Source-linked AI summary
Human-Centric Artificial Intelligence Architecture for Industry 5.0 Applications
Jože M. Rožanec, Inna Novalija, Patrik Zajec, Klemen Kenda, Hooman Tavakoli, Sungho Suh, Entso Veliou, Dimitrios Papamartzivanos, Thanassis Giannetsos, Sofia Anna Menesidou, Ruben Alonso, Nino Cauli, Antonello Meloni, Diego Reforgiato Recupero, Dimosthenis Kyriazis, Georgios Sofianidis, Spyros Theodoropoulos, Blaž Fortuna, Dunja Mladenić, John Soldatos
TL;DR
Manufacturing lacks an architecture that places safety, trustworthiness, and human-centricity at its core. The paper proposes and framework-aligns a modular architecture integrating AI, simulated reality, decision-making, and user feedback, then validates it through real-world use cases. The validation demonstrates feasibility and interplay among modules, while further work is needed on workers’ safety support.
Problem
Manufacturing lacks an architecture specification addressing trusted and secure AI while seeking human-machine synergies through humans-in-the-loop.
Method
The paper proposes a modular manufacturing architecture built around safety, trustworthiness, and human-centricity, integrating AI, simulated reality, decision-making, feedback, and reference-architecture alignment.
Results
The architecture’s feasibility was validated through real-world use cases, demonstrating AI applications and interplay among modules for a human-centric Industry 5.0 experience.
Takeaways & Limitations
The architecture supports a human-centric manufacturing experience aligned with the Industry 5.0 paradigm.
Takeaways & Limitations
Further research is required to hone and highlight how the architecture supports workers’ safety.
Abstract
from arXiv · showhide
Human-centricity is the core value behind the evolution of manufacturing towards Industry 5.0. Nevertheless, there is a lack of architecture that considers safety, trustworthiness, and human-centricity at its core. Therefore, we propose an architecture that integrates Artificial Intelligence (Active Learning, Forecasting, Explainable Artificial Intelligence), simulated reality, decision-making, and users' feedback, focusing on synergies between humans and machines. Furthermore, we align the proposed architecture with the Big Data Value Association Reference Architecture Model. Finally, we validate it on three use cases from real-world case studies.
1. Introduction
Manufacturing digitalization and diverse use cases create a need for interoperable standards and unified architectures. The paper develops a human-centered, trusted, and secure architecture aligned with established reference frameworks.
- Digitalization has enabled manufacturing technologies and Industry 4.0 functionalities such as mass customization, predictive maintenance, and zero-defect manufacturing.
- Manufacturing’s shared challenges require standards that support component interoperability and the application of best practices.
- Existing reference architectures divide responsibilities across Industry 4.0 building blocks, interoperable IoT systems, security, safety, and big-data guidance.
- The paper evolves an architecture for trusted and secure manufacturing AI that seeks human-machine synergies through humans-in-the-loop.
- The proposed architecture maps its modules to the BDVA reference architecture and ISSF framework to combine their complementary views.
2.1. Industry 5.0
Industry 5.0 shifts manufacturing beyond technology-centered efficiency toward a value-driven paradigm emphasizing human-centricity. Its realization requires systematically combining human and machine strengths, digital twins, and actionable AI outputs.
- Industry 4.0 emphasized technologies for operational efficiency, productivity, and competitiveness, while Operator 4.0 introduced assistance intended to relieve physical and mental stress.
- Industry 5.0 is presented as a co-existing industrial revolution with visions centered on human-robot co-working or a bioeconomy based on renewable biological resources.
- The paper focuses on Industry 5.0 as a value-driven manufacturing paradigm that highlights human-centricity.
- Realizing Industry 5.0 requires combining human and machine strengths, creating digital twins of entire systems, and using AI to generate actionable items for humans.
2.2. Considering standards and regulations
Industry 5.0 architectures must account for regulations and standards governing compatibility, interoperability, cybersecurity, data privacy, and AI use. The paper outlines these requirements but does not provide a systemic architectural solution for data-management regulations.
- Existing regulations and standards should be considered so Industry 5.0 building blocks can be universally understood, adopted, compatible, and interoperable.
- Cybersecurity is treated as a transversal concern, with ISO 27000, CISA, the EU Cybersecurity Act, and NIS II identified as relevant frameworks or regulations.
- Data management must address privacy through frameworks including GDPR, the ePrivacy directive, and the Data Governance Act, alongside anonymization and synthetic-data practices.
