Process industries such as metals, cement, and paper pose significant sustainability challenges due to their high resource demands and emissions. While benchmarking tools such as Best Available Techniques (BAT) help assess performance, existing approaches are often sector-specific, focused on isolated performance ranking, and influenced by subjective weighting assumptions. This work therefore shifts sustainability benchmarking from ranking-based evaluation to relational similarity-based analysis at the manufacturing-process level. This study introduces the Similarity-based Sustainability Assessment (SIMSA) framework to support transparent cross-sectoral benchmarking by comparing how unit processes relate to one another in sustainability terms. SIMSA integrates ratio-based normalisation relative to BAT values with structured similarity analysis using Relative Percentage Difference (RPD). Sustainability indicators are selected through a literature review and classified into the environmental, economic, and social pillars. A stochastic, simulation-based weighting and aggregation procedure, Optimisation-based Stochastic Weighting with TOPSIS (OSW-TOPSIS), is used to reduce reliance on directly elicited subjective preference weights. The framework is demonstrated through a case study comparing unit processes in the steel and cement industries. Results illustrate SIMSA’s ability to identify candidate transferable practices and distinguish between structural commonality and sustainability-performance similarity. This research advances transparent, relational benchmarking and supports the early-stage identification of candidate opportunities for cross-sector learning, industrial symbiosis screening, and sustainability optimisation.
The increasing relevance of circular manufacturing and repair oriented production requires adaptive and data-driven process planning approaches capable of handling highly variable products, uncertain component conditions, and incomplete process information. Conventional automation methods for assembly and disassembly are typically optimized for repetitive high volume manufacturing and are therefore not directly suitable for flexible repair environments. This work proposes a methodology for Digital Twin-based lifecycle management to support adaptive (dis-)assembly planning in circular repair factories. The methodology integrates heterogeneous lifecycle data into a unified Digital Twin environment and utilizes simulation-driven methods to evaluate robotic grasp feasibility, assembly accessibility, end effector suitability, and process constraints prior to physical execution.The novelty of this work lies in integrating lifecycle data, adaptive execution strategies, and simulation-driven planning into a unified framework for flexible circular repair operations. The framework combines robotic grasp planning, simulation-based (dis-)assembly analysis, and decision-making for robot-guided, collaborative, or manual execution. Lifecycle data further enhances efficiency through data-driven optimization methods such as assembly group and geometry-based grasp planning.While prototypical implementations are provided for robotic grasp planning and optimization methods, the overall lifecycle-driven framework and execution decision strategies remain conceptual. Validation results demonstrate grasp feasibility analysis, quantitative comparison of robot and end effector configurations, and geometric approximation through manufacturing data integration. Overall, the implemented modules indicate the potential to increase process success, and reduce replanning effort for flexible circular repair processes.
Throughput bottleneck detection is a well-established research topic. However, comparatively less attention has been paid to what should follow once the bottleneck station has been identified. In particular, limited attention has been given to diagnosing its underlying causes and identifying the internal areas whose improvement could most effectively reduce bottleneck-station cycle time. To address this gap, this paper makes two main contributions. First, it formalizes the notion of sub-bottleneck as the sub-cycle period within the selected bottleneck station with the greatest expected potential to reduce station cycle time and potentially improve the production rate when the station remains the effective throughput constraint. Second, it proposes a dynamic and data-driven framework for sub-bottleneck identification based on the Steptime methodology and a sub-bottleneck scoring algorithm. Under the Steptime methodology, each sub-cycle period is defined as the total activation time of its corresponding step in the sequential process, referred to as the Steptime. The framework is formulated for sequential processes represented according to the GRAFCET methodology standardized in IEC 60848. The framework was evaluated on an industrial production line at Fagor Arrasate S. Coop., located in Arrasate, Spain. A GRAFCET-consistent counterfactual benchmark showed that SBScore selected the candidate with the largest expected cycle-time-reduction potential in all nine product references, achieving the highest agreement among the evaluated ranking criteria. The results revealed performance-degrading patterns in the bottleneck station, including upper-tail spreads of up to 541.8 ms in some Steptimes and bimodal behavior in others.
