Geometric Similarity Retrieval of Industrial 2D CAD Drawings Using Deep Learning : Exploring similarity retrieval methods for industrial CAD drawings

Sammanfattning

The adoption of Artificial Intelligence (AI) at the commissioning organization remains limited.However, the increasing need to efficiently manage rubber profile blueprints and the associated pro-duction parameters has highlighted the need of a system capable of evaluating geometric similarity andretrieving relevant historical data from the enterprise system. Technical drawings of rubber profilesconstitute the foundation of production, yet these are stored without any advanced indexing or retrievalmechanisms, such as Computer-Aided Design (CAD) specific search engines or tools powered by AI.This limitation is shared across multiple facilities within the industry. Consequently, there is a cleardemand for a system that can identify and retrieve similar drawings, with the local facility additionallyrequiring access to historical production parameters linked to these profiles. This project proposes anovel retrieval-based system for 2D CAD drawings, where geometric primitives are extracted with thehelp of Density-based spatial clustering of applications with noise (DBSCAN) and transformed intolearned embeddings using custom Deep Learning (DL) models. Two architectures based on Trans-former and graph-based representations are implemented and compared to a baseline encoder model.The similarity retrieval is conducted using the K-nearest neighbors (k-NN) model. The results are eval-uated using domain experts at the company. The results show that the graph-based model achieved thehighest retrieval performance, while visual models, such as DINOv2, received the highest expert use-fulness ratings. This indicates a discrepancy between metric-based evaluation and human perceptionof similarity in an industrial context.