Development of multivariate statistical process monitoring using combination MLR-PCA method
The most popular types of process monitoring systems is Multivariate Statistical Process Monitoring (MSPM) which has the most practical method in handling the complicated large scale processes. This is due to the ability of the system in maximizing the usage of abundant historial process data, in su...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English English |
| Published: |
2015
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/12737/ http://umpir.ump.edu.my/id/eprint/12737/ http://umpir.ump.edu.my/id/eprint/12737/1/FKKSA%20-%20NORLIZA%20ISHAK%20-%20CD%209668.pdf http://umpir.ump.edu.my/id/eprint/12737/2/FKKSA%20-%20NORLIZA%20ISHAK%20-%20CD%209668%20-%20CHAP%201.pdf http://umpir.ump.edu.my/id/eprint/12737/3/FKKSA%20-%20NORLIZA%20ISHAK%20-%20CD%209668%20-%20CHAP%203.pdf |
| Summary: | The most popular types of process monitoring systems is Multivariate Statistical Process Monitoring (MSPM) which has the most practical method in handling the complicated large scale processes. This is due to the ability of the system in maximizing the usage of abundant historial process data, in such a way that the original data dimensions are compressed and data variations preserved to certain extent in a set of transformed variables by way of linear combinations. Thus, the composite model is generally flexible regardless of the amount of variables that utilized. In this regard, conventional Principal Component Analysis (PCA) has been widely applied to conduct such compression function particularly for MSPM. However, the conventional PCA is a linear technique which results sometimes inappropriately employed especially in modeling processes that exhibit severe non-linear correlations. Therefore, a new solution is demanded, whereby the number of original variables can be reduced to certain extent (in terms of scales), while it still can maintain the variation as maximally as possible corresponding to the original, which are then transferable into monitoring statistics. One of the potential techniques available in addressing the issue is known as Multiple Linear Regression (MLR). The main objective of the technique is to predict a set of output values (criterion) based from a specified of linear function, which consists of a set of predictor. Therefore, the main multivariate data will be divided into two groups, which are the criterion and predictor categories. The study adopts, Tennessee Eastman Process (TEP) and Multiple Output and Multiple Input (MIMO) Pilot Plant System for demonstration. The general finding is that MLR-PCA normally employs less number of PCs compared to PCA, and thus, this will perhaps reduce complication during diagnosis. By adopting such approach, the monitoring task can be made simpler and perhaps more effective, in the sense that only those selected criterion variables (predicted values) will be taken for monitoring, while preserving the rest of the predictor value trends in the form of linear functions in association with the criterion variables. This study also shows that MLR-PCA works relatively better in terms of fault detection and identification against the conventional system. |
|---|