用于结构损伤位置识别的两种神经网络性能比较

为了克服用BP网络进行结构损伤位置识别时网络结构确定难、网络训练易陷入局部极小、训练时间长以及处理带噪声数据需要大量的误码练样本等问题，提出用SOFM网络进行结构损伤位置识别的方法。分别用SOFM网络和BP网络对一桁架结构进行损伤位置识别，通过比较两种网络的性能发现SOFM网络不但网络结构容易确定，网络训练不存在陷入局部极小的问题，BP网络只有在大量训练样本条件下才能保证网络具有较好的抗噪声能力，若训练样本不足，则BP网络的抗噪声性能较差，而SOFM网络在较少训练样本情况下即可具有良好的抗噪声性能，因而SOFM网络更适合训练样本有限备件下的结构损伤位置识别。

Performance Comparison of Two Types of Neural Networks Used in Structural Damage Location Identification

BP neural networks used in the structural damage location detection have many shortcomings,such as difficult network structure deciding,training time consuming and easily trapped into local minimum,and a lot of training samples needed to identify the noise contaminated data.To overcome these shortcomings,a method using SOFM neural network to identify the damage location is put forward.The two types of networks are used to identify the damage location of a truss structure.The comparison of the performance between the two types of neural network shows that the structure of the SOFM neural network is easier to be decided and its training won't be trapped into local minimum,it has better antinoise capability even the training samples are limited,while the BP network can have good antinoise capability only when there are plenty of training samples,otherwise the BP network will have poor antinoise capability.It is concluded that the SOFM neural network is better than the BP network to identify the damage location when the training samples are limited.