基于DDNPE算法的间歇过程故障诊断

针对邻域保持嵌入(NPE)算法只通过欧氏距离挑选近邻带来的特征提取不充分导致故障诊断效果不佳的问题,将扩散距离(DD)与NPE算法相结合,提出了一种基于扩散距离的邻域保持嵌入(DDNPE)算法的故障诊断新方法.该方法首先发掘嵌入在原始高维数据的内在流行结构,进行数据降维,然后通过学习原始数据的潜在几何结构提取本征信息,并保持数据流行上的局部结构不变,避免了NPE算法只通过欧式距离挑选邻域带来的特征提取不充分的问题,最后利用T²和SPE统计量检测故障,并用变量贡献图法诊断出故障变量.通过青霉素发酵过程仿真结果验证了所提方法的有效性.

Fault diagnosis of batch process based on DDNPE algorithm

Traditionally, Euclidean distance determined only by neighborhood preserving embedding (NPE) algorithm is usually adopted to select the nearest neighbors for attracting characteristic features in fault diagnosis in industries. The troubleshooting effects in this way for the fault diagnosis are not effectively enough as well known. By employing a diffusion distance (DD) and combining it with NPE algorithm, a new fault diagnosis method based on the diffusion distance neighborhood preserving embedding (DDNPE) algorithm is proposed in this article by the author. The proposed new method explores the inherent popular structure embedded in the original high-dimensional data first, and reduces the data dimension, then extracts the characteristic information from the original data by learning the potential geometry information contained in the original data, while keeping local information structure in the original data popularity and unchanged. By doing so, it avoids the existing problem that the NPE algorithm selects the characteristics of the neighborhoods only by Euclidean distance. The problem of insufficient feature extraction can be avoided. It is now becoming possible to detect faults in fault diagnosis in industries by using T² and SPE statistics and the fault variable can be also diagnosed by the variable contribution graph method. Simulation results of penicillin fermentation process coming out from this research demonstrate that the proposed method is effective indeed.