Transfer Learning with PipNet: For Automated Visual Analysis of Piping Design

This paper presents an end-to-end learning approach based on latest CNN architectures and transfer learning to perform vision-based analysis of engineering designs. The specific application considered here is the design of pipe networks on-board ships or offshore platforms. Having a piping design in the form of an image, a framework known as Piping Net (PipNet) is introduced to understand the design and interpret if it complies with applicable engineering regulations. Designs and corresponding labels (compliant or non-compliant) are fed into an existing trained CNN in the form of images for transfer learning, with the subsequently obtained fine-tuned network called PipNet. Based on Regulation 12, Annex I Regulations for the Prevention of Pollution by Oil, International Convention for the Prevention of Pollution from Ships (MARPOL) and Rules for Classification of Ships of Lloyd's Register, two datasets containing 3,234 piping designs in the form of images were used for performance evaluation. The developed system demonstrates outstanding performance on these two challenging datasets.