Machine Learning of Part Variability for Predictive Maintenance

High Cycle Fatigue (HCF) characterization and maintenance accounts for a significant portion of the overall life cycle cost of most military propulsion systems.  A key variable that drives HCF margin is dynamic response which is directly related to the geometry of each part.  This is especially true of integrally bladed rotors (IBRs, or blisks).  It has been well established that HCF is a probabilistic phenomenon and only results in cracking for the most susceptible parts.  Such susceptibility can include geometric features and/or damage such as foreign object damage (FOD).  Nearly all of the effort put into HCF characterization testing is aimed at understanding dynamic responses of instrumented parts as a function of engine operation.  Little, explicit attention is given to how variations in part-to-part geometry affects the dynamic response however.    The proposed effort seeks to extend HCF assessment to include geometric variations by predicting the dynamic response of any part based entirely on its particular geometry.  To do this, this proposal will employ a sensitivity based method to compute mode shapes based on deviation of the part from the nominal geometry, and the dynamic characteristics and response history of the nominal part.  A machine learning routine will be implemented to parameterize the geometric variations and compute mode-shape sensitivities.  Integration of response histories from HCF characterization testing and predicted dynamic characteristics of some part of interest will be done by extension of a commercially marketed product called GageMap.   In the proposed STTR program, the University of Notre Dame will have primary responsibility for development of the machine learning, parameterization and sensitivity formulation.  APEX Turbine will provide integration of these technologies with the GageMap product to describe the response behavior of parts in the fleet based on data and geometry of parts tested under development or during diagnostic test programs.  An initial validation demonstration will be performed using data generated from a University of Notre Dame transonic rotor.