Objective Accurate identification of underwater cracks in dams is important for structural safety assessment and intelligent inspection. However, the generalization capability of intelligent models for dam underwater crack inspection is often limited by the scarcity of representative underwater crack samples. Moreover, domain differences among above-water crack images, generated underwater-style images, and real underwater dam crack images make it difficult for transfer learning and detection models to extract consistent crack features. To address these problems, this paper proposes a multi-source domain transfer learning method for underwater crack identification in dams, which combines CycleGAN-based image generation with an improved YOLOv8n detection model to enrich underwater crack samples and improve cross-domain identification performance.Methods Firstly, above-water crack images from campus buildings and bridge structures were assembled as multiple source domains, while field-acquired underwater dam crack images were used as the target domain. CycleGAN was then employed to translate the multi-source above-water crack images into underwater-style images through unpaired image translation. The translation process learned the color, illumination, and background characteristics of underwater environments while preserving crack morphology and structural information. Secondly, a cross-domain crack identification model was constructed based on YOLOv8n. An Asymmetric Cross-domain Attention Module (ACAM) was introduced at the backbone output to enhance directional crack features and suppress image-generation artifacts and background interference, while a Cross-domain Guided Fusion Module (CGFM) was incorporated into the neck network to strengthen the interaction between shallow spatial details and deep semantic features. Finally, the generated underwater-style samples were combined with real underwater dam crack images to train the improved YOLOv8n model, forming an integrated framework for multi-source sample generation, cross-domain feature enhancement, and underwater dam crack identification.Results and Discussions Experiments used 500 above-water crack images from two source domains and 500 real underwater dam crack images. For CycleGAN, the above-water and underwater image sets were each split into training and test subsets at an 8:2 ratio. For crack detection, the 500 real underwater images were further divided into 400 training images and 100 validation images, and an additional 91 independent underwater dam crack images were used to evaluate model generalization. The experiments were performed using an Intel Core i7-13700KF CPU and an NVIDIA GeForce RTX 4070 GPU with 12 GB of memory. Generated-image quality was evaluated using Fréchet Inception Distance (FID), Learned Perceptual Image Patch Similarity (LPIPS), and Structural Similarity Index Measure (SSIM), while detection performance was evaluated using mAP50, mAP50-95, GFLOPs, and FPS. The quality of the CycleGAN-generated images improved as the training duration increased from 50 to 150 epochs but deteriorated at 200 epochs. At 150 epochs, CycleGAN achieved the best generation performance, with an FID of 129.68, an LPIPS of 0.7636, and an SSIM of 0.5997, outperforming CUT and DCLGAN in image distribution consistency and crack structure preservation. Therefore, CycleGAN trained for 150 epochs was selected for sample generation. Among the 500 generated underwater-style images, 428 high-quality samples were retained, expanding the detection training set from 400 to 828 images. After incorporating the generated samples, the mAP50 values of YOLOv5n, YOLOv8n, YOLOv11n, YOLO26n, and RT-DETR-l increased by 2.63, 2.73, 3.01, 2.42, and 2.59 percentage points, respectively, demonstrating the applicability of the generated samples to different detection architectures. The proposed cross-domain identification model achieved mAP50 and mAP50-95 values of 88.19% and 56.74%, respectively, outperforming YOLOv5n, YOLOv8n, YOLOv11n, YOLO26n, Faster R-CNN, and RT-DETR-l. Compared with the baseline YOLOv8n model, mAP50 and mAP50-95 increased by 2.88 and 2.70 percentage points, respectively. The proposed model required 8.60 GFLOPs and achieved 213.63 FPS, indicating a balance between detection accuracy and computational efficiency. Ablation experiments quantified the contributions of ACAM and CGFM. Compared with the baseline YOLOv8n, ACAM increased mAP50 1.82 percentage points to 87.13%, while CGFM increased mAP50-95 by 2.09 percentage points to 56.13%. Combining ACAM and CGFM increased mAP50 and mAP50-95 by 2.88 and 2.70 percentage points to 88.19% and 56.74%, respectively. Grad-CAM++ visualizations showed that ACAM enhanced responses to crack bodies and extensions, CGFM improved the aggregation of crack features, and their combination focused more accurately on crack regions while suppressing background responses. Compared with CBAM, Coordinate Attention (CA), and SimAM, ACAM achieved the highest mAP50 and mAP50-95 values of 87.13% and 55.72%, respectively, while maintaining an inference speed of 275.99 FPS. Qualitative comparisons showed that the proposed model localized cracks more accurately and produced fewer missed and false detections under strong reflections, complex textures, low contrast, multiple cracks, and motion blur.Conclusions The proposed method integrates multi-source image translation with cross-domain feature learning to alleviate the scarcity of underwater dam crack samples and improve cross-domain identification performance. CycleGAN generated underwater-style samples while preserving crack morphology, and the generated samples consistently improved different detection models. ACAM enhanced directional crack features and suppressed image-generation interference, whereas CGFM strengthened the fusion of semantic and spatial information across domains. Their integration enabled the improved YOLOv8n model to achieve accurate and real-time underwater dam crack identification. The method can support automatic crack detection using optical images acquired by underwater inspection equipment, providing technical assistance for dam defect investigation and condition assessment.

