LMBP和RBF在ECS特性曲线拟合中对比研究

Comparative Study on Application of LMBP and RBF Neural Networks in ECS Characteristic Curve Fitting

为精确反映数字式涡流传感器的输入输出特性,为其非线性补偿提供可靠依据,对传统BP(BackPropagation)神经网络进行改进,利用LMBP(Levenberg-Marquart Back Propagation)神经网络和RBF(Radial BasisFunction)神经网络对涡流传感器的输入输出特性曲线进行拟合,并将两者拟合结果进行对比研究。仿真结果表明,在训练样本数量相等且中小规模网络的条件下,采用RBF神经网络比采用LMBP神经网络进行曲线拟合的误差更小、收敛速度更快且具有更高的拟合精度,为工程实际中一维数据的拟合方法选择提供了依据。

In order to accurately reflect the digital input and output characteristics of eddy current sensors and to improve traditional BP neural networks, LMBP （Levenberg Marquart Back Propagation） neural networks and RBF （Radial Basis Function） neural networks are first constructed. Then the two types of neural networks are applied respectively to the characteristic curve fitting of ECS （Eddy Current Sensors）. Finally a comparison is made to compare the fitting results of the two networks. The simulation results show that with the same number of training samples, the networks are small or medium sized, compared with LMBP, RBF neural networks are superior in fitting error, convergence speed and fitting precision. And this provides a basis for the choice of fitting method of one-dimensional data in practical engineering