This PDF file contains the front matter associated with SPIE Proceedings Volume 13403, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

This article proposes a convolutional neural network training algorithm based on frequency domain transformation to address the issues of high computational complexity and complex training processes in convolutional neural networks. Through experimental verification, this algorithm can further improve the accuracy of neural networks compared to existing frequency domain transformation algorithms.

For the path planning problem with constraints, the traditional space-time Astar algorithm have problems such as dealing with fewer types of constraints and running slowly. In this paper, the space-time Astar algorithm is further improved to cope with different types of constraints and the search efficiency of the algorithm under constraints is improved by the improvement of the search space and the search strategy.

Submodular functions characterize the diminishing returns effect and are widely applied in many fields such as economics, combinatorial optimization, information theory, and machine learning. Hence, research on submodular functions attracts significant attention. In recent years, the scale of problems that need to be addressed has gradually increased with the advancement of computer science. Moreover, the functions that need to be optimized in real-world scenarios often exhibit properties similar to submodular functions but do not strictly conform to the definition of submodular functions. Therefore, designing fast algorithms for set functions that are close to submodular properties has become a research focus. For the problem of optimizing monotone non-submodular functions subject to matroid constraints subject to matroid constraints, with a Diminishing-Return (DR) ratio of γ , this paper presents the γ-MatroidContinuousGreedy Algorithm (γ-MCG Algorithm). This algorithm is a nearly-linear time approximation algorithm with an approximation ratio of (γ²(1 − 1/𝑒)² − O(ϵ)). Notably, it is the first known nearly-linear time algorithm for graph matroid constraints and partition matroid constraints.

This paper studies the power transmission network planning problem using a hybrid GRASP algorithm, which combines the advantages of the Greedy Randomized Adaptive Search Procedure (GRASP) and the Simulated Annealing (SA) algorithm. By incorporating the Metropolis sampling criterion of the SA algorithm, the hybrid GRASP algorithm can probabilistically accept inferior solutions during the local search phase, effectively enhancing the algorithm's ability to escape local optima. We validate the correctness and superiority of this hybrid algorithm through simulation experiments on an 18-node system. The experimental results show that the hybrid GRASP algorithm not only optimizes computational efficiency but also significantly enhances the algorithm's convergence performance. Additionally, this study explores key parameters affecting algorithm performance, such as the number of feasible solutions constructed, temperature regulation, and the setting of the Restricted Candidate List (RCL) percentage, providing more efficient and feasible optimization strategies for power transmission network planning. The results of this study have significant theoretical and practical value for guiding the optimization design and operation of actual power transmission networks.

To solve the problems of quality management and control in equipment maintenance operation management, such as the carry out of quality management is not in time, and the implementation is not thorough enough, and the effectiveness is not significant enough, the key factors of quality management and control are determined, assessed and analyzed. The numerical function between quality management and control method and equipment maintenance operation management factor is established based on neural-network, the parameters of input layer, hidden layer, and output layer of neural-network are trained and optimized based on genetic algorithm. Hence, the assess model on key factor of quality management and control in equipment maintenance is accomplished based on genetic-neural-network algorithm. Sample set is generated based on contribution rate and interaction coefficient, and then the assess model is tested and analyzed using the sample set, the simulation results show the effectiveness and efficiency of the assess model with genetic-neural-network algorithm.

Aiming at the shortage of accuracy and feature extraction ability of aerial image behavior detection algorithm, we designed an improved YOLOv9 fight behavior detection algorithm. WE use drones to take aerial photos and build datasets of violent behavior, and we add Multidimensional Collaborative Attention (MCA) to the object detection header to improve the detection accuracy, and introduce maximum feature pooling to strengthen the ability of network feature extraction for self-built datasets. For unevenly distributed data sets, Focaler-loss function is used. The results show that compared with the original network, the detection accuracy is improved from 90.6% to 92.9%, and the FPS reaches 56 frames per second, which meets the requirement of high precision real-time detection.

In response to the drawbacks of traditional genetic algorithms, such as premature convergence, low optimization efficiency, and inadequate precision, improvements have been made to traditional genetic algorithms from two aspects: calibration of the fitness function and diversification of the population. This has prevented traditional genetic algorithms from prematurely converging to local optimal solutions and broadened the optimization space. The improved genetic algorithm has been applied to architectural structural optimization design by establishing an optimization mathematical model with the objective of minimizing mass, addressing discrete variable structural optimization problems with stress and cross-sectional size constraints. Comparative optimization design results have been conducted between the improved genetic algorithm and standard genetic algorithms. The results indicate that the evolutionary generations of the improved genetic algorithm are fewer than those of the standard genetic algorithm, with significantly better convergence performance. This has enhanced the computational speed and optimization effectiveness of genetic algorithms in structural optimization applications.

When evaluating the degree of air pollution, the PM2.5 index is often regarded as an important assessment indicator. The ability to accurately predict the future PM2.5 air pollution index at a certain time based on past data has significant research implications in areas such as public health protection and adjustments to production and economic activities. Based on this, a network model combining LSTM and ridge regression was proposed for predicting the air pollution index. Initially, operations such as handling missing data, preprocessing, data scaling, and standardizing data steps were carried out on the dataset to ensure better convergence during training of the neural network. It was observed that with an increase in the number of training epochs, overfitting occurred, leading to an increase in mean squared error. The use of ridge regression to assist in predicting with the LSTM model yielded better results. Emphasis was placed on the feature values of the air pollution index, with the feature values of the LSTM training model being extracted and used as input for the ridge regression model, resulting in improvements in the prediction errors (MSE, MAE, and SMAPE). This confirmed the high level of fitting of the model on the designated test set and actual data. Furthermore, it was found during the experimental process that in the absence of other auxiliary feature values (such as temperature, humidity, and precipitation conditions), more accurate results could be obtained by using ridge regression to assist the LSTM model.

In the traditional civil aviation transportation industry, airlines have limited means of obtaining aircraft operational data, lack algorithm models that match the health status, and lack application of analysis of fault modes and impacts of key aircraft systems, as well as solutions to fault problems, which prevents a comprehensive understanding of the aircraft's health status and prediction of health trends. To reduce the operational risks of civil aircraft, this study analyzes the current research status of civil aircraft operational risks and proposes a single-aircraft risk assessment method that combines the Fault Mode and Effect Analysis (FMEA) with Fault Tree Analysis (FTA) for individual aircraft. For the fleet, the study first estimates the exposure frequency of risks using historical operational data of civil aircraft through maintenance planning; then, it establishes risk calculation models and an uncensored fleet risk analysis method based on the Weibull bathtub curve's random failure period and wear-out failure period, and uses actual examples to obtain the fleet's risk change trend and the faced risk values. By analyzing and summarizing the above methods, three improvement directions for reducing operational risks are proposed, effectively reducing the operational risks of civil aircraft.

The Fast Fourier Transform (FFT) is one of the most important algorithm in digital signal processing. In the paper, we aim to provide an application which performance can be comparable with cuFFT, an efficient parallel FFT algorithm on the CUDA has been developed. The implementation of our FFT has been descripted in detail, the algorithm performance on CUDA has been verified by numerical tests, and the tests also indicate that our FFT algorithm is effective and efficient.

With the rapid development of artificial intelligence and optimization algorithms, the Ant Colony Optimization (ACO) has become one of the effective tools for solving path planning problems. This paper first introduces the basic principles and key concepts of the ant colony algorithm, and then elaborates on its application in path planning problems in detail. Several comparative experiments are conducted to demonstrate the advantages and limitations of the ant colony algorithm in solving path planning problems. The performance of the ant colony algorithm is verified through experiments, and its potential in practical applications is discussed.

This article considers the point estimations and interval estimations for a generally inverse exponential distribution on the basis of the progressive first failure censoring. We derive the maximum likelihood estimators along with corresponding asymptotic confidence interval at first. Next, we compute the maximum likelihood through the utilization of the Expectation Maximization (EM) algorithm. Afterwards, Bayesian inference is applied using both symmetric and asymmetric loss functions in scenarios involving either informative or non-informative priors. The M-H algorithm is utilized to compute point estimation, which leads to the highest posterior density credible intervals simultaneously. Eventually, a numerical simulation test is executed for the sake of evaluating the qualities of estimations mentioned and authentic dataset is investigated.

The strategic synchronization of multiple reservoir operations for flood management is a vital non-engineering strategy that enhances the flood mitigation capabilities of these reservoirs. It can enhance reservoirs' flood control ability, and play a role in regulating flood peak, saving flood, alleviating or even avoiding flood disaster. This paper takes the three large reservoirs of Chushandian, Nanwan and Suyahu above Wangjiaba in the main stream of the Huaihe River as examples, takes the floods in July 2020 as the input, changes their previous independent and regularized operation mode, and aims to minimize the sum of the maximum flood control storage capacity occupied by the three reservoirs as the target for joint optimal operation of reservoir group flood control under the premise of ensuring downstream flood control safety. The corresponding model is established and solved by improved PSO. The results show that the improved PSO algorithm has higher search efficiency and convergence accuracy. Compared with regular scheduling, optimal scheduling can ensure the safety of the reservoir and downstream control stations in the basin, and the total storage capacity occupied by the three reservoirs can be reduced by 123 million m³.

This paper takes Guizhou Province in China as an example to predict the future Development trend in underdeveloped regions combined with the influencing factors of Research sample mobility. Based on XGBoost (Extreme Gradient Lifting Tree) algorithm, several sample flow prediction models are constructed by quantifying the theme feature vectors as explanatory variables and taking the number of Research sample flow rate as explanatory variables. The results show that with the improvement of science and technology standards in China, in the five years 2021-2025, the gap between the less developed regions and the more developed regions is still large, in the state of net outflow, and the outflow number is increasing year by year.

Decision tree ID3 algorithm is a classical data mining classification algorithm. This paper improves the algorithm from three aspects: introducing experience value in professional field, simplifying the calculation amount of attribute gain value and setting pruning threshold. Selecting some sample sets of College Students' football training, referring to the attribute weights given by sports experts according to experience, the decision tree can be formed quickly and reasonably. Compared with the original ID3 algorithm, the classification result is consistent, but the efficiency is improved obviously.

Heuristic algorithms have shown excellent solutions to combinatorial optimization problems. This paper delves into an improved heuristic algorithm and its practical application in solving combinatorial optimization related problems. It elaborates on the operation process of the algorithm and verifies its performance through carefully designed simulation experiments. Further in-depth analysis was conducted on the advantages and disadvantages of this algorithm to comprehensively understand its performance in practical problems. This improved heuristic algorithm not only has a solid theoretical foundation, but also demonstrates remarkable results in practical applications.

The conventional hyperchaos security encryption algorithm for power big data center users' personal privacy information mainly uses Lorenz average index pre-diction center for random encryption coding, which is vulnerable to the exponential divergence of initial parameters, resulting in low encryption performance indicators. Therefore, an effective power big data center users' personal privacy hyperchaos security encryption algorithm is proposed. That is, to encode and de-code the personal privacy information of users in the power big data center with hyperchaos security encryption, so as to generate an effective hyperchaos security encryption optimization algorithm for the personal privacy information of users in the power big data center. The experimental results show that the designed hyper-chaos security encryption algorithm for power big data center users' personal privacy information has high encryption efficiency, encryption overhead, encryption access delay, and low encryption CPU occupancy, which proves that the de-signed hyperchaos security encryption algorithm has good encryption effect, reliability, and certain application value. In order to improve the transmission quality of power big data, It has made certain contributions to reducing user security risks.

Although search recommendation algorithms have been widely used, they also face a large bottleneck: it is difficult to improve recommendation accuracy and reduce user cost-effectiveness. Leaderboards adopts the popular goods priority recommendation strategy, without considering the user behavior data in the decision-making process, as well as differentiation factors between users make recommendations less effective, which in turn affects the conversion rate. Therefore, this paper proposes a personalized recommendation algorithm framework based on the B2C list, on the one hand, by considering the consumer user's current behavioral data to carry out the development trend of predictive recommendation, on the other hand, by information such as the average price and quantity of each major category of goods purchased by the consumer to categorize users, and then combined with the user's preference for the cross-information on the commodity to carry out the differentiation of commodity recommendation, so as to formulate a popular commodities list. The proposed framework has been analyzed and validated through legal advice website data to ensure its feasibility, but further analysis and research is needed due to the large number of B2C industry categories and the complexity of the data.

Survival data demonstrate a wide range of potential applications in fields such as biomedicine and economics, commonly utilized for prognosticating the presence of illnesses and predicting disease progression, thereby assisting healthcare professionals in formulating personalized diagnoses and treatment plans. In this study, using the METABRIC dataset, we employ Cox's model and random survival forest model to identify a series of influential factors while considering patients' clinical and genetic information. Subsequently, we introduce a stacking integrated model to combine the predictive capabilities of multiple models, resulting in significant optimization of prediction performance compared to individual algorithms. This innovative approach not only enhances understanding of complex disease mechanisms but also provides robust theoretical foundations and empirical evidence for developing more accurate medical prediction models and optimizing patient treatment pathways.

Collaborative filtering recommendation system is one of the core technologies for product recommendations on ecommerce platforms. The method used to measure user similarity directly impacts the accuracy of the recommendation algorithm. Addressing deficiencies in similarity calculation within collaborative filtering recommendation algorithms, this paper proposes a comprehensive similarity measurement method based on the cloud model's "mutual membership degree - shape". This method overcomes the limitation of strict matching of object attributes, effectively utilizes cloud model parameters and data statistics to more accurately reflect user preferences, thus providing more precise recommendations to users. Experimental results demonstrate that this approach has significantly improved recommendation accuracy to a certain extent.

The technique of measuring blood pressure (BP) through photoplethysmographic (PPG) signals has been widely applied. But this category of methods faces challenges such as the complexity of measurement devices, cumbersome feature selection and inadequate performance. To address these issues, this paper proposes an innovative MSCBU-Net for BP estimation. The method employs an end-to-end solution to predict mean arterial pressure (MAP), diastolic blood pressure (DBP) and systolic blood pressure (SBP) using PPG signal and intermediate continuous arterial blood pressure (ABP). Based on the evaluation of samples from the UCI-BP database, the average absolute errors determined for SBP, DBP and MAP are 4.87mmHg, 2.77mmHg and 2.41mmHg, respectively, demonstrating high accuracy in BP measurement. Furthermore, the method complies with standards set by the Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS), indicating its reliability and practicality in clinical applications. In conclusion, the MSCBU-Net model presents an effective BP measurement method with broad prospects for application in the medical field.

