• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 169-185.doi: 10.3901/JME.260320

• 特邀专栏:大模型驱动的高端装备智能运维 • 上一篇    

扫码分享

半监督学习视角下序列化变分自编码器驱动的高端装备智能剩余寿命预测

张九思1, 刘目1, 陈凯1, 邱根1, 程玉华1, 尹珅2   

  1. 1. 电子科技大学自动化工程学院 成都 611731;
    2. 挪威科技大学机械与工业工程系 特隆赫姆 7034 挪威
  • 收稿日期:2025-10-25 修回日期:2026-07-03 发布日期:2026-11-20
  • 作者简介:张九思,男,1997年出生,博士,副教授,硕士研究生导师。主要研究方向为高端装备的智能运维、数据驱动的监测/优化、工业大数据与人工智能。E-mail:jiusi.zhang@uestc.edu.cn;尹珅(通信作者),男,1981年出生,博士,教授,博士研究生导师。主要研究方向为故障诊断/预测和容错、数据驱动的监控和优化、系统和控制理论,健康诊断和信息物理系统的应用。E-mail:shen.yin@ntnu.no
  • 基金资助:
    国家自然科学基金(62503086,U2230206)、国家资助博士后研究人员计划、中国博士后科学基金(BX20250397,2025M771733)和中央高校基本科研业务费(ZYGX2025XJ020)资助项目。

Intelligent Remaining Life Prediction of Premium Equipment Driven by Sequential Variational Autoencoder from the Perspective of Semi-supervised Learning

ZHANG Jiusi1, LIU Mu1, CHEN Kai1, QIU Gen1, CHENG Yuhua1, YIN Shen2   

  1. 1. School of Automation Engineering, University of Electronic Science and Technology of China, Chengdu 611731;
    2. Department of Mechanical and Industrial Engineering, Norwegian University of Science and Technology, Trondheim 7034, Norwegian
  • Received:2025-10-25 Revised:2026-07-03 Published:2026-11-20

摘要: 剩余寿命预测是保障高端装备安全性与可靠性的核心内容。目前,一些传统的数据驱动方法需要海量的有标签数据进行监督式的模型训练,难以充分利用未标记的退化数据。传统的深度神经网络在可解释性和不确定性方面需要深入的研究工作。因此,从半监督学习的视角出发,提出了一种基于Transformer-序列化变分自编码器网络的剩余寿命预测方法。具体而言,考虑一种动态的序列化建模方式,设计基于Transformer的特征提取网络,聚焦输入时间窗口数据中的关键部位。在此基础上,探索一种贝叶斯深度概率生成网络和控制理论中状态空间模型的对应性关系,模拟剩余寿命预测值的概率密度分布,衡量退化过程中剩余寿命的不确定性,并尝试从隐含空间的角度提供可解释性。通过半监督视角下基于Transformer-序列化变分自编码器网络,能够从大量无标签的退化数据中学习具备潜在价值的特征表示,并依托NASA预测中心提供的飞机涡扇发动机数据集验证了方法的有效性。

关键词: 半监督学习, 序列化变分自编码器, 高端装备, 剩余寿命, 预测

Abstract: Remaining life prediction is the core content to ensure the safety and reliability of premium equipment. At present, some conventional data-driven approaches demand massive labeled data for supervised model training, which is difficult to make full use of unlabeled degradation data. Traditional deep neural networks need in-depth research in interpretability and uncertainty. Consequently, from the perspective of semi-supervised learning, a remaining life prediction approach based on Transformer-serialized variational autoencoder network is proposed. Specifically, a dynamic serialization modeling method is considered, and a feature extraction network based on Transformer is designed to focus on the key parts of the input time window data. On this basis, the correspondence between a Bayesian deep probability generation network and the state space model in control theory is explored. The probability density distribution of the remaining life prediction value is simulated. The uncertainty of the remaining life in the degradation process is measured, and an attempt is made to provide interpretability from the perspective of implicit space. Through the semi-supervised perspective based on Transformer-serialized variational autoencoder network, it is possible to learn feature representations with potential value from a large amount of unlabeled degradation data. The effectiveness of the approach is verified by relying on the aircraft turbofan engine dataset provided by NASA Prediction Center.

Key words: semi-supervised learning, sequential variational autoencoder, premium equipment, remaining life, prediction

中图分类号: