Predictive Maintenance For Rotorcraft: A Data-Driven Approach With Machine Learning Algorithms For Enhanced Availability

Abstract

Predictive maintenance represents a critical advancement in rotorcraft fleet management, offering substantial improvements in operational availability while reducing unscheduled downtime costs. This paper presents two complementary data-driven approaches utilizing machine learning algorithms for enhanced rotorcraft component monitoring. The first methodology addresses transmission system health monitoring through an unsupervised Principal Component Analysis algorithm that consolidates approximately one thousand health indices from accelerometer data into a single normalized health score. This approach successfully detected transmission failures an average of twenty-four flight hours in advance across a fleet of approximately seven hundred aircraft, achieving false positive rates below one per ten flight hours and false negative rates below one per sixty flight hours based on millions of flight hours of operational data. The second methodology employs virtual sensor technology combined with machine learning models to predict wear indices and remaining useful life for rotor components that cannot be directly instrumented during service operations. Gradient boosting and convolutional neural network approaches were implemented using flight data and inferred virtual sensor measurements to estimate component degradation. Both methodologies demonstrate fleet-wide applicability, providing maintenance organizations with actionable insights for proactive component replacement scheduling and logistics optimization.