面向机械设备通用健康管理的智能运维大模型  

Research on Large Model for General Prognostics and Health Management of Machinery

作　　者：雷亚国[1] 李熹伟 李响 李乃鹏[1] 杨彬[1] LEI Yaguo;LI Xiwei;LI Xiang;LI Naipeng;YANG Bin(Key Laboratory of Education Ministry for Modern Design and Rotor-Bearing System,Xi’an Jiaotong University,Xi’an 710049)

机构地区：[1]西安交通大学现代设计及转子轴承系统教育部实验室,西安710049

出　　处：《机械工程学报》2025年第6期1-13,共13页Journal of Mechanical Engineering

基　　金：国家重点研发计划(2022YFB3402100);国家杰出青年科学基金(52025056);国家自然科学基金重点(52435003);新基石科学基金会所设立的科学探索奖(XPLORER-2024-1036);国家留学基金委资助项目。

摘　　要：近年来,基于深度学习的各类机械设备健康管理模型取得了显著进展。然而,现有模型参数规模小,通常只能接受特定采频、转速、模态的数据,针对齿轮、轴承等特定零部件,执行监测、诊断、预测等特定任务,且难以适应新场景,缺乏持续进化能力。随着高端设备精密性、复杂度的不断提升,工业场景对高通用、易扩展、可进化的“一站式”健康管理服务需求日益迫切。受近年来Chat GPT等语言大模型在数据、任务、场景等方面通用化发展趋势启发,提出了面向机械设备通用健康管理的智能运维大模型。首先,将多模式数据通过角度域重采样和数据分割统一编码为词元序列;然后,输入基于Transformer的基底模型,提取健康信息和退化信息至特定词元;最后,将这些特定词元用于执行下游的监测、诊断、预测等多种任务。在故障数据和长期退化数据上对提出模型的基准性能、多任务协同性能和扩展性能进行了验证,结果表明:提出的智能运维大模型能够在轴承、齿轮等多种对象上联动实现状态监测、故障诊断和寿命预测;诊断与预测多任务能够有效协同,互相促进性能提升,相较于单任务模型表现更为出色;在小样本学习、持续学习等场景下,模型能够实现快速适配部署并持续进化。因此,提出的智能运维大模型具有高通用性、易扩展性、可持续进化等特点,有望为机械设备提供通用化“一站式”健康管理服务。In recent years,various deep learning-based health management models for mechanical equipment have made significant progress.However,existing models tend to be smaller in scale and are typically designed to handle data from specific frequencies,speeds,or modes,focusing on particular components such as gears and bearings to perform tasks like monitoring,diagnosis,and prediction.These models struggle to adapt to new scenarios and lack the capability for continuous evolution.With the increasing precision and complexity of high-end equipment,there is a growing demand for highly general,scalable,and evolvable"one-stop"health management services.Inspired by the trend of generalization in large language models like ChatGPT,which excel in handling diverse data,tasks,and scenarios,a large model for general prognostics and health management of machinery is proposed.First,multimodal data is resampled in the angular domain and segmented to token sequence.Then,the data is input into a Transformer-based information integration foundational model to extract health and degradation information into specific tokens.Finally,these specific tokens are used to perform downstream tasks such as monitoring,diagnosis,and prediction.The proposed large model's baseline performance,multitask synergy,and scalability were verified using fault and long-term degradation datasets.The results show that the proposed large model can simultaneously perform condition monitoring,fault diagnosis,and remaining useful life prediction for multiple objects like bearings and gears.Additionally,the diagnostic and predictive multitasks can effectively collaborate,mutually enhancing performance,and achieving better results compared to single-task models.In few-shot learning and continual learning scenarios,the large model can be rapidly deployed and continuously evolved.Therefore,the proposed large model features high generality,scalability,and sustainability,and is expected to provide universal"one-stop"health management services for mechanical equipment.

关 键 词：机械设备 故障诊断 健康管理 智能运维 大模型 

分 类 号：TH17[机械工程—机械制造及自动化]