Concluding Remarks | AIChE

Concluding Remarks

Visualizing big data sets in a reduced dimensional space is a key step for root cause identification in batch chemical processes. Persistent homology, a topological method, provides a novel way to summarize big data sets and link differences with process conditions by extracting characteristics of the high dimensional data space, such as the number of clusters. Visualization based on persistent homology brings similar information and requires less preprocessing of the data, such as an alignment, when comparing with traditional PCA and PLS multi-way approaches. The performance of the persistent homology method is also compared with UMAP (Uniform Manifold Approximation and Projection for Dimension Reduction) and t-sne (t-Distributed Stochastic Neighbor Embedding), demonstrating that persistent homology requires less hyper parameters for tuning. An industrial case study for chlorine production was utilized to evaluate performance of the visualization methods, where the source of run length underperformance was determined by examining specific process conditions during 7 years of operation.