Logistics serves as a critical operational backbone in offsite construction, serving as the physical conduit for prefabricated component transportation that bridges manufacturing plants and assembly sites. The current offsite construction logistics (OSCL) research overemphasizes building construction phases, neglecting multidimensional perspectives that limit logistics. This review paper investigated OSCL research and practices (2015-2024) through a lifecycle lens via academic publications, industry standards, and governmental policies. Building Information Modeling (BIM), as a lifecycle management tool integrating design, manufacturing, and construction processes, demonstrates transformative potential for OSCL when coupled with emerging technologies. The research encompassed three main steps: Firstly, the retrieved bibliographies on OSCL and smart BIM were categorized for thematic analysis. The research areas were divided into four subthemes: transportation, logistics dynamics, performance, and collaboration. Then, global OSCL practices were investigated, including laws, standards, academic programs, and industry innovations in representative countries from North America, Asia Pacific, and Europe. Two major global challenges were identified: (1) divergent specifications in transportation infrastructure, vehicle fleets, and prefabricated components, and (2) technological integration complexities for sophisticated and sustainable management. In response, this study proposed a smart BIM-enabled OSCL framework integrating digital twins and intelligent technologies to adapt to external forces. The paper contributes to bridging research gaps in OSCL through multidimensional investigations and establishing a roadmap for smart BIM-enabled OSCL in sustainable construction.

Surface defect detection is indispensable for managing product quality in industrial manufacturing. However, owing to the various defect morphologies, it remains difficult to improve accuracy in cross-domain surface defect detection. This paper proposes a defect-aware unequal network (DAU-Net) for detecting industrial surface defects with multi-morphology. In DAU-Net, the defect-aware global representation module is developed to overcome the limitation of local convolution for modeling large-scale surface defects. The defect-aware interactive computation strategy is designed to ensure complete aggregation of surface defect by perceive pixels of patches highly-correlated to surface defect. Moreover, an unequal loss is introduced to weight the regression error of important estimated keypoints, strengthening the model’s focus on learning highly-related bounding boxes of surface defects during training. The proposed DAU-Net is comprehensive evaluated on three cross-domain surface defect datasets, achieving better detection accuracy with multi-morphology surface defects compared with four state-of-the-art methods.

In engineering applications such as service robotics, human-robot collaboration, and industrial automation, robotic manipulators frequently operate in highly dynamic and partially observable environments. The presence of dynamic obstacles, unpredictable human behaviors, and rapidly changing task demands intensify the inherent conflicts between positioning accuracy, obstacle avoidance, and operational safety. At the same time, these systems must rely on incomplete and noisy sensor data to perceive and interpret their surroundings. Such dynamic scenarios significantly increase system complexity and require adaptive, real-time control strategies capable of making reliable decisions under uncertainty. To address these challenges, this paper proposes Trend Learning-Adaptive Reward Shaping-Temporal Difference Knowledge Distillation of Q-value(Action-Value Function), collectively referred to as TL-ARS-TDKDQ, a reinforcement learning framework designed to enable robotic manipulators to adaptively perform precise positioning and dynamic obstacle avoidance in dynamic environments. Trend Learning(TL) alleviates environmental uncertainty by extracting temporal dependencies from sequential data of the manipulator. Adaptive Reward Shaping(ARS) dynamically balances positioning accuracy and obstacle avoidance for the robotic manipulator while progressively increasing task difficulty via curriculum learning. To enhance stability during reward fluctuations caused by ARS,Temporal Difference Knowledge Distillation Q-value (TDKDQ) employs a dynamic teacher network and Temporal Difference(TD) error-based balancing, ensuring stable policy convergence in non-stationary scenarios involving robotic manipulator control. Experiments with a KUKA arm in CoppeliaSim demonstrate that TL-ARS-TDKDQ significantly improves convergence speed, control stability, and task success when integrated into mainstream continuous control reinforcement learning algorithms.

Lithography area scheduling is challenging due to its dynamic, real-time, multi-constraint and multi-objective characteristics. To address this issue, this paper proposes a lithography area scheduling method integrating Transformer with double deep Q-network (ITDDQN). A double deep Q-network model is constructed to make multi-objective real-time decisions on dynamic events in the lithography area. A feature enhancement module is proposed to improve the double deep Q-network's global optimisation capability. The unconstrained rules are designed to enhance the practicability of the scheduling approach. Experiments demonstrate that ITDDQN outperforms traditional reinforcement learning methods and heuristic rules in terms of makespan and wafer on-time delivery ratio.