With the development and widespread application of industrial information technology, greenhouse cultivation techniques have seen extensive use in agricultural production. Traditional manually managed greenhouses employ rudimentary methods, relying heavily on manual irrigation and temperature regulation. Effectively predicting the tomato greenhouse environment has emerged as a key research challenge. Accurate forecasting of future greenhouse conditions aids cultivators in efficient regulation, saving both time and labor management costs. Therefore, in this paper, to predict future environmental temperatures effectively, we propose a Time Multi Branch Attention Model (TMBAM). This model extracts strongly correlated factors of environmental temperature, captures the importance of various related features, and performs feature extraction for time series characteristics to predict environmental temperature and humidity. By comparing common algorithms such as LSTM, GRU, and other machine learning methods, our model demonstrates improved performance on evaluation metrics like MSE, MAE, and MAPE, showcasing excellent predictive capabilities.

This paper aims to extract keyframes and stage division of sports diving competition videos by advanced image processing and clustering algorithms to improve the efficiency of video analysis. Firstly, we cut the sports competition video into a series of images, store them in a folder, and perform Gaussian filtering and grayscale preprocessing. Then, a two-dimensional similarity matrix is constructed by comparing the Euclidean distance similarity of each frame with all other images in the cut order using a Siamese network. The positive characterization of the properties of the matrix is discussed in depth, which lays a theoretical foundation for the subsequent clustering analysis. An improved Jenks Natural Breaks clustering algorithm is used in the study to minimize intraclass variance and maximize inter-class separation. To optimize the clustering effect, this paper designs a new objective function, E, whose core idea is the exponent of the negative of the sum of interclass variance and intraclass variance. By comparing the values of E, we find the four-class clustering results that make the value of E more minor and thus identify the three keyframes in the video. These keyframes divide the video into four stages, effectively placing the change points between stages. Despite the interference of noisy points, the researcher further optimized the results through mean filtering processing. Ultimately, this approach improves video processing efficiency and lays the foundation for subsequent video evaluation and analysis through key frame extraction and stage division.

The multimodal data fusion algorithm for power marketing inspection based on knowledge graph is an innovative data processing method aimed at integrating power marketing inspection data from different data sources and various forms to improve inspection efficiency, accuracy, and intelligence level. Firstly, through data preprocessing, a new wavelet threshold function is used to achieve multimodal data denoising for power marketing inspection; Then, a knowledge graph in the field of electricity marketing inspection is constructed to represent and organize heterogeneous and multimodal data (text, images, videos, audio, etc.) in a unified form, connecting entities and relationships from different data sources to form a complete knowledge system. Further achieve deep data fusion and intelligent analysis. The experimental results show that this method can accurately integrate multimodal data of different types of power marketing inspections, improve the efficiency and accuracy of the inspection system, and provide strong support for the marketing inspection work of power enterprises.

Considering the challenges posed by the dung beetle optimization algorithm, which often gravitates towards local optimality and disproportionate global exploration and local development potentials, a fusion strategy driven dung beetle optimization algorithm (FSDDBO) is proposed. Firstly, the piecewise linear chaotic map (PWLCM) is used to initialize the population, so that the dung beetle population can better traverse the entire solution space; secondly, an improved spiral search strategy is added in the dung beetle breeding stage to accelerate the convergence speed and improve individual diversity; finally, a competitive reverse learning strategy is added in the dung beetle stealing stage to increase the randomness of the dung beetle and update the existing worldwide most unfavorable solution. In 11 benchmark function tests and comparisons with other algorithms, the method proposed in this research improves the algorithm's capacity to leap beyond its local optimum, ensuring strong robustness and precision in optimization, achieving good results.

Wordle, a popular daily puzzle offered by The New York Times, challenges users to guess a five-letter word within six attempts, with feedback provided through a color-coded system. This study explores the dynamic patterns of player engagement and result reporting on social media platforms such as Twitter, where users frequently post their game outcomes. The primary aim of our research is to understand and predict the variations in the distribution of these reports using a robust modeling approach that incorporates elements from epidemiological modeling. To address the fluctuations in reporting frequency and simulate the influences of game mechanics on player behavior, we employed the SIR (Susceptible, Infected, Recovered) model, commonly used to describe the spread of infectious diseases. This model helped us to conceptualize the spread of player engagement as analogous to an infectious disease, where interest peaks and then diminishes over time. Our adapted "Zombie Model" further refined this approach by including categories such as susceptible humans, infected zombies, and recovered individuals, along with a novel category of 'serum carriers' who exhibit mild infections and recover quickly, offering a nuanced understanding of the engagement lifecycle. Through rigorous data normalization and analysis, we observed patterns consistent with infectious disease spread—a sharp rise in activity followed by a gradual decline. The results from the model not only provided insights into how game design and external factors such as social media influence player interactions but also predicted future trends in player engagement. Our findings contribute to the broader understanding of digital media interaction and can inform future game design to enhance user engagement and retention.

With the increasing awareness of privacy protection, federated learning is increasingly applied to distributed training scenarios. However, in the process of executing federated learning on distributed clients, due to the non-independent and identical distribution between clients, the accuracy and convergence speed of the global model will decrease in the model aggregation process of ordinary federated learning. To solve this problem, a federated learning framework is proposed to improve the model aggregation method. During training, the server extracts the feature parameters of the local model, and performs entropy regularization on each layer of the local model to obtain the optimal transmission feature parameters. Finally, the optimal transmission and other federated learning global model feature parameters are generated by fusion. In the distributed training of the two data sets, the data distribution in four different cases was simulated for federated training comparison. The results show that compared with ordinary federated learning, the proposed algorithm significantly improves the model performance and the convergence speed of the global model, which can be used for distributed user training and commercial privacy-sensitive scenarios.

Aiming at the problems of poor visualization effect of recovered image, not rich in image details, too dark and unnatural image, etc., which exist in the dark channel a priori algorithm to deal with the large sky area, a dark channel a priori adaptive image defogging algorithm based on the RGB channel and the Retinex reflectance map is proposed. To address the issue where the restored image appears overly dark and lacks the clarity needed to reveal fine details, the pixel values of the three RGB channels of the original image are calculated and added to the original atmospheric light values to obtain an adaptive function to correct the brightness values. The reflection map in the Retinex algorithm is used to calculate the reflection component of the original image, and then the Gaussian filter is used to estimate the light component of the original image, and the transmittance is obtained by using these two components. The experimental results show that the proposed algorithm can effectively recover the detail information of the foggy image, the defogging is more thorough, the overall smoothness, the color brightness is better, the image is clear and natural, and the computational speed is ahead of the existing algorithms.

In this paper, an improved hybrid algorithm based on particle swarm optimization (PSO) and genetic algorithm (GA) is formed to solve the problem of "premature" and slow convergence. First of all, this paper improves the multi-task Collaborative assignment model (CMTAP), comprehensively considers the task completion, sequence, time, UAV type, quantity, energy consumption and other factors, and establishes a scientific and standardized assignment model that meets the task requirements. Secondly, for the problem of target allocation, this paper proposes an improved hybrid particle swarm optimization algorithm (PSO), which is based on particle swarm optimization (PSO), and uses genetic algorithm (GA) to increase the global search ability of the algorithm for the problem that the optimization of discrete nonlinear functions is very easy to fall into the local optimal solution. The mixed coding method of traditional coding and binary coding is used to solve the problem of uncertain number of unmanned aerial vehicles (UAVs). At the same time, the PSO algorithm is reduced by increasing the inertia weight, increasing the time-varying factor, and adopting the adaptive mutation probability, bit mutation and cross mutation methods for GA to increase the convergence speed and global search ability of the hybrid algorithm. Finally, the overall cost of the improved hybrid algorithm, hybrid algorithm and traditional algorithm is compared by simulation, and the effectiveness and accuracy of the proposed algorithm are verified.

To address the issue of low detection accuracy for surface defects in non-destructive apple inspection, a YOLOv8-AP algorithm for apple surface defect detection is proposed. Firstly, the Adown downsampling module is introduced into the feature extraction network to replace the traditional downsampling operation, thereby enhancing the model's feature extraction capabilities. Secondly, the FHLA structure is constructed in the neck network to integrate information from the backbone network and the feature pyramid, increasing the amount of feature information and enhancing the model's ability to detect multi-scale variations of apple surface defects. Lastly, LSD-DECD is proposed to improve the model's performance in localization and classification. Experimental results show that YOLOv8-AP achieves an mAP50 of 85.5% and an mAP50-90 of 52.9% for apple surface defect detection. Compared to the YOLOv8 algorithm, these values represent improvements of 2.6% and 1.4%, respectively. The improved YOLOv8-AP algorithm enhances the accuracy of the YOLOv8 algorithm, enabling precise recognition of apple surface defects.

This paper focuses on the channel coding technology in digital communication, especially convolutional code and turbo code. Convolutional codes improve error correction ability by introducing memory, while turbo codes combine interleaver and iterative decoding technology to further improve coding gain and decoding performance. This paper discusses the encoding and decoding algorithm of convolutional code in detail, including the optimal decoding method based on Viterbi algorithm, and introduces the basic composition and advantages of Turbo code. This paper aims to deeply analyze the common algorithms and optimization strategies of the two coding methods, and provide valuable reference for researchers and engineers in related fields.

The seaplane motion signal has strong nonlinearity, and it is difficult to make accurate predictions by traditional forecasting methods. For this reason, this paper proposes to use the combination of EMD and CNN-GRU to forecast two different motions of seaplane based on in-depth research on EMD and neural networks. Firstly, EMD is used to decompose the original non-smooth signal into smoother modal components, and then the CNN network, which has the strong ability to extract non-linear features, is combined with GRU, which has the function of long-term memory, and the CNN-GRU parallel neural network model is used to train each component, which improves the forecasting accuracy and reduces the complexity at the same time. Through comparison experiments with the CNN-GRU model, a single CNN model, and a single GRU model, the effectiveness of the proposed hybrid model in short-term forecasting is verified, which expands a new idea for the research of seaplane motion forecasting methods.

In recent years, low-light images have special properties and have been extensively studied. As deep learning technology advances, attention modules are increasingly employed in low-light image enhancement. We propose Discrete Wavelet Transform based Attention Network for Low-light Image Enhancement, our proposed Color Recovery Module initially recovers the color of low-light image, whereas the Color Adjustment Module focuses on useful information to adjust the color information. Finally, the overall details of the image are fine-tuned using 2D Discrete Wavelet Transform. The attention module is combined with the 2D Discrete Wavelet Transform to fully acquire the image information and ensure the information interaction in order to recover the image structure and details, and finally obtain a visually good and clean image. A series of experiments demonstrate that our network exceeds the state-of-the-art low-light image enhancement network to some extent.

In recent years, deep neural networks have made significant advancements, surpassing traditional models in computer vision. However, the vast number of parameters in deep models poses challenges such as increased network size, sluggish computation, high computational costs, and extensive storage requirements, thus hindering deployment on resource-constrained platforms like mobile devices. This paper focuses on compressing convolutional neural networks to address these challenges. This paper proposes a structured pruning method that integrates MobileNet features to prune the MobileNetV3-Large model, resulting in a more compact and efficient model. The approach involves sparse regularization training to obtain a sparser network model, which is followed up by structured pruning. This pruning is performed by leveraging the product of the sparsity values of convolutional layers and the scaling factors of batch normalization layers to identify and remove redundant filters. Experiments were conducted on the CIFAR-10 dataset. The experimental results demonstrate that the proposed compression method effectively reduces model parameters, decreases model size, minimizes runtime memory usage, and diminishes computational operations, while maintaining comparable performance levels despite the compression. Specifically, for MobileNetV3-Large, the model's parameter count is reduced by 40%, and computational operations are reduced by 35%, all while maintaining the same level of accuracy.

In order to understand the numerical simulation and prediction methods of ocean and wind waves based on ocean station data, a research on numerical simulation and prediction methods of ocean and wind waves based on ocean station data has been proposed. This paper first focuses on the statistical prediction of meaningful wave heights in waves, and conducts comparative analysis and research on typical prediction methods such as AR model, ANN model, and LSTM model. It establishes different meaningful wave height prediction models, verifies each model based on mixed wind and swell data in the East China Sea, and finally analyzes the prediction accuracy, providing the performance and error patterns of different prediction models. This study can provide theoretical reference for the optimization of wave forecasting models.

This study explores the influence of distribution line faults on forest fire risk, employing fuzzy theory and electric fire risk factor algorithms for comprehensive risk assessment. In this study, the effects of various fault types on forest fire risk were examined and analyzed by constructing a fuzzy rule base and an affiliation function. Through a detailed case study, this article demonstrates the application of these methods in assessing forest fire risk within a specific area. This study presents a systematic approach for assessing and managing the risk of forest fires resulting from electric power facilities.

This article analyzes the current research status of multi-agent reinforcement learning from the perspective of the training architecture and reward allocation of multi-agent reinforcement learning. From the analysis of the interaction process between fully cooperative multi-agent and environment, it can be seen that the focus is on solving the problem of reward allocation. However, common methods have problems such as low training efficiency, insufficient collaboration ability between multi-agent systems, and inability to adapt to large-scale scenarios. Therefore, this article proposes a collaborative multi-agent reinforcement learning method based on reward adaptive allocation, lists specific steps, and according to the analysis of experimental results, it can be concluded that this method can achieve higher and less volatile environmental rewards compared to MADDPG. Therefore, this method can explore samples more efficiently and improve training efficiency through the guidance of reward functions. In addition, the constraints of reward functions can enhance the stability of decision-making.

Artificial Intelligence (AI) has made remarkable advancements, surpassing human capabilities in various domains. This paper delves into cutting-edge AI technologies, analyzing significant models and techniques. However, current AI models have not fully captured the essence of human instincts, which we believe to be the core differentiator between humans and machines. We further explored simple human instinctive behaviors using a cellular automata (CA) model. While CA successfully emulated basic instincts, it fell short in replicating complex instinctive behaviors. We propose that future collaborations between computer science researchers and neuroscientists could offer the potential for gaining deeper insights into the underlying mechanisms of human instincts, developing advanced “artificial instinct” technology and producing the next generation of revolutionary machine systems.

