. Global warming has become a critical constraint on global sustainable development, with manufacturing widely regarded as the core entity undertaking the "zerocarbon" mission. Amid accelerating green and low-carbon transformation, enterprises must simultaneously deploy emission reduction strategies at both production and channel levels while navigating the interplay among government regulations, consumer preferences, and carbon market mechanisms. To address this challenge, this study constructs a two-level supply chain game model comprising a manufacturer and a retailer, where consumers make product choices based on their low-carbon preferences. Employing game theory and comparative static analysis, this research examines how carbon quota, carbon trading prices, and consumers' green preferences influence manufacturers' carbon reduction and channel decisions. Our analysis reveals the following findings. First, dualchannel operations combined with emission reduction investments generate the highest profits for manufacturers by enabling broader market coverage and capturing the demand premium from low-carbon preferences. Second, retailers' profits also increase with carbon prices, quotas, or consumer preferences under both single and dual-channel emission reduction scenarios. Third, the emission reduction rate remains higher in the dual-channel model than in the single-channel model under identical conditions, amplifying incentives for low-carbon investment.
To enhance the error-correction capability of neural network decoders for polar codes and minimize the required memory overhead, this paper proposes a low-precision successive cancellation (SC) decoding algorithm. The algorithm is assisted by deep neural networks and incorporates quantization and partitioning techniques. Specifically, it quantizes the weights of the deep neural network decoder to a limited number of bits, utilizing an 8-bit quantized input and output in the Q8.4 fixed-point digital format for all arithmetic operations, thereby achieving int8 computations. With quantization-aware training, quantization errors are accounted for within the training loss, allowing the deep neural network to effectively learn how to decode with reduced precision. The proposed scheme, by employing fewer floating-point weights, reduces both memory overhead and the number of floating-point operations. Simulation results indicate that the proposed decoder decreases memory usage and decoding latency while maintaining a slightly improved bit error rate (BER) performance.
. This article presents an open-source intrusion detection and notification (IDN) framework designed to detect attacks targeting Siemens S7-series programmable logic controllers (PLCs) in operational technology (OT) environments. The framework is validated using a simulated weight-based sorting system developed in Factory I/O, with TIA Portal control programs deployed on S7-300, S7-1200, and S7-1500 PLCs. Suricata serves as the core intrusion detection engine, while the Elasticsearch, Logstash, and Kibana (ELK) stack and LINE Notify are integrated to visualize alerts and provide realtime operator notifications. Simulated attacks are carried out using Snap7 to interact with each PLC during live process operations. The experimental results show that the framework reliably detects intrusions across all tested S7 PLC models and delivers timely alerts, demonstrating its effectiveness for real-time monitoring and incident response. In contrast to previous studies that focus primarily on protocol analysis or on individual PLC types, this work offers a practical and scalable intrusion detection solution validated on real hardware and designed to accommodate the coexistence of legacy and modern controllers within OT systems.
This article shows a model to determine the minimum contact area with the soil for trapezoidal combined footings considering that the area is partially compressed, i.e., a part of the contact area under the footing with the soil is subject to compression and the other part has zero pressure. Some works present the minimum contact area for trapezoidal combined footings, but the contact area of the footing with the soil is completely compressed. The methodology is developed by integration to determine the equations of the resultant force and the moments on the X and Y axes for the five cases of biaxial bending, and four cases of uniaxial bending (two on the X axis and two on the Y axis). Several studies are developed to determine the minimum contact area with the soil for trapezoidal combined footings subjected to biaxial bending and uniaxial bending in each column. The proposed model shows a significant reduction in the minimum contact area with the soil of up to 65.30% for biaxial bending and up to 48.68% for uniaxial bending compared to other studies. The model can be used as a review to minimize the allowable load capacity of the soil (objective function) and the same constraint functions for biaxial or uniaxial bending.
