Towards Multimodal Foundation Models for Integrated Manufacturing Operations: A Cross-Domain Transformer Approach

Research Project, 2026

The project investigates a cross-modal Industrial AI architecture that can learn shared and transferable representations from heterogeneous manufacturing data, particularly industrial time-series signals and unstructured textual data such as maintenance logs and quality reports. Current industrial AI solutions are typically developed as separate, task-specific models, limiting knowledge transfer across interconnected manufacturing domains.

The project will design a transformer-based shared embedding architecture to integrate time-series and textual data and investigate whether a pretrained backbone can transfer knowledge across downstream manufacturing tasks, such as predictive maintenance, quality control and process optimization. A focused proof of concept will assess the feasibility of the approach on representative industrial data and generate architectural insights toward future industrial multimodal foundation models. The work is connected to the AIMOps project and supports the development of more reusable, integrated, and scalable AI solutions for manufacturing operations.