Chemical Process Data Classification and Visualization for Process Monitoring

Publication Year:
2016
Usage 10
Abstract Views 6
Downloads 4
Repository URL:
https://digitalcommons.lsu.edu/gradschool_theses/4472
Author(s):
Gowri Shankar, Vikram
Tags:
SOM; Data Clustering; Fault Detection
thesis / dissertation description
Process monitoring is required for safety of operations, considerable reduction in downtime, and decrease in manufacturing costs. Chemical Engineers have a high responsibility in the proper functioning of a process plant as any deviations from normal operations might lead to a disastrous effect in loss of lives and infrastructure. The increased number of microprocessors due to the reduction in cost (Effect of Moore’s Law) has increased the speed of computers. This has led to the increase in the amount of data storage. Thus, creating a scope for us to train machines to identify representations of the data. Data clustering, an exploratory data analytics technique helps us in the process of unsupervised learning: unclassified datasets. Previously, process monitoring has been done using statistical techniques such as component analyses, however one of the challenge in practical applications is the difficulty to classify (cluster) the information from a high dimension data set commonly encountered in chemical process industry. There are many clustering algorithms such as K-Means, Mean Shift, Hierarchical, DBSCAN. In this study a detailed analysis of the alternative approaches for data classification was performed including conventional and novel techniques arising from Computer Science. This is the first reported instance of the use of HDBSCAN, a hybrid of Hierarchical and Density Based Spatial Clustering with Applications and noise, in a chemical process data. We have concluded that HDBSCAN outperforms the other clustering algorithms on chemical process dataset. The Chemical process dataset used was from the Tennessee Eastman Process. Since, in this study we compared clustering algorithms we realized the need for an easy automation for running clustering algorithms ,hence we built an application called “Mi ClustoRoma”, a graphical user interface built on python for classifying and visualizing chemical process datasets.