2
fundamental differences of the tests (gravimetric vs spec-
trometry). For instance, CDEM provides qualitative results
(Go/No-Go) while laboratories provide a quantitative anal-
ysis. In addition, the CDEM specifications state that the
accuracy is ± 2 %,however, this is only true when using
completely dried samples. One percent (1%) moisture can
drop the IC readings by as much as 7% [1].
The rapid development of spectral sensors within the
last decade has resulted in numerous portable spectrom-
eters being developed and brought to market. This has
allowed for the identification of materials in the field rather
than the laboratory. More recently, advances in regres-
sion modelling along with higher resolution sensors now
allow for portable quantitative analysis of materials. For
instance, portable X-ray fluorescence has been used for
the analysis of ash content in coal [5]. Recent advances in
micro-electro-mechanical systems (MEMS) have allowed
rapid technological advancement of miniaturized and rug-
ged spectrometers while the advancement in the field of
machine learning, and particularly chemometrics, has
allowed complex analysis of samples through spectroscopy
to be applied to everyday analysis. The advantages of using
MEMS sensors is that accurate and rapid material testing
can be conducted with handheld apparatus, at very low
power allowing the potential for intrinsically safe design to
be viable. Additionally, the use of near-infrared (NIR) spec-
trum sensors brings a unique advantage over photodiodes
as it can detect moisture within the NIR region.
Simtars is currently undertaking a viability study of
portable spectrometers for the rapid analysis of TIC as
part of a Coal Health and Safety Trust funded project (No.
20663) under laboratory conditions. The project is focusing
on devices available on the market, along with the applica-
tion of machine learning algorithms to determine TIC.
In this paper, the preliminary analysis to determine the
TIC of samples in real-time with chemometric modelling
using a portable instrument system employing near-infra-
red spectroscopy is presented. The classification model used
demonstrated a high accuracy of 76.92% for Category 2
(TIC =70%–80%), however, its overall accuracy of 56%
requires further analysis with the main difficulty encoun-
tered during the classification within the grouping of
Category 1 (TIC 70%).
METHODOLOGY
Spectrometer
The Stellarnet NIR ADK portable spectrometer was used
for this study. Utilizing a crossed Czerny-Turner optical
system, this NIR spectrometer operates within a wave-
length range spanning from 900 to 1700 nm, with a resolu-
tion of less than 5 nm and wavelength accuracy of less than
0.25 nm. It uses near-infrared light to analyze the vibra-
tional transitions within molecules, especially O-H, C-H,
and N-H bonds.
Sample Preparation
To compile a chemometric dataset, one hundred (100)
coal samples from Australian mines (Queensland and New
South Wales) were randomly selected for analysis. These 100
coal samples underwent proximate analysis for coal quality:
moisture, ash, volatile matter, and fixed carbon contents.
The results of the statistical analysis of the coal quality per
region (Queensland and New South Wales) are summa-
rized in Table 1. Levene Test and the One-Way ANOVA
determine if there are significant differences in indepen-
dent groups, providing insights into group variability and
potential statistical significance. Both the Levene Test and
the One-Way ANOVA yielded low test statistics and high
p-values for all coal quality parameters, except for moisture.
Specifically, the p-values for these parameters were consis-
tently greater than 0.05, indicating no statistically signifi-
cant variance among regions [6]. This suggests that the 100
coal samples were drawn from a similar sample group across
Queensland and New South Wales. The notable difference
observed in moisture content between the two regions, as
indicated by a p-value less than 0.05, may be attributed to
the preservation of the coal samples, considering that these
samples originated from a span of 20 years.
Table 1. Statistical Analysis of Coal Quality by Region
Coal Quality Levene Test One-Way ANOVA
Ash 0.21, p=0.65 0.19, p=0.67
Moisture 7.20, p=0.009 4.60, p=0.04
Volatile Matter 0.01, p=0.93 1.41, p=0.24
Fixed Carbon 2.32, p=0.13 0.16, p=0.69
The stone dust dosing regime was calculated based on
a polymodal distribution centered around the three regula-
tory limits of 70%, 80% and 85%, with a standard devia-
tion of 5%. This polymodal distribution was then randomly
applied to all coal samples resulting in the total of 300 coal/
stone dust samples with varying total incombustible con-
tent (TIC) values from the polymodal distribution. The
spread of the TIC in samples is displayed in Figure 1. For
quality assurance purposes, the TIC values were measured
by two independent laboratories and the results showed
good correlation.
fundamental differences of the tests (gravimetric vs spec-
trometry). For instance, CDEM provides qualitative results
(Go/No-Go) while laboratories provide a quantitative anal-
ysis. In addition, the CDEM specifications state that the
accuracy is ± 2 %,however, this is only true when using
completely dried samples. One percent (1%) moisture can
drop the IC readings by as much as 7% [1].
