Gaussian Mixture Model-Based Wasserstein Stationary Subspace Analysis for Process Monitoring
Hanwen Zhang, Qingqing Liu, Chao Jia
Abstract
Safe and stable production is a common concern in industrial processes, determining the manufacturing cost, product quality, energy consumption, pollutant discharge, etc. Timely detection of abnormal conditions and prompt counter-measures are vital for maintaining the safe operation of the industrial processes. However, the prevalent non-stationarity and non-Gaussianity in industrial process data present significant challenges to effective process monitoring. This paper introduces a novel approach, Gaussian mixture model-based Wasserstein stationary subspace analysis (GMM-WSSA), to address these challenges. The method focuses on extracting stationary features by minimizing the differences across distant time intervals, utilizing Wasserstein-type distances as a measure of these differences to overcome the limitations of Kullback-Leibler divergence. To adequately represent the non-Gaussian characteristics observed in practical processes, data from each time interval are modeled using Gaussian mixture models (GMM). The Wasserstein-type distance of these GMMs forms the basis of the constructed optimization objective, enabling more accurate and reliable process monitoring under non-stationary and non-Gaussian conditions. The effectiveness of the proposed method is demonstrated through its application to a real-world industrial process, specifically an ironmaking process, illustrating its potential to enhance operational safety and efficiency.