摘要
针对滚动轴承故障诊断中单一域特征表征不充分与模型抗干扰能力不足的问题,提出一种融合时频双分支图形推理(Fusion of Time-Frequency Dual-Branch Graphical Reasoning Network,TDG-Net)故障诊断网络。该网络对原始振动信号进行双域数据预处理:通过视图变换调整直接获取时域输入数据,同时采用快速傅里叶变换(Fast Fourier Transform,FFT)将原始信号转换为频域数据,构建时频双域输入样本集。设计双分支特征提取结构,将时域与频域数据分别输入嵌入相似感知注意力模块的ResNet34网络,实现时频域关键特征的深度挖掘与自适应强化。引入图卷积网络(Graph Convolutional Network,GCN)构建跨域特征推理机制:以双分支输出的特征向量作为图节点,通过余弦相似度计算构建邻接矩阵量化跨域特征关联强度,完成跨域特征的信息互补与全局优化。将优化后的特征向量输入由全局平均化层与softmax函数组成的分类模块,实现滚动轴承故障类型的精准识别。在公开的凯斯西储大学(Case Western Reserve University,CWRU)轴承数据集上进行实验,其准确率、精确度、灵敏度和F1分数分别为99.55%、99.62%、99.55%和99.57%。在变噪声和变负荷条件下,准确率和F1得分分别为85.27%和85.54%。实验结果表明,所提方法可有效捕获滚动轴承数据的全面特征,显著提升模型抗干扰能力,为滚动轴承故障诊断提供高效解决方案。
To address the problems of insufficient characterization of single-domain features and inadequate anti-interference capability of models in rolling bearing fault diagnosis,a fault diagnosis network named Fusion of Time-Frequency Dual-Branch Graphical Reasoning Network(TDG-Net)is proposed.The network operates through four key steps.First,it performs dual-domain data preprocessing on the original vibration signals by adjusting the time-domain input data through view transformation for direct acquisition,while converting the original signals into frequency-domain data using Fast Fourier Transform(FFT),thus constructing a time-frequency dual-domain input sample set.Second,a dual-branch feature extraction structure is designed.The time-domain and frequency-domain data are respectively input into a ResNet34 network embedded with a similarity-aware attention module,enabling in-depth mining and adaptive enhancement of key time-frequency domain features.Third,a Graph Convolutional Network(GCN)is introduced to build a cross-domain feature reasoning mechanism,the feature vectors output from the dual branches are used as graph nodes,and an adjacency matrix is constructed via cosine similarity calculation to quantify the correlation strength of cross-domain features,thereby achieving information complementation and global optimization.Fourth,the optimized feature vectors are input into a classification module composed of a global average pooling layer and a Softmax function,achieving accurate identification of rolling bearing fault types.The proposed method is validated using the public Case Western Reserve University(CWRU)bearing dataset.The results show that the accuracy,precision,sensitivity,and F1-score of the proposed method reach 99.55%,99.62%,99.55%,and 99.57%respectively.Under variable noise and load,the accuracy and F1-score are 85.27%and 85.54%respectively.The experimental results indicate that the proposed method can effectively capture comprehensive features of rolling bearing data,significantly improve the anti-interference capability of the model,and thus provide an efficient solution for rolling bearing fault diagnosis.
作者
何安军
徐涛
王浩
汤秋秋
曹建行
李亚成
卢长江
HE Anjun;XU Tao;WANG Hao;TANG Qiuqiu;CAO Jianhang;LI Yacheng;LU Changjiang(Hunan Shizhuyuan Nonferrous Metals Co.,Ltd.,Chenzhou 424300,China)
出处
《无线电工程》
2025年第12期2469-2478,共10页
Radio Engineering
基金
江西省教育厅科学技术研究青年项目(GJJ2200848)。
关键词
滚动轴承
时频双域融合
图形推理
抗干扰能力
rolling bearing
time-frequency dual-domain fusion
graph reasoning
anti-interference capability
