改进的加权t-SNE算法及在故障诊断中的应用

对随机邻域嵌入算法(stochastic neighbor embedding, SNE)中的距离进行改进，提出一种基于Manhattan距离的加权t-SNE(Mwt-SNE)算法。使用受空间维数影响较小的Manhattan距离作为度量方式，使用k均值聚类算法将高维空间数据样本点距离分为三类，基于表格法进行权重参数寻优与加权，以加权相对Manhattan距离代替欧式绝对距离计算相似度条件概率，从而增大数据对象之间的区分度，提升降维效果，增强分类显著性。提出基于Mwt-SNE算法的在线故障诊断模型，使用核密度估计(KDE)确定控制限并进行在线监控。TE化工过程实验表明Mwt-SNE算法能有效降低误报率和漏报率，从而提高故障诊断稳定性和准确性。

Improved weighted t-SNE algorithm and application in fault diagnosis

This paper proposed a novel Manhattan distance based weighted t-SNE (Mwt-SNE) algorithm on the basis of improved distance in stochastic neighbor embedding (SNE). Firstly, it calculated samples Manhattan distances rather than Euclidean distance from high dimensional space for their less affections of dimension. Next, divided these Manhattan distances into three groups with K-means clustering algorithm and implemented weighting processing separately with tabular parameter optimization method. Then calculated similarity conditional probabilities with weighted Manhattan distances according to values category distribution. The aim of weighted Manhattan distances is to enlarge the data distinction, promote dimension reduction and enhance classification significance. Finally, established an Mwt-SNE algorithm based on-line fault diagnosis model and the corresponding control limit with kernel density estimation (KDE). The experimental results on TE chemical process show that the proposed Mwt-SNE algorithm reduced FAR (false alarm rate) and MAR (missing alarm rate) as well as improved stability and accuracy.