基于特征挖掘的时间序列异常检测算法

Time series anomaly detection algorithm based on feature mining

为了有效地挖掘监控指标时间序列的周期性特征和节假日效应,本文提出了一种基于特征挖掘的时间序列异常检测集成算法。通过FFT算法检测序列的周期性特征,以及MODWT算法识别非周期序列的局部波动特征,以天为单位计算不同子序列的CSBD值,作为OPTICS聚类算法的距离半径,进而识别序列的节假日特征,并基于Holt-Winters算法所得的预测误差,构建余弦加形变量波动指标,以降低序列的时间、形状和量级等其余特征对异常点检测的影响。经过实际业务的应用验证,本文算法相对传统的算法准确率得到显著提升,且可以保持良好的稳健性。

We propose an integration algorithm for time series anomaly detection based on feature mining in order to effectively explore the cycle characteristics and holiday effects of the time series of monitoring indicators.Specifi cally,the FFT algorithm is utilized to detect the features of periodic sequences,and the local fl uctuation features of non-periodic ones are identified by MODWT algorithm.Afterwards the CSBD of different subsequences are computed in daily frequency as the distance radius of the OPTICS clustering algorithm to extract holiday features of the tested sequences.Based on the prediction error obtained by the Holt Winters algorithm,a volatility index measured by cosine plus shape variables is constructed to reduce the impact of other features such as time position,shape,and magnitude of the sequence on anomaly detection.The application of the proposed algorithm in practice convincingly verifi es its signifi cant accuracy as well as good robustness compared to traditional algorithms.