- To demonstrate the
beneficial application of Multivariate Statistical Process Control
(MSPC) and Plant Performance Monitoring in process manufacturing
and in pilot plant developments.
- To provide a contribution
to a better understanding of process operations.
- To predict the final
quality of the product from the more rapidly monitored process
variables.
- To establish a protocol
for the implementation of multivariate statistical process performance
monitoring.
- To provide showcase
demonstrators that will illustrate the power of the techniques
across a range of application sectors.
Achievements
The project objectives
were met and exceeded both in terms of undertaking internationally
recognised research, and catalysing the introduction of innovative
manufacturing methodologies within Quest International, BP, Pfizer, SmithKline Beecham and Zeneca. The research work focused
on meeting the industrial manufacturing challenges defined by the
end-user companies. This necessitated extending the state-of-the-art
in process performance monitoring and process diagnostics. Achievements
with individual CPACT companies, in collaboration with the research
team at Newcastle, are summarised below. The Research Associates
and Masters students employed on Project 6 have all attained good
positions in industry or else have gone on to study for a higher
degree.
1 University of Newcastle
(Centre for Process Analytics and Control Technology)
New developments in multivariate
data analysis, modelling and process performance monitoring for
both batch and continuous processes were all initially evaluated
on benchmark simulations prior to their being applied to industrial
challenges and data sets.
- Different approaches
for batch performance monitoring were evaluated - individual batch
analysis, end of batch analysis and observation level batch analysis.
It was concluded that the different techniques were complimentary
and that no one methodology outperformed the other approaches.
- A comparison between
the tri-linear technique of parallel factor analysis (PARAFAC)
and the bi-linear approach of multi-way performance monitoring
was undertaken on a benchmark simulation of a batch polymerisation
process. Advantages and limitations were identified with both
approaches.
- A new on-line batch-monitoring
algorithm was developed for the tri-linear technique of (PARAFAC).
This is based on a weighted scores in-filling procedure. Modified
confidence bounds were developed for the PARAFAC algorithm. The
methodology developed was shown to be equally applicable to bi-linear
(multi-way) approaches. Considerable benefits have been identified
with the revised algorithm in terms of the reduction in number
of false alarms and clarity of identifying the source of the process
change through contribution plots with associated bounds.
- PARAFAC models are
known to suffer from degeneracy, a comparative study that focused
on the various PARAFAC algorithms (ALS, ASD and ACOVER) has been
carried out and the advantages and disadvantages of the different
algorithms assessed. Based on the results of this study, a solution
to the problem of degeneracy has been researched.
- The issue of sensitivity
analysis for variable selection, for both predictive and performance
monitoring models, has been investigated through use of principal
component analysis. This approach complements that of process
understanding that is essential in model building. This was applied
to the Pfizer fermentation demonstration process.
- Investigations into
the modelling of non-linear processes using a number of local
linear models are on going. Local models are combined through
either through Bayesian validity or fuzzy membership functions.
The aim is to be able to use linear approaches to model and monitor
non-linear processes where the overall global non-linear model
comprises several local linear models. The methodology is being
evaluated on simulations of semi-batch processes including fermentation
and polymerisation.
- Plant wide, dynamic
multivariate statistical process control is being progressed through
simulation studies on the benchmark Tennessee Eastman problem.
Tools investigated included dynamic state space models. Comparisons
between dynamic PCA, dynamic PLS and canonical variate state space
approaches is being carried out.
- The dynamic modelling
of complex non-linear processes through non-linear dynamic partial
least squares (PLS) has been achieved. The methodology was demonstrated
by application to a benchmark simulation of a pH neutralisation
process. A new model-based predictive control approach based on
the use of a dynamic PLS model has also been proposed.
- The development and
application of dynamic partial least squares (PLS) and dynamic
multi-block PLS modelling methodologies for the prediction of
important parameters in a biological waste-water treatment process
has been progressed. Complimentary to this was the application
of multivariate performance monitoring techniques to the SmithKline
Beecham activated sludge treatment plant. This study demonstrated
the power of the techniques for enhanced monitoring of such processes.
