Integrated digital twin framework for adaptive diagnostics of complex technical systems

Abstract

This paper presents a comprehensive methodological framework for creating an integrated information environment for the monitoring, diagnostics, and prognostics of the technical condition of complex technical systems (CTS), exemplified by ship power plants. The relevance of the work is driven by the necessity of transitioning from fragmented and static analytical methods to dynamic, real-time lifecycle management systems for equipment. The primary focus is placed on the synthesis of classical approaches, such as probabilistic graphs and fault trees, with modern technologies including digital twins (DT), big data streaming analytics, and the "Model-as-a-Service" (MaaS) paradigm. The scientific novelty of the research lies in the development of an adaptive architecture where the traditional static fault tree is transformed into a dynamic ontological structure. A mathematical risk calculation model based on the Min Cut Upper Bound approximation is introduced, ensuring computational efficiency during the processing of high-intensity telemetry streams via Kafka and Flink. The paper describes an original Error Classifier architecture that functions as a semantic validator. The implementation of a "veto" mechanism, based on the physical constraints of the digital twin, reduced the diagnostic model's false-positive rate by 20%, ensuring the priority of physical consistency over statistical correlations. The practical significance of the research is confirmed by simulation results of various engine component degradation scenarios. Experimental data demonstrate that integrating prognostic models into the digital twin loop provides a systemic advantage: classification accuracy increases from 0.87 to 0.94, and the forecast update delay is reduced by more than half, from 5.0 to 2.1 seconds. A methodology for model lifecycle management is proposed through a closed-loop MLOps cycle, including physics-informed training, shadow deployment, and automated characteristic drift detection. The work formalizes a residual useful life (RUL) calculation algorithm based on a dynamic probability gradient, allowing the system to adapt to changes in operational intensity in real time. The resulting multilevel information environment architecture—comprising Data, Preprocessing, Diagnostics, Prognostics, and Decision Support layers—represents a complete methodological solution for proactive maintenance and the transition toward autonomous CTS monitoring systems.