- The EU Artificial Intelligence Act categorizes AI applications as unacceptable-risk, high-risk, or outside those categories, with different regulatory consequences.
- The listed regulations and concerns provide a high-level view rather than an exhaustive treatment.
2.3. Enabling technologies
The paper identifies active learning, explainable AI, simulated reality, conversational interfaces, and security as technologies supporting human-centric manufacturing. These technologies combine human expertise, interpretable decisions, synthetic or simulated data, and safer human-machine collaboration.
- Active learning: Active learning selects a small number of data instances so human expertise can improve AI models through a human-in-the-loop.Its assumptions include queryable learning, abundant unlabeled questions, and limited capacity to answer them.
- Simulated reality: Synthetic-data methods, including GANs, address expensive or scarce manufacturing data by generating samples resembling real data.GANs use a generator and discriminator, with discriminator feedback guiding higher-quality sample generation.
- Active learning: Active learning remains relatively scarce in manufacturing research despite applications in quality control, predictive modeling, and demand forecasting.Reported uses include gathering user input for visual inspection and informing logisticians about demand-relevant events.
- Explainable Artificial Intelligence: Explainable AI conveys model rationale so users can assess forecast trustworthiness and make decisions responsibly.The paper notes that manufacturing explanations should better account for human-environment interactions, including workers’ physical and psychological states.
- Simulated reality: Simulated reality generates synthetic data or alternative scenarios and supports reinforcement learning without costly or unsafe real-world interactions.Simulations can also project action consequences and validate desired outcomes before deployment.
- Intention recognition in manufacturing lines: Wearable sensors support worker activity, intention, fatigue, stress, and well-being analysis, while movement prediction helps cobots avoid collisions and injuries.Human movement prediction is especially important for collaboration in open workspaces where contact with moving equipment can cause harm.
3. Safe, Trusted, and Human-Centered Architecture
The proposed modular architecture centers Industry 5.0 on safety, trustworthiness, and human centricity. It maps modules to BDVA components and integrates simulation, AI, storage, interfaces, feedback, and transversal cybersecurity.
- Architecture values-based principles: The architecture is designed around safety, trustworthiness, and human centricity as desired characteristics for Industry 5.0 manufacturing environments.Trustworthiness includes transparency, reliability, availability, safety, and integrity, while human-centricity places people at the center of production.
- Architecture values-based principles: Cybersecurity is positioned at the intersection of safety and trustworthiness because attacks can disrupt manufacturing systems and data.The architecture treats cybersecurity as a transversal concern implemented through security policies and a policy manager.
- Architecture for safe, trusted, and human-centric manufacturing systems: The architecture complies with the BDVA reference architecture and organizes modules around AI, simulated reality, decision-making, feedback, user interaction, and storage.Its modules evolved from use cases and are depicted as interacting with physical and digital worlds, manufacturing platforms, and digital-twin capabilities.
- Architecture for safe, trusted, and human-centric manufacturing systems: The Simulated Reality Module generates synthetic data or alternative scenarios for data enrichment, reinforcement learning, and what-if analysis.Simulation can project outcomes based on potential user decisions and reduce dependence on complex real-world environments.
- Architecture for safe, trusted, and human-centric manufacturing systems: The Feedback Module collects explicit or implicit user feedback on forecasts, explanations, and recommended decisions, interacting directly with active learning.User interaction occurs through a multimodal interface, while storage may use databases, filesystems, or knowledge graphs.
- Architecture for safe, trusted, and human-centric manufacturing systems: Storage, simulation, and forecasting can provide digital twins with behavior models for humans, machines, and manufacturing processes.Forecasting also supports recognition and prediction of workers’ intentions and movement trajectories.
4. Validating use cases
The architecture is validated through demand forecasting, quality inspection, and human intention recognition use cases. Together, they address planning, defect detection, human-machine collaboration, and manufacturing safety.
- Use cases overview: Three validating use cases cover demand forecasting, quality inspection, and intention recognition, aligned with Industry 5.0 values of human centricity, trustworthiness, and safety.The first two use cases specifically target collaboration between humans and machines using active learning and explainable AI.
- Demand forecasting: Demand forecasting uses historical and other information to estimate future customer demand and reduce inefficiencies such as excess stocks or shortages.The study uses data from a European original equipment manufacturer serving the global automotive market.
