Abstract：In the context of computing-power and photovoltaic (PV) coordination, to address the issue that existing scenario generation methods struggle to capture the coordinated patterns of PV output response to computing load, a two-stage scenario generation method combining mixed-integer programming (MIP) and conditional diffusion models is proposed. First, the coupling mechanism between computing load and PV output is analyzed, and a MIP-based computing task scheduling model is constructed to generate joint PV-computing load scenarios. Second, using the MIP-generated scenarios as training data, a conditional diffusion model with U-Net as the backbone network is constructed, encoding temporal features, temperature, computing task levels, and electricity prices as conditional vectors to learn the high-dimensional joint distribution of PV-computing load, achieving controllable scenario generation. Results indicate that the proposed method can generate high-fidelity joint scenarios that provide reliable data support for distribution network planning.

Abstract：Difficult airway intubation is a critical risk in clinical anesthesia. This study proposes an automated, objective airway assessment framework named MMAAF (Multi-Modal Airway Assessment Framework) based on static mouth-opening images of patients. Using computer vision techniques, the framework automatically extracts six key anatomical features from the dataset of patient mouth-opening images. These features are combined with actual difficult intubation labels to construct a structured multi-modal prediction dataset, which is processed with a random oversampling technique. The study employs two integrated machine learning algorithms—random forest and gradient boosting decision trees—for model training and evaluation. Results show that the gradient boosting and random forest models achieve prediction accuracies of 92% and 93%, respectively, on an independent test set, demonstrating good assessment performance. Concurrently, a logistic regression model is trained within the MMAAF framework. Features are weighted according to their importance weights, and a personalized probability score for difficult intubation is calculated for each patient. This provides an intuitive quantitative reference based on multi-modal features to support clinical decision-making.

Abstract：In response to challenges in power-grid emergency supply assurance such as dispersed information, heterogeneous states, missing knowledge, and tightly coupled decision-making, this paper proposes an integrated hydrogen knowledge graph and collaborative decision-making framework for grid emergency supply assurance. Centered on the main line of “using hydrogen as the means and grid supply assurance as the objective”, the framework constructs a three-domain coupled knowledge representation architecture covering the hydrogen side, the grid side, and the dispatch side. At the knowledge completion layer, a hybrid completion strategy is developed by integrating R-GCN with symbolic rule reasoning. At the state perception layer, time-series sensor streams, equipment drawings, and operational data are fused to build a dynamically updatable multimodal knowledge graph. At the evolution layer, an entropy-based uncertainty active learning mechanism is introduced to support continuous knowledge iteration. Experimental results show that the proposed method improves NER F1 to 92.0%, RE F1 to 88.9%, and cross-sentence RE F1 to 65.4%, while achieving 58.2% Hits@10 in knowledge completion. It also outperforms multiple baseline methods in critical-load restoration rate, restoration time, and decision latency. The proposed framework provides a technical pathway that jointly supports knowledge representation, reasoning, and collaborative decision-making for hydrogen-enabled grid resilience enhancement and emergency power supply dispatch.

Abstract：In recent years, the global digital transformation has accelerated, the demand for computing power continues to increase, and the construction of computing power centers has accelerated, providing efficient computing resources for the development of the digital economy, but also bringing a large amount of energy consumption and carbon emissions. With the increasing global energy shortage and the growing attention to carbon emissions, green, low-carbon, and sustainable development have become industry consensus. Green computing power is a form of computing power that aims to minimize energy consumption and environmental impact, and provide maximum computing resources and services. It is an important form and direction for the future development of computing power. This article deeply analyzes the basic concepts and evaluation index system of green computing power, constructs a green computing power evaluation index model from the dimensions of energy conservation, carbon reduction, efficiency improvement, and empowerment, and proposes a green computing power implementation path to better guide the construction and evaluation of green computing power.

Abstract：Diffusion models show strong performance in images, speech, and scientific computing. Practical deployment, however, faces two bottlenecks: limited controllability at inference and high inference cost. Conditional alignment can fluctuate due to sampling randomness and time-step discretization. Multi-step sampling also increases latency. This survey focuses on Inference-Time Optimization (ITO). We organize recent work along two axes: controllability and efficiency. Without changing the pretrained backbone or the data, ITO uses constraint, scheduling, and selection strategies to improve conditional alignment and detail stability under few-step or one-step budgets, while reducing latency. We also outline transferable application scenarios. Finally, we present deployment-oriented prospects that aim to improve speed, coverage, and quality in a unified manner.

