MCS12 Luxor, Egypt, January 23–26, 2023

ACCURATE IGNITION DETECTION OF SOLID FUEL

PARTICLES USING MACHINE LEARNING
T. Li*, Z. Liang*, A. Dreizler* and B. Böhm*
[email protected]
*Technical University of Darmstadt, Department of Mechanical Engineering, Reactive Flows
and Diagnostics, Otto-Berndt-Straße 3, 64287 Darmstadt, Germany

Abstract
In the present work, accurate determination of single-particle ignition is focused on using high-
speed optical diagnostics combined with machine learning approaches. Ignition of individual
particles in a laminar flow reactor are visualized by simultaneous 10 kHz OH-LIF and DBI
measurements. Two coal particle sizes of 90 - 125 µm and 160 - 200 µm are investigated in con-
ventional air and oxy-fuel conditions with increasing oxygen concentrations. Ignition delay
arXiv:2305.00004v1 [cs.LG] 20 Apr 2023

times are first evaluated with threshold methods, revealing obvious deviations compared to the
ground truth detected by the human eye. Then, residual networks (ResNet) and feature pyrami-
dal networks (FPN) are trained on the ground truth and applied to predict the ignition time. Both
networks are capable of detecting ignition with significantly higher accuracy and precision. Be-
sides, influences of input data and depth of networks on the prediction performance of a trained
model are examined. The current study shows that the hierarchical feature extraction of the con-
volutions networks clearly facilitates data evaluation for high-speed optical measurements and
could be transferred to other solid fuel experiments with similar boundary conditions.

Introduction
Particle ignition is an essential stage for the flame stability of a pulverized fuel stream, which
has been a research topic since years [1–4]. In general, particle ignition can be classified into
two modes: homogeneous and heterogeneous ignition. In homogeneous ignition, volatile mat-
ters, including hydrogen (H2 ), hydrocarbons (Cx Hy ), and tars, are released from particles upon
increasing temperatures. These fuel gases, mixing with oxidizer, ignite above the flammable
temperature and mixture fraction limits, resulting in a gas-phase flame in the vicinity of the
particle. In heterogeneous ignition, oxidizers approach the particle surface followed by direct
surface reactions. The ignition mode can be influenced by coal rank, particle size, heating rate,
and gas composition.
On the one hand, particle surface temperature is a widely accepted parameter for deter-
mining heterogeneous ignition. A well-calibrated two- or three-color pyrometer can provide
conclusive information about particle temperature and heterogeneous combustion [5–7]. On the
other hand, indicators and corresponding measurement techniques are rather diverse for homo-
geneous combustion. Here, particle temperature is still an essential parameter in experiments,
see for example [8–14]. However, the pyrometric temperature closely relates to hot soot and char
particles. Recent advancements in high-speed imaging enables tracking the particle combustion
history with sufficient temporal and spatial resolutions [10, 15–17]. These investigations define
the gas-phase ignition on the first visible broad-band emission signal recorded by high-speed
cameras in the kHz range. Methylidyne (CH) chemiluminescence was also used as an ignition
indicator, which can be imaged in a time-integrated [18] or single-shot [7] manner. For single-
shot measurements, intensified cameras are required due to the weak CH* chemiluminescence
signals. In addition, soot particles were imaged by an intensified camera to derive the homoge-
neous ignition time [19]. Since the CH* correlates with the heat release in the gas phase, it is
a better indicator than the black-body radiation of soot particles. Unfortunately, CH* emission
(at about 430 nm) spectrally overlaps with the broad-band black-body emission.
Besides CH radicals, the hydroxyl (OH) radical has been used as an important flame marker,
which abundantly exists in the reaction zone and the burned gas. The first planar laser-induced
fluorescence of OH radicals (OH-LIF) on the single-particle level was reported in [20]. The ho-
mogeneous ignition process of individual particles was visualized at 10 kHz with a spatial reso-
lution of approximately 100 µm. The technique enabled the evaluation of single-particle volatile
flame structures with different particle sizes and atmospheres [21]. Further experiments, supple-
mented by simultaneous flame luminosity and diffused backlight-illumination (DBI), indicated
the distinctive appearance of OH-LIF and luminosity signals at the onset of ignition, suggesting
OH-LIF as a favorable diagnostic approach for homogeneous ignition detection [22].
Accurate ignition detection depends on combustion conditions, experimental techniques, as
well as the definition of the ignition event [23]. The onset of ignition is usually defined based on
a certain intensity threshold [7, 18, 22] above the background level or the evaluation of signal
topology [19, 24]. Parameter modification in the image analysis is unavoidable if a different
particle or particle size is investigated, or another detection system is used. This hurdle can be
tackled by deep learning for a better feature extraction. Deep learning approaches are capable
of learning image features and making predictions after training on a data set. Recently, con-
volution neural networks (CNN) have been applied in experimental studies to predict velocity
fields [25] and 3D flame structures [26]. In the present work, high-speed multi-parameter laser
diagnostics are assisted by deep learning approaches, providing accurate ignition delay time
based on object classification and detection architectures.
Considering its importance and challenges, the current study emphasizes on the accurate
ignition detection in solid fuel combustion, which was not fully addressed in the previous exper-
iment. Homogeneous ignition of high-volatile solid fuels at realistic heating rates is particularly
targeted. For this purpose, simultaneous OH-LIF and DBI measurements were applied for single
particles burning in a laminar flow reactor (LFR), producing a comprehensive database. Homo-
geneous ignition is evaluated for two particle sizes under air and oxy-fuel atmospheres. Two
main deep learning architectures, namely residual network (ResNet) and feature pyramidal net-
work (FPN) are implemented. The influence of training data, network depth, and pre-training
on the capability for ignition prediction is carefully examined.

