MBGA-FuseNet：双分支图网络与多域自适应融合的轴承故障诊断框架

doi:10.3969/j.issn.1004-132X.2026.09.009

摘要/Abstract

摘要：

针对复杂工况下滚动轴承故障信号存在多尺度特征提取不足、时空依赖能力弱及跨域特征融合不充分等问题，提出了双分支图网络与多域自适应融合（MBGA-FuseNet）的轴承故障诊断框架。构建双分支图神经网络架构：将改进位置编码的图级Transformer与自适应多尺度切比雪夫图卷积结合以增强节点拓扑性从而实现不同邻域信息的跨尺度自适应融合，并结合块间全局自注意力编码器提升特征的深度表达与层级关联。设计了非线性典型相关分析（NL-CCA）逐节点级跨通道特征融合模块，在节点层面实现跨分支特征的自适应关联与动态加权融合，并提出一种改进的通道融合注意力模块（CFAM），通过引入快速卷积门控线性单元（Faster-CGLU）机制，细化通道门控信号的细粒度。最后在时序特征建模方面，采用双向门控递归单元（BiGRU）结构捕获时间依赖关系，结合提出的多域交互融合框架（HIFTFS），将时域、频域与空间域特征进行深度交叉融合。将所提方法在多个公开轴承数据集上进行了实验验证，不同工况与信噪比条件下的准确率均优于其他对比模型。

关键词: 故障诊断, 图神经网络, 自适应融合, 注意力, 交叉融合

Abstract:

To address the problems of insufficient multi-scale feature extraction， weak spatio-temporal dependency， and inadequate cross-domain feature fusion in rolling bearing fault signals under complex operating conditions， a multi branch graph attention fusion network（MBGA-FuseNet） was proposed for bearing fault diagnosis. First， a dual-branch graph neural network architecture was constructed： a graph-level Transformer with improved positional encoding was combined with an adaptive multi-scale Chebyshev graph convolutional layer to enhance node topology and achieve cross-scale adaptive fusion of information from different neighborhoods. Furthermore， an inter-block global self-attention encoder was incorporated to enhance the deep representation and hierarchical correlation of feature representation. Second， a nonlinear canonical correlation analysis（NL-CCA） node-level cross-channel feature fusion module was designed to realize adaptive correlation and dynamic weighted fusion across branches at the node level. Third， a improved channel fusion attention module（CFAM） was proposed by introducing a faster convolutional gated linear unit（Faster-CGLU） mechanism， which refines the granularity of channel gating signals. Finally， for temporal feature modeling， a bidirectional gated recurrent unit （BiGRU） was employed to capture temporal dependencies. Moreover， a hierarchical inter-domain frequency-temporal feature synthesis（HIFTFS） was developed to deeply cross-fuse features from the temporal， frequency and spatial domains. Experimental results on multiple publicly available bearing datasets demonstrate that the proposed method consistently outperforms other comparative models under various operating conditions and signal-to-noise ratios.

Key words: fault diagnosis, graph neural network, adaptive fusion, attention, interactive fusion

中图分类号:

TH133.3

姜佩贺, 刘银河, 王晨旭. MBGA-FuseNet：双分支图网络与多域自适应融合的轴承故障诊断框架[J]. 中国机械工程, 2026, 37(9): 2166-2176.

JIANG Peihe, LIU Yinhe, WANG Chenxu. MBGA-FuseNet： Bearing Fault Diagnosis Framework Based on Dual-branch Graph Network and Multi-domain Adaptive Fusion[J]. China Mechanical Engineering, 2026, 37(9): 2166-2176.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://www.cmemo.org.cn/CN/10.3969/j.issn.1004-132X.2026.09.009

https://www.cmemo.org.cn/CN/Y2026/V37/I9/2166

图/表 19

图1 GRU内部结构图

Fig.1 Internal structure diagram of GRU

图2 BiGRU结构简图

Fig.2 BiGRU structure schematic

图3 MBGA-FuseNet轴承故障诊断总体框架

Fig.3 General framework for bearing fault diagnosis with MBGA-FuseNet

图4 ILAM示意图

Fig.4 Schematic diagram of ILAM

图5 节点级跨通道特征融合层

Fig.5 Node-level cross-channel feature fusion layer

图6 CAM模块

Fig.6 CAM module

图7 Faster-CGLU模块

Fig.7 Faster-CGLU module

图8 HIFTFS框架

Fig.8 HIFTFS framework

表1 CWRU滚动轴承数据集说明

Tab.1 Description of CWRU rolling bearing data set

故障位置	故障直径	标签	数据集
内圈	0.007	1	105
	0.014	2	169
	0.021	3	209
外圈	0.007	4	130
	0.014	5	197
	0.021	6	234
滚珠	0.007	7	118
	0.014	8	185
	0.021	9	222
正常	0	0	97

