判别结构保持的跨工况故障诊断方法OA
Cross-condition Fault Diagnosis Method Based on Discriminative Structure Preservation
针对跨工况轴承故障诊断中源域与目标域特征分布偏移、目标域类别判别结构退化等问题,提出一种判别结构保持协同融合网络.该方法以频域振动信号为输入,利用图卷积网络-时间卷积网络并联结构分别提取频谱局部关联特征和跨频带依赖特征,并通过双向交叉注意力机制实现异构特征协同融合;同时,引入基于余弦相似度的判别结构保持约束,增强跨域同类样本的特征一致性.在华中科技大学和东南大学 2 个轴承故障数据集上的实验结果表明,所提方法的平均迁移准确率分别达到 99.71%±0.13%和99.63%±0.11%,均优于对比方法,具有较好的跨工况迁移诊断性能.
To address the problems of feature distribution shift between source and target domains and the degradation of the category-wise discriminative structures in the target domain for cross-condition bearing fault diagnosis,this paper proposes a discriminative structure-preserved collaborative fusion net-work.The proposed method takes frequency-domain vibration signals as inputs,and employs a parallel ar-chitecture comprising a graph convolutional network and a temporal convolutional network to extract local-ly correlated spectral features and cross-frequency-band dependency features,respectively.These hetero-geneous features are then synergistically fused via a bidirectional cross-attention mechanism.Meanwhile,a discriminative structure preservation constraint based on cosine similarity is introduced to enhance the feature consistency of samples from the same class cross domains.Experimental results on two bearing fault datasets of Huazhong University of Science and Technology and Southeast University show that the proposed method achieves average transfer accuracies of 99.71%±0.13%and 99.63%±0.11%,respec-tively,outperforming the compared methods and exhibiting superior cross-condition transfer diagnosis performance.
张磊;缪秋循;刘川;李晓成;顾小虎;余洋;朱静;邓艾东
南京南自华盾数字技术有限公司,江苏 南京 210003南京南自华盾数字技术有限公司,江苏 南京 210003南京南自华盾数字技术有限公司,江苏 南京 210003南京南自华盾数字技术有限公司,江苏 南京 210003南京南自华盾数字技术有限公司,江苏 南京 210003南京南自华盾数字技术有限公司,江苏 南京 210003河南科技大学车辆与交通工程学院,河南 洛阳 471003东南大学能源与环境学院,江苏 南京 210003
机械制造
滚动轴承跨工况故障诊断判别结构保持特征融合图卷积网络
rolling bearingcross-condition fault diagnosisdiscriminative structure preservationfea-ture fusiongraph convolutional network
《机械与电子》 2026 (9)
1-8,8
国家自然科学基金资助项目(52576010)
评论