Proting from plant data

Predictive analytics is an evolving technology with many potential applications

in and gains for the process industries

DOUGLAS WHITE
Emerson Automation Solutions

ypical plant problems include:

T 1. You are a process engineer

and suspect that the catalyst
in one of the sites reactors is deacti-
Previous Today

Status information
vating faster than it has historically.
Is your suspicion correct, and if it
is, how much faster and what is the
cause? Will the catalyst last until the
next shutdown? If not, what operat-
Diagnostic information
ing changes does the plant need to
Travel deviation Valve signature
make to ensure it will last? Cycle counter Step response
2. You are a reliability engineer try- Air supply pressure Dynamic error drive
ing to understand why some pumps signal ...etc.
Output signal
need frequent overhauls and some
do not. What are the characteristics
of those that repeatedly require this
service? Are there leading indica-
tors that will allow you to identify
beforehand when a pump problem
is about to occur?
3. There have been several signif- Valve position
LFDQW DUH UHOHDVHV RYHU WKH SDVW
few years with associated visible
plumes. The source and composition
of the releases are not obvious. You
KDYH EHHQ DVNHG WR QG WKH FDXVH
and recommend changes to avoid or Figure 1 Valve feedback output data
at least reduce the incidence of these
releases in the future. In recent years, rapid progress has most recent three years of historical
What is the common factor in all been made in innovative algorithms one minute sampled data for these
of these issues? Their resolution is and approaches. How are the pro- points.
going to require deep analysis of cess industries taking advantage Even more data can be expected
plant data including data beyond of these new techniques today and from process plants in the future.
standard process measurements. how might they apply them in com- Much of the new equipment and
Another commonality is that time ing years? devices purchased for process
is an explicit variable and solu- In a related trend, process plants plants is already equipped with
tions should generate predictions of are producing more and more data. multiple sensors to monitor inter-
future plant and equipment behav- Big data is a popular topic and the nal conditions and performance,
iour that supports decisions. How process industries are no exception. embedded computing to analyse the
can these questions be addressed 2QH UHQLQJ FRPSDQ\ LQGLFDWHG LW data produced, and enhanced con-
PRUHHIFLHQWO\PRUHTXLFNO\DQG produced 80 billion data items for nectivity to transmit the results. This
with less manpower requirements? storage in one year from four sites. continuing evolution in the num-
In this article, some of the new Another process company refer- ber of sensors and their computing
methods that are available to assist enced a corporate historian with 10 and communication capabilities is
with this data analysis and develop- million tags across 15 sites and plans sometimes referenced as the
ment of solutions will be reviewed. to implement online access to the Industrial Internet of Things (IIoT).

