Mass customization accelerates the decentralization of manufacturing systems, rendering distributed scheduling a critical research challenge. However, heterogeneous demands and inter-factory collaboration substantially increase the complexity of resource matching and spatiotemporal coordination. Moreover, adaptive task splitting strategies are essential to balance efficiency and flexibility under diverse demands. This study investigates a collaborative lot-streaming scheduling problem in networked multi-factory systems with variable interleaved sublots for mass customization (MC-NMFLSP-VIS) and develops a multi-action parameterized deep reinforcement learning (MAPDRL) approach. A multi-Markov decision process with high-dimensional states, hybrid actions, and composite rewards is formulated, and multiple coupled sub-policies are jointly optimized via multi-Proximal Policy Optimization. An adaptive task splitting strategy based on a continuous action space and a squashed Gaussian policy is developed to handle large-scale heterogeneous orders. Prioritized experience replay and population-based training enhance training stability and convergence speed, while multi-scale feature fusion and retraining enable transferability and continuous refinement. Numerical results demonstrate that MAPDRL achieves stable convergence and outperforms benchmark DRL methods and metaheuristics by over 13.74%, 11.12%, and 13.19% in total cost, maximum completion time, and system utilization, respectively, confirming its effectiveness and scalability in large-scale and complex environments.

In mass customization, order tasks must strike a balance between batch production and customization, minimizing costs while ensuring fast delivery. Effective collaboration between enterprises and suppliers necessitates careful consideration of order allocation granularity and dynamic resource production capacities to achieve optimal task distribution. This paper introduces a multi-order collaborative allocation model based on a product-process mix, aimed to address the order allocation challenges faced by core manufacturing enterprises in cloud manufacturing under mass customization. First, the limitations of existing order decomposition and resource constraint strategies in the order allocation process are analyzed. A product-process mixed multi-order collaborative allocation model is then established, focusing on minimizing multiple-order costs while accounting for dynamic resource capacity constraints. Next, an improved ant colony algorithm (IACO-a&b) is proposed, tailored to the features of the model. This algorithm enhances the initialization of the ant search solution space by combining the optimal solution searched by the ants, while dynamically adjusts pheromone volatility coefficients and sets path pheromone concentration intervals, so as to avoid preventing the algorithm from precociousness. Finally, the performance of IACO-a&b is compared with several other algorithms, and the proposed order allocation model is evaluated against traditional models in this paper. The experimental results demonstrate that, in large-scale case studies, IACO-a&b improves the best fitness value by 11.89% and increases the fitness excellence rate by 60%. Furthermore, the proposed model reduces the leveling volatility index (LVI) of resources by 50.79% without significantly increasing order costs.

In intelligent manufacturing, collaborative production across multiple factories has become a key strategy for large enterprises to enhance flexibility and competitiveness. However, the proliferation of equipment-intensive manufacturing has made performance fluctuations and maintenance demands caused by machine degradation critical challenges that limit system efficiency and stability. Therefore, this paper investigates the networked multi-factory collaborative production scheduling problem with machine degradation and preventive maintenance (NMFCPS/MDPM) and proposes a multi-action deep reinforcement learning approach. Specifically, reliability-based deteriorating processing time and maintenance strategy are incorporated into the problem and modeled as a multiple Markov decision process with a high-dimensional state space, two-stage action space, and composite reward. A multi-Proximal Policy Optimization architecture is then constructed to optimize coupled sub-policies, and learning rate decay, prioritized experience replay, and retraining mechanisms are incorporated during training to enhance policy stability, precision, and adaptability. Numerical experiments show that the proposed approach rapidly reaches stable convergence, improving three scheduling metrics by 15.28%, 20.94%, and 5.91% over various reinforcement learning and metaheuristic algorithms, while also validating the effectiveness of the maintenance strategy in enhancing system efficiency and stability.

