KPI Anomaly Detection Based on LSTM with Phase Space

Abstract

Anomaly detection based on Key Performance Indicator (KPI) plays a crucial role in the cloud application monitoring platform. The methods for anomaly detection based on thresholds and rules are commonly used to monitor KPI. The key factor affecting the accuracy of these methods is the fitting accuracy of normal patterns. The Long Short-Term Memory (LSTM) Network can learn normal patterns and predict future values. However, according to the low-dimensional and non-stationary characteristics of KPI data, the prediction model merely on the basis of time dimension is limited. More specifically, the error occurs in the amplitude dimension and the time lag is present in the trend dimension, which further interfere with the accuracy of anomaly detection. In view of the problem mentioned above, we propose a model of anomaly detection based on LSTM with phase space (PSR-LSTM-AD). With the theory of phase space reconstruction, the motion trajectory of time series is recreated in the high-dimensional phase space, and the hidden spatial features of the original low-dimensional sequence are extracted. The model can extract features in both time and space dimensions. The accuracy of normal pattern prediction is improved in both amplitude and trend dimensions, thus enhancing the accuracy of anomaly detection. This model plays an effective role in the KPI dataset collected from the real environment, with average F1-score over 0.79. Besides, compared with another related model (KPI-TSAD) on this dataset, the proposed model works better, with average accuracy increasing by 7% and average F1-score increasing by 7%.