Learning Engineering Properties with Bag-of-Tricks. For the Automated Evaluation of a Piping Design

This paper presents a text based methodology to evaluate the piping design of a ship, which is usually in the form of a Piping and Instrumentation Diagram. The method starts with describing a piping layout or a design in the form of a paragraph of text, describing the essential properties of the design, including but not limited to the connectivity and the attributes. Next, vectorisation takes place to obtain a mathematical or quantitative representation of the text-based description. Technically, the piping design is represented as a point in d-dimensional space. Next, leveraging on the recent development in natural language processing and based on the vectorised representations, a bag-of-tricks based classifier is trained to evaluate if the design complies with the applicable rules of design. After the process of learning, the trained classifier can be used to evaluate if the piping layout or design of a ship is compliant. The developed method has demonstrated encouraging performance on a challenging dataset of 144 compliant and 1836 non-compliant piping designs collected, based on an existing case study around Regulation 12, Annex I Prevention of Pollution by Oil, International Convention for the Prevention of Pollution from Ships by learning in the space of 100 and 300 dimensions.