3.AVEVA Predictive Analytics Model Building Training

ASSET PERFORMANCE MANAGEMENT
AVEVA Predictive Analytics

Model Building Training
Gopikrishnan
© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

Model Building Introduction

Model Building Process
Test
Determine Review Customize Build Models Fine-tune
Scope Point List TDTs Models and Models
Deploy
Preparation Outside of AVEVA Predictive Model Building inside AVEVA Predictive

Analytics Analytics Client

Model Building Inputs Historian Tag List must contain Tag ID
(as it appears in Historian), Tag
Description and Engineering Units (C,
kg/s, bar….)
Tag Codification Glossary if not

Tag List Industry Standard (eg KKS)
P&IDs for all assets of interest
Also Useful: Plant

P&IDs Overall System
Plant
Diagrams, DCS Alert
Lists

DCS Screen Shots
Model Design Preparation
• TDT (Template Design Tool) spreadsheets are built by SMEs to define AVEVA Predictive Analytics models by a
collection of Metrics associated with a particular asset.
• Templates are build around a logical collection of engineering parameters.
• A Process Model would track flows, temperatures and pressures,
• A Mechanical Model would track vibration, bearing temperatures etc.
• Tag names, descriptions and units are assigned from the master tag list.

Historical Data Format for Offline Files
Note!! Header should contain 5 rows:
Point ID, Description, Extended ID,
Extended Description and Units Historical Data is requires to train AVEVA
Predictive Analytics models. Typically data is
required for 12 months of operation at 1 hour
intervals over a time of normal healthy
running of the plant.
Data should be provided in .csv or .xls format
Another data set containing known faults is

also useful to test the model.
IMPORTANT: Live connections to historians

Columns contain Tag 1, 2, 3 … N
are preferred (and in fact required for online
monitoring). These .csv files are only used
Date / Time in US format:
when a direct connection to a historian is not
MM/DD/YYYY HH:MM:SS possible.

Worked Example : FD Fan

Overall System Understanding
• Thermal power plant boilers operate on a ‘balanced draft’ principal where fans
supply combustion air from the atmosphere via heaters to the furnace and induced
draft (ID) fans transport the resulting flue gas through the boiler and flue gas
treatment systems to the stack.
• This AVEVA Predictive Analytics example concerns a ‘Forced Draft’ (FD) combustion
air fan. In this case the fan is a centrifugal type similar to that shown below.
GAS SIDE B GAS AIRHEATER FURNACE

1 HLA22 1 HLA22
AA001 AA002
1 HLA12 1 HLA12 HV OFA REAR WALL FRONT WALL

AA001 AA002 FAN
Silencer
1 HLA12 1 HFE60 1 HFE60
AA101 AA002
MILL
AA003
F
FD FAN 1 HFE02 1 HFE02 1 HFE02 1 HFE61 1 HLE42
AA001 AA002 AA003 AA101 1 HHL41 AA101
1 HFE50 1 HFE50 AA101
AA101 AA002
MILL
E
1 HFE51
Silencer
1 HFE40 AA1011 HFE40
Hot Air Crossover Duct

AA101 AA002
MILL
D
PA FAN Tempering Air
1 HFE41
1 HFA23
AA101
1 HFE31 AA001
MILL
AA101
C
PA FAN 1 HFE30 1 HFE30
AA101 AA002
1 HFE21 1 HHL42
AA101 AA101
MILL
Silencer
B
1 HFE20
1 HFE20 1 HLE41
1 HFE01 AA002
1 HFE01 1 HFE01 AA101 1 HFE11 AA101
AA002 AA003
AA001 AA101
MILL
A
FD FAN 1 HFE10 1 HFE10
AA101 AA002
HV OFA
1 HLA11 1 HLA11 FAN REAR WALL FRONT WALL
1 HLA11
Silencer AA001 AA002
AA003
1 HLA21 1 HLA21
AA001 AA002
GAS SIDE A GAS AIRHEATER FURNACE
Typical FD Fan with inlet guide vane Typical ‘twin stream’ arrangement of combustion air
control system atmosphere to furnace for a large utility boiler
FD Fan Tags
Air from atmosphere
FSAB FSAA
POINT NAME DESCRIPTION UNITS
06T31 06I03 #1FD Fan Motor Current Amps

