路灯监控系统中时序数据流的异常值检测研究

Research on Abnormal Value Detection of Time-Series Data Flow in Street Lamp Monitoring System

在智能城市的景观照明系统中,各种灯具、继电器和接触器等部件的实时状况经传感器采集后,往往会形成海量的时序数据流,为了构建对其更为高效实用的监控系统,提出一种改进的时序数据流异常值检测算法。首先针对路灯时序数据流具有均值缓变的特点,以滑动窗口的形式在线进行均值显著性判别、子序列划分以及分段去均值等处理;然后,基于自回归模型对各分段近似平稳的子序列进行建模,并估算出对应的模型参数;最后,基于自回归模型的建模残差定义有关统计量,并根据统计量迭代求出所有异常数据点。实验表明,改进算法在没有增加计算成本的前提下,其检测精度及算法的鲁棒性有了明显的提升。

In the smart city landscape lighting system,the real-time status of various lamps,relays,contactors and other components will often form massive time series data streams after being collected by sensors.In order to build a more efficient and practical monitoring system for them,an improved time series data stream outlier detection algorithm is proposed.At first,according to the characteristic that the street lamp time series data stream can slowly change in mean value,the mean significance discrimination,subsequence division and piecewise mean removal are performed online in the form of a sliding window.Then,based on the autoregressive model,the approximate stationary subsequence of each segment is modeled and the corresponding model parameters are estimated.Finally,the relevant statistics are defined based on the modeling residual of the autoregressive model,and all abnormal data points are iteratively obtained according to the statistics.Experiments show that the detection accuracy and robustness of the improved algorithm are improved obviously without increasing the computational cost.