Experiments
Multi-parameter optical diagnostics were employed to investigate the fundamental processes
of single coal particles during ignition and volatile combustion in laminar flows. The experi-
mental details were reported in the previous work [24] and will be briefly introduced here. An
in-house laminar flow reactor was designed to generate desired gas atmospheres simulating
temperatures and species concentrations in realistic conditions. Premixed lean methane (CH4 )
flat flames were stabilized on the surface of a ceramic matrix. By operating inlet gas mixtures of
CH2 /O2 /N2 or CH2 /O2 /CO2 , conventional (AIR) and oxy-fuel (OXY) atmospheres were gen-
erated in the exhaust gas with homogeneous temperature and velocity fields. These conditions
were denoted as AIR10/20/30/40 and OXY20/30/40, with the number indicating the volumet-
ric oxygen concentrations of the exhaust gas. Micrometer-sized bituminous coal particles were
seeded individually into the burner by carrier gases having the same molecular composition and
velocity as the flat flame inlet.
Figure 1 illustrates the multi-parameter optical measurement and a sketch of the burner. Gas-
phase volatile flames were visualized by two-dimensional high-speed OH-LIF at 10 kHz. OH-
LIF signals were excited at 283.01 nm by a dye laser system and collected by a CMOS camera
(HSS6, LaVision) coupled with a lens-coupled intensifier (HS-IRO, LaVision). A narrow band-
pass filter and a short gate time were applied to eliminate interfering thermal radiation and
chemiluminescence. Particle ignition time can be determined by tracking the temporal varia-
tion of OH-LIF signal topology [24]. Besides, particles positions and velocities are simulta-
neously detected by 10 kHz DBI measurements. A high-power LED illuminated coal particles
at 550 nm and projected particle shadows were imaged onto another CMOS camera (HSS6,
LaVision). This camera was equipped with a long-distance microscope to enhance the spatial
resolution estimated as approximately 20 µm at the burner center position [27]. The DBI tech-
 Optical Laminar flow reactor
multi-parameter
diagnostics
LIF
DBI
LED
Diffusor
Lens HS-IRO HSS6

λ=550nm

OH-LIF

x Filter
λ=283nm Microscope
y z
HSS6 DBI
Beam dump

Figure 1: A schematic experimental layout including optical diagnostics and the laminar flow
reactor.

nique allowed for an accurate determination of particle size, shape, and velocities [24, 27]. As
illustrated in Fig.1, the difference in resolutions led to a smaller field of view for DBI (i.e.
11(height) × 5(width) mm2 ) than for OH-LIF (i.e. 19 × 19 mm2 ). Since the homogeneous ig-
nition of single particles occurs within a few millimeters above the burner, the ignition process
can be fully captured by both imaging techniques. More details regarding experimental method-
ology are referred to [24].

