Multimodal deployment risks in manufacturing AI: a survey of boundary failures, taxonomy, and governance in Industry 5.0

Yang, Jialu ORCID: https://orcid.org/0000-0003-1463-2677, Cai, Taotao and Shi, Kaize 2026. Multimodal deployment risks in manufacturing AI: a survey of boundary failures, taxonomy, and governance in Industry 5.0. Presented at: WWW '26: The ACM Web Conference 2026, Dubai, United Arab Emirates, 29 June - 03 July 2026. WWW Companion '26: Companion Proceedings of the ACM Web Conference 2026. Companion Proceedings of the ACM Web Conference 2026. New York, NY: Association for Computing Machinery, pp. 888-897. 10.1145/3774905.3795460

PDF - Published Version Available under License Creative Commons Attribution. Download (1MB)

Abstract

Manufacturing AI systems increasingly rely on heterogeneous and interacting signals---from physical sensors and visual inspection streams to cyber/operational logs and human--machine interaction cues. In such multimodal settings, deployment risks rarely arise from a single model or dataset; instead, they emerge at cross-modal boundaries where signals are delayed, misaligned, or inconsistently interpreted across the socio-technical pipeline. We conduct a structured review of peer-reviewed studies, industrial reports, and international standards (2010--2025), synthesising evidence from over 80 relevant sources. Our analysis shows that safety, reliability, and accountability failures are often amplified by multimodal coupling (e.g., sensor--vision conflicts, log--control inconsistencies, and human-in-the-loop misunderstandings). Based on the synthesis above, we organise the identified deployment risks into a multimodal-aware taxonomy comprising five interrelated layers: data, model, system, organisational, and governance/ethical. Rather than treating these layers in isolation, the taxonomy explicitly links risk categories to mitigation approaches that operate across modalities, including technical controls, human-centred operational practices, and standards-driven governance mechanisms (e.g., NIST AI RMF, ISO/IEC~23894, ISO/IEC~42001). In addition, the review highlights several unresolved issues in multimodal manufacturing AI, including the lack of quantitative indicators for deployment risk, limited transferability across production contexts, insufficient empirical evidence on human factors, and the absence of machine-readable representations that connect risk signals to governance actions. Viewed through the lens of Industry 5.0, these findings suggest that systematic attention to multimodal interactions is essential for deploying AI systems that genuinely support worker well-being, operational resilience, and long-term sustainability in manufacturing practice.

Item Type:	Conference or Workshop Item - published (Paper)
Date Type:	Publication
Status:	Published
Schools:	Schools > Engineering
Publisher:	Association for Computing Machinery
ISBN:	9798400723087
Date of First Compliant Deposit:	10 June 2026
Last Modified:	10 Jun 2026 09:15
URI:	https://orca.cardiff.ac.uk/id/eprint/187499

Actions (repository staff only)

Edit Item