通过对故障特征的深入讨论，建立了故障特征在线提取系统，然后利用这些特征样本对前馈式神经网络（ＢＰ）进行训练。为了提高网络对故障的联想能力，减少网络的动态学习时间，提出了两阶段网络。采用工厂实际数据进行计算，并利用全息谱和Ｗｉｇｎｅｒ分布验算了方法的正确性。从验算结果来看，网络的分类效果是令人满意的。 Through in-depth discussion of fault features, an online feature extraction system for fault features is established, and then the feedforward neural network (BP) is trained by these feature samples. In order to improve the ability of the network to associate with the fault and reduce the dynamic learning time of the network, a two-stage network is proposed. The actual data of the factory is used for calculation, and the correctness of the method is verified by using the holographic spectrum and the Wigner distribution. From the checking results, the classification effect of the network is satisfactory.