基于GRNN的霍尔式位移传感器特性曲线拟合研究

摘要

A new method based on GRNN (General Regression Neural Networks)is proposed to fit the input-output characteristic curve of hall displacement sensor,in order to reflect its input-output characteristics precisely and to provide a reliable basis for its nonlinear compensation.To prove the feasibility and effectiveness of this new method, LM (Levenberg-Marquart)algorithm is first employed to improve the traditional BPNN(Back Propagation Neural Networks),then the fitting result of GRNN is compared with that of the improved BPNN.The simulation results show that for small-sized or medium-sized networks, when the amounts of training samples are the same, the method using GRNN produces smaller errors,faster convergence speed and higher fitting precision compared with the one using BPNN.%为了精确反映霍尔式位移传感器的输入-输出特性，为其非线性补偿提供可靠依据，提出了利用广义回归神经网络（General Regression Neural Networks, GRNN）对霍尔式位移传感器的输入-输出特性曲线进行拟合的新方法。为了证明此种新方法的可行性和有效性，利用LM (Levenberg-Marquart)算法对传统反向传播神经网络（Back Propagation Neural Networks, BPNN）进行改进，并将GRNN和BPNN的拟合结果进行对比。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GRNN比采用BPNN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。