一种新的高阶前馈神经网络及其在旋转机械故障诊断中的应用

剖析了基于 BP神经网络和径向基函数网络的故障诊断模型的诊断性能和应用中的局限性 ,针对这些诊断模型的局限性 ,提出了基于椭球单元 (Ellipsoid Unit)高阶网络的诊断模型 ,并对网络训练算法进行了研究 ,提出了基于模糊聚类算法的网络权重初始化方法和网络动态训练策略 ,有效地改善了网络的学习性能和诊断性能 ;最后对该网络在旋转机械故障诊断中的应用进行了研究。结果表明 :比之经典前馈网络 ,椭球单元网络在故障分类方面因其能形成封闭有界的决策区域而具有明显的聚类的优越性和分类的合理性 ,很适合故障诊断领域的分类问题

New higher order neural networks and their application in fault diagnosis for rotating machinery

To overcome the limitations of standard feedforward neural networks, a new type of higher order neural networks (i.e ellipsoidal unit networks) has been proposed recently, which is very useful for fault diagnosis applications because of its bounded generalization and extrapolation. This paper describes the theory and structure of such networks with respect to two problems arising in training processes, a method for initializing hyperellipsoids based on the fuzzy cluster algorithm and a dynamic training strategy. A case study is given for fault diagnosis for rotating machine. The research results show that, compared with standard feedforward neural networks, the ellipsoidal unit network is more reasonable and useful for fault diagnosis applications.