KNOWLEDGE-BASED APPROACH FOR 3D RECONSTRUCTION OF AS-BUILT INDUSTRIAL PLANT MODELS FROM LASER-SCAN DATA

摘要

The three-dimensional (3D) reconstruction of as-built industrial plant models plays an important role in revamping planning, maintenance planning, and preparation for dismantling during the lifecycle of industrial plants. Recently, the 3D reconstruction of existing industrial plants was conducted using laser-scan data to make surveying processes more efficient. However, the current 3D reconstruction process from laser-scan data is still limited due to the need for significant human assistance. Although a great deal of effort has been made to efficiently reconstruct 3D as-built industrial plant models, the presence of objects - such as equipment, pipelines, and valves of different sizes and shapes - in existing industrial plants significantly increases the complexity of laser-scan data and makes automating the reconstruction process more challenging in practice. The purpose of this study is to propose a knowledge-based approach for the 3D reconstruction of as-built industrial plant models from unstructured laser-scan data. First, pipelines were extracted from laser-scan data based on surface curvature information and knowledge about pipelines' sizes from existing piping and instrumentation diagrams (P&ID). Once entire pipelines were extracted, they were modeled based on skeleton features. Then, the remaining objects were clustered and grouped separately via the region grouping process. Afterward, clustered objects were retrieved and modeled based on global feature-based matching from the 3D database. Finally, the resulting model was checked to ensure that it was well-reconstructed according to the information regarding the relationships among objects abstracted from the existing P&ID. The preliminary results on actual industrial plants show that integrating knowledge into the reconstruction steps played an important role in the proposed approach and that this approach obtained accurate as-built industrial plant models from unstructured laser-scan data. The proposed approach could be successfully utilized to assist in many applications during the lifecycle of industrial plants.