基于CEEMDAN与RNN算法的离心泵空化状态识别研究  

Research on Cavitation State Identification of Centrifugal Pump Based on RNN and CEEMDAN Algorithm

作　　者：郎涛[1] 赵佳伟 陈刻强 徐恩翔 倪陈 印佳 张德胜[1] Lang Tao;Zhao Jiawei;Chen Keqiang;Xu Enxiang;Ni Chen;Yin Jia;Zhang Desheng(National Research Center of Pumps,Jiangsu University,Zhenjiang 212013,China)

机构地区：[1]江苏大学国家水泵及系统工程技术研究中心,镇江212013

出　　处：《水动力学研究与进展（A辑）》2023年第4期629-636,共8页Chinese Journal of Hydrodynamics

基　　金：国家自然科学基金联合基金重点项目(U2106225);江苏省杰出青年基金(BK20211547);2021年度江苏省高校优秀科技创新团队项目(SKJ(2021)-1)。

摘　　要：离心泵在发生空化时会产生蕴含特征信息的噪声信号,且随着空化程度的增加,噪声信号特征也会发生变化。该文针对离心泵空化状态识别的问题,提出了一种基于完全自适应集合经验模态分解(CEEMDAN)算法和循环神经网络(RNN)相结合的离心泵空化状态识别方法。通过空化性能试验获取不同空化状态的噪声信号数据,并进行频域分析获取其特征频段范围;引入CEEMDAN算法对信号数据进行分解获得相应的固有函数模态(IMF)分量,选取峭度较大的分量进行频域分析,并选择符合特征频段的分量进行功率谱熵计算以构造特征向量集合,发现空化与非空化状态下,功率谱熵值区分明显,在非空化时均值为0.55,空化时均值为1.05。最后将特征分量集合输入循环神经网络完成模式识别,选取最优模型。验证实验数据分析表明:采用CEEMDAN和RNN算法相结合的模型可以有效提取离心泵空化噪声的特征并进行空化状态识别,最高识别率达92.4%,为离心泵故障诊断提供了新的方法。The centrifugal pump will produce noise signals containing characteristic information when cavitation occurs,and as the degree of cavitation enhances,the characteristics of the noise signal will also change.This study proposes a cavitation state recognition method of centrifugal pump based on CEEMDAN algorithm and recurrent neural network aiming to the problem of cavitation state recognition of centrifugal pump.The noise signal data of different cavitation states are obtained by cavitation performance test,and the characteristic frequency band range is obtained by frequency domain analysis.The CEEMDAN algorithm is introduced to decompose the signal data to obtain the corresponding intrinsic function mode components.The components with larger kurtosis are selected for frequency domain analysis,and the components that conform to the characteristic frequency band are selected for power spectrum entropy calculation to construct the feature vector set.It is found that the power spectrum entropy values in cavitation and non-cavitation states are obviously different.The average value is 0.55 in non-cavitation and 1.05 in cavitation.Finally,the feature component set is input into the recurrent neural network to complete the pattern recognition and select the optimal model.The experimental data analysis shows that the model combining CEEMDAN and RNN algorithm can effectively extract the characteristics of cavitation noise of centrifugal pump and identify the cavitation state.The highest recognition rate is 92.4%,which provides a new method for fault diagnosis of centrifugal pump.

关 键 词：状态识别 空化噪声 功率谱熵 循环神经网络 

分 类 号：TH311[机械工程—机械制造及自动化] O427.4[理学—声学] TP181[理学—物理]