This PDF file contains the front matter associated with SPIE Proceedings Volume 13281, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

With the continuous development of digital economy and Internet of Things technology, it has brought huge benefits to people and generated massive amounts of data. As a result, users migrate their data to the cloud. Cloud storage not only provides a large amount of storage space and low storage costs for user data, but also provides users with more convenient access methods and more flexible capacity scalability. However, moving data to the cloud brings many benefits as well as security and integrity issues, which contains the data corruption, unauthorized access, and sensitive information compromise. In this work, we proposed a novel verification process to check the integrity of data in the cloud storage platform without revealing any sensitive information. The model uses block-chain technology to provide a decentralized verification mechanism in this protocol, ensuring the integrity of data by creating an immutable and transparent chain of records. As for data change or access is recorded on the block-chain, providing a verifiable, tamper-proof history of data that builds trust across the system. From our extensive experiments, we can observe that our proposed can achieve the verification process with reasonable protocols and acceptable computation costs.

Cloud manufacturing was a new type of manufacturing model, and a cloud task decomposition method and a service composition optimization method were proposed to address the problem of cloud manufacturing resource optimization allocation. Firstly, a correlation matrix was established based on the information correlation between atomic tasks, then hierarchical clustering algorithm was used to reorganize the atomic tasks, and finally the decomposition of the total tasks was realized. For the service composition problem of cloud manufacturing, an improved sparrow search algorithm (ISSA) was proposed. The ISSA adopted the Latin hypercube sampling method to generate an initialized population and incorporated the golden sine operator and the opposition-based learning strategy to avoid the algorithm from falling into a local optimum. Finally, an example of Bluetooth headset manufacturing was used to validate the resource optimization allocation method proposed in this paper, which demonstrated the feasibility of the method proposed.

In response to the inadequacies of the current methods for assessing the risk of coal mine gas overrun limits, including weak foundational techniques and limited visualization, this study integrates the AHP-PCA and cloud model algorithms to construct a more comprehensive and scientifically grounded model for evaluating the risk of coal mine gas overrun limits. The established coal mine gas overrun risk assessment system uses a combination of AHP and PCA to calculate the weights, which effectively solves the complex relationship and uncertainty between various assessment indicators and reduces the influence of subjective and objective factors on the assessment results. The cloud algorithm is used to conduct comprehensive evaluation, so that the evaluation results can be interpreted more visually. The results show that the method can accurately evaluate the risk factors of gas overrun, and provides a certain scientific reference for the prevention and control of coal mine gas overrun risk.

With the rapid development of Artificial Intelligence (AI) technology, the demand for various types of computing-power, such as Central Processing Unit (CPU), Graphics Processing Unit (GPU), Neural Processing Unit (NPU), Tensor Processing Unit (TPU), and Field-Programmable Gate Array (FPGA), has been increasingly growing to accommodate diverse data processing tasks. To enhance the efficiency of computing-power provisioning, addressing the current imbalance between supply and demand, the computing-power awareness architecture has been developed utilizing a domain-based collaborative topology. This architecture ensures a balanced network, robust fault tolerance, and straightforward system management and maintenance. Identification resolution and trusted authentication technology are applied to co-allocate heterogeneous com-puting-power resources (CPR), providing a unified interaction and sharing platform for various business needs.

Current outlier detection methods exhibit many limitations in high-dimensional settings. Traditional statistical approaches rely on strong assumptions and lack practicality and generality. Meanwhile, despite the better performance, machine learning methods suffer from low interpretability and reliability due to their complex mechanisms and the absence of confidence estimation. Although outlier detection methods based on principal component analysis (PCA) have shown some advantages by extracting important features from high-dimensional data, they have not entirely solved these problems. Conformal prediction, a finite-sample distribution-free uncertainty quantification method recently applied to outlier detection, produces a set-valued prediction with a Type-I error guarantee and false discovery rate control. This paper proposes a new distribution-free method for high-dimensional outlier detection, PCA-CP, combining principal component analysis and conformal prediction. PCA-CP overcomes the shortcomings of previous methods, demonstrating high generality and reliability while achieving high performance. Experiments on simulation and real data show that PCA-CP outperforms previous methods, achieving higher power and lower False Discovery Rate, thus proving its significant advantages.

In this paper, we propose a data generation method for background noise processing of spectral signals. Spectral signal data cannot be easily collected, so various data generation methods are being researched. The previous method implemented peak and background noise using kernel functions and polynomials. This method is similar to real spectrum data, but peaks are modeled globally, making peak identification difficult. Also, the polynomial method has limitations in implementing the complex background noise of the real spectrum. In this study, we focus on generating data using Generative Adversarial Network (GAN). GAN is a popular deep learning generation model using a generator and discriminator. In this study, data is generated using chemical Raman spectral library data. Afterwards, the designed background noise is added and trained on the baseline correction model using deep learning. Afterwards, it was applied to raw spectrum data and confirmed that it can be effectively applied to raw spectrum as well.

