ABSTRACT The transition from Industry 4.0 to the value‐driven human‐centric Industry 5.0 paradigm increasingly relies on autonomous AI agents for complex decision‐making. However, existing research remains fragmented within isolated operational silos. This systematic review aims to unify this landscape by analysing the deployment of agentic architectures encompassing classical multiagent systems, deep reinforcement learning and large language models (LLMs) across production scheduling, maintenance management and supply chain optimisation. Following PRISMA 2020 guidelines, we systematically analysed 118 peer‐reviewed studies published between 2015 and 2026. The findings reveal a critical maturity asymmetry: although agentic methods are highly advanced in diagnostic monitoring and localised execution, a structural ‘deployment gap’ persists at the real‐time prescriptive decision‐making step across all three domains. This bottleneck is primarily driven by unresolved vulnerabilities in algorithmic explainability, safety verification and cybersecurity. As a central novelty, this review introduces a cross‐domain taxonomy that categorises agent architectures and synthesises their shared cyber‐physical dependencies, including digital twins, edge‐distributed computing and semantic interoperability standards. By evaluating these traditionally isolated literature together, we expose performance discrepancies where identical architectures yield markedly different outcomes depending on the operational context. Finally, we propose a comprehensive research roadmap to bridge the deployment gap, emphasising the need for robust explainable AI (XAI), federated learning for cross‐organisational data governance and sociotechnical frameworks that ensure genuine human–agent collaboration. Ultimately, closing this gap is essential for realising the fully self‐organising resilient industrial networks envisioned by Industry 5.0.
ABSTRACT Intelligent fault diagnosis (FD) of mechanical equipment is vital for ensuring the safety and stability of industrial systems. As modern industrial environments continue to expand in scale and complexity, systematic FD increasingly requires not only accurate signal pattern recognition but also flexible human–machine collaboration for task configuration, model selection and decision optimisation. Traditional static diagnostic systems based on small‐scale specialised models (SSMs) are often inadequate for dynamic, multisource and heterogeneous diagnostic tasks. To address these challenges, this paper proposes a large language model (LLM)‐enabled human–machine collaborative fault diagnosis framework that combines multimodal task interpretation, signal‐source identification and dynamic selection of registered SSMs. The LLM operates at the task‐cognition level, whereas structured task checking, model applicability validation and human review for uncertain cases prevent uncertain cognitive outputs from directly triggering deterministic diagnosis. Experiments on two public datasets and an industrial spindle‐motor dataset show that the agent achieves 87.33% SSM selection accuracy and 75.00% end‐to‐end diagnosis accuracy without additional labelled samples for training a separate router, outperforming deterministic rule routing. These results indicate a feasible architecture for heterogeneous industrial diagnosis that combines flexible LLM‐based task understanding, deterministic SSM diagnosis and explicit human review for uncertain tasks.
ABSTRACT To effectively mitigate stochastic disturbances during production and ensure the on‐time delivery of orders, manufacturing systems must demonstrate resilience by rapidly generating revised production schemes within critical time windows. Although traditional scheduling methods facilitate system recovery, existing centralised approaches often suffer from computational paralysis and decision failure due to excessive rescheduling frequencies. Conversely, distributed multi‐agent methods are frequently constrained by limited local observability, leading to suboptimal convergence. To address these challenges, this study proposes a multi‐agent deep reinforcement learning (MADRL) scheduling method based on dynamic information entropy‐driven QMIX (DE‐QMIX). First, the flexible job‐shop scheduling problem is formulated as a partially observable markov decision process (POMDP), empowering individual machine agents to respond to disturbances based on localised perceptions. Second, a self‐attention mechanism is embedded within the agent architecture to optimise feature weight allocation, effectively suppressing policy degradation in high‐dimensional observation spaces. Finally, the cascading effects of disturbance events are inferred using a shop‐floor job network adjacency matrix. By integrating node‐level information entropy to quantify global state fluctuations, the proposed method significantly enhances the operational robustness of the system under multi‐disturbance scenarios.
ABSTRACT Digital twin technologies are gaining importance in modern intelligent manufacturing to facilitate increased efficiency, smartness in manufacturing, and sustainability. Nevertheless, the choice of a suitable digital twin solution is a complicated long‐term decision because of the high rate of technological development, integration issues, and financial limitations. To solve these problems, this research proposes a multi‐criteria decision‐making model based on the circular q‐rung orthopair fuzzy compromise of ideal solution (CqROF‐CoCoFISO). The proposed model uses circular q‐rung orthopair fuzzy sets, which is an effective model for capturing the uncertainty, hesitation and ambiguity that exists in smart manufacturing environments. The most important assessment criteria will be the technological flexibility, compatibility with the system, the cost of implementation, functionality of the system, intellectual property protection, adherence to security and sustainability. The CoCoFISO mechanism will allow ranking competing digital twin alternatives reliably based on conflicting criteria, which will improve the strength and stability of decision‐making results. The usefulness of the proposed framework can be proved through a comparative analysis. The findings offer useful information to manufacturing engineers, decision‐makers and policymakers aiming at informed and credible digital twin adoption policies in changing industrial conditions.
ABSTRACT In the realm of Industry 5.0, the concept of Operator 5.0 emphasises the role of technologies enhancing human capabilities and adaptability. Augmented Reality, which shows potential to improve productivity, cognitive support, and workplace efficiency is one promising technology, being able to provide real‐time access to instructions, facilitate assistance, and enhance operator performance. Nevertheless, its industrial adoption faces several challenges, including the need for specialised expertise in system design and concerns regarding the technology’s usability. To fulfil the gaps related to the development of human‐centric AR solutions, this study explores the impact of AR‐based guidance in assembly processes, through an experimental approach. A test conducted in laboratory involving 24 users demonstrates the AR's effectiveness in reducing errors and achieving high rates of usability, even if time increases in the tasks' completion were observed. The findings offer theoretical and practical insights to foster successful implementation of AR in manufacturing, understanding its benefits and challenges and highlighting its potential to optimise assembly operations and enhance workforce wellbeing and resilience.
