Abstract: Current piecewise linear representation methods just use heuristics method to extract the local feature points as the piecewise points. It usually falls into local optimization and cannot represent the global feature of the series very well. Besides, it usually uses the fitting error as the single threshold that cannot accurately predict the number of the sections. These methods cannot meet the application of time series analysis in late. This paper proposes a new representation method—PLR_BTBU which is based on the fixed section number method. The proposed method firstly uses the binary tree-traversal level method to divide the series into rough segments. Then, it uses the slope change characteristics to symbol the whole series. After that, the symbolic feature segments are initially used to determine the distribution of the piecewise points in each initial segment. Finally, an improved BU algorithm is used to fuse each sub sequence to the segment number. The experimental results show that the proposed method not only can preserve the global feature of time series, but also can get a smaller fitting error between the original time series and the new series compared with other algorithms.

Key words: time series, piecewise linear representation, binary tree-traversal level, symbolization, bottom-up

摘要： 针对目前的时间序列线性表示方法多采用启发式方法提取局部特征点作为分段点，容易陷入局部最优化，不能很好地表示时间序列全局特征，而且多采用单一的拟合误差作为阈值，不能准确预计分段数量，不利于后期进行的时间序列分析应用的问题。提出了一种新的固定分段数的表示方法——PLR_BTBU，首先根据二叉树层次遍历的思想，提取时间序列全局特征点将时间序列初始分段，再通过斜率变化特征将整个时间序列符号化，以各初始分段内的符号特征来确定各初始分段中的分段点分布，最后采用一种改进的固定分段数的自底向上融合算法，将各个子序列逐步融合到要求的分段数。实验结果表明，与已有的方法相比，该方法不仅较好地保留时间序列的全局特征，而且拟合后的时间序列和原时间序列之间的拟合误差更小。

关键词: 时间序列, 分段线性表示, 二叉树层次遍历, 符号化, 自底向上