基于广义回归神经网络的旋转机械振动特征预测

Rotating Machinery Vibration Characteristics Forecasting Based on Generalized Regression Neural Network

提出了一种基于广义回归神经网络(GRNN)模型的旋转机械振动特征预测策略,给出了一种快速的GRNN模型平滑参数的优选方法;采用滑动窗口的方法更新训练样本,以便在每步预测之前获得能反应振动最新变化趋势的网络结构,进而提高预测精度。将该方法应用于某600MW核电机组动静碰摩故障下的振动特征预测,并与粒子群算法优化的支持向量回归模型(PSO-SVM)、径向基函数神经网络(RBFNN)模型的预测结果进行了对比分析,结果表明:提出的振动预测模型的总体性能优于PSO-SVM、RBFNN模型,在学习样本数目较少的情况下也能够得到较为满意的预测结果。

A strategy of rotating machinery vibration characteristics forecasting is proposed based on generalized regression neural network （GRNN） model. And the determination optimization of smoothing parameters are introduced. The method of sliding window is adopted to update the training samples, thus, the newest network structure can refelect the changing trend of vibration responses would be achieved before every predition step which can reprove the prediction accuracy. The method is applied to forcast vibration characteristics of the contact rubbing fauh of a 600MW nuclear power unit. Prediction results are compared with support vector regression optimized by particle swarm optimization algorithm （ PSO - SVR） model and radial basis function neural network （RBFNN） model, respectively. Results show that the proposed GRNN vibration prediction model performs better both than PSO - SVR model as well as RBFNN model and forcast result can meet the requirment even with a few number of learning samples.