节点自删除神经网络及其在磨粒识别中的应用

Nodes self deleting neural network and its application in wear particles recognition

在前向多层神经网络的基础上，提出了一种新的节点自删除神经网络模型。该神经网络根据隐层节点输出的相似性能够自动地进行网络节点的删除。对网络中冗余节点的删除，使网络尺寸得到优化，这一定程度上也解决了前向多层神经网络的推广性和学习问题。在Ｐａｒｔｙ－４问题的实例中，通过对惯性ＢＰ算法网络和该节点自删除网络的比较，充分说明了该节点自删除神经网络在各方面的优越性。铁谱磨粒识别一直是一个困难的课题，应用自删除神经网络模型在该实例中，也取得了较好的识别效果。

This paper provides a new neural network model on the base of the forward multilay neural network. This network can itself deletes the nodes according to the similarity of hidden lay nodes output to optimize its construction. This constructive optimization also solves the questions on generalization and learning of the neural network on some degree. The nodes self deleting network has the better properties than the BP network when they are applied in the Party 4 problem. Some good results can be given when the model is used in wear particles recognition which is difficult in the ferrograph domain.