时序数据故障点检测方法分析比较及应用

Comparison and Application of Fault Point Detection Method used in Time Series Data Analysis

比较了3种不同的时序数据故障点检测算法.基于引力的孤立点检测算法考虑了数据对象周围的密度及数据之间的距离等因素.基于均值变点的检测算法则侧重于考察故障点周围统计量的变化,而非故障点的局部范围内统计量保持未定.第三种基于均值方差变点估计的检测算法则研究了时序数据中均值和方差两个统计量都存在变点且变点时刻不相同时的变点估计问题.试验表明基于引力的算法比其他两种效果要差,而基于均值变点检测算法的计算效率要比基于均值方差估计检测算法要高.

Three different fault point detection methods were analyzed and compared. Algorithm which is based on gravity mainly considered both the density around time series data and the distance between the point waited to be detected and the other neighbors. Faults detection method, based on change-point of mean, focuses on the statistical value around a point. In the area near the fault point, there may be more changes to statistical value of this area. The third approach both considered the problem of change point of independent observations whose mean and variance change respectively. Experiments showed that efficiency of algorithm based on gravity is less than the other two methods. And computational efficiency of the method based on change point of mean is better than the third approach.