Abstract

Modern gas turbine engines have been designed to have high levels of reliability and low incidences of failures. The high number of components, and the economic necessity to replace functional systems in a modular manner in the field to maintain operation, can make fault classification problematic if too few specific failures can be identified in historic datasets. Such cases present the problem of having to perform cluster analysis despite the true number of clusters being unknown a-priori—this constitutes a generic problem which is not unique to engine health monitoring. Whilst a variety of approaches to accommodate this deficiency have been presented in the literature, the majority seek a unique solution for the k-number of clusters. By contrast, this work applies a multi-objective approach that returns a Pareto optimal set of solutions. This approach is particularly suitable when the number of modes of operation (clusters) is unknown, and employs a new hybridized, multi-objective optimization algorithm that provides the solutions for k-means clustering. In the present work, the computational load of clustering gas turbine monitoring data using the proposed hybridized optimization algorithm is compared to the load imposed by using two alternative state-of-the-art multi-objective optimizers (differential evolution and particle swarm). It is shown that the new algorithm both accelerates convergence and, reduces the computational overhead for this classification task.