This study focuses on the Neolithic walled towns of Fenghuangzui and Zoumaling in the middle Yangtze River valley,integrating archaeological findings,macro⁃observations of pottery⁃making,and chemical composition analysis of ceramics to examine the similarities and differences in vessel forms,raw materials (clay and tempering materials),forming techniques,surface decoration,and intended function. The results reveal that the pottery production technologies at both sites are highly similar,employing a combination of wheel⁃throwing and hand⁃forming,with consistent forming techniques and surface decorations. However,significant differences exist in raw material use and vessel innovation: Fenghuangzui potters demonstrate greater flexibility,producing more diverse vessel forms,while Zoumaling adheres to the pottery production and use traditions of the core area,reflecting stronger cultural conservatism. Both sites exhibit localized production,with household access to pottery unaffected by status differentiation. The study suggests that the core area enhances community cohesion through shared ideology and ritual activities,driving cultural expansion. During social complexity development,pottery not only met domestic functional needs but also served socio⁃political purposes (e.g.,ritual activities or power symbolism) through specific forms,decorations,or contextual uses,yet its production and distribution were not subject to regional⁃level control. This finding provides a new perspective on the dynamics of social complexity in the Neolithic middle Yangtze River region,highlighting the unique relationship between pottery production and social structure.
A batch of Five Dynasties ceramics,including celadon,white porcelain,and glazed architectural components,was excavated from the palace site of the Nanyue Kingdom in Guangzhou,Guangdong Province. This article presents a technological study of celadon,using energy dispersive X⁃ray fluorescence (ED⁃XRF) and portable X⁃ray fluorescence (pXRF) spectroscopy for analyzing the composition of the body and glaze. The results reveal that the celadon wares were produced at at least four different kiln sites: the Meixian Shuiche Kiln,characterized by low titanium content in the glaze,producing celadon with a lustrous,translucent glaze and fine crackle patterns; and the Gaoming Dagangshan Kiln,Changsha Kiln,and Yue Kiln,all producing celadon with higher titanium content. Additionally,the low⁃titanium wares are compositionally consistent with those found at the Emperors' Tombs of the Southern Han Kingdom. This research provides important data for the technological study of ancient ceramics in Guangdong Province.
The Tangchaodun site is situated in Qitai County,Xinjiang Uygur Autonomous Region,along a critical transportation route of the Land Silk Road. To reveal the manufacturing techniques and origins of low⁃temperature glazed wares unearthed from this site during the 12th⁃14th centuries CE,we analyzed their chemical composition and microstructure using X⁃ray fluorescence (XRF) spectroscopy and micro⁃Raman spectroscopy. The results indicate that these sherds can be divided into glazed pottery and fritware,with glaze types including lead glaze,lead⁃alkali glaze,and alkali glaze. Lead glazed pottery exhibits green (copper) and yellow (iron) colors. The pottery with a low⁃Al glaze type is coated with a slip containing pyrophyllite,likely sourced from West Asia,whereas the pottery with a high⁃Al type originates from north China. Turquoise lead⁃alkali glazed pottery shows a distinct West Asian style,colored by copper without the use of tin⁃based opacifiers or slip. The opacifying effect is achieved through Mie scattering caused by diopside,quartz,and bubbles in the glaze. Alkali glazed fritware features blue and black; the underglaze⁃black contains chromite,and its low manganese content suggests raw materials sourced from northeast Iran. The study provides important evidence for the trade networks and cultural exchanges of low⁃temperature glazed ceramics along the Land Silk Road during the 12th to 14th centuries CE.
A substantial number of pottery artifacts dating from the Song⁃Yuan to the Ming⁃Qing dynasties have been unearthed at the Songjingqiaonan site in Lumu,Suzhou. This study selected pottery and soil samples for analysis using a ceramic densitometer,X⁃ray fluorescence spectrometer,X⁃ray diffractometer,and thermomechanical analyzer (TMA),complemented by historical literature review,to scientifically investigate their basic properties,body composition,phase composition,and firing temperature. The findings indicated that these potteries were fired from iron⁃rich,easily fusible clay. During the Song⁃Yuan period,potteries were predominantly made from locally sourced materials,whereas those from the Ming⁃Qing period utilized mainly lacustrine alluvial sandy silt clay. The physical properties of the Ming⁃Qing era potteries were superior compared to those of the Song⁃Yuan period. These potteries featured firing temperatures ranging between 1025 °C to 1125 °C (±10 °C),thus classified as high⁃temperature potteries. Potteries from the Song⁃Yuan period were fired in an oxidizing atmosphere,while during the Ming⁃Qing period,the firing atmosphere for Golden bricks was reducing but the firing atmosphere for cricket pots followed an oxidation⁃first⁃then⁃reduction process. Additionally,the carbon infiltration technique was skillfully applied.
