For Tunnel Boring Machines (TBM) main bearings, practical fault diagnosis extends beyond coarse fault-type recognition to characterize both fault mechanism and degradation severity for maintenance intervention and tunneling-parameter adjustment decisions. This fine-grained task is further complicated by three structural difficulties: severity assessment is physically meaningful only under a specific fault mechanism, the resulting mechanism–severity classes exhibit severe long-tailed imbalance, and hierarchical inference is susceptible to irreversible cascade errors. Conventional flat classifiers cannot adequately address these structural difficulties, as they treat fault type and severity as independent, parallel labels and offer no mechanism for revising intermediate errors. To address these limitations, this paper proposes a mechanism-conditioned hierarchical reinforcement learning framework, termed the Hierarchical Dueling Double Deep Q-Network (H-D3QN), which reformulates the diagnosis task as a goal-conditioned sequential decision process. A high-level Type Agent identifies the fault mechanism, and a low-level Severity Agent then infers the degradation level within the physically correct context. Three key designs support this framework: hierarchical decision decomposition for mechanism-conditioned severity inference, a frequency-aware reward design with a first-order analytical basis for approximately reducing class-frequency-induced optimization bias, and a mechanism-conditioned out-of-distribution (OOD) verification strategy that rejects physically inconsistent intermediate hypotheses and converts cascade errors into a revisable inference process. Experiments on a long-tailed TBM main bearing dataset show that H-D3QN achieves an overall accuracy of 88.11% while maintaining a balanced diagnostic profile across head, medium, and tail classes.

Semiconductor wafer fabrication systems (SWFSs) operate in large-scale, complex environments characterized by hundreds of machines and multi-step re-entrant processes. In practice, these systems must frequently adapt to dynamic production conditions, particularly due to the continuous introduction of new products and fluctuations in order volumes, which significantly alter product mix and workload distribution. This demand-side variability leads to substantial work-in-progress (WIP) accumulation and persistent idle lots (ILs) awaiting machine setup adjustments. As a result, scheduling becomes increasingly complex, and traditional strategies, which are often optimized for static settings, may degrade over time, resulting in suboptimal resource utilization and poor delivery performance. To address these limitations, we propose an adaptive reinforcement learning (RL) framework that dynamically integrates real-time WIP status and machine setup requirements into scheduling decisions. Our approach enhances the Proximal Policy Optimization (PPO) algorithm with an online bootstrap strategy to enable continuous adaptation to fluctuating product mixes and order quantities. Additionally, we design a hybrid reward mechanism that jointly optimizes short-term objectives (e.g., reducing setup-induced idle time) and long-term goals (e.g., maximizing on-time deliveries), ensuring balanced decision-making in dynamic environments. Experimental evaluations across two real-world production scenarios demonstrate the superiority of our method over state-of-the-art RL and heuristic algorithms. Experimental results demonstrate that our approach outperforms existing RL-based and heuristic scheduling methods, achieving relative improvements of 7.07% and 5.07% in total on-time delivery rates, and corresponding relative gains of 2.89% and 3.27% in the number of lots delivered on time across two test scenarios. These results highlight the method’s ability to mitigate inefficiencies from frequent product changes, demand volatility, and prolonged setup adjustments. By enabling responsive lot-to-machine assignments and adaptive policy fine-tuning, our solution provides a robust foundation for scheduling in evolving SWFS environments.

Mobile precision equipment requires efficient and accurate road surface perception for reliable operation. However, existing road classification methods, designed mainly for single-type surfaces, perform poorly on mixed surfaces, limiting system adaptability in complex environments. To address this gap, we constructed a dedicated dataset comprising images of four single-type and two mixed road surfaces, acquired using a custom-built intelligent vehicle equipped with multiple sensors. Each image was meticulously annotated to facilitate model training and evaluation. Leveraging this dataset, we propose a deep learning model, termed GS-ResNet, which integrates squeeze-and-excitation (SE) modules, a gating mechanism, and a region of interest (ROI) extraction inspired by the patch paradigm of Vision Transformers. The SE modules enhance feature discriminability through dynamic channel-wise recalibration, while the gating mechanism refines feature extraction to improve texture perception. The ROI strategy focuses computational resources on the most informative image regions. Our GS-ResNet achieves a classification accuracy of 97.23% with an average response time of 0.5 ms, outperforming state-of-the-art methods in both accuracy and efficiency. This study addresses a critical gap in mixed road surface classification and provides an effective solution for robust road recognition in complex scenarios, offering support for the stable operation of mobile precision equipment across diverse road conditions.

