The diffusion of Artificial Intelligence (AI) within the Industrial Internet of Things is reshaping industrial systems and accelerating the transition toward human-centered innovation. Yet research on AI adoption in industrial environments remains fragmented, predominantly technology-centered, and largely based on context-bound empirical settings. Existing studies often examine isolated variables, limiting generalizability and overlooking how employee-related antecedents, alternative human-AI integration mechanisms, and their positive and negative consequences jointly influence process innovation and information integration. This study addresses these gaps by developing an integrated Antecedent-Process-Outcome framework of employee-AI integration. A taxonomy of 70 constructs is consolidated from the literature and empirically examined through the coding of 162 AI-oriented industrial practices documented in peer-reviewed case studies and application-based research. Using multi-layer association analyses, the study reconstructs a highly interconnected network linking employee capabilities, organizational conditions, distinct integration mechanisms, and multidimensional effects, including performance, knowledge development, engagement, and adverse reactions. The findings highlight a systematic distinction between augmentation-oriented and automation-oriented configurations. Collaborative and augmentative mechanisms of human-AI collaboration show stronger alignment with enabling conditions and positive knowledge and performance dynamics, whereas automation-oriented arrangements display weaker connections with human-centric outcomes and learning trajectories. Network analysis further identifies structurally central constructs that differ from those most frequently emphasized in the emerging worker-AI integration literature, offering a more evidence-based prioritization of integration dimensions. Human-machine collaboration emerges as a socio-technical configuration in which preparatory conditions, agency distribution, and both beneficial and adverse effects jointly shape sustainable process innovation and long-term adaptive capacity.
Pipelines are susceptible to mechanical damage that can compromise structural integrity and flow performance, highlighting the need for digital twin frameworks capable of converting sparse field measurements into reliable performance predictions. This study develops a smart composite liner-integrated digital twin framework to address this need through three core methodological innovations. First, a machine-learning-enhanced Thin Plate Spline (ML-TPS) method reconstructs full-field circumferential strain from sparse distributed fiber optic sensor (DFOS) measurements. By combining FEA-informed dent state identification with learned residual correction, the approach adaptively conditions spline weighting to preserve sharp strain gradients while maintaining differentiability. Second, a strip-based Kirchhoff–Love Plate Theory inverse finite element method (KLPT-iFEM) reconstructs three-dimensional dent geometry from the strain field using curvature-based stripwise deformation and circumferential continuity constraints. Third, the reconstructed geometry enables integrated multi-domain performance evaluation, including finite element analysis (FEA) for structural integrity, computational fluid dynamics (CFD) for flow performance, and fatigue analysis for remaining useful life prediction. Two practical sensor layouts, 4-path and 3-path, are evaluated to balance reconstruction fidelity and deployment cost. Displacement-controlled indentation tests with 3D scanning validation demonstrate reliable dent depth reconstruction from 1 mm to 24 mm, corresponding to 12% of the pipe outer diameter. The reconstructed geometry supports FEA-based structural capacity assessment, CFD-based flow performance evaluation, and remaining useful life estimation for predictive maintenance.
Achieving efficient and reliable automation in design and manufacturing processes remains a key challenge for maintaining industrial competitiveness. While several previous simulation- and experiment-based approaches have played a crucial role in design validation and manufacturing optimization, their repetitive procedures, significant time and cost requirements, and insufficient automation across process stages constrain their applicability and system efficiency in industrial environments. This study proposes a large language model (LLM)-based multi-agent framework that autonomously integrates design and manufacturing through collaborative interactions among specialized agents. As a representative use case of automotive design and manufacturing, the proposed framework is designed to facilitate the entire workflow by enabling the design, evaluation, and demand agents in the design department to derive optimal designs that reflect user requirements, while the production and retrieval agents in the manufacturing department perform cost estimation and verification. Both quantitative and qualitative results demonstrate that the proposed framework improves design selection and manufacturing cost estimation performance compared with commercial LLMs. In particular, the design department reduces the target drag coefficient deviation by approximately 71% compared with the best-performing baseline, while the manufacturing department achieves an exact match accuracy of 96%. At the end-to-end level, the framework achieves a 100% task success rate and maintains at least 90% decision accuracy under design and manufacturing-related perturbations. These results demonstrate reliable workflow completion and decision-making through agent interaction and feedback. The proposed framework seamlessly connects the entire design-manufacturing process, highlighting its potential as a practical and intelligent pipeline for industrial applications.