Underground water conveyance tunnels (WCTs) are critical infrastructure in cross-basin water transfer projects, yet leakage remains a significant threat to their structural integrity and operational reliability. Existing inspection methods for detecting leakage in underground WCTs often suffer from limited accuracy, high costs, and poor adaptability to complex environments. To overcome these limitations, this study introduces infrared thermography (IRT) as a non-destructive technique for leakage inspection in underground WCTs and develops a novel theoretical framework based on inspection time windows (ITW), tailored to the environmental characteristics in underground WCTs. To thoroughly validate the method, a custom experimental platform simulating actual leakage scenarios is constructed, and extensive experiments are conducted to assess the feasibility, reliability and efficiency of the proposed method. Furthermore, relative thermal contrast (RTC) and time achievement rate (TAR) are defined as indicators to evaluate the visibility of thermal anomalies and the temporal feasibility of IRT deployment in specific tunnels. Finally, using the temperature monitoring data from an actual engineering project, the calculation process of ITWs that meet the detection condition and limitations of the proposed method are discussed. These findings provide practical insights for deploying automated cost-effective IRT leakage detection systems in large-scale WCTs.

The intelligent construction of water conservancy projects is a product of in-depth integration of digital technologies with conventional water conservancy projects.Its core value is reflected not only in the innovation of technical tools,but more importantly in promoting the upgrading of industry paradigms through the reconstruction of theoretical systems,thereby helping the water conservancy industry achieve its leap from"conventional construction"to"smart services."This study expounds on the basic concepts of intelligent construction of water conservancy projects,including its definition,core characteristics,and evolution.It examines the cognitive thinking,methodological logic,and practical framework of intelligent construction of water conservancy projects from the perspectives of systems philosophy and artificial intelligence(AI)philosophy.Moreover,the study constructs a theoretical system from the perspectives of complex systems theory,data science,systems engineering,and coordination of these three aspects.Specifically,the complex systems theory provides a cognitive paradigm for analyzing the complexity of water conservancy projects;principles of data science constitute the"engine"for intelligent decision-making in these projects;and systems engineering methods support the lifecycle collaboration of the projects.Furthermore,it proposes a technical system that covers elements such as intelligent perception,data fusion and analysis,intelligent decision-making and control,intelligent construction equipment,lifecycle collaboration,intelligent materials,and green construction,and clarifies the functions and logical relationships of these technologies.Looking forward,practical actions can be taken in the following aspects:expanding the depth and breadth of interdisciplinary integration,deepening the integration of data-driven and physical mechanisms,improving the adaptability of AI technology to complex working conditions,realizing the integration of design-construction-operation digital models,and promoting the standardization of intelligent construction technologies.These efforts will provide solid support for improving the national water security and sustainable development of China.

Due to the development of water conservancy projects in sulfate-rich environments in central and western China, concrete structures require improved durability against sulfate attack. Basalt fiber (BF) has been applied to enhance the mechanical performance and crack resistance of concrete because of its stable properties and eco-friendly production. This study investigated the deterioration behavior of basalt fiber reinforced concrete (BFRC) subjected to sulfate exposure. C30 BFRC specimens with BF volume fractions of 0%, 0.5%, 1.0%, and 1.5% were immersed in Na2SO4 solutions with concentrations of 0%, 5%, 10%, 15%, and 20%. The changes in macroscopic mechanical properties were evaluated, and the microstructural features were qualitatively examined using SEM. The results indicated that BF addition could mitigate performance degradation to a certain extent, and 1.0% BF exhibited the most favorable overall performance among the tested groups. The improvement is mainly attributed to fiber-induced crack bridging and stress transfer, whereas excessive fiber content may lead to agglomeration and increased porosity, thereby weakening the material. SEM observations suggested that the accumulation of sulfate-related products in pores and the interfacial transition zone may contribute to microcrack development and subsequent mechanical deterioration. Based on the Lemaitre strain equivalence hypothesis, a damage model was proposed to describe the degradation trend of BFRC under sulfate attack, and the model predictions were generally consistent with the experimental results. This work provides experimental evidence and a preliminary modeling framework for evaluating BFRC performance in sulfate environments.

