Abstract: Tool anomaly detection is the foundation and key to analyzing and judging tool health status.Aiming at the problem of low model detection accuracy caused by difficulty in distinguishing vibration signal status information,asynchronous spatiotemporal feature extraction,and insufficient consideration of the potential variables distribution in the latent space during tool anomaly detection,a method for unsupervised tool anomaly detection based on spatiotemporal feature extraction was proposed.An independent preprocessing method was adopted for each axial vibration signal to map them to the same range,which eliminated the impact of inconsistent signal fluctuation range.Then,the Temporal Convolutional Network(TCN)was embedded into Variational Autoencoder(VAE)to achieve synchronous extraction of spatiotemporal features and improve the learning ability of the model.In addition,the raw data was mapped to the latent space through nonlinear mapping,thereby learning the latent variables under each axial input and aligning them as closely as possible with the Gaussian distribution.Based on the publicly available tool wear dataset PHM 2010,the effectiveness of the proposed method was verified,and the results showed that the proposed method had high detection accuracy and better than other anomaly detection methods.

Key words: tools, anomaly detection, variational autoencoder, time convolutional network, unsupervised learning

摘要： 刀具异常检测是分析和判断刀具健康状态的基础和关键。针对刀具在异常检测过程中,振动信号状态信息难以辨别、时空特征提取不同步以及潜在空间中潜在变量的分布考虑尚不充分,导致模型检测精度低的问题,提出一种基于时空特征提取的刀具无监督异常检测方法。首先,对各轴向振动信号采用独立式预处理方法,将其映射到同一范围,消除信号波动范围不一致带来的影响。然后,将时间卷积网络(TCN)嵌入变分自编码器(VAE)中,实现数据时空特征的同步提取,提高模型的学习能力;同时,通过非线性映射将原始数据映射到潜在空间,从而学习到各轴向输入的潜在变量,并使其尽可能对齐高斯分布。最后,利用公开刀具磨损数据集PHM 2010验证了所提方法的有效性,结果表明,所提方法具有较高的检测精度,且性能优于其他异常检测方法。

关键词: 刀具, 异常检测, 变分自码器, 时间卷积网络, 无监督学习