Digital twins (DTs) could support directed energy deposition-arc (DED-arc) manufacture by connecting process monitoring, data management and simulation, but most implementations address isolated tasks rather than extensible process-level integration. This study develops and demonstrates a modular DT framework for DED-arc using NVIDIA Omniverse as the virtual environment and data-management platform. The framework combines Omniverse Nucleus for asynchronous storage of experimental and simulation data with a custom MQTT-based streaming layer for time-sensitive robot and deposition signals. A DED-arc manufacturing cell using a KUKA KR70 robot was reproduced in the virtual environment, with live synchronisation of robot motion and arc state. The implemented data pipeline achieved an average streaming latency of approximately 520 ms for robot and arc-status data, while a 192 MB transducer data file was uploaded to Omniverse Nucleus in 18 s, equivalent to approximately 10.7 MB/s. These results demonstrate that the architecture can separate low-latency monitoring data from larger asynchronous experimental datasets while maintaining a unified digital thread. The framework also demonstrates extensibility by using DT-acquired thermocouple and as-built data to support thermo-mechanical finite element simulation workflows for residual stress and distortion assessment. The results show that the proposed architecture provides real-time process monitoring, traceable data storage and simulation coupling within a single platform, without replacing the existing manufacturing-cell communication infrastructure. This establishes a practical data backbone for future surrogate modelling, quality prediction and closed-loop control in DED-arc manufacturing.
Quality prediction in metal additive manufacturing (AM) has conventionally relied on data-driven models that map process parameters to defect classes or quality metrics. However, these models often fail to generalize across different machines, materials, and process regimes. The emerging knowledge-driven approach based on large language models (LLMs) can interpret literature and expert guidance, yet struggles to deliver quantitative decisions tied to part-specific parameter sets. To bridge this gap, we propose AM-Agent, a neuro-symbolic LLM-agent framework that unifies data-driven and knowledge-driven quality control to enhance adaptivity in part-specific pre-build process planning for Laser Powder Bed Fusion (LPBF). AM-Agent treats both paradigms as independently callable services over a shared defect label space. A Data-driven Prediction Service (DPS) exposes melt-pool regressors and defect classifiers under a model-per-material design, in which the supervising LLM adaptively invokes the predictor matched to the current powder material. A Knowledge Service (KS) grounds its reasoning in process literature via retrieval-augmented generation and in dynamic digital-twin context retrieved from the Asset Administration Shell (AAS). To harmonize the two paradigms’ competing predictions, we formalize the fusion as a Linear Opinion Pool (LOP) with source-intrinsic reliability weights. The DPS weight is an entropy-and-margin calibration proxy modulated by a distribution-shift indicator, and the KS weight is the LLM’s evidence-grounded self-reported reliability. In-domain and out-of-domain experiments on the LPBF benchmark show that AM-Agent preserves DPS performance in-domain and recovers macro-F1 under cross-material distribution shifts. Conflict analysis further demonstrates how the LOP recovers from data-driven errors by deferring to retrieved physics-grounded evidence when the DPS flags uncertainty.
Modern railway maintenance increasingly requires diagnostic systems that not only detect anomalies but also align decisions with established engineering standards and remain consistent across inspection cycles; however, many existing solutions rely either on operational alarm thresholds or on purely data-driven detectors, with limited integration between numerical inference and domain semantics. This study proposes a Multi-Agent Cognitive Digital Twin (MA-CDT) architecture designed to distribute cognitive responsibilities across specialised agents linked through shared semantic representations to support traceable information flow between physical inspection processes and digital reasoning components. To provide an initial implementation of the proposed architecture, its diagnostic component was instantiated and empirically evaluated using real operational track-geometry inspection data. The implemented module integrates neural anomaly detection, engineering intervention thresholds and ontology-guided semantic validation within a single decision structure. Comparative experiments against an operational alarm baseline and a neural-only configuration show that the integrated design reduces unsupported alerts while preserving threshold-defined intervention detections under noisy and imbalanced sensing conditions. Ablation analysis isolates the contributions of neural inference, rule-based constraints and semantic validation, demonstrating how their structured interaction stabilises decision behaviour. The diagnostic case study illustrates how one cognitive agent can operate within the proposed semantic framework and exchange information through the architectural knowledge structures. The results provide an initial empirical validation of the diagnostic component as a proof of concept within the proposed MA-CDT architecture and suggest that embedding domain constraints into neurosymbolic diagnostic reasoning could enhance decision stability and traceability in threshold-driven maintenance systems.
