The Outlook for Machine Learning Application in the Problem of Propulsion Engine Monitoring and Diagnostics

抄録

This paper closely scrutinizes the application of machine learning techniques for the intelligent monitoring of ship propulsion engines. The information acquired online from the sensors is subjected to statistical processing with a factor analysis method. The resulting principal factor loadings characterize the contribution of every measured and analyzed parameter to the variance of the principal factor. Owing to the fact that the incipient failures cause deviation of different groups of propulsion engine parameters, the variation of principal factor loadings forms the anomaly pattern unique for every failure. A set of unique patterns is highly suitable for applying clustering and classification algorithms. Factor loadings sets, responsible for every failure, are used to train data-driven anomaly models based on Support Vector Machine (SVM) and Self-Organized Map (SOM). Once trained, these models are able to discriminate newly-acquired data samples as belonging to either a normal state or one of the trained fault patterns. In addition, the factor analysis method is used to derive a performance index recognizing engine state deviation from the normal condition. A combination of the performance index with the classification analysis provides a robust framework for the early detection and identification of incipient engine faults at an early stage.