To deal with the scheduling problem in rail vehicle assembly, where assembly line task allocation is complex and car body components require frequent cross-station transfers relying on trolleys, this study proposed an end-to-end hierarchical multi-agent deep reinforcement learning framework for scheduling optimization. Firstly, the allocation of assembly tasks across multiple assembly lines was modeled as a sequential decision problem. The high-level agent encoded the assembly task and line features using a Transformer and generated line assignment strategies with a Pointer Network. Secondly, the lower-level agents coordinated the selection of operations, station assignments, and dolly scheduling, and used Graph Attention Networks to extract relational features from heterogeneous nodes. Finally, multiple comparison experiments were conducted to validate the effectiveness of the proposed method. The results show that the method achieves optimal scheduling across different instance scales. The coordination of low-level agent strategies achieves an average maximum makespan gap of 11.36%, which outperforms the 15.00% achieved by the graph isomorphism network method, and the method provides high-quality scheduling with computation efficiency significantly higher than the Late Acceptance Hill Climbing algorithm. The proposed hierarchical collaborative scheduling framework achieves unified modeling and coordinated optimization of assembly task assignment and multi-resource scheduling, providing an efficient and adaptable intelligent optimization approach for rail vehicle assembly scheduling.

Active suspension systems can improve the operational stability of mobile precision equipment in the field while reducing equipment wear and maintenance costs. However, existing methods still exhibit limitations in generalization ability and forward-looking perception performance under real-world complex environments. Research on road surface classification in complex environments can provide new solutions for enhancing the forward-looking perception capability of active suspension systems. Firstly, this paper constructs a real-world multi-class road surface dataset named MTRSD, which includes image data of structured and unstructured road surfaces under various illumination conditions. On this basis, we propose the TF-ResNet road surface classification model. Its core components include an ambient illuminance compensation strategy and a texture feature embedding module. The illuminance compensation strategy adaptively adjusts image brightness to enhance the visibility of road surface features, thereby improving classification accuracy. The texture feature embedding module guides the model to focus on road texture patterns while suppressing background interference, thus increasing model stability. Experimental results show that the proposed method achieves an accuracy of 87.73% with a standard deviation of +/- 1.43% in the road surface classification task, outperforming existing mainstream methods.

New retail concepts that embrace a hybrid “online + offline” business paradigm promise super-fast order fulfillment of groceries within the next hour. In this ”online + offline” retail scenario, it is crucial to efficiently fulfill many online orders within the stipulated time while adhering to the layout rules of offline retail products. Zone-picking and overhead conveyors have been introduced to cope with the significant volume of orders batching, picking, and delivering fresh products. This has led to a new integrated order batching and picking decision problem, aiming for human-machine reconciliation in Industry 5.0. For such a problem, two new mixed integer linear programming models are developed, considering minimizing the number of picking task releases and the total delay time of all orders. The computational complexity of the two problems is provided. A customized two-stage heuristic framework is developed to solve the two models with distinct solution space structures. Numerical experiments have been conducted to test the performance of the proposed methods and provide solution analysis for practical insights. The results show that the proposed heuristic reduces the number of picking tasks for workers by 19% and the total delay in completing orders by 74% compared to prevailing store practices. The proposed framework complements the existing models in the literature. It contributes to developing a comprehensive analysis of order picking by integrating human factors into operational efficiency improvement in the new retailing industry.

