How can artificial intelligence (AI) be used effectively when only limited data is available? This question was addressed by the innovation project ‘AI for Scarce Data – Machine Learning and Information Fusion for the Sustainable Use of Laboratory and Customer Data’ (AI4ScaDa), which has now been successfully completed. The project aimed to develop practical AI solutions for data-poor scenarios in industry. The project aimed to make it easier for small and medium-sized enterprises (SMEs) to access AI.
The project partners' results and specific applications are published in the final report on the it's OWL innovation platform.
Testing in industrial application scenarios
The feasibility of the developed methods was demonstrated in collaboration with industry partners GEA, SU BIOTEC and Miele. The AI methods were used to optimise plant cultivation processes, enable machine-assisted decision-making based on expert knowledge and evaluate time series for condition monitoring in household appliances, among other things. In all cases, AI-supported methods achieved concrete improvements without relying on large amounts of data.
Figure 1: Goals of the project partners SU BIOTEC, GEA and Miele in the AI4ScaDa project.
A modular AI workflow for practical use
The project also resulted in a modular workflow that can be used to reliably and traceably create AI models, even from small and heterogeneous data sets. The platform-independent components are user-friendly and do not require in-depth prior knowledge. Through the targeted use of active learning, data annotation and information fusion, meaningful models can be created even with limited data, for example in the form of transparent decision trees.
Detailed instructions for the AI workflow are available on the innovation platform: https://doi.org/10.5281/zenodo.15598391.
Figure 2: The workflow for handling sparse data comprises 1) experimental design with optimised SLHDs, 2) tabular preparation of the data, 3) expert evaluation, 4) analysis using decision trees (M5’, GUIDE), iterative model improvement through 5) active learning, and 6) evaluation of new data points with visualised decision tracking.
Participation by inIT
Julian Bültemeier, a research assistant in the Discrete Systems working group led by Prof. Dr. Volker Lohweg, participated in the project on behalf of inIT. He was particularly involved in developing the modular workflow.
"The project was technically exciting and demonstrated the potential of even small amounts of data. It once again highlighted how important it is for research and practice to go hand in hand. The collaboration with partners from industry and research was constructive and positive throughout – it was a great project!” said Julian Bültemeier.
Thanks to everyone involved in the project!
Special thanks go to project partners GEA, SU BIOTEC and Miele, as well as to Bielefeld University of Applied Sciences, for their successful and trusting cooperation.
Little data, lots of insight!
Further information on the project can be found at: its-owl.de/projekte/wenig-daten-viel-erkenntnis-entwicklung-von-ki-anwendungen-fuer-small-data-ai4scada/.