Accurate prediction of cycle time (CT) is crucial for ensuring the timely delivery of wafer products, considering the inherent uncertainties in wafer fabrication systems. However, the concept drift problem arises when types of wafer products change due to the dynamic introduction of new orders, causing variations in the joint probability distribution of data. This drift challenges the performance of existing CT prediction models based on batch learning, potentially leading to degraded or failed predictions. To tackle this challenge, we propose a continual deep learning prediction method that continuously learns knowledge regarding new types of wafer products by memorizing and updating the wafer data stream. Specifically, we employ model-based concept drift detection on the incoming data flows and devise a dynamic sampling strategy to balance the sample distribution between minority and majority classes, thereby addressing the class imbalance problem. Within the continual learning framework, our proposed FD-MIR (Feature Distillation with Maximally Interfered Retrieval) method integrates a teacher-student feature distillation architecture and a Clustering-driven Adaptive Threshold Maximally Interfered Retrieval (CA-MIR) strategy with adaptive K selection: when concept drift is detected, the pre-drift model is frozen as the teacher to guide feature learning of the student model, while the CA-MIR dynamically determines the number of critical historical samples (K) for replay through clustering analysis, by clustering interference values into multiple clusters and selecting samples based on a cumulative interference ratio threshold. Experimental results demonstrate the superior performance of our method in accurately predicting CT under scenarios where the types of wafer products change. The proposed method holds great potential for enhancing the efficiency and effectiveness of CT prediction in wafer manufacturing systems.

To address the challenges of control latency and uncertainty in force control tasks involving soft-bodied operational robots, this paper proposes a physics-guided and uncertainty-aware feedforward force-admittance control method, referred to as PG-DGPRL. First, a feedforward admittance control strategy with online estimation of force-tracking error is introduced. By forecasting future deviations in interaction forces, the controller parameters are adjusted in advance, thereby mitigating the adverse effects of control latency. Second, a physics-informed deep Gaussian process reinforcement learning (DGPRL) method is developed. The policy is represented as a Gaussian distribution, and the constitutive mechanical model of flexible materials is embedded as a physical constraint, enabling both uncertainty awareness and physical consistency verification. In addition, a composite loss function is designed to achieve joint optimization among policy generation by the actor network, uncertainty estimation by the DGP, and physical regularization by a physics-informed neural network (PINN). Finally, experiments are conducted under various environmental damping and interaction force conditions, comparing PG-DGPRL with baseline methods including CAC, Ada-CAC, and A2C. The results indicate that, under the current experimental conditions, PG-DGPRL achieves superior trajectory tracking and force control performance, exhibiting strong stability and generalization capability.

Wafer defect detection is a crucial part of semiconductor manufacturing, requiring both high generalization across distribution shifts and interpretable explanations for diverse engineering roles. Existing deep learning methods often fall short in semantic reasoning and domain transferability. To address these challenges, we propose WaferDSR-RAG, a novel retrieval-augmented framework that integrates a bimodal wafer defect knowledge graph (BWDKG) with a three-stage semantic adaptation strategy, comprising defect visual-semantic alignment, defect-relevant knowledge retrieval and screening, and role-adaptive explanation generation. This design allows the system to dynamically adapt to unseen defect distributions and generate expert-aligned explanations tailored to different engineering responsibilities. Extensive experiments on both in-distribution and cross-fab wafer maps demonstrate that WaferDSR-RAG consistently outperforms state-of-the-art baselines in both detection accuracy and explanation quality. Compared to GPT-4o and Gemini 1.5 Pro, our method generates more semantically accurate and practically useful explanations for engineers with different roles, as validated by automatic metrics and domain expert evaluations. WaferDSR-RAG represents a scalable solution for wafer defect detection and reasoning in real-world semiconductor production.