In equipment-intensive manufacturing, machine degradation and breakdowns threaten system robustness and production continuity, while multi-layer couplings and dynamic collaboration in mass-customized networked multi-factory systems exacerbate these challenges. This study addresses the robust scheduling problem for networked multi-factory systems with machine degradation and breakdowns under mass customization (MC-NMFRS-MDB) and proposes a hierarchical multi-head deep reinforcement learning (HMH-DRL) approach. The problem is formulated as a multi-Markov decision process with a high-dimensional state space capturing degradation and breakdown dynamics, a dual-layer action space addressing coupled subproblems, and a multi-head reward balancing performance and robustness. A hierarchical multi-head policy architecture enables coordinated decision-making and multi-objective optimization. The training process employs prioritized experience replay to improve learning efficiency, nonlinear clipping to stabilize policy updates, and a retraining mechanism for continual adaptation. Numerical results indicate that HMH-DRL reliably converges, improving scheduling performance and robustness by over 10.59% and 10.98%, respectively, compared with four DRL methods, three metaheuristic algorithms, and one reliability-driven heuristic, confirming its effectiveness in complex and dynamic scenarios.

Predictive-reactive rescheduling combines pre-scheduling with real-time responses to maintain production stability under disturbances, and is extensively used in dynamic production environments. For multi-objective optimization of Dynamic Flexible Job Shop Scheduling Problem (DFJSP), the problem’s complexity increases significantly, challenging traditional algorithms to provide high-quality global solutions with low latency. This study proposes an innovative dual hyper-heuristic (DHH) method for rescheduling in DFJSP, aiming to generate rescheduling solutions in real-time and optimize two practical objectives: rescheduling stability and makespan. A rescheduling framework is proposed to decompose the global optimization problem into the sequence repair decision process based on the operation-to-interval transfer. The framework comprises two-stage hyper-heuristics: rule generation and selection. In the first stage, genetic programming is used to generate local repair heuristics as candidate action set. A double-tree encoding for selecting operations and intervals respectively is designed, and a terminal set combining real-time operating conditions and critical path characteristics is constructed to improve search efficiency. In the second stage, deep reinforcement learning is used to accurately relocate operations based on repair heuristics to achieve global optimization. A knowledge sharing pattern across scenarios is developed to enhance the generality of the algorithm. Comprehensive experiments on benchmark extensions demonstrate that the proposed algorithm outperforms widely used multi-objective meta-heuristic algorithms while significantly reducing computation time.

Dynamic gesture recognition is essential for natural human-computer interaction. Existing methods often struggle to jointly model fine-grained short-term motions and long-range temporal dependencies, limiting recognition of complex gesture sequences. To address this issue, a heterogeneous dual-branch framework, termed TS-GRN, is proposed for collaborative multi-scale spatiotemporal modeling. The Dynamic Snapshot (DS) component captures short-term motion dynamics by integrating a Temporal-Spatial-Channel Attention (TSCA) module into a 3D-MobileNetV2 backbone to enhance local spatiotemporal representation. The Temporal Continuity (TC) component models long-range temporal dependencies using ConvLSTM with a Spatio-Temporal Dynamic Memory Augmentation (ST-DMA) mechanism, which introduces auxiliary memory interaction to improve temporal information propagation. The two branches are fused in a shared embedding space to exploit complementary shortand long-term representations. Experiments on the Jester and EgoGesture datasets achieve Top-1 accuracies of 96.646% and 94.856%, respectively, demonstrating the effectiveness and robustness of TS-GRN for dynamic gesture recognition.

Amid escalating global energy and climate crises, the industrial sector faces increasing pressure to conserve energy and reduce emissions. Consequently, emerging technologies and policies (renewable energy sources, battery energy storage systems (ESS), and demand response (DR)) have been widely adopted to improve energy efficiency in industrial production. However, the limitation of integrated optimization tool tailored to workshop operations often result in suboptimal production and energy management decisions. This paper proposes an energy-efficient scheduling model for flexible job-shops that incorporates two DR policies and a photovoltaic energy storage system (EFJSP-DPE). The model simultaneously optimizes machine allocation, on-site energy flows, and ESS operations, targeting minimal makespan, total power consumption, and total energy cost. A learning-driven multi-objective memetic algorithm (LMMA), enhanced by embedding Q-learning into neighborhood selection, is developed for solving EFJSP-DPE effectively. Experimental results using random and benchmark instances confirm the algorithm's superiority. A real-world case study further illustrates the proposed method's effectiveness, achieving energy cost reductions of up to 55.9% and 51.5% under various supply strategies and policy scenarios. These findings underscore its potential to provide theoretical guidance and decision support for sustainable manufacturing operations.