The rapid development of spectral sensors within the
last decade has resulted in numerous portable spectrom-
eters being developed and brought to market. This has
allowed for the identification of materials in the field rather
than the laboratory. More recently, advances in regres-
sion modelling along with higher resolution sensors now
allow for portable quantitative analysis of materials. For
instance, portable X-ray fluorescence has been used for
the analysis of ash content in coal [5]. Recent advances in
micro-electro-mechanical systems (MEMS) have allowed
rapid technological advancement of miniaturized and rug-
ged spectrometers while the advancement in the field of
machine learning, and particularly chemometrics, has
allowed complex analysis of samples through spectroscopy
to be applied to everyday analysis. The advantages of using
MEMS sensors is that accurate and rapid material testing
can be conducted with handheld apparatus, at very low
power allowing the potential for intrinsically safe design to
be viable. Additionally, the use of near-infrared (NIR) spec-
trum sensors brings a unique advantage over photodiodes
as it can detect moisture within the NIR region.
Simtars is currently undertaking a viability study of
portable spectrometers for the rapid analysis of TIC as
part of a Coal Health and Safety Trust funded project (No.
20663) under laboratory conditions. The project is focusing
on devices available on the market, along with the applica-
tion of machine learning algorithms to determine TIC.
In this paper, the preliminary analysis to determine the
TIC of samples in real-time with chemometric modelling
using a portable instrument system employing near-infra-
red spectroscopy is presented. The classification model used
demonstrated a high accuracy of 76.92% for Category 2
(TIC =70%–80%), however, its overall accuracy of 56%
requires further analysis with the main difficulty encoun-
tered during the classification within the grouping of
Category 1 (TIC 70%).
METHODOLOGY
Spectrometer
The Stellarnet NIR ADK portable spectrometer was used
for this study. Utilizing a crossed Czerny-Turner optical
system, this NIR spectrometer operates within a wave-
length range spanning from 900 to 1700 nm, with a resolu-
tion of less than 5 nm and wavelength accuracy of less than
0.25 nm. It uses near-infrared light to analyze the vibra-
tional transitions within molecules, especially O-H, C-H,
and N-H bonds.
Sample Preparation
To compile a chemometric dataset, one hundred (100)
coal samples from Australian mines (Queensland and New
South Wales) were randomly selected for analysis. These 100
coal samples underwent proximate analysis for coal quality:
moisture, ash, volatile matter, and fixed carbon contents.
The results of the statistical analysis of the coal quality per
region (Queensland and New South Wales) are summa-
rized in Table 1. Levene Test and the One-Way ANOVA
determine if there are significant differences in indepen-
dent groups, providing insights into group variability and
potential statistical significance. Both the Levene Test and
the One-Way ANOVA yielded low test statistics and high
p-values for all coal quality parameters, except for moisture.
Specifically, the p-values for these parameters were consis-
tently greater than 0.05, indicating no statistically signifi-
cant variance among regions [6]. This suggests that the 100
coal samples were drawn from a similar sample group across
Queensland and New South Wales. The notable difference
observed in moisture content between the two regions, as
indicated by a p-value less than 0.05, may be attributed to
the preservation of the coal samples, considering that these
samples originated from a span of 20 years.
Table 1. Statistical Analysis of Coal Quality by Region
Coal Quality Levene Test One-Way ANOVA
Ash 0.21, p=0.65 0.19, p=0.67
Moisture 7.20, p=0.009 4.60, p=0.04
Volatile Matter 0.01, p=0.93 1.41, p=0.24
Fixed Carbon 2.32, p=0.13 0.16, p=0.69
The stone dust dosing regime was calculated based on
a polymodal distribution centered around the three regula-
tory limits of 70%, 80% and 85%, with a standard devia-
tion of 5%. This polymodal distribution was then randomly
applied to all coal samples resulting in the total of 300 coal/
stone dust samples with varying total incombustible con-
tent (TIC) values from the polymodal distribution. The
spread of the TIC in samples is displayed in Figure 1. For
quality assurance purposes, the TIC values were measured
by two independent laboratories and the results showed
good correlation.