Deep Hybrid Multivariate Time Series Anomaly Detection

• 1、School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044

Abstract：Time series anomaly detection is a hot-topic in time series data mining, and it occupies an important position in practical applications, such as network security, health care, and risk management. The time series in anomaly detection tasks are often multivariate high-dimensional data, and the number of anomalous samples is relatively less. Methods based on statistics and traditional machine learning cannot effectively deal with the problem of class imbalances, and they ignore the order of the data and local information, leading to misjudgment and omission of anomaly. This paper proposes a method for time series anomaly detection using a combination of Deep AutoEncoder (DAE) and traditional classifiers (such as support vector machines). The DAE is built by Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM).Thanks to the ability of CNN and LSTM to effectively capture the local spatial and temporal features,the correlation of the original data in the hidden layer subspace has significantly weakened the relationship between normal samples and abnormal samples. On this basis, traditional classifiers are easier to accurately distinguish abnormalities, and decrease the uncertain of detection schemes such as threshold screening. Experiments show that compared with various baselines, the proposed method shows better results on multiple datasets.

Keywords： Time Series Anomaly Detection LSTM CNN AotuEncoder