Transdermal drug development is substantially constrained by three major challenges, including the formidable skin barrier, complex nonlinear interactions among formulation components, and instability in manufacturing processes. Traditional empirical and mechanistic models often demonstrate limited generalizability across different protocols and chemical spaces. Objective This review aims to systematically map the applications of machine learning (ML) throughout the transdermal drug development pipeline, identify key challenges related to data quality, validation, interpretability, and regulatory translation, and propose a regulatory–aligned framework for integrating ML into quality–by–design (QbD) and process analytical technology (PAT) strategies. Methods A systematic literature review covering studies published between 1996 and 2026 was conducted, with critical synthesis of ML models, evidence–based applications, and emerging future directions. Results Tree–based ensemble models, including random forest (RF) and XGBoost, appear particularly suitable for small and heterogeneous transdermal datasets. Hybrid physics–informed ML approaches improve extrapolation capability, while deep learning demonstrates strong performance in image–based evaluation tasks. Major challenges include the lack of standardized datasets, variability in experimental protocols, insufficient external validation, and limited interpretability associated with “black–box” models. This review further proposes a problem–algorithm–solution framework, a minimum metadata schema for data harmonization, and a regulatory–oriented roadmap to support industrial translation. Conclusion ML has the potential to enable data–driven, end–to–end digitalization of transdermal development. This review provides a unified, practice–oriented, and regulatorily compliant guide to facilitate the reliable, interpretable, and scalable implementation of ML in transdermal drug delivery research and development.
Diesel engines in open-pit mines frequently experience compound faults due to long-term operation under harsh environments and high-power conditions, posing severe challenges to autonomous perception, dynamic decision-making, and intelligent operation and maintenance. To address the issues of severe feature redundancy, low model interpretability, and the incomplete embodied intelligence system construction in existing methods, a dynamic feature selection method that fuses hierarchical feature pre-screening with a Deep Q-Network (DQN) from the perspective of embodied intelligence is proposed. In the proposed method, a multi-domain fusion feature library is established, and a dual-screening mechanism based on Analysis of Variance (ANOVA) and correlation analysis is introduced to effectively eliminate redundant interference. The DQN is employed to reformulate the feature selection as a sequential decision-making process, adaptively searching for the optimal feature subset under a predefined feature budget constraint. Building on this, a closed-loop embodied intelligence system integrating dynamic diagnosis, risk assessment, and digital twin validation is established, creating a complete “perception-cognition-decision-execution” operation and maintenance workflow. Experimental results demonstrate that the proposed method achieves a diagnostic accuracy of 89.17% on the test dataset. Ablation experiments reveal that the proposed method exhibits superior performance in terms of accuracy, computational cost, and inference time compared to baseline schemes, including “using only the DQN without hierarchical pre-screening” and “employing only pre-screening without DQN-based dynamic decision-making”. The experimental results validate the effectiveness of the proposed method for the intelligent diagnosis of compound faults in diesel engines.
Nowadays, predictive maintenance is considered an essential tool bringing sustainability to industrial environments. It not only reduces the chances of failures but also ensures stability in the production systems. However, the traditional maintenance approaches often treat machines as an independent entity and overlook their interactive relation between machines, maintenance resources, and production processes. To address this issue, the present study proposes a simulation-based predictive maintenance framework for Sustainable Industrial Internet of Things (IoT) environments. The framework is modeled using Agent-Based Modeling (ABM), where machines, sensors, maintenance units, and production coordinators are treated as interacting agents. These agents continuously monitor machine conditions, exchange information, and coordinate maintenance decisions based on real-time conditions. IoT sensors collect real-time data, which is analyzed using a TCN to predict equipment degradation and its remaining useful life. Based on these predictions, agents dynamically prioritize maintenance tasks while considering production constraints. The proposed framework is evaluated with the help of simulation using metrics such as unplanned downtime, maintenance response time, and production throughput. The simulation results show that the framework reduces unplanned downtime by 32%, improves maintenance efficiency by 27.90%, and increases production throughput by 14%. In simulation, energy consumption falls by 18%, spare-part usage drops by 22%, and life-cycle cost decreases by 15%. These outcomes confirm measurable sustainability benefits. Combining predictive analytics with coordinated decision-making ensures improved reliability, operational performance, and production sustainability in industrial IoT environments.