Information extraction is a fundamental aspect of natural language understanding tasks. Previous works generally utilized a transformer-based architecture as a text encoder. However, gradient vanishing and attention dispersion tend to be inevitable when tacking document-level paragraph texts, which have negative effects on capturing global token relationships. To address these limitations, we propose CDM-IE, an information extraction approach specifically designed for lengthy text input. The CDM-IE is a hybrid CNN-Transformer architecture with dual-branch topology, respectively named a paragraph encoding branch and a dependence mining branch, which excels at learning comprehensive text representation by integrating both global context and local dependence. The two pathways converge at a dependence-guided attention module, which acts as a fusion bridge to feature alignment and synergy. Ablative experiment results on the Medical Entity Extraction (CCKS 2019) and Chinese Machine Reading Comprehension (CMRC 2018) datasets indicate that the proposed CDM-IE showcases improved performance and robustness on information extraction tasks, which provide a valuable solution for text modeling on long sequences.

Reversible information hiding in the encrypted domain has been utilized in numerous programs as it enables the embedding of secret information within the encrypted data while safeguarding the image content. Owing to its significance in the realm of privacy, this technology has garnered substantial attention and undergone significant development. Distinct from the traditional end-to-end reversible information hiding in the encrypted domain, the reversible information hiding scheme based on secret sharing is more applicable to applications in the cloud environment due to its multi-party security and fault tolerance. This paper undertakes a review of the development progress of Reversible information hiding based on secret sharing and conducts an analysis of the characteristics of existing schemes.

Aiming at the problem of high number of sensor nodes in the existing Data Link system and the inefficiency of message delivery, collection and processing within the sensor network, this paper proposed a message processing model based on improved TEEN routing protocol (MPMIT). Using the improved TEEN routing protocol in the sensor network, the message delivery path is determined by selecting cluster head nodes to form clusters. After the message is delivered to the sink node, the target trajectory data within the message is parsed, and the trajectory data is fused by the FCM clustering algorithm. The experimental results in the simulation environment showed that the target trajectory formed by the MPMIT model processing had a higher degree of coincidence with the actual trajectory, and the effect is better, which can effectively solve the message processing problem of sensor nodes in the Data Link system.

Medical image translation can provide auxiliary diagnosis for medical treatment, and existing methods are mostly based on GAN models. However, GAN models have limitations in learning image detail information, and therefore face many challenges. We mainly propose a new model for unpaired medical image translation based on wavelet transform and UNIT, WaveUNIT. This model decomposes the input image into low-frequency and high-frequency images through wavelet transform, and then inputs them into the low-frequency and high frequency generation parts of the generator, respectively. Finally, the obtained image is subjected to inverse wavelet transform to obtain a translated image. The cyclic network structure allows our model to be used for unpaired medical image translation tasks, and the additional high-frequency image loss constraint further enhances our model’s learning and extraction of high frequency image detail information. On this basis, we validated the effectiveness of WaveUNIT through experiments on different medical image translation datasets. We demonstrated the quality of generated images and the stability of the model from multiple dimensions, including quantitative analysis of performance indicators and qualitative analysis of visualization results. Finally, ablation experiments also demonstrated the superiority of the proposed method in unpaired medical image translation tasks.

Usually, mutation strategies and the corresponding parameter values play an important role in DE. Because different mutation strategies reveal different characteristic during the course of the evolution, combing different mutation strategies has been one of the research direction for DE improvement. In the work, a novel algorithm (eD-MPEDE) is introduced, which applies distance based parameter adaptation to produce competitive trial vectors. Furthermore, the candidate pool of eD-MPEDE contains three novel mutation strategies with different characteristics, which help to balance exploratory and convergence. Tests on CEC2017 show that eD-MPEDE is efficient to obtain promising solutions.

The permutation flowshop scheduling problems (PFSP) are present in various industrial sectors with broad engineering application background and theoretical value. To address this issue, an adaptive differential evolution algorithm with accompanying population (APADE) using differential evolution (DE) meta-heuristic optimization algorithm as the basic framework is proposed. The Largest-Order-Value (LOV) transform the newly generated vectors in the population into discrete ordered permutation vectors using the concept of discretization to make the APADE applicable to discrete coded. The initial set values of mutation factor and crossover rate are reset with the size of the PFSP to address the problem that improper settings of these two parameters can lead to the optimization search process falling into local optimums prematurely or the exploration phase being too extended. Each generation of successful control parameters are positioned in the historical memory, and this effective parameter adaptation technique guides the future control parameter values in the direction of advantageous evolution. In addition, a two-stage accompanying population is designed to slow down the rate of convergence and to enrich the diversity of the population at the later stages of the iteration. Finally, experiments comparing APADE with three other improved metaheuristics on 23 benchmark instances verify the performance of the APADE.

Named Entity Recognition (NER) is a critical task in Natural Language Processing (NLP). Compared to English NER, Chinese NER faces several issues due to differences in grammar between Chinese and English, which affect the performance of Chinese NER models. The current mainstream solution is to innovate lexical information into character-level models to achieve lexical enhancement. However, this process can introduce incorrect or irrelevant lexical information, leading to conflicts between words and affecting entity boundary segmentation and category annotation. To address this issue, this paper proposes a multi-feature fusion model based on Transformer. Building on the original model, which uses character vectors and word vectors as feature inputs, we add a new type of feature input: vectors obtained by re-weighting different words through adjusting lexical weights. This approach reduces the impact of incorrect lexical information, thereby enhancing model performance. Experiments on multiple datasets demonstrate the effectiveness of this method.

We study the decision-making problems encountered by two agents in parallel allocation and introduce the concept of risk attitudes. Risk attitudes reflect the likelihood of success in a competitive parallel allocation scenario. The probability of success is zero for risk-averse individuals, whereas it is certain for risk-seeking individuals. Individuals with different risk attitudes can accordingly adjust their strategies. When both agents are risk-averse, we provide the strategic characteristics that satisfy the Subgame Perfect Nash Equilibrium (SPNE) and design an algorithm to find the SPNE strategies. We prove the algorithm's correctness and resolve the selection conflict issues caused by even and odd items. When both agents are risk-seeking, we similarly provide the strategic characteristics that satisfy the SPNE and design an algorithm to find the SPNE strategies. Finally, we present a strategic analysis demonstrating the non-existence of relevant SPNE when one agent is risk-seeking and the other is risk-averse.

Sleep staging is a complex process that necessitates precise and robust analysis of polysomnographic signals. In this study, we introduce a novel approach aimed at effectively extracting crucial information from both the time and frequency domains of the signal. Our method is composed of three primary components: a multi-head spatiotemporal attention mechanism, and a module for multi-task learning. The spatiotemporal attention mechanism concurrently concentrates on various spatial locations and frequency ranges of the signal, enhancing the model’s attention allocation efficiency and precision. The multi-task learning-based module effectively leverages the correlation between sleep stages and can perform sleep staging and transition tasks simultaneously, thereby boosting the model’s learning capability and adaptability. The hyperparameter optimization method, which is based on the loss function, strikes a balance between the significance of the primary and secondary tasks to achieve optimal learning outcomes. By conducting a comprehensive analysis of various modules and hyperparameters, we examined the model’s performance and influencing factors, and validated the model’s rationality and robustness. The findings of this study offer significant support and direction for the ongoing advancement of sleep staging tasks.

In order to solve the problem that the traditional airborne radar multi-target tracking data association algorithm needs to obtain the prior information such as clutter density and target motion model in advance, a multitarget-tracking algorithm about data association based on Mamba network is constructed. The proposed algorithm combines the efficiency of selective state space and hardware-aware algorithms. This paper uses a data-driven approach to learn the mapping relationship between each target and measurement without providing various prior information to solve the problem of associating multiple targets with multiple measurements, and comprehensively considers the problems of missed detection and false alarm. Simulation results show that the optimal sub-pattern assignment (OSPA) distance error of the proposed algorithm is smaller, and it has better performance than the algorithm based on the Bi-LSTM network and the classic data association algorithm in the tracking experiment.

To address the challenges of occlusion in intelligent picking within unstructured environments, this paper proposes an occlusion perception algorithm utilizing knowledge distillation. A grape occlusion dataset was developed to support the study. The proposed method employs a lightweight MobileNetV3 student model, trained under the supervision of a high-accuracy ResNet50 teacher model, to achieve comparable performance with significantly fewer network parameters. Experimental results demonstrate that the knowledge-distilled student model operates with only 1.50% of the teacher model's operations while achieving an accuracy of 99.4%, representing a 2.4% improvement over the baseline model. These findings underscore the method's effectiveness in developing efficient, high-performance models suitable for deployment in resource-constrained environments.

Entity linking, the task of associating named entities mentioned in the text with their corresponding entries in a knowledge graph, addresses the inherent ambiguity in entity mentions. This process enhances semantic comprehension in both knowledge graph integration and natural language processing tasks. To address inefficiencies in processing lengthy texts and understanding complex contexts in existing models, this study proposes an innovative model, AT-LOM. The model integrates Longformer pre-training techniques and adversarial training mechanisms into text encoding, employs Extrapolatable Position Embedding (xPos) in the entity detection module and utilizes beam search algorithms in the linking module to rank text sequences generated by Long Short-Term Memory (LSTM). The AT-LOM model demonstrates significant performance enhancements on the standard AIDA-CoNLL dataset, achieving a 1.0% to 13.1% increase in Micro-F1 values compared to mainstream methods in recent years.

Word embedding is a widely used method for representing words in contemporary natural language processing, where words are depicted as low-dimensional dense vectors. However, the features of each dimension in the vector are difficult to interpret. Most existing interpretable word embedding methods enhance the interpretability of word embeddings through orthogonal or sparse transformations. Still, the semantics of each dimension need to be determined according to the knowledge base or manually assigned after transformation. Embedded topic model (ETM) can automatically acquire interpretable semantic space of words, but this topic space tends to be ambiguous and redundant. To address this issue, this paper proposes an Explainable Text Representation method with a Learnable and Explicit Semantic Space (ETRLESS), which autonomously learns an orthogonal explicit semantic space, allowing words and documents to be represented as interpretable vectors within this space. It obtains the document embedding representation through a BiLSTM model, initializes topic embeddings through pre-training ETM model, and imposes orthogonal constraints on the topic embedding to obtain a more interpretable topic semantic space. By using the reconstruction loss of documents, the document-topic distribution loss, and the orthogonal loss of topic embeddings as optimization objectives, it employs the backpropagation algorithm to learn an interpretable, orthogonal explicit topic semantic space and word representations based on this space. The results demonstrate that the embeddings generated by the ETRLESS model have clear semantic information in each dimension and maintain performance in downstream tasks.

Bare soil will cause the soil erosion and contribute to air pollution through the generation of dust, making timely and effective monitoring of bare soil an urgent requirement for environmental management. Though there have been some researches in bare soil extraction using high-resolution remote sensing images, great challenges still need to be solved, such as complex background interference and multi-scale problem. In this regard, the Hybrid Attention Network (HA-Net) is proposed for automatic extraction of bare soil from high-resolution remote sensing images, which includes the encoder and the decoder. In the encoder, HA-Net initially utilizes BoTNet50 for primary feature extraction, producing four-level features. The extracted highest-level features are then input into the constructed Spatial Information Perception Module (SIPM) and the Channel Information Enhancement Module (CIEM), to emphasize the spatial and channel dimensions of bare soil information adequately. During the decoder, the Semantic Restructuring-based Upsampling Module (SRUM) is proposed to leverage the semantic information of input features, and compensate for the loss of detailed information during the down-sampling in the encoder. Experiment is performed based on high-resolution remote sensing images from the China-Pakistan Resources Satellite 04A. The results show that HA-Net obviously outperforms several excellent semantic segmentation networks in bare soil extraction. The IoU and Recall of HA-Net in test scene can reach 81.4% and 87.4%, respectively, which demonstrates the excellent performance of HA-Net. It embodies the powerful ability of HA-Net for suppressing the interference from complex background and solving the multi-scale issue.

The anchor-based iterative deep graph representation learning (IDGL-anchor) method has the potential to yield excellent performance in node classification. Nevertheless, IDGL-anchor demonstrates simplicity when it comes to the selection of anchor points and overlooks the significance of these points. Additionally, throughout the learning process, it solely employs labeled nodes to direct the learning objectives, disregarding the information provided by unlabeled nodes and thereby squandering the valuable unlabeled information. To tackle these problems, a centrality-guided deep dynamic graph clustering (CGDDGC) method has been proposed. It enhances the strategy for choosing anchor points through the utilization of a centrality metric function. An unsupervised clustering module is incorporated to leverage unlabeled information for guiding the learning process. Simultaneously, the model is optimized by taking into account both labeled and unlabeled nodes, enhancing the accuracy and efficiency of node classification. Experiments conducted on five benchmark datasets demonstrate that our method surpasses IDGL-anchor and other state-of-the-art approaches.

The flight test data recorded in the airborne test are removed after the test and sent to the data unloading room for unloading. In order to improve the efficiency of flight test, it is necessary to achieve the level of quasi real-time processing, and get the processing results within 1~2 hours after the test. It is necessary to consider greatly shortening the time of data unloading and processing. By laser communication technology airborne record system test data cache to airborne laser communication terminal. Before aircraft landing, air-borne data is transferred to the ground laser communication terminal with the speed of 10Gbps, so that test data has been processed before aircraft landing, that save from waiting for landing to download the data a large segment time, significantly improve data produced speed and finally improve test efficiency.

Text-to-SQL is an important research direction in the field of artificial intelligence. This paper introduces a novel Text-to-SQL application framework based on large language models (LLMs), aiming to achieve more accurate Text-to-SQL generation. We propose three methods to effectively leverage the synergy between LLMs and Text-to-SQL prompts. Firstly, we employ prompt words in JSON format that include database query examples. Through interaction with LLMs, accurate SQL statements can be generated. Secondly, we utilize the Chain of Thought (CoT) prompt word method, prompting the model with query tables, fields, and conditions to generate SQL statements that comply with grammar rules. Finally, we enhance Text-to-SQL inference accuracy by querying an SQL knowledge base. We con- duct experiments on the Spider benchmark, which consists of 20 databases and 1034 query statements. The experimental results demonstrate that, under the single-domain database condition, the CreateTable+SelectCol+Normalized method achieves a Text- to-SQL accuracy of 77.9 with a 16-shot setting. This represents a 1.4% improvement compared to the Codex baseline.

