基于残差融合的时序数据离群点检测算法

物联网的快速发展产生了海量的高维时序数据,然而时间序列易受到外界变化的环境因素影响而产生离群点。针对现有的离群点挖掘算法不能兼顾时序数据的趋势性、季节性、循环性、不规则性的特点,从而导致检测效果不理想的问题,提出一种基于残差融合的时序数据离群挖掘(residual integration outlier,RIO)算法。首先利用线性自回归移动平均模型(autoregressive integrated moving average model,ARIMA)拟合数据,得到在相同时间粒度下的残差序列,并将该序列作为非线性模型长短期记忆网络(long short-term memory,LSTM)模型的输入,输出残差序列预测值,而后将经由ARIMA模型与LSTM模型处理的序列在相同时间粒度下融合,得到一条经由混合模型两次处理的残差序列。最后,利用基于直方图的离群点模型(histogram-based outlier score,HBOS)检测出该二次残差序列的离群点。实验表明,RIO算法的准确度得到了较为明显的提高,具备良好的实用价值。

Research on Outlier Detection Algorithm of Time Series Data Based on Residual Integration

Aiming at outlier mining in massive high-dimensional time series data, this paper proposes the time-series data outlier mining algorithm RIO(Residual Integration Outlier). Firstly, the data is fitted by the Autoregressive Integrated Moving Average Model (ARIMA) to obtain the residual series at the same time granularity. Then, the series are as the input of the Long?Short-Term?Memory network (LSTM ), and output the residual series prediction value. Both series processed by the ARIMA model and the LSTM model are integrated at the same time granularity to obtain the residual series. Finally, the outliers of the integrated residual series are detected using the HBOS(Histogram-based outlier score) algorithm. Experiments show that the accuracy of the RIO algorithm has been significantly improved and has good practical value.