基于上下文生成对抗网络的时间序列异常检测方法

时间序列的异常检测是网络服务保障、数据安全检测、系统监控分析等应用中所依赖的一项关键技术。为解决在实际场景的时间序列异常检测中由于时间序列上下文的模糊性、数据分布的复杂性以及异常检测模型的不确定性所带来的异常检测结果的有效性、合理性、稳定性等不足的问题，本文提出了一种新的基于上下文生成对抗网络的时间序列异常检测方法AdcGAN。首先，通过处理历史数据，提取用于生成时序数据的条件上下文；然后，采用条件生成对抗网络的设计策略，使用条件上下文，构建上下文生成对抗网络，实现对任意时刻数据的条件分布预测，同时AdcGAN采用Dropout近似模型不确定性，使用概率分布代替点估计作为预测结果；接着，从观测的差异(用期望偏差表示)和模型的不确定性(用预测方差表示)两个方面来衡量异常；最后，提出基于数据统计信息的异常阈值自动设置方法，减少手动调节的参数量。实验结果表明，与同类基准算法进行对比，在NAB数据集中的47个真实时序数据上，本文提出的AdcGAN可以有效地检测出时序数据中的异常，在大多数评价指标上都优于其他基准方法，并且具有更好的稳定性。

A time series anomaly detection method based on contextual generative adversarial network

Time series anomaly detection is a key technology relied upon in applications such as network service, data security, and system monitoring. In order to address the limitations of effectiveness, rationality and stability in anomaly detection results caused by the fuzziness of time series context, complexity of data distribution, and the uncertainty of anomaly detection models in practical scenarios, this paper proposes a new anomaly detection method (AdcGAN), based on contextual generative adversarial network. Firstly, AdcGAN extracts conditional context for generating time series data by processing historical data. Secondly, AdcGAN constructs a context generative adversarial network following conditional generative adversarial network strategy to achieve conditional distribution prediction of the data at any moment. Meanwhile, AdcGAN uses Dropout to approximate model uncertainty and replacing point estimates with probability distribution as prediction result. Then, anomalies are measured based on the differences in observations (represented by the expected deviations) and the uncertainty of the model (represented by prediction variances). Finally, an automatic method for setting anomaly thresholds based on statistical information of the time series data is proposed to reduce the number of manually adjusted parameters. Our experimental results on 47 real-time series data of the NAB dataset compared with baselines show that, compared to similar benchmark algorithms, the proposed AdcGAN method can effectively detect anomalies in time series data. It outperforms other benchmark methods in most evaluation metrics and achieves better stability.