ElgoeeAI - an Ai demonstration system
Author
Fleinert, Tobias Elgaard
Term
4. term
Education
Publication year
2020
Submitted on
2020-07-06
Pages
92
Abstract
This thesis presents ElgoeeAI, a demonstration system designed to illustrate the potential of artificial intelligence in manufacturing and to engage stakeholders in discussing opportunities and challenges. The work begins by exploring the manufacturing domain through interviews with industry practitioners, which informs a hypothetical production setup. A deep learning-based image recognition model is then developed and trained to classify product states within this conceptual process. In parallel, communication strategies are examined to make model outputs understandable and actionable for users; text, graphs, and icons are compared in lightweight user studies, and a combination of text and icons emerges as the most effective mediation approach. These insights are realized in a high-fidelity system with a Python/Django backend and statistical analyses in R, presenting classifications through an engaging interface. The system is briefly trialed in informal settings and, despite some bugs, sparks notable interest and debate around AI adoption in production. The thesis contributes a concrete demo setup, a rationale for key design choices, and empirical insights into how to best communicate AI outputs to human operators in manufacturing, while acknowledging the limitations of a hypothetical plant and the need for more systematic evaluation in real-world contexts.
Dette speciale præsenterer ElgoeeAI, et demonstrationssystem udviklet for at vise potentialet ved at implementere kunstig intelligens i fabrikationsmiljøer og for at engagere interessenter i muligheder og udfordringer. Arbejdet indledes med en kortlægning af fabrikationsindustrien gennem interviews med personer fra branchen, hvilket danner grundlag for et hypotetisk produktionssetup. På denne baggrund udvikles og trænes en deep learning-baseret billedgenkendelsesmodel, der klassificerer forskellige produktstadier i det konceptuelle forløb. Parallelt undersøges formidlingsformer for at gøre modelresultater forståelige og handlingsorienterede for brugere; her sammenlignes tekst, grafer og ikoner i simple brugerundersøgelser, og en kombination af tekst og ikoner viser sig at være den mest effektive mediator. Indsigterne omsættes i et HiFi-system med backend i Python/Django og statistisk analyse i R, der formidler klassifikationer i et engagerende interface. Systemet bliver kort prøvekørt i uformelle sammenhænge, hvilket – trods enkelte fejl – vækker betydelig interesse og debat om anvendelsen af AI i produktion. Specialet bidrager med et konkret demo-setup, en gennemgang af de væsentlige designelementer og empiriske pointer om, hvordan AI-uddata bedst formidles til menneskelige aktører i produktionskontekster; samtidig peger det på begrænsninger ved det hypotetiske setup og behovet for mere systematisk test i praksis.
[This apstract has been generated with the help of AI directly from the project full text]
Keywords