Chemical Engineering June 2018 - 85

erything by itself. It can work with the
data available, but it cannot create
meaningful insight without sufficient
information. As mentioned at
the
beginning of this discussion, manufacturers
do not always measure everything
they should. The temperature
at a specific point in the reactor
might be a very important CPP, but
unless there is a temperature sensor
capturing readings at that point,
there will be no data.
Similarly, such analysis cannot
suggest hypotheses, nor can it draw
conclusions, but with these tools
in hand, the engineers carrying out
the analysis can begin to consider
what-if scenarios. This is where the
human element becomes important.
Process experts can advance
theories of what might be happening,
and then use the data-analytics
tools to see if experience supports a
given cause-and-effect relationship.
The advantage of today's tools is the
relative ease with which such theories
can be tested.
Extending our example, let's say
Product A is a crystalline substance
that must be precipitated from a solution,
beginning 14 hours into the
unit procedure. The crystallization
process tends to vary from batch
to batch, which affects yield, but no
one has been able to identify exactly
what is happening during the process
to account for this variability.
Engineers looking at the situation
have different theories:
* Jim argues that it relates to the
temperature at the start of the
crystallization phase - he says it
should be higher
* Anne thinks the rate of temperature
change is key, and the cooling
action should be faster
* Chris sees it as a concentration
problem caused by an incomplete
reaction at an earlier stage
Each of these suggested remedies
can be evaluated by performing
basic extractions from the larger
block of historian data, and by then
performing various data analytics.
For Jim, it is possible to look for
specific solution temperature values
at the critical point in the process.
Here are the situations where we
were bumping against the upper
limit against those where we're at
the bottom. How do those compare
with overall yield? Is it a clear and
76
direct relationship? Does it suggest
there is still another factor influencing
the outcome?
For Anne, the slope of the temperature
line is critical, and it is easy
to separate the steep from the slow.
Beginning solution temperature can
also be factored in, along with the
cooling water temperature and flow.
If we can cool more quickly, does it
really affect yield?
Maybe Chris is right and we aren't
watching the earlier reaction closely
enough. What about concentration
at the beginning of the crystallization
phase? Does it have a significant
enough effect on yield to re-examine
how it works?
Testing these theories and even
hunches against the data requires
the ability to look at aggregate effects,
with the option of drilling
down into the details when necessary.
Comparisons must be easy to
make, even when multiple factors
are involved. Our group of engineers
may begin by looking at the unit procedures
with the highest yields, and
then back into the other variables. Or
the approach might be to test each
theory individually or in combination.
The software should be able to perform
regressions and other analytics
to show where correlations exist and
where they do not.
Jim, Anne and Chris may all be
right - the crystallization process
needs to start at a higher temperature,
but be brought down as quickly
as possible. The upstream process
may also need to be tweaked. Once
the unit procedures where those attributes
can be identified have been
examined, it's a simple matter to
make the necessary adjustments to
the process. The fix is easy; it's finding
the problem that's hard, unless
the right tools are used.
To make this practical, all the raw
data must be accessible by the dataanalytics
tool, and the mechanisms
to do the extractions and comparisons
must be easy to implement.
Data-analytics tools have come a
long way in this respect, by offering
more intuitive interfaces using dragand-drop
capabilities, rather than requiring
users to write equations and
macros. It is now easier than ever to
find the areas of interest in the data,
and to perform the kinds of analysis
necessary to test theories.
The human element is critical
Naturally there are limitations. Even
the best analytical tools are no better
than the quality and consistency
of the data being analyzed. The kind
of analysis discussed so far works
great with clean data. But that often
is not available. Real-world process
data have gaps, process values flatline
at times, process instruments
can go out of calibration or a given
variable might be missing entirely for
a unit process. Analytical tools need
to help engineers work with the data
as it is, providing mechanisms to
account for and work around
these issues.
Consistency is essential. In our example
where we are looking at 60 unit
procedures, all the variables must
have been measured and recorded
in the same way in every case. If the
solution temperature sensor was
moved from one part of the reactor
to another before Batch 28, it might
affect the measurements, thereby
reducing the confidence in any conclusions
built on that attribute. These
things happen, and analytical tools
should not be stymied when small
obstacles are encountered.
And most importantly, process
engineers and experts are needed
to advance theories and interpret
conclusions.
Automated
analysis
can find correlations, but will seldom
identify cause-and-effect, particularly
for complex processes. The
solution is to pick the right data-analytics
tool and put it in the hands of
those with a deep understanding of
the process.
n
Edited by Suzanne Shelley
Author
Jon Peterson is co-founder of
Seeq and serves as its senior vice
president of Product & Customers
(1301 2nd Ave, Seattle, WA
98101; Phone: (206) 801-9339).
He began his career as a process
engineer and software developer
for process industries, spending
21 years at software developer
OSIsoft (then called Oil Systems,
Inc.), where he held leadership roles as vice president
of Engineering and vice president of Marketing, working
in a wide range of areas, including asset optimization,
energy management, advanced control and process
equipment design. While at OSIsoft, he was
instrumental in the architecture, design and coding of
the PI Server. The PI Server 3.0 is recognized as the
industry standard and has nearly 15,000 installations
worldwide. Peterson holds a B.S.Ch.E. from Northwestern
University.
CHEMICAL ENGINEERING WWW.CHEMENGONLINE.COM
JUNE 2018
http://WWW.CHEMENGONLINE.COM
Chemical Engineering June 2018

Table of Contents for the Digital Edition of Chemical Engineering June 2018
