模拟电路故障诊断的约束聚类集成BP网络算法  

An Constraint-based Clustering Integrated BP Networks Approach for Analog Circuit Fault Diagnose

作　　者：王月海[1] 蒋爱民[1] 程冉[1] 王彤威[2] 

机构地区：[1]北方工业大学信息工程学院,北京100144 [2]北京航天测控技术开发公司,北京100037

出　　处：《计算机测量与控制》2011年第7期1592-1595,共4页Computer Measurement &Control

摘　　要：针对通用BP网络对于高纬度、大数据量训练收敛困难的问题,在使用动量因子、自适应调整学习速率等方法的基础,引入约束聚类,构造集成神经网络,以提高网络的训练速度及诊断效果;首先,采用约束聚类算法将训练样本集划分为若干个规模相当的子样本集,分别训练生成相应子网络;此外,在诊断过程中除各子网络的输出变量外,还加入了诊断数据相对各子训练样本集的隶属度因子;最后通过一个实际电路板25维采样数据、38类故障的BP网络诊断实例验证了算法的可行性。The training algorithm for BP network is hard to converge when the input data is high dimension and in large quantity.For this problem,with momentum and adaptive learning rate,a new integrated BP neural network based on constraint-based clustering was proposed to fasten the convergence of training process.Using the clustering algorithm,the training samples are firstly divided into several sample sub-sets with similar size and trained respectively.Moreover,the corresponding output variables as well as the diagnostic data＇s membership factors relative to one sub-training set are both taken into account during diagnosing.This approach was verified by an demo circuit board whose sample data are 25 dimension and faults categories are 38.

关 键 词：约束聚类算法 集成BP网络 模拟电路故障诊断 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]