Effective assistance in manual assembly requires perceiving the operator’s actions in real time and advising the operator flexibly. Traditional action recognition methods classify a video clip into a single holistic label for one action stream, failing to resolve the parallel dual-hand actions or expose the fine-grained semantics for downstream assistance reasoning. Moreover, conventional assistance systems retrieve predefined content keyed to recognized steps and cannot respond to an operator’s situated needs, whereas large language models (LLMs) can respond flexibly but are too slow and hallucination-prone for the real-time, safety-critical path. This paper connects the two through a compositional action representation that expresses each hand’s action as a structured tuple of action verb, manipulated object, target object, and tool. A compositional dual-hand action recognition model, CoDuAR, is proposed to predict the action tuples of both hands in a single forward pass through a shared encoder and hand-specific compositional decoders. On the HA-ViD benchmark, CoDuAR outperforms state-of-the-art recognizers that require a separate model per hand, while using 23% fewer parameters and roughly doubling the dual-hand inference throughput of the strongest parameter-comparable baseline, VideoMAE V2. Built on the compositional representation, a dual-process assistance system, Assembly Copilot, is developed. A deterministic System 1, driven by CoDuAR, maps the predictions to a symbolic task state, tracking the task progression, recommending next tasks, and detecting operational errors in real time, while an LLM-powered agent, System 2, reasons over the symbolic state to answer the operator’s queries and generate post-session performance reviews, gaining flexibility without placing the language model on the latency-critical path. A case study on a gear-plate assembly validates the functionality of the integrated system. The results indicate that the compositional action representation can serve as an effective interface between real-time perception and language-based assembly assistance within the evaluated scenario.
Industry 5.0 emphasises human-centric technologies (HCTs) as essential drivers of sustainable and resilient production. However, their specific contributions to Circular Economy (CE) strategies and the associated skill requirements are not well-defined. This paper investigates how HCTs support Circular Economy practices (CEPs) and which skills and competencies are needed for their effective implementation. A systematic literature review was conducted using Scopus and Web of Science, following established guidelines. The search employed a string that links Industry 5.0, human-centricity, and the 10 R framework of CE. After a multi-stage screening and snowballing process, 41 peer-reviewed contributions published between 2015 and 2025 were selected for analysis through a combination of bibliometric and qualitative content analysis. The review maps the main HCTs, such as AI, digital twin, XR, robotics, blockchain, and IoT, to CEPs and specific 10 R strategies. It identifies seven clusters of skills ranging from analytical and decision-making abilities to human-machine collaboration, CE-specific expertise, and green human resource management practices. A Sankey diagram visualises the primary linkages between technology and strategy. Then, the authors developed a framework (TSC framework) that links skill clusters, CE practices, and enabling technologies and validated it through an illustrative case study. Interpreting the findings through the Resource-Based View, the paper argues that value arises from socio-technical bundles that integrate technologies, circular practices, and human capabilities. The study concludes with implications for policymakers, educators, and practitioners and outlines potential avenues for future research on skills for human-centred circularity.
Determining appropriate recovery pathways for end-of-life (EoL) products is challenging due to limited information on their highly variable physical conditions and residual values. Existing recovery systems are highly dependent on static procedures, which hinder adaptive pathway selection and limit profitability. To address this limitation, this study proposes a multi-stage triage model inspired by medical triage protocols. This model formulates EoL recovery as a sequential decision process in three hierarchical stages: preliminary, product-level, and component-level triage. To drive this adaptive decision-making, the model uses a digital twin (DT) during the preliminary and product-level stages for structured state mapping and dynamic information updating. At the component-level stage, a knowledge graph (KG) is used to represent structural relationships and precedence constraints, supporting structurally feasible disassembly planning. The model incorporates progressively acquired evidence across all stages using fuzzy Bayesian updating. To evaluate the profitability of continued disassembly, the model embeds a partially observable Markov decision process (POMDP) with Bellman recursion, enabling sequential decision-making and optimal stopping under uncertainty. A gearbox case study involving four condition–uncertainty scenarios demonstrates that this dynamic, evidence-driven model adaptively updates recovery pathways and improves the resulting net recovery benefit.
Incremental sheet forming (ISF) enables flexible manufacturing of complex components, but localized thinning and ductile fracture still restrict its industrial application. Physical trial-and-error tests and finite element (FE) simulations are also time-consuming for repeated process refinement. To address this issue, this study proposes a Physics-Guided Digital Twin (PG-DT) framework incorporating a Physics-Guided Analytical Fracture Predictor (PG-AFP) for pre-machining fracture-risk evaluation. The analytical model calculates kinematic contact boundaries, extracts local strain increments, evaluates stress-state variables, and accumulates ductile damage using the Modified Mohr-Coulomb (MMC3) criterion, without requiring full-field FE discretization for fracture-risk screening. Experimental validation shows that the relative prediction errors for fracture depth remain below 8.5% under different geometric and kinematic conditions. In the benchmark case, the analytical engine completes global damage mapping in 13.35 s, corresponding to a 997-fold reduction in computation time compared with the explicit FE model. The rapid damage evaluation enables unsafe candidate strategies to be screened and refined before machine deployment. The parametric analysis explains how key process parameters affect the evaluated stress state and damage evolution. Material-specific formability evaluations are conducted for AA6061-T6 and Ti-6Al-4V. Finally, the practical viability of the PG-DT workflow is demonstrated through the fracture-free manufacturing of a twisted radial-groove freeform component and a customized cranial prosthesis.
