publication  MediaTryck, Lund 
Publisher 

Print ISBNs  9789178951338 
Electronic ISBNs  9789178951345 
Publication status  Published  2019 May 17 
Publication category  Research 
Bibliographic note
Defence details
Date: 20190614
Time: 09:15
Place: MH:Rieszsalen, Matematikhuset, Sölvegatan 18, Lund University, Faculty of Engineering LTH
External reviewer(s)
Name: Karlsson, Johan
Title: Associate Professor
Affiliation: KTH Royal Institute of Technology, Stockholm, Sweden.

Total downloads
No data available
Related research output
Johan Brynolfsson & Maria Sandsten, 2018 Sep 1, In : Signal Processing. 150, p. 2032 13 p.
Research output: Contribution to journal › Article
Johan Brynolfsson & Maria Sandsten, 2017 Oct 23, 25th European Signal Processing Conference, EUSIPCO 2017. Institute of Electrical and Electronics Engineers Inc., p. 326330 5 p. 8081222
Research output: Chapter in Book/Report/Conference proceeding › Paper in conference proceeding
Johan Swärd, Johan Brynolfsson, Andreas Jakobsson & Maria Sandsten, 2016, In : IEEE Transactions on Signal Processing. 64, 7, p. 17981807
Research output: Contribution to journal › Article