06G31 #1FD Fan Damper Position %
06G31
06GP34 #1FD Fan Outlet Pressure Bar
06T33-1
06T34 #1FD Fan Outlet Termperature deg C
06T32-1
06T31 #1FD Fan Inlet Temperature deg C
06T32-1 #1FD Fan Bearing Temperature deg C
M FAN
06T33-1 #1FD Fan Winding Temperature deg C
FSAA #1FD Fan Left Inlet Flow m3/hr
FSAB #1FD Fan Right Inlet Flow m3/hr
06I03 06GP34
06T34
Total 9 sensors / tags

Combustion air to boiler

Training / Test Data Files
Data 5th May 2012 through 31st Dec 2012 Data 1st Jan 2013 through 5th May 2013
at 10 minute intervals at 10 minute intervals

File Format
csv format (with , separator
not ;)
Point Name field is

mandatory
Other fields are optional

All tag values are
time aligned
Interpolated value
mm/dd/yyyy HH:MM:SS not average
. decimal point not ,

no 1,000 separator
• Note that this is the standard format when importing to AVEVA Predictive Analytics.
• Other custom formats can be configured.
• For real implementations, the training data would be retrieved directly from the historian,
with no need for these .csv files.
Model Building Steps

Model Building Steps in AVEVA Predictive Analytics Client
1 2 3 4 5 6 7
Define Project Set Filters Exclude Bad Create Set Warnings Test Model Deploy Model
Points and Data Operational and Alerts
Import Profile
Historical
Data

Model Building STEP 1 : Import Historical Data
• Open AVEVA Predictive Analytics Client with a new
Project
Right – Click Training Data
• In this example the data is imported from the csv
file. In this step both the Project Points and the
Training data are populated from the csv file.
Browse to ‘FD FAN

TRAINING DATA.csv’
Change the Time Zone to UTC –

(this avoids a conflict with
daylight saving time shifts) All Points are Selected
Select Import

Model Building STEP 1 : Import Historical Data
Right – Click Project Points Project Points dialogue –
details can be edited, order
can be changed etc
Project Points are
populated from the csv file
Training data is imported
Choose a Name followed by

the location in the Asset
Groups
Save the Project!

1 2 3 4 5 6 7
Import Profile
Historical
Data

Model Building STEP 2 : Set Filters
• Normally at least one filter is applied to switch the
Right – Click Filters
model on or off.
• In this case we will use the Motor Current to decide

when the fan is not running.
• The threshold value is set at 30 Amps.
• When the filter is active then the model will not

run (i.e. the OMR and deviations are set to zero and
no Alerts will be triggered). This avoids creating
nuisance Alerts when an asset is inactive.
Enter the filter description
Browse to the Fan Motor

Current
Choose a value of 30 Amps

Check ‘Apply filter to Profile
Generation’
Finally hit Apply and close

the filters dialogue
Add a New Filter

1 2 3 4 5 6 7
Import Profile
Historical
Data

Model Building STEP 3 : Exclude Bad Data
Right – Click Training Data
• The objective is to reduce the data to good, healthy
equipment behaviour.
1 Click Motor Current • Remove obvious anomalies, bad data (zero values
or flat lines) and outliers.
• First the data will be viewed as a time series trend.
• Where necessary, discuss with equipment

operators and maintenance personnel to identify
periods of bad operation which should not be
4 Change the Exclusions trained into the model.
drop-down to Hide – we
now only see the data
which is not excluded
2 Click Filters – this selects There is only a short period of

all the data which meets time where the fan is out of
the filter threshold service – the data is marked
by the vertical yellow
3 Click Exclude
highlight
This has now excluded the
data from the data set
which will be used to train
the model.