. Existing studies on the evaluation benchmarks and approaches of large language models (LLMs) primarily concentrate on accuracy, robustness, bias and security. As the application of LLMs in the metro domain intensifies, it is imperative to construct a specific LLM evaluation framework and benchmark for evaluating LLMs. We collected the official information released via metro companies' official websites, WeChat official accounts, Weibo, Douyin, as well as other professional sources in the metro domain. After data cleaning, the original data were categorized into four types: passenger service, line operation and maintenance, emergency management, and safety operation, and the training dataset, evaluation benchmark and criteria were constructed. Then, multiple agents are constructed through fine-tuning the Qwen-7B model and prompt engineering, and MetroEval framework is established by multi-agent collaboration to evaluate the ty, and ethics of LLMs in the metro domain. By evaluating and comparing the indicators of six models, it is discovered that retrieval-augmented generation (RAG) and fine-tuning can effectively enhance the performance indicators of the base models. It is expected that MetroEval can provide ideas and assistance for the future research and application of LLMs in the metro domain.
. Aggregation and decision-making of linguistic evaluation information in uncertain environments has garnered considerable research attention. This paper proposes a develops a multi-attribute group decision-making method based on this operator to tackle qualitative linguistic assessments. First, we define a support degree function between linguistic interval-valued spherical fuzzy numbers to quantify the correlation among expert evaluations. Subsequently, we construct a Linguistic Interval-Valued Spherical Fuzzy Power Weighted Average operator that incorporates information relevance to effectively fuse multi-expert evaluations. Finally, we develop an objective attribute weight determination method based on similarity analysis of expert assessments, establish a complete multi-attribute group decision-making method framework, and validate the proposed method's rationality and effectiveness through an illustrative example.
Prescription errors and adverse drug interactions remain critical challenges in healthcare, particularly in developing countries like Vietnam, where medical data is often fragmented and unstructured. This study proposes a symptom-based drug recommendation system to assist physicians and pharmacists in making accurate prescriptions. The system utilizes artificial intelligence and deep learning to analyze multidimensional medical data, including symptoms, drug side effects, comorbidities, and patient feedback. Three approaches are explored: Retrieval-Augmented Generation (RAG), Knowledge-Augmented Generation (KAG), and Graph Retrieval-Augmented Generation (GR AG). RAG integrates retrieval-based information with generative models, KAG incorporates domain knowledge for better reasoning, and GRAG leverages knowledge graphs with Graph Neural Networks (GNNs) such as Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), Message Passing Neural Networks (MPNN), Relational Graph Convolutional Networks (R-GCN), GraphSAGE, and Graph Transformers (GTs). These GNN models predict drug candidates and risk scores, which inform RAG queries for external knowledge. Evaluations on MIMIC-III (subset) and Medical Recommendation System datasets show GTs achieving 92.77% accuracy in symptom-to-disease classification, with GRAG yielding an F1-score of 0.407 in drug recommendation (19.0% improvement over RAG) and reducing DDI rate to 2.8%. This system could reduce errors, enhance personalized prescriptions, and optimize healthcare in resource-constrained settings.
. With the recent proliferation of Chinese historical text databases, the digitization of ancient documents in China has largely focused on shallow knowledge services such as document scanning, organization, and keyword-based retrieval. However, the rapid development of generative artificial intelligence offers new opportunities for advancing the depth and scope of digital humanities research. In this study, we construct a seismic dataset extracted from Chinese historical texts and present a domain-adaptive training framework based on the DeepSeek open-source large language model. Our approach includes continued pre-training, supervised fine-tuning to develop a generative dialogue model tailored for historical seismic analysis. We further evaluate the model through both automatic metrics and human expert reviews. While the BLEU and ROUGE scores across dialogue categories appear relatively low - suggesting a divergence from typical surface-level patterns - this reflects the model's robust domain creativity and adaptability. Training loss convergence also demonstrates successful model optimization. These results collectively highlight the model's potential to support nuanced understanding and intelligent interpretation of earthquake-related narratives embedded in historical Chinese texts.
