多层感知器 (multi-layerperceptronnetworks ,MLPN)是一具有多层神经元、前馈、误差反传结构的神经网络 ,它的学习和预测能力受多方面因素的影响。首先我们从理论证明和数值分析的角度研究了传输函数、神经元的数目、网络层数及网络误差的迭代方式等与MLPN学习和预测能力的关系 ,对常规的MLPN作了改进 ;然后结合一个理论模型分析的例子 ,讨论了改进的MLPN对非线性函数的学习能力 ;最后 ,以某地野外磁测数据的去噪为实例 ,将本文介绍的神经网络技术用于插值 ,从而达到去噪的目的。从理论和应用两方面证明了以随机全局优化方法更新网络权值的改进MLPN具有良好的学习和预测能力 Multi-layerperceptron networks (MLPN) is a neural network with multi-layer neurons, feedforward and error feedback structure. Its learning and predictive ability is affected by many factors. First of all, we study the relationship between MLPN learning and predicting ability, such as the transfer function, the number of neurons, the number of neurons, the number of network layers and the network error from the perspective of theoretical proof and numerical analysis, and then improve the conventional MLPN. Then, Theoretical model analysis, discusses the improved MLPN learning ability for nonlinear functions; Finally, taking the de-noise of some field magnetic data as an example, the neural network technology introduced in this paper is used to interpolate, so as to achieve the denoising purpose. It is proved theoretically and practically that the improved MLPN, which updates the network weights by stochastic global optimization, has good learning and predictive ability