Detecting Anomalies in Time Series Data via a Meta-Feature Based Approach

Anomaly detection of time series is an important topic that has been widely studied in many application areas. A number of computational methods were developed for this task in the past few years. However, the existing approaches still have many drawbacks when they were applied to specific questions. In this paper, we proposed a meta-feature-based anomaly detection approach (MFAD) to identify the abnormal states of a univariate or multivariate time series based on local dynamics. Differing from the traditional strategies of “sliding window” in anomaly detection, our method first defined six meta-features to statistically describe the local dynamics of a 1-D sequence with arbitrary length. Second, multivariate time series was converted to a new 1-D sequence, so that each of its segmented subsequence was represented as one sample with six meta-features. Finally, the anomaly detection of univariate/multivariate time series was implemented by identifying the outliers from the samples in a 6-D transformed space. In order to validate the effectiveness of MFAD, we applied our method on various univariate and multivariate time series datasets, including six well-known standard datasets (e.g. ECG and Air Quality) and eight real-world datasets in shield tunneling construction. The simulation results show that the proposed method MFAD not only identifies the local abnormal states in the original time series but also drastically reduces the computational complexity. In summary, the proposed method effectively identified the abnormal states of dynamical parameters in various application fields.