This paper proposes a defense framework against jailbreak attacks that exploit multi-language and multi-intent inputs. Research indicates these attacks are effective primarily due to two reasons: (1) LLMs may incorrectly capture key points and semantics in low-resource language inputs, generating malicious content; (2) multiple requests within a single input can cause attention flickering, leading to inadequate capture of implicit requests and incorrect responses. The proposed defense framework requires no additional training and works by mapping multi-language inputs to high-resource languages, guiding the model to think multiple times, decompose intents, and reflect. Experimental results show significant effectiveness of this framework in defending against attacks.

Traditional graphite ore production typically relies on carbon-sulfur analyzers to determine the carbon grade of graphite ore. This method, however, is cumbersome and lacks timeliness. To address these issues, we propose TRA-FFNet, a graphite carbon grade image recognition and classification model that integrates multi-level features from ResNet and Transformer. Initially, ResNet-50 is employed as the backbone feature extraction network, with model parameters initialized through transfer learning to accelerate convergence. Subsequently, a Transformer module based on spatial feature compression fusion is designed to capture the weight relation-ships among different channels of the graphite ore feature map. Finally, a multi-level feature fusion module is incorporated at the network's terminal position to enhance the joint learning of both global and local features in graphite ore images. Experimental results show that the pro-posed model achieves an accuracy of 93.473% and an F1 score of 94.023% on our self-constructed dataset containing 19 classes of graphite ore, outperforming classic models such as MobileNetv2, EfficientNet-b3, ConvNext, RepVgg, and Swin Transformer. The proposed model achieves high-precision recognition of graphite ore grade in an end-to-end manner, offering a valuable detection solution for smart mining operations.

Artificial Intelligence (AI)-based Internet of Things (IoT) applications benefit greatly from advanced deep learning models. However, the increasing complexity and resource requirements of deep learning models pose challenges to computational efficiency and deployment on resource-constrained devices. To address these challenges, the paper proposes a merging technique to efficiently merge multiple single-task deep learning models into a unified multi-task model. The proposed method merges models in a serial fashion, potentially replacing various layers with a smaller deep neural sub-network to connect the layers of the previous model and the latter model's layers. This method improves computational efficiency, and can create larger models to perform the functions of the original models with minimal training overhead. The paper designs and implements a merging model that supports end-to-end compressed sensing (CS) sampling and object detection, and the experimental results verify the efficacy of the proposed method in generating fine-tuned multi-task deep neural models with minimal training time and resource costs, making it a cost - efficient solution for edge-cloud collaborative inference.

Target recognition in SAR images is a key issue in remote sensing image processing, which is widely used in many fields. Traditional recognition methods face many challenges due to the complexity of SAR images. In this paper, we propose an improved method based on GoogLeNet and improved YOLOV5 algorithm to cope with the limitations of MSTAR dataset. In this paper, GoogLeNet is introduced to improve the performance of target recognition in complex backgrounds, and its unique module improves the efficiency while reducing the parameters, which is especially suitable for handling small targets and complex textures in SAR images. Meanwhile, YOLOV5 is improved to enhance the detection accuracy and reduce the computational complexity. The effectiveness and superiority of the method is verified through experiments on the MSTAR dataset.

Infrared technology plays a crucial role in various fields. However, defocus blur occurs in infrared images due to improper focusing, resulting in undesirable blurring effects. In recent years, deep learning-based methods have achieved remarkable success in image restoration. However, the spatial variability of defocus blur, combined with the low resolution and lack of textural details in infrared images, still presents significant challenges for deblurring. In this paper, we propose a CNN-based neural network that employs dynamic large separate kernel convolutions to adapt to real-world defocus blur patterns and to effectively extract blur features. Furthermore, we introduce an encoder-decoder feature fusion module that incorporates edge attention, spatial attention, and channel attention to enhance the network’s focus on edges while selectively processing relevant information, thereby improving the network’s deblurring performance. Experimental results demonstrate that our method outperforms recent advanced methods in handling defocus blur in infrared images. Additionally, ablation studies are provided to show the effectiveness of the proposed modules.

Satellite images play a pivotal role in meteorological studies and environmental monitoring, providing crucial data that contributes to our understanding of atmospheric conditions, weather forecasts, and climate patterns. One of the significant challenges in the analysis of satellite images is the accurate segmentation of clouds, which is imperative for refining the data used in various meteorological applications. This paper introduces an innovative approach to cloud segmentation by harnessing the capabilities of Deep Residual Learning combined with the UNet architecture. Our method focuses on optimizing the segmentation process by employing residual connections that facilitate the training of deep networks and enhance the flow of information through the network, allowing the model to learn hierarchical features effectively. The integration of the UNet architecture further assists in capturing intricate details and enables precise localization, resulting in improved segmentation accuracy. We present comprehensive experiments that demonstrate the efficacy of our approach in providing superior cloud segmentation performance, underscoring its potential as a robust tool for enhancing the quality and reliability of satellite image analysis for meteorological and environmental research. Through this contribution, we aim to propel the advancement of cloud segmentation techniques, facilitating improved data accuracy for a broader spectrum of geospatial and atmospheric studies.

Slurm is a resource management and job scheduling system widely used in the field of parallel computing. Kubernetes is an open source container orchestration platform widely used in cloud native and AI fields. With the development of technologies such as parallel computing, AI, and large-scale data processing, the demand for computing resources in business scenarios has become more complex and diverse. In order to better adapt to business needs and give full play to the advantages of both in different fields, this paper proposes a fusion scheduling solution based on Slurm and Kubernetes. The solution is mainly aimed at two scenarios: partitioned deployment and hybrid deployment. The fusion scheduling between the two is realized by developing the heterogeneous resource manager Unify. Dynamic node management function is provided for partitioned deployment, and unified resource view and unified scheduling function are provided for hybrid deployment. Application results show that the solution can effectively solve the problem of dynamic node division under partitioned deployment and the resource scheduling conflict problem of Slurm and Kubernetes under hybrid deployment. Through this solution, both can be applied to more complex demand scenarios and improve the overall resource utilization of the cluster.

Chaotic image encryption technology has become a key method for safeguarding sensitive image data and has garnered significant attention and research recently. To address the issues of limited key space and low encryption efficiency in conventional image encryption methods, this paper introduces a novel chaotic image encryption algorithm. This algorithm leverages Chen and Logistic-Tent chaotic mappings along with dynamic DNA encoding rules. Initially, the SHA-256 algorithm is employed to produce a 256-bit hash sequence, which serves as the initial key for the chaotic mappings. The chaotic sequences generated by these mappings are then used to rearrange the rows and columns of image pixels, while dynamically chosen DNA encoding rules are applied to encode each pixel. Following this, the DNA-encoded image matrix is XORed with the chaotic sequence matrix to produce a scrambled image matrix. In the final step, the encoded ciphertext image is decoded using DNA, and minor plaintext alterations are diffused across the entire ciphertext through multi-level diffusion processes. This approach enhances the complexity and security of the encryption method. Experimental results indicate that the proposed algorithm effectively withstands various types of attacks while maintaining high encryption efficiency, demonstrating promising application potential and practical value.

Compared with traditional computing devices, IoT devices are often smaller in size, resulting in weaker battery power, memory, and computing power. Additionally, IoT devices have a higher frequency of data exchange and require high-speed computation on limited platforms. Traditional cryptographic algorithms are unable to handle this task, and lightweight cryptographic algorithms have become the optimal choice for IoT devices. This paper takes a thorough sight into lightweight block cipher development, elaborates on six structures of the algorithms, and analyzes their indicators based on different implementation methods. Finally, the attack research status and implementation performance on both soft/hardware platform of each algorithm are reviewed, and the efficiency of their software and hardware implementation is compared.

In this study, we enhance emotion classification for Chinese text by modifying the NEZHA model and evaluating various architectures. We developed and evaluated six distinct classifiers. Using the SMP2020-EWECT and ChnSentiCorp datasets, we assessed the models based on accuracy, F1 score, and loss. The DeepFeatureFusionClassifier and DeepFeatureRegClassifier emerged as the most effective models, with the DeepFeatureFusionClassifier achieving the highest performance on the ChnSentiCorp dataset and the DeepFeatureRegClassifier excelling on the SMP2020- EWECT dataset. The study highlights the effectiveness of advanced model architectures and multi-fold cross-validation on enhancing emotion classification accuracy and robustness.

Dunhuang culture is a treasure house of human history and culture. As an integral part of Dunhuang culture, Dunhuang dance is famous for its unique national characteristics and artistic characteristics. In the form of Dunhuang dance, instrumental music and props are used to greatly enrich the stage expression, forming a unique dance form of "holding music while dancing, dancing with instruments". However, the phenomenon of homogenization of instrumental props occurs frequently, which hinders the overall presentation effect and beauty of Dunhuang dance. The development based on the generative diffusion model provides new ideas for the design of Dunhuang dance instrumental props. First, the role played by musical instruments and props in the presentation of Dunhuang dance art was sorted out, and the Dunhuang dance instrumental props were digitally collected and processed, and a special data set was constructed. Secondly, based on the recognition of related instruments, a diffusion model instrumental prop image generation that can be used in the Chinese context is trained. Experimental results show that the instrumental props generated by this method not only conform to realistic logic but also have strong applicability.

The field of aerospace control software has accumulated a large amount of reusable software assets in the long-term software development process. Developers have a demand to learn and retrieve code related knowledge from different sources and types during the software development process. Software knowledge graph can represent and apply multisource heterogeneous software knowledge. The raw data for building a code knowledge graph is very extensive, including functional descriptions of functions, modules, and projects from source code data, as well as requirement specifications, code, test cases, and so on. The entities and relationships in multi-source heterogeneous data have also become more complex. To address these issues, this article constructs a knowledge graph for aerospace control software and provides software knowledge retrieval based on the software knowledge graph. The main works are designing a software knowledge graph construction framework, proposing principles and methods for extracting software knowledge entities for different types of knowledge resources, proposing a heterogeneous data knowledge fusion method, implementing a software knowledge retrieval mechanism and presenting retrieval results in a graphical visualization format. Based on the above work, a software knowledge graph construction tool for aerospace control software projects is designed and implemented. The example proves that the constructed software knowledge graph can better assist software developers in retrieving and applying software knowledge.

In recent years, the number of coral reefs has decreased sharply, and the natural ecological climate is facing serious threats. Combined with the commonality of the public's cognition of eco-climate issues and the dilemma of communication, this paper uses technologies such as Touchdesigner, AIGC and Arduino to realize the development and design of interactive devices, integrates technical means into artistic innovation, and displays screen images of various visual forms of corals through external hardware configurations (various sensors, development boards, radar, Kinect, leapmotion, etc.) and screen images showing various visual forms of corals. Popularize the knowledge of corals and ecological climate to the audience, so that the audience can have a deep connection and resonance with the ecology, and inspire the audience to become a more responsible global citizen.

In the surge of the digital era, the role of educational information systems is increasingly prominent, serving as essential tools to enhance teaching quality and management efficiency. This paper addresses the limitations of traditional educational information systems in handling complex, multi-source data, particularly the challenges in data integration and information correlation, proposing an innovative design approach. Through deep learning technologies, notably a multi-source information correlation method based on convolutional neural networks (CNNs), this study aims to develop a system capable of effectively processing multi-source and multi-granular data within the educational domain. The system design encompasses the entire process from data collection and information correlation to system construction, striving to provide the educational sector with an efficient, intelligent information processing platform. Experimental results demonstrate significant advantages of the system in terms of educational information accuracy and processing speed, promising innovative changes for educational management and pedagogical research.

With the continuous development of Low Earth Orbit (LEO) satellite constellations, their application in the Global Navigation Satellite System (GNSS) has been increasingly emphasized. Minimizing the Geometric Dilution of Precision (GDOP), a key metric for evaluating positioning configurations, along with ensuring a sufficient number of positioning satellites, presents new opportunities and challenges for the selection of GNSS positioning satellites. While additional observation satellites can lower the GDOP value, an excessive number can increase the computational burden on receivers. The recent surge in the number of LEO satellites enhances the potential for GDOP minimization, bringing theoretical optimal spatial distributions within closer reach. In this paper, we address the challenge of GDOP minimization under observation-constrained conditions through a geometric analysis. We first develop a set of nonlinear algebraic equations to determine the conditions for minimized GDOP. These conditions are articulated by segregating satellites into high and low selection regions based on a predefined ratio, ensuring maximal satisfaction of various equation sets within each region. An intelligent optimization algorithm is then crafted to meet this geometric condition, with the concurrent fulfillment of all equation sets as its primary goal. This strategy successfully realizes a GDOP minimization algorithm that is resilient to occlusions. Finally, we assess the positioning performance under a range of occlusion scenarios, as well as with varying numbers and configurations of satellites. The findings offer valuable insights for the enhancement of future satellite positioning systems within expansive LEO satellite constellations.

Subgrid-scale (SGS) stress modeling based on filtered variables is one of the crucial scientific challenges in large-eddy simulation. With the rapid development of machine learning technologies in recent years, data-driven turbulence modeling methods have gained its popularity. In this study, an SGS stress model based on artificial neural network (ANN), with strain-rate tensor and modified Leonard tensor as inputs, is developed for incompressible isotropic homogeneous turbulence. The proposed ANN model demonstrates a substantial enhancement in the prediction of the SGS stress. Also, the ANN model could provide better predictions of turbulence statistics, as compared to the traditional models. It is suggested that the ANN methods exhibit obvious advantages and considerable potentials for the development of turbulence models with high accuracy.

Due to the disconnection between arithmetic demand and supply, and encountering the limitations of cloud blocking in the upgrading process, in order to solve the challenges of computational complexity and accuracy, as well as the problems of data security and privacy protection, we propose the research of parallel optimization and acceleration technology for new energy power generation large-scale model computation based on localized supercomputing platform. This study enhances efficient parallel algorithms in the prediction process by parallelizing the processing and improving the time complexity of the algorithms, dividing the large-scale computational task into multiple subtasks and executing these subtasks simultaneously, establishing regular performance monitoring and optimization plans, and periodically evaluating the performance of the system. After the optimization measures are implemented, performance testing and validation are carried out, and the experiments show that the computational parallel optimization acceleration of new energy generation capacity prediction is realized on the supercomputing platform. The optimization effect meets the expectation and achieves 100% of the set performance index.

This study aims to improve the accuracy and reliability of the initial phase angle measurement of power frequency line voltages. By adopting a new structure of power frequency initial phase angle high-voltage signal collector, a cloud platform-based data resource maintenance system, and the construction of base station networks, it achieves accurate verification and management of the initial phase angle of the power supply network. The experimental scheme design includes the selection of experimental subjects, experimental research, analysis, and pilot practices, to comprehensively and systematically advance the research and practical work of the project. In terms of results application and promotion, through technology transfer and cooperation, industry standard setting, professional training and promotional conferences, and the publication and promotion of scientific research results, the research findings will be effectively applied in practice, bringing direct and indirect benefits and promoting the modernization of the power industry. Finally, the conclusion summarizes the achievements and issues of the research work and proposes future research directions and prospects.

