Professional Documents
Culture Documents
Real Life
It’s really more than a demo… this EXAKT demo reveals the real data from a
client’s replacement analysis. The customer is Cardinal River Coals (CRC),
in Hinton, Alberta, Canada, a major producer of coal… and like all Coal
extractors, CRC uses costly, maintenance-prone heavy equipment.
The assets being analyzed, therefore, are worth $4,675,000. To say that
making the right replacement and repair decisions — deciding when to take a
unit out of service — is an important decision, is definitely an
understatement!
This demo will show you the steps taken to analyze CRC’s wheel-motor data,
and (with lots in between) will conclude with the impressive Pay-Off that
EXAKT delivered.
The Demo
Now… to the Demo. This Demo is designed to “walk” you, step-by-step,
through the sequences of EXAKT — from data input tables, through
modeling, displays, estimation… and finally to decision outputs, in everyday
language. We’ll show you what numbers went in, what happened to them…
and the results CRC obtained.
May we suggest you jot down any questions or comments as you go along.
Near the end of this demo we provide a link for an e-mail message to us for
the purpose of sending along your questions or comments. One of our
experts will promptly reply to you with answers to your questions, or
additional information that you may appear to require.
Now... let’s go to the Walk-Through demo, where, in just a few minutes you
can see how Cardinal River Coals’ data was handled, and the remarkable
results EXAKT produced.
You will see how both programs work, as we walk you through this Demo.
In everyday language: We’ll show you what numbers have to go in, what
happens to them... and the results you’ll get.
On request, a “Working Demo” is also available. It’s the real program, and
with it you can input a prepared database set, and navigate at will around the
entire software package. The Working Demo requires more time, but
provides hands-on experience with the program, and a live database.
This PDF Demo is the best place to start. If you wish to obtain a copy of the
Working Demo, contact us.
(IMAGE 1)
The pane on the left is a familiar file management and selection tool. Note
that our database has already been selected — CRC_WMOD.mdb. It is from
this panel that one can access EXAKT’s tables queries and programs.
This pane also includes the means to get to a variety of tables, most of which
will be explained later.
The pane on the right contains the Procedures flow-chart. This flow-chart —
and others like it — are “maps” to enable you to keep track of where you are,
and where you’ve been... and where you might be going.
This is the database we’ll be
using — the CRC Cardinal River
Coals data
(IMAGE 2)
This display provides an opportunity to get an overview of the entire sample and
to check for consistency and obvious anomolies. The Number of Histories, for
example, should equal the Number of Beginnings – beginnings are the start-
dates of histories. A history a lifetime record having a beginning, various events,
and an ending. The ending may be an actual failure or a preventive removal.
Mathematical Models
EXAKT uses database tables which it draws upon for its “models”. A model is
a mathematical representation of the relationships between the age-related
failure information, and the key condition indicators. EXAKT uses the term
“covariates” — the readings obtained from condition monitoring such as oil
analysis, vibration readings, thermography, current analysis, pressure
monitoring and so on.
To arrive at the ultimate Decision-Optimization Model we initially build
what is known as a Proportional Hazards Model (PHM). The EXAKT
program does the model building for you, but along the way the user will
contribute added information.
The PHM — or Proportional Hazards Model — provides the risk of failure as
a function of both... (1) time, and (2) the key significant condition indicators.
The Decision Model blends the risk of failure with the relative costs of a
failure. That blend produces a model that will generate the optimum decision
Here’s why.
EXAKT will provide you with many clues and in some cases informative
statements about the data requiring your attention. An example of a
data record in error is one whose working age for a given inspection is
lower than that of an inspection on a previous date. This is a common
error due to mis-transcribing the equipment's hour-meter. Other
common errors are missing "B" (beginning) events. EXAKT will point out
all such errors for you to correct. You will spend most of your modeling
effort in rooting out these errors, and in the process, becoming very
familiar, indeed intimate, with the data. Experienced maintenance
people should perform this task or be consulted often in this data
"cleaning" phase.
