减小点焊质量神经网络监测模型误差的措施  

Measures for decreasing errors of ANN models of spot welding quality monitor

作　　者：张忠典[1] 李冬青[1] 唐大平 李学军 

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]哈尔滨建成专用车有限公司,哈尔滨150030

出　　处：《焊接学报》2004年第5期113-116,共4页Transactions of The China Welding Institution

摘　　要：点焊过程监测信息与质量参数之间关系也含有较大的非线性 ,用线性模型去描述这样的关系将导致模型误差的增加。为了更好地描述点焊过程监测信息与质量参数之间的复杂关系 ,文中将神经网络理论用于点焊过程模型化。在建立点焊质量神经网络监测模型的过程中 ,发现 ,训练过程中的”假饱和”现象是减小网络模型误差的主要障碍。为此 ,分析了各种减小网络模型误差的可能途径 ,提出了相应的改善措施 ,并通过试验证明 ,所提出的观点是正确的。The relations between monitoring information and spot welding quality factors are substantially nonlinear, and it will make model error increasing to describe such relations by linear model. To describe the complex relations more reasonably, the artificial neural network (ANN) theory is applied to spot welding process modeling. During process modeling, false saturation phenomenon is a main obstacle to decrease errors of models. By analyzing this phenomenon, many valid measures to decrease the likelihood of false saturation are taken. The test results show that the measures to decrease errors of models are feasible and effective.

关 键 词：假饱和 神经元网络 误差反传算法 点焊质量监测 

分 类 号：TG453.9[金属学及工艺—焊接]