This paper presents a theoretical study of hunting bifurcation behavior in high-speed rail vehicles under modal coupling and compares it to the traditional uncoupled system. A dynamic simplified model that integrates the lateral and yaw motions of the rigid bogie and the lateral motion of the carbody is established to evaluate the modal coupled effect between the carbody and bogie. The stability and Hopf bifurcation of the trivial equilibrium are first analyzed qualitatively using the normal form theory. The linear stability analysis then reveals that modal coupling introduces a new unstable region known as carbody (primary) hunting, which is absent in the uncoupled system that only exhibits bogie (secondary) hunting. The double-parameter Hopf bifurcation analysis is further carried out, which considers the influence of suspension parameters on the bifurcation speed and stability region. Our findings indicate that the dynamical behavior of the coupled system can closely match that of the uncoupled system with suitable parameter configurations, effectively reducing primary hunting and enhancing the overall hunting stability of rail vehicles.

This survey presents a comprehensive analysis of the transformative role of Industry 5.0 technologies in advancing smart elderly healthcare services, focusing on China’s evolving digital landscape. By examining the opportunities and challenges unique to China’s aging population in the context of rapid technological integration, the study outlines a system architecture that underpins next-generation elderly healthcare ecosystems. Central to this framework are five pivotal enabling technologies: the Internet of Things (IoT), Edge Computing, Service Robots, Big Data Analytics, and Digital Twin, each systematically explored to highlight their contributions to personalized, proactive, and decentralized elderly healthcare. The survey identifies critical application challenges, including interoperability gaps, ethical concerns in human-robot interaction, and data heterogeneity, while forecasting future trends poised to accelerate the adoption of Industry 5.0 principles. Emerging directions such as context-aware robotic collaboration, federated learning for privacy-preserving analytics, and metaverse-integrated digital twins are underscored as catalysts for sustainable, human-centric solutions. By bridging technological innovation with socio-ethical considerations, this work not only maps current advancements but also advocates for a paradigm shift toward empathetic, adaptive, and equitable elderly healthcare systems. The findings aim to inspire interdisciplinary collaboration among researchers, policymakers, and healthcare providers, fostering a transition to intelligent, dignity-preserving care models that align with the human-centered ethos of Industry 5.0.

The fault identification of Tunnel Boring Machines (TBMs) main bearing is crucial for infrastructure safety and cost control. However, the long-tailed distribution of fault types causes conventional models to be biased toward head classes when identifying tail classes. To address this issue, we initially establish a long-tailed fault dataset for TBM main bearings. Specimens with scratches and scuffs of varying severity are fabricated, and vibration signals are collected on a simulation test rig. Through the constructed causal graph and feature decoupling analysis, this study proposes a method (CIC-CML) to identify the long-tailed faults of the TBM main bearing based on causal inference. The Causal Inference Classifier (CIC) is designed to eliminate the negative impact of the momentum. Concurrently, a Class-Mean Loss (CML) is introduced to prevent momentum updates from excessively favoring the direction of head classes, avoiding the disappearance of discriminative features. As evidenced by experiments on the long-tailed fault dataset of TBM main bearings, our method not only maintains the training efficiency of a single-stage process but also demonstrates a particular strength in the identification of tail classes. Notably, it attains an overall accuracy of 89.029% with a ResNet-50 feature extraction backbone.

Predicting the remaining useful life (RUL) of the aircraft engine based on historical data plays a pivotal role in formulating maintenance strategies and mitigating the risk of critical failures. None the less, attaining precise RUL predictions often encounters challenges due to the scarcity of historical condition monitoring data. This paper introduces a multiscale deep transfer learning framework via integrating domain adaptation principles. The framework encompasses three integral components: a feature extraction module, an encoding module, and an RUL prediction module. During pre-training phase, the framework leverages a multiscale convolutional neural network to extract distinctive features from data across varying scales. The ensuing parameter transfer adopts a domain adaptation strategy centered around maximum mean discrepancy. This method efficiently facilitates the acquisition of domain-invariant features from the source and target domains. The refined domain adaptation Transformer-based multiscale convolutional neural network model exhibits enhanced suitability for predicting RUL in the target domain under the condition of limited samples. Experiments on the C-MAPSS dataset have shown that the proposed method significantly outperforms state-of-the-art methods. Graphical Abstract

Due to the limited computational power of edge devices in distributed manufacturing systems, the challenge of meeting real-time computing requirements for industrial big data arises. Additionally, the significant number of computational tasks results in considerable energy expenses. Therefore, it is crucial to effectively tackle the multi-objective cloud-edge collaborative task offloading problem (MOCECTOP). This paper focuses on two optimisation objectives: total computation time delay and computational energy consumption, which are related to promoting work efficiency and lowering carbon emissions. The weights of these two objectives are hard to determine at different production stages. The challenge is to get multiple models with various possible weight pairs of multiple objectives within a single training session and to achieve high-quality schedules for MOCECTOP in real-time production line control. We address this challenge by proposing a hierarchical parameter sharing (HPS) multi-objective optimisation framework based on multi-agent deep reinforcement learning. The network model comprises a task selection agent and a computing node selection agent. The task selection agent prioritises tasks for computation based on their state features, while the computing node selection agent allocates available computing devices to a selected task. Parameter sharing and collaborative training are employed to obtain a solution for the multi-objective problem when considering the variations in computational capabilities between the cloud and the edge. The HPS optimisation framework effectively fulfills the near real-time requirements for task computation in distributed manufacturing. Numerical experiments demonstrate that our HPS-based strategy quickly produces better schedules superior to those of existing multi-objective solution methods.

Cloud-edge technology enables near-real-time optimization of production lines in group-distributed manufacturing systems. Offloading some tasks to the cloud and processing the remaining tasks on the edge side can improve efficiency of the production optimization. However, due to the complexity of the manufacturing environment and various constraints, an effective offloading strategy is crucial to reduce computing delays and minimize transmission requirements for large-scale optimization requirements. This paper proposes a mixed-integer programming model and a deep reinforcement learning (DRL) framework, based on a Transformer, to address the cloud-edge offloading problem. The DRL framework consists of an encoder and decoder, designed using Transformer. Task offloading decisions are translated into two options: cloud offloading or edge retention. The encoder extracts relevant features for each option, and the decoder generates the probability of selecting each option based on the encoded information. Extensive computational experiments demonstrate the effectiveness of the proposed framework in solving the task offloading problem with time windows, achieving near-real-time optimization of production lines within competitive computational time.

Nucleus instance segmentation is crucial in the digital pathology, serving as a foundational step for subsequent tasks like precision medicine and cancer prognosis. In this paper, we introduce MAP-SegNet, a new network architecture designed for the segmentation and classification of nuclei in H&E stained images from multiple tissues. The key innovation of MAP-SegNet lies in its Multi-Attention Feature Fusion Module, which enhances the network's ability to efficiently combine multi-scale feature aggregation mechanisms with local and global features. The core component is the dual attention fusion module, which possesses the capability to adjust spatial attention weights dynamically by the input contextual information. This functionality enables the model to concentrate on various image regions, effectively capture long-range dependencies, and mitigate the risk of information loss. MAP-SegNet's uplifting accuracy in nuclei segmentation and classification is validated by rigorous quantitative experiments on the standard PanNuKe dataset.

As the intelligent manufacturing paradigm evolves, it is urgent to design a near real-time decision-making framework for handling the uncertainty and complexity of production line control. The dynamic flexible job shop scheduling problem (DFJSP) is frequently encountered in the manufacturing industry. However, it is still challenging to obtain high-quality schedules for DFJSP with dynamic job arrivals in real-time, especially facing thousands of operations from a large-scale scene with complex contexts in an assembly plant. This article aims to propose a novel end-to-end hierarchical reinforcement learning framework for solving the large-scale DFJSP in near real-time. In the DFJSP, the processing information of newly arrived jobs is unknown in advance. Besides, two optimization tasks, including job operation selection and job-to-machine assignment, have to be handled, which means multiple actions must be controlled simultaneously. In our framework, a higher-level layer is designed to automatically divide the DFJSP into subproblems, i.e., static FJSPs with different scales. And two lower-level layers are constructed to solve the subproblems. In particular, one layer based on a graph neural network is in charge of sequencing job operations, and another layer based on a multilayer perceptron is used to assign a machine to process the job operations. Numerical experiments, including offline training and online testing, are conducted on several instances with different scales. The results verify the superior performance of the proposed framework compared with existing dynamic scheduling methods, such as well-known dispatching rules and metaheuristics.

