Hybrid As-Is Modelling of Existing Industrial Facilities

The aim of this project is to create a viable approach to automate the generation of as-is geometric Building Information Models (BIM) of existing industrial facilities. It is anticipated that the student will achieve this aim by devising, implementing and testing a novel hybrid (i.e. from facilities to structural objects, electrical equipment, the piping system etc.) approach that will take advantage of building priors to progressively segment point cloud data through fitting template objects and simultaneously assigning their connectivity relationships.The process of as-is infrastructure modelling involves the automated generation of infrastructure objects and their relationships (the geometric BIM) from low-level point cloud datasets. As-is modelling of a specific type of infrastructure (industrial facilities in this case), unlike a generic detection or recognition problem, is a process that can only lead to a limited set of objects and relationships. Geometric modelling is the "bottleneck" for the conversion of laser scanned point clouds of existing facilities to as-is BIMs. This conversion requires 90 % of the modelling time. Industrial plants are the most geometrically complex facilities due to the plethora and complexity of their objects. Therefore, the innovative nature of this research is two-fold: (a) which objects are most important to model and (b) how should we automatically model those to efficiently reduce modelling time of these assets? Top-down modelling is a novel rethinking of this process that is inspired by human vision and hypothesizes that many infrastructure object classes (e.g. pipe) are more uniquely distinguishable through their pose and relationships to other objects (e.g. beams/columns) than their own features (e.g. plane, white colour). In this case, top-down modelling will be combined with bottom-up (point-to-point) modelling that involves learning features of geometric shapes in a facility.
The research questions that will be answered in order to address the objectives of this research are: (a) which object types are most important to be automatically modelled in order to reduce modelling time based on the frequency of appearance, the modelling time and the value of each object type for safety, maintenance and retrofit of industrial plants, (b) how can these most important object types be detected and classified in cluttered point clouds of industrial scenes by integrating prior engineering knowledge (e.g. outer diameter distributions of piping elements) and (c) how can their connectivity relationships be defined to automatically generate the BIM model of industrial plants. The key novel idea that makes this possible is using a hybrid method that combines "bottom-up" geometric features using deep learning algorithms that learn geometric features of point clusters in a facility's point cloud (e.g. curvature, point normals) and high level engineering knowledge for the layout of the piping and structural system of the most important industrial objects. Priors, formulated as rule-sets (e.g. cylinders connected to the floor must be pipe supports) can be used to identify the individual components. Template priors and fitting methods could then be used to replace segmented high-level primitives with actual objects and their relationships. The applications of this research are two-fold. Firstly, these models will assist facility managers to be more proactive in decision making that involves maintenance, operations and safety. Secondly, this research will facilitate the reduction of the modelling time needed to generate the BIM models of these facilities. As a result, laser scanning to BIM model generation time will be reduced and continuous production flow of these assets will be achieved.
The datasets are provided by industrial partners (AVEVA, BP) and collected by the student.The student will then generate a benchmark dataset of the most important industrial shapes to train the proposed algorithms.