基于贝叶斯和有限集统计的多源多尺度异构统计信息融合

统计信息融合是最重要的数据处理手段之一，广泛用于军事和民用领域。在大数据时代，伴随传感手段的不断丰富，传感数据日益呈现出多源异构、多尺度等特点，这为信息融合理论的应用带来了极大挑战。具体表现为：数据多源异构，导致数据表示形式多样化，难以对其进行统一化、形式化表示；多源观测中含有的不确定、不精确性因素的成因也各不相同，难以用统一化模型语言对其进行统计分析建模；相互冲突观测和多尺度观测信息的存在，使得数据间的统计独立性假设不再成立，需要设计恰当模型用于刻划观测间以及信息尺度间的关联性，避免融合过程中出现不必要的信息丢失。本课题即在传统信息融合理论基础之上，结合有限集统计、贝叶斯非参、网络和分层模型以及隐变量分析模型，研究推导多源异构多尺度数据表示方法、信息融合机理、模型与算法，以解决阻碍信息融合理论直接应用于大数据分析的关键问题，进而实证信息融合概念在大数据分析问题中的潜在应用价值。

Statistical information fusion is one of the most important data processing approaches and has been widely used in both military and civil areas. In big data era, along with the ongoing abundance of the sensing approaches, the sensed data are increasingly characterized by the multisource, heterogeneous and multi-scale properties, which bring grand challenges for the applications of information fusion theories. Specifically, the multisource heterogeneous property leads to a diversity in data representations, thus the data are difficult to be represented in an unified and formalistic way. The sources of uncertainties and inaccuracies underlying the multisource data are also different, which makes it difficult to statistically analyze and model the observations using a unified modeling language. The existence of conflicting observations and the observations of different scales makes the assumption of statistical independence among data items not satisfied any more. So it is required to design appropriate models to characterize the relevances among the observations and among the information scales, in order to avoid unnecessary information loss during the information fusion process. This project is grounded on traditional information fusion theories, connecting finite set statistics , Bayesian nonparametrics, Bayesian networks and hierarchical modeling, and latent variable models to investigate and derive data representation approaches, information fusion mechanisms, models and algorithms for multisource heterogeneous observations. The purpose is to solve key problems that hinder the information fusion theories from being directly used for big data analysis, and thus to demonstrate the potential values of the information fusion concept in applications with respect to big data.