Inspections Table
DATE
Date of
observation
IDENT is the
identification number
of the item being
analyzed
(IMAGE 4)
You may then want to return to the Main Screen, and, to obtain a description
of your covariates you would click on VARDESCRIPTION.
(IMAGE 4)
(IMAGE 6)
Events Table
Let’s look at the Events table. To get to it, you would click on EVENTS in the
left pane.
(IMAGE 8)
(IMAGE 9)
You’ll note it has columns similar to the Inspections Table... but the focus
here is on the data in the Working Age and Event columns. This data is the
real data from Cardinal River Coals’ CMMS records.
The Events Table provides lifetime data which allows you to see how long
the item ran (the Working Age) before it failed or until it was “suspended”
(explained below). Other rows record other events such as an oil change
(OC) or minor repair.
Graphical Display
With the working version of EXAKT, to see a graphical display of the
inspections information you would go to VIEW on the Menu Bar and select
Inspections.
The Graphical Display is a most useful feature of EXAKT, and offers many
attractive options. It is a display that fully-synchronizes inspections data with
a graphical picture. For example, you can click on any inspection record in
the right-panel inspections table, and when you do, the graphical point
corresponding the record will be identified on the graph by a vertical line.
Or vice versa — you can click on any point on a graph and the information
corresponding to it will be highlighted in the inspection table to the right.
Various sizes, colors and other display options are at your fingertips as well.
By clicking on
any point on a
graph you may
highlight the
same value on
the table Clicking on any
point in the table
will highlight the
corresponding point
in the graph
(IMAGE 11)
Cross Graphs
Another of EXAKT’s powerful tools, the “cross-graph”, will reveal the data in
ways that will point out further possible problems or inconsistencies.
Furthermore the cross graph illustrates correlation between pairs of
numerical variables.
(IMAGE 12)
(IMAGE 13)
Modeling
To do this you would return to the Procedures Window, and click DATA
PREPARATION.
(IMAGE 14)
(IMAGE 15)
Once the General Project Data has been entered, a click on the OK button
will bring you back to the Procedures window, where you will click on the
INPUT DATA WITH COVARIATES button.
Click here
(IMAGE 17)
Once you do this the data from the Inspections Table and the Events Table
will be merged to create two new tables to be used in model calculations. In a
Data With Covariates is the sequence we are using for this demo. It
means that condition readings have been obtained and are available for
analysis.
After the data have been loaded, click on the MODELING button in the Data
Preparation for Modeling View.
Next…click on
this button
(IMAGE 18)
Your Data Preparation View will change to the Modeling View, and you would
click on the Weibull PHM button —
(IMAGE 19)
— to get to the Weibull PHM View. Here you would click on Select
Covariates
(IMAGE 20)
(IMAGE 21)
First, note that EXAKT requires a sub-model name. Why? It’s because the
user would likely look at several variations of a Proportional Hazards Model
(PHM) before finally settling on a particular one. Thus the models are
assigned different names to identify them throughout the modeling session.
In the dialogue box shown above, all the covariates are shown as “available”,
but unselected. In other words, no selection of any particular covariates has
been made yet. But to begin with actual modeling, it is recommended that all
the covariates be “selected” (i.e., moved into the selected box to the right).
(IMAGE 22)
With all the covariates now selected and in the right-hand box, you would
click the OK button. This initiates the model-building process.
What’s happening now is that the Weibull PHM (Proportional Hazards Model)
is being created through an intensive multiple regression procedure with the
purpose of finding significant correlations between condition data and event
data, particularly failure event data. Once the model has been created, a
report will appear — PHM Parameter Estimation.
This report shows the parameter estimates for the “All-Covariates” model for
that particular machine part, assembly or component. Two tables will be
shown on the screen.
(IMAGE 23)
A sample of 50 wheel motors was available for analysis. There were 176
“histories” related to these, since a given wheel motor might have more than
one history, due to its being repaired or renewed, and brought back into
service as “new”.
This table shows us that on 88 occasions the use of a wheel motor was
terminated by failure, and on 41 occasions the wheel motor’s life was
suspended. (See previous explanation of failures versus suspensions). This
information comes from the Events Table, described earlier.
