Högskolan i Skövde

AI-powered visual inspection is increasingly used in manufacturing quality control, yet its practical implementation in small and medium-sized enterprises raises questions about data availability, production scale, and the internal expertise needed to sustain such systems. This thesis investigates how a data-centric AI vision system can be implemented in a resource-constrained manufacturing context, where labelled defect data, specialist AI knowledge, computing resources, implementation time, and long-term maintenance capacity are limited, using Skandia Elevator as the industrial case company.The study builds on a longer collaboration between the University of Skövde and Skandia Elevator, where previous work established parts of the physical inspection infrastructure and an initial proof-of-concept system. The present project focuses on three connected aspects of implementation: sustainable knowledge transfer to company personnel, the trade-off between model quality and inference speed, and the suitability of anomaly-detection model candidates for temperature-sensor image inspection. The project therefore treats the AI system not only as a technical artefact, but also as an organisational learning process that must be understandable, maintainable, and useful after the thesis project ends.This thesis explains how data-centric AI can be implemented in a manufacturing SME context. The findings show that model performance depends not only on the algorithm, but also on data quality, user understanding, and organisational ownership. In resource-constrained settings, systematic work with data collection, labelling, retraining, and knowledge transfer may be more important than using the most complex algorithm. The study provides practical guidance for manufacturing companies that want to introduce AI-based visual inspection while maintaining local competence and long-term system ownership.