用于结构损伤识别的神经网络输入选取规则探究

Exploration of selection rules for inputs of neural network used on structural damage identification

将神经网络作为模式识别工具用于结构损伤位置识别时,其识别效果除了要受到网络隐层数目、各隐层神经元数目、神经元传递函数的形式、训练样本的数量与质量及训练方法的影响外,还会受网络输入性能的影响。在其他因素均相同的条件下,网络输入对网络性能起着决定性作用。为解决网络输入的选取问题,从网络功能、类别可分性和噪声的影响三个方面对网络输入的选取进行了分析研究,提出了用于结构损伤识别的神经网络输入选取的一般性规则,对采用神经网络处理模式识别问题具有参考价值。

When the diagnosis of damage location is taken as a problem of mode identification, neural network can be used as a pattern classifier. The identification effects will be affected by the number of the connotative layers and the nerve cells, the transfer function mode, the training samples, the training method and the performance of the inputs. Given the same conditions, the performance of the inputs is crucial to the performance of the neural network. To solve the problem of inputs sealecfion, three aspects ： the function of the neural network, the separability of the inputs and the effect of noise, are taken into consideration. The selection rules of neural network inputs that have been put forward in this article is instructive to those who will use neural network to solve the problem of mode identification.