Online weighted dynamic time-difference neighborhood preserving embedding for dynamic process monitoring

Background:

Process monitoring is an important technology for ensuring the safety of the operating process. Recently, many methods based on neighborhood preserving embedding (NPE) are designed for extracting the local structure of data space. However, data collected in actual industrial processes usually exhibit serial correlation. These methods lack attention to time information and ignore the discrepancy in the discriminability of latent variables to faults.

Methods:

In order to alleviate the drawbacks for better monitoring performance, a novel online weighted dynamic time-difference neighborhood preserving embedding (OW-DTNPE) algorithm is proposed. Considering the serial correlation, the DTNPE is designed to construct the low-dimensional space based on a new neighborhood selection criterion, which preserves the global–local time information and local space information of process data. In addition, the online weighting strategy is designed based on the deviation degree of each latent variable in real-time. The different weights reflect the correlation between latent variables and faults.

Significant Findings:

The effectiveness of the proposed algorithm is demonstrated by a real glycerol distillation process, a numerical case, and Tennessee Eastman (TE) process. The proposed OW-DTNPE outperforms the other comparison methods and shows great potential in real applications.