基于层次式多子网神经网络的缺陷识别

Flaw Identification Based on Layered Multi-subnet Neural Networks

针对单一神经网络在涡流无损检测中存在识别精度低、训练时间长和识别范围小的缺点,提出了一种适用于实时在线检测的神经网络结构——层次式多子网神经网络。该网络包括一个总网和各层子网,可以将一个复杂的任务分成多个小任务去完成,能快速识别出缺陷有无、走向以及大小。由于每个网络采用改进的径向基神经网络优化隐含节点数,利用小波多尺度边缘检测方法提取输入信号的特征值以简化输入节点数,网络结构得到极大简化。结果表明,层次式多子网神经网络适用于实时在线检测。

Pointed to the disadvantages such as low recognizing precision, long training time and limited recognizing range of single neural network in eddy current testing, layered multi-subnet neural network was presented. It was composed by a sumner and several layered subnets, and could divide a complex task into a series of subtasks, so it could quickly identify whether the defect was existed, and also the defect location and dimension. Because of the improved RBF and wavelet multi-scaling edge detecting were used in each network, the network structure was simplified much. The result showed that layered multi-subnet neural network was suitable to online eddy current testing.