Please focus on the “Sign. (Significance) column, where you will note “N” and
“Y” codes. “N” indicates that a given covariate is non-significant. The
recommended procedure is to de-select these covariates one-at-a-time in
This is done by closing the current Models Variables window, which will
display the Weibull PHM flow-chart, where, once again, the SELECT
COVARIATES button will be clicked.
This will bring you, again, to the Model Variables box where you will de-select
the covariate described above.
(IMAGE 24)
Once a model has been developed it is important to evaluate how “good” it is.
How closely does it “fit” the actual data from which it was generated in the
previous steps? If it does not fit well enough, the model should be rejected
since it will probably not provide maintenance decisions with an gadequate
level of confidence. The Goodness-of-Fit test provides an objective tool with
which to accept or reject the model at this point. To obtain this you would
click on Summary Report in the Weibull PHM flow-chart. The report looks
like this:
Before we leave this stage of the demo, let us refer to the right-hand
panel of buttons called “Residual Analysis”. Under Residual Analysis
you will see options that provide you with various graphical measures of
goodness-of-fit of the PHM we have decided to use. These options are
explained in the appendix to this Walkthrough Demo.
After closing the Proportional Hazards Model, you will come again to the
Modeling Weibull PHM View of the Procedures window.
(IMAGE 26)
(IMAGE 27)
(IMAGE 28)
What do the bands mean? These bands, or ranges, define the condition
measurement readings that can be interpreted as, for example, low,
medium, high and very high. That is, high contamination, low pressure,
medium voltage, low vibration — whatever the indicator is for the
condition that is being monitored.
In our demo, therefore, we have five bands for iron, and five bands for
Sediment. These are the only two covariates used in the final PHM.
We see the iron readings fall in the range between 0-to-2693 (see the
text at the bottom of the box) having intervals 0-50, 50-150, 0-400,
400-900, and 900 and greater. These are not default bands but were
selected on the basis of current practices at Cardinal River Coals and
reflect the severity of iron particles with respect to the condition of this
equipment.
EXAKT can help you decide whether or not the groups you set up are
statistically significant. This can be determined by clicking on the Test
Grouping button on the Analysis menu of the Transition Probability Model
Procedures Window.
For the purposes of this demo, we will not create and test covariate groups.
(IMAGE 29)
Click on Time Intervals. This option allows you to set a time-period for the
values you wish to analyze. One way this can be useful is in taking into
account that the behavior of the condition-measurements might vary during
different stages of the unit’s life. The early part of its life (the run-in period),
the normal working duration (the major part of the unit’s life), and the final
stage of its life (the wear-out period) may need to be analyzed separately.
This option gives you the opportunity of doing so.
Transition Probabilities
It is not necessary that the analyst use this information — it is available for
study, if necessary. Whether studied or not however, these “transition
probabilities” form a powerful part — together with the PHM — of the final
decision model.
(IMAGE 30)
(IMAGE 31)
Clicking on the DISPLAY MATRIX button brings up the Display Transition
Probabilities window.
From this window one may study the transition probabilities by using the
settings to display them in different ways. For example, we can ask EXAKT
to show us the probability of iron moving from any state to any others state.
Then we can set conditions to display iron’s transition probabilities, given
states of the other covariates. Once again, it is not necessary for the analyst
to study these transition matrices – but they are there for deeper probing, if
necessary.
(IMAGE 32)
You have the option of examining the transition probability matrix more
closely by considering also the influence of transitions of one covariate on
those of another. This is done by deactivating the radio button “ignore”. At
that moment you need specify a transition of the other variable.
If you deactivate the Ignore radio button, it means you want to see how the
transitions of one variable (covariate) affects another. For the demo, we will
provide no further detail here, but it is important to know that this option is
available.
Click OK. This will bring to the screen a report on the Markov Chain Model
Transition Probability Matrix.
64%
(IMAGE 33)
This is an example of a transition matrix for iron, where the other significant
covariate has been ignored.
Referring to the table above, we can see that the covariate under analysis is
iron, and that five levels of readings have been established. (That
establishment would have been done in the model-building phase we
covered earlier).