Data evaluation
Colombian high-volatile bituminous coal particles sieved to two size distributions of 90 -125 µm
and 160 -200 µm were investigated, which are referred as particles A and B. With an approxi-
mately 105 K/s particle heating rate [24, 27], homogeneous ignition of released volatiles dom-
inates the particle ignition mode. The ignition process was captured in the so-called single-
particle event, in which an individual particle moves through the probe volume without inter-
acting with other particles. In total, 1006 events and 512 events were detected for particles A
and B, respectively. Due to the larger diameter, particle B was more difficult to seed through the
0.8 mm injection tube, resulting in lower probabilities of particle detection.

tign-1 ms tign tign + 1 ms tign + 2 ms tign + 3 ms tign + 4 ms

(a)
OH
Intensity
1 IOH (a.u.) 3 5000 μm

(b)
Binary
OH
tign-1 ms tign tign + 1 ms tign + 2 ms tign + 3 ms tign + 4 ms

Figure 2: A time-resolved sequence of particle ignition with tign given by the ground truth
(manual labeling). (a) OH-LIF raw images. (b) binary OH-LIF images.

An example of a single-particle event is presented in Fig. 2, temporally conditioned on the

ignition instance at tign (every tenth images shown). Figure 2(a) shows the temporal evolution of
OH-LIF intensities normalized on the homogeneous background signal. By applying a constant
intensity threshold Ith of 1.2 [24], corresponding binary images are shown in Fig.2(b). In the
previously proposed signal and structure (SAS) analysis [24], the onset of ignition is defined if
the normalized intensity and the connected area exceed respective thresholds. The center points
of particles (red solid dots) are shot-by-shot determined from the DBI data. Since all cameras
are spatially mapped based on an accurate target calibration, DBI center points reveal as good
references to locate particles in OH-LIF images.
To assess the accuracy of different methods for ignition detection, a ground truth of ignition
delay time needs to be defined for each particle event. For this purpose, the ignition frame of all
1518 events is manually labeled the same person, who is instructed to select the first OH-LIF
image, which contains recognizable flame structures against the background. The performance
for predicting single-particle ignition can be evaluated using the ignition time difference (ITD):

ITDdetector = ti,detector − ti,gt . (1)

Here, ti,detector and ti,gt indicate the ignition delay time determined by a detector and ground
truth, respectively. If ITD > 0, ignition time is over-predicted, otherwise ignition time is under-
predicted. A statistical analysis was performed for each particle size to compare the accuracy
and precision of each detector or data processing method.
Two different deep learning networks are implemented, which belong to object classifica-
tion (i.e, residual networks, ResNet) and object detection (feature pyramidal networks, FPN)
in the context of computer vision. Image classification works on an entire image or an image
segment and classifies the image into a category, whereas object detection specifies multiple
objects within an image into categories (e.g. classification) with their locations represented by
bounding boxes (e.g. bounding box regression). For brevity, the architecture of deep learn-
ing networks are not elaborated here, and more details of related work are referred to [28] for
ResNet, [29] for R-CNN, [30] for Fast R-CNN, [31] for Faster R-CNN, and [32] for FPN.

Results and Discussions

Ignition detected by SAS
Figure 3(a) shows the ignition delay time determined by ground truth (ti,gt ) for particles A and B
under seven investigated atmospheres. Here, symbols and error bars indicate the mean value (µ)
and ± one standard deviation (σ), respectively. The ignition delay time is referenced to the flat
flame position (y ≈ 1.5 mm) where particles start to heat up by the hot exhaust gas. Basically,
the overall trend over particle diameters and atmospheres is very similar to ti,SAS in [24]. Effects
of particle size, oxygen concentration, and CO2 replacement can be observed, and more details
can be found in [24]. Differences between ti,gt and ti,SAS are observed. The over-prediction of
ignition times by the SAS analysis can be quantified by calculating the ignition time difference
ITD for each single-particle event, as shown in Figure 3(b). Overall good agreement is archived
by the SAS analysis for the small particles A, whereas some discrepancies can be identified
for the large particle B. Specifically, the SAS analysis over-predicts the ignition delay time of
particles B by 2 - 4 ms in all atmospheres, which is presumably caused by the improper selection
of intensity and area thresholds. This challenge motivated the application of machine learning
for the more generalized feature detection of ignited particles.