In hybrid configure-to-order and engineer-to-order (CTO–ETO) manufacturing, Bill of Materials (BOM) configuration for variant-rich products must translate structured order parameters into accurate and manufacturable component sets while accounting for legacy codes, semi-structured records, and engineering constraints. In industrial control-valve configuration, rule-based systems can provide traceable constraints but may have limited coverage for low-frequency or evolving component combinations, whereas purely data-driven predictors may generate BOMs that violate compatibility, dependency, or exclusion relations. This paper presents a deployment-oriented BOM configuration pipeline for consolidated hybrid CTO–ETO control-valve orders. The pipeline constructs two complementary order views: a semantic view used by a teacher model to learn cross-parameter and parameter–component cues, and a structured view used by a lightweight student model for MES/PLM-compatible inference. The student-generated BOM is then examined by a KG-structured auditing module that organizes mined positive and negative parameter–component evidence and applies confidence- and consensus-gated corrections instead of enforcing all triggered rules. Experiments on a real-world industrial dataset containing 938 orders and 3247 component labels show that the proposed pipeline achieves a micro-F1 of 0.9303 and a Hamming loss of 0.00202. The KG-structured auditing stage reduces the combinatorial violation rate from 44.68% to 35.11% and decreases the average conflict intensity from 20.30 to 11.49. These results indicate that, within the considered consolidated hybrid CTO–ETO control-valve setting, the proposed pipeline improves BOM prediction while providing auditable manufacturability-oriented correction.
Accurate and reliable tool wear monitoring (TWM) is essential for ensuring machining quality, production efficiency, and operational safety. However, data-driven methods generally lack physical interpretability, while physics-based methods often rely on simplified assumptions and fixed empirical equations. To address these limitations, this study proposes a physics-informed gated Mamba (PIGM) framework with adaptive feature reconstruction. First, a training-set-guided adaptive feature reconstruction mechanism reconstructs multi-domain sensor features through correlation-based and temperature-controlled weight allocation. A gating mechanism is then incorporated into the Mamba backbone to regulate information flow and enhance wear-related feature modeling. In addition, a physics-consistency surrogate branch (PCSB) introduces a learnable wear-rate consistency constraint, while an uncertainty-based adaptive weighting strategy balances monitoring accuracy and physical consistency during training. The proposed framework was evaluated on the PHM2010, carbon-fiber-reinforced polymer (CFRP), and titanium alloy (TiA) milling datasets. Compared with PIMamba, PIGM achieved average percentage reductions of 57.91% and 55.78% in RMSE and MAE, respectively, on PHM2010. Comparisons with recent time-series regression models further demonstrated improved monitoring accuracy across representative test subsets. Ablation experiments confirmed the contribution of each designed module. Additional analyses of noise interference and single-sensor failure showed that PIGM maintains stable monitoring under moderate signal disturbance and some redundant sensor failures, while remaining sensitive to process-specific key sensors. Shapley additive explanations (SHAP) and physics-consistency analyses further provided feature-level interpretation and evidence of improved physical plausibility. These findings indicate that PIGM provides an accurate and physically plausible framework for TWM across the evaluated milling cases without requiring an explicit empirical wear equation.
With the rapid adoption of new energy vehicles, the demand for disassembly of end-of-life battery continues to increase in maintenance, echelon utilization, and recycling. Under anomalous conditions, end-of-life battery disassembly heavily relies on the operational experience of skilled workers. When such experience is used for novice workers or machine adaptation, the extraction and structured representation of such knowledge remain difficult because the relevant knowledge is mainly held by individual workers. Therefore, the effective extraction and representation of operational experience provide an important basis for supporting human-AI collaborative evolution. To address the difficulties in extracting operational experience, representing fine-grained operational semantics, and organizing anomaly-related knowledge under anomalous operating conditions, this paper proposes a method for extracting and representing disassembly operation knowledge of end-of-life battery under anomalous operating conditions. Based on first-person disassembly videos, the proposed method performs temporal modeling and visual perception of disassembly actions and obtains action-tool-component triadic structures through semantic reasoning. Meanwhile, using the standard-condition disassembly sequence as the backbone, the method supplements anomaly types and anomalous operation branches to construct an operation knowledge graph for end-of-life battery disassembly under anomalous operating conditions, enabling the unified representation of standard procedures and anomalous operation knowledge. Experimental results show that the proposed method extracts key operational elements in anomalous disassembly scenarios more accurately. Specifically, Abnormal Operation Node Accuracy improves by about 8.0% compared with the conventional method. This improvement enhances the ability of the knowledge graph to represent anomalous operating conditions and supports knowledge accumulation, experience reuse, and human-AI collaborative evolution in end-of-life battery disassembly.
