The tightly coupled cyber network and service network in the industrial internet of things (IIoT) system are key factors influencing its reliability against cascade failures. This cross-layer coupling property can cause failures to spread rapidly within and between layers of the IIoT, thereby posing a serious threat to the manufacturing process. To reduce the vulnerability to cascade failures in the IIoT, a graph-embedding reinforcement learning-based topology optimization (GRLTO) scheme is proposed. In this scheme, targeting the topological heterogeneity between the two-layer networks, hierarchically adapted agents are introduced to perceive the network states of each layer and implement topology optimization. To address the issue of information sharing among agents in high-dimensional states, a graph neural network (GNN)-based agent information interaction scheme is proposed. This scheme uses graph convolutional network (GCN) to extract features from the local states of individual agents, and then employs graph attention network (GAT) to achieve global feature interaction among multiple agents. Furthermore, a multi-agent reinforcement learning (MARL) framework is presented to train agents to learn collaborative topology optimization strategies. Extensive experiments validate that the proposed GRLTO scheme significantly improves the reliability against cascade failures in IIoT systems.
The integration of Information Technology (IT) and operational technology (OT) in Industry 4.0 has made manufacturing systems more susceptible to cyber-physical attacks, where attacks occur in the form of both network issues and deviations in the real-world production process. The current security methods are still predominantly IT-based, whereas Retrieval-Augmented Generation (RAG) systems use textual similarity and do not consider the real-life limitations of industrial processes. Such disconnection may lead to operationally infeasible or unsafe AI-generated suggestions. This paper proposes a physics-based Agentic RAG framework for anomaly detection in cyber-physical manufacturing systems. Unlike conventional AI-based approaches that rely primarily on data-driven or semantic similarity methods, the proposed framework integrates deterministic Max-Plus Algebra (MPA) with constraint-aware knowledge retrieval to ensure physically consistent and operationally feasible decision support. The architecture consists of a five-layer pipeline that combines real-time system monitoring, MPA-based anomaly detection, semantic query construction, and RAG-driven reasoning. The framework is evaluated using a large-scale synthetic manufacturing dataset comprising 432,000 time-series observations with simulated cyber-attack scenarios, including distributed denial-of-service, sensor spoofing, and system freeze events. Experimental results demonstrate strong detection performance, achieving a precision of 0.967, recall of 0.777, and F1-score of 0.862. In addition, the integration of resource- and shift-aware constraints significantly improves the reliability and feasibility of AI-generated mitigation strategies. The findings highlight the effectiveness of combining physics-based modeling with generative AI for enhancing resilience, safety, and decision-making in smart manufacturing environments. The proposed approach provides a scalable and practical solution for next-generation Industry 4.0 systems.
In-line optical inspection on nanoscale process defects of patterned wafers is crucial for the yield management and control in advanced high-volume semiconductor manufacturing. Current industrial inspection modes mainly rely on die-to-die or die-to-database comparisons. These methods are faced with severe challenges due to limited optical resolutions compared to shrinking process nodes and comparison-induced efficiency loss. There is a strong need for high-efficiency and intelligent methods for the optical detection on nanoscale physical defects in patterned wafers. To address this gap, this paper suggests a deep learning driven approach without in-line reference comparisons. The suggested approach also includes methods for efficient bright-field inspection image simulation and pseudo-random defect dataset generation. To capture the nanoscale defect features from their low-resolution diffractive images, an improved YOLOv8 model is proposed by incorporating attention mechanisms to enhance its local learning capability while maintaining its inherent global vision. The proposed defect inspection approach can achieve simultaneous recognition, localization and classification of typical pattern defects. The experimental results demonstrate the effectiveness of our model and algorithm as a promising tool for the in-line detection of nanoscale defects in patterned wafers. This work provides a viable, reference-free inspection strategy that could enhance robustness and adaptability in real semiconductor manufacturing environments, supporting more intelligent and efficient yield management.
In next-generation intelligent manufacturing, process parameter design—crucial for product quality and efficiency—is evolving profoundly from experience-driven methods to intelligent, data-driven approaches. However, the expansion of high-dimensional, strongly coupled process windows, along with the “trust crisis” in engineering deployment caused by the “black-box” nature of data-driven models, has emerged as critical bottlenecks. This study provides a narrative and critical review of manufacturing process optimization research and proposes an integrated three-generation evolutionary framework describing the progressive evolution and integration of optimization paradigms: (1) Traditional Experimental and Statistical Optimization—Taguchi and response surface methodology mitigate blind trial-and-error but are limited in high-dimensional nonlinear contexts; (2) Prediction–Search Intelligent Optimization—surrogate models and deep learning enable efficient global optimization but lack interpretability; (3) Explainable Optimization Design—symbolic regression and post-hoc explanations based on explainable artificial intelligence quantify parameter contributions and interactions, bridging optimization and practical engineering. A comparative analysis highlights the limitations of each paradigm and points to emerging frontiers: transfer learning for small samples, multimodal knowledge fusion via large models, and online adaptive control through digital twin. This work provides a foundation for advancing intelligent process optimization in manufacturing.
Industry 4.0 manufacturing systems require intelligent control architectures capable of real-time adaptation, predictive maintenance, and autonomous resource coordination. This paper proposes MASIM (Multi-agent system for intelligent manufacturing), a four-layer hierarchical framework that integrates XGBoost-based predictive maintenance, double deep Q-network (DDQN) resource allocation, and Contract Net Protocol (CNP) negotiation within a distributed multi-agent architecture. The contribution is located in the coupling between these layers rather than in their juxtaposition: we introduce a health-aware sequencing index (ATCS-H), which discounts the classical Apparent Tardiness Cost with Setups rule by a calibrated failure probability, and a risk-adjusted Contract Net bid, through which a degrading machine withdraws itself from negotiation before it must be taken offline. The reinforcement-learning formulation is verified rather than assumed: the Markov property of the state vector is tested empirically, the reward is re-derived so that every term is O(1) with a potential-based shaping component whose policy invariance is proved, and the learned value function is shown to be uniformly bounded. Beyond the binary assign/defer gate, a factorised top-K action space is introduced so that the agent selects which job starts on which machine. The predictive-maintenance module is evaluated on two real industrial datasets and benchmarked against five alternative learners, including a CNN-LSTM; it attains a ROC-AUC of 0.735 on a predictive-maintenance dataset and a 5-fold cross-validated F1 of 0.279 on the UCI SECOM dataset, and its recall rises from 0.20 at the F1-optimal threshold to 0.45 at a 5 p < 0.001 in both cases), attaining the lowest tardiness in all four scenarios, at machine utilisation and energy consumption that are statistically indistinguishable from the centralised baseline and at a throughput cost of 8.7
