摘要
提出一种改进一维卷积神经网络(improved onedimensional convolution neural network,ICNN-1D),用于辨识汽轮发电机组轴系扭振模态参数。论文设计一种包含输入层、三组一维卷积层和池化层、全连接层与输出层的一维卷积神经网络;为了克服该网络模态参数定阶难的缺点,对其进行了无监督学习网络结构改造;并在Adam优化算法基础上,提出了动态衰减学习率优化策略,以避免网络学习过程出现震荡。以某600MW汽轮发电机组为研究对象,通过发电机5°角解并列、汽轮机甩60%负荷和甩100%负荷3种试验方案激发出机组轴系扭振,应用ICNN-1D进行轴系扭振模态参数辨识,结果表明ICNN-1D可准确辨识出机组在3种试验工况下轴系扭振固有频率和模态阻尼系数,证明了该方法的有效性。
An improved one-dimensional convolution neural network(ICNN-1 D)was proposed to identify the torsional vibration modal parameters of turbogenerator shafting,which included input layer,three groups of one-dimensional convolution layer and pooling layer,full connection layer and output layer.In order to overcome the difficulty in determining the modal parameters of network,the unsupervised learning network structure was built.Based on Adam optimization method,the dynamic learning rate optimization strategy was proposed to avoid the oscillation in learning process.Taking a600 MW turbogenerator set as the research object,torsional vibration of the shaft system was excited by three test schemes:generator 5-degree angle grid disconncetion and connection,steam turbine 60%load rejection and 100%load rejection.ICNN-1 D is applied to identify the modal parameters of the torsional vibration.The results show that ICNN-1 D can accurately identify the natural frequency and modal damping coefficient of torsional vibration under three test schemes,which proves the validity of the method.
作者
何成兵
王润泽
张霄翔
HE Chengbing;WANG Runze;ZHANG Xiaoxiang(School of Energy,Power and Mechanical Engineering,North China Electric Power University,Changping District,Beijing 102206,China)
出处
《中国电机工程学报》
EI
CSCD
北大核心
2020年第S01期195-203,共9页
Proceedings of the CSEE
基金
北京市自然科学基金项目(3132015)
关键词
一维卷积神经网络
无监督学习
汽轮发电机组
扭振
模态参数辨识
one-dimensional convolution neural network
unsupervised learning
turbogenerator set
torsional vibration,modal parameter identification
