A game of outliers

Pipeline integrity is a game of outliers.  Best integrity management practices catch almost all cases of impending failures.  Statistically, transmission pipes never fail…  The number of failures per pipeline mile is infinitesimally small.  (I calculated the actual value at one point, will try to dig it up.)  The trouble is the pipeline industry is held to the same standard as airlines.  Single failures are significant.  Single failures are inevitably outliers.

A statistical approach like extreme value analysis attempts to coax outliers from data using the assumption that we can never perfectly know a system.  This approach requires an exhaustive understanding of the natural laws of the system being studied. Given what we know about pipeline failures, what are the odds that we can predict something that we have never observed?  This technique is a bit spooky. It will only tell you that something rare can occur.  It will not tell us where or when.  As a result, we have to apply super conservative maintenance practices knowing there could be a problem but not knowing where.

What about a more direct approach?  What about looking for developing outliers as they occur.  Can we see river scour as it happens during a flooding event? The most conservative models will never catch a super aggressive bacteria population in a random spot of the pipe. But maybe with enough suitable sensors, we could see it happening? Direct monitoring would allow us to identify a runaway corrosion pit and, most importantly, know where this pit is.  Then, we only have to apply conservative practices near the trouble spot.  We can leave the rest of the pipeline alone, knowing that with direct monitoring, the rest of the pipe is okay.

I am looking for better ways to directly monitor pipe deterioration.  I have gotten my masters in wireless sensing, and am working on my PhD in information systems. My specialization is combining sensor readings to achieve time-based situational awareness. I am adapting tools like SLAM from autonomous car navigation and pattern-matching techniques from image registration. I am also developing a quantitative method for identifying which combination of sensors is best at identifying specific failure modes.

This blog entry represents my re-entry into industry.  I have finished my course work.  I have passed my PhD qualifying exams.  I am now exclusively working on the problem.  So, time to dust off this blog again.  I have enough material that I can produce a fairly steady stream of interesting pipeline integrity topics.  I will keep you posted about the progress of my research, topics I have explored, and random thoughts about near-real-time pipeline monitoring.