(314a) Comparison of Machine Learning Techniques for the Modeling and Monitoring of Cryogenic Air Separation Unit Startups

For complex industrial processes with flexible operation and/or demand fluctuation in their products, such as cryogenic air separation units (ASUs), efficient startups are vital in productivity while reducing energy consumption. ASU startups are often challenging because they involve complex sequential actions with variable starting conditions. While there have been efforts in the industry to automate startup through a recipe-based approach, its proper execution needs the supervision and discretion of an expert operator to take corrective actions based on the state of the plant, which results in variations in the duration of different startups.

In this work, we investigate different machine learning techniques for their potential applications in modeling and monitoring of the startup of ASUs. The machine learning techniques studied in this work include linear principal component analysis (PCA), nonlinear PCA of principal curve and kernel-PCA, and neural network (NN) based self-organizing map (SOM) and variational autoencoder (VAE). We compare their performances in terms of early detection of bad or slow startups, as well as their diagnosis capabilities in terms of pinpointing the root causes. We also compare their practical implementation considerations and results interpretabilities.

To take the characteristics of ASU startups into account, we investigate the strategy of stage division for addressing significant nonlinearity and nonstationarity. We also investigate subunit division to address multimodality. Both knowledge-based and data-driven stage and subunit divisions are studied and their pros and cons are discussed. Their impacts on the performances of the abovementioned linear and nonlinear modeling techniques are compared.