误差反传算法被广泛用于多层前馈神经网络的训练．但该算法的收敛性问题并没有解决，这导致训练后的网络泛化能力一般很差．本文研究了这一问题，并基于神经网络映射定理提出一种用于训练网络逼近单输出函数的隐节点校正（HNR）算法．这在神经辨识领域是有用的，因为大多数工业对象都是多输入单输出的．我们对HNR算法的收敛性和泛化能力作了理论上的研究．仿真实验和在催化重整过程建模中的应用实例表明该算法在一定条件下具有很高的学习速度和较强的泛化能力． Error back propagation algorithm is widely used in the training of multi-layer feedforward neural network. However, the convergence problem of the algorithm is not solved, which leads to poor generalization ability of the trained network. This paper studies the problem and proposes a hidden node correction (HNR) algorithm for training network approximation of single output function based on the neural network mapping theorem. This is useful in the field of neural identification because most industrial objects are multi-in, single-output. We theoretically study the convergence and generalization ability of HNR algorithm. The simulation experiment and the application in the process of catalytic reforming process show that the algorithm has high learning speed and strong generalization ability under certain conditions.