Eureka! Over the course of four projects, a method and tool were created that accurately predict abnormal events prior to the event happening. The tool works but implementing it can be tricky.
Dr. Michael Baldea of the University of Texas created a method to display complex 3D data in an easy to visualize 2D format. Previous 3D data for early detection of process upsets was converted by calculating and presenting the center of the polygon moving over time. Violation of boundaries indicating an abnormal situation was easy to detect with this centroid approach. Individual tags contributing to the violation are highlighted to aid in troubleshooting the cause. Further development of the technique needs to be done by software developer to ensure long-term support.
A significant number of undesirable process events can be prevented by the use of empirical indexes and visualization techniques preconfigured for recorded historian data. The development of such indicators depends of the data features captured during abnormal operating conditions. The proposed work will determine how to make use of data that characterize abnormal conditions to identify and present such indicators to operators, and how to best apply these decision aids to assist the operators in an industrial environment.
• Which visual data mining tools can be implemented to assist operators during process monitoring?
• How to extract meaningful statistical features from normal and abnormal operating data?
• How can such features or indicators represented visually in a graphical environment?
• Which data features can be used to predict the advent of an undesirable operating condition?
• How to reduce the number of false alarms when predicting the coming of a potential fault?
• What is the probability of successfully predicted faults?
• How to incorporate the proposed methodology in new operator decision support tools?
• Can these methods be used to select the parameters required to support key decisions (i.e., adding new key variables to the operator interface)? and, if so, What is the best way to use the methods to select parameters to support key decisions?
In Phase I of the project (EPM III), researchers from University of Texas have introduced three-dimensional radial coordinates (3DRCs) for the visualization of multi-variate time dependent process data. In 3DRCs the time dimension of data is represented explicitly, whereas in representations such as parallel coordinates (as used in CPM/PPCL) time is obscured in the plot. For 3DRCs, normal process conditions were defined in terms of a region lying between the inner and outer hulls of the 3DRC plot. Using compressor surge and column flooding data sets it was shown that a violation of the normal operating region by the data plot is indicative of process conditions leading to abnormal events. The detailed description of these developments is documented in a paper, which has been submitted for publication in the Journal of Process Control.
• Visualization algorithm for process monitoring, operator display and abnormal event detection and prediction
• Mechanism for visual classification of types of abnormal events as a precursor for root cause analysis
• Mechanism for early isolation of the variables causing abnormal events
• Extension of visualization approach to monitoring transitions between process states and detecting abnormal events during transitions
• Matlab implementation of the above algorithms, tested and validated with aforementioned data sets
• Journal publication of the research findings
To access the tool, use the first five files below: