基于ARN和BiLSTM的轴承剩余寿命预测方法
Bearing Remaining Life Prediction Method Based on ARNs and BiLSTM
徐嘉杰 1沈艳霞1
作者信息
- 1. 江南大学 物联网工程学院,江苏 无锡 214000
- 折叠
摘要
针对深度学习方法进行轴承剩余使用寿命(Remaining Useful Life,RUL)预测时出现的网络退化和噪声信号干扰问题,提出一种基于注意力残差降噪模型(Attention and Residual Network,ARN)和双向长短时记忆网络(Bidirectional Long-Short-Term Memory network,BiLSTM)的轴承剩余使用寿命预测方法.ARN融合了卷积注意力机制(Convolution Block Attention Module,CBAM)和残差网络,利用通道和空间双维度注意力降低噪声特征的权重,结合软阈值函数进行降噪处理,能够同时提取到更多全局和局部的振动特征来构建健康指标(Health Indicator,HI).以健康指标作为输入,通过BiLSTM网络映射得到RUL预测值.在IEEE PHM 2012轴承数据集上进行所提方法与其他健康指标构建模型和RUL预测模型的对比实验,结果表明在6种不同信噪比下(-5、-3、-1、1、3、5 dB),所提方法的抗噪能力最强,预测误差最小.
Abstract
In view of the problem that the accuracy of bearing remaining useful life(RUL)prediction with previous deep learning methods is affected by network degradation and noise interference,a new bearing RUL prediction method based on Attention and Residual Network,(ARN)and Bidirectional Long-Short-Term Memory(BiLSTM)network is proposed.The ARN incorporates the Convolution Block Attention Module(CBAM)attention mechanism and residual networks,uses both channel and spatial dual dimensional attention to reduce the weight of noisy features,and applies soft threshold functions for noise reduction.This method can extract more global and local vibration features simultaneously to construct health indicators(HI).Then,with the HI as input,the prediction value of RUL is obtained through BiLSTM network mapping.The experiments for comparing the proposed method with other health metric construction models and RUL prediction models are carried out on the IEEE PHM 2012 bearing dataset.The results show that for six different signal-to-noise ratios(-5,-3,-1,1,3,5 dB),the proposed method has the best noise immunity and the lowest prediction error.
关键词
故障诊断/剩余使用寿命/轴承/注意力机制/残差网络/双向长短时记忆网络Key words
fault diagnosis/RUL/bearing/attention mechanism/residual network/BiLSTM引用本文复制引用
基金项目
国家重点研发计划(2020YFB1711100)
国家重点研发计划(2020YFB1711102)
出版年
2024
国家科技期刊平台
NSTL
万方数据