Model Building STEP 3 : Exclude Bad Data 4 The chart automatically
re-scales to show the
remaining data
Click through the points

and inspect the data
looking for obvious outliers
or bad values
The fan bearing

temperature has some
obvious bad values towards
the end of the data
1 Drag a rectangle around
3 Set Exclusion = Hide to
the data
automatically mask
excluded data
2 Click Exclude

4 It may be hard to see at
1 Select both the fan flows first, but this data also
by using Ctrl key while right looks suspicious
clicking
2 This data looks like bad

values: Drag a rectangle
3 Click Exclude around the data

5 Double Left-Click
anywhere on the chart
to Undo Zoom
2 Drag a rectangle around
this time period to zoom in
1 Change from Select

to Zoom
Tip: Right-click on
the chart and 3 We can confirm this is bad data
select Mark Data – it looks like the source data is
Points to indicate missing from 1st – 3rd September
4 Switch back to Select – drag a
the location of the and so the historian has
rectangle around this data and
data interpolated between the two
then Exclude
adjacent ‘valid’ points

Model Building STEP 3 : Add a Calculation Point
• For this example we have two flow measurements
and we would prefer to use just the total flow in
our model.
In the Project points • In order to do this we can add a AVEVA Predictive
dialogue select Add -> Analytics Calculation Point to sum the two flows.
AVEVA Predictive Analytics
Calculation
1 Fill out the Point Name,
Description and Units fields
3 Adapt the calculation code to 2 Fill out two variable names

return FlowA + FlowB FlowA and FlowB
Then assign the corresponding
points to those names using the
drop-down
4 Select Test
5 In the test table insert some

6 Click Save numbers to check the calculation
Model Building STEP 3 : Add a Calculation Point
2 The calculated total fan flow is
now added to the list of points.
Use the Ctrl key to display the
individual and total flows
1 Return to the Trend View tab
and then click the Refresh icon
on the top toolbar

Select Visual Comparison • We have cleaned obvious bad data points in the
from the Analysis Menu trend view, but most often the underlying
relationships in the data are not obvious in a time
Select Motor Current
on the X axis
series.
• Next, use the comparison view to look at the

relationship between pairs of points in our data.
Notice the data is
Select Total Flow on nicely correlated as
the Y axis expected but we have
some outliers
2 The algorithm detects outliers
based on the density of the data
points – experiment with the
1 Change Outliers to density threshold (change up or
Mark down)
Hide the data already

excluded 3 Click the blue Outliers link to
select the data and then click the
Exclude link
Motor Current
Damper Position
There a few data points here where it

appears the fan was running with the
damper nearly closed – drag a
rectangle around those points and
Exclude

Outlet Pressure
Fan Total Flow
Use the polygon tool to select these

outliers to the top of the data then
Select the Polygon Exclude
button
Click once to start

the polygon – and
then double click to
close
Note the time periods highlighted in yellow
Model Building STEP 3 : Exclude Bad Data on the Overview trend corresponding to the
data points that have been selected on the
Comparison trend
Inlet Temperature (i.e.

ambient temperature)
Use the Outliers Tool to Mark, Select,

then Exclude
Bearing Temperature
The bearing temperature correlates with

the ambient temperature (the spread
corresponding to change in load of the
fan)


Use the Outliers Tool to Mark, Select and

then Exclude
Winding Temperature
Notice we have two distinct groups of data –

the upper group corresponding to the time
period between August and October. In this
model we will include both groups.


Outlet Temperature
It appears the fan outlet temperature

measurement is faulty during most of the
data. The fan outlet temperature is expected
to be greater than the inlet temperature by
3-4C. In most of the data the outlet
temperature is actually lower than the inlet.
Consequently the fan outlet temperature
will not be used in the model.

