4
procedure to automatically determine the most influencing
airways correlated to each fire location while maintaining
acceptable prediction accuracy. We used MATLAB’s built-
in function to determine the most important airways. The
importance of an input variable, airway airflow, is calcu-
lated by the ML algorithm in terms of its ability to predict
the output.
Based on the feature importance function applied to
the training dataset in MATLAB, the obtained feature
importance of all 766 airways was sorted in ascending order.
After that, it was determined how the number of selected
features can affect the model performance in predicting the
fire location. To accomplish this, the training and testing of
the ML-based fire location was repeated for each set of the
selected features. Figure 2 shows how the model prediction
accuracy improves by increasing the number of used fea-
tures. Airflow data from the most important seven features/
airways lead to about 92% accuracy in fire location predic-
tion. Increasing the number of features to 20 can increase
this accuracy to about 96%.
Inspection of the ventilation network and the location
of these most significant airways confirms that they are well
distributed throughout the ventilation network. This indi-
cates a good representation of strategic airways for monitor-
ing airflow changes during a fire emergency as part of fire
characterization for size and location.
A future field test will be needed to test the predic-
tive model obtained from this study in collaboration with
the partner mine. The seven important airways needed to
achieve over 90% accuracy account for only 1% of the
main airways, marked with a circle in Figure 2. Increasing
the number of important airways to 20 will increase this
percentage to only 3% of total main airways, marked with
an arrow in Figure 2.
Model performance is assessed based on the speed of
training and prediction and how accurately the trained
model predicts the response variable, which in this case is
the fire location.
Using only the features corresponding with these seven
airways, the selected algorithm was trained against the
training dataset using the fire size as the response variable
and tested using the testing dataset. The model training
with seven features took about 23 seconds. The predicted
fire sizes for the test dataset are only using seven features or
important airways. The model prediction takes only 1 sec-
ond. However, preparing the data set which is based on
many fire simulations using Ventsim, 83000 simulations
in this case, requires a significant amount of computer
time but only required whenever the ventilation model is
updated.
The performance of the trained fire size model using
the seven important airways is shown in Figure 3. Fire
size prediction provides an R2 of 0.88 with seven signifi-
cant features. In addition to the R2 performance metric,
the average percentage error was calculated as an additional
prediction performance metric. The point percentage of
predicted values versus actual fire sizes was calculated for
each fire scenario and then averaged as an absolute value
also led to an average point accuracy percentage of 88%.
Adding more features or important airways can improve
the fire size prediction slightly by about 2-3 percent.
The ML-based modeling technique used in this study
relies on the airflow distribution obtained from the ventila-
tion network. Consequently, the model’s performance and
accuracy depend on how well the ventilation network is
calibrated.
Despite excellent prediction accuracy for such a large
network of airways, it can be seen from Figure 3 that smaller
fires tend to be overestimated while larger fires tend to be
underestimated. However, overestimating a small fire size
can be considered a favorable (conservative) outcome. In
other words, it would be on the safe side to have a small fire
reported as a somewhat larger fire. Additionally, this model
does not account for normal operating airflow fluctuations
as the model relies on significant changes in the airflow that
are larger than the normal operating airflow fluctuations or
measurement error.
It can also be seen from the results that the fire loca-
tion prediction provides better performance with the same
Figure 2. Fire location prediction accuracy as a function of
the number of important airways
procedure to automatically determine the most influencing
airways correlated to each fire location while maintaining
acceptable prediction accuracy. We used MATLAB’s built-
in function to determine the most important airways. The
importance of an input variable, airway airflow, is calcu-
lated by the ML algorithm in terms of its ability to predict
the output.
Based on the feature importance function applied to
the training dataset in MATLAB, the obtained feature
importance of all 766 airways was sorted in ascending order.
After that, it was determined how the number of selected
features can affect the model performance in predicting the
fire location. To accomplish this, the training and testing of
the ML-based fire location was repeated for each set of the
selected features. Figure 2 shows how the model prediction
accuracy improves by increasing the number of used fea-
tures. Airflow data from the most important seven features/
airways lead to about 92% accuracy in fire location predic-
tion. Increasing the number of features to 20 can increase
this accuracy to about 96%.
Inspection of the ventilation network and the location
of these most significant airways confirms that they are well
distributed throughout the ventilation network. This indi-
cates a good representation of strategic airways for monitor-
ing airflow changes during a fire emergency as part of fire
characterization for size and location.
A future field test will be needed to test the predic-
tive model obtained from this study in collaboration with
the partner mine. The seven important airways needed to
achieve over 90% accuracy account for only 1% of the
main airways, marked with a circle in Figure 2. Increasing
the number of important airways to 20 will increase this
percentage to only 3% of total main airways, marked with
an arrow in Figure 2.
Model performance is assessed based on the speed of
training and prediction and how accurately the trained
model predicts the response variable, which in this case is
the fire location.
Using only the features corresponding with these seven
airways, the selected algorithm was trained against the
training dataset using the fire size as the response variable
and tested using the testing dataset. The model training
with seven features took about 23 seconds. The predicted
fire sizes for the test dataset are only using seven features or
important airways. The model prediction takes only 1 sec-
ond. However, preparing the data set which is based on
many fire simulations using Ventsim, 83000 simulations
in this case, requires a significant amount of computer
time but only required whenever the ventilation model is
updated.
The performance of the trained fire size model using
the seven important airways is shown in Figure 3. Fire
size prediction provides an R2 of 0.88 with seven signifi-
cant features. In addition to the R2 performance metric,
the average percentage error was calculated as an additional
prediction performance metric. The point percentage of
predicted values versus actual fire sizes was calculated for
each fire scenario and then averaged as an absolute value
also led to an average point accuracy percentage of 88%.
Adding more features or important airways can improve
the fire size prediction slightly by about 2-3 percent.
The ML-based modeling technique used in this study
relies on the airflow distribution obtained from the ventila-
tion network. Consequently, the model’s performance and
accuracy depend on how well the ventilation network is
calibrated.
Despite excellent prediction accuracy for such a large
network of airways, it can be seen from Figure 3 that smaller
fires tend to be overestimated while larger fires tend to be
underestimated. However, overestimating a small fire size
can be considered a favorable (conservative) outcome. In
other words, it would be on the safe side to have a small fire
reported as a somewhat larger fire. Additionally, this model
does not account for normal operating airflow fluctuations
as the model relies on significant changes in the airflow that
are larger than the normal operating airflow fluctuations or
measurement error.
It can also be seen from the results that the fire loca-
tion prediction provides better performance with the same
Figure 2. Fire location prediction accuracy as a function of
the number of important airways