2
knowledge about the location and size of the fire source.
For these reasons, a ventilation diagnostic methodology is
proposed to be used in this study that applies the use of the
machine learning (ML) technique that can be considered as
reverse modeling to determine the location and size of an
underground fire. ML-based methods have been utilized in
other industries to perform predictive maintenance, health
monitoring, financial portfolio forecasting, and advanced
driver assistance systems [8]. The foundation of ML-based
modeling is the simple idea of getting computers to learn
from data [9]. Many ML and artificial intelligence appli-
cations address mining problems, including predicting
mechanical failures on equipment, production optimiza-
tion, and ore body delineation [10]. Moreover, NIOSH
researchers conducted machine learning analysis to exam-
ine whether a model could be built to assess the likelihood
of dynamic ground failure occurrence, leading to seismic
events, based on geochemical and petrographic data [11].
An ML-based model, once properly trained using
many fire simulations using fire and ventilation software to
build the required dataset, can be used to diagnose impend-
ing ventilation issues in response to a mine fire without the
need for a lengthy simulation process to locate and charac-
terize a fire source. Updating the required dataset will be
needed whenever significant changes in the ventilation net-
work occur. NIOSH researchers have previously developed
tools to characterize an unknown fire source by analyzing
the CO arrival time data from the AMS installed in a mine
[12-14].
The objective of this paper is to present and demon-
strate the applicability of a previously NIOSH-developed
novel methodology [3,15] to characterize an underground
mine fire in large ventilation networks. The methodology is
based on the application of an ML technique using only the
ventilation network airflow response under the influence of
a fire source. Previous work [3,15] showed results with rea-
sonable accuracy from the application of this methodology
to the SRCM experimental mine at the NIOSH Pittsburgh
Mining Research Division.
The evaluation of the performance and efficiency of
the developed method when applied to a large ventilation
network is the goal of this paper and is presented in the
following section.
MACHINE LEARNING-BASED MODEL
The ML-based modeling used in this work has been demon-
strated using the ventilation network of the Safety Research
Coal Mine (SRCM), an underground experimental mine
facility located at the Bruceton Campus of the Pittsburgh
Mining Research Division of NIOSH [3,15].
The selected example mine is a metal mine with mul-
tiple zones and several ventilation controls such as main
fans, booster fans, doors, and regulators that bring fresh air
to the work zones comprising over 4,000 airway segments.
The quantity of the airflow distribution is controlled by the
ventilation controls. Figure 1 shows the partial layout of the
large example mine ventilation network. The complexity of
such a model can be seen from the layout, using true 3D
coordinates of each airway.
The model was provided in the Ventsim ventilation and
fire simulation software which is widely used in metal/non-
metal mines. The model is then imported into MATLAB
where a macro-based data generator is used to generate
and process data for the ML-based modeling. Each airway
is known by its airway number in the ventilation model,
which is also used to identify a fire source location as well
as all airflow data in the dataset used for the ML model
development. The very low flow-rate airways, as nonpartic-
ipating airways, are automatically excluded from the data-
set. These airways are typically the ones that are dead-end
airways or block sections. The focus of the model is on the
active airways with numerically significant airflows, larger
than 0.05 m/s based on a typical point airflow sensor mea-
surement error.
Furthermore, a continuous airway that has been split
into multiple segments for geometry compliance and more
realistic visualization can be represented by one of those
segments. For consistency, the middle segment airway is
selected, rendering the remaining airways along the same
continuous path redundant. This will reduce the size of
the dataset used in the proposed ML-based model as each
remaining airway is a feature considered for fire simulation.
After excluding the inactive and redundant airways, the
remaining 766 active airways were considered in this study
as potential fire sources.
The ventilation model input was fed into a MATLAB
macro to build a large set of fire scenarios of different fire
sizes placed in any airway of the network. For every airway,
50 fire source scenarios were randomly selected from a range
of 6.7 MW to 36.7 MW. The maximum fire size of mine
equipment was determined based on the literature available
on the heat release rates of equipment tires, diesel fuel, and
hydraulic oil, used to estimate the heat release rate (HRR)
for wheel loaders and drilling rigs [7,16-17]. For diesel fuel,
the total fuel tank capacity is used to estimate the HRR. For
a typical wheel loader used in the example mine, the maxi-
mum HRR was estimated to be 7–15 MW. For the drilling
rig, the maximum HRR was estimated to be in the range of
16–35 MW. A total of 38,300 fire scenarios were generated.