Abstract：Real 3D modeling is essential for building digital twins and smart rural infrastructure. To tackle geometric distortions and texture artifacts from moving objects in UAV oblique photography, this study proposes a lightweight framework combining enhanced small object detection with image inpainting. An improved YOLO11s model with a spatial attention and pyramid downsampling module boosts detection accuracy, while a Transformer-augmented inpainting algorithm ensures semantic and texture consistency in removed areas. On the VisDrone2019 dataset, the model achieves 46.4% mAP@0.5, an 8.0% gain over the baseline, with only 9.7M parameters. This approach enables automated removal of dynamic objects in 3D modeling applications.

Abstract：Electroencephalogram (EEG) signals objectively reflect genuine emotional states, making them a vital tool in emotion research. Significant progress has been made in developing emotion recognition algorithms based on EEG signals. However, the non-stationary nature and low signal-to-noise ratio of EEG signals, coupled with the complex spatiotemporal dependencies of emotional representations, pose challenges to model recognition performance. Traditional methods struggle to capture correlations in distant time series data across both temporal and spatial dimensions, thereby compromising emotion classification accuracy. To address these challenges, this study proposes a method that performs multiscale decomposition of raw EEG signals using a tunable Q-factor wavelet transform (TQWT). Multiscale principal component analysis (MSPCA) is then applied at each scale for denoising and dimensionality reduction. The decomposed multiband signals are constructed into a dynamic brain functional network, and a graph neural network (GNN) is introduced to uncover emotion-related connectivity features between brain regions. Finally, the time-frequency features extracted via TQWT are fused with the spatial relational features extracted by GNN to capture temporal and spatial characteristics for classification. Experimental results on the SEED dataset demonstrate that the feature-reconstructed data are effective for EEG emotion classification tasks, achieving an accuracy of 98.15% ± 0.63% for recognizing three emotions.

Abstract：The accurate measurement of silicon carbide epilayer thickness is the key to the fabrication of semiconductor devices. This paper conducts research based on the measured data of Problem B in the 2025 National Undergraduate Mathematical Modeling Competition. Firstly, Savitzky-Golay smoothing filtering, wavelength window function and Huber Loss robust loss function are used to preprocess high-frequency noise, baseline drift and outliers in the original data. Secondly, a dual-beam interference model is constructed. Combined with the Sellmeier dispersion equation, the quantitative relationship between thickness and related parameters is derived. High-precision inversion is achieved through initial estimation by Fast Fourier Transform (FFT), iterative optimization by nonlinear least squares and inverse variance weighted average, and its effectiveness is verified by back-injection test. Thirdly, a multi-beam interference model is established. Thickness inversion is realized based on the transfer matrix method, and a method of oscillation component extraction and baseline recovery correction is proposed to reduce interference.

Abstract：In order to study the application characteristics of three typical pressure scanning valve system in wind tunnel environments and provide technical support for related experiments, a systematic research and comparative analysis method was adopted.Wind tunnel tests were conducted based on the PSI 9116 electronic pressure scanning valve, DTC Initium multi-parameter data acquisition system, and Optimus intelligent measurement and control system. At the same time, a universal force and pressure measurement software platform for high-speed wind tunnels was developed, and a wind tunnel test performance evaluation system was established to verify the system accuracy. The quantitative performance comparison results of the three systems were obtained, providing a scientific basis for the selection of wind tunnel test equipment.

