LMBP神经网络改进算法的研究  

Research on improved algorithm of LMBP neural network

作　　者：张妮[1] 田学民[1] 

机构地区：[1]中国石油大学信息与控制工程学院,山东东营257061

出　　处：《计算机与应用化学》2010年第8期1111-1114,共4页Computers and Applied Chemistry

摘　　要：在LMBP算法训练过程中,大型矩阵的求逆运算限制算法的收敛速度,本文针对这一特点,在训练网络的权值和偏移值时采用求解大规模线性方程组的超记忆梯度算法,避免矩阵求逆耗时的缺点,同时对原有的步长因子进行自适应改变,并通过网络修剪对隐层神经元结构进行优化。最后以某型号设备齿轮箱为例进行仿真。结果表明,本文的改进算法能够明显缩短训练时间,并且经过此算法训练的网络有较高的故障诊断性能。During training the BP neural network with LMBP,getting the large inverse matrix restricted the convergence speed.The paper directed against it and introduced super memory gradient algorithm to solve large linear equations during calculating the weights and offsets of the network.It adopted adaptive change to the original step and optimized the hidden neurons by net pruning in the network at the same time.In the end,the algorithm was applied into simulation of one device＇s gear box.The result showed that the improved algorithm could shorten training time remarkable.The net trained could also get the relatively high performance to diagnosis faults.

关 键 词：LMBP算法 超记忆梯度算法 网络修剪 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]