The
Hilbert diagnostics of phase structures is the important result of a fruitful
synthesis of methods developed in optics and radio engineering [1, 2].
This is an integral operation that signal energy redistributes in a given band
of spatial frequencies of the probing field perturbed by the medium under
study. The energy loss of the optical signal is minimized in this case. The
Hilbert transform in the frequency space is reduced to a certain type of phase
transformation of the signal Fourier spectrum with energy conservation in a
wide frequency band.
The
Hilbert diagnostics results of reacting jets were previously presented in [3, 4].
The possibility of polychromatic Hilbert visualization of phase optical density
fields with temperature profile measurement in selected sections of the medium
under study is shown using the example of an axisymmetric hydrogen-air
diffusion flame and a candle flame using the Abel transform. Iterative
selection of radial temperature profiles fitted by Bezier curves, followed by
spatial structure calculation of the refractive index and phase function, is
performed. Comparison of hilbertograms obtained in the experiment with
hilbertograms modeled in the approximation of axial symmetry of the flame is a
criterion for the research results reliability.
The
Gauss-Newton method application [5, 6] is proposed to iterative algorithm
optimize for reconstructing the phase function in Hilbert diagnostics in this
paper. This is an iterative numerical method for solving the least squares
problem, a Newton's method modification for finding the objective function
minimum. It does not require the second derivatives definition, unlike Newton's
method, which greatly simplifies and reduces the calculations number [7].
Various modifications of the Gauss-Newton method that increase the convergence
rate and reduce of the ill-posedness influence in the formulation problem (in
particular, the Levenberg-Marquardt algorithm [8]) are presented in [9–11]. The
Gauss-Newton method is effective in solving many optimization problems, it is
easy to implement and is present in most software packages for applied
mathematics.
Let us turn to Fig. 1,
where the selected section
of the studied
phase object with a radial distribution of the refractive index
is shown in a
simplified way, the
axis coincides
with the direction of the probing light beam.
Fig. 1. Scheme for diagnosing a phase object
(axisymmetric approximation):
–
radial distribution of the refractive index.
Phase
perturbations
of the probing
light field that has passed through the medium under study depend on the
geometric path length and the refractive index
:
where
is the
wavelength,
is the
refractive index of air,
and
are the entry
and exit points of the beam from the object.
The
nonlinear integral operator of the first kind is a mathematical model of the
Hilbert visualization of the phase shift
:
|
(1)
|
The
inverse problem consists in reconstructing the function
from the
values, which are recorded in the experiment.
The
method based on the successive selection of the profile
and
calculation of the hilbertogram
is proposed for the solution. The local extrema
coincidence of the experimental and reconstructed hilbertograms serves as a
criterion for stopping the procedure.
The
desired phase perturbation function
in the section
on the
interval
is modeled by
Bernstein polynomials of the
n-th order (Bezier curves) in the smooth
fields case [12]:
|
(2)
|
where
are the
vectors components of the reference vertexes,
are the basis
functions of the Bezier curve, called Bernstein polynomials:
is the
polynomial degree,
is the ordinal
number of the reference vertex. The function
= 0 outside
the interval
.
The
equality
is valid in the section
for the phase
function (2) and any parameter
,
while
depends on the
coordinates of the reference points
.
Denote its
Hilbert image by
.
It is
calculated taking into account formula (1) as follows:
|
(3)
|
As
,
where
,
equation (3)
can be represented as
|
(4)
|
The
optimization problem is to determine those values of the parameters
(, …,
)
and
(, …,
),
at which the
objective function minimum is reached:
|
(5)
|
where
is the
hilbertogram recorded in the experiment (reference data).
In
general terms, the optimization problem is to find the extremum (minimum) of
the objective function:
The
iterative Gauss-Newton method, which uses the Jacobian matrix
of first-order
derivatives of the function
to find the
vector
that minimizes
,
is one way to
solve this problem.
The
Jacobian matrix is determined by the formula [13, 14]:
The
Gauss-Newton method consists in performing successive approximations
according to
the expression:
|
(6)
|
where
is the
iteration number,
is the
coefficient used to regulate the optimization "step" [6],
is the
transposed matrix.
