Automated Annotation of Hydraulic Schematic Diagrams for ETO Manifolds Using Deep Learning and Computer Vision

Sanjay PARAJULI

Master's thesis

Abstract:

Many Industrial Hydraulics (IH) applications demand custom-designed manifolds, known as Engineer-to-Order (ETO) manifolds. The design and validation of hydraulic schematic diagrams for such manifolds require manual annotation, which is a time-consuming and error-prone process that delays the design-to-production pipeline. Although many methods exist for digitizing engineering diagrams, such as Piping and Instrumentation Diagram (P&ID) and electrical circuit diagrams, only limited research focuses on automated annotation of hydraulic schematics. This thesis addresses this research gap by investigating automated symbol recognition in hydraulic schematics using computer vision and deep learning techniques. To address class imbalance and high inter-class similarity, a multi-stage symbol recognition framework was developed. The research investigated the effect of modularizing the recognition task by employing Faster R-CNN and RF-DETR architectures for symbol localization and ResNet-18 for symbol classification. Using the proposed symbol recognition method, an end-to-end image-to-graph pipeline was developed to extract semantic relationships between symbols. A comparative study of classical computer vision methods and deep learning architectures was conducted for symbol recognition and line detection. To address the limited availability of labeled data, various training strategies such as data augmentation and transfer learning were evaluated. The proposed multi-stage symbol recognition method achieved a mean average precision (mAP@0.5) of 86.70% and a Recall@0.5 of 91.20% across 44 test diagrams. The solution comprising symbol recognition and the image-to-graph pipeline for automated annotation of manifold schematics is estimated to reduce the human annotation effort by at least 50% and enable downstream automation.

Abstract:

Many Industrial Hydraulics (IH) applications demand custom-designed manifolds, known as Engineer-to-Order (ETO) manifolds. The design and validation of hydraulic schematic diagrams for such manifolds require manual annotation, which is a time-consuming and error-prone process that delays the design-to-production pipeline. Although many methods exist for digitizing engineering diagrams, such as Piping and Instrumentation Diagram (P&ID) and electrical circuit diagrams, only limited research focuses on automated annotation of hydraulic schematics. This thesis addresses this research gap by investigating automated symbol recognition in hydraulic schematics using computer vision and deep learning techniques. To address class imbalance and high inter-class similarity, a multi-stage symbol recognition framework was developed. The research investigated the effect of modularizing the recognition task by employing Faster R-CNN and RF-DETR architectures for symbol localization and ResNet-18 for symbol classification. Using the proposed symbol recognition method, an end-to-end image-to-graph pipeline was developed to extract semantic relationships between symbols. A comparative study of classical computer vision methods and deep learning architectures was conducted for symbol recognition and line detection. To address the limited availability of labeled data, various training strategies such as data augmentation and transfer learning were evaluated. The proposed multi-stage symbol recognition method achieved a mean average precision (mAP@0.5) of 86.70% and a Recall@0.5 of 91.20% across 44 test diagrams. The solution comprising symbol recognition and the image-to-graph pipeline for automated annotation of manifold schematics is estimated to reduce the human annotation effort by at least 50% and enable downstream automation.

Keywords

ETO manifold, hydraulic schematics, symbol recognition, computer vision, deep learning

Jazyk práce: English

Datum vytvoření / odevzdání či podání práce: 11. 8. 2026

Obhajoba závěrečné práce

• Vedúci: prof. Dr. Ing. Marcus Barkowsky

UNIVERSITY OF SOUTH BOHEMIA IN ČESKÉ BUDĚJOVICE

Faculty of Science

Master programme / odbor:
Artificial Intelligence and Data Science / Artificial Intelligence and Data Science