www.eptq.com PTQ Q1 2017 61

 McKinsey3 has ranked big data
analytics as one of the top potential
What is
forecast to
technologies that can increase pro-
happen ductivity and GDP over the next
Predictive
few years. Among business sectors,
analytics manufacturing was the lead with
What is an estimate of overall GDP yearly
happening increase of $125-$250 billion.
Real-time 'DWD DQDO\WLFV FDQEH FODVVLHG LQ
analytics three categories based on the time
Future
scale of interest as illustrated in
What has
happened Figure 2.
Descriptive/historical analytics
Descriptive
analytics is analysis of what has happened,
Present when it happened and why it hap-
pened. This includes producing ret-
rospective performance measures.
An example is calculation of actual
Past versus planned energy usage for
plant equipment as well as the over-
Figure 2 Analytics time line all site for the past day, week or
month.
But how can plant staff extract value practice of capturing, organising, Real-time analytics is the use of
from all of this data? and analysing data to determine the most recent and historical data
Figure 1 shows one example. In patterns, correlations, and conclu- to interpret current plant conditions.
years past, there might be one (or sions. Engineers, of course, have An example is comparison of plant
even zero) feedback signal from a been doing analytics, in practice, for operating conditions against the
HOG FRQWURO YDOYH 7RGD\ WKHUH FDQ as long as they have been employed. approved operating window and
be more than 100 including sophis- However, the continuing increase issuing an alert if the operating win-
ticated diagnostic information and in computing capabilities has led to dow is being violated.
generated graphical performance development of practical algorithms Predictive analytics is prediction
representations. WKDWFDQDGGUHVVVLJQLFDQWO\ODUJHU of what will happen and when it is
Of course the increase in data is data sets and more heterogeneous likely to happen based on an anal-
not limited to the process industries. data types. ysis of current and past data trends
Much larger increases are reported and patterns. An example is estimat-
in other industries. Walmart is ing the likelihood that a pump will
reported to collect and store 2.5 need servicing before the next shut-
pentabytes of unstructured data Refinery down. Predictive analytics is the
every hour from the more than focus of this article.
one million customers around the
Action
world currently purchasing items.1 Predictive analytics applications
Algorithms
Amazon is estimated to have more Figure 3 shows a typical predictive
than 1.5 million servers in their data analytics application in the process
centres.2 Acquire industries.
In this article the focus is on the Dynamic event detection Relevant sensor and event data
and sensor data capture
plant operational impact of these is captured in real time. They are
developments and their potential automatically analysed for patterns
business impact. One characteristic Analyse RI VLJQLFDQFH DQG LI RQH RU PRUH
of most of these operational issues is Extract information patterns is detected a prediction is
patterns
that they have a temporal/dynamic generated from the model. If appro-
character which leads to special data priate, an alert is generated or spe-
requirements. There has also been Prediction FLFFRPSHQVDWLQJDFWLRQLVLQLWLDWHG
an equivalent and important impact Models An example is the pump condi-
on other corporate activities such as tion question at the beginning of this
trading, customer and market anal- End users article. The spectral properties of
\VLV DQG QDQFLDO VHUYLFHV ZKLFK the vibration measurement from the
Present:
has been covered in many other HMI pump and the statistical properties
publications. Distribution of the outlet pressure measurement
PLJKW EH WKH LQSXWV RI VLJQLFDQFH
Data analytics If both of these variables increase
Data analytics is the theory and Figure 3 Predictive analytics implementation VLJQLFDQWO\ DW WKH VDPH WLPH WKH

62 PTQ Q1 2017 www.eptq.com

digitalrening.com is the most extensive It provides a constantly growing
source of freely available information on database of technical articles,
all aspects of the rening, gas and company literature, videos, industry
petrochemical processing industries. news and events.
 Common predictive model algorithms

Problems types Plant application Types of algorithms Types of models General application Comments
examples examples
Regression Predict energy use in Generalised linear Linear continuous Marketing campaigns Still the most common
plant based on planned regression variables output algorithm used
operation and models
Spectral data Neural nets Non-linear Image recognition Training requires lots of data
interpretation continuous variables
Video/photograph Convolutional neural Non-linear Fingerprint identication;
characterisation nets continuous variables photo sorting
Anomaly detection Fault detection; Principal component Linear continuous Human gene to
product quality tracking analysis (PCA) variables disease correlation
Classication Fault detection Logistics regression Binary outcome Yes/No Fraud detection
Sorting documents Decision trees Discrete outcome Recommend movies Random forest often most
random forest A,B,C,D to watch; music accurate but hard to interpret
Product quality prediction, Support vector Binary outcome Image recognition Handles high dimensionality
data preclassication machine (SVM) Yes/No and large data sets
Clustering Spectral data encoding; Nearest neighbour; Clusters similar data Recommend movies
characterisation K-means, to watch; music
Natural language Text sorting Text analytics Textual analytics Email spam
processing identication

Table 1

model might suggest a potential Statistical methods for regres- distinct groups within data sets with
future pump problem and suggest sion have been and remain the most similar characteristics. It can be used
action by operations and mainte- popular for plant applications with to sub-divide large data sets into
nance to address the problem. a further subdivision into those more manageable and meaningful
employing linear models versus sub-sets. Natural language process-
Predictive model development non-linear ones. Principal compo- ing is used to convert unstructured
Central to the success of any predic- nent analysis (PCA) is a subset of lin- text into structured indices that can
tive analytics project is the devel- HDUDQDO\VLVWKDWLGHQWLHVDUHGXFHG then be used for further analysis and
opment of an appropriate model set of derived variables from the to segregate the source text into rele-
which accurately forecasts the data set that approximate the entire vant sub-groups.
occurrences under investigation. set. These variables can be used to Most applications to date in the
Many of the popular algorithms for identify the normal operating region process industries have been in the
model building have been known to for a piece of equipment. Deviations equipment availability/reliability
statisticians for many years. What outside this region are often char- area with the system intended to
has changed is the ability to apply acteristic of faults. For non-linear warn of impending faults or of dete-
these algorithms to extremely large models, neural net algorithms have riorating performance. These appli-
data sets and widespread availabil- been gaining importance because cations will often use a combination
ity of easy-to-use software for these of their ability to model quite com- of the algorithms listed above.
applications. This has resulted in plex data such as photo recogni- For example, clustering can be
many recent applications in the con- tion. However, in contrast to linear used to identify groups of pumps
sumer area including such famil- regression, the resulting models with similar maintenance histo-
LDU RQHV DV QJHUSULQW UHFRJQLWLRQ have no easy physical interpreta- ries and these groups can then
on smart phones, recommenda- WLRQ &ODVVLFDWLRQ DOJRULWKPV DUH have individual predictive models
tions for music and movies from often used when the results are dis- with resulting higher accuracy than
previous selections, and driverless crete: binary such as yes/no, or mul- lumping all pumps together.
cars. Table 1 lists some of the major ti-class such as A,B,C or D. Logistics The methods discussed above are
classes of algorithms and their exist- regression is regularly used for primarily examples of supervised
ing and potential use in the process binary cases with if-then splitting model building. In this case there is
plant area and in more general con- (CART trees) often used for mul- a data set where the output is gen-
sumer applications. There is much ti-output problems and large data erally known and the desire is to
overlap in terms of potential appli- sets. Support vector machine algo- determine if the other variables in
cations and the approach used often rithms are popular for binary prob- the data set can predict the outcome.
depends on the preferences and lems with a large number of input An area of current research is what
experience of the user. variables and large data sets and is termed unsupervised learning
For more background on these non-linear boundaries. Clustering is where the system processes a large
algorithms see references 4 and 5. a technique to determine if there are amount of data to determine if there

64 PTQ Q1 2017 www.eptq.com

 break through
the

NOISE

Better Signal-to-Noise Ratio Means

Better Level Control Performance
While transmit pulse amplitude (signal size) has helped to make
guided wave radar technology the standard for accurate, reliable level
4.40 measurement, the fact is signal-to-noise ratio (SNR) represents a far
Model 706 more critical indicator of level control performance. For superior SNR
Higher SNR in all process conditions, no other GWR device beats the Eclipse
1.57 Model 706 transmitter from Magnetrol.
Competitor
To learn more about the breakthrough ECLIPSE Model 706
SNR
GWR transmitter visit radar.magnetrol.com or contact your
The ECLIPSE Model 706 transmitter has a MAGNETROL representative today.
signal-to-noise ratio nearly 3 times higher
than competitors.

magnetrol.com 1-630-969-4000 info@magnetrol.com 2016 Magnetrol International, Incorporated