The five levels indicate the severity of the condition such as light
contamination to very heavy contamination. If the second level of severity,
where the readings are 50 to 150 indicate. say, “medium” contamination, the
table would tell you that there is a 64% (0.638362) probability of the
measurement being in the same range at the next inspection, a 19.3%
probability of it having worsened to the next higher level of severity (150-to-
400), a 3.3% chance of it having reached the next state (400 to 900) and a
0.6% of chance of it having reached the worst state (above 900) before the
next inspection.
Click on
Decision
Model button
(IMAGE 34)
Clicking Decision Model causes EXAKT to build its most important output —
the Optimal Decision Replacement Model. The process incorporates all of
the work that went on earlier. — the data-preparation effort, the PHM, and
the TPM…together with the cost information. The Replacement Decision
Model emerges. The maintenance decision-makers may now confidently
apply this model as a “watchdog” to monitor future condition data, and to
deliver the optimal decision each time.
Click here
(IMAGE 35)
(IMAGE 36)
It is already partly filled in from prior information when we were in the Data
Preparation and General Data phase, much earlier.
We have chosen VARIANT ONE, because the working reality of the equipment
studied at Cardinal River Coal was such that it was not possible to carry out a
repair or replacement except at regular inspection times. EXAKT will assume
that reality when it renders future replacement decisions.
In the “Cost” panel, 20 was entered as the replacement cost (parts and
labor). The value 60 for was entered for failure costs (parts, labor and
downtime). For the Cost Unit, $1000 was entered. This means that a
Planned replacement cost is $20000 and a Failure Cost is $60000. Note that
250 was entered for the duration of the inspection interval, and 500 for
regular maintenance event such as an Oil Change, corresponding to the
interval recorded in the Events Table.
(IMAGE 37)
This is a kind of “sensitivity” graph. In our demo, the flat shape of the line
shows that the replacement policy does not have to be strictly followed.
Costs are not greatly affected if the replacement time varies. The Cost
Function graph shows the minimum costs per hour we can achieve and the
cost per hour if we follow a failure-replacement policy.
Note the dip in the Cost curve. The deeper the dip, the more savings you
derive from using the Optimal decision model. The difference between the
low point on the curve and the dashed line equals the savings relative to a
run-to-failure policy.
You can also get at this graph by clicking Cost Function in the Analysis
window. For even more information, after opening the Cost Function Graph
you can click VIEW on the Menu Bar to choose VIEW FULL REPORT and get a
full report (including summary of cost analysis) along with the graph. You will
see this when you get to Image 40. Close the window.
(IMAGE 38)
(IMAGE 39)
Our publication, “The CBM Optimizer News”, explains the graph, this way:
In real situations, the actual ratio of failure and planned repair costs, may
not be well known. Furthermore the dynamics of industry are such that costs
can change with changing technology, production, and market conditions.
Therefore one would like to know, to what degree one's true costs and one's
optimal policy would change with changes in costs. Sensitivity analysis is
the method to determine this. The software generates a graph and
corresponding tabular data
For example, if the actual cost ratio is 5 and we are using a model which
was based on CR=3, then the increase in optimal cost incurred by following
that (wrong) policy is around 6%. In other words the red graph represents
the sensitivity of costs to changes in CR.
Blue Dashed Line — Again, assume the actual cost ratio has strayed from
what was assumed when the model was built. Assume we were to rebuild
the model using the new ratio. The blue line tells how much the new optimal
cost would differ from that of the original model. In other words the blue line
represents the sensitivity of the optimal policy to changes in CR. (Note that
the sensitivity graphs assume that only Cf (failure repair cost) changes and
Cr (planned repair cost) remains unchanged.)
You can also call up a full report for this, as with the Optimal Replacement
Decision Graph, by clicking on VIEW on the Menu Bar, and choosing VIEW
FULL REPORT.
Incidentally, these savings were calculated using a cost ratio — (the ratio of
the cost of a operating replacement versus a preventive replacement) — of
“3”. If the ratio had been “5” the savings would have been nearly 44%, and if
the ratio had been “6” the savings would have been 50%!
