摘要
针对单一传感器数据在复杂工程系统的故障诊断中难以全面获取机械设备的状态信息、且易受工况变化干扰的问题,提出了一种基于卷积神经网络(convolutional neural network,简称CNN)的多模态数据融合轴承故障诊断方法。首先,构建小波滤波器组,对振动信号和声发射信号自适应选择时频变换尺度区间,进行连续小波变换生成时频图样本;其次,搭建深度学习诊断模型,设计多尺度模块、密集耦合模块以及融合与决策模块,提取不同模态数据的故障特征,并引入相似性约束学习联合特征信息;最后,利用Softmax分类函数实现滚动轴承不同故障位置和程度的精确分类。实验室多模态数据集的验证结果表明,当测试集中加入未训练过的转速数据时,所提出的多尺度特征密集耦合卷积神经网络(multi-scale feature dense coupled convolutional neural network,简称MFDCCNN)的平均识别准确率达到99.21%,且在分类准确性、诊断稳定性和泛化能力这3个方面均优于经典深度学习模型、消融实验模型和单一源数据诊断方法。
Single sensor data in complex engineering systems are limited in their ability to capture equipment health information and are highly susceptible to operating conditions.To address these limitations,a convolutional neural network(CNN)based multi-modal data fusion method for rolling bearing fault diagnosis is proposed in this study.First,a wavelet filter bank is constructed to adaptively select time frequency representations scale intervals for vibration and acoustic emission signals,performing continuous wavelet transforms to generate time frequency pattern samples.Then,a deep learning diagnostic model is built,incorporating multi-scale module and dense coupling module to extract fault features from each modality,while a similarity constraint is employed to enhance joint feature learning.Finally,a softmax classifier achieves precise classification of fault location and severity.Experimental results on a laboratory multi-modal dataset demonstrate that,when testing on previously unseen rotational speeds,the proposed multi-scale feature dense coupled CNN achieves 99.21%average accuracy.It outperforms classical deep learning models,ablation variants,and single modality methods in the aspects of classification accuracy,diagnostic stability,and generalization ability.
作者
寇梓良
张西宁
李兵
成志辉
张雨洁
KOU Ziliang;ZHANG Xi'ning;LI Bing;CHENG Zhihui;ZHANG Yujie(State Key Laboratory for Manufacturing Systems Engineering,Xi'an Jiaotong University Xi'an,710049,China;School of Mechano-Electronic Engineering,Xidian University Xi'an,710068,China)
出处
《振动.测试与诊断》
北大核心
2025年第6期1082-1089,1269,共9页
Journal of Vibration,Measurement & Diagnosis
基金
国家自然科学基金资助项目(51275379)。
关键词
数据融合
故障诊断
卷积神经网络
多模态
滚动轴承
data fusion
fault diagnosis
convolutional neural network
multi-modal
rolling bearing
