With the rapid increase of edge network scale and the complexity of service interaction, it takes more time for operation staff to analyze anomalies from complex scenarios. To maintain the normal network operation, various key performance indicators, such as link delay, throughput, and memory usage, are monitored for timely anomaly detection and troubleshooting. We introduce artificial intelligence for IT operations to assist operators in performing anomaly detection, anomaly localization, and root cause analysis, and building an intelligent operation and maintenance platform over the software-defined networking (SON)-based edge network. In this article, the graph-based gated convolutional network for anomaly detection (GAD) is first proposed to solve the anomaly detection problem of time series data with topology information. Specifically, GAD uses a gated convolutional encoder to encode spatial-temporal time series, and a graph convolutional network is developed to capture the spatial dependence. Then, based on the features, a convolutional layer is used to decode features and reconstruct input sequence. Finally, the residual between input and reconstructed sequences is further utilized to detect anomalies. Our experimental results demonstrate that GAD outperforms the state-of-the-art anomaly detection baselines in terms of F-scores on the datasets collected by an SDN simulation platform.

中文翻译：

---

基于时空学习的人工智能在边缘网络中的IT运营

随着边缘网络规模的迅速增加和服务交互的复杂性，运维人员需要花费更多时间来分析复杂场景下的异常情况。为了维持正常的网络运行，将监控各种关键性能指标，例如链路延迟，吞吐量和内存使用情况，以便及时进行异常检测和故障排除。我们为IT运营引入人工智能，以帮助运营商执行异常检测，异常定位和根本原因分析，并在基于软件定义网络（SON）的边缘网络上构建智能的运维平台。本文首先提出了基于图的门控卷积异常检测网络（GAD），以解决具有拓扑信息的时间序列数据的异常检测问题。具体来说，GAD使用门控卷积编码器对时空时间序列进行编码，并开发了图卷积网络来捕获空间相关性。然后，基于特征，使用卷积层对特征进行解码并重建输入序列。最后，输入序列和重构序列之间的残差还被用来检测异常。我们的实验结果表明，在SDN模拟平台收集的数据集上，GAD的F评分优于最新的异常检测基线。输入序列和重构序列之间的残差进一步用于检测异常。我们的实验结果表明，在SDN模拟平台收集的数据集上，GAD的F评分优于最新的异常检测基线。输入序列和重构序列之间的残差进一步用于检测异常。我们的实验结果表明，在SDN模拟平台收集的数据集上，GAD的F评分优于最新的异常检测基线。