Abstract：This paper addresses the critical challenges in achieving Ultra Reliable and Low Latency Communications (URLLC) for Industrial Internet of Things (IIoT) within non-coherent massive Single-Input Multiple-Output (SIMO) systems operating over generally correlated channels. A novel constellation design algorithm grounded in the Kullback-Leibler (KL) distance between constellation points is proposed. The methodology begins with decorrelating the received signal via Karhunen-Loève Transform (KLT). Subsequently, under an average power constraint and leveraging high signal-to-noise ratio (SNR) asymptotic analysis, the optimization objective is to maximize the minimum KL distance. This non-convex problem is efficiently decomposed into two independent subproblems: optimizing the KL distance between non-zero constellation points, and optimizing the KL distance between zero and non-zero constellation points. This paper derives the geometric progression structure for optimal non-zero constellation points, and based on this, develops an efficient one-dimensional bisection algorithm to solve for all optimal constellations. Experimental results demonstrate that the proposed optimization method is applicable to arbitrary correlated channels, significantly enhances the Pairwise Error Probability (PEP) performance, and notably outperforms traditional benchmark schemes, thereby validating its effectiveness. The findings confirm that the proposed approach effectively improves system performance and exhibits greater generality, offering a practical solution for URLLC communication.

Abstract：Against the backdrop of large-scale integration of distributed energy resources, such as photovoltaic and wind power, into the power grid, traditional methods for calculating loop-closing currents face challenges of insufficient accuracy and difficulty in characterizing the uncertainty of loop-closing characteristics. To address this, this paper proposes an interval prediction model for loop-closing currents that considers uncertainty. Specifically, the research process includes the following key steps. First, given the numerous influencing factors and complex data structure in loop-closing current prediction, this paper adopts the Adaptive LASSO (ALASSO) regression method to screen influencing factors and construct a multivariate dataset suitable for loop-closing current prediction. Second, to enhance the interpretability of the loop-closing current prediction model, the study introduces the Extreme Gradient Boosting algorithm to evaluate the feature importance of the influencing factors selected by LASSO, clarifying the contribution of each factor to the prediction results. Then, to address the temporal characteristics of loop-closing currents, this paper proposes a prediction model based on an attention-mechanized Gated Recurrent Unit (GRU) for time-period prediction of loop-closing currents. This model dynamically allocates weights through the attention mechanism to capture key features in the time-series data of loop-closing currents, thereby improving prediction accuracy. Finally, to more accurately represent the uncertainty of loop-closing current predictions, the study employs the Bootstrap method to scientifically calculate the confidence interval, thereby quantifying the potential fluctuation range of the prediction results. Subsequently, the calculated confid

Abstract：Existing fire detection technologies mostly rely on smoke sensors or cameras, which are not only costly in terms of deployment and maintenance, but also face problems such as privacy leakage and limited recognition range, especially difficult to cover non-line-of-sight (NLoS) areas. In addition, traditional methods are generally unable to determine whether there are trapped people in the fire area, making it difficult to provide effective decision support for emergency rescue. To solve the above problems, this paper proposes a human-fire detection method based on channel state information (CSI), which uses the wall penetration and attenuation characteristics of Wi-Fi signals to achieve joint perception of human status and fire conditions in complex indoor environments. A subcarrier weighted fusion algorithm is proposed to compress high-dimensional CSI data, and a double-layer gated recurrent unit (GRU) network is constructed to extract time series features to achieve accurate recognition of human activities and fire status. Experimental results show that the accuracy of this method in detecting whether a fire has occurred and whether there are people to be rescued in a fire environment in non-line-of-sight scenarios is 94.93% and 94.53%, respectively. In addition, the experimental dataset developed in this paper has been made public for further exploration by relevant researchers. The dataset can be obtained at https://github.com/T-bjq/Wi-HFC-dataset.

Abstract：To address the problem of substantial noise introduced during the actual measurement of pulse signals due to environmental changes, subject motion, electromagnetic interference, and other factors, this paper proposes a denoising algorithm based on Variational Mode Decomposition (VMD) combined with an improved wavelet thresholding method. Using the Pearson correlation coefficient as the evaluation criterion, the optimal value of k is determined, and the signal is decomposed into k modes via VMD. Low-frequency components are selected from the resulting Intrinsic Mode Functions (IMFs), while the remaining IMFs containing useful information are denoised using the improved wavelet thresholding algorithm. Finally, the denoised IMFs are reconstructed with the low-frequency components to obtain the final denoised signal. Simulation experiments using synthetic data were conducted to compare the proposed method with several other denoising approaches, including VMD, VMD-SSA, VMD-SG, ICCEMDAN-WT, and MEEMD-WT. The results demonstrate that the proposed algorithm achieves superior performance in terms of Signal-to-Noise Ratio (SNR), Mean Squared Error (MSE), and Normalized Cross-Correlation Coefficient (NCCC). Validation on actual measured pulse signals further confirms that the proposed algorithm effectively removes noise while preserving the key features of the pulse signal.