The purpose of this study is to develop an autonomous and controllable voltage real-time calculation system based on Flink framework to improve the efficiency and accuracy of power equipment monitoring and management. The preprocessed data results are input to the least square support vector machine, and the Lagrange multiplier is introduced to construct a prediction model for the voltage prediction of the substation, thus completing the early warning of the low voltage substation. The experimental results show that the prediction error of the model is less than 1.5%, the delay time is 67ms, and the resource utilization ratio is 78%. This shows that the voltage real-time calculation system based on Flink framework can effectively improve the efficiency and accuracy of power equipment monitoring and management, and has important application value in the operation of power system.

The maritime industry is on the cusp of a revolution with the advent of Unmanned Surface Vehicles (USVs), which are redefining the paradigm of autonomous maritime operations. These sophisticated vessels are transitioning from human-operated systems to autonomous entities, courtesy of advancements in robotics and artificial intelligence (AI). This paper examines the role of embodied intelligence, the integration of AI into the operational fabric of USVs, in enhancing the capabilities of these vessels. It delves into the current state of USV applications, the challenges encountered in their deployment, and the prospective future developments that promise to further augment their functionality. Key areas such as intelligent navigation, collision avoidance, swarm behavior, and decision-making systems are explored to underscore the potential of USVs in transforming maritime tasks. The paper also addresses the hurdles impeding USV integration into the maritime domain, including navigational uncertainties, energy constraints, and the need for standardized regulatory frameworks. Looking forward, it is posited that embodied intelligence will be instrumental in propelling USVs to new heights of autonomy, efficiency, and operational scope, thereby shaping the future of maritime operations.

Signature verification efficiency is one of the key issues affecting blockchain performance. This paper designs a blockchain signature acceleration engine based on FPGA (Field Programmable Gate Array). The signature algorithm adopts SM2 elliptic curve public key cryptography with higher efficiency and security. The FPGA optimization design of SM2 algorithm and User Datagram Protocol (UDP) is also carried out. The experiments show that the IO control and area power consumption are further optimized, the response time is shortened, the network speed is significantly improved, parallel processing can be realized, and the power consumption is lower, and the performance is significantly improved compared with that achieved on the CPU.

Power work ticket is an indispensable working document in the electric power industry, the current safety measures in the power work ticket are mainly filled out manually by the practitioners based on their experience, which lacks consistency and has the risk of omission, in order to reduce the dependence on the front-line practitioners, this paper proposes a model based on the knowledge graph and the large language model of the safety measures generation. Firstly, based on the knowledge graph of work tickets, similar work tickets are found and preliminary safety measures are generated according to the rules, and then relevant safety specifications are queried based on the semantic similarity, and the model inputs, preliminary safety measures, and relevant safety specifications are inputted into the large language model together to get the complete safety measures. From the experimental results, it can be seen that this method outperforms other models in terms of expert assessment and the accuracy of security measure generation, and the security measures generated by the model can meet the needs of invoicing, ensuring the accuracy and efficiency of filling out work tickets.

Affected by load balancing and synchronization and mutual exclusion problems among threads, domestic supercomputing platforms are prone to reduced parallel efficiency. In this study, MPI combined with OpenMP hybrid parallel technology is adopted. First, the data is divided into multiple subtasks according to the time period, and a set of parallel processes is created using the MPI library function, with each process corresponding to a processor. Second, the MPI environment is initialized and communication channels and synchronization mechanisms are established. At the same time, the distributed memory system is utilized for cross-node communication and data exchange. Finally, by reasonably dividing tasks, data and communication, combined with appropriate thread management and synchronization mechanisms, efficient hybrid parallel computing can be achieved. The experimental results show that the use of MPI combined with OpenMP hybrid parallel technology can not only accelerate the computation speed, but also have good scalability.

With the rapid development of Internet technology, network traffic anomaly detection has become an indispensable part of network security, which is of great significance for preventing network attacks and maintaining data security. This article focuses on the in-depth research of XGBoost based network traffic anomaly detection technology, aiming to improve the accuracy and efficiency of network traffic anomaly detection through an efficient gradient boosting decision tree algorithm called XGBoost under the integrated learning framework. XGBoost stands out in multiple machine learning competitions with its outstanding performance. Its advantages in processing high-dimensional sparse data, providing parallel computing power, and model interpretability make it an ideal candidate for solving network traffic anomaly detection problems. In the experimental section, we selected real network traffic datasets for verification and demonstrated the superiority of XGBoost model in anomaly detection tasks through comparative experiments, especially in handling nonlinear relationships and capturing complex patterns. In addition, through in-depth analysis of the importance of features, key factors affecting abnormal network traffic were revealed, providing a basis for the formulation of security strategies.

In conventional single image super-resolution (SISR) models utilizing deep learning, information is generally transmitted in a feedforward fashion. Nonetheless, these models have not fully exploited the feedback mechanism. Drawing inspiration from the feedback process in human vision, we introduce a Multiscale Feature Extraction Feedback Network (MFEFN), which utilizes high-level information to enhance low-level details. In particular, we employ the principles of Recurrent Neural Networks (RNNs) to reintroduce part of the output back into the model as input, thereby optimizing the input feature representations. We design a multiscale feature extraction channel attention feedback module to handle the feedback connections and generate rich detail texture features. Additionally, to address the issue of limited receptive fields and singular convolutional kernels in models, we propose a multiscale feature extraction module to capture detail texture features at different scales. To further enhance the model's cross-channel learning capability, we incorporate a channel attention mechanism. Comprehensive experimental results showcase the proposed MFEFN's superiority over current state-of-the- art methods.

The rise in network security awareness has led to an exponential increase in the use of encrypted traffic. While encryption protects user privacy, it also presents significant challenges for network security detection. This paper introduces a novel encrypted traffic classification method combining Convolutional Neural Networks (CNN) and multi-layer Bidirectional Gated Recurrent Unit networks (BiGRU). The CNN extracts detailed and global information from traffic data frames, while the temporal attention mechanism in BiGRU captures the temporal relationships between frames. Experimental results indicate that the proposed CNN + BiGRU model achieves over 97.5% accuracy on test sets, outperforming existing deep learning models in both accuracy and F1 score.

To address the issue of low intrusion detection accuracy due to incomplete feature extraction by a single model, an SSCBLM model detection method is proposed. This method combines Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) to extract and integrate features. SE modules and self-attention mechanisms are incorporated into the CNN and Bi-LSTM, respectively, to enhance the model's ability to extract important features. Finally, a Multilayer Perceptron (MLP) model is used as a classifier for multi-classification. Additionally, class merging, RandomUnderSampler, and Synthetic Minority Over-sampling Technique (SMOTE) are employed to address the class imbalance issue. The model is trained on the CIC-IDS2017 dataset, and experimental results show that the model achieves an accuracy of 99.94%, effectively detecting abnormal traffic.

Aiming at the problems of frequent occurrence of domestic operating system robustness problems, insufficient test coverage and low defect detection rate at this stage, we carry out the research on operating system robustness testing and propose a method of constructing operating system robustness test cases based on FMEA (Failure Mode and Effect Analysis) technology, which applies the FMEA technology to software robustness testing. In this paper, we design FMEA analysis guidelines using the requirements of relevant standards in the testing field, and guide testers to analyze the potential failure modes in the relevant documents of the operating system according to the guidelines, and carry out the failure mode and impact analysis to further construct the set of robustness test cases. Finally, through experimental verification, this method gives full play to the advantages of FMEA in the field of robustness testing, and the number of defects found has been greatly improved, which effectively improves the quality of the robustness test cases and further improves the robustness of the domestic operating system, and it has strong engineering application value.

Due to occlusion and other problems during point cloud scanning, the collected point clouds are often incomplete. Therefore, reconstructing a complete point cloud based on a partial point cloud with missing information is of great significance in practical work. In this paper, we propose a point cloud completion method (PointVQDM) based on the VQ-Diffusion model, which uses a discrete diffusion model to model in the latent space for shape reconstruction. Specifically, we obtain the corresponding vector combination based on the partial point cloud, and use the point cloud quantization network to decode it to obtain a complete point cloud. Experimental results on multiple datasets show that our PointVQDM outperforms the most advanced completion network. Moreover, thanks to our feature fusion method, we achieve diverse and high-quality generation results.

As computer systems become increasingly complex, deep learning methods for rapidly analyzing and pinpointing anomalies in system logs are gaining widespread application to ensure smooth system operation. Addressing the issues of many existing methods that require substantial amounts of labeled data and insufficient utilization of temporal features in logs, we propose a semi-supervised log anomaly detection method. Initially, log templates are extracted using the Drain template parsing technique. Subsequently, BERT is employed to extract deep semantic feature vectors from logs and to derive time feature vectors. Unsupervised clustering algorithms are then used to estimate labels for unlabeled samples, tackling the problem of insufficiently annotated data in practical log anomaly detection scenarios. Finally, anomaly detection is achieved using a Attn-based Bi-LSTM model. Experimental results on two datasets, HDFS and BGL, demonstrate that our proposed method achieves notable improvements in terms of accuracy and recall, thereby validating the effectiveness of our work.

Construction projects typically consist of multiple processes. Therefore, process allocation needs to be conducted before project begins. Traditional process allocation requires the construction party to have extensive experience. However, with the rapid development of artificial intelligence, AI-based automated process allocation for construction projects has become possible. In this paper, we identify that the problem of process allocation in construction projects is essentially a partial multi-label recognition problem. To address this issue, we propose an automated process allocation method for construction projects based on partial multi-label recognition. This method contains a label feature decoupling module and a missing label inference module. First, we use a cross-attention mechanism to construct the label feature decoupling module and extract label features from project descriptions. Then, the missing label inference module employs graph convolution networks to capture label correlations and infer missing labels from existing partial labels. We validated the effectiveness of our method by using a private dataset of power construction projects.

As a new type of device discovered in recent years, memristors have many advantages such as non-volatile and multi value storage. With the improvement of research on their preparation, characterization, and testing, memristors have broad application prospects in non-volatile memory, computing/storage fusion architecture computing systems, new neural computing systems, and other fields, and are highly likely to change the physical basis of existing IT technology. Accurate weight control is the key to the application of memristors, but there are still some technical difficulties that are difficult to overcome in current research, such as the incomplete understanding of the conduction mechanism of memristors, the coexistence of multiple conduction mechanisms and their mutual transformation, which to some extent limit the largescale practical application of memristors. Therefore, this article focuses on the characteristics of memristors and designs a gate voltage feedback memristor weight precise control system using Moku: Go, a universal instrument. The system is controlled through a GUI interface and some devices are selected for testing in the prepared TiN/TaOx/HfOx/TiN memristor 1T1R array. The final test results show that the system achieves fast and accurate weight control.

Continuous action segmentation is a challenging task in video semantic understanding, aims to temporally segment unedited long videos. Current state-of-the-art methods combine time-domain convolution with self-attention mechanisms to capture temporal correlations, achieving high-accuracy frame-level classification and reducing over-segmentation during prediction. However, these models rely on multiple decoding modules and complex self-attention mechanisms, which, while improving frame accuracy, incur high computational costs, particularly when dealing with long video sequences. To address this issue, we propose a novel hybrid model that optimizes the temporal convolutional model with a simplified extended linear self-attention layer. This design enables the model to focus more effectively on action changes at key moments in a video sequence while significantly reducing computational complexity. Furthermore, we introduce an attention enhancement module with a pyramidal pooling structure to improve the model's ability to capture actions at multiple scales. By leveraging these innovations, our hybrid model achieves a better balance between accuracy and computational efficiency, making it more suitable for real-world applications involving long video sequences. The experimental results demonstrate that our model achieves an accuracy of 86.0% on the 50Salads dataset and 84.6% on the GTEA dataset, outperforming existing techniques such as MS-TCN++ and ASFormer in terms of both accuracy and computational efficiency. Notably, our model requires only 0.98M parameters, a significant reduction compared to ASFormer's parameter count. These advantages make our model particularly well-suited for practical applications involving long video sequences, where computational efficiency and accuracy are crucial.

Various automated code summarization techniques enhance the efficiency and accuracy of code annotations, enabling succinct natural language comments for code snippets. Recently, large language models (LLMs) have significantly improved natural language processing tasks. Among them, ChatGPT, based on the GPT-3.5 architecture, has gained widespread attention in academia and industry. Previous studies have also tested the ability of ChatGPT in code summarization, designing heuristic questions to explore an appropriate prompt that can guide ChatGPT to generate comments. In contrast, we have designed a more targeted and adaptive suggestion word strategy to study the impact of prompt design on model generation summary. Additionally, we have made extensive data fine-tuning to enhance ChatGPT's ability in code summarization tasks. The experimental results demonstrate that our prompt strategy has significantly improved the quality of code summaries generated by ChatGPT compared to previous studies, but still falls short of the SOTA model.

Segmentation of medical images is pivotal in the progression of proficient healthcare, especially the segmentation of heart images, which is one of its essential research contents. The U-Net architecture has demonstrated remarkable success across diverse segmentation of medical images tasks. Due to the inherent locality of convolution operations, the classical U-Net architecture exhibits challenges in image segmentation, manifesting as reduced resolution and information loss. Inspired by Transformers, the Vision Transformer model applied to classification tasks has good results. In this paper, a refined medical image segmentation model, HPNet, is proposed, and an encoder module that fuses convolutional network and Transformer structure is designed. The modules are paralleled to extract image features. Can grasp the global and local information of the image wellhouse modules are interconnected in parallel to extract image features, thereby enabling effective capture of both global and local information within the image. In the HPNet model, the encoder structure part of the transformer module and the CNN module are connected in parallel to process the input image. After upsampling the encoded features, the decoder merges them with the high-resolution feature map from the encoder section (the output of CNN and ViT). Secondly, this paper uses pyramid pooling operation in the Transformer module to reduce the complexity of the process and uses Shift MLP to improve the segmentation accuracy. The experimental findings indicate that through reduced computational complexity, the proposed HPNet model can improve the performance of the ACDC heart dataset by 1%.

