Adaptive similarity assessment metric for intelligent failure diagnostics in ship power plants

Prompt and accurate diagnosis of failures in ship power plants (SPPs) is essential for ensuring maritime safety, minimising operational risks, and optimising maintenance strategies. With increasing system complexity, heterogeneous data sources, and limited historical failure records, conventional diagnostic methods often prove insufficient, especially in scenarios involving rare or ambiguous faults. The purpose of this study was to develop an interpretable, adaptive, and probabilistically grounded methodology for assessing similarity between failure cases within SPPs for use in intelligent decision support systems. The proposed method integrates Euclidean, Jaccard, and logistic similarity metrics with Bayesian inference, temporal degradation modelling, frequency-based weight correction, and contextual smoothing of affected subsystems. The model employs L-BFGS-B optimisation to automatically adjust metric weights according to diagnostic relevance. Numerical experiments based on synthetic case data revealed high classification accuracy: 96% for failures related to cooling system overheating, 84% for bearing degradation cases, and 92% for fuel supply irregularities. Even with a 40% reduction in training data volume, the performance drop did not exceed 7%, indicating strong resilience to data sparsity. The visualisation of decision boundaries demonstrated the model’s ability to distinguish overlapping failure classes while preserving semantic interpretability. Weight optimisation results identified “failure type” as the dominant factor, while “risk category” and “affected subsystems” had negligible impact and were excluded. Bayesian aggregation further improved the credibility of diagnostic conclusions by combining local similarity with global statistical priors. The developed methodology can be effectively applied by marine engineers, ship operators, and developers of intelligent diagnostic platforms for fault detection, root cause analysis, and predictive maintenance under conditions of uncertainty and incomplete information. Its modular structure also allows extending it to other complex technical domains beyond SPPs