Abstract：To address the challenges of complex backgrounds, occlusion, and illumination variations in Chinese herbal medicine image recognition, this paper constructs a large-scale dataset containing over 20 000 images of 76 herbal categories and applies various data augmentation techniques to enhance model generalization. Based on YOLO11n, an improved model named SEC-YOLO is proposed, which incorporates the C3k2_Star module, ECA_SR attention mechanism, and C2PSA_CGLU module to improve recognition performance under complex conditions. Furthermore, the feature fusion structure and detection head are optimized to enhance the detection of small and overlapping herbs. The improved model achieves a mean average precision (mAP) of 98.0%, with a weight size of 3.7 MB, 4.7 GFLOPs, and 1.80M parameters. Compared with YOLO11n, the weight size is reduced by 30.1%, FLOPs by 28.7%, and parameters by 31.0%. Both detection accuracy and speed meet the requirements of real-time detection. Experimental results demonstrate that SEC-YOLO achieves high accuracy and real-time recognition while maintaining lightweight characteristics, providing strong support for the automation of Chinese herbal medicine identification.

Abstract：In today's energy field, wind power generation is convenient and environmentally friendly, but its power generation efficiency is affected by weather conditions. Therefore, in order to optimize energy utilization, this paper proposes a wind power forecasting method based on bipartite echo state network. This model consists of three reservoirs. The first reserve pool extracts information features from the input data. Then, based on the different extracted features, the features are input side by side as input information into the latter two reserve pools to further fuse the data features. In order to ensure that the model can operate stably in the actual prediction, a sufficient stability criterion is given, and the structural parameters of the prediction model can be better selected. Furthermore, in order to obtain better prediction accuracy, the gradient descent algorithm is used to optimize the reservoir parameters of the bipartite echo state network. Finally, through a set of simulation numerical experiments, the validity and feasibility of the wind power prediction model based on bipartite echo state network are verified.

Abstract：In response to the problem of "over repair" or "disrepair" in traditional maintenance strategies for refining equipment, this paper constructs an Equipment Health Index (EHI) model based on multi-source data fusion. This model integrates multidimensional data such as defects, alarms, and repairs, and utilizes deep neural networks for nonlinear fusion and real-time evaluation. Engineering verification shows that this technology can accurately quantify the health status of equipment, with higher evaluation accuracy and early warning than traditional methods, providing effective support for intelligent operation and maintenance.

Abstract：With the rapid advancement of ultra-high-voltage direct current (UHVDC) transmission technology, accurate monitoring of ground-level composite electric fields has become increasingly important. Traditional measurement methods are easily affected by noise, reducing accuracy. To address this issue, a signal conditioning circuit based on parallel amplifier optimization and a signal processing method using Morlet wavelet decomposition are proposed. By enhancing the signal-to-noise ratio, the developed approach significantly improves the anti-interference capability and measurement precision of the field mill sensor. Experimental results demonstrate that the proposed method effectively suppresses noise, improves the signal-to-noise ratio, preserves signal details and dynamic response, and greatly enhances the accuracy of ground electric field measurements under UHVDC transmission lines.

Abstract：In response to the driving requirements of high-power brushless DC motor, a design of an intelligent power drive module using thick film hybrid integrated process is proposed. The module features a maximum operating voltage of 500 V and a maximum output current of 30 A. The module includes anti-shoot-through dead-time circuit design, input-output electrical isolation design, as well as an isolated DC/DC converter and high-side floating power supply circuit, etc. These features prevent damage to the preceding circuitry from the power stage. Furthermore, the module is encased in a fully hermetic metal package, ensuring reliability and enhancing its heat dissipation capability. In addition, its simple peripheral circuitry effectively shortens the development cycle of motor servo drive systems.

