On Generative Topographic Mapping and Graph Theory combined approach for unsupervised non-linear data visualization and fault identification

Process monitoring of chemical plants relies on two steps: discriminating anomalies (fault detection) and characterizing them (fault identification). This work proposes a combined Generative Topographic Mapping (GTM) and Graph Theory (GT) approach. GTM highlights system features, reducing variable dimensionality and providing a strategy for calculating similarity between samples. GT then clusters them using networks, discriminating normal and anomalous entries. Because of biased normal and anomalous labeling, however, the methodology proposed is unsupervised, meaning that labels are inexistent. Three case studies were considered: a simulation data set, Tennessee Eastman process and an industrial data set. Principal Component Analysis (PCA), dynamic PCA and kernel PCA indexes (Q and T2) alongside GTM and GT independent monitoring methodologies were used for comparison, considering supervised and unsupervised approaches. For the industrial scenario, soft sensors were used for assessing discrimination performance. The proposed method, while unsupervised, discriminated normal states similarly to supervised strategies, justifying its development.