PURPOSE:To enhance the temporal feature learning capability of the laparoscopic cholecystectomy phase recognition model and address the class imbalance issue in the training data, this paper proposes an Xception-dual-channel LSTM fusion model based on a dynamic data balancing strategy. METHODS:The model dynamically adjusts the undersampling rate for each surgical phase, extracting short video clips from the original data as training samples to balance the data distribution and mitigate biased learning. The Xception model, utilizing depthwise separable convolutions, extracts fundamental visual features frame by frame, which are then passed to a dual-channel LSTM network. This network is composed of a temporal mapping bidirectional LSTM structure and a sequence embedding LSTM structure, both working in parallel. The dual-channel LSTM network models the temporal dependencies between adjacent frames, capturing the contextual temporal information to perceive the dynamic feature changes of the surgical phases. Finally, the surgical phase is determined by combining the prediction scores from both channels. RESULTS:Experimental evaluation on the public dataset Cholec80 demonstrates that the proposed model outperforms traditional single-channel LSTM models. Moreover, compared to the model without the dynamic data balancing strategy, the F1-scores for all surgical phases have been improved. CONCLUSION:The experimental results validate the effectiveness of this strategy in extracting temporal feature information, alleviating the data class imbalance issue, and enhancing the overall detection performance of the model.

With increased market competition and the development of Industry 4.0, production operations is being driven towards increased flexibility. Multi-product flow and variable production processes on the shop floor complicate waste identification, analysis, and reduction. To continuously improve based on real-time information in dynamic production processes, this paper proposes an integrated approach of dynamic value stream mapping and hybrid simulation(DVSM-HS). The dynamic value stream mapping(DVSM) is extended from time temporal and economic dimensions respectively to decompose and integrate work-in-process sequences and processes. A dynamic cost model is constructed based on time-driven activity-based costing to extract the variable costs of production resources. By combining hybrid simulation of discrete events and system dynamics, variations in multi-product flows are captured and mapped, and their economic impact over time is quantified. A case study in an automotive painting workshop demonstrates the practical application of waste reduction and decision-making across different scenarios. The results show that the proposed DVSM-HS can address process variability by identifying the temporal and spatial characteristics of waste. Based on waste localisation and economic evaluation, this study supports managers diagnose problems and conduct Lean improvements in real-time operations, optimising resource and capability allocation to enhance production profitability.

The detection of defects on automotive coated surfaces is of paramount importance to ensure the quality of automotive appearance. However, due to the challenges posed by the lack of sufficient samples of product-specific coated defects, the uneven number of species, and the difficulty of detecting defects in small targets due to the complex background of the body surface. The detection of defects on automotive coated surfaces still relies on experienced manual labor to a significant extent. To address these issues, Poisson fusion is employed to enhance the data for generating defect images, thereby mitigating the limitations of insufficient coated defect samples and an imbalanced distribution of defect species. Based on this, a new YOLOv8-SC defect detection model is proposed, in which the C2f-Star is used to replace the original C2f structure. This aims to improve the model's ability to extract defects of varying sizes in complex backgrounds, while reducing the number of model parameters and the amount of computation. The content-guided attention fusion (CGAFusion) module, situated before each detection head of the model, has been designed to enhance the model's capacity to extract defects of varying sizes and complexity. This is achieved through the adaptive fusion of low-level and high-level features. The proposed model's superiority in terms of detection metrics has been demonstrated through the use of example validation. The experiments were conducted on the enterprise dataset CPD-DET and the public dataset NEU-DET. The results demonstrated that the defective image generation and data enhancement method could significantly improve detection performance and have good generalization. The proposed YOLOv8-SC reduces the model parameters by 12.2% compared to the normal model. Additionally, the mean average precision (mAP) at 50% and 50-95% thresholds are enhanced by 4.7% and 2.3%, respectively. Furthermore, the accuracy of the model exhibits a discernible positive growth trend with an increase in the sample set size. The design of an automotive coated surface defect detection system is presented at last. When deployed in industrial settings, this system can facilitate the intelligent detection of coated defects.

ObjectiveThis study aimed to examine the impact of non-driving-related tasks (NDRTs) on drivers in highly automated driving scenarios and sought to develop a deep learning model for classifying mental workload using electroencephalography (EEG) signals.MethodsThe experiment involved recruiting 28 participants who engaged in simulations within a driving simulator while exposed to 4 distinct NDRTs: (1) reading, (2) listening to radio news, (3) watching videos, and (4) texting. EEG data collected during NDRTs were categorized into 3 levels of mental workload, high, medium, and low, based on the NASA Task Load Index (NASA-TLX) scores. Two deep learning methods, namely, long short-term memory (LSTM) and bidirectional long short-term memory (BLSTM), were employed to develop the classification model.ResultsA series of correlation analyses revealed that the channels and frequency bands are linearly correlated with mental workload. The comparative analysis of classification results demonstrates that EEG data featuring significantly correlated frequency bands exhibit superior classification accuracy compared to the raw EEG data.ConclusionsThis research offers a reference for assessing mental workload resulting from NDRTs in the context of highly automated driving. Additionally, it delves into the development of deep learning classifiers for EEG signals with heightened accuracy.

Multi-scale defect features, blurred edges and inability to locate geometric features have been the three key factors limiting the detection of surface defects on quality control system in the industrial manufacturing process. In this study, a method based on the fusion of multi-scale features and pixel-level semantic segmentation is proposed for the detection of surface defects. The proposed method firstly fuses multi-level feature maps to balance the expressiveness of multi-scale features, then adds a boundary refinement module to enhance the accurate inference of edge fine-grained, and finally adopts an en-decoder architecture to locate geometric features at the pixel-level for each type of defects, realizing intelligent detection of geometric features of end-to-end multi-scale defects on the surface of parts. We conduct experiments using the collected parts datasets to evaluate the effectiveness of our framework. The experimental results show that the proposed model achieves MIoU of 80.1%, the recognition accuracy reaches more than 95 %, and a detection rate of up to 29.64 FPS, demonstrating the advancement and effectiveness of the proposed method with less misclassification and superior generalization performance and has progress and effectiveness in detecting surface defects of multi-scale features. It provides a research idea for the subsequent realization of surface quality inspection in the manufacturing process system.

A simple yet sensitive colorimetric method based on in situ formation of AuNPs in liquid bead-headspace microextraction of arsine has been developed for ultrasensitive sensing of trace As(III). Taking advantage of our self-designed sensing device, the strong reductant AsH3 generated from the hydride generation of As(III) passed through a platform and reacted with a liquid bead (60 ??L, containing HAuCl4, KI, PVP and PVA) for the formation of AuNPs inside. Due to the synergistic reduction of iodine ions and the stabilization of polymers, AuNPs could be formed quickly, and the color of liquid-bead changed from yellow to red, which deepened in proportion to the As(III) concentration. Under the optimal experimental conditions, a linear range of 10???500 ??g/ L with a correlation coefficient up to 0.999 was achieved. Trace As(III) as low as 5 ??g/L could produce visible color change. Combining the advantages of hydride generation, headspace liquid-phase microextraction and AuNPs-based colorimetric assay, including efficient matrix separation and analyte enrichment, rapid mass transfer, and visual observation of color change, the proposed sensor features high sensitivity, great selectivity, strong stability and simplicity, enabling successful applications for As(III) detection in water samples.

