Design Knowledge Base Using Natural Language Processing

Abstract

The rise of big data and machine learning in engineering has brought a high hope that designers can learn from past successes and failures. However, this has generally only been possible when the available data is in numerical or graphical form where time series data analytics or convolutional neural networks (CNN) can be used. Very little work has been done toward utilizing the massive amount of textual data that is collected during the early-stage design process such as solution concepts generation and systematic design decisions. Design documentation at this stage contains extremely valuable information and expert knowledge that has been iterated upon for hundreds of years, but so much of that knowledge can only be stored in textual form and is therefore generally unused in big data and machine learning methods for engineering design. This thesis aims to improve the availability and accessibility of knowledge stored within engineering design documentation and to facilitate the learning process of junior designers by enabling big data from textual design knowledge. State of the art models in Natural Language Processing (NLP) are trained on and applied to large corpora of heterogeneous forms of technical design documentation to enable accurate information retrieval via semantic search capabilities. Text analysis algorithms are applied to connect relevant textual and non-textual design data present within documentation to enable searchable representations of non-textual design data. The goal of this thesis is to provide systems and direction for improving human design and engineering practice by learning from past design successes and failures. These systems are applied to case studies of massive corpora of design documentation in the fields of climate change research and micro/nano research to develop semantic knowledge management systems for each domain.