Objective Dams are critical infrastructure ensuring national economic development and regional stability. Deformation is the most intuitive structural response that reflects the comprehensive service behavior of concrete dams under the coupling effects of external and internal loads. Reliable analysis and prediction of dam deformation are essential for scientifically diagnosing operating conditions and detecting early structural damage. Given the strong nonlinearity of dam deformation monitoring data, directly transplanting conventional data analysis methods cannot guarantee predictive accuracy. To address the susceptibility of existing models to data noise, as well as their deficiencies in prediction accuracy and generalization capability, this paper proposes a data- and knowledge-driven divide-and-conquer prediction model for dam deformation.Methods First, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) combined with the extreme wave extension method is employed to decompose the measured displacement data, effectively overcoming the end effect. Based on the frequency randomness and amplitude normality of noise, a criterion of denoising (COD) based on Euclidean distance is constructed. Through statistical analysis, the denoising threshold is scientifically determined to be 0.25 to achieve precise noise elimination. Subsequently, by deeply integrating the physical mechanisms of dam deformation with a divide-and-conquer (D&C) strategy, a comprehensive prediction framework is established. The effective separation of time-varying, water level, and temperature components is achieved using CEEMDAN and a K-means based isothermal clustering method, followed by the establishment of kernel extreme learning machine (KELM)-based prediction models for each component. Finally, the predicted outputs of all components are superimposed and reconstructed to yield the final dam deformation prediction, featuring both high precision and interpretability.Results and Discussions The effectiveness of the proposed model is thoroughly validated through simulation experiments and actual monitoring datasets from a high arch dam in southwestern China. In the denoising simulations, the proposed COD method significantly improves the signal-to-noise ratio (SNR) by 58.2% and 37.7% under single-power and multi-power composite noise interferences, respectively, demonstrating excellent denoising robustness. In the engineering case analysis, the D&C prediction model achieves extremely high accuracy on the test sets of four typical monitoring points. The maximum RMSE is controlled within 1.75 mm, and the correlation coefficient stabilizes above 0.99. In the long-term prediction performance evaluation, the single KELM model without the D&C strategy can only maintain acceptable results over a short prediction horizon, showing insufficient long-term forecasting capability. In contrast, the proposed model achieves stable and reliable predictions across different prediction steps, exhibiting robust long-term prediction capability and generalization. Furthermore, multi-model performance comparisons reveal that the prediction curve of the proposed model aligns best with the actual deformation sequence, showing superior capability in fitting local extrema and nonlinear fluctuations compared to other baseline models. Taking monitoring point PL13-1 as an example, the average RMSE of the proposed model in 10 independent repeated trials decreased by 41.7%, 40.5%, and 70.6% compared to the MLR, LSTM, and GRU models, respectively, highlighting its outstanding predictive performance.Conclusions The proposed data- and knowledge-driven divide-and-conquer prediction model provides a novel technical approach for the early warning of dam deformation anomalies. The COD denoising criterion, based on multi scale noise characteristics, effectively filters out interference noise while maximizing the retention of true high-frequency structural response information in the measured sequences. By deconstructing complex deformation sequences into sub-sequences with simpler modalities, the proposed D&C model framework strips away the mutual interference of multi-scale features, effectively reducing the difficulty of the overall prediction task and enhancing model interpretability. Compared to traditional statistical models and complex deep learning models, the KELM based D&C model not only significantly improves long-term prediction accuracy but also possesses superior computational stability and generalization capability, fulfilling the stringent stability requirements for early warning of dam deformation anomalies.