CPACT Phase II will progress these initial studies through to
a full plant application with GSK.
- Comparison of a range
of clustering techniques on shake flask and fermenter data in
new drug development with Zeneca Pharmaceuticals. This analysis
demonstrated the power of multivariate statistical data analysis
and visualisation techniques in the discovery of new bio-therapeutics.
A new algorithm for the equalisation of batch lengths based on
dynamic time warping was developed and applied to the analysis
of the Zeneca drug development data where the batches differed
considerably in length and data was only available on a limited
number of batches.
- A methodology previously
proposed to evaluate prediction intervals for neural network models
has been modified and extended to linear and non-linear algorithms.
The intervals are computed using a first-order Taylor series expansion
and the Jacobian matrix of the functional mapping provided by
the PLS algorithms.
- The continued development
of MATLAB© based Multivariate Statistical Process
Control software, PreScreen, MultiDAT and BatchDAT. The software
was issued to those partners actively involved in the process
performance monitoring project.
2 BP (Grangemouth)
- Data from three plants
were analysed using multivariate statistical data analysis techniques
- Innovex (low density polymer); Rigidex (high density
polymer); and PIB (high viscosity polymer).
- The standard process
performance monitoring techniques of principal component analysis
(PCA) and partial least squares were applied to data collected
during the manufacture of high-density polyethylene (Rigidex)
for the early detection of deviations from specification of melt
index and density. A similar set of objectives was identified
with respect to the Innovex process, i.e. early warning of process
malfunctions. This study has also demonstrated how final product
quality can be efficiently monitored using partial least squares
(PLS). PLS also provides a technology for the on-line estimation
of difficult, or expensive, to measure process variables, such
as melt flow index.
- For the PIB process,
PCA models for each polymer grade were built, corresponding to
the normal operating conditions for different production grades.
Process control and other disturbances as well as grade transitions
were identified as data points that exceeded the confidence bounds
in the squared prediction error and T2 plots.
Contribution plots were used to identify the variables responsible
for different operating conditions and the sources of the disturbances.
For modelling the polymer viscosity, PLS models were developed
based on each grade separately as well as using a combination
of grades to form generalised models.
3 BP (Hull)
- The aim of this analysis
was to identify the potential of multivariate statistical process
control for optimising both production and control parameter values
for the A4 reformer process at the BP Salt End Site.
The main objective of this work was to stabilise and optimise
the production of CO by providing a deeper understanding of the
relationship between the NG composition and the CO produced. PCA
and PLS were initially applied to the data. A comparison was then
made between these individual monitoring models and those based
on the novel generic modelling approach using a pooled-covariance
matrix. The comparison was based on how quickly the disturbance
(N2 and CH4 in the NG) could be detected. The success of this
work is reflected in that a working process performance model
has been developed for the A4 reformer process and is to be validated
at-line using a real-time, on-line MATLAB based performance monitoring
tool-kit and subsequently using MDC’s MSPC+® as
part of CPACT Phase II.
Overall the work in Project
6 with BP has shown that multivariate statistical projection
based data analysis and data visualisation techniques are powerful
tools that can provide additional insight into the manufacturing
process and manufacturing issues. These complement the experience
of the plant engineers, at the same time enhancing process supervisors
and operators involvement in the drive towards manufacturing excellence.
4 Pfizer
- With increasing global
competition, it is of critical importance that each fermentation
produces high quality product at maximum yield. The Pfizer challenge
was to improve the consistency and yield on a large-scale fermentation
process. Initially a sensitivity analysis was carried out to identify
the key variables that determine process behaviour. The results
were confirmed from process understanding. Two strategies to the
monitoring of a batch type process were then investigated, multi-way
principal component analysis (MPCA) and batch observation level
analysis. Both methods identified those batches that had low yields.
Of particular interest was the fact that the two approaches identified
different batches as non-conforming. This is not surprising as
the process is particularly complex as well as being highly regulated.
The analysis showed that MPCA and batch observation level analysis
was complementary in terms of process performance monitoring.
A further outcome of this work was the conclusion that the data
visualisation methods developed are appropriate for use by fermentation
personnel as process deviations can be highlighted and dealt with
immediately, thereby ensuring that the process is returned to
good operational conditions with minimum loss of yield. The success
of this work is reflected in the development of an on-line performance
monitoring scheme in MATLAB for evaluation on the production process.
- A second study addressed
the Pfizer waste water plant with the aim of demonstrating the
potential of multivariate statistical data analysis and visualisation
techniques to provide an enhanced understanding of an activated
sludge wastewater treatment plant.
5 Zeneca (Alderly
Edge)
- A major challenge
in the development of new bio-therapeutics is the extraction of
hidden information in data collected during the innovation process.
A number of statistical feature extraction techniques were investigated
including linear and non-linear Principal Component Analysis,
Factor Analysis, K-Means Clustering, Fuzzy C-Means, Projection
Pursuit and Hierarchical Clustering and the Kohonen Self-Organising
Map. The data clustering methods were applied in an exploratory
manner to data gathered from a series of experiments under known
conditions on shake-flasks and laboratory scale fermentations.
None of the methods investigated were found to significantly outperform
the others and the cluster results were consistent with one another.
In terms of ease and clarity of interpretation, Factor Analysis,
slightly outperformed the other approaches. It is believed that
these tools, used in conjunction with biologists’ knowledge, can
offer major benefits to laboratory scientists in terms of improved
understanding of experiments during the design of new bio-therapeutics.
The information extracted was shown to enhance the understanding
of these complex processes and help biologists optimise process
development.
6 SmithKline Beecham
- The early detection
and monitoring of the presence of impurities that can form in
reacting mixtures is of key importance in the bio-chemical, fine
chemical, polymer and pharmaceutical industries. The study with
SB identified a number of challenges that require to be addressed
to isolate the key factors that potentially contribute to the
presence of process impurities. Multi-way PCA was applied to batch
data collected on a drug intermediate. A number of interesting
features were identified and addressed leading to an enhanced
understanding as to the cause of increased impurity levels. The
inclusion of agitator speed resulted in clustering related to
different speeds causing the monitoring model to be insensitive
to the detection of subtle process events. By applying a recently
developed multi-product modelling approach, this issue was addressed.
Through the application of a novel data augmentation approach
more subtle features in the data were extracted, allowing a more
representative model of the process to be constructed. The study
succeeded in providing an enhanced understanding of the process
and the identification of the key parameters and operational procedures
related to reactant impurity levels. The success of this work
has lead to the award of an EPSRC Mathematics CASE PhD studentship
with SB.
- Biological wastewater
treatment plants are increasingly being used for the on-site processing
of industrial effluent streams prior to discharge. The application
of multivariate statistical techniques to these processes is an
attractive solution to achieving a deeper understanding of their
complex operation, thereby achieving enhanced operational performance.
The aim of this work was to demonstrate the applicability of multivariate
statistical methods for the modelling of the process outputs and
monitoring of the performance of an industrial activated sludge
wastewater treatment plant. Nominal models based on normal operating
conditions were built. These successfully detected faulty process
operation. Satisfactory one-step ahead predictions of important
process output variables were obtained through dynamic PLS models.
Improved modelling and the identification of those particular
plant areas (blocks) responsible for faulty operation were achieved
using multi-block PLS. Such results can be used as a feedback
to process engineers to enhance process knowledge, reduce the
time taken to identify the causes of process problems and abnormal
operations as well as identifying important variables that might
be additionally monitored. The success of this work is reflected
in SB wishing to move towards monitoring the process on-line in
real time as part of CPACT Phase II.