. This paper presents an advanced biomechanical model with eight degrees of freedom (8-DOF) to rigorously characterize the dynamic interactions between wheelchair occupants and seating systems. In contrast to traditional single- or two degree of freedom (SDOF/TDOF) approaches that oversimplify occupant-seat coupling, the proposed 8DOF framework explicitly resolves the complex mechanical pathways governing vibration transmission through anatomical and structural subsystems. By incorporating precisely calibrated stiffness and damping parameters derived from seven commercial wheelchair cushions, the model achieves high fidelity in predicting seat-to-occupant transmissibility across the critical 0-20 [Hz] frequency range. Experimental validation using real-world chanical responses under realistic excitation. Notably, cushions with low stiffness and low damping produced amplified transmissibility peaks near 3-4 [Hz] - a range corresponding to human vertical resonance - while designs with optimized damping-stiffness combinations effectively mitigated these vibrations. These findings emphasize the pivotal role of energy dissipation in preventing resonance amplification and promoting long-term user comfort. The validated 8-DOF framework thus offers mechanical, computational, and translational advantages, serving as a robust tool for designing next-generation seating systems that balance biomechanical protection with engineering performance.
. Disturbance observer-based control has been recognized as one of the most promising approaches for disturbance attenuation, and many papers on design methods of disturbance observers have been published. Recently, the parameterization of all disturbance observers for plants with any disturbances has been clarified. However, these methods generally require the availability of the control input. Since there are systems which the control input of system is not unavailable, it is essential to design unknown input disturbance observers. In this paper, we clarify the parameterization of all unknown input disturbance observers for plants with general exogenous output disturbances. Furthermore, we present a design method applicable to such observers.
This paper presents an on-line identification of continuous-time nonlinear systems using a moving-window type Gaussian process (GP) model and genetic algorithm (GA) with variable number of generations. The GP is a Gaussian random function and is constructed by its mean function and covariance function. During the training phase of this model, the hyperparameters of the covariance function and the moving-window length are optimized by the modified GA, whereas the system parameters of the linear terms and the weighting parameters of the mean function are estimated by the recursive least-squares (RLS) method. In order to reduce the computational burden, the number of generations of the GA is adjusted according to the time changes of the system and only the input-output data extracted by the k-means method are utilized in both GA and RLS method. The effectiveness of the proposed method is demonstrated through numerical simulations for a simplified power system. Simulation results show that the computational time of the proposed method is reduced to about 40% of that of the conventional method without deteriorating the identification accuracy.
At present, the current national standards for China's tobacco industry classify tobacco leaves into 42 grades based on 7 indicators. However, these indicators lack a unified quantitative standard, resulting in prominent problems of strong subjectivity and low efficiency in grading. To solve the issues of difficult image feature extraction caused by the curling and folding of redried tobacco leaves, and limited classification accuracy restricted by small inter-class differences during the sorting and grading process of redried tobacco leaves, this article proposes an intelligent tobacco leaf grading method that integrates multi-attention fine-grained features with color and texture features. Firstly, the RGB/HSI dual-space color statistical strategy, Local Binary Pattern (LBP), and Gray-Level Co-occurrence Matrix (GLCM) are used to extract color and texture information, respectively. Meanwhile, a fine-grained classification network model based on multi-attention sampling is designed, which utilizes the attention mechanism to capture global features and local texture details, making up for the defect that traditional convolutional networks are insufficient in distinguishing subtle features. Finally, the Particle Swarm Optimization (PSO) algorithm is introduced to optimize feature weights, and a multi-model integration framework is constructed by fusing color and texture features. In addition, an image dataset of tobacco leaves at various grades is established. Experimental results show that the fine-grained classification network fusion model designed in this paper achieves recognition accuracies of 84.9% and 79.6% for binary classification and five-class classification, respectively, both of which are superior to those of the ResNet and InceptionV3 models. The intelligent tobacco leaf sorting and grading system developed based on this method realizes fully automatic grading of tobacco leaves and effectively improves production efficiency.
To improve model interpretability in clinical medical datasets such as chronic kidney disease data, this study utilizes information granular technique to propose a data reconstruction and feature selection method. First, the information granulation method grounded in the principle of justifiable granularity is applied for data reconstruction. Then, a modified Davies-Bouldin index is developed as an evaluation criterion for a set of information granules. Furthermore, a novel task-oriented loss function integrating classification objective is designed, and a heuristic feature selection algorithm is developed through cross-validation procedures. Experimental validation on chronic kidney disease datasets demonstrates that the proposed information granulation-based data reconstruction method effectively identifies critical features while maintaining satisfactory classification performance.