Efficient detection and timely clearance of underwater blockages are crucial for ensuring the safe operation of hydropower stations. However, traditional detection methods are constrained by low automation and high operational costs, while deep learning models struggle to generalize across diverse underwater blockages such as branches and stones. To address these limitations, this paper proposes an advanced framework for efficient underwater blockage detection termed EUBDNet, which is characterized by two designs, namely the Re-Calibration Feature Pyramid Network (RCFPN) and the Task-Dynamic Alignment Detection Head (TDADH). Extensive experiments were conducted on a self-built on-site dataset, and the results demonstrated the advantages of the proposed model, achieving a mAP50 of 75.46% with only 11.98 M parameters and an inference speed of 210.92 FPS. Based on this model, the ROV platform can rapidly identify underwater blockages and guide manipulators in clearing operations, providing an intelligent and efficient visualization solution for practical underwater clearance.

Utilizing recycled ceramic waste as aggregate in concrete not only alleviates the environmental burden of waste landfilling but also reduces the consumption of natural aggregates. This study integrates digital image correlation (DIC) and acoustic emission (AE) techniques to investigate the fracture performance and crack propagation of concrete incorporating recycled ceramic as coarse aggregate. The effects of aggregate size (3 mm, 6 mm, and 9 mm) and notch-to-depth ratio (0.1, 0.25, and 0.5) on the fracture process zone (FPZ) are analyzed in detail. The experimental results indicate that an increase in ceramic aggregate size leads to a more tortuous crack path and enhanced aggregate interlocking, along with increases in nominal fracture strength, fracture energy, and ligament transition length. The presence of boundary effect results in longer and wider FPZ for ceramic aggregate concrete with smaller notch-to-depth ratio. AE results show that crack mechanisms vary with the notch-to-depth ratio and aggregate size.

Unsupervised domain adaptation (UDA) can effectively address the two main drawbacks of transfer learning: the requirement of a large number of samples collected from different working conditions, and the inherent defects of convolutional neural networks (CNNs). In the realm of UDA, it is essential to leverage three types of information: class labels, domain specifications, and data organization. These components play a vital role in linking the source domain with the target domain. A technique aimed at identifying issues in rolling bearings is presented, employing an integration of CNN-KAN and GraphKAN structures to support the UDA methodology. A cohesive deep learning architecture is employed to represent the three types of information involved in UDA. The initial two types of information are represented through the roles of classifier and domain discriminator. To begin with, an architecture leveraging CNN-KAN is employed to extract features from the incoming signals. Following this, the features obtained from the CNN-KAN architecture are input into a specially developed graph creation layer that constructs instance graphs by analyzing the relationships among the structural characteristics found within the samples. In the following step, an innovative GraphKAN model is applied to illustrate the instance graphs, concurrently employing CORrelation ALignment (CORAL) loss to assess the structural discrepancies among instance graphs from different domains. Results from experiments conducted on two separate datasets demonstrate that the proposed framework surpasses alternative approaches and successfully recognizes transferable characteristics that are advantageous for domain adaptation.

The optimization scheduling model of the hydro–solar complementary system has the characteristics of high dimension, nonlinearity, strong constraints, etc., and it is difficult to solve. In view of this problem, this paper proposes an Improved Beluga Whale Optimization to solve the model. The local development strategy of the IBWO is replaced by the spiral movement of the whale algorithm to enhance the local development ability of the algorithm. In addition, an elimination mechanism is added after the whale fall stage of the original algorithm to increase the population diversity and improve the ability of the algorithm to jump out of the local optimum. This paper compares the solution effect of the IBWO algorithm with several well-known algorithms on 24 classic test functions and 29 CEC2017 test functions; the superior performance of the IBWO algorithm is verified. With the maximum power generation as the goal, the power generation scheduling model of the Beipan River hydro–solar complementary system is constructed and solved by the BWO algorithm, the IBWO algorithm, and the SCA algorithm, respectively. The results show that the IBWO algorithm can effectively improve the power generation of the hydro–solar complementary system and has a faster convergence speed than the BWO algorithm and the SCA algorithm, providing a new optimization tool for dealing with complex engineering optimization problems.

The physical and mechanical properties of rockfill materials significantly deteriorate when subjected to cyclic wetting-drying. Based on the three-dimensional scanning results of rockfill particles, a comprehensive template library featuring a wide range of particle sizes has been established. Considering the effects of wet-dry cycles on particle strength, effective modulus, and inter-particle friction, a particle crushing simulation scheme has been proposed, taking into account the degradation of rockfill particles caused by wet-dry cycles. On this basis, using the coupled FDM-DEM method, a series of triaxial compression tests have been conducted on rockfill specimens subjected to cyclic wetting and drying. The results reveal that the initial modulus and peak strength of the rockfill materials decrease nonlinearly with an increase in the number of wet-dry cycles (N). Wet-dry cycles exert an inhibitory effect on volumetric strain during the initial stages of shearing and dilation in later stages of loading. However, as N increases to a certain value, the impact of wet-dry cycles on the macroscopic deformation and strength of rockfill materials becomes less significant. As the number of wet-dry cycles increases, both the average coordination number and contact slip ratio within the specimen rise. Additionally, the deviatoric fabric inside the specimen exhibits a nonlinear decrease, indicating a gradual reduction in anisotropy. This suggests that wet-dry cycles have an inhibitory effect on the development of anisotropy within the specimen.