Due to the differences in the characteristic states of factors affecting carbon emissions in different scenarios, there are certain differences between the accounting results and the actual situation. For this reason, a study on the design of an adaptive carbon accounting model that combines transfer learning and dynamic feature selection is proposed. The dynamic characteristics of carbon emissions were analyzed according to material categories and energy consumption categories, and the dynamic characteristics of carbon recycling were analyzed according to carbon sink categories and resource categories; during the model construction process, the dynamic characteristics of carbon emissions and carbon recycling were learned and transferred, and utilized Scenario transfer learning method optimizes the sequential decisionmaking process. Among the test results, the test results of the design method have the highest degree of fit with the actual situation, and the errors corresponding to different stages are within 1.0*104 t.

The global quantity of cigarette butt waste is staggering. According to data from the World Health Organization, approximately 4.5 trillion cigarette butts are discarded annually. This cigarette butt waste poses a serious threat to both the environment and human health. The filters in cigarette butts are made of non-biodegradable plastic material, which releases harmful substances such as heavy metals when decomposing in natural environments, leading to pollution of soil, water sources, and air. Therefore, the timely collection and treatment of cigarette butt waste has become a pressing issue. This paper introduces a cigarette butt waste suction robot combined with a real-time detection system based on an improved YOLOv5 algorithm. Experiments have shown that the recognition accuracy can reach 97%. The cigarette butt waste suction sanitation robot can efficiently collect cigarette butt waste, achieving automatic collection and processing of cigarette butt waste, improving cleaning efficiency, and reducing labor costs. This contributes to addressing the environmental and health issues caused by cigarette butt waste.

Parallel programs often struggle to achieve ideal computational speed and parallel efficiency on parallel computers, with communication overhead considered one of the primary limiting factors. However, existing communication performance analysis tools struggle to meet programmers' needs in terms of practicality and efficiency, making it difficult to analyze performance issues. In order to analyze the MPI communication performance of parallel programs more conveniently and comprehensively, a lightweight and low-overhead communication analysis software MPI Toolkit is designed and implemented, which consists of four modules, namely MPI function tracing, MPI function data statistics, log output, and data visualization, and encapsulates MPI functions based on the PMPI analysis interface to intercept MPI function calls, collect communication data during program operation, output it in the form of a log file, and provide a variety of visual views. To evaluate the effectiveness of the MPI Toolkit, comparative experiments between the MPI Toolkit and the TAU performance analysis tool are performed on typical parallel applications HPL and VASP. The results show that the tracing results of MPI Toolkit are basically the same as those of TAU, and the performance overhead is lower, which can easily and efficiently help users to analyze communication performance and locate potential bottlenecks.

This study mainly discusses the application of deep neural network model in improving the traditional source code auditing tool of hypertext preprocessor (PHP). Under the traditional static and dynamic conditions, there are still many restrictions on the use of PHP vulnerability mining technology. This study will build a deep neural network model on the basis of improving the traditional PHP source code auditing tools, and take the improved S-ASTNN deep neural network as the starting point to fundamentally improve the PHP vulnerability mining method. This study will use the special structure of PHP abstract syntax tree to adjust it. It is found that the semantic information contained in the improved traditional ASTNN deep neural network can be effectively preserved and the model efficiency is improved. After analyzing the experimental results, it is found that the improved S-ASTNN deep neural network model has higher accuracy and recall rate than the traditional vulnerability mining method, and the improved S-ASTNN deep neural network can play a more significant role in the field of PHP language vulnerability mining.

Multiple clients can co-train high-performance global models using federated learning (FL) without revealing their personal information. The main obstacle to federated learning, however, is the considerable statistical variability across local data distributions of clients, which results in clients optimizing contradictory local models. We suggest a novel adaptive federated learning method with local drift decoupling and correction (AFEDDC) to tackle this basic problem. We experimentally demonstrate the efficiency, optimality and robustness of AFEDDC and show that AFEDDC outperforms existing algorithms and provides better convergence for FL.

With the sharp increase in the global prevalence of diabetes, early diagnosis and prediction have become key to improving patient management and reducing the disease burden. This study investigates the use of explainable machine learning models for diabetes prediction, specifically employing the Pima Indian Diabetes dataset to develop a model that integrates explainability. Several advanced machine learning models—including the Bayesian classifier, Decision Tree, XGBoost, LightGBM, Random Forest, K-Nearest Neighbors (KNN), and CatBoost—were assessed using the 2019 Diabetics dataset. Notably, the SHAP framework was applied to the Pima Indian dataset to enhance the transparency and interpretability of the models. Through a comprehensive evaluation of different models, we found that the CatBoost model performed excellently across multiple performance metrics, particularly achieving an AUC value of 0.867, indicating superior sensitivity and specificity compared to other models. Additionally, the application of the SHAP framework revealed that glucose levels and BMI are the main factors influencing diabetes prediction, providing valuable insights for medical professionals to better understand and explain the prediction results. This study confirms the effectiveness of explainable machine learning methods in improving the accuracy of early diabetes prediction and model transparency. Our results offer new perspectives and approaches for applying machine learning technologies to diabetes and other chronic diseases in clinical practice, promoting the development of personalized and precision medicine.

Collaborative robots in the petrochemical industry often use electro-hydraulic servo control due to explosion-proof requirements. This paper establishes a mathematical model of the electro-hydraulic servo system for explosion-proof collaborative robots and studies the system control strategy. In response to the problems of slow response speed and resonance peaks in conventional PID control, a three state feedforward compensation and structural filtering design were adopted to eliminate the influence of system resonance peaks and improve the performance of system position tracking and speed tracking.

Power grids, as essential infrastructures of modern society, are vital for economic progress and the well-being of the populace due to their stability and reliability. However, the occurrence of natural disasters often leads to severe disruptions in the power grid, causing a sharp increase in material demand. Conventional material allocation models, due to their deficiency in capturing the complex dynamics of long and short-term mixed correlations in time series data, struggle to meet the requirements for rapid and precise material allocation in disaster scenarios. This paper presents a Gated Recurrent Unit model with a Dynamic Gating Mechanism (DGM-GRU) to enhance the predictive accuracy and responsiveness of power grid material allocation. Through the Dynamic Gating Mechanism, the DGM-GRU can effectively capture information of the complex dynamics of long and short-term correlations in time series data. The experimental outcomes indicate that the DGM-GRU model shows superior performance in both training and validation loss, demonstrating strong generalization capabilities and swift convergence. The model is adept at capturing the trends in material demand, offering scientific decision-making support, and thereby enhancing the efficiency and effectiveness of emergency responses.

Wireless self-powered sensors are necessary for long-term, large-scale, real-time environmental monitoring systems. Triboelectric nanogenerators can capture energy from the environment to power sensors. However, fewer works that directly transform energy into wireless sensing signals, and the majority of existing environmental sensors require energy storage devices to store energy. Direct-driven wireless self-powered sensing is further limited by low energy density in the environment. In this paper, we present a wireless humidity sensor system that is suitable for ambient energy harvesting. This method can be used in self-powered ambient energy harvesting devices to convert and modulate environmental energy into a sensing signal without any energy storage devices. The signal is processed by using a machine learning algorithm, and the humidity recognition accuracy can reach 98.7%.

With the changes in the ecological environment and the increase in human activities, the identification and prevention of poisonous species have become an important issue in public safety. However, traditional identification methods rely on a large number of samples and expert experience, which is inefficient and costly. This paper proposes a popular science platform for poisonous species identification based on small sample learning, aiming to solve the problems of scarce poisonous species samples and low identification accuracy, help users quickly and accurately identify poisonous species, and instantly obtain relevant popular science knowledge and preventive measures. In terms of platform design, we adopted a modular idea and divided the system into a front-end display layer, a back-end service layer, and a data processing layer. In response to the problem of sample scarcity, we used small sample learning methods such as data enhancement technology, transfer learning technology, and metric learning to achieve high-precision identification. In addition, we also designed a popular science module, including basic knowledge introduction of poisonous species, identification skills, preventive measures, etc., to help users improve their awareness of prevention and master self-protection skills. In summary, the popular science platform for poisonous species identification based on small sample learning designed in this paper is an innovative solution, which provides new ideas and methods for solving the problem of poisonous species identification. In the future, we will continue to optimize platform functions, improve recognition accuracy and efficiency, and explore more application scenarios and promotion models to better serve the public and the cause of ecological and environmental protection.

In modern emergency management and training, Virtual Reality (VR) technology has become an indispensable tool due to its safety and high simulation capabilities. However, in complex scenarios such as nuclear emergency response, traditional VR systems often require the collaboration of multiple participants, increasing the complexity of organization and implementation. This paper proposes an optimized Finite State Machine (FSM) method based on state aggregation and hierarchical design to enhance the autonomous decision-making capability and task execution efficiency of VR intelligent agents. By decomposing complex tasks into different levels and aggregating similar states, the optimized FSM effectively reduces the number of states and transitions, significantly improving the system's response speed and flexibility. This provides a new solution for intelligent training systems in emergency management. The experimental results demonstrate the optimized FSM's superior performance in state reduction and transition efficiency, offering significant improvements in system response and flexibility for VR intelligent agents in nuclear emergency scenarios.

Fire risk prediction is crucial for urban firefighting deployment, as it can reduce the damage and fatalities caused by fires. Therefore, we propose an urban fire risk prediction model, FIRE-CLA, to predict fire risks in urban areas. This model aids firefighting departments in prioritizing fire inspections at specific locations, including commercial and property areas, based on the predicted fire risks in different urban regions. FIRE-CLA calculates the fire risks for over 6,000 streets in the city, achieving a prediction accuracy of up to 90%. Additionally, FIRE-CLA presents fire risks at specific locations through an interactive map in a visualized form, making the model more intuitive and practical. This helps firefighting departments enhance their decision-making process for fire inspections.

In network intrusion datasets, the normal traffic is usually magnitude hundreds or even thousands of times more than the minority class attacks, and this dataset imbalance creates greater difficulties in deep learning model training, often leading to the model not being able to effectively learn the data features of the minority class, which in turn reduces the detection efficiency. Therefore, an effective solution is proposed to balance the normal traffic by utilizing WGAN for minority class data augmentation and random undersampling technique for random censoring, aiming to improve the overall model and the detection rate of minority class attacks. Finally, an intrusion detection model based on hybrid neural network for feature extraction is designed to address the shortcomings of existing intrusion detection models in terms of detection capability and structure. In the convolutional part, three different scales of convolutional kernels are used to perform feature extraction on the input data in order to obtain richer feature representations. Meanwhile, a bidirectional long short-term memory (BiLSTM) network is utilized to learn the temporal relationship between features, and a Transformer encoder is introduced to establish the connection between different features.

In order to solve the stability problem of intelligent vehicles, which is that high-adhesion roads are prone to rollover accidents and low-adhesion roads are prone to side-slip accidents under sharp turning conditions, the vehicle adaptive control model that considers vehicle yaw and roll stability is proposed. The vehicle dynamics model was constructed by comprehensively considering many factors that cause vehicle instability such as yaw, sideslip, and roll during the operation of the smart car. At the same time, the vehicle dynamics model is optimized to realize adaptive control of the model's time domain parameters, which significantly improves the tracking accuracy and the calculation efficiency of the controller. In order to further verify the performance of the controller, a vehicle dynamics model predictive controller was built based on the vehicle dynamics simulation software Carsim, and the effectiveness of the adaptive predictive control model was verified through J-turn test and fishhook test analysis. And The simulation results is show that the adaptive model predictive controller proposed in this article can effectively improve the calculation efficiency and control accuracy of the model, and the average single calculation time is shortened by approximately 16.76%. Compared with traditional rolling prediction methods, the adaptive simulation model proposed in this paper has a faster response speed and higher control accuracy, which can effectively ensure the safe and stable operation of intelligent connected vehicles without collision in complex environments such as time-varying curvature and roll slope.

This study aims to predict the corrosion fatigue crack propagation rate in 7050 aluminum alloy using machine learning methods, specifically models based on the XGBoost algorithm, BP neural network, and random forest. Due to the widespread application of 7050 aluminum alloy in aviation, automotive, and other industries, accurate prediction of its corrosion fatigue performance is crucial for enhancing the reliability of engineering structures. By thoroughly analyzing the research background of 7050 aluminum alloy, trends in fatigue life prediction, and machine learning feature selection algorithms, significant features are identified for the XGBoost algorithm, BP neural network, and random forest. The performance of these three algorithms is compared, and the XGBoost algorithm is ultimately selected for its excellent predictive performance, achieving a mean squared error of 0.94 and a coefficient of determination of 0.98. The results provide valuable insights into corrosion fatigue life prediction for aerospace-grade 7050 aluminum alloy and establish a theoretical foundation for subsequent model development.

In complex inspection procedures, there is a significant variation among different personnel. However, the existing personnel scheduling methods often treat individuals as similar or identical entities, which can result in low overall yield and substantial losses for the company. In this context, a study was conducted to improve the yield from the perspective of personnel scheduling in inspections. Based on the interception capabilities of different personnel for various types of defects, and in combination with the concentrations of different defect types for different models, a hybrid algorithm combining genetic algorithms and particle swarm optimization algorithms with elite decision-making was proposed. This algorithm allows for the selection of suitable personnel from the personnel pool, followed by targeted personnel scheduling based on the defect concentrations specific to each personnel and model. This ensures that the appropriate personnel are assigned to the corresponding models, thereby maximizing the yield for all models. Through experiments and result analysis on different scenarios, as well as comparisons with other algorithms, the test results demonstrate the superiority of this method in terms of convergence and optimization capability. This approach offers a new perspective on scheduling based on the multidimensional testing capability of employees, providing valuable insights for optimizing production and scheduling in enterprises.

The purpose of this paper is to construct and validate a low-carbon dispatch model for the power system, which combines the design of the objective function and the formulation of multiple constraint conditions to reduce carbon emissions during the process of power production and distribution. The objective function takes into consideration both the reduction of energy consumption and carbon emissions, and selects an evaluation index that balances production efficiency and environmental impact. The constraint conditions ensure the practical applicability of the model, covering key indicators such as power balance, generation capacity limitations, and load demands. Additionally, the model adopts a fuzzy processing strategy to deal with the uncertainty of input parameters. To solve the complex low-carbon dispatch problem, a Particle Swarm Optimization algorithm with simulated annealing mechanism (PSO-SA) is proposed. The integration of the algorithm's global search capability and the avoidance of premature convergence enhances the optimization efficiency. Finally, through simulation experiments, the paper demonstrates the effectiveness of the low-carbon dispatch model in practical power systems, validating the potential of the proposed method in reducing carbon emissions and improving the operational efficiency of the power system.

