The objective of this post-doc is to develop new computer vision and machine learning approaches and models for physics-based representation learning in the domain of tunnel construction.
In recent years, the issue of resource efficiency has also become increasingly important in construction engineering, as soil and rock account for more than 50% of mineral construction waste. Tunnel projects play a special role in this regard, as large quantities are generated at specific times and places. Due to the high degree of mechanisation and the associated advantages in terms of construction performance and safety at work, almost the half of tunnels is built with Tunnel Boring Machines (TBM). For documentation and control of the construction process, these are equipped with various sensor systems that provide comprehensive data sets. But in this area, modern data-driven processes have not yet found a wide application.
This 24-month post-doc position is funded by the French-German ANR project REMATCH. The overall objective of this project is to use the data sets from TBMs and conceive novel machine learning-based models to enhance the recycling of the large quantities of tunnel excavation material. In this regard, an innovative real-time measurement system for material characterisation is to be developed which gives decision support on the question if soil is “usable” or “not usable” for other purposes and thus has to be disposed of in a landfill. This system will base on several approaches using Computer Vision (CV) and state-of-the-art machine learning methods: firstly, on video data of excavated material on a conveyor belt, and secondly, on sensor data recorded simultaneously by the TBM.
The objectives of this post-doc position are to develop a computer vision system that analyses the videos captured from one or several cameras installed at the TBM and filming the excavated material on the conveyor belt. In order to decide on the possible reuse of the material, different geophysical properties need to be estimated from visual features extracted in real-time from the video stream(s) coming from RGB cameras. This is challenging due to the mediocre acquisition conditions under low lighting and fast motion inducing some motion blur. More specifically, after some preprocessing, a first step is to develop a machine learning solution based on appropriate CNN models that are trained for either classification and/or regression tasks in a supervised manner. Different models should be developed and evaluated in terms of robustness and precision.
To go further, novel innovative neural network-based architectures and weakly supervised learning schemes should be proposed to learn a latent representation that reflects the meaningful similarities of relevant soil characteristics. Then, potentially other physical properties should be incorporated more explicitely into this semantic latent representation (either via a specific CNN, an autoencoder-type model and semi-supervised or contrastive learning depending on the amount and type of available annotation and the possibility to exploit the TBM sensor data for labelling). The goal is to make the learnt features and models more explainable. After evalating these models, a final approach would be to incorporate TBM sensor data in a combined and multi-modal architecture.
The position is in the IMAGINE team of the LIRIS laboratory in Lyon (Campus La Doua) under the supervision of Catherine Pothier and Stefan Duffner. The funding is for 24 months. Additional teaching activities may be conducted at INSA Lyon if the candidate desires to.
Stefan Duffner firstname.lastname@example.org
Catherine Pothier email@example.com
(c) GdR 720 ISIS - CNRS - 2011-2022.