3
run these fire simulations using the Ventsim software [18].
The results of each scenario in terms of airflow distribution
in every airway, the fire location airway, and fire size were
all recorded in a tabular format to form a dataset with each
scenario’s data stored in a row.
This dataset contains 38,300 rows or records of data
each representing a fire simulation for a fire size in an air-
way. The fire location, in terms of airway ID, is considered
categorical response data. With the 766 airways, each air-
way flow rate is considered as one predictor or feature, lead-
ing to 766 features in this dataset.
An ML model relies on the input/output data to
learn the underlying physical model behind the data. It
is hypothesized that knowing the flow distribution can
only lead to at least one solution in terms of fire location.
With this concept in mind, we developed an ML-based
predictive model that can use the flow distribution data to
determine the fire location causing such flow distribution.
The generated dataset was split into a training set (70%)
and a testing set (30%). Two different prediction models
were considered, one for the fire location and one for fire
size. For each model, the training set was used to train the
model, and then the test dataset was used to test the trained
model and determine the prediction accuracy. The accuracy
of the fire location was calculated as the ratio of the number
of correctly predicted fire locations to the total number of
fire locations.
Since the response variable is categorical data, the suite
of ML algorithms in MATLAB was tested. The random
forest was chosen from the MATLAB statistical and ML
toolbox following the performance of the selected algo-
rithm in previous research by the authors [3]. The fire size is
a numeric value, and a regression-based ML algorithm was
used to predict the fire size using the changes in the airflow
caused by a fire. The model performance is then calculated
using the test data set by calculating the R2, coefficient of
determination, between the predicted fire size values and
the fire size used in the fire simulation model.
RESULTS AND DISCUSSION
The complexity of an ML model depends on the number
of features used. It is important to investigate and develop a
Figure 1. Partial layout of the example mine’s ventilation network
run these fire simulations using the Ventsim software [18].
The results of each scenario in terms of airflow distribution
in every airway, the fire location airway, and fire size were
all recorded in a tabular format to form a dataset with each
scenario’s data stored in a row.
This dataset contains 38,300 rows or records of data
each representing a fire simulation for a fire size in an air-
way. The fire location, in terms of airway ID, is considered
categorical response data. With the 766 airways, each air-
way flow rate is considered as one predictor or feature, lead-
ing to 766 features in this dataset.
An ML model relies on the input/output data to
learn the underlying physical model behind the data. It
is hypothesized that knowing the flow distribution can
only lead to at least one solution in terms of fire location.
With this concept in mind, we developed an ML-based
predictive model that can use the flow distribution data to
determine the fire location causing such flow distribution.
The generated dataset was split into a training set (70%)
and a testing set (30%). Two different prediction models
were considered, one for the fire location and one for fire
size. For each model, the training set was used to train the
model, and then the test dataset was used to test the trained
model and determine the prediction accuracy. The accuracy
of the fire location was calculated as the ratio of the number
of correctly predicted fire locations to the total number of
fire locations.
Since the response variable is categorical data, the suite
of ML algorithms in MATLAB was tested. The random
forest was chosen from the MATLAB statistical and ML
toolbox following the performance of the selected algo-
rithm in previous research by the authors [3]. The fire size is
a numeric value, and a regression-based ML algorithm was
used to predict the fire size using the changes in the airflow
caused by a fire. The model performance is then calculated
using the test data set by calculating the R2, coefficient of
determination, between the predicted fire size values and
the fire size used in the fire simulation model.
RESULTS AND DISCUSSION
The complexity of an ML model depends on the number
of features used. It is important to investigate and develop a
Figure 1. Partial layout of the example mine’s ventilation network