建立点焊质量神经网络监测模型时作用函数的选取  

Selection of Action Function When Establishing the Neural Network Monitoring Model on Quality of Spot Welding

作　　者：张忠典[1] 李冬青[1] 赵洪运[2] 范伟光[2] 

机构地区：[1]哈尔滨工业大学现代焊接生产技术国家重点实验室,哈尔滨150001 [2]吉林工学院,长春130012

出　　处：《焊接学报》2002年第3期49-62,共14页Transactions of The China Welding Institution

摘　　要：多层前向神经网络是最常用、最流行的神经网络模型 ,其逼近能力和训练算法是其应用的关键。误差反传算法 (BP)以诸多的优点而成为多层前向神经网络训练的首选算法 ,但却存在收敛速度慢的缺点。研究发现 ,“假饱和”是导致BP算法收敛缓慢的主要原因之一 ,也是减小点焊质量监测模型误差的主要障碍。为了减少BP算法学习过程中出现“假饱和”的可能性、加快学习速度 。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.

关 键 词：点焊质量 监测模型 神经元网络 误差反传算法 作用函数 

分 类 号：TG453.9[金属学及工艺—焊接] TP183[自动化与计算机技术—控制理论与控制工程]