Abstract: ObjectivesAgainst the background of shipping digital transformation, "dual carbon" target implementation and the 15th Five-Year Plan advancement, adaptive technical schemes, core engineering implementation contradictions and development directions of Large Language Model (LLM)-centered artificial intelligence large model systems for marine electromechanical equipment intelligent operation and maintenance (O&M) are systematically investigated. This work targets industry pain points of fragmented research, insufficient multi-factor integration, and inherent limitations of traditional experience-driven O&M modes, single data-driven modeling and deep learning small models, to provide theoretical and practical support for intelligent O&M paradigm innovation and green digital transformation of the shipping industry.MethodsA full-chain analytical framework of "limitation analysis - status review - contradiction identification - path prospect" is established. The development status and technical bottlenecks of intelligent O&M technology in two core fields (state prediction & energy efficiency optimization, fault diagnosis & health management) of marine electromechanical equipment are systematically reviewed, a horizontal comparison of four mainstream technologies (mechanism model, shallow machine learning, deep learning small model, LLM-driven method) is completed from six core dimensions, and the technical schemes, scenario adaptation difficulties and core implementation contradictions of LLM and its supporting technologies are deeply analyzed to clarify future development directions.ResultsTechnical bottlenecks of the two core fields and corresponding LLM-based solutions are clarified. Four essential contradictions restricting engineering implementation, namely data-mechanism mismatch, black box-interpretability conflict, computing power-resource constraint, and generation-security risk, are condensed, and the core gap from technical feasibility to engineering applicability is revealed. Three implementable technical paths adapted to marine scenarios, including few-shot learning, cross-ship domain knowledge aggregation, and zero-shot fault identification, are sorted out. Finally, a marine scenario-specific three-in-one evolution paradigm of "mechanism fusion - edge deployment - security construction" is proposed, with its inherent driving logic of mutual premise, constraint and collaborative closed loop defined.ConclusionsThe new-generation LLM-centered artificial intelligence large model system can effectively break through the inherent bottlenecks of traditional intelligent O&M for marine electromechanical equipment, and its engineering implementation will promote the profound transformation of ship O&M from post-fault maintenance to pre-fault prediction, and from local optimization to global intelligence. The proposed three-in-one evolution paradigm provides a dedicated technical route for compliant engineering implementation of LLMs in marine scenarios, and has important theoretical and practical significance for improving O&M reliability and operation efficiency of marine electromechanical equipment, as well as promoting high-quality transformation of the shipping industry.