|KGE road cavity survey vehicle|
While that process has been effective, KGE’s senior management has been keen to explore new approaches that could make it less time-consuming and burdensome for its highly experienced engineers. And in recent years artificial intelligence has become the prime candidate for bringing efficiencies to the challenge.
On the road to AI
For that it turned to one of its existing technology partners Fujitsu Traffic & Road Data Service, a unit of the global digital transformation company Fujitsu, which was able to deploy its core AI technology at KGE in less than a month. Fujitsu’s Zinrai Deep Learning platform was used to create and run AI algorithms, alongside data management servers and an analytics engine that could handle the huge volumes of image data coming in from road surveys.
|KGE: Road cavity wavesforms|
For KGE, the results have been impressive — but only after development teams grasped that they needed to extend the scope of training required by the AI engine so it could become adept at distinguishing cavities from other underground materials.
Refining the model
“Our original objective was to limit the training of AI to cavities,” KGE’s Imai points out. And while that training allowed it to find almost every occurrence of a cavity, it also meant that the AI was picking up waveforms of other objects that it could not distinguish from cavities. “It turned out that training it [to just recognize cavities] was not sufficient. That is because stones and pieces of concrete embedded under the road can too easily be interpreted by the AI engine as cavities.
|Toshimune Imai, a specialist engineer at KGE|
“Sometimes even our engineers make that mistake. So we decided that if a trained human eye can do so, then the AI could have the same problem,” he says. By providing a more sophisticated set of training data, the tendency towards such false positives was dramatically reduced and the AI became more accurate at identifying genuine cavities.
That was very evident in a test phase which worked with data from a 10km stretch of road. KGE technicians manually identified 50 locations that they regarded as cavities and the AI was able to accurately spot 82% of those, even though it was working with a previously unseen data set. As Imai notes, that accuracy rate only increases as the AI absorbs more data.
While levels of accuracy were high, the speed of processing was even more striking. A three-person team of KGE technicians would typically take one hour to go through the image data for a 2km stretch of road. For the same data set, the AI took less than a minute — effectively a 90% reduction in primary detection time.
Boosting human competency with AI
But there was another important aspect. When a human technician analyzes the data, there is a tendency to only look at areas of specific interest in the wider data set, whereas the AI “added objectivity to the analysis” by scanning all available data, says Imai.
Over the next few years, the company has plans to create a compact road survey unit which is more mobile and could be used by less trained personnel and in non-specialist vehicles. If sensors are mounted on the patrol vehicles local governments use on a daily basis, analysis could be effectively performed on the fly.
Moreover, with the advent of 5G communications, KGE hopes to offer survey results from its road patrols almost instantaneously to its customers — civic authorities and their contractors — so they can act before a cavity turns into a dangerous sinkhole.