(191a) Visualizing Chemical Process Data Via Dimensionality Reduction

The complexity of a chemical plant often requires multivariate analyses to identify the root cause of process interruptions, key variables influencing product quality, etc. Visualization of the dataset early in the data analytics journey can provide preliminary insights to guide an efficient analysis. Dimensionality reduction techniques, such as principal component analysis, provide a concise visualization in just a few dimensions. The techniques differ in how they attempt to preserve local or global properties of the entire dataset in mapping from the original, high dimensional space to the reduced, low dimensional space. The visualization readily enables cluster analysis in the reduced space that can assist in quickly identifying the key performance indicators. While there is an abundance of dimensionality reduction techniques, most are seldom applied to chemical process data. Therefore, we present case studies in which we analyze process data with a variety of these techniques, such as Isomap, t-SNE, and Laplacian Eigenmaps, among others. We outline their strengths and weaknesses and make comparisons against the industry standard approach – principal component analysis.

References:

L. van der Maaten, E. Postma, J. van den Herik. Dimensionality Reduction: A Comparative Review. Tilburg University Technical Report, 2009

Z. Ge, Z. Song, S.X. Ding, B. Huang. Data Mining and Analytics in the Process Industry: The Role of Machine Learning. Access IEEE. 2017.