摘要
时间序列数据的有效异常检测对现代工业应用非常重要。然而,由于缺乏异常标签、数据高波动性、训练不稳定,导致建立一个能够准确地进行异常检测的系统是一个具有挑战性的问题。尽管异常检测的深度学习方法最近有所发展,但其中只有少数能够应对所有这些挑战。本文提出了CAE_AD,这是一种基于卷积自编码器(CAE)的无监督异常检测模型。为了尽量地放大异常,避免错过异常,笔者引入了两阶段的对抗训练。同时,为了提高训练稳定性,笔者引入了第一阶段的重建误差以作为第二阶段卷积自编码器的输入。笔者将CAE_AD与先进的时间序列异常检测方法在多个数据集上进行了比较。实验结果表明,本文提出的模型性能优于这些对比方法。在SMAP数据集上,相比于其他模型,CAE_AD模型的f1领先了4%,Precision领先了8%。
Effective anomaly detection of time series data is crucial for modern industrial applications. How-ever, due to the lack of anomaly labels, high data volatility, and unstable training, establishing a system that can accurately detect anomalies is a challenging problem. Although deep learning methods for anomaly detection have recently developed, only a few of them can address all of these challenges. This article proposes CAE_AD, which is an unsupervised anomaly detection model based on convolutional autoencoder (CAE). In order to maximize the amplification of anomalies and avoid missing them, the author introduced two-stage adversarial training. Meanwhile, in order to improve training stability, the author introduced the reconstruction error from the first stage as the input for the convolutional autoencoder in the second stage. The author compared CAE_AD with advanced time series anomaly detection methods on multiple data sets. The experimental results show that the model proposed in this article performs better than these comparison methods. On the SMAP dataset, compared to other models, the CAE_AD model has a 4% lead in f1 and an 8% lead in Precision.
出处
《图像与信号处理》
2024年第1期21-32,共12页
Journal of Image and Signal Processing