表2 西安交通大学滚动轴承数据集

Tab.2 Description of XJTU-SY rolling bearing data set

工况	标签	数据集	故障位置
35 Hz/12 kN	0	Bearing 1_1	外圈
35 Hz/12 kN	1	Bearing 1_4	保持架
37.5 Hz/11 kN	2	Bearing 2_1	内圈
	3	Bearing 2_2	外圈
	4	Bearing 2_3	保持架
40 Hz/10 kN	5	Bearing 3_1	外圈
40 Hz/10 kN	6	Bearing 3_3	内圈

表3 消融实验结果

Tab.3 Results of the ablation experiments

模型	WGT	MFCN	NL-CCA	CFAM	BiGRU	HIFTFS	精确率/%	召回率/%	F1分数
A（基准）	√	√					88.68	88.50	0.8859
B	√	√					94.12	93.20	0.9367
C	√	√			√		94.36	93.45	0.9390
D	√	√	√		√		95.12	94.87	0.9491
E	√	√		√	√		94.52	94.00	0.9426
F	√			√	√		91.16	90.96	0.9106
G		√		√	√		89.09	89.02	0.8899
H			√	√	√		86.72	86.74	0.8673
I	√	√	√	√	√		97.32	97.29	0.9731
本文方法	√	√	√	√	√	√	99.28	99.46	0.9937

图9 不同消融模型的混淆矩阵

Fig.9 Confusion matrices of different ablation models

表4 对比实验结果

Tab.4 Results of the comparative experiments

模型	准确率/ %	F1分数	召回率/ %	参数量/10⁶	浮点运算次数/ $109$
CNN-BiGRU^［19］	97.50	0.9775	98.00	0.780	1.23
BILSTM^［3］	87.60	0.8815	88.65	0.672	1.08
Informer^［20］	89.63	0.8998	90.34	8.96	18.6
SGCN^［21］	94.99	0.9524	95.50	1.09	2.64
Graph-Transformer^［15］	91.54	0.9216	92.79	1.46	2.14
SpGAT^［22］	79.60	0.8005	80.45	0.830	2.91
本文方法	99.52	0.9942	99.46	3.83	7.36

表4 对比实验结果

Tab.4 Results of the comparative experiments

模型	准确率/ %	F1分数	召回率/ %	参数量/10⁶	浮点运算次数/ $109$
CNN-BiGRU^［19］	97.50	0.9775	98.00	0.780	1.23
BILSTM^［3］	87.60	0.8815	88.65	0.672	1.08
Informer^［20］	89.63	0.8998	90.34	8.96	18.6
SGCN^［21］	94.99	0.9524	95.50	1.09	2.64
Graph-Transformer^［15］	91.54	0.9216	92.79	1.46	2.14
SpGAT^［22］	79.60	0.8005	80.45	0.830	2.91
本文方法	99.52	0.9942	99.46	3.83	7.36

表5 不同负载和信噪比下的实验结果 (%)

Tab.5 Experimental results with different loads and signal-to-noise ratios

模型	负载为0					负载为2237.1 W（3 HP）
模型	R=2 dB	R=4 dB	R=6 dB	R=8 dB	R=10 dB	R=2 dB	R=4 dB	R=6 dB	R=8 dB	R=10 dB
CNN-BiGRU	85.23	88.30	92.52	94.70	96.43	82.26	86.27	91.13	93.50	96.13
BILSTM	69.83	75.10	79.88	83.50	85.59	67.96	74.10	78.47	82.68	84.80
Informer	76.83	82.25	85.47	87.14	88.21	74.71	80.94	84.24	86.83	87.94
SGCN	83.17	86.21	89.98	91.72	93.20	81.08	84.68	88.71	91.08	92.76
Graph-Transformer	78.52	84.40	87.97	89.50	90.27	78.04	84.27	86.94	89.32	90.32
SpGAT	67.43	72.07	75.56	77.17	78.64	65.83	70.80	74.34	76.67	78.40
本文方法	93.41	94.09	95.58	96.95	98.71	91.32	93.82	95.34	96.81	98.58