Hepatoid adenocarcinoma of the duodenum is a rare special type of adenocarcinoma, featured by hepatocyte components in primary adenocarcinoma of the duodenum. It has the characteristics of high malignancy, invasiveness, rapid progress, and poor prognosis. An abnormal elevation of serum alpha-fetoprotein (AFP) may occur in most cases. The diagnosis is mainly based on pathological morphology. Here, we reported a case of hepatic adenocarcinoma of the duodenum. The middle-aged female patient had an ampulla mass at diagnosis and received radical pancreaticoduodenectomy. The postoperative pathology was stage IIIA duodenal adenocarcinoma. At 1 month after surgery, she had multiple intrahepatic metastases and retroperitoneal lymph node metastasis; the AFP level was 300 ng/ml at that time. As she refused target therapy, two cycles of capecitabine-oxaliplatin (XELOX) chemotherapy were performed. However, the AFP elevated from 300 to 1,931.90 ng/ml, and the disease progressed rapidly. Immunohistochemistry (IHC) of tissue samples from presurgical endoscopic ultrasound guided fine needle aspiration (EUS-FNA), surgery, and liver biopsy showed positive AFP staining. Combining the abnormal elevation of serum AFP and microscopic pathological morphology, this case is diagnosed as hepatoid adenocarcinoma of the duodenum with liver metastasis. The physical condition of this patient was too poor to receive follow-up treatment. She died of the rapid disease progression with an overall survival time of 161 days. Considering that in most patients with hepatoid adenocarcinoma the abnormal elevation of serum AFP occurs preoperatively and returns to normal postoperatively rather than normal before surgery and increased after surgery, the primary lesion is located in the stomach rather than the intestine, and the patients are more often older men rather than middle-aged women; this case is rare particularly. Therefore, reporting this case with complete case data may be helpful to further study, so as to improve the understanding of this special type of malignant tumor.

Fatigue driving is one of the main causes of traffic accidents. In order to solve this problem, a new Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) based real-time driver fatigue detection method is proposed. First of all, using simple linear clustering algorithm (SLIC), the driver's image is divided into super pixels of uniform size, which are used as input of CNN, and CNN is trained to automatically learn the features of eyes and mouth contained in the image, and then the location and area of eyes and mouth are obtained by using the trained CNN. On this basis, the eye feature parameter Perclos, mouth feature parameter MClosed and face orientation feature parameter Phdown are extracted, and the above feature parameters on the continuous time series and steering wheel angle feature parameter SA are taken as the input of LSTM, and the fatigue level is taken as the output to detect the fatigue state of the driver in real time. Experimental data shows that this method can not only overcome the influence of illumination, background, angle and individual differences, but also the accuracy of detection can reach 99.78%, and the average detection time is 16.94 ms/frame.

Road detection algorithms with high robustness as well as timeliness are the basis for developing intelligent assisted driving systems. To improve the robustness as well as the timeliness of unstructured road detection, a new algorithm is proposed in this paper. First, for the first frame in the video, the homography matrix H is estimated based on the improved random sample consensus (RANSAC) algorithm for different regions in the image, and the features of H are automatically extracted using convolutional neural network (CNN), which in turn enables road detection. Secondly, in order to improve the rate of subsequent similar frame detection, the color as well as texture features of the road are extracted from the detection results of the first frame, and the corresponding Gaussian mixture models (GMMs) are constructed based on Orchard-Bouman, and then the Gibbs energy function is used to achieve road detection in subsequent frames. Finally, the above algorithm is verified in a real unstructured road scene, and the experimental results show that the algorithm is 98.4% accurate and can process 58 frames per second with 1024×960 pixels.

•Image classification under different unfavorable visual conditions is carried out based on gray and definition features.•Based on adaptive dynamic image enhancement algorithm, the images under different unfavorable visual conditions are enhanced.•The adaptive gray features and statistical morphological features are used to get the candidate regions of the lane lines, and the improved probabilistic Hough transform algorithm is adopted to extract accurate Lane line from the candidate regions.•The region growth algorithm based on Gaussian model is used to obtain the drivable regions.

Remanufacturing has been regarded as a key technology for the sustainable development of the world economy. However, the quality of remanufactured products is difficult to meet the needs of customers, which has become a bottleneck restricting the development of remanufacturing industry. Therefore, remanufacturing assembly management and technology has more important engineering significance to improve production efficiency and quality of remanufactured products. Aiming at the lag of the research on management and control of remanufacturing assembly, this paper summarizes the research status of remanufacturing assembly on the basis of studying the characteristics and connotation of remanufacturing assembly system. Intelligent robot of remanufacturing assembly driven by large data and intelligent manufacturing will become an important development direction in the future. The research direction of remanufacturing assembly is summarized, and the future research trend of remanufacturing assembly is pointed out. It can be predicted that the intelligent robot of remanufacturing assembly will be the key to the flexibility, intellectualization, precision, and integration of remanufacturing assembly system in the future. This study provides a direction for the development of remanufacturing industry.