Understanding and anticipating consumer decision-making is increasingly important in digital markets, yet traditional survey-based and rule-based simulation approaches often struggle to capture the complexity, heterogeneity, and evolving nature of real consumer behavior. This study proposes a scalable framework for building and validating knowledge-driven synthetic consumer agents with large language models (LLMs). The framework combines data-driven prototype extraction from large-scale online reviews with theory-informed persona synthesis to construct a heterogeneous population of artificial consumer agents that capture diverse consumer motivations, preferences, and decision patterns. Each agent is instantiated as a knowledge-driven consumer representation capable of generating decisions and reasoning that reflect its underlying motivational profile. To evaluate the quality of the generated population, we introduce a multi-level validation protocol that examines decision consistency under rhetorical framing variations, persona-level behavioral discriminability through a novel Persona Sensitivity Index (PSI), and population-level thematic alignment with real-world consumer discourse. Results from a controlled validation experiment show that the generated agents maintain coherent reasoning patterns while exhibiting differentiated responses to motivation-aligned stimuli. At the aggregate level, the generated synthetic consumer population exhibits discourse-level and preference-structure alignment with review-derived consumer discourse. These findings provide preliminary evidence for the potential of LLM-based synthetic consumer agents as a scalable and reproducible testbed for consumer decision modeling, consumer response analysis, and innovation-oriented behavioral simulation.
As a core ground test platform for aerodynamic research, the wind tunnel plays an irreplaceable role in aircraft design and development. However, traditional wind tunnel testing suffers from limitations such as long cycles, high costs, and low data utilization efficiency. Although numerical wind tunnels offer advantages in flexibility and cost, they are still restricted by model credibility and real-time performance. Digital twin technology provides a new approach for the digital transformation of wind tunnel testing by establishing real-time mapping and bidirectional interaction between physical and virtual spaces. This paper systematically reviews digital twin-driven wind tunnel test technologies. It first elaborates on the definition and connotation of digital twin, and distinguishes among digital model, digital shadow, and digital twin. On this basis, a five-dimensional framework tailored to wind tunnel testing is proposed, comprising physical wind tunnel, virtual wind tunnel, data perception, data fusion, and application decision-making. Enabling technologies are comprehensively sorted out around the “perception-fusion-decision” information chain. Technologies for acquisition, processing, and transmission of inherent and real-time data are analyzed, multi-source data fusion methods are summarized, and research advances in intelligent applications such as visualization, flow field prediction, and fault diagnosis are reviewed. Finally, five major challenges restricting engineering applications are discussed: model credibility and verification mechanisms, computing resources and real-time performance, multi-source data fusion quality, interdisciplinary collaboration and standardization, and engineering application value and deployment. Research directions including lightweight modeling, multi-fidelity fusion algorithms, collaborative development platforms, and real-time interaction technologies are proposed, providing strong support for digital twin-driven wind tunnel testing.
As modern manufacturing systems become increasingly complex, flexible flow shops (FFS) face challenges posed by tightly coupled temporal and buffer constraints, dynamic disturbances, and evolving data distributions. To address these challenges, a novel bi-level multi-agent framework for dynamic scheduling is developed, integrating real-time decision-making with adaptive model maintenance. At the decision level, a graph-based state representation is used to encode production semantics and structural relationships. A two-stage predictive model, consisting of a genetic algorithm-based feature selection process and a multi-head graph attention network combined with a multilayer perceptron, is employed to accurately estimate global optimization objectives. These predictions are then used as reward signals to guide a graph-based proximal policy optimization policy, where a constraint-aware reference mechanism is incorporated to ensure feasibility under coupled temporal and buffer constraints. At the model maintenance level, a sliding window-based multi-level statistical test is used to detect concept drift, and a reinforcement learning-based adaptive retraining strategy is applied to update the model using weighted historical data, enabling sustained performance over time. The framework is validated using 1198 industrial scheduling instances collected from a textile manufacturing workshop. Experimental results demonstrate that the proposed framework outperforms representative metaheuristic and reinforcement learning methods, including GA-SAGE, DCN, D3QN, and industrial scheduling rules. It achieves a high prediction accuracy (R2 = 0.93) and a scheduling feasibility rate of 92% in real-world workshop deployment. The results highlight the effectiveness of integrating predictive modeling, constraint-aware policy learning, and drift-adaptive model updating into a unified framework for robust industrial scheduling.
The advancement and increasing reliance on information technology and cyberspace applications in both the industrial and private sectors necessitate investment in online application services, which has become imperative. This study develops an advanced risk management framework for an online service supply chain model that incorporates dual-channel distribution, addressing both B2B and B2C markets. This study develops five models that incorporate various risk types and mitigation strategies, addressing demand risk, general supply chain risk, and cybersecurity risk under conditions of information asymmetry and Stackelberg game theory between two players, the infrastructure provider and the service provider, to enhance realism. Each model is quantified through numerical experiments that demonstrate the financial impact of various risks and the effectiveness of the proposed mitigation strategies, whose optimality is validated both mathematically and graphically. The results indicate that information asymmetry and risks collectively reduce total profit by 6.1%, while the proposed mitigation strategies enable the recovery of 5.7% of this loss.
