 |
 |
 |
 |
 |
 |
 |
||
|
|
|||||||||
|
This presentation
describes a CBM (condition based maintenance or condition monitoring)
project. The title contains the word "Optimal". What does optimal
mean in regard to condition monitoring? It can really mean
many things. It depends on your objective.
If your objective is to run at very low
life-cycle cost then you may react to condition monitoring data in one way. But what if your objective is to run at very
high availability? Will your behavior (as to how you interpret monitored
data) be the same? Or you may need to operate at some particular combination
of availability and cost. And for good measure you may require a specified
reliability (mean time to failure). Each of these
objectives will require you to manage your CBM program differently. They will
depend on the context in which the asset operates. |
Slide 1 |
|
The phrase "P-F
Interval" was coined by the late John Moubray. He used the term to
highlight the requirements of a CBM program in this well-known diagram (on
the left of Slide 2). However, this
empirical diagram is deceptively
simple. Deceptive, for at least two reasons. First it assumes that the monitored
data resembles the “Ideal” graphs on the slide – monotonically increasing
trend lines with the red alert limit set, presumably, to the level of the potential
failure “P”. How many of us,
involved in CBM, believe that data, generally, looks like these ideal plots?
Are not the random fluctuation and contradictory trends of the “real” graphs
(on the far right of Slide 2) more familiar? |
Slide 2 |
|
That is one
problem. We illustrate the second
problem in the study of 11 gearboxes that were run to failure (in accelerated
life tests). Even with appropriate signal processing (described in Slide 3),
we must still decide upon the level at which to declare that a
potential failure has occurred. Determining the
decision point, (i.e. the potential failure) is seldom obvious. Unless the decision
process can be automated, CBM, is simply not workable. There are insufficient
human resources at our disposal to pore over (ever increasing numbers of)
graphs and tables to reach practical day-to-day condition monitoring decisions. How do we solve both these
problems? |
Slide 3 |
|
We need to recognize
that there are two major sources of data for a decision model. Event
data (defined at the top of Slide 4), and Inspection data. The table at the
bottom of Slide 4 contains the inspection data variables or features that
were extracted from the raw vibration spectrum (using the wavelet
based signal processing algorithm). |
Slide 4 |
|
Once you find an
appropriate signal processing method that targets the failure mode of
interest, you will get graphs that are similar to those of Slide 5 and Slide 6. |
Slide 5 |
|
Well behaved (little
random scatter) as these graphs are, we still need a decision making
procedure or model. To emphasize that point, gearboxes 12 to 15
had a different geometry – different gear sizes
and different gear ratios from those of gearboxes 5 to 11. Will the same
decision policy (i.e. the same declaration rule/alert limit for a potential
failure) apply to these deliberately modified sets of gearboxes? |
Slide 6 |
|
From the data
(Inspection and Event) of the first set of gearboxes we investigated several
candidate (potential failure declaration) models. How do we know which model
(of the six that were built using the proportional hazard modeling method) is
the best with regard to our specific condition monitoring objective? |
Slide 7 |
|
In this case our
objective happens to be lowest average maintenance cost. That is, we
want to attain the best compromise between 1) acting too quickly, and
2) waiting too long. “Best” means that the policy would result in the minimum
life-cycle cost. |
Slide 8 |
|
What does such an
optimal policy look like? Slide 9 shows the policy (FGP1) applied retroactively
to gearboxes 5, 6, 7, and 8. If you operate your CBM program according to the
policy as directed by these graphs, you will, in the long run, achieve your
CBM objective for that asset (or fleet of assets). Note the curvature
of the limit boundaries. The curve of the optimal decision policy is telling
us that age counts. That is, for Gearbox Type A, a younger gearbox can
“tolerate” higher levels of monitored values than an older unit. |
Slide 9 |
|
Now what about the
second group (Type B) of gearboxes? Those familiar with the Weibull model
will quickly note (in the equations of Slide 10) that the value of Beta (the
shape parameter) is equal to “1” in all 5 of the candidate models. What,
then, will the (potential failure) declaration policy for Type B gearboxes
look like graphically? |
Slide 10 |
|
That’s right. No
curvature. Age is not a key significant factor to the risk of failure of
gearboxes of type B. Hence we acknowledge
that the intepretation policy that one chooses is not obvious. We can
rarely observe graphical trends of condition monitoring values, and,
unassisted by analysis, arrive at an appropriate decision. If we do, that
decision will be subjective and almost certainly sub-optimal. On the other
hand, analyzing historical data will allow us to understand the key risk
factors influencing failure and enable us set optimal policies for
declaring a potential failure. |
Slide 11 |
|
A trial (see http://www.omdec.com/)
on the experimental Advanced Amphibian Assault Vehicle was a practical
problem in which the wavelet algorithm and an EXAKT optimal interpretation
model were used. They were embedded as an “intelligent agent” in a
single-board computer and a digital signal processor (DSP). Each of 17 gears
could thus be monitored through the use of an optimal decision policy. |
Slide 12 |
|