We have now built our statistical decision model, and we are confident we
can use it for decisions. This will eventually be “saved”. Close your window.
In the case of the Cardinal River Coals data, this will display a list of wheel
motor idents from which you can choose an item on which you would like to
make a decision. By clicking on Decisions you will actually run the model for
any of the equipment in the database. Later, by using only the EXAKT for
Decisions Program you will be able to apply the model to any unit for
"automated" monitoring. This will display the Replacement Decision graph.
Again, you have the opportunity of getting a full report if you wish.
(IMAGE 42)
For the everyday maintenance practitioner who just wants to get an answer
to a simple question, i.e., should we keep on running, or should we replace
right now?”, he or she will turn to the EXAKT Replacement Decisions chart and
immediately find the answer on a clear unambiguous graph.
EXAKT assumes that regular inspections are being carried out, so the
position of the right-most dot on the chart will provide one of three
possible actions related to the inspection interval policy.
If your dot is in the green region, then the decision is to continue operating
until the next scheduled inspection.
If your dot is in the yellow region — Replace Before Next Inspection — it’s a
warning signal which says although you’re getting “close” to a possible failure
event, the risk of failure combined with the benefits of running a while longer
means that you may be able to delay replacement. It’s an invaluable help in
planning because it allows you to forecast the maintenance work-load and
down-time eventualities in the near-term.
If your dot is in the red area, EXAKT is telling you to Replace Immediately
because the risk of imminent failure is very high.
The answer is that users have suggested they wanted a portion of EXAKT —
the modeling portion — to be un-available to persons other than the
modelers in the organization. This way, decision-makers — mangers,
planners, schedulers, maintenance supervisor, and so on — can freely use
EXAKT with no danger of altering or destroying the model. EXAKT for
Decisions, uses the established models for future decision making, but
leaves their originals intact.
Users of the “EXAKT for Decisions” program can generate all the output that
really matters. For example they can generate the Replacement Decisions
Graphs as already illustrated. They can also generate a useful Report that
displays all the assets, individually or grouped by, say, location, or type.
Here’s an example of such a report as it pertains to Cardinal River Coal.
(IMAGE 43)
Earlier we suggested you note any questions that arose as you walked
through the demo. Here, now, is your opportunity to send them along to us.
Please include, at the beginning of your message – Your Name, Job-Title,
Organization, Telephone-Number, and Geographical Location (address not
necessary). We look forward to hearing from you. To e-mail us, click here.
Please note our e-mail address is expertquestion@oliver-group.com.
The Estimates are used in the Proportional Hazards Model. The PHM is
as follows:
β −1
βt γ 1Z1 ( t ) + ... +γ p Z p ( t )
h( t ) = e
η η
where t denotes time, Z1 (t ) ,..., Z p (t ) are p condition variables and
β , η and γ 1 ,..., γ p are parameters to be estimated from the data.
To show how the values in the estimates column are used in the PHM
model, we will now substitute these values in the model, above.
3.693−1
3.693 t
h( t ) = e 1.337 nerf ( t ) +1.701 pitation( t ) +.0107 gobbing ( t )
18230 18230
Sign — Indicates whether, from a statistical viewpoint, one or more (or
perhaps, none) of the parameters — (one in this case — gobbing) — is
not statistically significant in the context of estimating risk based upon a
PHM..
Standard Error — The values here are used to establish whether or not
the estimates in the second column are statistically significant.
Wald — The values here are obtained from the values in the Estimates
and Standard Errors columns through a transformation process.
DF — Degrees of Freedom. Tells us the Wald statistic is distributed
through a chi-square distribution with one degree of freedom.
p-value — This column is obtained from tables of the Chi-square
distribution and indicates whether the Wald statistic is significantly large.
If it is, the p-value will be small, demonstrating that the previously-
calculated estimate is significant, and a Y (for “yes”) would be produced
in the 3rd (Sign) column.
exp of est — Exponent of the Estimate... Along with the associated 95%
confidence interval on that exponent, provides the statistician with
additional insights into the significance of the estimates.