Denote
,
.
|
|
Then
the objective function components (5) in the case of hilbertogram processing
can be written as
|
(7)
|
The
partial derivatives values with respect to the components of the vector
of the
function (4) will be found to determine the Jacobi matrix
.
First, we
write down the values of derivatives with respect to
coordinates:
Then
Since
and
it follows that
where
Thus
or
where
|
|
|
(8)
|
We
transform (8) by integration by parts:
Since
then
As
a result, we arrive at the equality
Similarly,
one can show that
Now
we write down the values of derivatives with respect to
coordinates:
Then
Since
we get
or
Similarly
As
a result, the Jacobian matrix will have the following form:
|
(9)
|
The
Gauss-Newton algorithm (6), taking into account (7) and (9), will be defined as
|
(10)
|
Thus,
optimization begins with setting the vector components
,
which
determine the phase function initial profile, then applying formula (10) step
by step, and terminating the calculation process when the squared distances sum
between the extrema coordinates of the reference and optimized hilbertograms
becomes less than the specified value.
It
is sufficient to determine the phase function only on the interval
in the case of
axisymmetric objects.
Let's
simulate an example when an axisymmetric object with a section radius
= 30 mm is
diagnosed. Define
within 0
≤
≤
as a
parametric Bezier curve of the third degree:
for
the value
= 0, and we
continue it in an even way to the region
≤
≤ 0.
The
pixel size in the image corresponded to 1/30 mm when registering the optical
signal in the experiments [3, 4]. Therefore, the number of samples
= 1801 in the
interval
].
The
function graph
and its
hilbertogram is shown in Fig. 2 for the reference points values
Fig. 2. The phase function represented by the
Bezier curve (with reference vertices) and the hilbertogram calculated from it.
The hilbertogram is presented in a dimensionless form with indication of the
local extrema points.
The hilbertogram shown in
Fig. 2 will be taken as the reference
,
and the phase
function corresponding to it will be denoted by
.
Let us
consider the biased function
as an
initial approximation, which must be optimized to
(Fig. 3).
Fig. 3. Reference hilbertogram
and
initial (optimized) hilbertogram
,
and their corresponding phase functions.
In
this case, the parameter
for
derivatives with respect to the vector
will have the
values presented in Table 1.
Table 1.
Derivatives
with respect to
and
coordinates
The
Jacobian matrix will have the following form:
The
root-mean-square error between the hilbertogram
and
after 30
iterations as a result of applying the Gauss-Newton method (10) has reached the
value
while the maximum deviation was
The phase function
obtained as a
result of optimization is shown in Fig. 4.
Fig. 4. Optimization result.
The
maximum deviation of
from
was
The
initial approximation
in the
considered example was chosen in such a way that the number of Hilbert bands
and
coincided. Let's
try to estimate the “limit” deviation of the initial approximation
, at which the
Gauss-Newton algorithm begins to work incorrectly (Fig. 5 and 6).
a
b
Fig. 5. (a) – Reference hilbertogram
and
initial (optimized) hilbertogram
,
and their corresponding phase functions (second approximation); (b) – optimization
result.
Let
the initial approximation
be smaller by
one Hilbert band in the ranges
and
than the
hilbertogram
(as shown in
Fig. 5.
a). As a result, the following values were achieved after 12
iterations (Fig. 5.
b):
Now let’s set the
so that the
number of Hilbert bands is greater by one in the same ranges
and
(Fig. 6.
a).
a
b
Fig. 6. (a) – Reference hilbertogram
and
initial (optimized) hilbertogram
,
and their corresponding phase functions (third approximation); (b) – optimization
result.
As
a result, after 41 iterations we obtained (Fig. 6.
b):
Thus,
it was found that for the convergence of the Gauss-Newton algorithm it is
necessary that the number of Hilbert bands
and
be equal. In
practice, of course, it is necessary that the Hilbert bands of the experimental
and optimized hilbertograms be as “close” to each other as possible.
