• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2021, Vol. 57 ›› Issue (21): 88-95.doi: 10.3901/JME.2021.21.088

• 机械动力学 • 上一篇    下一篇

扫码分享

一种基于空间卷积长短时记忆神经网络的轴承剩余寿命预测方法

王久健1,2, 杨绍普2, 刘永强2, 文桂林1   

  1. 1. 湖南大学汽车车身先进设计制造国家重点实验室 长沙 410082;
    2. 石家庄铁道大学省部共建交通工程结构力学行为与系统安全国家重点实验室 石家庄 050043
  • 收稿日期:2020-11-18 修回日期:2021-06-18 出版日期:2021-12-28 发布日期:2021-12-28
  • 通讯作者: 杨绍普(通信作者),男,1962年出生,博士,教授,博士研究生导师。主要研究方向车辆系统动力学与控制、非线性系统动力学理论与应用、振动与噪声控制、交通环境与安全工程。E-mail:yangsp@stdu.edu.cn
  • 作者简介:王久健,男,1990年出生,博士研究生。主要研究方向为机械设备状态监测和剩余寿命预测。E-mail:wangjiujian@hnu.edu.cn
  • 基金资助:
    国家自然科学基金(11790282,12032017,11802184,11902205,12002221)、河北省科技计划(20310803D)和河北省自然科学基金(A2020210028)资助项目。

A Method of Bearing Remaining Useful Life Estimation Based on Convolutional Long Short-term Memory Neural Network

WANG Jiujian1,2, YANG Shaopu2, LIU Yongqiang2, WEN Guilin1   

  1. 1. State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan University, Changsha 410082;
    2. State Key Laboratory of Mechanical Behavior in Traffic Engineering Structure and System Safety, Shijiazhuang Tiedao University, Shijiazhuang 050043
  • Received:2020-11-18 Revised:2021-06-18 Online:2021-12-28 Published:2021-12-28

摘要: 传统的基于数据驱动的轴承剩余预测方法仍需要一定的先验知识,比如:特征指标选取、健康指标构建、失效阈值选定等等。预测结果严重依赖人工经验,为了克服这一缺点,基于深度学习方法提出了一种用于轴承剩余寿命预测的新方法,该方法的核心包括健康指标构建和剩余寿命计算。首先提出了一种无需先验知识的基于空间卷积长短时记忆神经网络(Convolutional long short-term memory neural network,ConvLSTM)的健康指标生成网络模型,该网络利用卷积神经网络的局部特征提取能力和长短时记忆网络的时间依赖特性,可直接从采集到的原始信号中挖掘反映退化程度的特征,构建健康指标,实现了高维原始数据向低维特征的映射转化,并利用Sigmoid函数将其归至[0,1]区间内,实现了阈值的统一;然后,利用粒子滤波更新双指数寿命模型,实现剩余寿命结果的输出。利用轴承全寿命试验对所提方法进行了验证,并与其他相关方法进行对比,结果表明本文方法所构建的健康指标具有更好的趋势性、单调性和鲁棒性,同时剩余寿命预测的准确率更高。

关键词: 滚动轴承, 剩余寿命预测, 健康指标, 深度学习, ConvLSTM

Abstract: Traditional prediction methods of bearing remaining useful life estimation based on data driven method still need some prior knowledge, such as feature index selection, health index construction, failure threshold selection and so on. In order to overcome this shortcoming, a new method for bearing residual life prediction based on deep learning method is proposed. The core of this method includes health index construction and remaining useful life calculation. First proposed a without prior knowledge of health indicators generated neural network based on ConvLSTM, the network combines the local feature extraction ability of convolutional neural network and long length of time dependent characteristics of long short-term memory neural network. This network can directly mining the characteristics of degradation degree from the original signal and build health indicators to realize the original high-dimensional data to lower dimensional feature mapping. And it uses the Sigmoid function to unify the threshold to[0, 1] interval, realized the unification of the threshold value; Then, the particle filter is used to update the double exponential life model to realize the output of the remaining life results. The method is verified by bearing life test and compared with other related methods. The results show that the health index constructed by this method has better trend, monotonicity and robustness, and the accuracy of remaining useful life prediction is higher.

Key words: rolling bearing, RUL, health indicator, deep learning, ConvLSTM

中图分类号: