Abstract To avoid the blind spots and brittleness of imperatively simulating a physical system, modelers often turn to declarative methods that can autonomously execute a model. Such solvers rely on a model structure that allows procedures for transforming inputs to outputs to be automatically discovered. However, the model structures used for traditional declarative solvers are insular, in that they are often isolated to specific modeling domain. Furthermore, these specific model types are unable to be understood by the advanced computing applications required to simulate complex systems. Here we provide a general purpose modeling framework that can integrate software functionalities into the declarative simulation of a model, offering for the first time the ability to define a multi-physics, multi-scale system independent of its eventual simulation. This is accomplished by representing a system as a constraint hypergraph. System models are deconstructed into state variables and relationships in the graph. The APIs of external tools are integrated into the model as inter-variable functions. By encoding these functionalities in the graph, a solver is able to autonomously arrange these relations into executable simulation processes, enabling fully declarative simulation. This is demonstrated by integrating the capabilities of three software platforms together into a single model of a crankshaft from a piston engine: solid geometry (Onshape), structural mechanics (Ansys Mechanical), and kinematic analysis (MATLAB). The result is a holistic modeling framework that allows for flexible simulation of a complex system, integrates directly with otherwise sequestered platforms, and reveals cross-cutting interactions between system elements.
Abstract In aircraft stringer assembly, flatness out-of-tolerance is caused by the cumulative propagation of multisource coupled deviations, and it is difficult to perceive this phenomenon dynamically. To address this problem, this study proposes a dynamic prediction method that integrates assembly deviation mechanisms and machine learning. First, based on screw theory, a full-process deviation propagation model for manipulator positioning–clamping–transplanting is established. This model quantifies the coupling effects of positioning errors, flexible clamping deformations, and inertial disturbances. On this basis, a physics-informed machine learning (PIML) framework is designed. The Lagrangian dynamic equation is embedded into the spatiotemporal graph convolutional network (ST-GCN) as a physical constraint. A dual-stream architecture of Attention-long short-term memory (LSTM) is combined to process multisensor time-series data, and an online Bayesian update mechanism is introduced to adapt to changes in working conditions. Finally, case verification is carried out on the fuselage assembly line of a certain type of civil aircraft. The results show that the root mean square error (RMSE) of the proposed method for stringer flatness deviation prediction is reduced to 0.21 mm, which is about 19% higher in accuracy than the traditional LSTM model. Moreover, this method can reduce the assembly out-of-tolerance rate by about 23%, providing a theoretical tool and technical support for high-precision aircraft assembly.
In the assembly process design, knowledge question-answering is a crucial scenario for promoting the sharing of knowledge resources and enhancing the process design accuracy and efficiency. Simultaneously, knowledge graph technology enables efficient semantic modeling of knowledge, allowing for more accurate capture of user semantics and intentions. This, in turn, enhances the accuracy and flexibility of knowledge question-answering. However, due to the high complexity and specialization of assembly processes, current knowledge graph question-answering for assembly processes still faces challenges, such as difficulty in understanding complex queries. In response to this, this article proposes a large language model (LLM)-enhanced knowledge graph multi-hop reasoning method for assembly process question-answering. This method decomposes the multi-hop knowledge graph question-answering task into three subtasks: LLM-tuning-based question-answering chain generation task, which transforms the question into one or more question-answering chains, multi-hop question-answering chain reasoning task, and LLM-based natural language answer generation task. Among them, a graph path-based multi-hop reasoning model for assembly processes is constructed for question-answering chain generation. This model employs a "core reasoning + attribute constraints" strategy and a task-oriented negative sample setting method to enable rapid and precise reasoning between the knowledge graph and question-answering chains. The effectiveness of the proposed method is validated through the comparative experiments with existing mature knowledge graph question-answering models. In the comparative experiments, three datasets were constructed for LLM fine-tuning and question-answering chain reasoning of multi-hop questions, and the performance index HITS@5 of 0.913 surpassed the existing mature Hete-MF model and could fully meet the process designer needs.
As a critical stage in helicopter manufacturing, the assembly process relies on effective scheduling to ensure both efficiency and quality. Traditional, experience-based scheduling methods are often inadequate for complex shop-floor environments, especially given the heterogeneity of worker skills and complex technological constraints. Therefore, this article proposes a deep reinforcement learning approach based on proximal policy optimization with an attention mechanism (PPO-AM) to solve the helicopter assembly workshop scheduling problem (HASP) with consideration for worker skill proficiency. First, an assembly time prediction model was established that integrates workers' skill levels, task criticality, and the dynamic evolution of proficiency. This model provides a precise foundation for task-time estimation in assembly workshop scheduling. Based on this foundation, a shop scheduling model incorporating multiple process constraints was constructed. Furthermore, the state space, action space, and reward function were dynamically adjusted according to production progress, thereby formulating the problem as a Markov decision process (MDP). Within this framework, a PPO-AM method incorporating a self-attention mechanism was proposed, and an assembly shop scheduling agent was developed based on this method to enable flexible and efficient worker allocation in practical scenarios. The PPO-AM method leverages the self-attention mechanism to assess the importance of state features, enabling the agent to adaptively focus on critical information, thereby enhancing its state awareness and policy generalization capability. Experiments were conducted on assembly shop cases of various scales as well as in real-world engineering scenarios, and the results demonstrated that PPO-AM exhibits superior performance and practical value in complex assembly scheduling tasks.
Managing modern engineering processes is increasingly challenging as they evolve into socio-technical systems where human experts and artificial intelligence (AI) agents coexist. The efficiency of such environments is highly vulnerable to unanticipated stochastic interruptions-such as urgent design changes, field issues, and equipment malfunctions-which share a common operational signature: they are unplanned, time-critical, and induce cascading bottlenecks that increase project makespan, exposing the limitations of conventional static scheduling. This study proposes a generalized human-AI collaborative dynamic reallocation framework to minimize the makespan and enhance the operational flexibility of disruption-prone engineering workflows. The proposed framework addresses dynamic workflow characteristics through two complementary policies: (i) task offloading from overloaded human experts to AI agents, and (ii) resource sharing to lend computational capacity to bottlenecked agents. These two policies are operationalized through an integrated dynamic allocation process that derives resource distribution via a long short-term memory-based workload predictor refined by a reinforcement learning calibration layer. The framework is evaluated via simulation using industrial issue-log data from large-scale software development, where urgent design changes serve as the representative interruption class. The proposed method reduced the makespan by approximately 12% compared to a first-in-first-out, first-available baseline heuristic under uncertain interruptions, and its robustness was verified across 27 diverse workload scenarios. This study presents a new dynamic resource allocation model that strengthens operational resilience by mitigating bottlenecks and proactively utilizing AI resources in engineering collaboration systems.
Modern manufacturing systems demand cognitive digital twins (CDTs) capable of not only mirroring physical processes but also interpreting context and reasoning about system behavior. Traditional digital twins (DTs), while effective in replication, lack these cognitive abilities-limiting their usefulness for trustworthy, transparent, and adaptive decision-making in smart manufacturing environments. This article presents a neuro-symbolic CDT framework that unifies ontology-based modeling with large language models (LLMs) through a retrieval-augmented generation (RAG) architecture to enable explainable fault diagnosis and decision support. The ontology captures domain knowledge on faults, corrective actions, and rule-based logic using the semantic web rule language (SWRL), forming a structured layer for cognitive reasoning. Simulation data from a virtual machine cell are semantically mapped to the ontology to construct a knowledge graph (KG) that contextualizes real-time operational data. Leveraging this graph, the RAG pipeline allows LLMs to retrieve structured insights and generate human-interpretable explanations as well as machine-actionable commands, effectively bridging neural and symbolic reasoning. Demonstrated use cases show that integrating rule-based reasoning with neural generation enhances fault detection, interpretability, and adaptive control, charting a path toward CDTs that are both transparent to users and operationally effective in real time.
Abstract Industry 4.0 and smart manufacturing increasingly require secure, decentralized, and collaborative data-driven systems. However, challenges such as data privacy, cybersecurity, trust among stakeholders, and heterogeneous data and data sources continue to limit the scalability and efficacy of conventional artificial intelligence (AI) and machine learning (ML) solutions. This review investigates the integration of blockchain (BC) and federated learning (FL) as a framework to address these challenges within the manufacturing domain. We aim to explore the existing application scenarios and technical approaches adopted for BC–FL integration in manufacturing, identify technical and organizational challenges, and uncover cross-domain innovations that may be adapted to industrial settings. Key identified applications in manufacturing include cybersecurity, predictive maintenance, supply chain optimization, digital twin (DT) systems, and quality control. Cross-domain insights offer promising strategies to support secure collaboration, improve model performance, and enhance trust in manufacturing environments. This review provides a reference for researchers and practitioners seeking to design secure, scalable, and collaborative AI systems in smart manufacturing environments using FL and BC technology.
Design optimization has important applications in many engineering fields, where the goal is to find the best design within the available means. In many practical applications, finding the best design can bring significant profit, quality, and performance advantages. However, popular design optimization methods, such as gradient-based methods and heuristic methods, often become trapped in local optima and fail to find the global optimum that corresponds to the best design, leading to inconsistent or incorrect assumptions and applications. In this article, a novel global optimization method designed for graphics processing unit (GPU)-based massively parallel computing is introduced to efficiently enclose the global optimum for continuous design optimization problems, where the objective and constraint functions have analytic expressions. Using interval arithmetic, coupled with the computational power of GPU, the method iteratively rules out the regions in the design space where the global optimum cannot exist and leaves a finite set of regions where the global optimum must exist. Because of the rigor of interval arithmetic, the method is guaranteed to enclose the global optimum in the regions within a user-specified width tolerance for design optimization problems, even in the presence of rounding errors. The GPU-based global optimization method is validated through two case studies of the Ackley function and launch vehicle design. The results show that the method successfully encloses the global optimum that corresponds to the best design in each case study, while both a typical gradient-based method and a popular heuristic method become trapped in local optima that correspond to inferior designs.
Metamaterials (MMs) offer a promising solution for footwear design by enabling spatially tunable mechanical responses associated with comfort-related considerations. However, the current application of MMs in footwear manufacturing remains relatively limited. Furthermore, effective optimization of MM structures under comfort-related mechanical considerations remains a largely unexplored space. This study explores a framework that teams finite element analysis (FEA) and deep learning (DL) to optimize MM-based shoe sole designs by using the individual plantar pressure distribution data. Specifically, the FEA is deployed to analyze the mechanical response of the sole, and it establishes links among pressure, deformation, and geometry, while a DL model is developed to map plantar pressure to MM lattice geometry. In our proposed method, a multimodal convolutional neural network (CNN) model is designed and trained to predict the optimized lattice-rod radii in MM based on multimodal input, including the pressure image, rod coordinates, and deformation data. The DL-optimized MM structures are subsequently validated through FEA simulations. With this proposed approach, this study conducted a case study relevant to potential footwear applications under a deformation-based surrogate objective. The results show that the proposed method achieves significant improvement in deformation-pressure alignment compared with a benchmark design. In summary, the proposed FEA-DL framework enables direct generation of optimized MM designs from user-specific plantar pressure and a surrogate design objective. This synergistic approach offers a promising pathway to achieve a customizable, efficient, high-performance, resource-conserving inverse design.
With rapid advances in three-dimensional (3D) metrology, point cloud data are increasingly available for surface quality inspection in additive manufacturing (AM). Compared to images, point clouds capture richer geometric information for characterizing surface anomalies, enabling more comprehensive defect diagnosis and mitigation. However, it remains challenging to extract anomaly-pertinent information from scanned point clouds, due to (1) the scarcity of annotated point cloud data for training robust anomaly detection models and (2) the inherent complexity of point cloud processing, stemming from their high dimensionality, high volume, and unstructured nature. To address the challenges, this study develops a new framework for self-supervised representation learning of point clouds to glean anomaly-pertinent features. Specifically, a graph contrastive learning scheme is constructed by integrating & ell;-hop subgraphs, hard-negative sampling, and graph neural networks (GNNs) to explore the self-similarity of AM-fabricated surface patterns and highlight anomaly-induced variations. Unlike most existing approaches, it requires no external training samples or manual annotations. The framework has been evaluated using simulations and real-world data collected from wire arc additive manufacturing (WAAM). Results demonstrate that it outperforms the state-of-the-art benchmarks in accurately locating and characterizing surface defects, including subtle ones. The developed framework has strong potential for broader applications in differentiating surface textures and geometric patterns across diverse AM processes.
The design and management of complex products require engineers to synthesize vast amounts of information from disparate sources like design manuals, specification sheets, and technical reports. Manually constructing coherent and structured product information models is a primary bottleneck in the product development lifecycle. Product information modeling captures functional requirements, physical component hierarchies, interconnections, and critical parameters. This process is not only time-consuming and labor-intensive but also hinders scalability. To address this critical engineering challenge, this study explores the integration of large language models (LLMs) with expert-guided verification to streamline the product information modeling process. Our approach aims to automatically extract and structure product knowledge from large volumes of unstructured technical documents, providing both formal semantics and visual graphical representations to support domain engineers and IT professionals. We demonstrate the approach through a case study on the Tennessee Eastman Process (TEP), a well-established benchmark in process control and industrial systems research. The results highlight the method's effectiveness in capturing relevant product knowledge and aligning it with design intent. Guided by a defined set of evaluation metrics, further experimental results validate a substantial increase in performance, evidenced by an 82% saving in modeling time and comprehensive information coverage (100% recall). This work paves the way for automated, knowledge-driven product modeling and offers promising advancements in the design and management of complex, large-scale products.
The rapid development of large language models (LLMs) and robotics is fundamentally reshaping the manufacturing sector. However, there are challenges for applying LLMs in robotic control, particularly in understanding the spatial and logical relationships between different objects. To address the problem, this article proposes a novel industrial robot control scheme called the LLM knowledge graph powered intelligent agent framework (LKPI). The goal is to leverage the external knowledge and logical relationships provided by knowledge graph (KG) to enhance robotic perception and decision-making in complex production environments. In experimental validation, LKPI demonstrated superior logical reasoning and environmental adaptability in a dynamic and noise-prone pecan processing case study. Even in the presence of noise, LKPI utilized environmental background information and relationship information stored in the KG to accomplish transportation and navigation. By integrating real-time recognition and depth measurement from an RGB-D camera, the robot continuously adjusts its path in response to changes in the position of moving targets. By leveraging Internet of Things technologies to continuously update the knowledge graph in real time, LLMs can achieve real-time environmental awareness through the use of retrieval-augmented generation. This research demonstrates the reasoning control capability, dynamic perception capability, and real-time adaptation capability of the LKPI framework. Experimental results show that the LKPI framework achieved a 96.67% success rate and successfully realized dynamic tracking of target objects. These results demonstrate LKPI's capability for dynamic information acquisition and robotic control in complex industrial environments, thereby improving the overall task completion rate.
Segment is a key component of a continuous casting machine, which frequently experiences faults. The straightening force is an important indicator for the working state of the continuous casting machine, effectively reflecting the segment fault. To provide data support for the identification of segment faults quickly and accurately, an approach for straightening force prediction based on the model-agnostic meta-learning-Long Short-Term Memory (MAML-LSTM) is proposed. First, the characteristics of the straightening force for the segment are analyzed, and the corresponding straightening force time-series data are preprocessed using wavelet transform adaptive threshold for reducing the interference of external factors. Second, through analysis of influencing factors for straightening force, the straightening force dataset is constructed. After that, to address the issue of a poor sample of segment fault, the MAML-LSTM model for straightening force prediction is built, where the straightening force dataset is employed for meta-training and meta-testing. Finally, the results of the experiment verify the effectiveness and feasibility of the proposed approach, which can effectively improve the accuracy of straightening force prediction. It can provide valuable data support for segment fault identification in continuous casting.
This article introduces a domain-agnostic simulation framework for evaluating the resilience and robustness of network-based systems subjected to degradation and partial recovery. Built upon a previously developed mutation-only genetic algorithm (GA) for network optimization, the framework integrates agent-based modeling to simulate full-lifecycle behavior. Functional, degradation, and recovery agents operate concurrently over time, enabling dynamic representation of system performance loss, cascading effects, and restoration efforts. The simulation evaluates ten optimized network configurations, including in-depth analysis of three representative cases (best overall, balanced, and vulnerable) to explore how structural design influences degradation response. Results demonstrate that GA-derived fitness and topological centrality measures fail to predict long-term resilience outcomes. In contrast, four new behavior-driven metrics introduced in this work (the normalized performance loss per unit degradation, robustness index, sustained functionality index, and structural integrity index) offer interpretable and transferable insights into system survivability. These metrics are shown to be independent of traditional indicators and capture both node-level sensitivity and system-wide adaptation over time. As a result, they provide a more complete foundation for resilience assessment in mission-critical systems. This study completes a three-part research effort aimed at advancing mission assurance by bridging design-time optimization with operational survivability analysis. Future work will explore cross-domain applications of the metrics, integration of resilience objectives into the GA process, and the extension of simulation dynamics to support probabilistic degradation, adaptive pathing, and intelligent repair strategies.
Fault diagnosis is important for the complex equipment, and physics-informed fault diagnosis has become an emerging trend. While physics-informed fault diagnosis is hard to realize, unless several problems are addressed, one of the limitations is searching for the best architecture, which influences the performance greatly. Neural architecture search (NAS) has been a research hotspot. However, limited by computing resources and the deviation of supernet prediction, NAS might miss the best architecture, and impedes the application of NAS in physics-informed fault diagnosis greatly. Thus, this article proposes an evolutionary one-shot NAS method based on single-path cells (SPC-NAS) for physics-informed fault diagnosis. The proposed method develops a new supernet based on single-path cells, to reduce the computing resources and improve the reusability. An improved supernet training method is introduced to reduce the deviation between the one-shot model prediction and the stand-alone model accuracy. Finally, an evolutionary search strategy with constraint is developed to find the best architecture. The experimental results show that the proposed method can automatically find the best architecture for different tasks, which achieved an accuracy of 100% in Case Western Reserve University (CWRU) dataset with only 0.292 M parameters. All the results indicate that the proposed method can address the limitation to search the best architecture and provides a foundation for future integration with physics-informed methods.
The design of datum systems for body-in-white (BIW) assembly is critical but reliant on expert experience, leading to inefficiency and inconsistency. This article presents a knowledge-driven framework to automate this process. The core is a hierarchical topology mapping model (HTMM) that formally represents the spatial and functional relationships among part features within the geometric dimensioning and tolerancing (GD&T) context. This model is implemented via a four-layer architecture centered on key reference points (KRPs), encompassing component, feature, point cloud, and datum levels, integrating manufacturing and structural constraints into a computable graph. A reasoning mechanism using KRP-based analysis and multicriteria evaluation is developed to generalize datum logic across components. Validated in an industrial case study, the framework generated a datum scheme for a body side outer panel that achieved a 4.5% higher composite quality score than an expert baseline. In a generalization test across three different components (A-pillar, B-pillar, sill), it maintained 100% compliance with all core engineering constraints. This work contributes a structured, model-based method that enhances automation, consistency, and reuse in datum design for complex assemblies.
