早期工作阶段滚动轴承剩余寿命预测算法  

Remaining Useful Life Prediction Algorithm for Rolling Bearing in the Early Stage

作　　者：郝金骁 王龑 郭倩宇 张文强 HAO Jinxiao;WANG Yan;GUO Qianyu;ZHANG Wenqiang(Academy for Engineering and Technology,Fudan University,Shanghai 200433,China;School of Computer Science,Fudan University,Shanghai 200433,China;Intelligent Robotics Research Institute,Fudan University,Shanghai 200433,China)

机构地区：[1]复旦大学工程与应用技术研究院,上海200433 [2]复旦大学计算机科学技术学院,上海200433 [3]复旦大学智能机器人研究院,上海200433

出　　处：《计算机工程》2024年第12期48-58,共11页Computer Engineering

基　　金：上海市科委科研攻关计划(22511102202)。

摘　　要：传统寿命预测算法在包含退化阶段数据的滚动轴承寿命预测方面已取得不错的效果,但是由于刚运行和运行一段时间数据相似,因此在只有正常工作阶段数据的情况下难以准确预测。储备池计算(RC)可根据之前时刻数据预测多个时间步长之后的数据,通过数据模拟补充退化数据,提高了将早期预测转化为传统预测的可能性。回声状态网络(ESN)可在充分利用时序信息的基础上输出当前时刻的相关维度。针对早期阶段轴承寿命预测,提出一个基于RC和ESN的递归可重构神经(RRN)网络的算法。首先设计一个基于RC的特征模拟网络,根据早期特征模拟包含退化数据的全寿命周期数据;然后提出一个基于ESN的寿命预测网络,根据输入的模拟特征输出剩余寿命。在PHM 2012数据集上验证了该算法的有效性,实验结果表明,与目前效果较好的算法相比,该算法在原测试数据实验与早期阶段剩余寿命预测的实验平均误差分别降低了61.35%和53.14%,具有较优的预测性能。Traditional useful life prediction algorithms have achieved good results in predicting the useful life of rolling bearings containing degradation stage data.However,as data just running are similar to data running for a period of time,accurate prediction using only normal-working-stage data is difficult.The Reservoir Computer(RC)can predict future data after multiple time steps based on previous time data,raising the possibility of converting predictions in the early stage into traditional predictions by supplementing the degraded data through data simulation.An Echo State Network(ESN)can output the relevant dimensions of the current moment while fully utilizing the temporal information.In this study,a Recursive Reconstructible Neural(RRN)network algorithm based on RC and ESN is proposed for bearing useful life prediction in the early stage.First,an RC-based feature simulation network is designed to simulate the entire lifecycle of data containing degraded data based on early features.Subsequently,a useful life prediction network based on ESN,which outputs the Remaining Useful Life(RUL)based on the simulated features of the input,is proposed.The effectiveness of the algorithm is validated on the PHM 2012 dataset,and the experimental results showed that compared with current algorithms with good performance,the proposed algorithm reduced the average error of the RUL prediction in the original test data and early stage experiments by 61.35%and 53.14%,respectively,demonstrating superior prediction performance.

关 键 词：早期剩余寿命预测 滚动轴承 数据模拟 储备池计算 回声状态网络 递归可重构神经网络 

分 类 号：TP306.3[自动化与计算机技术—计算机系统结构]