 The other issue is that few of the
,X\PWTLU[ <UZ[Y\J[\YLK[L_[ 6WLYH[PVUHSYLHS[PTL analytic algorithms were developed
JVUKP[PVUHUK Daily operating plans HUKOPZ[VYPJHSKH[H to model dynamic systems, that is
WLYMVYTHUJL Targets Process systems that are time varying. As
Notes, etc. Offsites
9LSH[PVUHSKH[H Email Utilities emphasised here, that is exactly the
Maintenance Terminal data type of models that are needed for
Inspection Analysers (NIR)
Lab data
most plant problems. Fortunately,
Work orders 4VKLS there has been a recognition of this
...etc. I\PSKPUN =PKLV GHFLHQF\ DQG VXLWDEOH H[WHQVLRQV
:WLJ[YHSKH[H Flares
Process equipment of the algorithms have been and are
Rotating machinery
Vibration :WH[PHS ...etc. being developed.
Thermography GIS/location
Infrared/acoustic/ What are the expected benets?
ultrasonic
7R HYDOXDWH WKH EHQHWV IURP WKH
increased use of predictive analyt-
Figure 4 Typical plant predictive analytics data requirement LFVUVWFRQVLGHUWKHRSHUDWLQJREMHF-
tives of a typical large continuous
is any structure to the data or rela- torical data, process plants often process plant. The goals can be sum-
WLRQVKLSVZLWKRXWSULRUVSHFLFDWLRQ have residence times that are sig- marised as The Four Zeros:
of the desired answer. QLFDQWO\ ORQJHU WKDQ W\SLFDO GDWD Safety: zero safety incidents
VDPSOLQJ WLPHV )RU WKH UHQHU\ Sustainability: zero environmen-
Process plants and analytics crude unit shown in Figure 5DXLG tal incidents, excess energy use, and
Much of the work on analytics has element entering the unit can take waste
taken place outside the process one to two hours to exit. If product Availability and reliability: zero
plant environment. Why are plant quality is measured at the current unscheduled downtime
SUREOHPVPRUHGLIFXOW" time and you want to correlate pro- Financial optimisation: zero lost
7KH UVW DUHD RI GLIFXOW\ LV WKH cess conditions with the quality, you SURWRSSRUWXQLWLHV
character of the data required. Figure have to use the conditions at the time How can predictive analytics con-
4 shows typical data inputs for plant WKH XLG HOHPHQW ZDV WUDQVLWLQJ LQ WULEXWH WR PHHWLQJ WKHVH REMHFWLYHV"
predictive analytics problems. Note that part of the plant. To accomplish Although there are many possibili-
the diverse data types. this requires selecting the measure- ties, the most common areas are:
For the catalyst example, opera- ments from that part of the plant at
tional real time and historical data past intervals corresponding to the Improving safety performance
is required but also types of feed- residence time between that point Avoiding incidents through early
stocks and their analyses. To predict and the sampling point. This is detection of potential hazardous sit-
whether the catalyst will last requires called time shifting the data. If the uations multivariable equipment
data on the planned feedstocks, rates product was stabilised naphtha and operating window
and modes of operation. The pump the time of sampling is t, you might Reducing personnel exposure
example requires time synchronisa- have to use data from the stabiliser to hazardous conditions through
tion of operational data with mainte- column at t-10 minutes, data from enhanced detection of likely hazards.
nance records and use of equipment the crude column overhead system
data sheets and history. As men- at t-20 minutes, data from the crude Supporting sustainability
tioned earlier, time is an explicit var- heater at t-55 minutes, and crude Comparing current utilisation of
iable which means that all of these composition data at t-85 minutes to resources, such as energy and other
records have to be accurately time properly analyse the unit relation- utilities, to their expected usage
V\QFKURQLVHG )RU WKH DUH H[DPSOH ships. Identifying the actual delay under current conditions with a
how can the historical video footage times empirically from the data is determination of possible causes of
RIWKHDUHEHDXWRPDWLFDOO\VHDUFKHG FRPPRQO\RQHRIWKHUVWVWHSVWKDW variation
to identify when there was a visible has to occur in analysis. (DUO\ LGHQWLFDWLRQ RI SRWHQWLDO
plume and coordinate the incidents In addition, process data is environmental emission events to
with operational data and equipment natively of poor quality. It is often facilitate mitigation.
work orders at preceding time peri- corrupted by noise and has poor sta-
RGV" 0XFK XVHIXO LQIRUPDWLRQ PD\ tistical qualities. It can be auto-corre- Increasing plant availability and
be in unstructured text such as oper- lated, cross-correlated, collinear, and reducing unplanned downtime
ational logs that need to be analysed. non-stationary. Data cleaning is the Anomaly detection identifying
To further complicate the issue each term used to describe the necessary precursor events to unscheduled
of these data sources typically has and very important data preparation equipment outages or problems,
its own physical storage, data organ- step in analysis. It includes handling allowing required maintenance
isation, search programs and user missing data, removing outliers, activities to occur before the event
interface. scaling, compensating for non-sta- Performance monitoring for con-
For operational real time and his- tionarity, and so on. dition based and preventative main-

66 PTQ Q1 2017 www.eptq.com

 Stabiliser data at Stabiliser outlet
time (t10) mins sample at time t mins
Column overhead data Off gas
at time (t20) mins 9LMS\_
KY\T
([TVZWOLYPJ 5HWO[OH
[V^LY Z[HIPSPZLY

Light
naphtha
7\TWHYV\UKZ :[YPWWLY
JVS\TUZ
;