Modern engineering design requires high-fidelity simulations, which can impose an enormous computational burden and slow the speed of design iteration. Data-driven up-sampling methods like physics-informed neural networks (PINNs) help reduce the computational resources required. However, machine learning model capacity and hardware limitations still pose challenges when evaluating large engineering simulations with complex physics dynamics. Recently, methods have been proposed to enforce the principle of locality in physical systems to neural network layers, allowing for concurrent inference on smaller subdomains with improved efficiency and accuracy. Based on such an idea, we extend the theory of domain decomposition to complex three-dimensional geometries using graph neural networks (GNNs). We developed a graph decomposition method to improve the training and inference efficiency of machine learning models. Super-resolution GNNs are then trained on individual subdomains distributed among GPU nodes. This approach significantly reduces computational overhead while maintaining simulation accuracy. We validate the method's performance on two engineering applications: a variable inlet-angle mixing elbow junction and a low-pressure bleed duct from an Airbus A350 aircraft. For the elbow geometry with varying inlet angles (0-60 deg), the framework achieves R-2 values exceeding 0.995 for velocity predictions across all configurations while demonstrating adaptability to different flow geometries and the presence of secondary flows. For the larger-scale Airbus duct system, we achieve 0.9947 in R-2 metric in velocity and 0.9996 in pressure compared with high-fidelity simulations, with a 5.5 & times; computational speedup and reciprocal scaling with GPU count. These results demonstrate that our approach can effectively bridge the gap between computational efficiency and simulation fidelity across different scales of complex engineering design tasks.
Modern manufacturing enterprises face growing challenges due to frequent production disturbances and intensified demand fluctuations. As a resilient manufacturing paradigm in the era of Industry 5.0, the matrix manufacturing system (MMS) can effectively accommodate multivariety and multibatch production requirements. However, traditional layout algorithms lack effective dynamic response mechanisms and have limited capability for real-time optimization of cell configurations under a volatile environment. To address this issue, we propose a dynamic cell layout planning method based on a consensus-enhanced fruit fly optimization algorithm (CE-FOA), in which physical constraints are embedded into the optimization process through feasible layout representation, fitness evaluation, and constraint-guided search. A consensus-driven evolutionary mechanism is incorporated into the conventional FOA to enhance global search efficiency and convergence stability, while a logistics relationship dimensionality reduction strategy is devised to lower computational complexity during optimization. Case study results from an MMS-based optoelectronic-pod (OP) workshop show that CE-FOA outperforms traditional FOA and simulated annealing (SA) in solution quality and convergence rate. These results validate the effectiveness and superior performance of the proposed approach, demonstrate the practical value of embedding manufacturing physical knowledge into dynamic layout optimization, and provide a new solution for dynamic workshop layout planning under the MMS paradigm.
Abstract Data augmentation plays a critical role in improving the performance, generalization, and robustness of deep learning models, particularly when dealing with limited or imbalanced datasets. This study presents an approach to monitoring cutting forces in milling operations by integrating deep learning models with structured data augmentation. Time–frequency analysis methods are employed to extract dynamic signal characteristics, and the resulting representations are converted into image-based formats to enable the use of vision-based learning models. Multiple deep learning architectures are evaluated for classification and monitoring under varying cutting conditions. The results demonstrate that augmentation significantly enhances model robustness and performance. Among the evaluated models, the proposed convolutional neural network achieves the highest validation accuracy of 87.42% under specific augmentation settings, outperforming other architectures while maintaining low loss. These findings highlight the effectiveness of combining signal transformation, image-based representation, and augmentation strategies for reliable cutting-force monitoring.