With the increasing use of digital technology across industries, standardization is undergoing a digital transformation. Machine-readable standards, characterized by their digital and structured nature, support this transformation. Standard tag sets, as specific applications of machine-readable standards, play a crucial role in creating information formats that comply with these standards, thereby improving efficiency in information processing. This study focuses on designing standard tag sets for the oil and gas pipeline domain. Through comprehensive requirements analysis, including overall and personalized needs assessment, we identify and address the challenges in developing, implementing, and maintaining these tag sets. Our research outlines the construction of both general and extended standard tag sets for oil and gas pipelines domain. The general standard tag sets mainly identify the structure of standards and technical indicators defined in standards. The extended standard tag sets achieve unification of technical element data in standards of different type by constructing domain ontology models. By proposing and implementing these standard tag sets, stored in the Neo4j database, we lay a strong foundation for knowledge organization and applications such as building oil and gas pipeline knowledge graphs. In conclusion, the standard tag sets that we put forward and design inject vitality into the digitization of the oil and gas pipeline industry, fostering innovation and development while providing essential support for its long-term progress.

This article provides a comprehensive examination of the digitalization process in nuclear power plant switchboards design, emphasizing its significance in enhancing efficiency, precision, and minimizing design alterations risks. Utilizing VBA technology, the switchboards design workflow has been effectively digitized and automated, covering tasks from feeder table processing to system diagram creation and revision. It begins by highlighting the intricate nature of switchboards design in nuclear power plants and the stringent safety requirements involved. Subsequently, it discusses the digitalized design framework, outlining key functionalities such as feeder table generation, process load data processing, electrical component and cable selection and verification, and system diagram publication and revision. Through case studies, it showcases the substantial benefits of digitalized design in streamlining design processes and improving accuracy. Notably, the feeder table processing module significantly reduces processing time and enhances accuracy in handling process load data, while the cable and component selection module simplify the selection and verification process. Additionally, the automated operations of the system diagram publishing and revision module contribute to efficient drawing processes. Overall, digitalized design offers innovative insights and solutions for switchboards design in nuclear power plants, ensuring dependable technical support and assurance throughout the design journey.

The research focuses on the buffeting of transmission line tower system under the action of nonstationary typhoon. Using a field-monitoring-based typhoon model, the wind field for line tower structures is generated. Then, typhoon-induced dynamic responses of the system are simulated by the nonlinear loading. According to the resultant time histories and evolutionary power spectrum density (EPSD) of responses, the vibration characteristics of the line tower are discussed in both time and frequency domains. The dynamic loads transferred from the conductor to the tower are also of concern. The results indicate that the vibration of line tower system exhibits non-stationarity, and the dynamic behavior of conductor is significantly influenced by wind fluctuating while that of tower is closely related to conductor dynamics.

In order to improve the accuracy and efficiency of detecting degraded insulators in distribution networks, a detection method for degraded insulators in distribution networks based on sound and light collaborative detection technology is proposed. This method first captures the sound wave signal generated during the operation of insulators through an acoustic sensor system, and synchronously obtains the image data of insulators using high-definition optical imaging equipment. Secondly, preprocess the collected acoustic and optical data to remove potential noise interference and extract key feature information. Integrating acoustic and optical data to obtain more comprehensive and reliable insulator status information. Finally, the Kalman filtering algorithm analyzes the fused data to accurately identify degraded insulators. The experimental results show that the proposed method is more efficient than traditional detection methods, reducing the workload and subjective errors of manual detection.

The accurate identification of large and medium slag in power plant slag transport system has an important impact on the safety, efficiency, environmental protection and economic benefits of power plant. The timely processing of large slag can effectively improve the operation efficiency and maintenance cost of power plant, and with the rapid development of image recognition technology, the recognition of large slag in power plant slag transport system is gradually intelligent. In this paper, based on CenterNet image recognition algorithm, through the adaptive spatial feature fusion module is added in the backbone network, and combined with the feature of the pyramid networks structure. The low-level feature information is fused with the high-level feature information. The loss function is replaced by the distributed focus loss, and the effectiveness of the system is verified by experiments. The results show that the system has good convergence, can accurately identify the size and position of large slag block, can respond to the slag removal command quickly, improve the operation efficiency of the power plant slag delivery system, reduce the manual maintenance cost, and improve the operation safety of the power plant.

In the face of the huge number of pictures of electrical UAV patrol electrical equipment, the efficiency of human eye recognition is low, and it is impossible to quickly locate the high temperature defects of electrical equipment. Therefore, this paper proposes a screening method for high temperature defect images of electrical equipment. Firstly, the target data set is constructed, and the target sample is enhanced for the problem of unbalanced target samples. Then, the network training of improved YOLOv5 is completed, and compared with YOLOv5, YOLOv4 and other methods. The experimental results show that the proposed method can effectively improve the detection ability of the network to the target, and realize the recognition of electrical high temperature defects in the case of too many infrared interference factors.

Ship power system has the characteristics of poor self-regulation ability and low system inertia. Meanwhile, the application of hydrogen energy increases the complexity of the system, making it more difficult to control. The proportion of propulsion load in ships is high, which can easily affect the overall system stability of hydrogen ship. This paper presents a propulsion load forecasting framework for hydrogen ships, which is based on Temporal Convolutional Network (TCN)-Long Short-Term Memory (LSTM) and combined with classified data processing method. Firstly, TCN is used to extract the time series characteristics in the long history range, and then the load forecasting is carried out through LSTM. Data set is divided into two parts: discrete data and stable data. Complementary Ensemble Empirical Mode Decomposition (CEEMD) is used to decompose the discrete data. After integration with the stable data, the dimension is reduced by Principal Component Analysis (PCA). Finally, taking the operation data of a hydrogen fuel cell ship as an example, the results show the superiority of the proposed method.

The oil pipeline system faces numerous safety hazards and operational management challenges, such as leaks, blockages, and pressure fluctuations, which may lead to environmental pollution, economic losses, and even casualties. To address these issues, this study proposes a predictive PID control method for oil pipeline infrastructure based on cloud servers and Industrial Internet of Things (IIoT) technology, coupled with advanced data analysis and control algorithms, enabling comprehensive monitoring and management of oil pipeline systems. A risk assessment and control framework based on online model fusion is developed, integrating diverse data sources and optimization model algorithms to achieve intelligent monitoring and control of pipeline systems. Additionally, PID control algorithms are employed for real-time adjustment of flow rates and pressures in the pipeline system to ensure stable operation. Experimental results demonstrate that the proposed risk assessment and control framework effectively identifies anomalies in the pipeline system, while the PID control algorithm enables real-time adjustment of flow rates and pressures, enhancing the safety and stability of the pipeline system. Comparative analysis shows that the proposed control method offers a 28.96% improvement in feedback control effectiveness and can reduce the risk rate by up to 4.2 times through the predictive modeling method.

Nuclear power safety has always been a key focus within and outside the industry. To ensure the safe operation of nuclear power equipment, various methods have emerged, including empirical judgment, regular inspections, and expert systems. However, with the demand for automation and intelligence, traditional diagnostic methods and the inability to meet the current new situation, intelligent diagnosis has become a mainstream and trend. In recent years, intelligent diagnostic technology based on artificial intelligence has gradually attracted people's attention due to its unique advantages. This technology can accurately identify fault patterns and provide effective decision support by mining and analyzing a large amount of data, greatly improving work efficiency. This article aims to review the development of intelligent diagnostic technology, introduce the practical application of intelligent diagnostic technology, and provide suggestions for future improvements. Firstly, the core components, main diagnostic steps, and classification of diagnostic methods of intelligent diagnostic technology were elaborated. Secondly, some research and applications of intelligent diagnostic technology in nuclear power equipment were demonstrated, demonstrating its enormous potential in achieving automation and precise services. Then, the existing problems and shortcomings were pointed out, and corresponding improvement suggestions were proposed. This article focuses on the enormous potential of intelligent diagnostic technology in this field, and analyzes the challenges and development trends faced by nuclear power equipment diagnostic technology from three perspectives: model performance improvement, multimodal fusion, and interpretability.

In response to the problem of difficult detection of loose bolts in transmission tower bolts, this paper proposes a method for detecting loose bolts in power tower bolts based on the Chirplet decomposition algorithm of laser vibration signals. This method obtains vibration signals of power towers through laser vibration measurement technology, and uses the Chirplet decomposition algorithm to finely process and analyze the vibration signals. Chirplet, as a basis function with adaptive time-frequency characteristics, can accurately capture nonlinear and non-stationary features in vibration signals, especially in details such as frequency changes and mode transfer caused by bolt loosening. By extracting and identifying the signal features after Chirplet decomposition, effective monitoring and diagnosis of the loosening status of power tower bolts can be achieved. The experimental results show that this method has high accuracy and reliability, providing strong technical support for the safe operation of power towers.

Communication base stations and data centers represent an emerging industry derived from the existing Information and Communication Technology (ICT) sector and have become integral components of the digital society's infrastructure. This study uses the random forest model to predict the development trends of these facilities in China, analyzing data from 2014 to 2023 on development scale, electricity consumption, and load. The results show that economic development and infrastructure investment, market demand are the main drivers of the scale of these facilities. Forecasts indicate substantial growth in their scale under high economic growth, with corresponding increases in electricity consumption and power load. However, traditional industries will still dominate power load. This research supports government power regulation and grid adaptation, providing a scientific basis for planning the development of communication base stations and data centers, highlighting its practical significance.

Due to the discrete nature of electricity consumption data of typical power users, the reliability of direct analysis and mining is difficult to guarantee. Therefore, a graph neural network-based algorithm for mining electricity behavior patterns of typical power users is proposed. Construct a Markov matrix of quantile boxes for discrete typical electricity user behavior data, and encode dynamic transition probabilities into a Quasi Granian matrix to map the user behavior graph structure data; Using GCN to aggregate the polarity of neighbor information of typical user electricity behavior graph structure data nodes, based on setting the attention coefficient between nodes, and following the addition rule propagated layer by layer downwards, output the characteristics of typical power user electricity behavior patterns. In the comparative test results, the designed algorithm achieved convergence the fastest, and the corresponding DBI value was the lowest, only 1.44. The PCC between different load states was below 0.1.

Aiming at the problem of the automatic layout of optimal energy-efficient piping in the high-density layout environment of an intelligent workshop, a method of optimal energy-efficient piping system layout in intelligent workshop based on an improved ant colony is proposed. Firstly, the friction loss of multi-specification pipes is measured through the friction experiment as the layout evaluation index. Secondly, the layout scene is reconstructed based on the grid and enclosing box methods. Thirdly, initial guidance information and improved update strategy based on the workshop piping layout rules to differentiate the area guide; Reward and correction rules are introduced to control the change of path information, balancing the intensity of environmental guidance in later stages; and the heuristic information of inclination to the target is introduced to the ant colony state transfer strategy to enhance the efficiency of exploring the target solution. Finally, the optimal energy-efficient piping layout of the intelligent workshop is obtained by the optimization Mechanism for main and branch pipelines. Experiments show that this method can achieve higher energy efficiency and faster layout results in dealing with complex intelligent workshop piping scenarios.

The cross server room cluster environment is prone to failures due to network latency, node failures and other factors. In order to ensure the smooth progress of fault detection in cross server room cluster environment, a distributed storage cross server room cluster fault detection method based on improved Gaussian mixture model is proposed. First, the data in the cluster is modeled by using GMM; second, the storage resources on multiple independent devices are integrated through the network to form a unified virtual storage device. Finally, according to the running environment of the application program and the current state of the network, the distributed storage cross server room cluster fault detection is dynamically adjusted to further improve the adaptive performance of the fault detector and the accuracy of the result determination. The feasibility and performance of the model are evaluated. The experimental results show that compared with the traditional fault detection model, the distributed storage cross-room cluster fault detection method based on improved Gaussian mixture model has significantly improved the reliability and effectiveness of storage cross-room cluster.

This study addresses the detection of dislodged fault in photovoltaic (PV) power stations by proposing a visible light image processing method utilizing line scanning. The method employs HSV (Hue-Saturation-Value) color space conversion and morphological processing to preprocess images, which achieves precise segmentation and shape simplification of PV module areas through flood fill and polygon contour fitting techniques. For fault detection, the study introduces a line-scanning-based algorithm. By setting appropriate thresholds and continuous occurrence parameters, the algorithm effectively identifies and marks dislodged areas of PV modules in the original image. Experimental results show that this method significantly improves detection accuracy and computational efficiency, providing technical support and solutions for the safe operation and maintenance of PV systems.

With the advancement of industrialization and the development of the social economy, electric power resources have become an essential guarantee for ensuring the efficient operation of society. As a critical implementation step in the design and development of power systems, power load forecasting not only ensures the safe and stable operation of power systems but also assists in accomplishing reasonable power distribution tasks. It has significant technical and economic importance. However, existing research on power load forecasting primarily relies on expert systems or general time series forecasting methods, which seldom consider the differences between power load data and other time series data. This presents difficulties in effectively leveraging the spatiotemporal attributes of power load data for forecasting purposes. To tackle this challenge, this paper introduces a data-driven model for power load prediction utilizing the attention mechanism. Firstly, the model incorporates multi-source heterogeneous data, deeply exploring the spatiotemporal correlations of load user behavior data. Secondly, a covariate dimensionality reduction module based on residual neural networks is designed, significantly improving the model's computational efficiency. By constructing the Fourier transform, the model can effectively extract and embed the periodic characteristics of power data. The model is tested on a regional dataset and three public datasets. The findings indicate that the proposed approach surpasses baseline models across all evaluation metrics, offering dependable predictive support for the stable functioning of power systems.

Electric power operations have high safety requirements for human operators. Human fatigue is a significant factor affecting operational safety. Various objective physiological and behavioral indicators of the human body provide certain characterization effects on the degree of human fatigue. However, single indicators have strong limitations in measuring human fatigue, whereas models based on machine learning and the fusion of multiple indicators are expected to mitigate these limitations. Through large-sample experiments, this study collected various objective physiological and behavioral indicators of operators under different working and fatigue states (including: α, β, and θ wave power spectral energy in EEG, blood oxygen saturation, heart rate variability, end-tidal carbon dioxide concentration, PERCLOS value, reaction time, and critical flicker fusion frequency). Then, based on deep learning theory, fatigue evaluation models were established using five different machine learning algorithms, and the algorithms were compared. The conclusions are as follows: (1) The nine objective physiological and behavioral indicators mentioned above are correlated with changes in fatigue levels and can well reflect different fatigue states; (2) When the nine indicators were integrated into the machine learning classification model for classification evaluation, the highest classification accuracy, recall rate, and F1 score reached 0.996, which appeared in the decision tree model. This proves that multi-indicator fusion for fatigue evaluation has excellent performance. The conclusions of this study can provide practical references for improving the efficiency and level of safety management in enterprises.

