High-Dimensional Clustering-Driven Performance Evaluation and Mutation-Centric Early Warning for Marine Diesel Engines

Abstract

To proactively identify performance anomaly evolution and potential operational risks of marine main engines and reserve a sufficient time window for maintenance intervention, this paper proposes a data-driven multi-algorithm fusion framework for performance evaluation and anomaly early warning of marine main engines. The framework first adopts a steady-state detection strategy to filter valid operating conditions and introduces the CLIQUE clustering algorithm to realize adaptive partitioning of high-dimensional operating parameters; comparative experiments with the classical K-means++ clustering algorithm demonstrate that CLIQUE achieves finer-grained operating condition classification without pre-defining the number of clusters and better adapts to the uneven distribution of actual marine engine operating conditions, which addresses the limitations of traditional single-parameter analysis and conventional dimensionality reduction methods in practical shipboard scenarios. On this basis, the Mahalanobis distance evaluation model is constructed under each partitioned operating condition, which further improves the stability and anti-interference performance of quantitative performance assessment for the main engine. Meanwhile, by integrating cumulative anomaly trend analysis and the Yamamoto mutation test, the framework accurately captures statistical mutation characteristics of the performance deviation trajectory and identifies the first mutation point as the retrospective candidate change point, forming a systematic anomaly detection mechanism. Validation using field measurement data from a 6RT-flex82T marine main engine shows that the proposed framework can comprehensively characterize the overall operating state of the main engine and capture long-term performance deviation evolution patterns. Retrospective analysis indicates that the first statistical mutation of the multivariate performance deviation precedes the significant abnormal fluctuation of a single parameter by approximately 20 days, showing the potential of providing a maintenance buffer period. The proposed method can provide technical reference and support for condition-based maintenance of marine main engines and intelligent operation and maintenance of shipboard power equipment.