Abstract: Multilayer feedforward neural network is the most popular network model,and the approach ability and training algorithm are the key of its application. The Back-propagation algorithm is the first choice algo rithm to multilayer feedforward neural network because of its many merits, but its convergence rate is slow, which is its shortcoming. This article found that "false saturation" is one of the main reasons which causes BP algorith's convergence slow and is also the main obstruction to decrease the error of quality monitoring model for spot welding. In order to reduce the appearance possibility of "false saturation" in the learning process of BP algorithm and accelerate learning rate, this article puts forward the principle about selecting the type of neural action function.