Handling Short Segments in FWP Optimisation
Some observations about the difficulties with handling short segments in FWP optimisation models
I wrote in an earlier post about my involvement in tuning triggers for a Juno Cassandra deterioration model on a road network for a large New Zealand City.
In that analysis, we were trying to refine our triggers for selecting candidate treatments such that there would be maximal correspondence between the Field Validated programme the client suggested and the candidate treatments put forward into the optimisation algorithm by our model.
One of the interesting challenges that cropped up in this process was how “short” sections seemed to be a fly in the ointment when it comes to interpreting percentage distress from network surveys.
What we often found is that a trigger such as “if mesh cracking is greater than 20%, then consider treatment X” would tend to put forward many short elements that had only localised distress (often at one end of a segment, such as at a Cul-de-Sac). Treating short segments also has a lower cost which means short segments may be unduly favoured in an optimisation algorithm based on Benefit and Cost unless special precautions are taken.
In the figure below I try to conceptualise my evolving view on this topic. Here, I define a “short” segment as anything less than 150m, although other limits can also be chosen.
As the figure shows, situation A, representing a short 100m segment, has only 30m of distress at one end of the element, yet the percentage distress is a rather alarming 30%. In my view, this is something that can be fixed with routine maintenance and does not warrant one of the coveted places in a three year Forward Works Programme (FWP) - unless there are other mitigating factors such as a very old surfacing in need of a resurfacing.
Situation B shows that the same length of distress on a longer segment will only map to 12% distress - also clearly not a systemic problem that should warrant a place in the FWP.
For longer segments, I feel more confident to use percentages. The key challenge is to find the percentage distress that would suggest a systemic rather than isolated pavement performance problem. That is, for longer segments we are looking to distinguish situations E from B (and possibly also D).
My experimentation suggested that a sum of the most relevant distress types (more about what “most relevant” means in a later post) over around 20% to 30% applies, although these thresholds should be expressed as a function of surface age.
That is, if the surface age is more than 100% of the expected life, then a lower threshold applies. When a segment with a relatively young surface (say 50% of expected life) is selected, then normally the sum of relevant distress threshold is closer to 60%-80%.
In our developing Juno Cassandra model for NZ Local Authorities, we have included a sliding scale threshold that is a function of Surface Life Achieved percentage. This scale can be calibrated from one network to the next.
At this stage, this component of our candidate treatment selection trigger logic looks like this:
Of course, whether the segment is short or long, a percentage distress cannot tell us whether the distress is isolated (situations A, B, C and D above) or extensive (situation E). This is why I am really excited about the potential of identifying distresses using Video Recognition (VR) and Artificial Intelligence (AI) technology.
These technologies are now available through Juno Intelligence and other similar products and makes it possible to pinpoint distress locations, get a relatively accurate estimate of distress quantity and you get a video survey of your network as an important bonus.
For me, one of the most exciting benefits of identifying distresses using VR and AI is that it now provides us with a means to determine more rigorously whether the (say) 30% distress all occurs in one spot, or whether it is spread out over a segment’s length, thereby confirming whether or not systemic distress is present. I plan to write more about this detection of systemic distress using Juno Intelligence in a future post.