摘要
在机械设备典型转动部件智能诊断与剩余寿命预测中,针对多变量时序输入的神经网络泛化能力差、特征域不匹配的问题,提出可解释小波网络的多变量混合监督剩余寿命预测方法。首先,构建小波卷积模块,将Morlet小波初始化一维卷积权重作为预训练方法,得到具有可解释性的自适应小波时频特征;其次,提出多变量混合监督策略,对多变量数据应用通道独立参数共享的特征提取方法,通过自监督学习分支约束小波卷积权重,学习退化与剩余寿命的定量特征;最后,通过全寿命实验数据集进行对比实验与消融实验。实验结果表明,相比典型模型,可解释小波网络的多变量混合监督剩余寿命预测方法对多变量数据有较好的剩余寿命预测效果。
In intelligent diagnosis and remaining life prediction of typical rotating components in mechanical equipment,there exist the problems of poor generalization ability and feature domain mismatching in multivariable-timeseries-input neural network.To address these problems,a interpretable wavelet multivariate hybrid-supervised network method for remaining useful life prediction was proposed.Firstly,a wavelet convolution module initialized by Morlet wavelet was put forward.As a pre-training method,it can extract interpretable and adaptive time-frequency features.Secondly,a multi-variable hybrid-supervising strategy was proposed,and the feature extraction method based on channelwise independence and parameter sharing was applied to the multi-variable data.The self-supervising branch was used to constrain the weights in wavelet convolution and learn the quantitative features of degradation and remaining useful life.Finally,comparative experiments and ablation studies were conducted using a full lifespan dataset.Experimental results demonstrate that the proposed method outperforms typical models in predicting the remaining useful life of multi-variable data.
作者
王一鸣
周军
石生超
王正伟
李富才
徐肖磊
WANG Yiming;ZHOU Jun;SHI Shengchao;WANG Zhengwei;LI Fucai;XU Xiaolei(State Key Laboratory of Mechanical System and Vibration,Shanghai Jiaotong University,Shanghai 200240,China;State Grid Qinghai Electric Power Company,Xining 810000,Qinghai,China)
出处
《噪声与振动控制》
北大核心
2025年第6期196-202,共7页
Noise and Vibration Control
基金
国网青海省电力公司科技资助项目(522807230005)。
关键词
故障诊断
剩余寿命预测
多变量
MORLET小波
小波重构
混合监督
fault diagnosis
remaining useful life prediction
multi-variable
Morlet wavelet
wavelet reconstruction
hybrid-supervising