Deformation is a critical indicator for the safety control of high-arch dams, yet traditional statistical regression methods often exhibit poor predictive performance when applied to long-sequence time series data. In this study, we develop a robust predictive model for deformation behavior in high-arch dams by integrating signal dimensionality reduction with deep learning (DL)-based residual correction techniques. First, the fast Fourier transform is employed to decompose air and water temperature sequences, enabling the extraction of temperature cycle characteristics at the dam boundary. A data-driven statistical monitoring model for dam deformation, based on actual temperature data, is then proposed. Subsequently, an improved Bayesian Ridge regression model is used to construct the dam deformation monitoring framework. The residuals that traditional statistical methods fail to capture are input into an enhanced Long Short-Term Memory (LSTM) network to effectively learn the temporal characteristics of the sequence. A high-arch dam with a history of long-term service is used as a case study. Experimental results indicate that the data dimensionality reduction method effectively extracts relevant information from observed temperature data, reducing the number of input variables. Comparative evaluation experiments show that the proposed hybrid predictive model outperforms existing state-of-the-art benchmark algorithms in terms of predictive efficiency and accuracy. Additionally, this approach combines the interpretability of statistical regression methods with the powerful nonlinear modeling capabilities of DL-based models, achieving a synergistic effect.

Ensuring the safety of water networks is a research hotspot in the current water conservancy industry, and dams are an important part. However, over time, the dam is prone to varying degrees of aging and disease, most of which are structural cracks. If they cannot be discovered and repaired in time, the normal operation of the dam will be affected, and even catastrophic accidents such as dam failure will occur. However, complex backgrounds and blurred images can easily lead to misjudgments by machine vision detection models, and high-efficiency and accurate detection and evaluation technology are urgently needed. This paper combines the deep semantic segmentation network and the model hyperparameters optimization algorithm to propose a data-intelligent perception method of dam underwater cracks driven by knowledge coupling. Taking the underwater detection of a concrete face rockfill dam as an example, the effectiveness of the model is verified by using the underwater vehicle as the carrier. Experimental results indicate that the developed method achieves an intersection-union ratio of 0.9301, a precision rate of 0.9678, a precision rate of 0.9472, and a recall rate of 0.9577 in the test set. This shows that the constructed method has a high crack fine detection performance. In addition, the developed method has better segmentation performance in different complex underwater crack scenes, which further illustrates the high performance of the developed method.

Geopolymer concrete (GPC) is a potential alternative to ordinary Portland concrete (OPC) owing to its inherent benefits of low-carbon footprint and eco-friendliness. A comprehensive understanding of the deformation behavior of GPC under cyclic tensile loading is crucial to promote its engineering applications. This study first presented a detailed investigation into the nonlinear behavior of fly ash-slag based geopolymer concrete under cyclic tension. The proportions of slag in the precursors of GPC were designed at 20 %, 30 %, and 40 %, respectively, and the test results were compared with those of OPC. The experimental results indicate that the increase of slag content enhances the ultimate tensile strength, ultimate tensile strain and elastic modulus of GPC, but the elastic modulus of GPC is approximately 36 % lower than that of OPC with a similar strength level. The nonlinear characteristics of the stress-strain curves under cyclic tension for GPC and OPC are highly similar, including residual deformation, reloading strain, and stiffness degradation. Furthermore, the cyclic tensile unloading-reloading curve models with explicit functional forms suitable for engineering analysis are established for both GPC and OPC, which show good agreement with experimental data. Damage evaluation and damage localization based on acoustic emission technology broaden the understanding of the progressive failure process of geopolymer concrete under axial tension.