This study addresses the limitations of cross⁃group communication constraints,personalized goal absence,and unconscious interactions in task scheduling⁃oriented multi⁃agent structural collaborative communication within intelligent manufacturing. We propose a Goal⁃Oriented (GO) learnable multi⁃agent structural dynamic collaboration model (GOLSC) integrating Deep Q Network (DQN) with Dijkstra's algorithm,which enhances Learning Structural Communication (LSC) by incorporating autonomous communication awareness. The framework establishes a dynamic task scheduling model by pairing each machine with an agent,while a dedicated manager agent tracks workpiece states and monitors dynamic events. By implementing GOLSC⁃enhanced static allocation rules and dynamic job scheduling strategies for the machine agents to select unfinished workpieces,the model achieves coordinated optimization of collaboration efficiency and production responsiveness. As the scale of the agents continues to grow,the tardiness rate of our model decreases by 20%~70% compared to traditional communication⁃free models and 10%~40% compared to structural communication models,and the average bandwidth occupancy rate is reduced by 10%~15%,effectively addressing the production inefficiencies caused by the lack of adaptability to dynamic events and interaction among agents in conventional intelligent manufacturing workshops.
Programmability recovery is an effective method to ensure network quality of service after node failure in Software Defined Wide Area Network (SD⁃WAN). To address the problem of a single form of failed node in programmability recovery,a heuristic algorithm⁃based primary backup controller deployment method is proposed to ensure network programmability when switches and controllers fail. When deploying primary controllers,control path reliability in the domain is measured by control path density and strength to improve path programmability in the event of switch failure. When deploying backup controllers,a certain number of adjacent primary control domains are divided into multiple backup regions,and a backup controller is deployed in these regions to establish their mapping to switches. This aim is to maximize the recoverability of offline switches and offline flows to improve the programmability of offline flows in the event of controller failure. Further,a discrete Wolf Pack algorithm with redefined intelligent behaviors is presented to tackle the placement problem of the primary and backup controllers. Experiments are conducted on four real⁃world network topologies,and the results show that the proposed method can reduce the impact of failed control paths in the case of switch failure and maximize the recovery of offline flows in the case of controller failure.
Due to the non⁃linear and noisy nature of stock price sequences,stock price prediction has always been a challenging task. Many studies use decomposition algorithms to improve prediction accuracy,but these studies only focus on overcoming stock price nonlinearity and do not consider other price factors. To address the above issues,this paper proposes a stock price prediction method based on an improved empirical mode decomposition and A⁃LSTM hybrid neural network. This method introduces multiple data indicators and combines complementary set empirical mode decomposition algorithm with attention enhanced Long Short⁃Term Memory (LSTM). Firstly,this method utilizes the complementary set empirical mode decomposition method to decompose the original closing price of the stock,obtaining multiple Intrinsic Mode Functions (IMFs) and a trend term to reduce the nonlinearity of the stock price,while extracting multi⁃scale features of the IMF; secondly,the obtained IMF,trend term,as well as the highest,lowest,and closing prices are input into an attention enhanced LSTM to learn multiple stock influencing factors and mine their feature information; finally,the attention enhanced LSTM is utilized to learn long⁃term dependencies in features and dynamically adjust the weights of input features,highlighting key information,and outputting prediction results. The experimental results on two stock markets and four stock datasets show that the predictive performance of our research method is higher than that of the benchmark model,with good accuracy and stability,which can provide support for financial market analysis and investment decision⁃making.
To address the challenges of causal discovery in high⁃dimensional data,this paper proposes a significance⁃weighted divide⁃and⁃conquer causal discovery approach (SWCD).Traditional causal discovery methods suffer from high computational complexity,ambiguous differentiation of Markov equivalence classes,and crude conflict resolution mechanisms in high⁃dimensional scenarios.To overcome these limitations,this work integrates divide⁃and⁃conquer strategies with significance weighting to refine the causal discovery process.Specifically,the method achieves a synergistic optimization of efficiency and accuracy through a three⁃tiered design.In the partitioning phase,path significance values (PSV) and path importance scores (PIS) are defined to dynamically quantify the statistical reliability of causal paths.By combining topological features and adaptive partitioning strategies,the framework prioritizes retaining high⁃confidence causal chains to protect critical structures while dynamically optimizing decomposition paths.In the solving phase,the PC algorithm is enhanced using residual⁃based conditional independence testing (ReCIT),which distinguishes Markov equivalence classes by analyzing the independence of regression residuals.In the merging phase,a confidence score⁃driven conflict resolution mechanism is designed to resolve edge conflicts during subgraph merging,where edge reliability is quantified through confidence scores.Experimental results demonstrate that the proposed method significantly outperforms existing baseline approaches such as CPBG on high⁃dimensional datasets,achieving superior performance in efficiency,robustness,and interpretability.Future research can focus on optimizing significance quantification metrics,refining dynamic partitioning strategies,and exploring adaptability to nonlinear causal relationships.
