Large Models for Machinery Fault Diagnosis: Current Advances and Future Directions

Machinery fault diagnosis (MFD) is critical for the intelligent operation and maintenance of advanced equipment, contributing to enhanced safety, improved efficiency and reduced operating costs. Early machine learning-based methods partially reduced human involvement, whereas deep learning methods enabled end-to-end fault diagnosis directly from raw monitoring data. Despite notable progress, these approaches remain limited in cross-scenario generalization under data scarcity. Recent advances in large models (LMs), including GPT, DeepSeek, Qwen, and other large language and multimodal models, offer significant potential in feature extraction and decision-making, enabling few-shot or zero-shot fault diagnosis, and substantially reducing manual feature engineering and human intervention. However, comprehensive reviews that systematically trace this evolution and outline future research directions remain limited. To address this gap, this article presents a review and roadmap for LM-driven MFD. Future developments are expected along several interrelated directions: Reasoning-driven approaches will exploit emergent inference capabilities to handle complex fault patterns; knowledge-enhanced methods will incorporate domain expertise to improve interpretability and diagnostic reliability; multimodal strategies will integrate vibration, acoustic, and visual sensing for comprehensive monitoring and awareness, and agent-empowered systems will enable autonomous monitoring and adaptive maintenance. Continual learning mechanisms will support adaptation to evolving operating conditions and new fault types, and lightweight-optimized designs will facilitate deployment on edge devices under industrial constraints. This review provides a systematic reference for researchers and practitioners, offering a roadmap to guide the next generation of LM-driven MFD research.