摘要
轴承作为机械系统中关键的部件,其状态监测和剩余使用寿命(remaining useful life,RUL)预测变得日益重要。为此,提出了一种基于Transformer的时频域双通道融合网络(time frequency domain dual channel fusion network,TFDN)以解决深度学习的RUL预测方法中存在难以考虑数据的空间关系和样本贡献度,以及未能充分挖掘时频域数据的关联性和互补性等问题。该方法利用二维卷积提取时频域信号的空间特征,通过Transformer编码器层对卷积层输出进行位置编码学习退化特征。设计了一种时频域融合回归算法,对时频域信号进行权重特征融合实现RUL预测。通过在西安交通大学轴承数据集和青岛港实际采集数据集上进行测试,试验结果显示TFDN在预测精度上超过了其他网络结构和现有方法。
Bearing is a key component in mechanical systems,and its condition monitoring and remaining useful life(RUL)prediction are becoming increasingly important..To address the challenges in deep learning⁃based RUL prediction methods,such as the difficulty in considering spatial relationships and sample contributions of data,and the insufficient exploration of the correlation and complementarity of time⁃frequency domain data,a Transformer⁃based time frequency domain dual channel fusion network(TFDN)was proposed.two⁃dimensional convolution was employed to extract spatial features from time⁃frequency domain signals and a Transformer encoder layer was used to perform positional encoding and learn degradation features from the convolutional layer outputs.A time⁃frequency domain fusion regression algorithmwas designed to achieve RUL prediction by performing weighted feature fusion on time⁃frequency domain signals.Experimental results on the Xi’an Jiaotong University bearing dataset and the Qingdao Port real⁃world collected dataset demonstrate that TFDN surpasses other network structures and existing methods in prediction accuracy.
作者
徐浩
高乾
王铭榜
吕成兴
杨智博
陈健
XU Hao;GAO Qian;WANG Mingbang;LÜChengxing;YANG Zhibo;CHEN Jian(School of Information and Control Engineering,Qingdao University of Technology,Qingdao 266000,China;Qingdao High-Tech Industrial Development Co.,Ltd.,Qingdao 266000,China)
出处
《振动与冲击》
北大核心
2025年第4期322-332,共11页
Journal of Vibration and Shock
基金
青岛市关键技术攻关及产业化示范类项目(23-1-2-qljh-6-gx)
青岛市自然科学基金(23-2-1-154-zyyd-jch)。
