本项目主要研究间歇过程数据驱动故障诊断中的海量数据的降维处理和数据统计模型准确.性的提高问题。提出将高维空间数据建立与之对应的Markov随机游走邻近关联，得到任意数据点之间近似度的一种加权距离，通过数据的扩散映射最大化保留扩散距离统计量信息并获取数据的低维流形特征，提取关联数据空间转移概率扩散映射下的低维数据空间，即数据的本征维度空间。数据本征维度是最大化保留数据空间信息的最小维度，降低了数据的冗余度，提高了数据存储量和计算速度。本征维度空间的k近邻考虑数据间的稀疏分布和数据近邻的关联度和相似度，同时在数据扩散游走过程中，提取数据点密度扩散核，建立数据局部特征关系、捕捉数据局部几何特性，得到本征维度空间下的数据间的扩散距离,将扩散距离作为k近邻统计分布的标定距离进行故障诊断。上述方法将应用于东北制药集团的磷霉素发酵间歇生产过程进行验证实验，可进一步提高产品质量和生产的安全性与可靠性。

This research is mainly focus on a large amount of data dimensionality reduction and statistical model accuracy improvement for data driven fault diagnosis in batch process. It is proposed that the data in high dimension is relevance to Markov neighbor random walk and the weighted distant based degree of approximation between any data is gained. By data diffusion maps, the distant statistic maximum information is reserved and manifold features in data low dimension is gained, and data in low dimension is extracted based relevant data transition probability diffusion maps, then this is called data intrinsic dimension. The data intrinsic dimension is a minimum dimension that optimally preserve data original information and reduce the data redundancy and enhance data memory size and calculating speed. k-nearest-neighbor in data intrinsic dimension is based not only sparse distributed and degree of relevant and approximation but also extracting data density diffusion kernel and local characteristic relationship and distributed geometrical features in diffusion random walk, then diffusion distant in data intrinsic dimension is gained and used in fault diagnosis based k-nearest-neighbor. These theories and methods will be applied into fom fermentation batch process in Northeast Pharmaceutical Group, then product quality and safety and reliability will be improved.