基于QSE-ResNet的浮环密封摩擦振动信号分类方法  

Classification method for frictional vibration signals of floating ring seal based on QSE-ResNet

作　　者：刘伟 张书尧 翟志兴 朱书海 李双喜[1] LIU Wei;ZHANG Shuyao;ZHAI Zhixing;ZHU Shuhai;LI Shuangxi(Collage of Mechanical and Electrical Engineering,Beijing University of Chemical Technology,Beijing 100029,China)

机构地区：[1]北京化工大学机电工程学院,北京100029

出　　处：《振动与冲击》2024年第21期194-201,共8页Journal of Vibration and Shock

基　　金：国家重点研发计划项目(2018YFB2000800,2022YFB3303600)。

摘　　要：针对浮环密封装置在不同摩擦工况下振动信号特征微弱且难以识别的问题,传统深度学习网络在性能上取得了显著成果。然而,由于存在模型超参数多、训练时间长、迭代次数多、网络精度低以及计算成本过高等问题,其在实际使用时存在一定的局限性。因此,该研究提出了一种基于深度残差网络(residual network,ResNet)的方法,即基于快速注意力机制的残差网络(quick squeeze excitation ResNet,QSE-ResNet),以解决浮环密封摩擦振动信号分类中的问题。该方法通过引入注意力机制、调整网络残差块的连接方式并选择特定的优化器来提高模型性能,并与其他四种模型在同一个梅尔频谱图(Mel Spectrogram)数据集上进行对比测试。研究结果显示,QSE-ResNet的准确率达到了97%,比传统卷积神经网络高出13%,同时缩短了55%的模型迭代次数,节约了30%的网络训练时间。此外,QSE-ResNet成功地解决了过拟合、梯度爆炸和梯度消失等问题,显著缩短了迭代次数、节省了网络训练时间并提高了测试精度,使得浮环密封摩擦振动的信号状态监控及设备部署更为便利。该研究提出的QSE-ResNet使得浮环密封摩擦振动模型具备更便捷的部署能力,为浮环密封摩擦振动信号的研究提供了新的思路。Aiming at the problem of vibration signal characteristics of floating ring seal device being weak and difficult to identify under different friction conditions,traditional deep learning network achieves significant performance results.However,due to problems of many hyperparameters,long training time,many iterations,low network accuracy and excessive computational costs,the model has certain limitations in practical use.Here,a method based on deep residual network(ResNet)called the fast attention mechanism-based quick squeeze excitation(QSE)-ResNet was proposed to solve the problem of classifying friction vibration signals in floating ring seal.This method could improve model performance by introducing attention mechanism,adjusting connection form of network residual blocks and selecting specific optimizers.It was tested contrastively with the other 4 models on the same Mel Spectrogram dataset.The study results showed that the accuracy of QSE-ResNet reaches 97%,it is 13%higher than traditional convolutional neural network(CNN),meanwhile,it shortens the number of model iterations by 55%and saves 30%of network training time;in addition,QSE-ResNet successfully solves problems of overfitting,gradient blast and gradient disappearance to significantly reduce number of iterations,save network training time,improve testing accuracy and make status monitoring and equipment deployment of floating ring seal friction vibration signals be more convenient;the proposed QSE-ResNet can provide a more convenient deployment capability for friction vibration model of floating ring seal,and new ideas for studying friction vibration signals of floating ring seal.

关 键 词：浮环密封 摩擦振动 声发射 深度学习 特征分类 

分 类 号：TH212[机械工程—机械制造及自动化] TH213.3