Simultaneous Localization and Mapping (SLAM) based on 3D Gaussian Splatting (3DGS) have demonstrated significant advantages in high-fidelity map reconstruction and real-time rendering. In particular, Inertial Measurement Unit (IMU) has proven to be highly effective in enhancing SLAM performance by providing robust motion constraints in visually degraded environments. However, existing visual-inertial SL-AM systems using 3DGS often suffer from poor initialization and insufficient IMU bias correction, which limits their robustness in challenging scenarios. To address these issues, we propose MTCGS, a tightly coupled multi-modal SLAM framework that integrates 3DGS with inertial measurements. Our method employs joint optimization of IMU pre-integration residuals and 3DGS rendering loss to refine both camera poses and inertial noise parameters, enhancing pose estimation especially in the presence of visual degradation. For map reconstruction, we introduce a hybrid keyframe selection strategy combining a sliding window with adaptive optimization to prevent catastrophic forgetting. Furthermore, 3D Gaussian smoothing and an improved isotropic regularization loss are applied to suppressing high-frequency artifacts and enhancing reconstruction quality. We evaluate our system on the OpenLORIS and TUM-RGBD datasets. Experimental results show that MTCGS outperforms the representative method MM3DGS in both localization accuracy and mapping robustness, particularly under conditions of noisy inertial data.
. This paper proposes one innovative control strategy - multi-switching control with supervised learning for unmanned helicopter's automatic takeoff and landing process. More specifically, firstly before showing our proposed control strategy, the detailed description and analysis on automatic takeoff and landing of unmanned helicopter are given, respectively, meaning they are all divided into many subprocesses, so the controller design problem is turned to devise each subcontroller for each subprocess. Secondly, to better understand the control principles for automatic takeoff and landing process, three different points are analyzed from constraint analysis, automatic takeoff control and landing control, bringing the idea of model predictive control to consider these different constraints. Thirdly, based on our above detailed analysis about automatic takeoff and landing process, i.e., both entire processes can be separated into many subprocesses, we combine multi-switching control and supervised learning together to be supervisory multi-switching control. After separating the total time interval for automatic takeoff and landing process into five subintervals, corresponding to five subprocesses, five subcontrollers are designed for each subprocess, while introducing one supervisor mechanism to achieve that switching function. Finally, one practical platform is constructed, and some simulation results are given to prove our theoretical results.
Belt tears can lead to equipment failure, fire, and other accidents. To address the diverse forms of belt tears and the interference caused by harsh working conditions on identification accuracy, this study introduces a lightweight target detection algorithm, DDC-YOLO11, to improve the robustness of tear detection in complex scenarios. Firstly, the Dilation-wise Residual and Dilated Reparam Block (DWB) module is embedded in the C3k2 module of the YOLO11 backbone network. It utilizes a complex feature extraction and fusion mechanism to enhance the model's ability to detect targets of different scales and locations in the image. Additionally, the attention module in the C2SPA module is replaced with a variable self-attention mechanism (DAttention), which boosts the model's capacity to extract image features through adaptive adjustment of the attention area. Furthermore, the Context Anchor Attention-High-level Screening-Feature Fusion Pyramid Network (CAA-HSFPN) is adopted in the neck network. It uses adaptive pooling technology to reduce the feature map dimension and significantly decrease the number of parameters. Images acquired from the production site are preprocessed using the Adaptive Enhancement method (AdaEnhance) to improve recognition performance. Experimental results show that the DDC-YOLO11 model achieves 92.5% mAP50, with 4.2M weight parameters, 1.957 & times; 106 total parameters, and a detection speed of 100.8 FPS.
