改进DenseNet的干气密封摩擦润滑状态识别研究  

Study of dry gas seal friction lubrication state identification with improved DenseNet

作　　者：张帅 丁雪兴[1] 王世鹏 力宁 张兰霞 ZHANG Shuai;DING Xuexing;WANG Shipeng;LI Ning;ZHANG Lanxia(School of petroleum and chemical engineering,Lanzhou University of technology,Lanzhou 730050,China;State Key Laboratory of Tribology in Advanced Equipment,Tsinghua University,Beijing 100084,China;AECC Hunan Aviation powerplant Research Institute,Zhuzhou 412002,China)

机构地区：[1]兰州理工大学石油化工学院,兰州730050 [2]清华大学摩擦学国家重点试验室,北京100084 [3]中国航发湖南动力机械研究所,湖南株洲412002

出　　处：《振动与冲击》2025年第4期313-321,共9页Journal of Vibration and Shock

基　　金：国家重点研发计划(2020YFB2010001);国家自然科学基金项目(51565029);甘肃省重点研发计划(24YFGA029)。

摘　　要：为了克服干气密封运行中端面接触状态参数(膜厚、端面开启时间)测量困难的问题,提出自注意力机制融合稠密连接网络(DenseNet-convolutional block attention module,DenseNet-CBAM)的干气密封端面摩擦润滑状态识别方法。根据斯特里贝克曲线和干气密封运行规律分析端面可能出现的摩擦润滑状态:流体润滑,边界润滑、混合润滑。通过声发射传感器采集密封系统运行时的声发射信号,通过滤波、时域分析、频域分析得出能够表征各种摩擦润滑状态的特征分量,获取三维连续小波(3D continuous wavelet transform,3D-CWT)时频图,最终基于深度学习模型Densenet-CBAM识别时频图,实现密封系统摩擦润滑状态识别。与其他二维时频特征图作为输入端相比,3D-CWT时频图提高了状态识别的准确率。同时,相较于其他深度学习模型,该方法对干气密封摩擦润滑状态识别精度高,达到了99.27%。In order to solve the problem that it is difficult to measure key contact state parameters(such as membrane thickness and open time of end faces)during the operation of dry gas seals,a method for recognizing the friction lubrication states of dry gas seal end faces was proposed by integrating a self⁃attention mechanism with a DenseNet⁃CBAM(convolutional block attention module)network.Based on the Stribeck curve and the operating laws of dry gas seals,the potential friction lubrication states of the end faces were identified as fluid,boundary and mixed lubrication.Acoustic emission signals generated during the operation of the seal system were collected by using sensors,and characteristic components representing various friction lubrication states were extracted through filtering,time⁃domain and frequency⁃domain analysis.A 3D continuous wavelet transform(3D⁃CWT)was applied to generate time⁃frequency maps,and ultimately,the deep learning model DenseNet⁃CBAM was employed to recognize these maps and identify the friction lubrication states of the seal system.The results show that the 3D⁃CWT time⁃frequency map improves the accuracy of state identification compared to other 2D time⁃frequency feature maps as inputs.Compared to other deep learning models,this approach achieves higher accuracy in identifying the friction lubrication state of dry gas seals,reaching 99.27%.

关 键 词：干气密封 稠密连接网络 自注意力机制 声发射 状态识别 

分 类 号：TB523[理学—物理] TH113.1[理学—声学]