Adaptive case-based reasoning with probabilistic integration for the diagnosis and prognosis of the technical condition of complex systems

Abstract

This study introduces an advanced adaptive Case-Based Reasoning (CBR) framework designed for real-time diagnosis and prognosis of the technical condition of ship power plants, achieved through the seamless integration of Bayesian networks, Markov process modeling, and cognitive simulation within a dynamically adaptive environment. Traditional CBR approaches, while effective at retrieving analogues from historical case archives, often lack the capability to capture complex stochastic dependencies among system components, dynamic degradation patterns under varying operational loads, and real-time contextual variations in sensor data. To address these limitations, the proposed methodology incorporates six integrated phases: data acquisition and normalization, ensuring consistent standardization of heterogeneous sensor readings and operational parameters; probabilistic failure analysis utilizing Bayesian networks to compute conditional failure probabilities and adjust case relevance weights in light of intercomponent dependencies; scenario-driven forecasting based on discrete-time Markov process models to simulate state transition dynamics and predict degradation trajectories; decision adaptation and fusion, which combines classical CBR retrieval outcomes, probabilistic inference results, and forecasted degradation estimates via dynamically normalized weighted coefficients (α, β, γ) that reflect current risk levels; knowledge base maintenance through the incorporation of newly acquired real cases and synthetically generated cases from cognitive simulation, thus enhancing retrieval accuracy and mitigating data scarcity; and automated generation of preventive maintenance recommendations aligned with predicted remaining useful life. Validation experiments conducted on a comprehensive dataset of more than 11 000 historical and synthetic cases demonstrated a diagnostic accuracy of 91 % compared to 79 % achieved by traditional CBR, a 6.7 % reduction in false alarms, a 5–7 % improvement in remaining useful life prediction accuracy, and a 4.7 % decrease in forecast error attributable to the cognitive simulation module, which also improved rare-failure detection rates by 5.1 %. These empirical results confirm the proposed system's high reliability and robustness under fluctuating operational loads and cascading failure scenarios, as well as its seamless integration into onboard monitoring architectures for optimized maintenance scheduling, reduced unplanned downtime, and enhanced operational safety of maritime power plants.