Magnetic resonance imaging (MRI) is paramount for clinical diagnosis and treatment of brain tumours. However,extant methods primarily focus on feature extraction and fusion from MR images (image domain),with a paucity of research exploring different feature extraction approaches. This paper proposes an MRI brain tumour segmentation method integrating features from both the image domain and the K⁃space domain. The proposed method exploits the global properties of the K⁃space domain in MRI to achieve independent extraction of global features. The method consists of an image domain feature extraction module,a K⁃space domain feature extraction module,an adaptive affine fusion module,and a decoder module.First,MR images and K⁃space domain data are separately input to two feature extraction modules to extract local and global features.Then,the adaptive affine fusion module establishes an affine mechanism to effectively fuse the two types of features. Finally,the decoder,enhanced by deep supervision,generates the final segmentation mask by exploiting the fused feature information.The proposed method was evaluated on the BraTS public brain tumour dataset,and compared to other methods,it achieved improvements of 1.12%~2.47% in Dice coefficient and 17.5%~52.8% in HD95 metric. In addition,the proposed method demonstrated excellent performance in terms of computational complexity,thus suitability for clinical diagnostic applications.
Multi⁃label learning assigns multiple labels to data instances,but real⁃world data often suffers from missing labels,increasing model complexity and prediction bias. Existing methods recover missing labels using predefined label correlations,yet neglect the compatibility between original label spaces and correlation matrices,introducing noise and spurious dependencies. To address this,we propose DDLC(Data⁃Dependent Dynamic Label Correlation Learning)method. By preserving label correlations through manifold regularization,DDLC employs a dynamic mapping function to recover the missing labels while suppressing the noise interference in the output space,adaptively adjusting the label associations across scenarios. Experiments on benchmark datasets demonstrate DDLC's superior performance and generalization capability.
Latent factor models aim to learn implicit embeddings of users and items from historical behavior data,serving as a core technology in modern recommender systems. However,the lack of interpretability in implicit embeddings significantly limits the trustworthiness of recommendations. To this end,we propose a novel method called Prompt Ensemble⁃based Explainable Embedding for Review⁃aware Rating Regression (PE3R3),which jointly leverages textual reviews and numerical ratings to learn explicit embeddings with clear semantics,thereby enhancing the interpretability of recommendations. First,PE3R3 employs pre⁃trained language models with diverse prompt templates to extract meta codebooks with explicit semantics from textual reviews. Then,using numerical ratings as supervisory signals,PE3R3 represents users and items as linear combinations of multiple meta codes through a residual quantization mechanism,yielding explicit embeddings rich in semantics and enabling interpretable recommendations. PE3R3 is a plug⁃and⁃play method,allowing seamless integration with existing rating regression models. Experimental results show that PE3R3 yields an average improvement of 5% and a maximum improvement of 16% in predictive accuracy. In addition,both quantitative analysis and qualitative analysis demonstrate that incorporating PE3R3 effectively enhances the interpretability of recommendations.
With the rapid development of social media, the spread of fake news poses a serious threat to public opinion and social stability. Traditional methods for detecting fake news often focus on the fusion of text information and single scale image features,and most of these methods only focus on binary classification of news authenticity,lacking more detailed and accurate classification of the fake news. This paper proposes a multimodal fake news detection method (multi⁃scale and multi⁃task learning for multimodal fake news detection,MML⁃MFD) that integrates multi⁃scale and multi⁃task learning. This method introduces multi⁃scale techniques in the image feature extraction stage,which can capture rich semantic information of the images at different resolutions and overcome the limitations of single scale feature extraction in processing complex scenes. At the same time,the model introduces tasks of image text matching and image text similarity judgment,which can not only accurately identify the specific types of the fake news,but also improve the robustness and generalization ability of the model. In addition,through fine⁃grained image text matching and similarity judgment subtasks,the model can provide evident interpretative basis for the detection results,enhance users' trust in the detection model,and provide more targeted guidance for subsequent debunking work. The experimental results show that the MML⁃MFD model exhibits excellent performance on both the Twitter and Weibo datasets,validating its effectiveness in multimodal fake news detection.