The
Gauss-Newton method is adapted to the problem of phase function determining
from Hilbert diagnostic data of gaseous, condensed and reacting media. The
phase structure reconstruction algorithm is based on the sequential selection
of the phase profile specified by the Bézier polynomial and subsequent
calculation of the hilbertogram. The local extrema coincidence of the reference
and reconstructed hilbertograms is a criterion for the results reliability. The
Jacobian matrix calculation for the nonlinear integral operator of Hilbert
visualization has been completed, and an operation example of the algorithm for
the test function in the case of an axisymmetric formulation of the problem is
given.
Direction
of further research: application of the algorithm to the experimental data
processing. Adaptation of the Gauss-Newton method to the possibility of
specifying the desired phase function using several Bezier curves in cases of
diagnosing complex structures.
The work of the first
author was carried out within the framework of the state assignment of IM SB
RAS (No. FWNF-2022-0009), and the work of the other authors was carried out
within the framework of the state assignment of IT SB RAS (No. 121031800217-8).
1.
Arbuzov
V. A., Dubnishchev Yu. N. Metody gil'bert-optiki v izmeritel'nykh
tekhnologiyakh [Hilbert-optics methods in measurement technologies], NSTU
University Publ., Novosibirsk, 2007,
316 p. (in Russian)
2.
Soroko L. M.
Gil'bert-optika [Hilbert optics], Nauka Publ., Moscow, 1981, 160 p. (in
Russian)
3.
Arbuzov
V. A., Arbuzov E. V., Dubnishchev Yu. N., Lukashov V. V., Zolotukhina
O. S. Method of polychromatic Hilbert diagnostics of phase and temperature
perturbations of axisymmetric flames //
CEUR Workshop Proceedings, Vol. 3027, 2021, pp. 369–378 (doi:
10.20948/graphicon-2021-3027-369-378).
4.
Arbuzov
V. A., Arbuzov E. V., Dubnishchev Yu. N., Lukashov V. V., Zolotukhina
O. S., Tupikin A. V. Hilbert-optic diagnostics of hydrogen-oxygen
inverse diffusion flame // Energies, Vol. 15 (24), 2022, P. 9566 (doi: 10.3390/en15249566).
5.
Floudas
C. A., Pardalos P. M. Encyclopedia of optimization. New York:
Springer, 2008,
4622 p.
6.
Levenberg-Marquardt
algorithm for non-linear least squares method and its implementation in Python,
2016.
URL:
http://habr.com/ru/articles/308626/.
7.
Lovetsky
K. P., Sevastyanov L. A., Bikeev O. N., Paukshto M. V.
Matematicheskiy sintez opticheskikh nanostruktur [Mathematical synthesis of
optical nanostructures], RUDN University Publ., Moscow, 145 p. (in
Russian)
8.
Vasin V. V.
The Levenberg-Marquardt method for approximation of solutions of irregular
operator equations // Automation and remote control, Vol. 73, 2012, pp. 440–449
(doi: 10.1134/S0005117912030034).
9.
Yudin
N. E. Modified Gauss–Newton method for solving a smooth system of
nonlinear equations // Computer research and modeling, Vol. 13, pp. 2021, pp.
697–723 (doi: 10.20537/2076-7633-2021-13-4-697-723).
10.
Loke
M. H., Dahlin T. A comparison of the Gauss-Newton and quasi-Newton methods
in resistivity imaging inversion // Journal of Applied Geophysics, Vol. 49
(3), 2002, pp. 149–162
(doi:
10.1016/S0926-9851(01)00106-9).
11.
Bergou E. H., Diouane
Y., Kungurtsev V. Convergence and complexity analysis of a Levenberg-Marquardt
algorithm for inverse problems // Journal of Optimization Theory and
Applications, Vol. 185, 2020, pp. 927–944
(doi:
10.1007/s10957-020-01666-1).
12.
Rogers D. F., Adams J. A.
Matematical elements for computer graphics (Second Edition), McGraw-Hill, New
York, 1990.
13.
Zorich V. A.
Matematicheskiy analiz [Mathematical analysis], part 1, 6st. ed., ICNMO Publ.,
Moscow, 2012, 702 p. (in Russian)
14.
Bakhvalov N. S.,
Zhidkov N. P., Kobelkov G. M. Chislennyye metody [Numerical methods],
Binom Publ., Moscow, 2003, 632 p.
(in Russian)