A Data-Driven Multidimensional Visualization Technique For Process Fault Detection and Diagnosis.

摘要

This thesis describes a multidimensional visualization technique for fault detection and diagnosis of a multivariate process by principal component analysis (PCA) of historical data. The visualization technique uses a parallel coordinate system to visualize data that allows for detection of abnormal process events and fault propagation. The technique enables the visualization of multiple principal components effectively and facilitates the study of how variation of each principal component changes with respect to time. Furthermore, we propose the use of principal component and residual space control limits for fault detection and "Random Forests" machine learning for fault diagnosis. The validity and usefulness of the techniques are demonstrated through a comparative study of the benchmark Tennessee Eastman process.;Keywords: Fault detection, Fault diagnosis, Random Forests, Singular value decomposition, Principal component analysis (PCA), Multivariate statistical process monitoring, Multidimensional visualization, Big Data, Historical data analysis, Tennessee Eastman process.