Session: 26-01 Probabilistic and Machine Learning Methods Development and Applications

Paper Number: 125838

125838 - Industrial Applications of Transfer Learning Methods for Aircraft Propulsion Systems 

Sparse data problems challenge engineers throughout the lifecycle of aircraft propulsion systems. For instance, data from legacy systems may not be applicable to novel high-pressure turbine blade configurations, driving substantial computational and test efforts to design and validate the blade. Once the blade is fielded, some manufacturers implement predictive maintenance approaches to reduce downtime and associated costs of sustaining the fleet. Shortly after entry-into-service for a new design or re-design, limited field observations are available with which to construct prognostic models for the life of the blade. In this paper we explore transfer learning methods in machine learning to overcome these challenges. The goal is to borrow strength from one setting with abundant data—available for legacy systems that may have different configurations and/or operations—and transfer the embedded knowledge to model new systems in a different setting where fewer data are available. Transfer learning methods have been implemented for a broad range of machine learning approaches, and in this paper we highlight several industrial propulsion applications from various phases of the propulsion system lifecycle where transfer learning is applied with three frameworks that GE has developed: Bayesian Hybrid Modeling, Probabilistic Deep Neural Networks, and Physics Discovery. The added benefit of leveraging data from similar contexts is examined for each framework.

Presenting Author: Ryan Jacobs GE Aerospace Research

Presenting Author Biography: As the Military Sustainment Platform Leader at GE Aerospace Research, Ryan advances innovative technologies for inspection, repair, cleaning, and Condition Based Maintenance Plus (CBM+). Building on 10+ years of experience in improving engineering decision making through the development and application of probabilistic machine learning and design methods, he has led efforts spanning aviation services, military sustainment, multi-mission design, and materials informatics for GE Aerospace and government customers.

Prior to joining GE Aerospace Research, Ryan was a Principal Systems Engineer at MITRE Corporation. During his time at MITRE he led teams conducting systems analysis across government sponsors including the U.S. Air Force, USSTRATCOM, U.S. Navy, and DARPA. He graduated with MS and PhD degrees in Aerospace Engineering from Georgia Tech and a BS in Aerospace Engineering from Embry-Riddle Aeronautical University.

Authors:

Ryan Jacobs GE Aerospace Research
Sandipp Krishnan GE Aerospace Research
Anindya Bhaduri GE Aerospace Research
Lele Luan GE Aerospace Research
Piyush Pandita GE Aerospace Research
Sayan Ghosh GE Aerospace Research
Liping Wang GE Aerospace Research