融合路径聚合网络的Swin Transformer的故障诊断方法研究  

Fault diagnosis method based on Swin Transformer with path aggregation networks

作　　者：刘晨宇 李志农[1] 熊鹏伟 谷丰收[2] LIU Chenyu;LI Zhinong;XIONG Pengwei;GU Fengshou(Key Laboratory of Nondestructive Testing of the Ministry of Education,Nanchang Hangkong University,Nanchang 330063,China;Centre for Efficiency and Performance Engineering,University of Huddersfield,London HD13DH,UK)

机构地区：[1]南昌航空大学无损检测技术教育部重点实验室,南昌330063 [2]哈德斯菲尔德大学效率与性能工程中心,英国伦敦HD13DH

出　　处：《振动与冲击》2024年第18期258-266,共9页Journal of Vibration and Shock

基　　金：国家自然科学基金项目(52075236);江西省自然科学基金重点项目(20212ACB202005);南昌航空大学研究生专项基金资助项目(YC2023-046)。

摘　　要：针对Transformer在航空发动机故障诊断中存在空间信息特征建模能力不足、计算复杂度较高的问题,提出一种基于路径聚合网络的Swin Transformer的故障诊断方法。该方法将路径聚合网络嵌入到Swin Transformer网络中,提高模型多尺度融合特征金字塔顶层信息和底层信息的效率,并采用窗口多头自注意力模块和移动窗口多头自注意力模块,有效降低提取空间信息特征的计算复杂度,并促进信息的流动和特征的传递。最后,将提出的方法应用到航空发动机滚动轴承故障诊断中。试验结果表明,提出的方法明显优于Transformer和传统Swin Transformer方法,在保证识别精度的同时,提高了模型的识别速度。To address the insufficient spatial information feature modeling capability and high computational complexity of the Transformer in aero-engine fault diagnosis,a fault diagnosis approach was proposed based on the Swin Transformer with path aggregation networks(PANet).In the proposed method,the Swin Transformer with PANet improves the efficiency of fusing the multiscale feature pyramid top and bottom informations.Then,window-based multi-head self-attention and shift window-based multi-head self-attention modules were used to reduce the computational complexity in spatial information feature extraction.Therefore,the information flow and feature transmission can be promoted effectively.Finally,the proposed method was applied in fault diagnosis of the aero-engine rolling bearings.The experimental results show that the proposed method is better than the Transformer and traditional Swin Transformer methods.While guaranteeing the recognition accuracy,the recognition speed of the model is improved.

关 键 词：故障诊断 Swin Transformer 路径聚合网络 航空发动机 滚动轴承 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置] TP183[自动化与计算机技术—控制科学与工程]