Cellular KPI Anomaly Detection with GAN and Time Series Decomposition

Abstract

Detecting anomalies in key performance indicator (KPI) is crucial for cellular network management. With the 5G deployment, the network is becoming much more complicated, making manual anomaly detection tedious or even impossible. In this paper, we propose LSTM-GAN-G for anomaly detection in cellular KPI time series. We build a deep generative adversarial network (GAN) with long short term memory (LSTM) cells to better capture temporal characteristics of the input data. The LSTM-GAN is trained with purely normal data so that the trained generator learns the distribution of the normal cellular KPI time series. During anomaly detection, the trained generator (G) is exploited to detect anomalies in testing samples by computing the difference between the testing sample and the generator reconstructed samples. Additionally, we propose to decompose the input time series and feed the de-seasoned data to the LSTM-GAN to further improve the anomaly detection performance. The proposed algorithm is evaluated on a real-world cellular KPI dataset. Our results show that the proposed method is able to detect both point anomaly and segment anomaly accurately, and significantly outperforms benchmark algorithms.