Rolling bearings are vital components in rotary machines where even minor faults can lead to the collapse of the entire mechanical system, resulting in significant economic losses. The condition monitoring and fault diagnosis of rolling bearings are crucial for safe operation of complex industrial systems. Traditional fault diagnosis methods mainly depend on expert experience, which leads to obvious deficiencies in diagnostic efficiency, especially in terms of accuracy. To address the issues of traditional bearing fault diagnosis relying on expert experience and low efficiency and accuracy of fault diagnosis, a bearing fault diagnosis model combining synchrosqueezed wavelet transform (SWT) and SE-ResNet is proposed. Firstly, the one-dimensional non-stationary bearing vibration signal is converted into a high-frequency two-dimensional time-frequency image through SWT, which serves as the input for the convolutional neural network; Secondly, an attention mechanism is introduced to construct an SE-ResNet model for extracting two-dimensional image features, which enhances the attention of important features and thus enhances the network's representation ability; Finally, the results of fault diagnosis are output through Softmax layer. To validate the capacity of the proposed model, experiments with the bearing dataset collected from Case Western Reserve University (CWRU) were conducted. The results show that the transfer learning accuracy of the CWRU dataset is 86.25% to 99%. This method has good feasibility, good convergence performance, and has high practical application value.

As the most direct and effective evaluation indicator, dam displacements are often utilized to reflect the behavior of dam structures under external environments and loads. The statistical regression dam monitoring model has the disadvantages of limited generalization ability and weak robustness in dealing with complex dam deformation behavior prediction problems. To solve these problems, this study combines artificial intelligence and deep learning (DL) methods to propose a high-precision hybrid forecasting model considering residual correction. Specifically, kernel principal component analysis is utilized to reduce the data dimension of high-dimensional prototypical monitoring data. The statistical regression algorithm and improved hydraulic-thermal-time are combined to develop the overall trend of dam deformation sequences. Then, the residual components that cannot be effectively explained by traditional statistical models are fed into DL-based algorithms to learn the underlying relationship. Specifically, a bidirectional long short-term memory neural with the self-attention mechanism network is used to learn the residual distribution law, and the random search optimization algorithm is used for determining the optimal parameters. A high arch in long-term service with massive prototypical monitoring data is introduced as the case study, and the three typical dam displacement monitoring points are utilized as research items. The experimental results show that the method can fully combine the interpretability of the traditional statistical regression method and the nonlinear modeling ability of the DL-based method, and has achieved good performance in the deformation prediction of high arch dams.

建立健全长江中下游水生态工程体系,是理顺保护与发展关系、促进构建和谐人水关系的重要途径.围绕荆江—洞庭湖、皖江—鄱阳湖、长江口、沿江湖区等重点区域水生态工程的内涵与特征,探讨了长江中下游水生态工程体系构建的路径与主要任务,认为应尽量恢复水生态系统的完整性和自我组织能力,使其能够在承受一定胁迫的情况下维持在自然演替的弹性区间内,避免出现物种灭绝等永久性损失;并提出长江中下游水生态工程体系建设应从开展水生态工程的论证制度改革、推进重大水生态工程建设及加强理论技术研究等方面着手进行.

Abstract In recent years, more and more attention has been paid to the refined operation of reservoirs. Precise operation of reservoirs requires accurate calculation of reservoir capacity. Although hydrodynamic models can accurately calculate dynamic storage capacity, they cannot be used in reservoir operation models due to computational efficiency. This study uses a hydrodynamic model based on ELM to guide the training of PSTA-TCN as a proxy model. The dynamic storage capacity of the reservoir is included in the calculation when generating training data. The experimental results show that the proxy model can be used as a method to accurately calculate the reservoir capacity during the refined reservoir operation.

Inland waterways in China flows from west to east, affects the economy of South and North China, communicates with the ocean.

The segregation in a binary mixture of particles with different sizes has been extensively investigated. However, there are few researches on the flow and deposit characteristics of binary mixtures of particles with different densities. The granular column collapse experiments of binary mixtures of steel and aluminum particles are conducted in the laboratory and simulated by using the discrete element method (DEM). Five granular mixtures with different centroid heights are prepared by adjusting the distribution of the two types of particles along the column height. The results indicate that the spatial distribution of particles with different densities significantly affects the flow and deposit characteristics of granular column collapse. During the collapse of granular mixture, there are two different motion patterns of granular flow: the first is that high-density particles and low-density particles move simultaneously, forming the front end of granular flow; the second is that the high-density particles squeeze the low-density particles, and the low-density particles move faster to form the front end. With the increase of the centroid height of granular column, the second mechanism gradually dominates the collapse and flow of granular mixture.