- Demand forecasting: Forecasts remain decision-support tools: planners retain responsibility, while explanations of model rationale support responsible decisions and may be legally required.Forecast adjustments and their reasons can be recorded for retrospective evaluation and model improvement.
- Quality inspection: Quality inspection uses shaver-logo images to detect double-printing and interrupted-printing defects.The Philips dataset focuses on printed logos and the two classified printing-quality defects.
- Quality inspection: The quality-inspection study targets automated visual inspection, class imbalance, model-rationale understanding, and improved manual revision.Defective samples become scarcer as manufacturing quality improves, limiting data available for defect-detection training.
- Intention recognition and robot safety: SmartFactoryKL combines intention recognition, dynamic robot reconfiguration, and path planning to create safer navigation around workers and equipment.The system uses human activity prediction, changing layouts, object positions, robot speed, and safety zones to support collision-free paths.
5. Experiments and Results
The experiments apply the architecture to demand forecasting, quality inspection, and human activity recognition, combining prediction, explanation, decision support, synthetic data, active learning, and user feedback. Results include improved forecasting precision, strong inspection performance, reduced unidentified defects, and wearable-sensing feasibility, alongside data-quality and experimentation challenges.
- Demand Forecasting: Demand forecasting combined model development, explainability, decision-option recommendation, and a voice interface for supply-chain use.Forecasting models covered smooth, erratic, lumpy, and intermittent product demands using real-world automotive-manufacturer data.
- Demand Forecasting: More than 30% greater precision in predicting demand occurrence came from a two-fold approach for lumpy and intermittent demand.The improvement also produced gains under Stock-keeping-oriented Prediction Error Costs.
- Demand Forecasting: XAI explanations mapped forecast-relevant features to intelligible high-level concepts while preserving feature-importance rankings.The explanations also incorporated media-news information and potentially useful open datasets; the models achieved state-of-the-art performance.
- Challenges: Data acquisition and data quality were major challenges in developing demand-forecasting models and their explanations.The work required multiple environments, application programming interfaces, query constraints, and repeated validation iterations.
- Quality Inspection: Quality inspection used automated visual models, active learning, simulated reality, and defect-location hints from GradCAM, DRAEM, or similar labeled images.Synthetic images addressed class imbalance and supported a manual revision prototype that collected user feedback.
- Quality Inspection: 0.9792 AUC ROC was achieved by the best batch model, while the best streaming model lagged by at least 0.16 points in active learning.The batch model was a multilayer perceptron and the streaming model was streaming kNN; both used ResNet-18 embeddings.
- Quality Inspection: More than 80% fewer unidentified defects were reported after using the manual revision process with defect-identification cues.Future work will examine discovering new defects from user feedback and detecting fatigue to suggest task alternation or breaks.
- Human Behaviour Prediction and Safe Zone Detection: A wrist-worn sensing prototype and neural models were used to evaluate human activity recognition and movement-trajectory prediction.An adversarial encoder-decoder with maximum mean discrepancy was tested on four open datasets, and reported results outperformed state-of-the-art methods.
6. Conclusions
The proposed architecture supports human-machine collaboration through forecasting, explainable AI, active learning, simulated reality, decision-making, and human feedback. Its feasibility was validated across real-world manufacturing use cases, while worker-safety support remains an area for further research.
- Conclusions: The architecture is intended to combine human and machine strengths through modules for forecasting, explainability, active learning, simulated reality, decision-making, and human feedback.This systemic approach is positioned as support for human-machine collaboration in Industry 5.0.
- Conclusions: Three real-world use cases validated the architecture’s feasibility across demand forecasting, quality inspection, human behavior prediction, and safe zone detection.The reported experiments showed interplay between architecture modules in delivering a human-centric experience.
- Future Work: Future work targets human intention recognition, cybersecurity-oriented active learning, and machine-learning-based fatigue monitoring.These directions address workers’ safety, cyberattack assessment, and worker well-being in manufacturing settings.
Funding
The work was supported by the Slovenian Research Agency and the European Union’s Horizon 2020 programme through the FACTLOG and STAR projects.
- Funding: Funding came from the Slovenian Research Agency and European Union Horizon 2020 projects FACTLOG and STAR.The projects were supported under grant agreements H2020-869951 and H2020-956573.
- Disclaimer: The document belongs to the STAR consortium and requires formal approval for distribution or reproduction.The report states that its content reflects only the authors’ view and limits European Commission responsibility.