Requirements for AI-Based Predictive Analytics Systems in an Industrial Manufacturing Context - Selection of the Most Suitable Data Fusion Method

Abstract

The digitalisation of modern manufacturing has led to an increased volume of heterogeneous data, leading to operational problems such as disconnected systems, data overload, and a lack of data-driven insights. While AI-based predictive analytics systems attempt to solve these problems, they are often limited by the need to process homogeneous data, which clashes with the naturally heterogeneous data found within manufacturing; an issue that data fusion methods can rectify. This thesis, through its method combining requirement engineering and solution selection, outlines the process used for selecting the most appropriate fusion method for fusing heterogeneous data in manufacturing. The fusion method was selected by first identifying a set of 10 requirements that the manufacturing industry has for such systems. These were identified through thematic analysis based on a combination of a literature review and interviews with industry professionals. 3 different fusion methods: early, intermediate, and late, were then evaluated against 6 of the derived requirements, which were deemed relevant to fusion method selection, and the most suitable method was selected. Late fusion was selected with an aggregated score of 4.27 compared to the other alternatives’ 3.71 and 2.60, largely due to its fulfilment of the industry’s two key requirements: processing a large number of heterogeneous data sources, and allowing traceability to motivate the source of its predictions. The difference between late and intermediate fusion was small and is subject to change as the technology and industry evolve. Late fusion was implemented in a test-bed artefact able to process heterogeneous data sources. The test-bed can scale with growing sources, with latency being bounded by the slowest source and not increasing as other, faster sources are added. It also demonstrated the ability to provide both modality level and feature level traceability.