The present study formulates an integrated dynamic control policy that jointly optimizes production, quality sampling, and maintenance in an uncertain Hybrid Manufacturing-Remanufacturing (HMR) system facing progressive aging. We study remanufacturing strategies since they anchor modern circular economies, and so addressing their specific dynamics is essential to ensure sustainable economic viability. Specifically, we study how progressive wear in the remanufacturing unit alters defect rates and failure risk, necessitating an age-dependent switching rule to decide when to favor manufacturing over remanufacturing and vice versa. To address this, the core contribution lies in developing a novel age-dependent production switching priority rule within the HMR structure, dynamically alternating between production sources to mitigate the impacts of equipment deterioration on product quality. Through extensive numerical analyses, we found that the proposed integrated model performs substantially better than traditional decoupled models. By actively linking the dynamic production switching rule with adaptive sampling and maintenance, the system becomes much more cost-effective. In fact, the results show total cost reductions up to 33% compared to decoupled strategies. Ultimately, these savings highlight just how important an integrated control policy is for keeping circular manufacturing operations economically viable over time.
Heavy-duty gas turbines are core equipment in modern energy systems, and their assembly accuracy directly affects aerodynamic performance and operational safety. However, the assembly of large semi-cylindrical casings faces significant challenges, including restricted measurement fields, visual blind spots, pose ambiguity caused by non-closed rotational structures, and low efficiency in conventional manual adjustment processes. To address these problems, this paper proposes a digital intelligent assembly framework integrating distributed measurement, multi-casing pose solving, and quasi-static cooperative adjustment. First, a distributed large-scale measurement network is constructed based on feature-section discretization sampling to improve the coverage and reliability of key geometric feature acquisition. Second, a robust multi-casing pose estimation method is developed by combining SVD/PCA-based geometric feature extraction, Rodrigues rotation-based axis alignment, and an opening direction consistency detection module, thereby reducing pose ambiguity in semi-cylindrical casing alignment. To improve robustness against local measurement disturbances and micro-deformation effects, a residual-bounded weighted correction strategy is further introduced for sectional circle-center fitting and central axis extraction. Third, Automated Guided Vehicles (AGVs) and four-support vertical hydraulic cylinders are employed for cooperative posture adjustment, where adjustment displacements are calculated using small-angle linearization and the Jacobian matrix, and constraint-based safety verification is incorporated to ensure operational safety. Experimental validation on ten heavy-duty gas turbine units demonstrates the effectiveness of the proposed system. Compared with conventional methods, the average adjustment cycle is reduced from 36.3 h to 17.8 h, and the first-pass yield reaches 80%. Additional multi-index posture error verification, parameter sensitivity analysis, and software-level time statistics further demonstrate the accuracy, robustness, and operating efficiency of the proposed quasi-static assembly framework. This study provides a feasible system architecture for the digital and automated assembly of large-scale high-end equipment under complex geometric and industrial constraints.
To address the challenges encountered in the automated assembly of aeroengine blades, where non-conformal insertion, irregular boundary features, and insufficient pose alignment accuracy occur between the blade root and disk slot, this paper develops a force-vision-guided robotic assembly platform. Based on the 3D point cloud measured by structured light, a pose estimation method integrating geometric priors (GP) and reinforcement learning (RL) strategies is proposed. Firstly, the key tasks and coordinate transformations in the automated blade assembly process are systematically defined, the geometric constraint mechanisms during assembly are analyzed, and a mathematical chain linking point cloud input to robotic motion control is constructed for pose estimation. Subsequently, during the coarse registration stage, three types of geometric priors—planar constraint, coaxial constraint, and boundary contour constraint—are incorporated to improve the accuracy and robustness of initial pose alignment. In the fine registration stage, the pose refinement of point clouds is formulated as a Markov Decision Process (MDP), and a hybrid sparse–dense reward mechanism is designed. The Proximal Policy Optimization (PPO) algorithm is then employed to efficiently optimize the registration policy. Finally, assembly validation experiments were performed on both dovetail and fir-tree blades using the CoppeliaSim simulation platform and a self-developed structured- light measurement system. The results indicate that the proposed method achieves superior performance over existing registration algorithms in pose estimation accuracy, alignment stability, and assembly feasibility. This study provides an effective and generalizable solution for the intelligent assembly of complex structural components.
