Towards Development of Virtual Strain Gauges for Rotorcraft using Neural Networks and Transfer Learning

One of the modernization strategies proposed by the Army under Future Vertical Lift is the digitization of the cockpit and reduction of dependence on analog instrumentation. The Army possesses a large amount of flight test data from heavily instrumented Chinook rotorcraft that includes strain gauges which was produced over a span of many years at a significant cost. This work seeks to develop a methodology for predicting the stress loads measured by these strain gauges using properties available through standard data bus channels present in both highly instrumented flight test as well as operational Chinooks. The initial stages of this work were focused on the development of virtual strain gauges for the flight test dataset using neural networks, with the target application being usage for operational rotorcraft which do not have these strain gauges. In this talk we will discuss the development of a modular neural network and transfer learning pipeline used to develop virtual strain gauges for operational Chinook rotorcraft using data from both the test and operational rotorcraft. We demonstrate that a single common output signal can be used to learn a mapping from the input features of the test rotorcraft to the operational model, potentially allowing for virtualization of strain gauges not present on operational Chinooks. Neural network design was found to heavily impact the stability of training in the transfer learning step. We will present the results from this work and discuss lessons learned as they may pertain to other efforts to utilize machine learning for condition-based maintenance throughout the DoD.

PRESENTER

Wissink, Andrew (Presenting for Boyer, Mathew)
Boyer, Mathew
mathew.j.boyer.ctr@army.mil
570-294-8957

DoD HPCMP PET

CO-AUTHORS

Brackbill, Christian
christian.r.brackbill.civ@army.mil

Finckenor, Jeffrey
jeffrey.l.finckenor.ctr@army.mil

Wissink, Andrew
andrew.m.wissink.civ@army.mil

CATEGORY

Artificial Intelligence / Machine Learning usage for HPC Applications

SYSTEMS USED

Narwhal

SECRET

No