1 2 3 4 5 6 7
Import Profile
Historical
Data

Model Building STEP 4 : Create Operational Profile
Right – Click Operational • Having cleaned the data, the Operational Profile
Profile
can be created.
• The Operational Profile is the name given to AVEVA

Predictive Analytics’s data cluster model of the
data.
Keep the default LSH

(Locality Sensitive Hashing) Right – Click FD Fan Demo
Each row in the OP Table is

one cluster of data in this
case with 7 dimensions
De-select Fan Outlet
Temperature, Left Flow and
Right Flow
Click OK to create the

Profile

Model Building STEP 4 : Add Constraint
• Constraining a point means that, for a new piece of
data, AVEVA Predictive Analytics will only select
clusters which match the value of that point.
• A deviation on the constrained point will only occur

if the value moves outside of the training range for
that point.
Re-open the Project Points • In this case we will add a constraint to Motor
dialogue and tick the Current – the driving parameter in this piece of
Constrain box against Motor
Current. Note that only one
equipment
point can be constrained.

1 2 3 4 5 6 7
Import Profile
Historical
Data

Model Building STEP 5 : Set Warnings and Alert Thresholds Right – Click Alert
Thresholds • Overall Model Residual is by default 5% Warning and
10% Alert
• Set absolute deviation Alerts on each tag as

appropriate for the engineering unit and magnitude
of the value being measured versus the magnitude
Check the Alert persistence of deviation to be detected
criteria
• Training data statistics (i.e. standard deviation) is
Expand the Absolute Signal useful as a guide to set Alert thresholds if you’re
Deviation section unsure what reasonable, actionable values would be
• Values can always be adjusted later when testing the

model
Apply Alert thresholds as

shown
OK to close dialogue

1 2 3 4 5 6 7
Import Profile
Historical
Data

Model Building STEP 6 : Test Model
Trend Menu -> Data • Test the Operational Profile on the historical data
Playback (or another test data set) using Data Playback
Currently the training data set • Verify that the predicted value tracks well (low
we have just created is being deviations) when the equipment is in normal
played back through the OP
operation, but that deviations increase (and create
including all data There are a few spikes on the Alerts) in periods of abnormal operation
OMR chart corresponding to
the data which has been
excluded • Use absolute deviation and ‘contribution’ to
understand which values in the model are the root
cause of anomalies
• If improvements are needed repeat steps 3, 4 and 5

adjusting training data selection and Alert
thresholds can easily impact model performance –
model testing can be an iterative process
Select Alerts = Threshold

and Mark

Import Test Data
Browse to ‘FD FAN TEST

DATA.csv’
Change the Time Zone to UTC –

(this avoids a conflict with All Points are Selected
daylight saving time shifts)
Select Import

Large deviation on
Large deviation on Magnitude of
re-start, but then
re-start, but fluctuations
remaining deviations
diminishes to zero – increases over time
Refresh and select the test indicating some issue
data good operation
Fan Off
(filter met) Fan Off
© 2020 AVEVA Group plc and its subsidiaries. All rights reserved. (filter met)
Model Building STEP 6 : Test Model Magnitude of
Large deviation on fluctuations
Large deviation on re-start, but then increases
re-start, but remaining deviations
Motor current diminishes to zero – indicating some issue
good operation
From April onwards there are

positive deviations in motor
There are various current eventually with spikes
ways to display the up to 20 Amps. Recalling that
results – here we the typical operation was in
will use Predictions the range of 35 – 45 Amps
and Absolute then this is a significant
Deviation increase. Note that since we
constrained motor current
then this positive deviation is
outside of the training range
Hide filtered data if

it is getting in the
© 2020 AVEVA Group plc and its subsidiaries. All rights reserved. way
Model Building STEP 6 : Test Model Magnitude of
Large deviation on Large deviation on
re-start, but re-start, but then fluctuations
Damper Position diminishes to zero – remaining deviations increases
good operation indicating some issue
From the re-start at the end of

February the damper position
shows a negative deviation
with some fluctuation just
before the end of the data

Model Building STEP 6 : Test Model Large deviation on
Large deviation on Magnitude of
re-start, but then
re-start, but fluctuations
remaining deviations
diminishes to zero – increases
indicating some issue
good operation
Total Flow
From the re-start at the end of

February the fan flow shows a
negative deviation with some
fluctuation just before the end
of the data

