Predicting a Maximum Stress using Machine Learning and Parametric Flight Data

Abstract

Mercer Engineering Research Center (MERC) developed a neural network-based regression method for predicting maximum stress per flight values at four structural tracking locations on the United States Air Force HH-60G helicopter airframe using Individual Vehicle Health and Usage Monitoring (IVHMS) data. Maximum stress per flight is utilized when evaluating a failure criterion within the HH-60G service life analysis, so an accurate, fleet-wide estimation of maximum stress magnitude and likelihood is critical for accurate service life determinations. The model was trained using parametric flight data time histories (from IVHMS) and stress time histories from a strain survey aircraft. The stress time histories were developed from the strain signals using two different methods depending on the location of strain gauges in the vicinity of the tracking locations. For two of the tracking locations, they were derived from a global finite element model using a collection of strain gauge signals throughout the strain survey aircraft. At the other two tracking locations, the strain time histories were derived from single strain gauges installed in close proximity to the tracking locations. Multiple regression methods and input data configurations were evaluated in order to identify an appropriate regression method that predicts a maximum stress per flight accurately without over-fitting the training data. MERC identified that the relationship between parametric flight data and aircraft component strain can be exploited to a high level of accuracy using machine learning regression tools. Achieving a high level of accuracy required an extensive review of independent and dependent variable data quality and thoughtful consideration of model inputs.