Wafer defect recognition is crucial for semiconductor manufacturing. During wafer fabrication, mixed-type defects with various morphologies, along with random noise, introduce data uncertainty into the wafer defect recognition process. However, most existing recognition methods fail to account for this uncertainty, leading to potentially unreliable results. In this paper, we propose a trustworthy mixed-type wafer defect recognition method (TMWDM) based on evidence theory, which models and leverages the evidential support of individual features during training to enhance recognition robustness. TMWDM consists of two main components: a directional evidential feature discrimination network and an evidential feature differentiated learning strategy. The former converts extracted features into evidence representations, enabling the evaluation of the directional alignment and strength of each feature's evidential support under varying input and noise conditions. The latter employs an uncertainty-sensitive loss function that incorporates a penalty term measured using evidential features to optimise model learning under data uncertainty. Extensive experiments demonstrate that TMWDM consistently achieves over 98.80% in different recognition metrics across all mixed-type wafer defect scenarios, outperforming state-of-the-art methods. It also generalises well to WM-811 K and shows significant gains in ablation study.

Abstract Feature extraction from high-speed winding machines is crucial in diagnosing equipment failures in chemical fiber production. However, current methods often face challenges in prediction accuracy and adaptability, especially when accounting for varying degradation modes and operational conditions. This article proposes a data-knowledge fusion evolutionary generative programming method for feature representation of winders. First, the operational data of winders is preprocessed. Subsequently, the adaptability of these features is improved through knowledge-informed feature extraction and correlation analysis. Finally, an evolutionary generation method integrates these enhancements with data-knowledge fusion, resulting in the gene expression programming (GEP) with correlation analysis fault feature set, which facilitates adaptive remaining useful life prediction for winders. The proposed method is validated using both the PHM2012 challenge bearing dataset and real-world data from high-speed winding machines. Experimental results demonstrate that the proposed method significantly improves fault prediction accuracy and stability compared to traditional techniques and standard GEP approaches, demonstrating its practical value for industrial applications.

Most research on 3D object shape recognition is unimodal and lacks interpretability. To address issues of missing point cloud and poor interpretability of visual perception when acquiring the object shape information under occlusion, a novel multi-modal model based on the fusion of visual and tactile point cloud information for 3D shape recognition and interpretability is proposed. First, the experimental acquisition platform for visual and tactile point clouds is constructed, which facilitates the collection of visual and tactile point clouds of objects under self-occlusion conditions. Second, a shape recognition model based on the fusion of multiple attention mechanisms for visual and tactile point clouds has been established, which is used to extract features from the preprocessed visual and tactile point clouds. The instance and class accuracies of the Dual-VT-Multi-attention model on the self-built dataset are 80.32% and 83.32%, respectively, which are significantly higher than single visual or tactile modal. Finally, to provide an intuitive interpretation of the classification decision process of the Dual-VT-Multi-attention model in each modal, an interpretable method based on the recognition model of visual and tactile point clouds is proposed. The contribution of each point can be calculated by weighting the summation of its feature vectors, which allows the generation of the Class Attention Response Map to visualize the points that are important for the model’s classification decision. The Class Attention Response Map makes the shape recognition result of Dual-VT-Multi-attention model in each modal more transparent and interpretable.

Fabric defect detection is essential for ensuring the quality of textiles, particularly when addressing lustre issues in silk materials and tiny, inconspicuous flaws. This paper introduces a saliency detection framework for silk surface defects, named EFRNet, which aims to overcome the limitations of existing technologies in handling such fabric imperfections. Firstly, a multilayered spatial domain edge perception method was designed. This method enhances the perception of defect edge details through edge‐aware units and integrates these detailed features with high‐level semantic information using cross‐layer connection strategies, thereby achieving precise detection of defect edges. Secondly, a context‐aware multilayer feature fusion technique is proposed, which includes attention‐induced feature aggregation units and global context differential units, aimed at optimising feature selection, reducing background interference and emphasising the salient features of fabric defects through differential analysis. Experiments have demonstrated that EFRNet excels in improving the accuracy of small defect detection and local detail recognition. Compared with existing methods, it significantly enhances the quality of saliency maps, proving its effectiveness in enhancing the adaptability and precision of fabric defect detection.