The quality of process planning, linking product design and manufacturing, critically depends on selecting suitable manufacturing processes. In recent years, several deep learning approaches have been proposed for selecting manufacturing processes, frequently relying on synthetically created datasets. While such datasets allow models to capture general manufacturing patterns, they fail to reflect the complexity and variability of industrial 3D CAD models. As a result, the effectiveness of deep learning for predicting manufacturing processes is constrained by the limited availability of high-quality, annotated industrial data. To mitigate the effort for annotation of industrial data, this paper investigates deep transfer learning for the selection of manufacturing processes. Models are first pre-trained on a synthetically designed dataset and subsequently fine-tuned on smaller sets of real industrial data. The findings demonstrate that deep transfer learning enables improved performance in the selection of manufacturing processes.
The transition to Circular Manufacturing (CM) demands interoperable coordination of heterogeneous Product, Process, and Resource (PPR) data across organizational boundaries and shopfloor systems. Existing semantic technologies for interoperable knowledge representation, such as ontologies and the Asset Administration Shell (AAS), remain passive data carriers decoupled from data integration and operation, while shopfloor orchestration approaches lack grounding in interoperable data standards, leaving a critical gap between structured lifecycle knowledge and closed-loop manufacturing decisions. To resolve this isolation, this paper proposes the Agentic Active AAS (A4S), a neural-symbolic architecture that unifies representation, integration, and orchestration within a single framework. A4S makes three core contributions. First, A4S extends the AAS standard (IEC 63278) from a passive interface to a proactive interoperability layer by leveraging Large Language Models (LLM), which implement the well-known concept of “Active AAS” or “Type3 AAS”. A graph-like recursive encoding of AAS instances is proposed to preserve the strict parent-child topology of AAS shells, submodels, and submodel elements as a symbolic graph traversable by LLM retrieval. Building on this representation, a suite of AAS tools is formalized as the LLM-agent’s bounded action space, enabling the agent to dynamically construct, query, and manipulate AAS instances. Second, A4S introduces Agentic Operation Delegation, which abstracts heterogeneous shopfloor operation interfaces into unified AAS operation schemas and mathematically bounds the LLM’s token generation through schema-constrained decoding. Third, building upon the Active AAS layer, an LLM-based Multi-Agent System (LMAS) is developed, in which the Orchestrator Agent and specialized sub-agents interact through the unified AAS substrate. This simultaneously enables automatic creation of Digital Product Passports (DPP) for cross-organizational value chain exchange and intra-enterprise PPR integration for CM operations. The proposed approach is evaluated in a battery remanufacturing scenario with a dynamic system environment. Experimental results show that A4S achieves over 95% success rate in complex multi-step shopfloor orchestration tasks, consistently outperforming state-of-the-art LMAS approaches across all model scales. Notably, open-source models (e.g., Qwen3.6–35B) under A4S surpass large proprietary models such as GPT-5.2 in orchestration tasks, demonstrating the effectiveness of the proposed neural-symbolic constraints. For DPP generation, A4S-GPT5.2 achieves 90% success rate, where tool execution correctness reaches 100%. These results highlight that A4S effectively enables more robust orchestration in the constrained CM environment. The code is released at https://github.com/quickhdsdc/aas-mcp-server.git.
Autonomous factory inspection increasingly relies on spatially distributed sensing systems to provide consistent and repeatable monitoring in industrial environments. However, inspection outcomes are strongly influenced by the combined effects of carrier platforms, localization accuracy, and sensing characteristics, which are often evaluated in isolation. This paper presents a comparative assessment of carrier platforms and localization systems for autonomous factory inspection, supported by a unified system architecture that integrates heterogeneous carriers, localization methods, multi-modal sensing, and factory layout references. A representative implementation and factory case study evaluate SLAM-based and RTLS-based inspection using mobile and fixed carriers for ambient sensing of noise, illumination, temperature, humidity, and CO₂ concentration. The results show that higher localization accuracy improves the spatial consistency of ambient maps, while different ambient variables exhibit different responses to mobile and static sensing, as sensing behavior varies with carrier mobility and localization uncertainty.
