A new technique using principal component analysis for predicting abnormal events. The technique was proven to not only identify when an event would occur but also the sources (tags) of the problem.
This software tool was the culmination of several COP projects looking to improve the visualization of process data to help operators recognize and prevent potential abnormal operating conditions. The analysis workflow associated with the radial plots framework (data preprocessing and filtering, identifying steady-state regions in the data, model development, visualization, and fault detection) was implemented into a single standalone software that can be applied to any set of data.
The following files are available:
- Radial Plot Tool
- Windows-Based Radial Plot Tool
- Matlab and Python Source Code
- Training Presentation
- Training Videos
Note: Use of this tool requires understanding of principal component analysis. Every intention was made to ensure that the provided files are comprehensive, but it was created in 2014 and updates to the software have not been made.
Application of the tool are front-loaded in the sense that the preparation of an accurate model is the most time consuming task. Once developed, this model can be, through the use of the developed software, quickly and easily accessed by multiple engineers and operators for fault detection on any given data set.
Regarding model building, the developers make specific recommendations as follows:
- Perform aggressive selection of the data (following recommendations from Nova Chemical). This includes,
- retrieving several years of data,
- inspecting alarm logs and other available information to identify the time of process events,
- eliminating a day or more of data, depending on the process time constant, from before and after the identified events to remove data affected by the cause and any residual effects of these events,
- using hourly averages to diminish the impact of noise, and
- retaining only the variables/tags that are consistently present throughout the operation of the process.
- Maintain a larger number of components in the PCA analysis that is typical in most PCA applications. We found that retaining sufficient components to capture 95% variance is appropriate (vs. about 80% variance captured in typical PCA applications)
- Multivariate confidence limits of around 95% when building the fault detection models tend to offer the best balance between low false alarms and sensitivity to events/faults.