Professional Documents
Culture Documents
Measurementsof Laminar Flame Speed Goulier 16
Measurementsof Laminar Flame Speed Goulier 16
© 2016 Author(s).
Numerical Assessment of Accurate Measurements of
Laminar Flame Speed
Joules Goulier1,2, Katarzyna Bizon3, Nabiha Chaumeix1, Nicolas Meynet2, and
Gaetano Continillo4,a)
1
ICARE-CNRS, 1C av. Del a Recherche Scientifique 45000 Orléans, France
2
Institut de Radioprotection et de Sureté Nucléaire IRSN, 92262 Fontenay aux Roses, France
3
Cracow University of Technology, ul. Warszawska 24, 30-155 Kraków, Poland
4
Università degli Studi del Sannio, Piazza Roma 21, Benevento, Italy
a)
Corresponding author: continillo@unisannio.it
Abstract. In combustion, the laminar flame speed constitutes an important parameter that reflects the chemistry of
oxidation for a given fuel, along with its transport and thermal properties. Laminar flame speeds are used (i) in turbulent
models used in CFD codes, and (ii) to validate detailed or reduced mechanisms, often derived from studies using ideal
reactors and in diluted conditions as in jet stirred reactors and in shock tubes. End-users of such mechanisms need to have
an assessment of their capability to predict the correct heat released by combustion in realistic conditions. In this view,
the laminar flame speed constitutes a very convenient parameter, and it is then very important to have a good knowledge
of the experimental errors involved with its determination. Stationary configurations (Bunsen burners, counter-flow
flames, heat flux burners) or moving flames (tubes, spherical vessel, soap bubble) can be used. The spherical expanding
flame configuration has recently become popular, since it can be used at high pressures and temperatures. With this
method, the flame speed is not measured directly, but derived through the recording of the flame radius. The method used
to process the radius history will have an impact on the estimated flame speed. Aim of this work is to propose a way to
derive the laminar flame speed from experimental recording of expanding flames, and to assess the error magnitude.
EXPERIMENTAL METHODOLOGY
Experimental Setup. The spherical bomb is a jacketed stainless steel vessel. The internal diameter is 476 mm with
a thickness of 12 mm, for a maximum operating pressure of 50 bar. Heated thermal fluid circulates in the jacket to
ensure constant wall temperature, measured up to a maximum of 500 K with a precision of 1 K. A Kistler 601A
quartz pressure transducer located on the inner wall is used to record the pressure. Four quartz windows on the
equatorial plan of the vessel allow the flame to be recorded up to a radius of 46 mm, when only 0.85 % of the total
volume is burned. Hence, the pressure can be considered constant during flame propagation and no wall-effects take
place. The visualization of the flame front is obtained via a Z shape Schlieren diagnostic (Fig. 1a) with a high speed
camera ensuring acquisition up to 25,000 Hz at a resolution of 768x768 pixels. The flame is ignited via two tungsten
electrodes connected to a high voltage generator, delivering about 2 mJ for a gap distance of 2 mm. The spark
triggers pressure and camera measurements at the same time via a TTL generator. The mixtures were prepared
directly inside the spherical vessel using compressed hydrogen, methane, ethane from Air liquid, ethanol from
Sigma-Aldrich, and laboratory dry air (0.21O2 + 0.79N2). Partial pressures of the gases were measured using
capacitive manometers (MKS) of two different scales (133 mbar and 1333 mbar). According the precision of the
manometers, the mixtures were obtained with an accuracy of 1%.
Flame Radius Measurement. The images of the growing flame are acquired using a high speed camera, then
transferred on a PC to be analyzed. Figure 1b gives an illustration of the recorded images. The images were
processed using MATLAB 2013b and the Image Processing toolbox. The first image of the series, where usually no
flame or a small kernel is present, is defined as the background image. Figure 2 shows the successive steps of the
International Conference of Computational Methods in Sciences and Engineering 2016 (ICCMSE 2016)
AIP Conf. Proc. 1790, 110006-1–110006-4; doi: 10.1063/1.4968713
Published by AIP Publishing. 978-0-7354-1454-9/$30.00
110006-1
image processing: The first step is to perform an absolute subtraction of each image of the series (Figure 2.b) with
the background (Figure 2.a) followed by contrast adjustment. The resulting image (Figure 2.c) highlights the flame
sphere and removes the background. The next step is an edge detection using the Canny method [1] which
underlines all edges of the flame (Figure 2.d), the electrodes and, at high temperatures, thermal gradients inside the
bomb. Unwanted areas are then removed from the image by applying masks. A boxed mask is manually applied
over the electrodes, so that every edge detected inside the box is discarded. Two elliptic masks are sequentially
applied to each image: an inner ellipse will mask every pixel that is under the ellipse fitted to the previous image
minus a small variation (1 to 3 pixels). A second ellipse will mask every pixel that is above the ellipse fitted to the
previous image plus a variation of 10 to 20 pixels. Figure 2.e shows (blue area) the region where the edges in white
are kept; red edges are discarded. All pixels kept are used to fit an ellipse using the method described by Pilu et al.,
giving the major and minor axis , and orientation [2]. The radius of the flame is then computed as √ · .
Mirror
Pinhole
pressure (bar)
Light Source
Screen
Knife
Mirror
FIGURE 2. Image processing: (a) Background; (b) image of an 11.8 mm flame; (c) absolute subtraction of the background; (d)
edge detected on the subtracted image; (e) edge cleaned with a satisfactory region; (f) ellipse fitting added to the flame image.
NUMERICAL PROCEDURE
The unstretched flame speed, and the Markstein length, were determined using both linear (Markstein and
Karlovitz) and non-linear (Romney and Sivashinsky) models. For the linear model, given by:
(1)
the unknown parameters, and , are determined by simple linear regression. The nonlinear model is given by:
L L L
ln 2 , where: and (2)
L L L
Using basic properties of logarithms and using the definition of , the model can be rewritten as:
ln 2 (3)
By replacing , the problem takes the form of a fully implicit parametrized ordinary differential equation:
ln 2 (4)
110006-2
The associated initial condition is given by:
; where , (5)
Parameter estimation procedure. The parametrized implicit initial value problem of Eqs. (4,5) can be written as:
, ;
, ; 0 (6)
; (7)
where ; , ,…, is a state vector, , ,…, is a vector of unknown parameters
and is a know function. In practice, in experiments we observe only a surrogate of ; , namely:
, 1, … , (8)
where e(ti) is a measurement error. The vector of parameters can be estimated by minimizing a cost function:
∑ ̃ ; (9)
At each time , is the experimentally observed value, and ̃ ; is the numerically predicted value.
Uncertainties in multi-variable functions. To evaluate the influence of the measurement error e(ti) onto the
uncertainty of the estimation of parameter vector , ,…, , consider a general function of variables,
, ,…, , 1, … , . One needs to map a series of measurements, , , … , , and their associated
errors though the function . Assume that the best estimate of is made through the mean values of the
measurements, hence , ,…, . The error in is then a function of both the mean values and their
errors, i.e. , , … , ; Δ , Δ , … , ∆ . In particular, the uncertainty in due to the uncertainty in is:
| , ,…, Δ ,…, , ,…, ,…, | (10)
110006-3
FIGURE 3. (a): Simulated flame radius, and (b): corresponding flame velocity plotted as a function of stretch,
calculated for ̃ . (c) Flame velocity as a function of stretch, for 0.8499 and 3999.99.
FIGURE 4. (a): Absolute error of the estimates of the unstretched flame speed and (b): the Markstein length (b), for
varying minimum radius . (c) Flame velocity as a function of stretch, , for different values of and .
Finally, the influence of the measurement error introduced in the numerically derived data was evaluated,
assuming the maximal measurement error on the radius 0.198 . In the first step, all components of ̃ were
translated by , to create new data vectors containing a constant systematic error, namely:
̃ , 1, … , ; ̃ , 1, … , .
The values of the absolute error for the parameters estimated using as input vectors ̃ , ̃ and ̃
(remember that ̃ , 1, … , ) are reported in Table 1. The error of the estimates obtained for
̃ is much smaller - by four orders of magnitude for and by three orders of magnitude for - than the error
obtained for ̃ and ̃ . This shows that the consequence of the error introduced by the numerical
procedure on the parameter estimates is much smaller than the consequence of the expected minimal experimental
error (corresponding to one pixel in the observation).
TABLE 1. Influence of the measurement error on the absolute error of the parameter estimates.
̃ ̃ ̃
, 2.19 · 10 0.0164 0.0165
, 0.0011 3.2350 3.2156
Successively, to evaluate the total error in , , the uncertainty of the estimate due to the
uncertainty in , 1, … , , was calculated according to Eq. (10). Relatively high values of uncertainty were
obtained when altering by the measurement error the first point , that is 64.4219 and
0.1935. They are caused by the fact that, when solving the initial value problem within the optimization procedure,
this point is set as initial condition. The total error is:
∑ 65.8562 ; ∑ 0.1954 .
REFERENCES
1. J. Canny, “A Computational Approach to Edge Detection,” in IEEE Transactions on Pattern Analysis and
Machine Intelligence,Vol. PAMI-8, No. 6, 679-698, (1986).
2. M. Pilu, A. Fitzgibbon and R.Fisher, “Ellipse-specific Direct least-square Fitting, ” in IEEE International
Conference on Image Processing, Lausanne, September (1996).
110006-4