图10 对比实验损失与准确率

Fig.10 The loss and accuracy of comparison experiments

图11 t-SNE可视化对比结果

Fig.11 The t-SNE visualization comparison results

图12 不同信噪比下的结果对比

Fig.12 Comparison of results under different signal-to-noise ratios

表6 不同信噪比下的对比试验结果 (%)

Tab.6 Comparison results under different signal-to-noise ratios

模型	信噪比
模型	2 dB	4 dB	6 dB	8 dB	10 dB	未添加噪声
CNN-BiGRU	84.53±0.42	86.38±0.78	88.55±0.50	88.74±0.41	91.10±0.32	93.32±0.35
BILSTM	68.76±0.12	72.26±0.58	77.98±1.08	79.93±0.62	81.85±0.25	82.63±0.26
Informer	54.88±0.61	57.00±0.72	59.02±0.74	62.90±0.41	69.84±1.89	75.66±0.82
SGCN	74.80±0.71	76.61±0.37	80.27±0.51	82.11±0.61	82.63±0.33	89.76±0.68
Graph-Transformer	68.62±1.12	70.93±0.28	72.91±0.62	75.07±0.59	79.61±0.84	80.39±0.45
SpGAT	51.34±1.38	56.23±2.34	61.55±0.76	64.31±0.45	66.67±0.38	69.70±1.08
本文所提方法	92.79±0.19	94.67±0.15	96.77±0.03	97.56±0.06	97.93±0.02	98.59±0.02

