M, Venkata Ramana and Detroja, Ketan P
(2018)
Fault Diagnosis In Batch Process Monitoring.
Masters thesis, Indian Institute of Technology Hyderabad.
Abstract
Every process plant nowadays highly complex to produce high-quality products and to satisfy de-
mands in time. Other than that, plant safety is also crucial event had to be taken care to increase
plant e�ciency. Due to poor monitoring strategies leads to huge loss of income and valuable time to
regain its normal behavior. So, when there is any fault occurs in the plant it should be detected and
need to take supervisory action before propagating it to new locations and new equipment failure
leads to plant halt. Therefore process monitoring is very crucial event had to be done e�ectively.
In Chapter 1 Importance of fault detection and diagnosis(FDD) in plant monitoring, what are
the typical situations will leads to fault and their causes of fault is discussed. How data will be
transformed in di�erent stages in diagnostic system before certain action, desirable characteristics
for good diagnostic systems are discussed brie
y. And in �nal part of this chapter what are the basic
classi�cations of FDD methods are discussed. Principle component analysis is multivariate statistical
technique helps to extract major information with few dimensions. Dimensionality of reduced space is
very low compared to original dimension of data set. Number of principle component(PC) selection
depends on variability or information required in lower dimensional space. So PCA is e�ective
dimensionality reduction technique. But for process monitoring both PC and residual space are
important. In chapter 2 mainly discussed about PCA and its theory.
Batch Process Monitoring is relatively not easy to monitor compared to Continuous process be-
cause of their dynamic nature and non-linearity in the data. So there are methods like MPCA(multi-
way Principle component analysis), MCA(multi-way correspondence analysis) and Kernal PCA, Dis-
similarity Index based(DISSIM) etc., are there to monitor batch process. Kernal based methods need
to choose right kernal based on the non-linearity in the data. Dissimilarity Index based methods
well suits for continuous process monitoring since it can able to detect the changes in distribution of
data. Extension of DISSIM method to batch process monitoring is EDISSIM, which is discussed in
chapter 3. And also MPCA is very traditional method which can able to detect abnormal sample but
these cannot be able to detect small mean deviations in measurements. Multi Way PCA is applied
after unfolding the data. Batch data Unfolding discussed in section 3.2 and selection of control lim-
its discussed in 3.2.3. Apart from these methods there is another strategy called Pattern matching
method introduced by Johannesmeyer. This method will helps to quickly locate the similar patterns
in historical database. In Process industries we frequently collect the data so that there will be lot
of data available. But there will be less information containing in it, used PCA to extract main
information. In pattern matching strategy to detect the similar patterns in historical data base we
need to provide some quantitative measure of similarity between two data sets those are similarity
factors. So by using PCA method we are extracting high informative data in lower dimensional
space. So Using PCA method similarity factors are calculated. Di�erent similarity factors and their
calculation is shown in chapter 4. On-line monitoring of Acetone Butanol batch process discussed
using pattern matching strategy. Acetone Butanol fermentation process mathematical model will
be simulated to di�erent nominal values with di�erent operating conditions to develop historical
database. In this case study there will be 500 batches with �ve operations conditions like one NOC
and 4 di�erent faulty operation batches. In each batch there will be 100 batches. After calculation
of similarity factors instead of going for candidate pool selection directly we are trying to detect
the batches which are similar to snapshot data. Performance of On-line monitoring using pattern
matching strategy is discussed. On-line monitoring strategy will change the way we are anticipating
iv
the un�lled data. Here we are trying to �ll with reference batch data. Reference data will be average
of NOC batches. The performance of this method veri�ed in MATLAB as shown in section 4.3.
In Chapter 5 described average PC's(Principle components) model. This method will helps to
decrease the e�orts in candidate pool selection and evaluation to �nd snapshot data in historical
database. And also Incremental average model building and model updating will improves the quality
of model ultimately.In incremental average model building If any of the snapshot data classi�ed as
any of the already existed operating condition data set it will be used in building average model.
If not existed in any of the operating condition data set utilized to update average model. This
method applied on Acetone Butanol fermentation process data and veri�ed. Because of the fact
that batch data highly non linear in nature So PCA not able to handle non-linear correlations.
And pattern matching approach using PCA average model not give good discrimination. For better
discrimination ability and self aggregation can be possible using Corresponding Analysis because of
non-linear scaling. In chapter 6 pattern matching approach using corresponding analysis has been
discussed brie
y. Results obtained using CA based similarity factor displayed for Acetone Butanol
fermentation process case study.
Actions (login required)
 |
View Item |
Altmetric
Altmetric