High slopes are susceptible to rainfall, weathering, earthquakes and human engineering construction, resulting in the originally stable high slopes are very prone to landslides and collapses, etc. In order to prevent such phenomena, in addition to optimizing the existing monitoring system, it is also crucial to establish a scientific and accurate prediction model for high slopes. In order to improve the prediction accuracy of high slope deformation data, it is proposed to establish a deformation data prediction model coupled with autoregressive integrated moving average model (ARIMA model), genetic algorithm (GA) and BP neural network. The model takes into account the linear and nonlinear parts of the data, and the ARIMA model is used for linear prediction, while the GA-BP algorithm regressively corrects the residuals of the ARIMA model prediction for nonlinear prediction. The results show that the residual-corrected ARIMA-GA-BP model has higher accuracy and extrapolation ability than other prediction models, and can be effectively applied in the research field of high slope deformation prediction.

Rolling bearings occupy a pivotal position as one of the most crucial and frequently used components in machinery and equipment. Traditionally, diagnosing faults in these bearings involved the reduction of the original signal's dimension and the extraction of a limited set of characteristic values from the vibration signal. However, this conventional approach often overlooked vital fault information embedded within the original signal. To overcome this limitation, this paper introduces a groundbreaking approach that seamlessly integrates the Improved Deep Convolutional Neural Network (IDCNN) with the Continuous Wavelet Transform (CWT) method. This innovative hybrid technique facilitates a comprehensive multiscale analysis of vibration signals, generating wavelet time-frequency domain graphs that effectively highlight fault features. Additionally, a three-channel multi-scale convolutional network is utilized to learn from these diverse fault features, enabling precise fault identification. To validate the superiority of this proposed model, experiments were conducted using the bearing dataset from Case Western Reserve University. The findings indicate that, in comparison to traditional neural networks and deep learning technologies, the model proposed in this paper demonstrates superior performance in extracting bearing fault features and diagnosing faults, ultimately achieving an average recognition accuracy of 99.50%.

Detecting wood surface flaws is a highly tough task due to the wide range of different types and sizes of defects, as well as the similarity between some defective and non-defective areas. To solve this problem, The proposed approach utilizes an enhanced version of YOLOv5s to detect defects on wooden surfaces. To begin with, a Channel Attention (CA) mechanism is incorporated into the main network in order to improve the exchange of information across channels and the capacity to identify faulty targets. Meanwhile, A DCN_Bottleneck is specifically engineered to enhance the model's capability to extract features from faulty targets that vary in terms of scales and shapes. The enhanced YOLOv5s algorithm demonstrates superior performance in comparison to the original YOLOv5s method. It exhibits a notable improvement in mean Average Precision (mAP) by 5.52 percentage points, Precision by 3.3 percentage points, and Recall by 3.8 percentage points. This indicates that the enhanced algorithm effectively enhances the accuracy of defect detection on wooden surfaces.

This report selects the rice dataset from UCL to compare the performance of several classic classification algorithms in the rice classification task, including linear discriminant analysis, logistic regression, K-nearest neighbor KNN classification, and naive Bayes classification. Through data preprocessing and feature engineering, we run naive Bayes classifiers under different prior distributions and analyze the classification results in detail. In addition, we select evaluation criteria such as accuracy, precision, recall, and F1 score to compare and discuss the effectiveness of each classification algorithm. The final results show that the choice of different prior distributions also has a certain impact on the classification results, and the classification effects of linear discriminant analysis, logistic regression, and Gaussian Bayes are better. This article details the experimental process and results analysis, providing some reference value for how to classify rice.

There are many factors related to sports performance, which brings many difficulties to the prediction of sports performance. To improve the prediction effect of sports performance, we propose a sports performance prediction method based on time series analysis. This method selects the singular value decomposition filtering algorithm to decompose the original data sequence into two parts: random components and trend components; The GM (1,1) model and SVM model were used to predict the random and trend components respectively, and the predicted results of the two models were fused to obtain the final sports performance prediction results. The practical application results show that our proposed method improves the accuracy of sports performance prediction and significantly reduces the error of sports performance prediction, which has certain advantages.

