Estimating The Health of Helicopter Turbine Engines by Means of
Regression and Classification Using a Probabilistic Neural Network
Tyler Romano
1
, Nathan Siegel
2
, Samuel T. Willis III
3
, William Henn
4
, Rishie Seshadri
5
1,2,3,4,5
Belcan, Windsor, CT,06095,United States
tromano@belcan.com
nsiegel@belcan.com
stwillis@belcan.com
whenn@belcan.com
rseshadri@belcan.com
ABSTRACT
This paper presents team Mad SoftMax’s approach to the
2024 data challenge presented by the Prognostics and Health
Management Society. The competition tasked participants
with estimating the health of helicopter turbine engines by
calculating torque margin via regression and classifying en-
gines as either healthy or faulty. Probabilistic regression was
employed to estimate the torque margin at each measurement,
and a neural network classifier was used to categorize each
observation within the dataset as belonging to a healthy or
faulty engine. Both the regression and classification networks
were developed using open-source libraries such as Tensor-
Flow. These networks were tested in isolation using training
data and evaluated for performance before integration for use
on evaluation data. The team was able to successfully con-
struct a system of models that achieved a final score of 0.849
out of a maximum score of 1.
1. INTRODUCTION
The 2024 PHM Society Data Challenge tasked participants
with estimating the health of helicopter turbine engines
(PHMSociety, 2024). The data for this year’s challenge was
gathered from seven helicopter engines of the same make and
model. These engines each captured metric data such as: out-
side air temperature oat, mean gas temperature mgt, pressure
altitude pa, indicated airspeed ias, net power np, compressor
speed ng, and torque measured trq
meas
. The data was split
into three datasets, a training set, a daily test dataset, and a
final validation dataset. The engines in all sets of data were
shuffled and had any asset identifiers removed to anonymize
the data. The initial four engines made up the training dataset
while the remaining three were withheld for the testing and
Tyler Romano et al. This is an open-access article distributed under the terms
of the Creative Commons Attribution 3.0 United States License, which per-
mits unrestricted use, distribution, and reproduction in any medium, provided
the original author and source are credited.
validation phases. The test data was available to test against
daily to allow for teams to validate their approach methods
of classifying engine health and determination of torque mar-
gins. The validation data, however, was only available to run
a single time at the end of the challenge.
The teams entered in the challenge were asked to provide
three items for each data point. Teams were asked to classify
the health of the engine as either nominal or faulty, as well
as assigning a confidence to that engine state determination.
In addition, teams were tasked with determining an estima-
tion for torque margin expressed as a probability distribution
for each data point. Daily scoring of models was available
to allow teams to compare against testing data to ensure the
accuracy of their approach.
2. METHODS
Upon entering the competition, the team sought to understand
the importance of each parameter within the dataset. The first
step in this process was to separately learn about the environ-
mental and engine performance parameters to determine if
and how these parameters may influence each other. The en-
vironmental parameter set made up of oat, pa, and ias helped
to understand the flight envelope without additional data such
as weight and true altitude (FAA, 2019). The engine perfor-
mance parameters included: mgt, np, ng, and trq
meas
. While
trq
meas
and the torque margin trq
mar
were provided in the
training set, the target torque trq
tg t
was not. The relation be-
tween the three parameters is shown in Eq. (1), which was
provided in the problem statement (PHMSociety, 2024).
1
