Student Paper
Transfer Learning for Flight Loads Estimation
H.G. Aydogan
1,4
, H. Fayek
2,4
, X. Zhang
2,4
, K.E. Niessen
3
, D.O. Franke
3
and P. Marzocca
1,4
1
RMIT University, School of Engineering, Melbourne, Victoria, 3000, Australia
2
RMIT University, School of Computing Technology, Melbourne, Victoria, 3000, Australia
3
Defence Science and Technology Group, Melbourne, Victoria, 3207, Australia
4
RMIT University, Sir Lawrence Wackett Defence and Aerospace Centre Victoria, 3000, Australia
Abstract
Developments in flight test data systems have led to an abundance of data which can be
harnessed via machine learning for flight load estimations. Main measurements used for the
load estimations are strains. However, the transition from strains to loads requires load
equations obtained from calibration load tests. Estimated loads obtained by strain data are
used for several purposes, from predictive maintenance to development projects. The fidelity
of the obtained loads highly affects these activities. Prior research showed that higher
accuracy could be achieved with artificial neural network models instead of classical
regression analysis. However, the models provided some inaccurate results for some load
components during the transition to predicting flight data, and further development is required
to improve the flight loads predictions.
Using the idea that calibration test data consists of the knowledge of the loads, our aim is to
pass on this knowledge to the flight domain to increase the fidelity of the estimated loads.
This is possible with machine learning algorithms classified under transfer learning. Transfer
learning deals with the transfer of knowledge obtained in one domain to another domain. A
load model developed using the calibration test data as a source for training artificial neural
networks can be fine-tuned using flight test data, which also includes additional information
to increase the accuracy. The research in this paper focuses on applying transfer learning to
convey the knowledge in the calibration domain to the flight domain and shows the potential
to overcome the poor generalisation in transition to flight in order to achieve high fidelity
loads outputs when using such a method.
Keywords: flight loads prediction, flight test, load calibration test, transfer learning
Introduction
Load calibration and flight load tests are used for validation and verification of the loads
calculated in aircraft design. They also assure the aircraft structural integrity under severe and
repeated load cases over its life. Loads cannot be directly measured by strain gauges installed
on aircraft. To obtain the loads under different conditions, strain gauges are installed in
several locations on the aircraft to observe the behaviour of the related aircraft component
under shear, bending and torsional loads. Load calibration tests provide the data to establish
relationships between measurements from strain gauges and the applied loads. These
relationships are given as inputs to calculate the loads during subsequent flight tests.
As the next step in the load estimation process, flight tests need to be performed. During the
flight test, the aircraft performs different manoeuvres, which causes different types of loads
over the aircraft structure. In addition, shear, bending moment and torsional loads are also
related to each other. Measurements are taken from the flight test instrumentation (strain
