Events

Automated Extraction of Complexity Measures from Engineering Drawings

Abstract: An engineering drawing is a detailed representation of an object used to communicate complex information for the purposes of design, manufacturing, and maintenance. These line drawings typically consist of multiple 2D orthographic views of a 3D object, along with dimensioning information and metadata about specific properties. Over the past decades, engineering drawings have evolved from hand-drawn sketches to highly standardized documents created with the help of CAD software. The large variety of engineering drawings makes it difficult to automatically extract abstract information in a robust way. The emergence of additive manufacturing (AM) promises companies that they can produce spare parts on demand for maintenance, potentially increasing the operational time of their infrastructure. Evaluating the AM potential of spare parts is essential from both an economic and technical perspective. This analysis of economic and technical viability requires the interpretation of complexity measures that can be derived from the engineering drawing of a spare part. The external dimensions of an object are key complexity measures to facilitate an AM potential analysis. In this thesis, we propose a processing pipeline that automates the extraction of complexity measures from engineering drawings, focusing on the external dimensions of the depicted objects. An in-depth examination of engineering drawings from different eras forms the basis of our methodology. Our pipeline is designed to be adaptable and consists of interpretable stages for specific tasks. We segment important entities in the input drawing to detect candidate dimension lines that are subsequently filtered by a sequence of processing steps. The grid structure of the orthographic views is determined, which allows us to assign axis labels to each view. We run optical character recognition (OCR) on detected dimension numbers and use the results to optimize the ratio between the OCR values and the length of dimension lines in pixels, providing us with a solution that is resilient to errors in the OCR predictions. A prototypical implementation of our pipeline demonstrates its capabilities in handling a large variety of drawings. We conduct a basic quantitative and qualitative evaluation of our methodology. The results confirm the effectiveness of our approach in automatically extracting abstract information from real-world engineering drawings.