The Netherlands is a densely populated country which has a high rate of development. By consequence, there has been a lot of construction activity over the years. This resulted in the creation of a substantial underground infrastructure. We therefore cannot ignore the complexity of the cables and pipelines that are in place throughout the Dutch underground.

In a previous blog, we showed that the complexity of the infrastructure has even today not been fully administered. This results primarily from the pre-2008 situation: only then, the WION – the law requiring proper administration of underground infrastructure – was introduced into Dutch legislation. Before, a system of self-regulation was in place, which did not work out properly. Additionally, the equipment that was used to determine exact locations much less accurate in the 1950s and 1960s, an era in which there were many construction activities resulting from a growing economy. Today, the WION has been extended into the WIBON, in order to include above-surface utilities as well.

In sum, the challenges described above have yielded one big consequence: annually, many utilities are struck by excavators, with millions of Euros in damages as a result. And then we don’t even include the immaterial damage, because what if a gas pipeline is hit?

Avoiding excavation damage with Ground Penetrating Radar

In a collaboration between TerraCarta B.V., Aime investigates since January whether Artificial Intelligence can be used to prevent excavation damage with Ground Penetrating Radar (GPR). We do this for a master’s graduation assignment at the University of Twente. GPR provides many opportunities for this purpose, since it allows practitioners to look into the subsurface. This is possible because the device emits electromagnetic waves, which reflect on objects and of which the echoes are picked up by antennas. The data is subsequently interpreted by an expert, whose experience allows them to identify what is located in the underground. Acquiring such experience often takes many years.

Based on an extensive literature review, we identified the state-of-the-art in academia with respect to machine learning and Ground Penetrating Radar. Much work has been done in the area of recognizing underground objects. In short: with contemporary algorithms, it is not too difficult to detect the presence of underground objects. However, what also emerges from our review is that those algorithms cannot be used to detect the characteristics of underground utilities. Whereas the algorithms are highly capable of recognizing the hyperbolic shape of the object (which we see in the image below), they cannot accurately recognize its structure.

Three new algorithms for recognizing cables and pipelines

Yet, the scientific fields related to our research goals are moving slowly but surely towards recognizing the characteristics of underground utilities. The object we’re seeing, is it made of steel? Or is it perhaps PVC? The quantity of studies related to those questions is scarce, but nevertheless highly relevant for practice.

In the past few weeks, together with TerraCarta’s geophysical expert, the professorate at the University of Twente and an expert in the field of signal processing, we designed three novel algorithms which we will test in the upcoming period. We strive to pre-process the signal from the GPR device  in a smart way, which allows the signal characteristics of different structures to become very distinct. Using the input data to train a specialized neural network specialized in handling such data, we hope to set a new step forward. Neural networks like those, which attempt to mimic human information processing in an artificial way, have in previous years gained much traction and are therefore the prime candidate for research like ours.

The precise configuration of the three neural networks, which we obviously make based on our own experience augmented with insights of other machine learning engineers reported online, is a task on its own. Currently, we’re finishing the configuration with which we aim to start the training process. If luck is with us – like a child cannot be forced to learn a new language perfectly, neither can a machine learning model be forced to learn the complexities underlying data perfectly – we will find positive results that are usable for practice.

Aime & TerraCarta … and a multinational organization!

We can thus conclude that we have worked very collaboratively with TerraCarta in the past period. The result is three novel algorithms that we will start validating soon. TerraCarta’s experts shared various innovative ideas which are grounded in their knowledge of geophysics, the underground and the GPR-based analysis market. We’re working in a playing field for testing new ideas! It’s a win-win for all.

We also wish to end our update positively. Thanks to outstanding work by TerraCarta in a previous project, a multinational organization is currently interested to be involved with our future developments. Recently, a delegation from TerraCarta and Aime visited this party, during which interesting ideas were discussed to ensure larger maturity of the analysis market. To be continued!