Human-in-the-loop approach to digitisation of engineering drawings

In the nuclear power industry, the high-performance black-box systems which are prevalent in modern AI research are difficult to match to applications which take advantage of their strengths. These systems generally require volumes of labelled or well formatted data, and will provide a high level of performance which cannot be easily understood, explained or audited. Indeed, most AI systems deployed in this industry are done so with constant oversight by a skilled human operator, negating many of their advantages in cost, speed and reliability. In many cases, the time taken to format data for these systems is prohibitive even when automation might be desirable. This paper presents a framework for deploying a variety of AI techniques to industries where human oversight is required. Instead of treating the user as an external element while automating the task, this framework incorporates them as an active participant in the process, augmenting their performance while leveraging their strengths to make the AI systems more reliable and user-friendly. The framework is concerned primarily with the problem of digitising Elementary Wiring Diagrams, an important class of engineering drawing. These are in regular use even as low-quality scans of paper documents, and while digitising them is desirable and has many potential benefits, the level of time investment required by skilled engineers is prohibitive. Mistakes in digitisation are also potentially costly, meaning that the skilled engineer must remain involved in the process