Deepfake detection aims to authenticate the authenticity of facial images and videos, thereby providing crucial theoretical and technical support for portrait rights protection, false information identification, and online fraud prevention. Recent research has introduced a deep neural network Transformer, leveraging a self-attention mechanism to model long-range dependencies and global receptive fields. This advanced architecture enhances the ability to capture contextual relationships in images and temporal dependencies in videos, thereby improving the representational capacity of the detection system. In this paper, we propose a dual-branch feature extraction framework, incorporating a spatial domain feature extraction branch based on EfficientNet and a frequency domain feature extraction branch, to enrich the feature representation compared to single-branch methods. The extracted features are subsequently integrated with the Transformer's encoder structure and cross-attention mechanism, enabling effective modeling of feature correlations across global regions. Our proposed method achieves detection accuracies of 83.5%, 70.25%, and 78.5% for Deepfake, Face2Face, and NeuralTextures forged images, respectively.

In order to reduce the economic losses caused by pest infestation during mulberry leaf cultivation, this paper proposes a YOLOv8 detection model integrating lightweight and multi-scale, named YOLOv8s-SLS. The channel-to-feature-to-space channel (C2FSC) module is first introduced in the backbone network to compensate for feature information lost due to model deepening by using complementary information between neighboring regions. Then, the neck structure and the detector head (NLN) were redesigned to improve the recognition of target pests at multiple scales while removing redundant connections in the model. Finally, the LSKA module enhanced the feature representation of the model by dynamically adapting to the sensory field. In addition, a mulberry leaf pest dataset containing different target sizes, named MPD1, consisting of 1705 raw images of three pests, was constructed for model training and validation. The experimental results on the test dataset showed that the parameters of the enhanced and multi-scale versions of the model were reduced by about 15% and the mAP50 was improved by 3.7% compared with the original YOLOv8 model. The experiments proved that the model can quickly and accurately identify pests in mulberry gardens, providing feasible technical support for real-time detection of pests in the sericulture industry.

Automated bird identification systems are becoming increasingly important in biodiversity research and ecological conservation. I n this study, we present a bird recognition approach utilizing an enhanced YOLOv8 model. The detection accuracy and robustness are boosted by integrating the Global Attention Mechanism (GAM) and NAS-FPN into the model. GAM effectively enhances the model's ability to focus on important features, while NAS-FPN optimizes the fusion and extraction of multi-scale features. We tested the enhanced model using the CUA-200-2011 dataset, and the experimental findings indicate that the upgraded YOLOv8 model demonstrates improvements in both recall and accuracy metrics. This research not only enhances the precision of bird recognition but also offers a novel approach for detecting small targets in complex environments.

Road damage detection is a key step in maintaining road safety and extending road life, and traditional detection methods depend on manual in spection, which is both labor-intensive and lacks efficiency, with limited accuracy. With the development of UAV technology, UAV inspection has become an efficient and low-cost means of road inspection. In this study, we introduce a UAV-based road breakage detection method leveraging an enhanced Detr model. By adding feature pyramid network (FPN) after the CNN feature extraction module and adopting the ViT encoder structure, the model's capability to detect targets of various sizes and identify them in complex backgrounds is significantly improved. The experiments utilized the UAV-PDD2023 dataset, comprising 2,440 road pavement images captured through UAV inspections and annotated with six common types of road damage. The experimental results demonstrate that the enhanced Detr model can efficiently and accurately identify six types of road damage, including longitudinal cracks, transverse cracks, alligator cracks, diagonal cracks, repairs, and potholes. In particular, the detection performance is significantly improved for damage targets in small scales and complex backgrounds. In summary, the UAV road breakage detection method based on enhanced Detr introduced in this study provides effective technical support for UAV road inspection and has important practical application value.

Surrogate model plays a vital role in both the engineering design optimization and design space exploration as it can provide fast analytical approximations of the quality of interest. While, current modeling method can only use end-to-end data from the numerical simulation, leaving the valuable distributed physic information of the physic field being unused. To get the full benefit of both the integrated and distributed information, a physic-embedded Hierarchical Kriging modeling method is proposed and utilized to predict the aerodynamic performance of a tandem cascade for axial flow compressor. This modeling method extracts low-dimensional physical features from the pressure distribution of the cascade surface and then improves the modeling accuracy of the static pressure of the cascade effectively via feature embedding. Compared with traditional modeling approach, the proposed method can improve the modeling accuracy more than 20%.

This paper investigates power load forecasting models and introduces an advanced method that integrates Ensemble Empirical Mode Decomposition (EEMD), Whale Optimization Algorithm (WOA), and Long Short-Term Memory neural network (LSTM). The proposed EEMD-WOA-LSTM model is validated through experiments, demonstrating superior prediction accuracy and efficiency compared to traditional methods, particularly in managing nonlinear and non-stationary data. The model outperforms traditional LSTM and WOA-LSTM models in various evaluation metrics, showing a closer alignment with actual values when handling complex load data. Additionally, the paper discusses future research directions, such as model optimization, multi-source data fusion, real-time prediction systems, and extending applications to broader areas. This study not only highlights the significant potential of the EEMD-WOA-LSTM model in power load forecasting but also offers valuable technical references for the intelligent transformation of the power industry, aiming to develop more advanced and practical solutions for smart grids and related fields.

A sub-pixel edge detection method based on improved polynomial interpolation is proposed to solve the problem of rapid positioning and real-time detection of the position and welding offset of the workpiece during electron beam welding. In the process of image processing, the Canny edge detection algorithm is first used for rough localization. This algorithm effectively extracts the edge information of the weld seam through Gaussian filtering, gradient calculation, non maximum suppression, dual threshold detection, and edge connection steps. In order to improve the accuracy of edge detection, an adaptive adjustment method of dynamic threshold is introduced to adapt to images with different brightness and contrast. In the edge refinement stage, polynomial interpolation is used for sub-pixel coordinate detection. By using cubic interpolation method and combining the grayscale information of local areas, the sub-pixel positions of edges are accurately calculated, thereby improving the accuracy of edge detection. The experimental results show that the improved method has significant accuracy advantages in edge extraction and sub-pixel coordinate detection, providing an effective solution for quality control of electron beam welding.

In this study, a real-time secondary electron-based weld seam localization method is proposed to improve the accuracy and reliability of weld seam localization during electron beam welding (EBW). By using secondary electron imaging technology, detailed image information of the weld area is acquired in real time and these images are analyzed to accurately locate the weld position. Compared with the traditional optical observation method, the secondary electron-based real-time weld seam localization method has the advantages of high real-time performance and accuracy, high automation and quantification, and strong resistance to environmental disturbances. The experimental results show that the method can effectively improve the welding quality and ensure the stability and consistency of the welding process.

Fused Deposition Modeling (FDM) is one of the most widely used additive manufacturing processes, but its printing process defects such as first-layer defects, drawing, wrinkling and other printing failure problems will often occur, and the printing first-layer defects are the most common and the most influential kind of defects in 3D printing process. In this study, for this first layer defect identification problem in 3D printing, a 3D printing first layer defect detection method based on improved YOLOv8s is proposed, which adopts the EfficientViT model as the backbone network, and improves the diversity of attention while reducing the number of overall parameters in the model. Secondly, MPDIoU is used to replace the CIoU of the original model for the problem of bounding box regression loss. to improve the detection performance of the model, it is experimentally verified that the improved model improves Precision and Recall by 3.5% and mAP 50 by 3.3% in the first-layer defect detection compared with the original YOLOv8s model, respectively. The model size is reduced by 47.4%, which has the advantages of high accuracy in first layer defect recognition and less memory consumption.

In today's high-speed logistics era spawned by e-commerce, with the increase of the circulation volume of goods, warehousing is urged to innovate to improve the efficiency of commodity delivery. However, the traditional methods of warehouse renovation and reconstruction will lead to huge cost consumption, which may even lead to the stagnation of warehouse operations and the backlog of orders. This makes many enterprises avoid the replacement of system structure and mode when they are faced with improving warehousing performance. This is inefficient in the long run. Though AutoStore has avoided these disadvantages of traditional warehouse upgrading, it has never received much attention since its birth for more than 20 years. Therefore, we think it is necessary to systematically elaborate this efficient warehousing system. In this study, chapter 1 introduces the structure and history of AutoStore, and sorts out the related literature. It also points out the optimization problem with the most research necessity and prospect: Bins Relocation Problem for AutoStore (BRP-AS), which is the research content of this paper; Then in chapter 2, we give a detailed definition of it and point out its principle; In chapter 3, we put forward a set of abstract methods to transform BRP-AS into mathematical problems. Meanwhile, in the chapter 4, according to the similarity of the problems, we compare BRPAS with the containers relocation problem (CRP) and the traveling salesman problem (TSP), and point out the referential solutions. Lastly, in chapter 5, we summarize the whole paper.

Aiming at the low efficiency of manual input of electronic scales in supermarkets, wasting labor costs and unable to adapt to the current technological development, this study therefore introduces a lightweight detection of fruits based on the improved YOLOv5s for electronic scales in supermarkets to achieve fast and accurate identification of relevant fruit targets. First, the dataset is obtained from the publicly available fruit dataset and organized and then manually labeled, mainly labeling the location and category of the fruit. Second, the backbone of YOLOV5s uses MobileNetV3 to reduce parameters for easy deployment in mobile devices. Then, BiFPN is introduced in the process of neck cross-scale fusion, which can be able to fuse multi-scale features more effectively and further improve the accuracy to the network. Finally, to further increase the precision of anchor boxes, the CA coordinate attention mechanism is introduced before the head prediction. The experiment shows that the improved YOLOv5s improves 17.3,12,10.3,and 1.2 percentage points compared with SSD, YOLOV3, YOLOv4, and YOLOv5s respectively, and the size of the improved YOLOv5s has decreased from 13.8MB to 7.2MB, and the improvement of YOLOv5s in this experiment can be better applied and deployed for the superstore's electronic scales and better meet the practical needs.

In recent years, research on multi-modal object detection has garnered significant attention due to the comprehensive information obtained from multi-modal data. However, most object detection studies focus on outdoor scenarios, such as autonomous driving, with relatively few investigations addressing the characteristics of indoor scenes. In this paper, we identify the shortcomings of similar object detection performance in indoor scenes and propose MODCL: Multi-modal Object Detection with End to End Contrastive Learning in Indoor Scene. Within MODCL, we focus on two aspects of improvement: first, the fusion of multi-modal context based on the mapping relationship between point clouds and images; second, the incorporation of supervised contrastive learning in an end-to-end manner, eliminating the need for pre-training. Furthermore, we conducted experiments on the SUN RGB-D dataset, and the results indicate that MODCL outperforms existing detection methods that utilize both point clouds and images compared to those that rely solely on point clouds.

As bridge infrastructure ages and load increases, Bridge Health Monitoring Systems (BHMS) have become increasingly important. Sensors play a crucial role in BHMS, but sensor failures may result in inaccurate data, thereby reducing the reliability of the monitoring system. Due to the interference from the vehicles on the bridge, current bridge sensor fault detection algorithms cannot detect the faults of sensors accurately. To tackle this challenge, we combine the advanced State Space Model (SSM, known as Mamba) with Transformer and propose a novel bridge sensor fault diagnosis method MSMV-MT. Firstly, MSMV-MT samples the sensor sequences at different scales and adopts Past Decomposable Mixing (PDM) for multi-scale fusion to capture at different-scaled features. MT block is constructed by replacing the Multi-head Attention with bidirectional Mamba in Transformer. We introduced a Full Sequence Channel Attention (FSCA) mechanism in the MT block to weight the attention of sequences of different channels and perform multi-view fusion with Self-Attention, thereby enhancing the accuracy and robustness of fault diagnosis. We constructed a simulation dataset with collected bridge sensor data and conducted experimental analysis. The experimental results show that the MSMV-MT outperforms all compared algorithms and achieves the highest accuracy of 92.7% on the bridge sensor fault diagnosis dataset. Comprehensive comparative experiments and ablation analysis have demonstrated the effectiveness of the proposed scheme.

In response to the vibration signals of rolling bearings containing strong background noise, this paper uses multi-scale fuzzy entropy as the threshold for the dynamic mode decomposition mode to effectively distinguish low rank feature information from sparse noise components in the original signal, thereby reducing the impact of noise on fault extraction. At the same time, the IPSO algorithm is used to adaptively select thresholds and truncation ranks in dynamic mode decomposition, avoiding the shortcomings of traditional truncation rank hard thresholds that cannot effectively select truncation ranks, and achieving effective processing of fault signals. The fault feature frequency is successfully extracted, and the fault diagnosis of rolling bearings is achieved.

With the rapid development of artificial intelligence technology, large language models are increasingly applied in the power industry, providing strong support for the intelligent management of power systems. This paper proposes an evaluation system for large language models in the power domain, aiming to evaluate their performance and limitations in solving specific tasks within the power industry. Through testing on the evaluation dataset, we analyzed the performance of large language models in both general domain and the power industry. The experimental results show that domain-adapted power large language model should balance both power industry-specific and general domain capabilities during evaluation. This research has significant theoretical and practical value for the evaluation of large language models in other industries.

The transmission line image defect diagnosis method based on a large model of power inspection uses a large model pre-training algorithm to design a complex, high-parameter network architecture and an industry big data training network to improve the model's perception capabilities and capture industry data characteristics more accurately. Improve the recognition accuracy of later application models. The self-supervised learning of massive data in the pre-training stage accumulates a large amount of background knowledge, allowing the model to infer richer semantic information from the image space context information when processing small samples, effectively improving the model's learning ability on small samples. and generalization ability. Design a foreground and background screening model, use 3D modeling with random backgrounds to produce inspection images, automatically segment and generate foreground and background image blocks, realize rapid training of the foreground and background screening model, realize pre-training data classification selection in real scenarios, and improve Large model pre-training speed. In the transfer learning process, a non-quantitative region-of-interest pooling layer is designed to achieve target positioning and classification optimization.