The following questions were received recently (July 2004) in response to the "Elusive P-F Curve" article that was published at Reliabilityweb.com. http://www.reliabilityweb.com/art04/p-f_curve.htm

 

A: Only if we have a policy for declaring the P point. That policy can come from:

-         1) common sense  (if we are dealing with simple situations),

-         2) engineering models, or

-         3) experiential models

For example, How do you know the change from 0.1 bar to 0.12 bar is the P point? Why isn’t it  the change from 0.1 bar to 0.15 bar, or something else? Your policy (for declaring that a developing failure mode has attained ‘P’ status) will have been derived from one of these 3 types of models.

 

Every individual unit behaves differently, but they, as a population, follow a certain rule – a distribution. This is why we need the statistical models and tools. There is no practical physics model that can describe the whole population in which each individual behaves randomly differently.

 

A: There is a P point. Sometimes (in well understood physical situations) it is definable as an unwavering alert limit  for a CBM program. Other times it is non-deterministic. Meaning it depends in unknown ways on a variety of operational factors. In those cases we have to use statistical means to seize its nature.

 

A: No argument with your logic for this case, which is similar to that of a tire wearing down. For a monotonically increasing indicator, and a good conceptual model of a physical phenomenon (such as the accumulation of debris on a filter or the wearing down of the treads of a tire), and a condition indicator that clearly tracks resistance to failure, CBM is applicable without recourse to statistics.  However, the logic does not "scale well" with increasing complexity. Often, we need some help in defining the residual life estimate where multiple internal processes are not so obvious, and, where the signals we are monitoring are not clearly reflective of those physical processes.

 

Take, even, the oil filter, for example. How do you know (how do you estimate) that it will take 2 months before the delta-pr becomes 0.5 bar from the time of a slight rise in the delta-pr? For that matter, how do you know that you have indeed arrived at the F point when the delta-pr. becomes 0.5 bar? Is it you who has defined the point at which to declare F? How have you done so? You (the process engineer, the filter manufacturer, or the system designer) have, most likely, used method 1, 2, or 3. EXAKT is a methodology that assists the process for establishing/declaring a potential failure and providing an estimate of the residual life until failure. In an interview, Dr. Dragan Banjevic describes the process by which such a “P” declaration model is arrived at.

 

A: We would probably not apply EXAKT to such a situation where the differential pressure signals are so directly and unmistakenly related to failure resistance, and the signals contain little random noise and there are few complicating factors. Of course, EXAKT can apply to this example. The delta-pr. would be the covariate in the PHM. But we would still need to know your (as the filter owner or user) definition of failure. It seems that you defined failure (“clogged”) as "delta-pr. having reached 0.5 bar". If you can provide us with the data (see the article Using your CMMS for Reliability Improvement and Tutorial 1), we can built the model for you. The model  would be somewhat similar to that of the case history described in the article Seal leakage in nuclear fuel handling system whose data and model building procedures are given in Tutorial 4.

 

A: The statistics will only reflect the physics. But if we have the “physics” (meaning a good engineering model) that adequately describes the physical situation, then a statistical model may be superfluous. It is generally preferable to use the engineering model if one is available. There are many real-world situations, however, that are very complex. We don't have an analytical model that adequately describes such situations and we are forced to use statistical or simulation tools.

 

However, even when we have a good engineering model, an EXAKT model can show you whether the signal you are using is a right one or not, and how good it is. That is, how much confidence can you place in your model and the decisions that it generates. It can show you if there are other possible physical indicators that you should include in the model.

 

A: No argument. If you understand the situation well enough to predict F with sufficient confidence. Go ahead. Don't waste your valuable time with "fancy mathematics and statistics". In fact EXAKT's methodology agrees with you, in the sense that it does not attempt to recognize a deterministic point P. But it does require the user to have defined F in some way. “F” for EXAKT should not be a catastrophic functional failure involving loss of life, injury, or serious economic consequences. That (as Resnikoff would say) would be a bad way to build a model. OMDEC subscribes to the "staged intrusiveness" approach described in the interview with Dr. Dragan Banjevic.

 

A: Not all of these papers are "academic". Some of them are quite practical. Please have a look. In fact every model built at the CBM lab over the past decade has addressed data that came from an industrial site. And every project was performed collaboratively with a team member from the Lab and one from the site.

 

A: OMDEC cannot address the "down-to-earth practical situation" better than you have excellently done. However we might be able to do better[1] in more difficult situations that are characterized by random noise in the measured signals, conflicting effects from multiple internal phenomena, minor maintenance affecting condition indicators and resistance to failure, changes in operating profile, and sensor data that does not directly nor clearly capture ground-truth resistance to failure.

 

I might mention that the foundation work for EXAKT was done in the CBM lab at U of T and is funded by 10 companies all concerned about excellence in condition-based maintenance. The research program is in its 10th year and it was with much consideration of EXAKT's practical value that we decided to spin-off OMDEC in order to bring the methodology to the mainstream of physical asset management.

 

You might look at the lab web site for reports of some pilot studies undertaken with Consortium members. The 10 members supporting the CBM optimization  research program are:

 

From Canada:

 

ABB

Department of National Defence

Dofasco Steel

Hydro One

INCO

Irving Pulp and Paper

Syncrude Canada

 

and international member representation comes from :

 

EDF (Electricite de France)

Ministry of Defence(U.K.)

Zachry Construction (USA)

 

You will find details of the Lab’s work at  the web site