A team of researchers from East China University of Science and Technology has introduced a novel visual monitoring technique aimed at enhancing the visualization and monitoring of chemical processes. The findings are detailed in their study, “Supervised Projection with Adaptive Label Assignment for Enhanced Visualization and Chemical Process Monitoring,” published in Frontiers of Chemical Science & Engineering.
In industrial settings, data-driven monitoring techniques play a crucial role, as they allow operators to gain an intuitive understanding of operational conditions. This understanding is essential for optimizing safety and efficiency in production. However, the high-dimensional nature of industrial data often leads to intricate structures, which traditional two-dimensional visualization methods struggle to interpret effectively, particularly when distinguishing various fault types.
To address these challenges, the research proposes a methodology that combines supervised uniform manifold approximation and projection with an innovative label assignment strategy. Initially, the method enhances the visualization process by utilizing label information to steer the nonlinear dimensionality reduction, thereby improving the separation between classes and increasing the compactness within the same class.
Recognizing that online samples often lack sufficient label information, the study introduces a strategic approach for label assignment. This approach leverages kernel Fisher discriminant analysis and Bayesian inference to categorize online samples based on their confidence levels, assigning them appropriate labels. The integration of this label assignment strategy with the supervised projection method significantly boosts the separability of online projections and facilitates the visualization of previously unknown data.
The proposed method has undergone validation through applications on the Tennessee Eastman process and a real continuous catalytic reforming process. Results indicate that this new approach surpasses existing state-of-the-art methods in visual fault monitoring and diagnosis, demonstrating particularly effective performance in real industrial applications.
For those interested in a more comprehensive understanding of this research, the full paper is accessible at this link.