Heavy
; naphtha

Jet fuel /
; kerosene
Crude oil feed quality
at time (t85) mins
Diesel
;
7YLOLH[
*Y\KL Intermediate
; gasoil
+LZHS[LY
Heavy
gasoil

Steam
Desalter outlet data /LH[LY ;
at time (t70) mins
Heater outlet data at Atmospheric
time (t55) mins residue

Figure 5 Time shifting data

WHQDQFHGHWHFWLQJORVVRISURFHVV 'HWHUPLQLQJ UHDVRQV IRU SURGXFW DQG WKH HVWLPDWHG QDQFLDO RU
HTXLSPHQW SHUIRUPDQFH EHIRUH LW TXDOLW\\LHOGLVVXHV RWKHU EHQHWV :KLOH VRPHWLPHV
LPSDFWV SURGXFWLRQ FDSDFLW\ DOORZ- GLIFXOWWKLVLVDSUHUHTXLVLWHWRJDLQ-
LQJ IRU VFKHGXOHG UDWKHU WKDQ Typical analytics project LQJDSSURYDODQGVXSSRUWIURPVHQ-
XQVFKHGXOHGPDLQWHQDQFH :KDWDUHWKHW\SLFDOVWHSVLQH[HFXW- LRUPDQDJHPHQWZKLFKLVLPSRUWDQW
 5HGXFLQJ WKH LPSDFW RI FRUUR- LQJDSUHGLFWLYHDQDO\WLFVSURMHFW" IRUHYHQWXDOSURJUDPPHVXFFHVV
VLRQDQGIRXOLQJRQHTXLSPHQWZLWK
HQKDQFHGFRUURVLRQPRQLWRULQJDQG Problem denition Inventory and acquire available data
IRUHFDVWLQJ 7KH UVW VWHS LV WR GHYHORS D FOHDU ,GHQWLI\LQJ ORFDWLQJ DQG JDLQLQJ
 5HGXFLQJ WXUQDURXQG GXUDWLRQ LGHD RI WKH REMHFWLYH RI WKH SURMHFW DFFHVV WR DYDLODEOH GDWD LV XVXDOO\
DQGIUHTXHQF\WKURXJKEHWWHUNQRZO- :KDWGR\RXZDQWWRXQGHUVWDQGDQG WKHQH[WKXUGOH:KLOHSURFHVVGDWD
HGJHRIDFWXDOHTXLSPHQWFRQGLWLRQV SUHGLFW" +RZ ZLOO UHVXOWV EH XVHG LQ D KLVWRULDQ PD\ EH HDVLO\ DYDLOD-
WR LPSURYH EXVLQHVV SHUIRUPDQFH LQ EOHLWLVVHOGRPDOOWKDWLVUHTXLUHG
Increasing margin and WKHSODQW":KDWDUHWKHVSHFLFSHU- )RU WKH H[DPSOH DULQJ HYHQWV
productivity IRUPDQFH PHDVXUHV RU PHWULFV WR EH PD\EHGRFXPHQWHGLQHQYLURQPHQ-
 (QHUJ\DQGXWLOLWLHVGHPDQGIRUH- XVHG" )RU WKH DUH H[DPSOH OLVWHG WDO UHSRUWV SODQW RSHUDWLQJ LQVWUXF-
FDVWLQJ  SUHGLFWLQJ SRWHQWLDO FRQ- DW WKH VWDUW RI WKH DUWLFOH WKLV REMHF- WLRQVLQGDLO\LQVWUXFWLRQGRFXPHQWV
WUDFWSHQDOWLHVEHIRUHRFFXUUHQFH WLYH PLJKW EH LGHQWLI\ WKH FDXVHV DQGXQXVXDORSHUDWLQJVLWXDWLRQVLQ
 'HWHFWLQJDQGGLVVHFWLQJFRPSOH[ RI UHFHQW DULQJ HYHQWV LQ WKH SODQW ORJ ERRNV 0DLQWHQDQFH LVVXHV RQ
LQWHUDFWLQJ FRQVWUDLQWV VRPHWLPHV DQG SRVVLEOH PLWLJDWLQJ VWUDWHJLHV WR HTXLSPHQW PD\ EH LQ ZRUN RUGHU
PXOWLXQLW RQ SURGXFWLRQ WR KHOS DYRLGWKHPLQWKHIXWXUH UHFRUGV
RYHUFRPHERWWOHQHFNV
,QFUHDVLQJ WKH \LHOG RI PRVW YDO- Cost and benets estimation Data cleaning, consolidation and
XDEOH SURGXFWV E\ LGHQWLI\LQJ RSWL- 1H[W LW LV QHFHVVDU\ WR GHYHORS DQ transformation
PXPRSHUDWLQJFRQGLWLRQV H[SHFWHG FRVW IRU WKH SURJUDPPH ,W LV RIWHQ WKH FDVH WKDW JDWKHULQJ

