Objective In response to the pressing need for intelligent transformation in equipment support for China Coast Guard vessels, this paper presents a systematic review of research progress in ship fault detection and diagnosis (FDD) systems based on large language models (LLMs). Shipboard environments pose unique challenges including weak communication links at sea, limited onboard computing resources, high levels of mechanical noise, and stringent reliability requirements for safety-critical systems. Traditional data-driven FDD methods, while leveraging accumulated industrial IoT data, suffer from a scarcity of high-quality labeled data and poor cross-scenario generalization. LLMs offer new opportunities with their capabilities in multi-modal understanding, knowledge fusion, and zero-shot or few-shot reasoning.

Method A systematic literature review method was adopted to analyze 95 relevant publications from 2022 to 2026. The search was conducted across Web of Science, IEEE Xplore, ScienceDirect, and CNKI databases using keyword combinations including "fault diagnosis", "fault detection", "large language model", "LLM", "GPT", "PHM", "intelligent operation and maintenance", and "agent". After initial screening of titles and abstracts and backward snowballing, 95 papers were finally included for analysis. Optimization strategies of LLMs in industrial FDD were summarized into four dimensions: input-side enhancement (including RAG, knowledge graphs, multi-modal LLMs, and digital twins), model-side optimization (including fine-tuning or retraining, prompt engineering, and multi-model collaboration), output-side refinement (feedback optimization), and architectural design (integrating LLMs with signal processing, machine learning, neural networks, and multi-agent systems). In response to the characteristics of shipboard data heterogeneity, limited resources, and high reliability requirements, a hierarchical, decoupled, multi-agent collaborative reference architecture for shipboard FDD systems is proposed. The architecture consists of five layers from top to bottom: the interaction layer (supporting multi-modal input and adaptive output), the Agent layer (comprising eight specialized agents for reception, scheduling, query, process, reasoning, generation, data processing, and feedback), the model layer (including a domain-specific fine-tuned LLM, an embedding model, and a lightweight multi-modal model), the tool layer (featuring time-series signal processing with dual pipelines for feature extraction and semantic mapping, along with text processing, vector embedding, multi-modal alignment, and knowledge graph tools), and the data layer (containing a raw database, a vector knowledge base, and a knowledge graph).

Results The number of research publications on LLMs in the industrial FDD field increased from 4 in 2023 to 22 in 2024, further surging to 61 in 2025, with approximately 76% focused on input-side enhancement and model-side optimization. This trend indicates rapidly growing research interest and the significant potential of LLMs in this domain. For shipboard scenarios, the proposed architecture supports localized deployment of lightweight LLMs through quantization and pruning techniques, enabling multi-source data fusion via vector knowledge bases and knowledge graphs. Based on multi-agent collaboration, it realizes automatic task decomposition for fault diagnosis and closed-loop decision-making. The architecture also incorporates mechanisms for conflict resolution with uncertainty calibration (including historical accuracy-based confidence correction, external hard constraint anchoring, and conservative priority strategies), multi-agent overhead optimization (on-demand wake-up, hierarchical context management, agent role merging, and asynchronous task queuing), and adaptation to unknown faults (zero-shot reasoning with LLM prior knowledge, cross-device fault pattern transfer, active learning with human labeling, and synthetic data enhancement).

Conclusion The proposed hierarchical reference architecture provides a feasible path and framework solution for transforming ship FDD systems toward an intelligent mode driven by the synergy of data and knowledge. Future research should focus on trustworthy output, adaptive capabilities, and system integration to enhance equipment support efficiency and navigation safety. Key future directions include robust intelligence under strong constraints (multi-modal signal characterization under high noise, and LLM decision trustworthiness assessment meeting IEC 61508 functional safety requirements), continuous self-adaptation in offline environments (ontological incremental learning and knowledge distillation during communication blackouts, and few-shot or zero-shot cross-device fault transfer diagnosis), and system integration with standardization (standardized interfaces with existing ship cyber-physical systems, and verification and validation methods compliant with China Classification Society specifications).