Agentic AI for autonomous condition monitoring and predictive maintenance of marine vessels

Modern marine vessels generate large volumes of high-frequency operational data, yet translating this data into reliable maintenance intelligence remains challenging. Existing solutions rely on threshold alarms, rule-based logic, or isolated machine learning models, which lack contextual reasoning, adaptability, and interpretability in safety-critical environments. This paper presents an Agentic AI architecture for condition monitoring and predictive maintenance of marine vessels. The proposed system integrates a Transformer-based Autoencoder for multivariate anomaly detection with a Large Language Model (LLM) that enables interactive, explainable analysis through structured tool-calling. The autoencoder learns normal operational patterns from high-dimensional time-series data and detects deviations associated with abnormal loading, inefficient generator utilisation, and potential equipment degradation. Rather than ingesting raw numerical streams, the LLM invokes domain-specific tools to retrieve structured measurements, compute derived metrics, and query anomaly outputs before generating responses. This mechanism mitigates hallucination risks and avoids the brittleness of fine-tuned models as system behaviour evolves. The approach is validated on 91 days of data from an offshore construction vessel with a hybrid diesel-electric propulsion system and six gensets under dynamic positioning constraints. Four fault scenarios (slow drift, load imbalance, temporary reduction, spikes) were all correctly detected, with prognostic lead times between 97 and 637 min.