www.eptq.com PTQ Q1 2017 67

 and preparing the data for analysis racy and increase uncertainty in the
consumes from one-third to one-half model predictions. Accuracy should
the total project hours. What data be audited regularly and model
should be collected? How should Plant 1 parameter updates performed when
Surge
it be stored, particularly non-nu- basin
appropriate.
meric data? How should search Plant 2 Current data has to be cleaned and
and retrieval of data of interest be transformed prior to use by the pre-
Plant 3
implemented? diction model just as the training data
Next is the data cleaning phase, Plant 4 was. This procedure is automated in
Denitrification
as discussed above. Data needs to be basin the production environment.
brought to a consistent format with Plant 5
a uniform timescale. Missing data Project review
Plant 6
and outliers need to be excluded. At the conclusion of the project the
For numeric analysis, scaling is Plant 7 actual return should be evaluated
Clarification
often required. The better the train- basin against the initial objective. What
ing data, the better the model. Plant 8 ZDV WKH DFWXDO QDQFLDO UHWXUQ YHU-
)RU WKH DUH H[DPSOH GDWD WLPH sus the investment required? What
Plant 9
synchronisation is likely to be are the lessons learned and best
important and consistent data for- Plant 10 practices?
Filtration
matting an issue.
Plant 11 How to get started
Model building and evaluation Plant 12
,I \RX DUH FRQVLGHULQJ \RXU UVW
As discussed, there are many algo- application, how do you get started?
rithms that can be used for model Generally a phased approach that
Settling pond
building. Before embarking on com- LV WDUJHWHG DW D VSHFLF OLPLWHG
plex models and analysis it is often Outfall
but important problem that can be
worthwhile to start with visual- addressed in a few weeks or months
isation and simple analysis with Figure 6 Wastewater treatment plant is better than a big bang type pro-
linear models, to see if any rela- gramme. A pilot project that gen-
tionships seem to exist. Confusion there may be relatively few actual HUDWHV VRPH EHQHWV FDQ KHOS EXLOG
between correlation and causation is events from which to build a model. support for more extensive future
common and needs to be addressed. This typically requires special algo- programmes. Finding an executive
At the end of the project, the model rithm approaches. sponsor who is interested in the
has to make sense for it to be used programme and the results is much
ZLWKFRQGHQFH Model deployment, monitoring and better than conducting an isolated
Previously, use of the various pre- updates study. Experienced consultants can
dictive analytics algorithms required For the model to be useful it needs help.
VLJQLFDQWH[SHULHQFHDQGH[SHUWLVH to be moved into the daily deci-
More and more easy-to-use software sion making process. Who will take Case studies
is becoming available which allows action based on the results of the pre- 1. Reducing efuent emission
those interested to build analyti- diction? What are their needs for a violations
FDO PRGHOV ZLWKRXW UVW EHFRPLQJ user interface? Standard questions A US chemical company installed
an expert. These also often provide include: what infrastructure should ELRORJLFDO GHQLWULFDWLRQ LQ WKHLU
an infrastructure for implementa- be used?; what tools should be used wastewater treatment plant to
tion, which will be a requirement for for deployment?; should the data reduce ammonia in their water
eventual use. be stored on premises or off-site (in discharge (see Figure 6). The pro-
Testing and validation of the model the cloud)? As with model building, cess uses aerobic bacteria to con-
are the next step. Ideally there is there has been recent development of vert ammonia to nitrates. However,
enough data to segregate testing and a number of new software offerings the company experienced occa-
validation data sets from the training in this area. In contrast to the past, sional breakthroughs of ammonia
data. After the model is developed it is possible to deploy a solution that caused the water discharge to
with the training data, it is tested using these new infrastructures with- H[FHHG VSHFLHG HIXHQW DPPR-
against the test data set. If parameter out requirement for an experienced nia limits. To assist in identifying
changes are then made to the model H[SHUWRQWKHVSHFLFVRIWZDUH7KHUH issues, they decided to conduct a
the improved model is checked is a standardised documentation for data analytics study to see if the
against the validation data set and models Predictive Model Markup breakthroughs were associated with
the accuracy reported. This test- Language (PMML) that can facili- VSHFLF SODQW FRQGLWLRQV 2QH RI
ing and validation step is extremely tate movement from a development WKH LVVXHV LQ WKH LGHQWLFDWLRQ ZDV
important in developing an estimate environment to deployed status. the large residence time in the sys-
of the accuracy of the model. There are often changes in the tem, from one to two weeks for the
7KHDUHH[DPSOHLVDFDVHZKHUH plant over time that degrade accu- total system depending on individ-

