Data-driven prognostics using a combination of constrained K-means clustering, fuzzy modeling and LOF-based score

Today, failure modes characterization and early detection is a key issue in complex assets. This is due to the negative impact of corrective operations and the conservative strategies usually put in practice, focused on preventive maintenance. In this paper anomaly detection issue is addressed in new monitoring sensor data by characterizing and modeling operational behaviors. The learning framework is performed on the basis of a machine learning approach that combines constrained K-means clustering for outlier detection and fuzzy modeling of distances to normality. A final score is also calculated over time, considering the membership degree to resulting fuzzy sets and a local outlier factor. Proposed solution is deployed in a CBM+ platform for online monitoring of the assets. In order to show the validity of the approach, experiments have been conducted on real operational faults in an auxiliary marine diesel engine. Experimental results show a fully comprehensive yet accurate prognostics approach, improving detection capabilities and knowledge management. The performance achieved is quite high (precision, sensitivity and specificity above 93% and κ=0.93), even more so given that a very small percentage of real faults are present in data.