MACHINE LEARNING TECHNIQUES FOR DEVELOPMENT OF A CONDITION-BASED MAINTENANCE PROGRAM FOR NAVAL PROPULSION PLANTS

摘要

In this thesis, we investigate a specific type of machine learning (ML) algorithm, specifically a support vector machine (SVM) regressor, as the foundation behind a condition-based maintenance (CBM) program for the major components affecting a naval propulsion system (NPS). This program is designed to specifically monitor the degradation of the ship’s engines, the propeller, and the hull. Simulated data generated in previous work by modeling a combined diesel electric and gas NPS is applied to design the SVM and optimize its hyperparameter values—insensitivity, penalty parameter, and kernel spread. Our results show that an optimally tuned and trained SVM algorithm can make predictions with error rates below 0.5%. Results also show our SVM algorithm outperforms the SVM algorithm discussed in previous work. In this work, we established a good base for developing a CBM program for the U.S. Navy.