MultiSR: A SR-Based Framework for Multivariate Time Series Anomaly Detection in Complex Unmanned Scenarios

Abstract

Nowadays, more and more unmanned systems are adopting time series anomaly detection algorithms as the primary tool for KPI monitoring. However, many anomalous data in real environments often have complex features such as large data fluctuations, irregular data fluctuations and large data drift. Traditional univariate or multivariate anomaly detection algorithms can hardly meet the demand for anomaly detection of time series with many features and large fluctuations. Therefore, we propose a multivariate anomaly detection framework adapted to complex environments, which can effectively reduce the complexity of time series data and improve the overall performance of anomaly detection algorithms in complex real systems. First, data preprocessing is performed on the time series before anomaly detection to reduce the impact of temporal bias of different data items in multivariate time series data. Second, to intuitively analyze the anomalous fluctuation characteristics of the time series, we use Spectral Residual to process the significance of the time series. Then, we combine the Spectral Residual processed data and Logistic Regression algorithm to construct anomaly classifiers for anomaly detection of real-time data in complex real systems. Finally, by experimenting with a prototype system on a private cloud platform, our approach obtains superior results under multiple anomaly patterns compared with traditional univariate and multivariate anomaly detection algorithms.