KPI anomaly detection method for Data Center AIOps based on GRU-GAN
Hang Su, Qian He, Biao Guo · 2021
The system architecture and application services of the data center are becoming increasingly large. To ensure the stable operation of the systems and businesses carried by the data center, the operations engineer needs to collect and monitor the generating KPIs during the operation of the systems and services. Traditional KPI anomaly detection methods are faced with the challenges of the huge amount of KPIs and constantly changing data characteristics, which are gradually no longer suitable for highly dynamic systems and services. With the popularity of artificial intelligence algorithms, machine learning and deep learning methods have also begun to be applied in operation and maintenance scenarios, that is the emergence of Artificial Intelligence for IT Operations (AIOps). KPI anomaly detection is the underlying core technology of AIOps. This paper proposes a hybrid model based on GRU-GAN (GGAN) for KPI anomaly detection in data center AIOps. The Gated Recurrent Unit (GRU) network is selected as the generator and discriminator of Generative adversarial network (GAN) in this model, which get the time correlation and data distribution of KPI through the adversarial training between the generator and the discriminator to make use of the reconstruction ability of the generator and the discriminant ability of the discriminator at the same time. At the anomaly detection stage, the anomaly score is formed by integrating reconstruction difference and discrimination loss to complete the anomaly detection task. Experimental results show that the proposed method can more accurately capture the variable data characteristics of KPI compared with the traditional KPI anomaly detection method and the general unsupervised method, as well as achieve better performance in the KPI anomaly detection task.