Engineering Document (P&ID) Digitization

Introduction:

• P&IDs are essential in the manufacturing industry, but the manual process of digitizing them is a roadblock in the transition to digitally connected factories.
• The goal is to automate the identification of symbols, text, and connections within P&IDs.
• Research papers such as "Digitization of chemical process flow diagrams using deep convolutional neural networks" and "Digitize-PID: Automatic Digitization of Piping and Instrumentation Diagrams" provide insights for the solution.

Types of Symbols in P&IDs:

1. Equipment Symbols: Represent process equipment used in a plant or process.
2. Instrumentation Symbols: Represent various instruments used to monitor and control process parameters.

Symbol Detection Model:

• Azure Custom Vision was initially used, but the limitation of algorithms led to using AutoML Image Object Detection with YOLOv5 for training the model.
• A synthetic dataset with 50+ symbols recognition, including JPEG images with label annotations and bounding boxes, was used.
• New symbols can be recognized by feeding the model with new training data and labeling them using Azure ML Data Labeling tool.

Automated Training Pipeline Workflow:

• Data is pulled from blob storage and aggregated into a single annotation file to avoid bias in the model.
• The training pipeline generates a new model.

Text Detection Module:

• Text detection faces challenges such as crowded symbols and poor resolution quality.
• The OCR service used is optimized for text-heavy documents and engineering diagrams like P&IDs.
• A single OCR pass approach is used.

This digitization process eliminates the manual effort in converting P&IDs into digital format and enables more efficient onboarding of new customers and factories in the manufacturing industry.