融合GRU单元的CNN网络在石油旋转机械故障诊断中的应用  

Application of CNN network fused with GRU units in fault diagnosis of petroleum rotating machinery

作　　者：苗俊田[1] 鹿德台 李卓军 刘冬冬[1] 赵博[1] MIAO Jun-tian;LU De-tai;LI Zhuo-jun;LIU Dong-dong;ZHAO Bo(China University of Petroleum,Qingdao 266400,Shandong Province,China;Qingdao Shida Huatong Technology Co.,Ltd.,Qingdao 266400,Shandong Province,China)

机构地区：[1]中国石油大学(华东),山东青岛266400 [2]青岛石大华通科技有限责任公司,山东青岛266400

出　　处：《信息技术》2024年第10期7-13,共7页Information Technology

基　　金：国家自然科学基金面上项目(52074340)。

摘　　要：针对现有石油旋转机械故障诊断算法存在的问题,提出一种融合GRU单元优化改进的CNN网络算法模型。先利用小波包算法保留高频区间的弱故障信号特征,选择ReLU函数作为CNN网络卷积层的激活函数,提升算法的运行速率并抑制模型梯度值的过快衰减;基于GRU单元整合输入门和遗忘门,抑制CNN网络存在的梯度弥散问题,改善了CNN网络模型的故障数据训练性能和分类检测性能。实验结果显示:融合GRU单元的CNN网络对训练集和测试集的故障诊断精度和故障分类能力均优于现有故障诊断算法,而且MAE函数的预测值和真实值更接近。A CNN network algorithm model that integrates GRU unit optimization improvement is proposed to address the problems of existing fault diagnosis algorithms for petroleum rotating machinery.Firstly,the wavelet packet algorithm is used to retain the characteristics of weak fault signals in the high frequency interval,and the ReLU function is selected as the activation function of the convolution layer of the CNN network to improve the running speed of the algorithm and suppress the excessive attenuation of the gradient value of the model.By integrating input gates and forgetting gates based on GRU units,the gradient dispersion problem in CNN networks is suppressed,and the fault data training performance and classification detection performance of CNN network models are improved.The experiment results show that the CNN network that integrates GRU units has better fault diagnosis accuracy and classification ability than the existing fault diagnosis algorithms in both the training and testing sets,and the predicted values of the MAE function are closer to the true values.

关 键 词：门控循环单元 深度卷积神经网络 转盘轴承 dropout网络 MAE函数 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]