Semiconductor wafer fabrication, as a large-scale and complex discrete manufacturing system, presents significant challenges in shop floor scheduling due to its scale, uncertainty, and re-entrant processing. Additionally, effectively leveraging historical scheduling decision data remains a challenge, limiting the ability of scheduling algorithms to accurately assess the current system state. To address these issues, this paper proposes a reinforcement learning-based optimisation method guided by decision graphs. First, we introduce a multidimensional heterogeneous disambiguation graph to comprehensively represent the operational state of the wafer manufacturing system. Second, we design a graph neural network to characterise the multidimensional disambiguation graph and learn from historical decision-making experiences. Finally, we propose a decision graph-guided action strategy that optimises the reinforcement learning policy, improving training efficiency and the accuracy of action selection. Experimental results demonstrate that our method achieves superior generalisation performance and outperforms traditional approaches. This study provides an effective solution for optimising scheduling in semiconductor wafer manufacturing systems.

Defect detection in striped fabrics has always been an important task for fabric production companies. Due to the interference caused by different background stripe color combinations and minimal period variations, visual small-sized detection of defects on fabrics with varying background textures and stripes becomes a challenging engineering problem. To address the issue of missing small-sized targets in striped fabric defect detection tasks, this study investigates a multi-scale detection method for striped fabric defects enhanced by coordinate attention. A multi-scale detection module with small-scale detectors was designed to enabling the detection model to identify small-sized defect features. Additionally, a coordinate attention-based defect feature extraction module was developed, which integrates coordinate information in both vertical and horizontal directions of the feature matrix to learn the spatial coordinate relationships of elongated defects, rigorous experiments have shown that using this method can improve precision by 0.47

Rapid and accurate detection of surface defects has become critical with the increasing demand for highly reliable carbon‑fiber composite plates (CFCPs) in advanced manufacturing. This study proposes a dual-stage enhancement framework to detect subtle defects in CFCPs, addressing the limitations of conventional down‑sampling and feature‑extraction methods, especially when samples are limited and defects are subtle. The framework highlights the critical role of expert knowledge in defect detection and allows effective parameter transfer between a task-specific super-resolution reconstruction module and a residual, multiscale fusion semantic segmentation network. Experiments on a digital‑radiography CFCPs data set demonstrate that this method markedly amplifies weak defect signatures and pinpoints their locations with high fidelity. The findings exhibit significant gains in precision, recall, F1‑score, and mean intersection‑over‑Union relative to U‑Net, SegNet, and other baselines. In small‑sample conditions, the proposed model nearly doubles the performance of the canonical U‑Net. This framework offers a broadly applicable solution for automated micro‑scale defect inspection across composite‑material systems and other advanced‑manufacturing contexts.

In the chemical fibre production process, the winding stage has a significant impact on the final fibre quality. Ensuring accurate and reliable fault diagnosis of the winder is crucial for maintaining production efficiency. However, current fault diagnosis of winder chucks faces the challenge of a lack of effectively labelled fault data. Contrastive learning, as an effective solution, has been gradually applied. However, existing contrastive learning frameworks have not been specifically designed for vibration signals of rotating equipment, where spindle vibration signals exhibit high similarity between samples and strong background noise. To address these challenges, this paper first proposes a multi-layer wavelet attention convolutional signal adaptive sample augmentation method to generate diverse positive sample pairs, in order to accommodate the actual working conditions of vibration signals. Secondly, a progressive pseudo-negative sample correction contrastive learning method is proposed to resolve the sampling bias caused by incorrectly treating samples with the same actual label as negative samples. Experimental results demonstrate that the proposed method outperforms existing methods in fault diagnosis accuracy. Further diagnostic experiments on vibration signals from a winder in a chemical fibre plant validate the strong adaptability of the proposed method in real-world operational scenarios.

Shape prediction of deformable linear objects (DLO) plays critical roles in robotics, medical devices, aerospace, and manufacturing, especially in manipulating objects such as cables, wires, and fibers. Due to the inherent flexibility of DLO and their complex deformation behaviors, such as bending and torsion, it is challenging to predict their dynamic characteristics accurately. Although the traditional physical modeling method can simulate the complex deformation behavior of DLO, the calculation cost is high and it is difficult to meet the demand of real-time prediction. In addition, the scarcity of data resources also limits the prediction accuracy of existing models. To solve these problems, a method of fiber shape prediction based on a physical information graph neural network (PIGNN) is proposed in this paper. This method cleverly combines the powerful expressive power of graph neural networks with the strict constraints of physical laws. Specifically, we learn the initial deformation model of the fiber through graph neural networks (GNN) to provide a good initial estimate for the model, which helps alleviate the problem of data resource scarcity. During the training process, we incorporate the physical prior knowledge of the dynamic deformation of the fiber optics into the loss function as a constraint, which is then fed back to the network model. This ensures that the shape of the fiber optics gradually approaches the true target shape, effectively solving the complex nonlinear behavior prediction problem of deformable linear objects. Experimental results demonstrate that, compared to traditional methods, the proposed method significantly reduces execution time and prediction error when handling the complex deformations of deformable fibers. This showcases its potential application value and superiority in fiber manipulation.

