Abstract: This paper proposes a novel autoencoder deep neural network with memory unit for abnormal identification of installation status of key components under a train. This method does not require labeled fault samples, and by minimizing the reconstruction input, the network takes the reconstruction error as a factor for abnormal identification. It has been verified that the detection rate of the overall abnormal identification of key components can reach more than 92%.

Key words: autoencoder, deep neural network, abnormal identification

摘要： 提出一种新颖的带有记忆单元的自编码器深度神经网络，用于列车车底关键部件安装状态的异常识别，此方法不需要带标签的故障样本，网络通过最小化重构输入，把重构误差作为异常判别的因子，经过验证表明，关键部件整体异常识别的检出率可以达到92%以上。

关键词: 自编码器, 深度神经网络, 异常识别