Machine Learning of Part Variability for Predictive Maintenance

Extensive testing at substantial cost is a major part of military propulsion development programs. These programs involve numerous component and engine tests aimed at producing safe and reliable engines. Despite these efforts, engines still experience unexpected failures. Geometric variability of parts within design tolerances is a well-known source of uncertainty in overall reliability. USAF sustainment activities increasingly include geometric measurements of parts to quantify variations, but these data are generally ignored in the test process. In the proposed STTR program, the overall technical objective it to use the geometry and response data from test of one set of parts to describe the response behavior of a second population of parts of known, or formulated geometry. The University of Notre Dame will extend methods for parameterizing geometry and computation of response sensitivities for utilization of measured geometry data by the test community. APEX Turbine Testing Technologies will integrate these methods in GageMap, a commercially developed FEA post-processing product, together with response measurements, to describe the response for other geometries. This allows the USAF to identify geometric features that can produce higher responses. This can assist in focusing inspections, adjustment of tolerances, or perhaps culling parts as needed for fleet management.