Human-centric assembly is emerging as a promising paradigm for achieving mass personalization in the context of Industry 5.0, as it fully capitalizes on the advantages of human flexibility with robot assistance. However, in small-batch and highly customized assembly tasks, frequently changes in production procedures pose significant cognition challenges. To address this, leveraging computer vision technology to enhance human cognition becomes a feasible solution. Therefore, this review aims to explore the cognitive characteristics of human beings and classify existing computer vision technologies in a manner that discusses the future development of cognition-augmented human-centric assembly. The concept of cognition-augmented assembly is first proposed based on the brain's functional structure - the frontal, parietal, temporal, and occipital lobes. Corresponding to these brain regions, cognitive issues in spatiality, memory, knowledge, and decision-making are summarized. Recent studies conducted between 2014 and 2023 on visual computation of assembly are categorized into four groups: position registration, multi-layer recognition, contextual perception, and mixed-reality fusion, all aimed at addressing these cognitive challenges. The applications and limitations of current computer vision technology are discussed. Furthermore, considering the rapidly evolving technologies such as the metaverse, cloud services, large language models, and brain-computer interfaces, future trends on computer vision are prospected to augment human cognition corresponding to the cognitive issues.

In response to the issues of incomplete local data degradation characterization and multi-scale, non-stationary characteristics exhibited during the degradation process of rotary machinery vibration signals, a remaining useful life (RUL) prediction method based on trend encoding and multi-scale spatio-temporal feature fusion is proposed. Firstly, a trend encoding method is introduced to compensate for the missing temporal information and long-term degradation information in vibration signal samples, enhancing the degradation characterization ability of local data. Subsequently, a soft threshold self-attention mechanism is proposed for the adaptive fusion of trend-encoded features and vibration signal features, to prevent imbalanced weight distribution. Finally, a spatio-temporal feature fusion network, MACNN-Informer, is designed. It possesses multi-scale spatial feature extraction capabilities and can effectively capture long-distance dependencies in sequential features, thereby better revealing the degradation characteristics of vibration signals at different degradation stages. Experimental results show that, compared with methods such as MSCNN, the proposed method–despite being slightly slower in inference speed–achieves the lowest prediction error and oscillation amplitude, making it well-suited for RUL prediction of critical rotating machinery components such as rolling bearings.

The manufacturing process of the aircraft flat-tail assembly is complex and discrete. It typically involves manual assembly at fixed stations with variable shift teams. However, uncertainties can arise even after a scheduling scheme is created, leading to non-optimal or even infeasible schedules. To address this issue, a new scheduling strategy called ‘inverse scheduling’ has been proposed by incorporating the concept of inverse optimization. Notably, this is the first application of inverse scheduling in the complex manufacturing process of aircraft flat-tail assembly. This paper presents a multi-objective optimization model for the inverse scheduling problem of flat-tail assembly production. The scheduling objectives include minimizing the maximum delay penalty cost and minimizing the assembly time adjustment cost. To address the limitations of traditional mathematical planning methods in terms of efficiency and solution quality, an improved genetic algorithm is proposed. This algorithm combines the genetic algorithm with a local search strategy to solve the large-scale inverse scheduling problem. Additionally, an inverse scheduling strategy based on the self-adaptive tolerance-driving mechanism is designed to enhance the algorithm’s efficiency and effectively handle order delay exception events. The effectiveness of the self-adaptive tolerance driving mechanism and the inverse scheduling method is verified through case studies in enterprises. Comparative analysis demonstrates that the proposed method significantly outperforms traditional rescheduling strategies by avoiding high sequence adjustment and material handling costs, offering a more practical and efficient solution for managing disruptions in complex assembly systems.