This paper investigates the bifurcation characteristics of a railway wheelset under active yaw control to evaluate the control laws for hunting motion. The creep forces between wheels and rail expressed as nonlinear functions of creepages are further modified. The piecewise exponent model handled by regularization method can accurately simulate the flange force. The control force provided by actuators is represented as a nonlinear function of state variables, including primary terms, quadratic terms and cubic terms. At first, the increase of linear control gains can observably increase the speeds of Hopf bifurcation and limit point of cycle bifurcation and reduce the limit cycle amplitude. The subcritical Hopf bifurcation can be transformed into supercritical by the linear control gains. The effect of displacement and velocity control gains on hunting frequency is opposite. Secondly, the limit point of cycle bifurcation speed can be remarkably affected by the quadratic nonlinear control gains. The quadratic terms can maintain the Hopf bifurcation speed constant. Only one term can achieve the transition of Hopf bifurcation form. In particular, the quadratic term disrupts the symmetry of limit cycles and induces complex nonlinear behaviors such as period-doubling bifurcation. Thirdly, increasing the cubic nonlinear control gains can slightly increase the limit point of cycle bifurcation speed and change the Hopf bifurcation type. The Hopf bifurcation speed is not affected by the cubic terms. Notably, there is a significant discrepancy in the order of magnitude required for control gains to reduce the control force and limit cycle amplitude. Finally, the combination control of linear and cubic nonlinear gains can noticeably increase the Hopf bifurcation speed while reducing the limit cycle amplitude. Hence, the reasonable combination control of linear and cubic nonlinear gains is extremely helpful to maintain the hunting stability.

A major challenge in structural reliability is that computational accuracy and efficiency are hard to balance, especially the time-variant hybrid reliability problem. To improve the reliability analysis efficiency under the condition that the accuracy is assured, this paper develops a new learning function named Regional Learning with Weighted Simulation (RLWS) to estimate the failure probability. The Kriging model is first employed as a response surface to fit the response of limit state function. An adaptive regional learning model as the core of the most probable point is then proposed to update the Kriging model, where the boundary of the updated region is determined by an attenuation function. Furthermore, based on the RLWS model, this paper develops a continuous maximin salp swarm algorithm to calculate the upper bound of the response value for time-variant hybrid performance function. Meanwhile, the Salp Swarm Algorithm is employed to calculate the lower bound of the response value. The Monte Carlo simulation can thus facilitate evaluation based on the final generated Kriging response surface. Several case studies are performed to test and validate the effectiveness of the proposed method and its applicability to practical engineering problems.

Due to the harsh working environment of storage stacking machinery, the fault information of important components is significantly complex, which leads to the problem of low classification accuracy and high computational complexity of existing deep learning-based fault diagnosis methods. To alleviate the problem, this paper presents a novel architecture named attention-based adaptive multimodal feature fusion networks for intelligent fault diagnosis of storage stacking machinery, which is aimed at improving the diagnostic precision and robustness of feature fusion network and learning the broader feature representation. Firstly, the long short-term memory layer is introduced to extract the feature information of multiple time steps to improve the self-extraction ability of multi-temporal features. Then, the maximum temporal feature fusion module is utilized to highlight the recognizability of deep fusion features. Finally, a residual layer with spanning connections is added to increase the utilization and characterization capability of deep fusion features. Experimental results demonstrate the effectiveness and superiority of the proposed method in intelligent fault diagnosis of storage stacking machinery under variable working conditions compared with some state-of-the-art deep learning-based methodologies.