68 PTQ Q1 2017 www.eptq.com

ual plant discharge rates, although centered maintenance and predic- bending-size-of-amazon-s-cloud.html
there were intermediate measure- tive diagnostics. The results were a 3 Bloom A, McKinsey on Big Data Analytics:
ments that did identify likely break- 12% decrease in maintenance costs The #1 Key to US Economic Growth, 2013,
through. Another was that there while achieving a 2% increase in blog.pivotol.io/pivotal/p-o-v/mckinsey-
were multiple plants discharging to overall plant availability. on-big-data-analytics-the-1-key-to-us-
economic-growth
the waste water system, each with
4 Hastie T, Tibshirani R, Friedman J, Elements
different product production sched- Conclusion
of Statistical Learning Data Mining, Inference
ules. More than 200 measured pro- In summary, predictive analytics is and Prediction, 2nd Edition, Springer, 2008,
cess variables were considered as an evolving technology with many Free download at statweb.stanford.edu/~tibs/
candidates and a thorough corre- potential applications in the process ElemStatLearn/
ODWLRQ VWXG\ RYHU YH \HDUV RI GDWD industries. However, implementa- 5 Rudin C, Prediction: Machine Learning
was performed using PCA. The con- tion in these industries has distinc- and Statistics, Course (free) under
clusion of the study was that a drop tive issues due to the time series MITOpenCourseware, ocw.mit.edu/courses/
LQ WKH S+ LQ WKH HIXHQW IURP RQH character of the data and associated sloan-school-of-management/15-097-
of the plants preceded many of the complex data requirements. Current prediction-machine-learning-and-statistics-
breakthrough indications by one to applications are demonstrating spring-2012/index.htm
two days. This drop in pH corre- EHQHWV LQ SURFHVV IDXOW GHWHFWLRQ
sponded to a particular type of pro- increasing equipment availability, Doug White is a Principal Consultant with
cess operation in the plant. improving safety performance and the Process Systems and Solutions Group of
VXSSRUWLQJ QDQFLDO RSWLPLVDWLRQ Emerson Process Management. Previously
2. Reducing maintenance, increasing There will be more of these applica- he held senior management and technical
positions with MDC Technology, Protpoint
plant availability tions in the future.
Solutions, Aspen Technology, and Setpoint
A large Middle East chemical com-
References where he was responsible for developing
pany commissioned a major new and implementing advanced automation
1 McAfee A, Brynjolfsson E, Big Data: The
site with over 6000 important Management Revolution, Harvard Business and optimisation systems in process plants
equipment assets requiring regular Review, Oct 2012. around the world. He has published more than
maintenance. After initial commis- 2 Clark J, 5 Numbers That Illustrate the Mind- 50 articles on these subjects and holds a BChE
sioning of the plant, an integrated Bending Size of Amazons Cloud, Bloomberg, from the University of Florida, an MS from
maintenance programme was 2014, www.bloomberg.com/news/2014- California Institute of Technology, and an MA
instituted centred on reliability 11-14/5-numbers-that-illustrate-the-mind- and PhD from Princeton University.

EXPOEUROPE
28 - 30 MARCH 2017 - AHOY, ROTTERDAM

Organised by

Register now qq www.stocexpo.com

www.eptq.com PTQ Q1 2017 69