一种采用串行自编码器的时序数据异常检测方法

基于深度学习的时序数据异常检测模型大多采用循环神经网络或长短期记忆网络捕捉时序依赖性,并利用自编码器重构数据,进而实现时序数据的异常检测.虽然此类检测模型实现了较高的异常检测率,但它们的网络结构复杂,导致模型的计算效率较低.为提高模型的计算效率,提出一种基于串行自编码器的异常检测模型SAE-AD.该模型仅包含两个结构简单的自编码器(AE_1$和AE_2$),其所含参数较少,且训练目标较为简单,从而加快了模型的计算效率.通过将自编码器AE_1$和AE_2$串行拼接,即AE_1$的输出作为AE_2$的输入,可有效提高AE_2$的解码器对正常数据特征的解码能力,有助于提升模型的检测准确率.实验结果表明,相较于其他新近提出的异常检测模型,SAE-AD模型具有更高的精确率、召回率和$F_1$值.

A serial autoencoders based method for detecting time series anomalies

Aiming to detect time series anomalies, deep learning methods generally use the recurrent neural network or long short term memory to capture temporal dependency, and adopt autoencoder to reconstruct data. Although they work well for detecting anomalies, the network structures of these methods are complex, resulting in slow computational efficiency. In order to improve the computational efficiency, this paper proposes a method called serial autoencoders based anomaly detection (SAE-AD) which contains two autoencoders (AE_1$ and AE_2$) with simple structure. Due to the simplicity, there are a few training parameters and its training objectiv function is relatively simple, which speeds up the computation. In addition, the output of AE_1$ is fed into AE_2$ to improve the decoding ability of the decoder of AE_2$. This way of serial training makes SAE-AD achieve better detection accuracy. Experiment results show that the proposed method has better precision, recall, $F_1$ score than several state-of-art anomaly detection methods.