基于双路神经网络多尺度特征提取的轴承故障诊断  

Fault Diagnosis of Bearings Based on Dual-Path Neural Network with Multi-Scale Feature Extraction

作　　者：宋蒙恩 罗敏[1] SONG Meng′en;LUO Min(School of Robotics and Automation,Hubei University of Automotive Technology,Shiyan 442002,China)

机构地区：[1]湖北汽车工业学院机器人与自动化学院,十堰442002

出　　处：《组合机床与自动化加工技术》2025年第2期163-170,共8页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金项目(51575211)。

摘　　要：针对现有滚动轴承故障诊断方法对振动数据中蕴含信息挖掘不够充分的问题,提出了一种基于双路卷积神经网络(CNN)多尺度特征提取的滚动轴承故障诊断方法。首先,利用连续小波变化(CWT)将一维振动信号转换为二维时频图,获取对轴承数据的不同视角表达,分别作为双路CNN的输入;然后,分别构建双路CNN的多尺度特征提取模块,1D-CNN的特征提取模块由大卷积层和并行卷积层组成,并在每个并行层后加上GRU提取不同尺度的时序特征;2D-CNN的特征提取模块使用多分支连续卷积结构从输入中提取不同尺寸和抽象层次的特征,并引入CBAM注意力机制来增强模型对重要特征的关注;最后,对双路CNN提取到的特征进行多尺度特征融合,利用融合特征训练分类模块,实现轴承的故障诊断。实验结果表明,所提模型10次测试的平均准确率为99.95%,在每类故障训练集仅含有24个样本时平均准确率依然可达95%左右,对比其它诊断模型,所提方法在小样本条件下具有更高的诊断准确率、更强的特征提取能力。To enhance information extraction from vibration data in existing rolling bearing fault diagnosis methods,we propose a dual-path CNN-based fault diagnosis approach with multi-scale feature extraction.Initially,employing continuous wavelet transform,one-dimensional vibration signals are transformed into two-dimensional time-frequency spectrograms,serving as distinct inputs for the dual-path CNN.Then,construct the multi-scale feature extraction module of the dual-path CNN.The feature extraction module of 1D-CNN consists of a large convolution layer and parallel convolution layers,with a GRU added after each parallel layer to extract temporal features of different scales.The 2D-CNN module uses a multi-branch continuous convolution structure for extracting features of different sizes and abstraction levels,introducing the CBAM attention mechanism to enhance focus on crucial features.Finally,features extracted by the dual-path CNN undergo multi-scale fusion,training the classification module for bearing fault diagnosis.Experimental results show an average accuracy of 99.95% over 10 tests.With only 24 samples in each fault training set,the accuracy remains around 95%.Compared to other diagnostic models,our method demonstrates higher accuracy and stronger feature extraction under small sample conditions.

关 键 词：卷积神经网络 连续小波变换 GRU 注意力机制 

分 类 号：TH133.3[机械工程—机械制造及自动化] TG66[金属学及工艺—金属切削加工及机床]