A novel unsupervised framework for time series data anomaly detection via spectrum decomposition

Time series is a common type of data that widely exists in various real-world scenarios, such as traffic flow data, network KPI, financial data, which can be regarded as time series. Anomaly detection in time series is an interesting research topic with a wide range of real-world applications, such as network intrusion detection, traffic situation monitoring and sensor error detection. In the real-world scenario, the frequency of anomalies is very low and little anomalous sample is available for analysis. Therefore, unsupervised methods are usually used for anomaly detection. In this paper, based on spectrum analysis and time series decomposition, an unsupervised deep framework for anomaly detection in time series data is designed. First, we decompose the original time series into trend series, seasonal series and residual series based on spectrum analysis. Then, prediction models based on long short-term memory (LSTM) networks and convolutional neural networks (CNNs) are designed to predict the trend series and seasonal series, respectively, and the residual series is reconstructed based on a Gaussian distribution. Next, the time series data are reconstructed by superimposing the predicted trend series and seasonal series and reconstructed residual series. Finally, we compare the original time series with the reconstructed time series and detect anomalies according to a certain threshold. The method proposed in this paper integrates prediction-based and reconstruction-based methods, and the experimental results on four datasets demonstrate the excellent performance of our method.

Automated summary

This paper proposes an unsupervised deep framework for anomaly detection in time series data based on spectrum analysis and time series decomposition. It decomposes the time series into trend, seasonal, and residual series, and detects anomalies through prediction models and reconstruction.