基于长短时记忆网络的旋转机械状态预测研究  

State Prognosis of Rotary Machines Based on Long/Short Term Memory Recurrent Neural Network

作　　者：赵建鹏[1] 周俊[1] ZHAO Jian-peng;ZHOU Jun(School of Mechanical Engineering, Shanghai University of Engineering Science,Shanghai 201620, China)

机构地区：[1]上海工程技术大学机械工程学院,上海201620

出　　处：《噪声与振动控制》2017年第4期155-159,共5页Noise and Vibration Control

基　　金：教育部留学回国人员科研启动基金资助项目(教外司留[2012]940号)

摘　　要：作为深度学习算法的一种,长短时记忆网络越来越成为时间序列预测的重要手段,简要阐述长短时记忆网络的基本原理,并将其应用于旋转机械状态监控领域,以轴承数据为样本进行仿真,针对轴承数据的非平稳性,运用经验模态分解方法将其分解为平稳信号,并计算其本征模态分量能量熵作为状态特征,通过计算长短时记忆网络对旋转机械状态单步预测的结果,并与支持向量回归机模型的预测结果进行比较,证明长短时记忆网络在旋转机械状态预测方面可以取得比支持向量回归机更好的效果。As a method of deep learning algorithm,long/short term memory network(LSTM)has been playing an increasingly important role in the field of time series prediction.In this paper,a new method for state prediction of rotary machines based LSTM is proposed.And the principle of LSTM applied to rotary machine state monitoring and prognosis is introduced and applied to state monitoring of the rotary machinery.In order to avoid the instability of the bearing data,the empirical mode decomposition(EMD)is used to decompose the bearing data into several steady signals,and the intrinsic mode function(IMF)energy entropy is calculated as the feature of the machine state.Furthermore the analysis of simulation results based on LSTM is compared with those of support vector regression machine(SVRM).LSTM has achieved a stateof-the-art prediction performance in the single-step prediction of rotary machine states.It is shown that the proposed prognosis technology is superior in rotary machine state prediction and monitoring.

关 键 词：振动与波 长短时记忆 故障预测 状态监控 经验模态分解 能量熵 

分 类 号：TP206.3[自动化与计算机技术—检测技术与自动化装置]