轴承剩余使用寿命预测的IDSA-LSTMNN  

IDSA-LSTMNN for Prediction of Bearing Remaining Useful Life

作　　者：梁天添 刘健 梁贺焱 郑皓谦 王茂[2] LIANG Tiantian;LIU Jian;LIANG Heyan;ZHENG Haoqian;WANG Mao(School of Electrical Engineering,Dalian Jiaotong University Dalian,116028,China;Center for Control Theory and Guidance Technology,Harbin Institute of Technology Harbin,150001,China)

机构地区：[1]大连交通大学电气工程学院,大连116028 [2]哈尔滨工业大学空间控制与惯性技术研究中心,哈尔滨150001

出　　处：《振动.测试与诊断》2025年第2期273-280,410,共9页Journal of Vibration,Measurement & Diagnosis

基　　金：国家自然科学基金资助项目(U24B20159);辽宁省教育厅基本科研资助项目(JYTMS20230037);辽宁省交通科技资助项目(202243)。

摘　　要：针对长短期记忆神经网络提取特征信息相关性和时间信息依赖性不足的问题,提出基于改进双多头注意力机制的长短期记忆神经网络(improved dual stage attention-based long short-term memory neural networks,简称IDSA-LSTMNN),以提高滚动轴承剩余使用寿命(remaining useful life,简称RUL)的预测精度。首先,采用改进的蜘蛛蜂优化器(improved spider wasp optimizer,简称ISWO)优化变分模态分解(variational mode decomposition,简称VMD)的关键参数,以提取更优的时频域特征;其次,结合时域和频域特征,构建特征数据集;然后,利用非线性成分改进核主成分分析(kernel principal component analysis,简称KPCA)的核函数,以优化特征降维过程;最后,利用多头注意力权重改进特征注意力机制和时间注意力机制提出IDSA,设计了轴承RUL预测的IDSA-LSTMNN,并进行了对比试验、抗噪性试验和泛化性试验。结果表明,相较于仅具有单头注意力机制的LSTMNN,不同工况、不同噪声环境下,IDSA-LSTMNN在预测精度、抗噪性和泛化性上具有显著优势。Addressing inadequacies of long short-term memory neural networks in extracting feature information relevance and time information dependency,an improved dual stage attention-based long short-term memory neural network(IDSA-LSTMNN) is proposed to improve the prediction accuracy of rolling bearing remaining useful life(RUL).An improved spider wasp optimizer(ISWO) is utilized to optimize the key parameters of the variational mode decomposition(VMD) to extract superior time-frequency domain features;combining time-domain features with frequency-domain features,a feature dataset is constructed.Nonlinear components are used to improve the kernel function of kernel principal component analysis(KPCA),which is to optimise the feature dimensionality reduction process.IDSA is proposed by refining feature attention mechanism and time attention mechanism using multi-head attention weights.IDSA-LSTMNN is devised for bearing RUL prediction,and comparative experiments,anti-noise experiments and generalization experiments are implemented.The results show that compared with LSTMNN which only has the single-head attention mechanism,IDSA-LSTMNN has prominent advantages in accuracy,anti-noise capability and generalization under different working conditions and different noise environment.

关 键 词：轴承寿命预测 变分模态分解 特征降维 长短期记忆神经网络 改进的注意力机制 

分 类 号：TH133.33[机械工程—机械制造及自动化]