Abstract
The Internet of Production (IoP) promises to be the answer to major challenges facing the Industrial Internet of Things (IIoT) and Industry 4.0. The lack of inter-company communication channels and standards, the need for heightened safety in Human Robot Collaboration (HRC) scenarios, and the opacity of data-driven decision support systems are only a few of the challenges tackled in this work. We outline the communication and data exchange within the World Wide Lab (WWL) and autonomous agents that query the WWL, which is built on Digital Shadows (DS). We categorize our approaches into machine level, process level, and overarching principles. This paper surveys the interdisciplinary work done in each category, presents different applications, and offers actionable items and guidelines for future work.
Introduction: Bridging the Industrial AI Gap
While Industry 4.0 promised a revolution, the reality on the factory floor is often characterized by data silos, opaque "black-box" algorithms, and a lack of communication standards. This research identifies three critical gaps preventing AI from becoming truly actionable in manufacturing:
- The Transparency Gap: Engineers often lack trust in AI models they cannot understand or verify.
- The Data Scarcity Gap: High-quality labeled data for specific manufacturing defects is rare and expensive to produce.
- The Collaboration Gap: There is no secure, standardized way for different companies to share insights without exposing sensitive intellectual property.
To address these, the Internet of Production (IoP) framework leverages Digital Shadows—purpose-driven, aggregated data views—to bridge the gap between raw data and real-world manufacturing decisions.
A Multi-Layered Framework for Production AI
The proposed approach is structured across three distinct domains: the machine level, the process level, and overarching principles that tie them together.
Machine Level
At the machine level, the focus is on low-level control and human-machine interaction. Rather than relying on pure deep learning, the authors advocate for Gray-Box Modeling. By combining Newton-Euler physics equations with neural networks, robot dynamics can be modeled to ensure AI respects physical constraints. Additionally, Behavior Trees (BTs) are utilized for robot control, offering superior modularity and safety compared to traditional state machines in human-robot collaboration scenarios.
Process Level
The process level handles the product life cycle and the optimization of production flows. Using Process Mining (specifically the PM2 methodology), researchers analyze event logs to identify and eliminate bottlenecks. A unique challenge here is capturing "tacit" expert knowledge—the intuitive know-how of veteran workers. The authors developed a serious game-based approach to extract this decision-making logic in a structured, formal format.
Overarching Principles
These are omnipresent tools for data collection, analysis, and visualization. Key methods include Generative Adversarial Networks (GANs) for data augmentation in data-scarce environments and novel Concept Extraction methods for eXplainable AI (XAI). Concept extraction mines the latent space of a CNN to identify specific features (like pinholes or malformed edges) the model uses to make decisions, ensuring it isn't relying on background noise.
Digital Shadows and the World Wide Lab
A core innovation of this framework is the distinction between Digital Twins and Digital Shadows. While a Digital Twin is an exhaustive digital replica, a Digital Shadow (DS) is a task-specific, aggregated view of data.
This task-specific nature makes Digital Shadows faster to compute and easier to share across the World Wide Lab (WWL). The WWL acts as an infrastructure for inter-company collaboration, allowing different entities to query shared knowledge without revealing proprietary machine details. This creates a feedback loop where data extraction leads to decision support, which in turn optimizes the physical production process.
Key Results and Validation
The framework's efficacy was validated across several industrial use cases, including electric vehicle assembly and textile manufacturing:
- Bottleneck Detection: Process mining successfully identified station GA 16 as the primary bottleneck in an electric vehicle assembly line for e.GO Mobile AG.
- Data Augmentation: Using GANs to synthesize realistic training images for textile manufacturing significantly improved defect classifier accuracy in environments where real-world samples were scarce.
- User Acceptance: A user-centered planning tool for the textile industry achieved a System Usability Scale (SUS) score of 91.8, indicating that AI-driven recommendations are highly effective when the human remains the final decision-maker.
- XAI Validation: Concept extraction proved that a model trained on metal casting parts was correctly identifying "pinholes" (Importance = 1.0) rather than being biased by background shadows.
Conclusion and Future Outlook
While individual data-to-knowledge pipelines have proven successful, the next challenge is holistic integration. Current solutions for robot control, process mining, and data management remain somewhat fragmented. Future research will focus on developing the "Human Digital Shadow" to better model ergonomics and safety, as well as scaling the World Wide Lab concept to interconnect research and production sites across international borders.
Summary updated