A computer workstation with 24 physical cores was used to
knowledge about the location and size of the fire source.
For these reasons, a ventilation diagnostic methodology is
proposed to be used in this study that applies the use of the
machine learning (ML) technique that can be considered as
reverse modeling to determine the location and size of an
underground fire. ML-based methods have been utilized in
other industries to perform predictive maintenance, health
monitoring, financial portfolio forecasting, and advanced
driver assistance systems [8]. The foundation of ML-based
modeling is the simple idea of getting computers to learn
from data [9]. Many ML and artificial intelligence appli-
cations address mining problems, including predicting
mechanical failures on equipment, production optimiza-
tion, and ore body delineation [10]. Moreover, NIOSH
researchers conducted machine learning analysis to exam-
ine whether a model could be built to assess the likelihood
of dynamic ground failure occurrence, leading to seismic
events, based on geochemical and petrographic data [11].
An ML-based model, once properly trained using
many fire simulations using fire and ventilation software to
build the required dataset, can be used to diagnose impend-
ing ventilation issues in response to a mine fire without the
need for a lengthy simulation process to locate and charac-
terize a fire source. Updating the required dataset will be
needed whenever significant changes in the ventilation net-
work occur. NIOSH researchers have previously developed
tools to characterize an unknown fire source by analyzing
the CO arrival time data from the AMS installed in a mine
[12-14].
The objective of this paper is to present and demon-
strate the applicability of a previously NIOSH-developed
novel methodology [3,15] to characterize an underground
mine fire in large ventilation networks. The methodology is
based on the application of an ML technique using only the
ventilation network airflow response under the influence of
a fire source. Previous work [3,15] showed results with rea-
sonable accuracy from the application of this methodology
to the SRCM experimental mine at the NIOSH Pittsburgh
Mining Research Division.
The evaluation of the performance and efficiency of
the developed method when applied to a large ventilation
network is the goal of this paper and is presented in the
following section.
MACHINE LEARNING-BASED MODEL
The ML-based modeling used in this work has been demon-
strated using the ventilation network of the Safety Research
Coal Mine (SRCM), an underground experimental mine
facility located at the Bruceton Campus of the Pittsburgh
Mining Research Division of NIOSH [3,15].
The selected example mine is a metal mine with mul-
tiple zones and several ventilation controls such as main
fans, booster fans, doors, and regulators that bring fresh air
to the work zones comprising over 4,000 airway segments.
The quantity of the airflow distribution is controlled by the
ventilation controls. Figure 1 shows the partial layout of the
large example mine ventilation network. The complexity of
such a model can be seen from the layout, using true 3D
coordinates of each airway.
The model was provided in the Ventsim ventilation and
fire simulation software which is widely used in metal/non-
metal mines. The model is then imported into MATLAB
where a macro-based data generator is used to generate
and process data for the ML-based modeling. Each airway
is known by its airway number in the ventilation model,
which is also used to identify a fire source location as well
as all airflow data in the dataset used for the ML model
development. The very low flow-rate airways, as nonpartic-
ipating airways, are automatically excluded from the data-
set. These airways are typically the ones that are dead-end
airways or block sections. The focus of the model is on the
active airways with numerically significant airflows, larger
than 0.05 m/s based on a typical point airflow sensor mea-
surement error.
Furthermore, a continuous airway that has been split
into multiple segments for geometry compliance and more
realistic visualization can be represented by one of those
segments. For consistency, the middle segment airway is
selected, rendering the remaining airways along the same
continuous path redundant. This will reduce the size of
the dataset used in the proposed ML-based model as each
remaining airway is a feature considered for fire simulation.
After excluding the inactive and redundant airways, the
remaining 766 active airways were considered in this study
as potential fire sources.
The ventilation model input was fed into a MATLAB
macro to build a large set of fire scenarios of different fire
sizes placed in any airway of the network. For every airway,
50 fire source scenarios were randomly selected from a range
of 6.7 MW to 36.7 MW. The maximum fire size of mine
equipment was determined based on the literature available
on the heat release rates of equipment tires, diesel fuel, and
hydraulic oil, used to estimate the heat release rate (HRR)
for wheel loaders and drilling rigs [7,16-17]. For diesel fuel,
the total fuel tank capacity is used to estimate the HRR. For
a typical wheel loader used in the example mine, the maxi-
mum HRR was estimated to be 7–15 MW. For the drilling
rig, the maximum HRR was estimated to be in the range of
16–35 MW. A total of 38,300 fire scenarios were generated.
A computer workstation with 24 physical cores was used to