Hybrid FWFI/AI process for the automatic detection, 3D location and classification of embedded utiliities by electromagnetic techniques – PROMETHEUS

The detection and localization of utility networks in an urban setting has over the past few years become a topic of major interest. Standards (i.e. NF S 70-003) require a recognition of utility lines and an accurate location to within 11 cm conducted by certified service companies. According to feedback and evaluations from the scientific and technical teams within the Ministry of Ecological and Solidarity Transition, no solution currently meets the need for mapping underground networks over a large area at an acceptable cost for communities.
For such an undertaking, a precise mapping of buried networks through combining physical methods, artificial intelligence (AI) methods and innovative technologies adapted to hybridization, offers an undeniable advantage for optimizing work in terms of both time and costs. This step will also lead to quality gains and help reduce the risks associated with sensitive networks.
The PROMETHEUS project seeks to derive such a non-invasive methodological and technological solution, based on 3D radar technology, to structure the urban mapping of underground utility networks. This project is organized in five Work Packages (WPs), including project management (WP0) to coordinate the other WPs. WP1 focuses on the state-of-the-art and specifications for listing and selecting the influential indicators describing utilities and their environment. This selection step will generate output values or classifications for the machine learning techniques developed in subsequent WPs.
WP2 is devoted to developing a hybridization approach (Deep Learning & Matrix Pencil Method) for automatic utility detection and classification applied to raw C-scan data acquired from a multi-antenna GPR device. EM signal processing can be introduced for target identification given its sole dependence on geometry and physical properties. This proposed identification step thus entails applying the high-resolution method (Matrix Pencil Method) to frequency responses. Next, a deep learning segmentation will enable automatically detecting and classifying the utilities.
In parallel, WP3 focuses on a hybridization approach using GPR processing (3D migration and full-wave form inversion) prior to a deep learning process, implemented for high-yield and automatic investigations, as well as utility location and classification.
The final WP (WP4) addresses the constitution of various experimental GPR databases to complement the data modeled in WPs 2 and 3 towards developing methodological approaches. These databases will then be demonstrated on: a controlled test site with several homogeneous soils, a full-scale test site offering water table level control, and several actual sites proposed by the St Quentin Metropolitan Water Authority. The use of a commercial step-frequency 3D radar array system and the design of a laboratory multi-antenna radar prototype will also contribute to database compilation in the effort to devise a global methodology of automatic detection, localization and classification.
From an economic standpoint, this operational research proposal is part of the industrial partner’s (Logiroad) technical and commercial roadmap, calling for a solution that provides local authorities with access, at all times, to the full set of information characterizing road resources under their purview. The innovation resulting from this project will greatly improve positioning, thanks to a 3D platform regularly updated with information on road surfaces and structures, ancillary facilities and underground networks, in at least three markets, namely: public works contractors, network specifications, and communities seeking to optimize their road assets.
To carry out this project, the five partners (including the St-Quentin Authority as an external partner) will be aided by a research engineer, two PhD students and several Master interns.