电子测量与仪器学报

吕轶,祝志荣,赵天宇.基于跨传感器特征融合的滚动轴承故障诊断[J].电子测量与仪器学报,2026,40(3):114-123

基于跨传感器特征融合的滚动轴承故障诊断

Fault diagnosis of rolling bearings based on cross-sensor feature fusion

DOI：

英文关键词:fault diagnosis cross-sensor strong noise attention mechanism feature fusion

基金项目:国家自然科学基金(52075086)、辽宁省教育厅项目(JYT2020049)、中国航空工业空气动力研究院项目(2022020600018)资助

作者	单位
吕轶	沈阳航空航天大学电子信息工程学院沈阳110136
祝志荣	沈阳航空航天大学电子信息工程学院沈阳110136
赵天宇	东北大学理学院沈阳110819

Author	Institution
Lyu Yi	School of Electronic Information Engineering, Shenyang Aerospace University, Shenyang 110136, China
Zhu Zhirong	School of Electronic Information Engineering, Shenyang Aerospace University, Shenyang 110136, China
Zhao Tianyu	School of Sciences, Northeastern University, Shenyang 110819，China

摘要点击次数: 365

全文下载次数: 221

中文摘要:

神经网络处理复杂场景下的故障诊断任务会面临强噪声干扰和单个传感器信息不完整，针对其导致诊断性能下降问题，提出了一种跨传感器特征融合的卷积神经网络模型。首先，提取数据特征时结合卷积神经网络与双环残差模块来缓解训练过程中的梯度消失问题；然后，引入卷积注意力模块进而提升模型对关键特征的关注能力；随后，使用特征优化重构模块改善特征学习效率与表达能力；之后，利用自适应特征融合模块将不同传感器所提取的高级特征进行自适应融合；最后，通过全局平均池化层、全连接层和Softmax函数对融合后的特征进行分类，实现故障诊断任务。结果表明，该方法能有效融合多传感器数据特征，且对不同强度噪声表现出良好的鲁棒性；模型平均诊断准确率在-2~-18 dB噪声下达到97.48%，相对于单个传感器提升1.88%，为解决复杂场景下的故障诊断提供有效参考。

英文摘要:

Neural networks employed for fault diagnosis in complex scenarios often face challenges like strong noise interference and incomplete information from individual sensors, leading to degraded diagnostic performance. To address this issue, a cross-sensor convolutional neural network with dual residual and feature adaptation model is proposed. Firstly, during the data feature extraction process, a dual-ring residual module is utilized to alleviate the gradient vanishing problem during training. Subsequently, the convolutional block attention module attention mechanism is introduced to enhance the model's ability to focus on critical features. Then, a feature optimization and reconstruction module is utilized to improve the efficiency of feature learning and the capability of feature expression. Thereafter, an adaptive feature fusion module is employed to adaptively fuse high-level features extracted from different sensors. Finally, the fused features are classified through a global average pooling layer, a fully connected layer, and a Softmax function to accomplish the fault diagnosis task. The results demonstrate that the proposed method effectively integrates multi-sensor data features and exhibits robustness against noise of varying intensities. The average diagnostic accuracy of the model reaches 97.48% under noise levels ranging from -2 to -18 dB, showing an improvement of 1.88% compared to using a single sensor. This study provides an effective reference for solving fault diagnosis problems in complex scenarios.

查看全文查看/发表评论下载PDF阅读器