Multimodal Retrieval and Product Structure Proposal from 2D Technical Drawings
3.16 MB
avoin
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
Pysyvä osoite
Verkkojulkaisu
DOI
Tiivistelmä
Manufacturers manually produce Enterprise Resource Planning (ERP) product structures, bill of materials (BOM) and work phase plans for every new two-dimensional technical drawing. Although they usually have an archive of past drawings with linked ERP data records, it is impractical for the engineers to find similar drawings and data within it, especially if the archive is expansive. This thesis develops and evaluates a proof-of-concept computer aided process planning system that leverages this archive. Given a new drawing, it retrieves similar past drawings, and proposes a preliminary bill of materials and work-phase plan. The system employs a two-layer pipeline architecture. The first layer performs format-specific data extraction to produce a format-neutral representation. The second layer embeds this representation using a Multi-Layer Perceptron on the textual data and a frozen DINOv2 encoder on the visual data. These embeddings drive the retrieval, classification and proposal tasks. The system is evaluated on a 1829 PDF corpus with linked ERP data. The results are distinct and divided by task. ERP prediction is primarily driven by text. Work-phase routing is predicted with multi-label classification over a 21-class production line taxonomy, and scores 0.82 macro-F_1, 0.92 micro-F_1 with 0.60 exact set match rate while staying vendor neutral. The long-tailed BOM record is addressed using a classifier for recurring components and a retrieval pool for the tail. Visual channel proves stronger for retrieving similar drawings that are human relevant. Measured against human judgement, a frozen mean-pool encoder attains a top 5 hit-rate of 0.84, but it does not improve ERP prediction. Visual similarity does not seem to suggest shared product structures, and multimodal fusion yields no performance increase. The thesis provides a vendor-agnostic, retrieve-and-propose framework and its prototype, together with a realistic discussion of its capabilities on real industrial data.