图13 噪声环境下不同负载的结果对比

Fig.13 Comparison of results under different loads in a noisy environment

参考文献 22

[1]	HUANG Delin， SU Xiangdong， YANG Jinghui， et al. An Improved Dual-channel CNN-BILSTM Fusion Attention Model for Fault Diagnosis of Aero-engine Bearings［J］. Measurement， 2025， 253： 117761.
[2]	REN Zelin， JIANG Yuchen， YANG Xuebing， et al. Learnable Faster Kernel-PCA for Nonlinear Fault Detection： Deep Autoencoder-based Realization［J］. Journal of Industrial Information Integration， 2024， 40： 100622.
[3]	LUO Yijing， FENG Chen， ZHANG Yuquan， et al. Vibration Fault Diagnosis of Hydroelectric Generator Sets Based on Multi-dimensional Feature Extraction and CNN-BiLSTM［J］. Measurement， 2026， 264： 120034.
[4]	邹筱瑜，孙国庆，王忠宾，等. 基于时频融合深度网络的矿用钻机轴承故障诊断［J］. 中国机械工程， 2024， 35（8）： 1405-1413.
	ZOU Xiaoyu， SUN Guoqing， WANG Zhongbin， et al. Bearing Fault Diagnosis of Mining Drilling Rig with Time-frequency-fused Deep Network［J］. China Mechanical Engineering， 2024， 35（8）： 1405-1413.
[5]	LEI Chunli， WANG Lu， ZHANG Qiyue， et al. Rolling Bearing Fault Diagnosis Method Based on DSICNN under Strong Noise Background［J］. Nondestructive Testing and Evaluation， 2025， 40（12）： 5804-5820.
[6]	LIANG Haopeng， ZHAO Xiaoqiang. Rolling Bearing Fault Diagnosis Based on One-dimensional Dilated Convolution Network with Residual Connection［J］. IEEE Access， 2021， 9： 31078-31091.
[7]	YANG Aoyu， ZHONG Zhidan， ZHAO Yao， et al. Rolling Bearing Fault Diagnosis Method Based on Dynamic Convolution Capsule Network［J］. Journal of Physics： Conference Series， 2023， 2419（1）： 012091.
[8]	阴艳超，施成娟，邹朝普，等. 基于深度时间卷积神经网络与迁移学习的流程制造工艺过程质量时序关联预测［J］. 中国机械工程， 2023， 34（14）： 1659-1671.
	YIN Yanchao， SHI Chengjuan， ZOU Chaopu， et al. Time-series Correlation Prediction of Quality in Process Production Processes Based on Deep TCN and Transfer Learning［J］. China Mechanical Engineering， 2023， 34（14）： 1659-1671.
[9]	LI Tianfu， ZHOU Zheng， LI Sinan， et al. The Emerging Graph Neural Networks for Intelligent Fault Diagnostics and Prognostics： a Guideline and a Benchmark Study［J］. Mechanical Systems and Signal Processing， 2022， 168： 108653.
[10]	宋李俊，刘松林，辛玉，等. 基于轴承退化状态评估和改进图注意力双向门控循环单元网络的轴承剩余寿命预测［J］. 中国机械工程， 2025， 36（7）： 1562-1572.
	SONG Lijun， LIU Songlin， XIN Yu， et al. Residual Life Prediction for Bearings Based on Bearing Degradation State Assessment and IGAT-BiGRU Network［J］. China Mechanical Engineering， 2025， 36（7）： 1562-1572.
[11]	ZUO Dong， TANG Tang， CHEN Ming. Rolling Bearing Fault Diagnosis Based on Multi-scale Weighted Visibility Graph and Multi-channel Graph Convolution Network［J］. Measurement Science and Technology， 2023， 34（11）： 115019.
[12]	FAN Xiaoxuan， DUAN Lixiang， ZHANG Na. A Multi-scale Graph-guided Dynamic Enhanced Alignment Network for Mechanical Fault Diagnosis Considering Domain Shift and Data Imbalance［J］. Neurocomputing， 2025， 625： 129546.
[13]	WEN Haoran， GUO Wei， LI Xiang. A Novel Deep Clustering Network Using Multi-representation Autoencoder and Adversarial Learning for Large Cross-domain Fault Diagnosis of Rolling Bearings［J］. Expert Systems with Applications， 2023， 225： 120066.
[14]	WANG Zhe， WU Zhiying， LI Xingqiu， et al. Attention-aware Temporal-Spatial Graph Neural Network with Multi-sensor Information Fusion for Fault Diagnosis［J］. Knowledge-Based Systems， 2023， 278： 110891.
[15]	吴百礼. 基于图卷积神经网络的轴承故障诊断与剩余寿命预测方法研究［D］. 广州：华南理工大学， 2023.
	WU Baili. Research on Bearing Fault Diagnosis and Remaining Life Prediction Method Based on Graph Convolutional Neural Network［D］. Guangzhou： South China University of Technology， 2023.
[16]	WANG Jinhua， CAO Wenbao， CAO Jie， et al. Fault Diagnosis Method Based on Multi-scale Adaptive Super-graph Convolutional Neural Networks［J］. Applied Acoustics， 2025， 240： 110889.
[17]	NI Peihao， ZHANG Yuanyuan， XIONG Xiaoyun， et al. MSGAFN： Multi-scale Graph Attention Fusion Network for Machine Fault Diagnosis［J］. Proceedings of the Institution of Mechanical Engineers， Part C： Journal of Mechanical Engineering Science， 2024， 238（15）： 7894-7907.
[18]	张永宏，许鑫豪，尹贺峰，等. MFSF-DETR：一种基于多尺度特征移位融合的PCB缺陷与元件检测算法［J］. 仪器仪表学报， 2025， 46（8）： 266-285.
	ZHANG Yonghong， XU Xinhao， YIN Hefeng， et al. MFSF-DETR： a PCB Defect and Component Detection Algorithm Based on Multi-scale Feature Shift Fusion［J］. Chinese Journal of Scientific Instrument， 2025， 46（8）： 266-285.
[19]	XU Ziwei， LI Yanfeng， HUANG Hongzhong， et al. A Novel Method Based on CNN-BiGRU and AM Model for Bearing Fault Diagnosis［J］. Journal of Mechanical Science and Technology， 2024， 38（7）： 3361-3369.
[20]	ZHANG Xin， YANG Kaiyue， ZHENG Liaomo. Transformer Fault Diagnosis Method Based on TimesNet and Informer［J］. Actuators， 2024， 13（2）： 74.
[21]	HUANG Zhenhua， ZHOU Wenhao， LI Kunhao， et al. SGCN： a Scalable Graph Convolutional Network with Graph-shaped Kernels and Multi-channels［J］. Knowledge-Based Systems， 2023， 279： 110923.
[22]	CHANG Heng， RONG Yu， XU Tingyang， et al. Spectral Graph Attention Network with Fast Eigen-approximation［C］//Proceedings of the 30th ACM International Conference on Information & Knowledge Management. Online， 2021： 2905-2909.