Ignition detected by ResNets

In machine learning approaches, only 35% of the entire events are used to train ResNet models
[28], which corresponds to 462 particle events. The remaining 65% (1056 events) are test-
ing data, allowing for sufficient statistical comparisons with ground truth and the performance
evaluation of different architectures. However, if the 35% data were further divided into train-
ing data and validation data, the number of training data would be insufficient to obtain con-
verge networks. Therefore, the k-folds cross-validation approach is consistently used to train
object classification networks. With k-folds cross-validation, the entire training set is split into
 20 8
(a) A B (b) A B
6
15

ti,SAS - ti,gt (ms)

4

ti,gt (ms)
10
2
5
0

0 -2
AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40 AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40
Conditions Conditions

Figure 3: Comparison of ignition delay times by the SAS method ti,SAS and the manual label
ti,gt for two particle sizes A and B in seven atmospheres.

k sets. One by one, a set is selected for validation, and the k − 1 other sets are combined into the
corresponding training set. 5-folds cross-validation is used, and 10 epochs are trained for each
fold. A predicted probability higher than 50% is classified as an ignited particle in testing data.
3 3
(a) Nev = 14 A B (b) Nev = 56 A B
2 2
ti,RN - ti,gt (ms)

ti,RN - ti,gt (ms)

1 1

0 0

-1 -1

-2 -2

-3 -3
AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40 AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40
Conditions Conditions
3 3
(c) Nev = 140 A B (d) Nev = 462 (35% of all) A B
2 2
ti,RN - ti,gt (ms)

ti,RN - ti,gt (ms)

1 1

0 0

-1 -1

-2 -2

-3 -3
AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40 AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40
Conditions Conditions

Figure 4: Ignition time difference ti,RN - ti,gt by using ResNet-18 with the amount of particle
events (a) Nev = 14, (b) Nev = 56, (c) Nev = 140, and (d) Nev = 462.

Several ResNet-18 models are trained with an increasing number of particle events, which
are equally selected from different operating conditions, in order to examine the influence of
training data. Figure 4 shows the ignition time difference between ground truth tt,gt and ti,RN
predicted by ResNet-18 which is trained with increasing particle events Nev = 14, 56, 140, and
462. They correspond to 1, 4, 10, and 33 (equivalent) particle events selected from each atmo-
sphere and particle size. Note that new data are always added into the training by retaining the
existing data, e.g., the original 14 events are included in the 56 events and so on. For a reason-
able comparison, 1056 particle events are consistently used for the prediction purpose. Gener-
ally, ResNet-18 achieves a higher precision with less scattered results than the SAS methods.
With more data fed into the training process, mean values of ti,RN - ti,gt (ITD) approach 0 ms
indicating an continuously improving accuracy in predicting the ground truth. With a small
amount of data, e.g. Nev ≤ 140, ITD distributions shift in either negative or positive side. It
implies that the networks are still fragile to the newly added training data, as weight and bias
parameters are not sufficiently trained with a limited number of events. With Nev increased to
462, the predictive accuracy becomes equally high for particle A and B. Furthermore, error bars
in Fig. 4 constantly narrow with increasing data involved in the training process. It is obvious
that the number of images used for training improves the precision of ignition prediction. For
ignition detection, 35% of data for training is proposed as a reasonable compromise between
the performance of trained networks and the expanse of manual labeling whenever a new data
set is under evaluation.

