Prediction of marine engine wear using topological data analysis of lubricating oil

Predicting marine engine wear based on the analysis of used lubricating oil samples is an important task in predictive maintenance. However, conventional forecasting methods are often ineffective in this context because of irregular sampling intervals, the limited number of observations available for establishing reliable relationships, and the apparent “improvement effect” in oil condition caused by oil replenishment or changes in engine operating conditions. This study proposes a methodological approach based on topological data analysis (TDA). Its distinctive feature is the identification of hidden relationships under conditions of critically incomplete data and a “blind start,” when no sufficiently representative historical dataset is available. The proposed integrated algorithm includes preliminary parameter normalization and feature-space dimensionality reduction using principal component analysis. The DBSCAN algorithm is applied to identify dense regions corresponding to stable operating states and to detect noise points. The resulting multidimensional space of diagnostic indicators is then transformed by the Mapper algorithm into a topological graph representing the chronological degradation trajectory. Mapper identifies persistent groups of observations interpreted as graph nodes. These nodes correspond to characteristic system states, while the graph edges represent possible transitions between them. Geometric graph metrics are used to calculate the cumulative wear path, the system condition index, and the remaining useful life (RUL). The introduction of an additional calculated parameter, degradation intensity, makes it possible to estimate the topological rate of wear per unit of operating time. The proposed method makes it possible to develop an interpretable model of the engine state space. Such a representation facilitates accurate assessment of the current condition and prediction of wear progression. Modeling results obtained from real-world data demonstrated that the approach is capable of identifying normal operating states and anomalous transitions. The resulting information provides a reliable basis for predictive wear assessment and the optimization of maintenance processes.