Abstract: To solve the fault data restriction problem in the ball screw condition monitoring,a ball screw anomaly monitoring method was proposed based on the Convolutional Variational Autoencoder (CNVAE) and the dynamic kernel density estimation model.The one-dimensional convolutional neural network was used to construct the variational autoencoder model,and it was trained by the normal vibration data of the ball screw collected in the early stage.Hence,a CNVAE model capable of reconstructing the normal samples could be learned.Then,the real-time signals were input into the CNVAE model,and the reconstruction error could be obtained,which was used as a health indicator to represent the degradation state of the ball screw.With a sliding window on the time scale,the reconstruction errors in different time periods were input into the kernel density estimation model,and the ball screw anomaly could be automatically detected through observing the probability distribution variation of the reconstruction errors in the sliding window.Experimental results showed that the proposed method could recognize the ball screw degradation stage effectively,characterize the deterioration evolution process and detect the abnormality earlier than traditional methods.

Key words: anomaly monitoring, convolutional variational autoencoder, kernel density estimation, ball screw, degradation stage identification

摘要： 为解决滚珠丝杠副状态监测中面临的异常状态数据缺少问题,基于卷积变分自编码器(CNVAE)和动态核密度估计模型,提出一种无需故障数据的滚珠丝杠副异常监测方法。首先,采用一维卷积神经网络构建变分自编码器,以早期正常阶段采集的信号作为输入,训练得到能够对正常数据进行重构的CNVAE模型。然后,将实时信号输入CNVAE模型中得到重构误差,作为表征滚珠丝杠副退化状态的健康指标。最后,采用一个在时间尺度上滑动的窗口选取不同时间段内的重构误差构成时间序列,输入核密度估计模型中,通过观测滑动窗口内重构误差的概率分布变化自动判定滚珠丝杠副是否出现异常。实验结果表明,提出方法能够区分滚珠丝杠副的不同退化阶段,表征滚珠丝杠副的退化演变过程,相比于传统方法能够更早地检测到滚珠丝杠副的异常。

关键词: 异常监测, 卷积变分自编码器, 核密度估计, 滚珠丝杠副, 退化阶段识别