Ignition detected by FPN

For FPN [32], training loss converges, and 5 epochs are trained to reduce the computational
cost. A particle will be identified as ignited for the final classification if its predicted probabil-
ity is higher than 50%. Figure 5 compares predictive results of FPN by implementing ResNets
with an increasing depth in the bottom-up pathway. All models are trained with 462 particle
events. The ignition time difference ti,FPN - ti,gt is represented by its mean and standard devia-
tion. An overall good agreement between FPN prediction and ground truth is observed, which
is independent on atmosphere but on particle size. Compared to the SAS method in Fig.3,
evident improvements are achieved in predictive accuracy and precision. However, compared
to ResNet-18 in Fig 4(d), the ITD scatters more in FPN by using the same training data. In
addition, it can be noted that increasing backbone convolutions layers narrows the error bars
and improves the prediction precision. But no further improvements can be noticed comparing
backbone ResNet-101 with ResNet-50.
3 3
(a) Backbone = ResNet18 A B (b) Backbone = ResNet34 A B
2 2
ti,FPN - ti,gt (ms)

ti,FPN - ti,gt (ms)

1 1

0 0

-1 -1

-2 -2

-3 -3
AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40 AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40
Conditions Conditions
3 3
(c) Backbone = ResNet50 A B (d) Backbone = ResNet101 A B
2 2
ti,FPN - ti,gt (ms)

ti,FPN - ti,gt (ms)

1 1

0 0

-1 -1

-2 -2

-3 -3
AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40 AIR10 AIR20 AIR30 AIR40 OXY20 OXY30 OXY40
Conditions Conditions

Figure 5: Ignition time difference ti,FPN - ti,gt using different ResNet models in the bottom-up
pathway of FPN networks.

Conclusions
The current work presents an experimental investigation of single particle ignition using high-
speed optical diagnostics in a laminar flow reactor. Homogeneous ignition is visualized by
10 kHz OH-LIF measurements with the simultaneously detected particle location in DBI mea-
surements. Accurate detection of ignition delay times is focused on with both conventional
threshold methods and advanced machine learning approaches. The prediction performance of
different approaches for ignition detection are conclusively compared in Fig. 6. The mean µ and
standard deviation σ of ITD are evaluated by including all atmospheres and further used to gen-
erate normal distributions N (µ, σ) approximating the overall prediction performance for parti-
cles A and B. The previously introduced SAS method provides satisfactory predictions for small
particles but substantially over-estimates large particles’ ignition. It is because the accuracy and
precision inherently relate with the threshold selection. Owing to the changing characteristics of
OH-LIF signals with particles and atmospheres, fixed thresholds for area and intensities might
induce errors and restrict the detection quality. Sensitivity analyses (not shown) clearly indi-
cates the distinctive signals, especially the intensity levels at the onset of ignition, which makes
conventional threshold methods incapable for thus a task. Although optimum thresholds can
be statistically obtained by minimizing the difference between prediction and ground truth, this
method is not really viable when dealing with a new set of data, owing to the in-prior knowledge
required for optimal thresholding. To avoid the difficulties in optimizing algorithm parameters,
convolutions networks with hierarchical feature extraction are implemented.

1 (a) SAS A
B
0.5

Relative Probabilities
1 (b) ResNet-18
0.5

0
1 (c) FPN
0.5

0
-4 -2 0 2 4 6 8
ti,det - ti,gt (ms)

Figure 6: Relative probability distributions of ignition time differences ti,det - ti,gt , in which ti,det
represents predicted ignition delay times by different approaches.

Figure 6(b) and (c) show the results provided by the best ResNet model (i.e. pre-trained
ResNet-18 fine-tuned with 462 events) and the best FPN model (i.e. backbone ResNet-50
trained with 462 events), respectively. Evidently, enhanced quality of ignition detection is
achieved, which is superior to the conventional processing approach investigated in this study. This
can be explained by the feature recognition over different scales, which is inherently included
in the hierarchy of convolutional layers in their architectures. As a result, ResNet-18 achieves
the most accurate and precise ignition delay time compared with ground truth. More complex
networks such as FPN promote no further improvements but slight under-estimation of igni-
tion delay times. However, training ResNet-18 involves an additional processing step of RoI
extraction from the particle center. Despite FPN models also require pre-estimated particle po-
sitions but is able to work on an entire image. Although these models are heavier and need a
longer training time, they could be further developed to detect multi-particle ignition in the fu-
ture. Regarding the simple features of ignited coal particles, it can be concluded that both object
classification and detection approaches of machine learning are valuable for solid fuel combus-
tion analysis. Residual and feature pyramidal networks are appropriate architectures for such
evaluation tasks and have the potential to be transferred to other experimental investigations.

