Salient time steps selection from large scale time-varying data sets with dynamic time warping
Salient time steps selection from large scale time-varying data sets with dynamic time warping
复制标题
DOI：
10.1109/ldav.2012.6378975
复制
发表时间：
2012-12
期刊：
影响因子：
--
通讯作者：
Xin Tong;Teng-Yok Lee;Han-Wei Shen
中科院分区：
文献类型：
--
作者：
Xin Tong;Teng-Yok Lee;Han-Wei Shen
由于高性能计算机体系结构和软件的快速发展，科学家现在可以以前所未有的精度执行高分辨率模拟。如今，来自大规模仿真的数据的总大小可以很容易地超过数百TB甚至PB，分布在大量的时间步长上。数据的庞大规模使得在计算完成后很难进行后期分析和可视化。通常情况下，从模拟中产生的大量有价值的数据被丢弃，或留在磁盘中未分析。在本文中，我们提出了一种新的技术，可以检索最显着的时间步长，或关键的时间步长，从大规模的时变数据集。为了实现这一目标，我们开发了一种新的时间规整技术与一个有效的动态规划方案映射到用户指定的任意数量的时间步长的整个序列。我们的动态规划方案的一个新的贡献是整个时间序列和关键时间步之间的映射是全局最优的，因此信息损失是最小的。我们提出了一个高性能的算法来解决动态规划问题，使选择的关键时间运行在真实的时间。基于该技术，我们创建了一个可视化系统，允许用户浏览时变数据在任意层次的时间细节。由于该算法的计算复杂度低，该工具可以帮助用户交互式和分层地探索时变数据。我们证明了我们的算法的效用，从不同的随时间变化的数据集显示的结果。
Empowered by rapid advance of high performance computer architectures and software, it is now possible for scientists to perform high resolution simulations with unprecedented accuracy. Nowadays, the total size of data from a large-scale simulation can easily exceed hundreds of terabytes or even petabytes, distributed over a large number of time steps. The sheer size of data makes it difficult to perform post analysis and visualization after the computation is completed. Frequently, large amounts of valuable data produced from simulations are discarded, or left in disk unanalyzed. In this paper, we present a novel technique that can retrieve the most salient time steps, or key time steps, from large scale time-varying data sets. To achieve this goal, we develop a new time warping technique with an efficient dynamic programming scheme to map the whole sequence into an arbitrary number of time steps specified by the user. A novel contribution of our dynamic programming scheme is that the mapping between the whole time sequence and the key time steps is globally optimal, and hence the information loss is minimum. We propose a high performance algorithm to solve the dynamic programming problem that makes the selection of key times run in real time. Based on the technique, we create a visualization system that allows the user to browse time varying data at arbitrary levels of temporal detail. Because of the low computational complexity of this algorithm, the tool can help the user explore time varying data interactively and hierarchically. We demonstrate the utility of our algorithm by showing results from different time-varying data sets.