Estimation of confidence margins for Direct Load Recognition (DLR) using supervised and unsupervised machine learning
Abstract
One of the key challenges of any helicopter manufacturer lies in the competiveness of its products compared to its competitors. Operating costs is clearly one of them. Taking into account the complexity to maintain in operational and airworthy conditions a helicopter, substantial gains are achievable through the optimization of the Direct Maintenance Cost (DMC). Over the last decades, significant improvements have been made on mandatory enablers of such technologies, as for example Data Based Models (DBM), Predictive Maintenance (PM) or Conditions (Usage) Based Maintenance (CBM). More and more complex approaches were elaborated, getting always closer from true helicopter behavior, but certification has often been a hindrance to development, as infallible evidences are difficult to build. Thus, instead of pushing for additional improvement to gain limited performances, this paper presents an alternative way, taking uncertainties as an intrinsic component of the prediction. Similarly to fatigue metallic, choice is made to live with uncertainties, but to confine them thanks to statistics approaches based on the estimation of confidence intervals around a prediction. Through the paper, three methods have been explored, Euclidean proximity (K-Nearest Neighbors ,KNN), unsupervised learning (Self-Organizing Map) and supervised learning (Conformal Prediction). This broad, largescale exploratory analysis, has demonstrated that reasonable intervals, which would not prevent any gain from a Condition Based Maintenance approach, can be defined and shall be now compared with certification authority's preliminary targets.