Acknowledgements
This work was founded by the Hessian Ministry of Higher Education, Research, Science and
the Arts - cluster project Clean Circles.

References
[1] Howard, J.B., Essenhigh, R.H., “Mechanism of solid-partical combustion with simulta-
neous gas-phase volatiles combustion”, Symposium (International) on Combustion 11(1):
399–408 (1967).
[2] Essenhigh, R.H., Misra, M.K., Shaw, D.W., “Ignition of coal particles: A Review: A
review”, Combustion and Flame 77(1): 3–30 (1989).
[3] Annamalai, K., Ryan, W., “Interactive processes in gasification and combustion—II. Iso-
lated carbon, coal and porous char particles”, Progress in Energy and Combustion Science
19(5): 383–446 (1993).
[4] Du, X., Annamalai, K., “The transient ignition of isolated coal particle”, Combustion and
Flame 97(3-4): 339–354 (1994).
 [5] Levendis, Y.A., Estrada, K.R., Hottel, H.C., “Development of multicolor pyrometers to
monitor the transient response of burning carbonaceous particles”, The Review of scientific
instruments 63(7): 3608–3622 (1992).
[6] Bejarano, P.A., Levendis, Y.A., “Single-coal-particle combustion in O2 /N2 and O2 /CO2
environments”, Combustion and Flame 153(1-2): 270–287 (2008).
[7] Yuan, Y., Li, S., Zhao, F., Yao, Q., Long, M.B., “Characterization on hetero-homogeneous
ignition of pulverized coal particle streams using CH* chemiluminescence and 3 color
pyrometry”, Fuel 184: 1000–1006 (2016).
[8] Khatami, R., Levendis, Y.A., “On the deduction of single coal particle combustion tem-
perature from three-color optical pyrometry”, Combustion and Flame 158(9): 1822–1836
(2011).
[9] Levendis, Y.A., Joshi, K., Khatami, R., Sarofim, A.F., “Combustion behavior in air of
single particles from three different coal ranks and from sugarcane bagasse”, Combustion
and Flame 158(3): 452–465 (2011).
[10] Khatami, R., Stivers, C., Joshi, K., Levendis, Y.A., Sarofim, A.F., “Combustion behavior
of single particles from three different coal ranks and from sugar cane bagasse in O2 /N2
and O2 /CO2 atmospheres”, Combustion and Flame 159(3): 1253–1271 (2012).
[11] Maffei, T., Khatami, R., Pierucci, S., Faravelli, T., Ranzi, E., Levendis, Y.A., “Experimen-
tal and modeling study of single coal particle combustion in O2 /N2 and Oxy-fuel (O2 /CO2 )
atmospheres”, Combustion and Flame 160(11): 2559–2572 (2013).
[12] Khatami, R., Levendis, Y.A., Delichatsios, M.A., “Soot loading, temperature and size
of single coal particle envelope flames in conventional- and oxy-combustion conditions
(O2 /N2 and O2 /CO2 )”, Combustion and Flame 162(6): 2508–2517 (2015).
[13] Lei, K., Ye, B., Cao, J., Zhang, R., Liu, D., “Combustion Characteristics of Single Particles
from Bituminous Coal and Pine Sawdust in O2/N2, O2/CO2, and O2/H2O Atmospheres”,
Energies 10(11): 1695 (2017).
[14] Vorobiev, N., Valentiner, S., Schiemann, M., Scherer, V., “Comprehensive Data Set of
Single Particle Combustion under Oxy-fuel Conditions, Part I: Measurement Technique”,
Combustion Science and Technology pp. 1–22 (2020).
[15] Cai, L., Zou, C., Liu, Y., Zhou, K., Han, Q., Zheng, C., “Numerical and experimental
studies on the ignition of pulverized coal in O2/H2O atmospheres”, Fuel 139: 198–205
(2015).
[16] Khatami, R., Levendis, Y.A., “An overview of coal rank influence on ignition and com-
bustion phenomena at the particle level”, Combustion and Flame 164: 22–34 (2016).