7 Quest International
- Quest International
operates a fed-batch yeast propagation process. This involves
growing Saccharomyces cerevisiae on a complex substrate
using a relatively sophisticated control system. A high degree
of process knowledge is required to understand and control the
interactive response of the yeast to the substrate matrix such
that the desired yeast characteristics match required quality
parameters. Current understanding by Quest personnel is extensive
but due to the complex interaction of matrices, problems with
quality parameters or yield results can take too long to track
to source. The multivariate statistical projection techniques
of principal component analysis and projection to latent structures
were applied to extract the information inherent within the data.
The benefits of the technology were seen in the recognition of
a number of previously unidentified features. For example differences
were identified between the two fermenter houses, these were traced
initially to capacity differences and then more subtly, to their
respective control systems; clustering occurred between recipes
which was attributable to recipe variability and for certain recipes,
differences in recipe adjustments; and fermenter vessel and molasses
effects were identified for individual recipes. Overall the potential
of the methodology to explain process variability was realised
and new insights into variable interactions was achieved. This
study resulted in increased process knowledge and understanding.
8 Software Development
- MATLAB software developments
in data pre-screening (PreScreen) and multivariate
statistical process modelling and control (MultiDAT)
to Project 6 members under a University Evaluation Agreement which
does not allow their use for commercial gain. BatchDAT,
a Batch Process MSPC package is in the final stages of validation
prior to being available to CPACT companies.
Manufacturing
Excellence and CPACT
At recent meetings of
Projects 5 and 6, views were expressed that there is a strong interdependence
between Projects 5 and 6 and parts of Project 2. Through their conjunction,
it is possible to tackle "plant-wide manufacturing problems".
The "3P" issues of Product development, Process development
and Product manufacturing (vertical integration) combined with academic
research and in association with analytical and control systems
providers provide access to new technologies across the "3P"
industrial problems (horizontal integration).
In this respect the CPACT
contributions come from:
- The importance of
using multivariate statistical methods for data exploration, feature
extraction and to obtain an enhanced understanding of the process
(Project 6).
- The development of
advanced linear and non-linear multivariate statistical methodologies
(Project 6).
- The application of
plant-wide process control and optimisation, as opposed to single
unit optimisation (Project 5).
- The need to make real
use of more detailed information on the chemical and physical
properties of the material being processed (Project 2).
- The need to be able
to develop process models and calibration models from small data
sets (Project 2).
- The importance of
being able to monitor the whole manufacturing route from product
inception and development, through pilot studies to final product
manufacturing and customer feedback (Project 6).
It was suggested the
CPACT company members all had these issues in one form or another
– i.e. they were generic and CPACT was unique in being able to address
these with a range of tools capable of being used to provide generic
solutions.
The Japanese manufacturing
engineer will first look for a statistical tool to identify the
cause of the problem. The European Process Engineer will attempt
to solve the problem by modeling from first principles. CPACT is
unique in the UK, and possible Europe, in that it embraces both
of these approaches
Projects 5 and 6 recommend
that CPACT embraces the best of the Eastern and Western approaches
to problem solving. This can be achieved through the establishment
of an industrially led generic programme focussing on contributions
to "Manufacturing Excellence".
Project Objectives
and Deliverables
- To demonstrate the
beneficial application of MSPC in process manufacturing and in
pilot plant developments. (Delivered).
- To provide a contribution
to the better understanding of process operations. (Delivered)
- To predict the final
quality of the product from the more rapidly monitored process
variables. (Delivered).
- To provide showcase
demonstrators that will illustrate the power of the techniques
across a range of application sectors. (On-going).
- To establish a protocol
for the implementation of MSPC. (Delivered outside CPACT).
- To explore the need
for, and potential applications of, non-linear multivariate statistical
process monitoring. (Delivered outside CPACT).
Document Ref: 01/P5/P6/Execsum
Issued: 26 February 2001
|