Abstract：With the advancement of the intelligentization and informatization of aircraft, especially the increasing application of drones, the requirements for the size and weight of airborne equipment have become increasingly stringent. The contradiction between the high cost, single functionality, and large volume and weight of airborne data storage equipment has become more severe. To address this issue, a domestically produced compact airborne recording module platform has been designed and developed. This platform adopts an integrated design approach, reducing the storage recording unit to the size of a standard electronic disk. It not only meets the capabilities of data storage, browsing, playback, and management but also incorporates data analysis functions. The system utilizes the domestically produced FMQL10S400 as the main processor to achieve a compact design. Through testing and simulation, it meets the requirements of small size, low power consumption, and low heat dissipation, making it suitable for airborne applications. It can fulfill the needs of various scenarios, including both manned and unmanned aircraft.

Abstract：At present, most of the analog circuits of integrated circuits use behavior-level models to verify the basic connectivity during digital verification. However, as the complexity and scale of circuits in integrated circuits in the world become higher and larger, there are more and more analog modules integrated by SoCs or MCUs, which are becoming more and more complex, and the complexity of mix-signal verification has increased dramatically. For most companies, the function of digital-to-analog interaction, timing, etc., can only wait for the return test to verify the correctness of the design. In this paper, a hybrid mixed-signal verification method based on wreal model is used to realize the verifiability of most functions of mixed-signal interaction, improve the comprehensiveness of verification, and verify the timing between mixed-signal circuits.

Abstract：This paper presents an FPGA-based bridge design between the SpaceWire interface and the RGMII interface. The host computer can inject data packets into the SpaceWire interface at a specified throughput, which are subsequently output from the FPGA’s RGMII interface in the form of Ethernet MAC-layer protocol data. Concurrently, the design supports receiving Ethernet data packets via the RGMII interface and transmitting them to the host computer over the SpaceWire bus. The proposed bridge is compatible with Ethernet data rates of 10/100/1000 Mb/s and supports both full-duplex and half-duplex modes. The design is implemented in VHDL and synthesized and placed-and-routed using Xilinx Vivado 2018.3, simulated with ALDEC Active-HDL 13 and Vivado 2018.3, and deployed on a Xilinx Kintex-7 FPGA device. Validation is conducted on a hardware platform comprising a processor system, FPGA board, and peripheral devices. The Ethernet MAC layer is realized using the Xilinx Tri-Mode Ethernet MAC IP core version 9.0, with the FPGA logic supporting parsing and responding to ICMPv6 Echo Request/Reply (PING) messages, as well as encapsulation and forwarding of UDP packets. Furthermore, a partial register Triple Modular Redundancy strategy is incorporated to mitigate the impact of single-event upsets in aerospace environments.

Abstract：This paper focuses on reasoning for Temporal Knowledge Graph (TKG) in periodic scenarios and proposes the EI-SAR model. The model integrates the temporal decomposition capability of the SARIMA model with Temporal Knowledge Graph Embedding (TKGE), decomposing the time series of entities and relations into non-seasonal trends, seasonal trends, and random residuals. Meanwhile, it designs a symmetric KL divergence scoring function to integrate residual characteristics, and constructs a hybrid loss optimization model that combines negative sampling loss with SARIMA residual penalty term.​ Experimental results show that the EI-SAR model exhibits excellent performance in link prediction on three public TKG datasets: YAGO11k, Wikidata12k, and ICEWS05-15. In the link prediction on the dedicated ecological industry dynamic dataset EITKG, the model significantly outperforms current mainstream TKGE models. Thus, the EI-SAR model can provide technical support for periodic scenario tasks such as resource recycling and ecological restoration.

Abstract：The core task of scene graph generation is to mine structured relationships between multiple objects in an image. Its essence is the semantic organization and association of visual data elements. Existing methods fail to fully filter and utilize object context information during data flow, resulting in limited robustness and accuracy in relationship prediction. To this end, this paper proposes a Message Context-Aware Network (MCAN). By constructing a coarse-to-fine data element fusion mechanism, it improves the quality and utilization of contextual information in scene graph generation. This model utilizes a Gated Recurrent Unit (GRU) to capture long-range dependencies between objects, implements a multi-head attention mechanism for fine-grained filtering of contextual data elements, and incorporates a residual fusion strategy to enhance the stability of visual representations. Experiments on the Visual Genome dataset demonstrate that MCAN achieves an average performance improvement of 2.1% across three subtasks, effectively validating its noise suppression and information enhancement capabilities in multimodal data flow. Ablation experiments further reveal the contribution of each module to data element fusion.