[17] Adewale Adeosun, Zhenghang Xiao, Akshay Gopan, Zhiwei Yang, Xuebin Wang, Tianx-
iang Li, Qiang Yao, Richard L. Axelbaum, “Pulverized coal particle ignition in a com-
bustion environment with a reducing-to-oxidizing transition”, Journal of energy institute
(2018).
[18] Molina, A., Shaddix, C.R., “Ignition and devolatilization of pulverized bituminous coal
particles during oxygen/carbon dioxide coal combustion”, Proceedings of the Combustion
Institute 31(2): 1905–1912 (2007).
[19] Shaddix, C.R., Molina, A., “Particle imaging of ignition and devolatilization of pulverized
coal during oxy-fuel combustion”, Proceedings of the Combustion Institute 32(2): 2091–
2098 (2009).
[20] Köser, J., Becker, L.G., Vorobiev, N., Schiemann, M., Scherer, V., Böhm, B., Dreizler, A.,
“Characterization of single coal particle combustion within oxygen-enriched environments
using high-speed OH-PLIF”, Applied Physics B 121(4): 459–464 (2015).
[21] Köser, J., Becker, L.G., Goßmann, A.K., Böhm, B., Dreizler, A., “Investigation of ignition
and volatile combustion of single coal particles within oxygen-enriched atmospheres using
high-speed OH-PLIF”, Proceedings of the Combustion Institute 36(2): 2103–2111 (2017).
[22] Köser, J., Li, T., Vorobiev, N., Dreizler, A., Schiemann, M., Böhm, B., “Multi-parameter
diagnostics for high-resolution in-situ measurements of single coal particle combustion”,
Proceedings of the Combustion Institute 37(3): 2893–2900 (2019).
[23] ZHANG, D., WALL, T., “Ignition of coal particles:The influence of experimental tech-
nique”, Fuel 73(7): 1114–1119 (1994).
[24] Li, T., Farmand, P., Geschwindner, C., Greifenstein, M., Köser, J., Schumann, C., Attili,
A., Pitsch, H., Dreizler, A., Böhm, B., “Homogeneous ignition and volatile combustion of
single solid fuel particles in air and oxy-fuel conditions”, Fuel 291(1): 120101 (2021).
[25] Barwey, S., Hassanaly, M., Raman, V., Steinberg, A., “Using Machine Learning to
Construct Velocity Fields from OH-PLIF Images”, Combustion Science and Technology
194(1): 93–116 (2022).
[26] Huang, J., Liu, H., Cai, W., “Online in situ prediction of 3-D flame evolution from its
history 2-D projections via deep learning”, Journal of Fluid Mechanics 875 (2019).
[27] Li, T., Li, B., Farmand, P., Dreizler, A., Pitsch, H., Böhm, B., “Motion and swelling
of single coal particles during volatile combustion in a laminar flow reactor: accepted”,
Proceedings of the Combustion Institute (2022).
[28] He, K., Zhang, X., Ren, S., Sun, J., “Deep residual learning for image recognition”, in
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
(2016).
[29] Girshick, R., Donahue, J., Darrell, T., Malik, J., “Rich feature hierarchies for accurate
object detection and semantic segmentation”, in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR) (2014).
[30] Girshick, R., “Fast r-cnn”, in Proceedings of the IEEE International Conference on Com-
puter Vision (ICCV) (2015).
[31] Ren, S., He, K., Girshick, R., Sun, J., “Faster r-cnn: Towards real-time object detection
with region proposal networks”, in Advances in Neural Information Processing Systems,
edited by C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, R. Garnett, vol. 28, Curran
Associates, Inc. (2015).
[32] Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S., “Feature pyra-
mid networks for object detection”, in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR) (2017).