With the continuous expansion of the Starlink system and the continuous improvement of its networking application performance,the communication based on the satellite constellation networks has once again attracted the attention of the academia and industry. A number of constellation systems with similar architecture and service groups have emerged at home and abroad. Research is needed to make good use of these satellite constellations for overall benefits. In view of the challenges of incomplete information sharing between constellation network clusters and unbalanced workload distribution among multi⁃agent satellites,this paper studies the access technology of hybrid satellite constellation networks based on cross⁃domain collaboration,designs a space⁃ground integrated hybrid constellation network structure based on inter⁃satellite links,and constructs a cross⁃domain interconnection service model of hybrid satellite constellation networks. This paper also studies the dynamic routing algorithms based on historical information and conducts the comparative analysis of different routing algorithms based on network simulations. Mathematical analysis and simulation results show that cross⁃doamin collaboration of hybrid constellations can shorten the average queuing time of users in each constellation in the system,improve the robustness of the hybrid satellite constellation networks,reduce the payload imbalance among the satellite nodes,and effectively promote the overall efficiency of the whole system with 30%.
With the continuous development of human space exploration and aerospace engineering,space missions exhibit an increasing variety of operations,a continuous growth in the total amount of data,and increasingly complex space networking trends.OTN technology,with its ultra⁃high bandwidth transmission capability,robust error resilience ability,and the benefits of a mature industrial chain advantages,has emerged as one of the viable inter⁃satellite laser⁃bearing solutions for future integrated air⁃space⁃ground networks.In high⁃speed motion scenarios,the relative motion between low⁃earth⁃orbit (LEO) satellites and between low⁃orbit satellites and geosynchronous orbit (GEO) satellites induces Doppler frequency shifts that change periodically over time.This causes link rate fluctuation effects,leading to a bit rate mismatch on the receiving side,which severely affects the communication efficiency of the integrated space⁃terrestrial networks.This study focuses on the link rate fluctuation degradation effects induced by laser links based on Optical Transport Network (OTN) transmission technology and proposes a general mechanism framework for link rate fluctuation.Based on this framework,a rate fluctuation suppression mechanism based on the overhead channel adjustment is proposed,and specific details of frame structure modifications are provided.Focusing on the key parameters in the framework,solutions are derived based on Walker constellation link simulation results and continuous⁃time mathematical models,demonstrating that the suppression mechanism can achieve a rate adjustment of 61 parts per million (ppm),which is sufficient to achieve the adjustment targets in both LEO⁃LEO and LEO⁃GEO scenarios.
This study employs a high⁃resolution coupled numerical simulation integrating the Weather Research and Forecasting Model (WRF) and the Hybrid Coordinate Ocean Model (HYCOM). By utilizing the best⁃track dataset of typhoon observations and float observation data from the Propagation of Intraseasonal Tropical Oscillations (PISTON) project,sensitivity experiments are conducted to evaluate the impacts of three microphysics parameterization schemes (Morrison,WSM6,and Thompson) on the track,intensity evolution,and upper⁃ocean responses of Super Typhoon Mangkhut (1822). The results demonstrate that the air⁃sea coupled model can reasonably simulate the typhoon's track and intensity characteristics,albeit with an overestimation of intensity during the initial development phase. The coupled model successfully reproduces the asymmetric spatial distribution of sea surface temperature (SST) and salinity (SSS) induced by the typhoon. However,compared to buoy observations,the model slightly overestimates the magnitudes of oceanic cooling and salinity increase. The choice of microphysics parameterization schemes exhibits discernible influences on the simulation of Mangkhut. The Morrison scheme yields lower errors in simulating the Minimum Sea Level Pressure (Min SLP) and Maximum Wind Speed at a Height of 10 m compared to the WSM6 and Thompson schemes. Nevertheless,the Morrison scheme overestimates surface cooling,leading to an underestimation of wind speed during the transition from the mature to the early decay phase. In terms of three⁃hour cumulative precipitation (R3),the Morrison scheme produces higher values than WSM6 in the early stage of Mangkhut but lower values in the middle to late stages,while showing minimal differences compared to the Thompson scheme. Although the microphysics parameterization schemes moderately affect the simulation of the ocean surface responses to the typhoon,their sensitivity remains limited. The thermohaline changes at the ocean surface are jointly influenced by typhoon intensity and translation speed. Both the magnitudes of SST decrease and SSS increase intensify with higher typhoon intensity but diminish with faster translation speed,indicating a correlated interplay between intensity and speed in modulating sea surface responses.