中文摘要：

在多粒度时间序列研究中不确定性问题是时间序列数据挖掘研究中的重要课题．时间序列时序粒度本身的不稳定是一种广泛存在现象，也是时间序列数据挖掘困难的一个重要原因，然而这种情况却较少文献进行过讨论．对于这个问题首先建立了多粒度时间序列的基础数据模型及相关时序粒度的定义．其次对时间粒度不确定性现象的不同成因进行了讨论，并建立相应的不确定性时间序列数据模型．最后基于上述理论和粒计算的思想，多粒度时间序列的最优粒度获取和不确定性粒度时序粒度的基本稳定策略分别进行了研究和讨论．由于聚类分析是时间序列数据挖掘中的最重要的理论研究和应用基础之一，不确定性多粒度时间序列数据的聚类成为一个典型的时间序列数据挖掘难题．一个引入稳定粒度策略的聚类算法框架被提出来解决这类不确定性时间序列数据的聚类问题．最后一个典型的具有不稳定粒度时间序列特点的重症监护病房生理指标数据集和病人存活率预测实验被应用于验证上述理论．实验结果表明在时间序列数据挖掘中选择不同的时间属性粒度对于数据挖掘的效果符合粒计算的计算规律，同时选择了粒度稳定性处理策略聚类算法的实验能够获得更好的预测效果．

英文摘要：

Uncertainty is a major and important property in the study of time series data mining under the background of multi-granular research. It is found that the instability of temporal granularity itself is a widespread phenomenon, but due to the variability of granularity,it is a non-trivial task to develop feasible solutions. As a result,few papers report efficient idea for this problem. To describe it systematically,a basic model of the uncertain multi-granularity time series with related basic definition is proposed initially. Based on the above theory, the reason of uncertain time series granularity is analyzed, and a strategy of stabling time granularity algorithm is then proposed. Finally, an optimal strategy of granularity selection for multi-granularity time series and a tactic for decreasing the variability of uncertain granularity are studied and discussed respectively. Uncertain multi-granularity time series data clustering requires cluster analysis since it is a fundamental job for further process such as prediction and analysis in time-series data,but due to the complexity, it becomes a special challenging. We introduce a stable temporal granularity clustering algorithm framework to solve the uncertain time series data clustering problem. An Intensive Care Unit Survival prediction experiment with the property of uncertainty in temporal granularity is conducted to verify the above theory. Results show that different choices of time granularity produces different outcome which is in line with the computational law of granular computing, and stable strategy for granularity can achieve satisfied prediction.