Large deviation on
Large deviation on re-start, but then Magnitude of
re-start, but remaining deviations fluctuations
Outlet diminishes to zero – indicating some issue increases
Pressure good operation
Positive fluctuation in outlet

pressure
Fluctuations reduce until just
before end of data

This bar graph shows which signals

contributed most to the OMR over the
timespan displayed. Zoom in on the trends to
see how these bars change. Click on a bar to
go to the trends for that sensor.
Click Signal
Contribution

Overall Diagnosis
• Damper control linkage became damaged during the period when the fan was off perhaps
during some maintenance. (maintenance time). The fluctuating relationship between the
outlet pressure, flow and motor current demonstrates that the Damper actuator is not in
proper control.
• At some point the damper may have become unresponsive and the motor is compensating
for the changes in demand causing significant spikes in current.
• AVEVA Predictive Analytics identified the issue around 3 months before the fan was
stopped.

1 2 3 4 5 6 7
Import Profile
Historical
Data

Model Building STEP 7 : Deploy Model
1 Select the Project
Points dialogue
3 Select the
Simulator Service Note the Real-Time
(has to be set up in Service is now changed
eDNA) and the Point Type
changes from Offline to
Historian
2 Select Deploy
Offline Points
4 Save Changes

Model Building STEP 7 : Deploy Model
1 Select Deploy 5 Notice that the deployment
Profile from the status indicator in the bottom
Deploy Menu 3 Select the right hand side now turns to a
appropriate model red cross
run frequency
2 Check the
Operational Profile
6 Save the Project
7 The deployment status

indicator in the bottom right
hand side now turns to a green
tick indicating the project is
now deployed and running
4 OK
8 View the results in AVEVA

Predictive Analytics Web

THANK YOU
linkedin.com/company/aveva
@avevagroup
ABOUT AVEVA
AVEVA is a global leader in engineering and industrial software driving digital transformation across
the entire asset and operational life cycle of capital-intensive industries.
The company’s engineering, planning and operations, asset performance, and monitoring and
control solutions deliver proven results to over 16,000 customers across the globe. Its customers are
supported by the largest industrial software ecosystem, including 4,200 partners and 5,700 certified
developers. AVEVA is headquartered in Cambridge, UK, with over 4,400 employees at 80 locations in
over 40 countries.
aveva.com

3.AVEVA Predictive Analytics Model Building Training

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

3.AVEVA Predictive Analytics Model Building Training

Uploaded by

Copyright:

Available Formats

ASSET PERFORMANCE MANAGEMENT

AVEVA Predictive Analytics

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

Preparation Outside of AVEVA Predictive Model Building inside AVEVA Predictive

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

Tag Codification Glossary if not

P&IDs for all assets of interest

Also Useful: Plant

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

Data should be provided in .csv or .xls format

Another data set containing known faults is

IMPORTANT: Live connections to historians

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

GAS SIDE B GAS AIRHEATER FURNACE

1 HLA12 1 HLA12 HV OFA REAR WALL FRONT WALL

Hot Air Crossover Duct

06T31 06I03 #1FD Fan Motor Current Amps

Total 9 sensors / tags

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

Point Name field is

Other fields are optional

. decimal point not ,

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

Browse to ‘FD FAN

Change the Time Zone to UTC –

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

Training data is imported

Choose a Name followed by

Save the Project!

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

• In this case we will use the Motor Current to decide

• The threshold value is set at 30 Amps.

• When the filter is active then the model will not

Browse to the Fan Motor

Choose a value of 30 Amps

Finally hit Apply and close

Add a New Filter

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

• First the data will be viewed as a time series trend.

• Where necessary, discuss with equipment

2 Click Filters – this selects There is only a short period of

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

Click through the points

The fan bearing

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

2 This data looks like bad

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

1 Change from Select

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

3 Adapt the calculation code to 2 Fill out two variable names

5 In the test table insert some

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

• Next, use the comparison view to look at the

Hide the data already

There a few data points here where it

© 2020 AVEVA Group plc and its subsidiaries. All rights reserved.

Fan Total Flow

Use the polygon tool to select these