Deep Learning Based Anomaly Detection for Muti-dimensional Time Series: A Survey
Zhipeng Chen, Zhang Peng, Xueqiang Zou, Haoqi Sun
Abstract
Abstract Multi-dimensional time series are multiple sets of variables collected in chronological order, which are the results of observing a certain potential process according to a given sampling rate. It also has the ability to describe space and time and is widely used in many fields such as system state anomaly detection. However, multi-dimensional time series have problems such as dimensional explosion and data sparseness, as well as complex pattern features such as periods and trends. Such characteristics lead to rule-based anomaly detection methods suffer from poor detection effects. In the big data scenario, deep learning method begins to be applied to anomaly detection tasks for multi-dimensional time series due to its wide coverage and strong learning ability. This work first summarizes the definitions of anomaly detection for multi-dimensional time series and the challenges it faces. Related methods are sorted out, and then the deep learning-based method is emphasized. The existing work and its advantages and disadvantages are summarized. Finally, the shortcomings of the existing algorithms are clarified and the future research direction is explored.