| 摘 要: 局部加权回归是非参数学习方法，可自动规避在数据拟合过程中异常值对近邻点的影响。通过使用基于 局部加权回归的时间序列分解方法，我们对时间序列进行特征分解，将原始时间序列分解为趋势项、周期项和残差项； 在给出合理的检出水平阈值后，我们使用改进的格雷布斯检验法在得到残差项后对残差项进行异常值检测。经过实验证 明，该方法相比传统的时间序列方法三次指数平滑法可减低离群值对模型拟合的影响，更能有效适应数据中潜存的趋势 项的复杂变换，从而更加精准地找到数据中的真实异常点。 |
| 关键词: 局部加权回归；时间序列分解；假设检验；异常检测 |
|
中图分类号: TP309
文献标识码: A
|
| 基金项目: 本文系基于人工智能技术的信息系统状态诊断与辅助管理决策，中国铁路上海局集团有限公司2019年重点科研课题，课题编号2019036. |
|
| Research on Time Series Anomaly Detection Based on Locally Weighted Regression |
|
JIANG Xinle,LONG Jun,CHEN Gang,XIA Lei,LIANG Duozi,LIU Limin,FAN Huilong1,2
|
1.( 1.Information Technology Department, China Railway Shanghai Group Co., Ltd., Shanghai 200071, China;2.
2.School of Computer Science and Engineering, Central South University, Changsha 410083, China)
|
| Abstract: Locally Weighted Regression is a non-parametric method,which can automatically avoid the influence of anomaly values against adjacent data in the process of data fitting.By means of time series decomposition method based on Locally Weighted Regression,the original time series can be decomposed into trend,seasonality and residual.After reasonable threshold of detection level is given,we implement anomaly detection for the residual gained by the improved Grubbs testing method.Experiments show that compared with traditional time series method-Holt Winters,this method can reduce the effect of outliers during model fitting process,and adapt the complex variation of the trend in the data more effectively so as to find the real anomaly in the data accurately |
| Keywords: Locally Weighted Regression;time series decomposition;hypothesis testing;anomaly detection |
