Vessels typically house large sets of different, complex types of equipment; functional failures in them lead to operational stoppage or downgrade with impacts on performance, quality and/or cost. Preventive maintenance schedules are commonly employed, the optimization of which relates to the need of maintenance, the specific component where a problem is detected, the identified fault type, the severity, the expected remaining life within acceptable performance (confidence) limits, etc. Recent advances in sensors and in Machine Learning (ML) methods, have boosted both the fault diagnosis and prognosis, thus incenting companies to invest on the development of efficient Predictive Maintenance (PdM). In this work, we explore the PdM problem for a family of equipment, namely, compressors, through the application of ML techniques on large datasets obtained from on-board sensors. We first deal with the problem of identifying the most useful features in the frequency and time domains, that enable efficient classification and we demonstrate results on data pre-processing and feature extraction. We apply two different clustering and classification algorithms, namely, k-Nearest Neighbor (KNN) Support Vector Machines (SVM) on big datasets obtained from laboratory and industrial setups. We demonstrate that early failure prediction and fault classification is feasible and provides ample opportunities for the development of PdM tactics that reduce cost and minimize risk. Finally, we comment on the appropriateness of features and evaluate the classification accuracy for simple fault cases.

This content is only available via PDF.
You can access this article if you purchase or spend a download.