Intelligent Recognition of Geometrical Tolerancing from Engineering Drawings toward LLM-Driven Semantic Understanding

Geometrical tolerancing is critical to ensuring the quality and performance of mechanical products. The rise of intelligent manufacturing presents new challenges for accurate and efficient recognition of geometrical tolerancing from engineering drawings. Traditional manual and rule-based Optical Character Recognition (OCR) approaches struggle to effectively interpret the intricate graphical symbols and their semantic relationships. To address these limitations, this paper proposes an intelligent geometrical tolerancing recognition framework that integrates deep learning-based detection with rule-driven structural reasoning. A YOLO-based detector is employed to localize dimension arrows, symbols, values, datums, and textual annotations, supported by targeted data augmentation to enhance robustness against variations in line styles and orientations. The detected primitives are then organized into structured geometrical specification elements through a lightweight post-processing algorithm that combines spatial-proximity matching with geometric constraints derived from drafting standards. This hybrid strategy enables accurate reconstruction of dimensional elements and tolerance relationships while preserving their semantic associations. Beyond recognition, the structured output produced by the proposed method provides a machine-readable representation that can serve as a foundation for Large Language Model (LLM)-driven semantic understanding in intelligent manufacturing workflows.