Reinforcement learning applied to InfraBIM

In the last decades, artificial intelligence and machine learning have become very popular showing incredible success in applications ranging from speech recognition to computational biology. However, few have been done in the building or infrastructure field, where the potential of such methodologies is still unexplored. The introduction of BIM (Building Information Modelling) and InfraBIM (in the case of infrastructural projects) for the design, analysis and maintenance of projects makes available tons of data, along with a collection of modelling strategies that can be embedded in machine learning algorithms to solve complex problems. The construction of the tunnels by tunnel boring machine (TBM) starts from the design of a theoretical allignment, which is used by the TBM as guidance to calculate at each step a new as-built allignment through a laser-based machine that deviates as little as possible from the theoretical one. However, there is a lack of information during the design phase because of the locations of the rings are unknown before starting the cunstruction phase. This thesis aims to use machine learning algorithms, from the reinforcement learning branch, to calculate the position of the rings making the tunnel lining and allow to know the possible as-built allignement during the design phase. This can be done by extrapolating the coordinates and altimetries of the theoretical allignment and combining them with the data relative to the geometry of the rings to carry out simulations that allow the algorithm to understand, according to structural connections, what ring’s rotations are better in order to minimize the distance between the as-built allignment and the theoretical alignment. Subsequently, through the use of modelling software, the tunnel lining is modelled in order to enable future analyses.