. As diversity of characteristics, quantitatively analyzing the pulmonary nodules in CT images remains a big challenge. Existing CNN-based segmentation methods are hindered by redundant characteristics of convolutional operation, and scarcity of annotated data for the peripheral tissues. In this paper, we proposed a novel network for automatic pulmonary nodule segmentation, named RICSBU-Net, which models the relationship between the encoder and decoder via a bidirectional convolutional LSTM. The main contribution includes the following. Firstly, we replace U-Net's encoder with a pretrained ResNet-18 to overfit and prevent semantic feature loss typically associated with pooling layers. Secondly, an ICS-Attention module is introduced to mitigate performance degradation caused by variations in nodule morphology, location, and density. This module merges fine-grained spatial and semantic information to amplify the weight of important features. Thirdly, the ICS-BiConvLSTM module replaces the skip connection and concatenator, with the ICS-Attention module emphasizing key features to reduce parameters, and BiConvLSTM enhancing the extraction of local features. Various experiments on the LUNA16 dataset demonstrate that RICSBU-Net surpasses other advanced methods in both quantitative assessment and qualitative metrics, achieving a DSC of 95.84%. Comparing to the other up-to-date segmentation methods, RICSBU-Net delivers more refined and consistent boundary delineation, particularly for nodules near adjacent tissues.
This study investigates the impulsive synchronization control of positive complex dynamical networks with nonlinear couplings. A generalized network model is formulated by incorporating nonlinear interactions, preserving the positivity constraint inherent in such systems. An isolated node is designated as the reference trajectory for synchronization convergence. Building upon this framework, two distinct control strategies are developed: an impulsive synchronization controller requiring full-state measurement, and an observer-based synchronization controller for scenarios with partial state observations. Through systematic integration of these control architectures, both network positivity and global asymptotic stability are theoretically ensured. The main contributions of the paper are as follows: (i) A novel impulsive synchronization controller and an observer-based synchronization controller are constructed, (ii) A tractable approach is given to obtain the gains of controller and observer, and (iii) Linear programming and copositive Lyapunov function are employed to analyze the synchronization and stability of positive complex dynamical networks. Finally, several illustrative examples are given to verify the effectiveness of the proposed theoretical results.
Fuzzy Cognitive Mapping (FCM), as a soft computing methodology, has emerged as a prominent research area in recent years. However, existing FCM construction methods exhibit significant limitations in weight learning, particularly when applied to long-term or complex time series prediction tasks, demonstrating insufficient prediction accuracy and limited capability for logical knowledge representation and reasoning. This paper proposes a novel approach for constructing high-order FCMs that extracts concepts from data using triangular membership functions and transforms the FCM learning problem into a constrained convex optimization problem, thereby effectively determining the weights of high-order FCMs. The proposed method is applied to large-scale time series prediction tasks. Ten diverse real-world time series datasets are employed to validate the predictive performance of the proposed approach. Experimental results demonstrate that the proposed time series numerical prediction method is effective, achieving superior prediction accuracy: it yields an average accuracy improvement of 55% compared to baseline methods and 25% compared to the latest methods. Additionally, it exhibits enhanced FCM representation capabilities and improved inference performance relative to existing approaches.
. Diffusion models, as a paradigm shift in artificial intelligence, have demonstrated outstanding capabilities in image processing and modeling, marking a significant breakthrough in knowledge discovery and visual content creation. However, achieving precise alignment between textual prompts and the semantic and visual attributes of generated images remains a key challenge. To address this, we propose a diffusion-based text-toimage generation framework - AMAC-T2I - that aims to improve fine-grained attribute control and text-image consistency under textual guidance. The framework consists of Module (ARM). ACM incorporates prior image information via the multimodal encoder Q-Former, enabling a more targeted and guided generation process. It also uses adversarial loss as supervision to dynamically correct attribute discrepancies. ARM combines Q-Former with a cross-attention mechanism to effectively extract correspondence features between text and image and leverages the UNet network to generate category masks and calculate attention loss. This design addresses attribute omissions and mapping errors, thereby enhancing the generalization and consistency of generated results. Extensive experiments conducted on two authoritative attribute-level datasets, T2I-CompBench and art methods in attribute alignment accuracy, generation controllability, and text-image alignment flexibility, highlighting its superior performance and broad application potential.