To address the issue of low diagnostic accuracy of convolutional networks for bearing faults under noisy conditions, a Convolutional Neural Network model (WMCNN) incorporating wavelet weight initialization and multi-scale attention is proposed. Firstly, considering the crucial role of the first convolutional layer in model robustness, a wavelet weight initialization layer is introduced to selectively filter features, imparting them with multi-scale characteristics. Subsequently, for extracting features at different scales, channel attention is employed to adaptively select channels containing fault features, thereby enhancing the model's noise resistance and suppressing noise interference. Additionally, adaptive one-dimensional convolution is utilized to adjust channel weights of features at different scales, facilitating adaptive fusion of multi-scale features. Finally, feature classification is performed through fully connected layers. Experimental results on bearing datasets demonstrate that, under varying signal-to-noise ratios and noise interference, WMCNN exhibits superior bearing fault diagnostic capability compared to other methods.

In order to reduce the power generation load of water-light-storage cluster and ensure the balance between energy storage and power supply, the mixed integer programming and mathematical model are introduced, and an analysis method of mutual regulation potential of water-light-storage cluster is designed. The overall structure of water-light-storage cluster microgrid is designed, and the mathematical models of hydro-generator set, photovoltaic power supply and energy storage battery are constructed. On this basis, the analysis characteristics of mutual regulation potential of water-light-storage cluster are extracted and analyzed, and the potential analysis is realized under mixed integer programming. The experimental results show that the power output of water-light-storage cluster mutual regulation becomes significantly more stable after the application of this method, and the load of hydro-generator, photovoltaic power supply and energy storage battery shows relatively stable values, which is more effective in optimizing cluster mutual regulation and can distribute and balance the load of various energy sources more reasonably.

In response to the issues of extremely unbalanced sample distribution, high-dimensional samples, and large volumes of data in the problem of credit card transaction fraud detection, this paper proposes a hybrid model that combines DBSCAN, SVMSmote, and Artificial Neural Networks (ANN), namely the DB-SVMSmote-ANN model. This model can generate samples of the minority class, addressing the issue of too few positive samples in credit card transaction fraud datasets. Subsequently, the balanced dataset is used to train the ANN classification model, solving the classification problem between fraudulent and normal transaction samples. In the experimental process of this paper, the differences between balanced and unbalanced datasets were first compared, followed by a comparison of the effects of using the DBSVMSmote-ANN and other classification models. Ultimately, the experiments demonstrated that the DB-SVMSmoteANN model can excellently solve the problem of credit card transaction fraud detection.

Enhancing the precision of predicting bus arrival times can improve passenger travel efficiency and bus service quality, and save bus operating costs. Long Short Term Memory Network (LSTM) models are diffusely applied to time series analysis of bus travel time prediction tasks, but they may have shortcomings in spatial feature research. To address this issue, a prediction model mixing Convolutional Neural Network (CNN) and Long Short Term Memory Network (LSTM) comed up, which uses CNN to draw spatial features from the data, then uses LSTM networks to learn time dependencies in time series data, and introduces Attention mechanism to enhance the performance of LSTM models. Finally, the effectiveness of the prediction model is verified using bus GPS data. The test results express that the model is provided with low error.

In engineering practice, multi-objective optimization problems (Multi-objective Optimization Problems, MOPs) are common, but multi-objective optimization problems usually cannot be solved directly. In order to solve the multi- objective optimization problem, the goals to be optimized are usually analyzed, and the problems to be studied are transformed into computable mathematical models by using mathematical theories and methods. At the same time, the mathematical model and calculation method are studied, and a more appropriate algorithm is selected to further solve the mathematical model established. After obtaining all feasible schemes that meet the research objectives, the best scheme is selected according to the needs of the research objectives. In the process of multi-objective function optimization, the problem of constraint complexity and high dimension is encountered, so the choice of the algorithm has a great impact on the accuracy of the solution. Therefore, in this paper, the hybrid algorithm of NSGA-II and WOA is adopted to solve the optimization scheduling problem, which is a mixture of NSGA-II and WOA. The hybrid algorithm of NSGA-II and WOA can obtain the optimal solution set in each calculation and select the optimal solution.

Human society has entered the era of big data, and data has become a production factor alongside labor, land, technology, and other things. Public data is an important part of data resources. Privacy computing can fully tap the potential value of the original public data under the premise of ensuring that it will not be leaked to the outside world, and promote economic development and social progress in the era of big data. However, there are still technical loopholes in privacy computing, and the relevant legal supporting measures are lagging, which may lead to technical and legal risks in practical application. In this regard, it is necessary to start from the technical principle of privacy computing, analyze the impact of this technology on the commercial utilization of public data, and put forward effective countermeasures.

Environmental sensors play a crucial role in modern society. They provide strong support for environmental protection, resource management, and sustainable development by monitoring and perceiving various parameters in real-time. With the development of big data and artificial intelligence technologies, intelligent management of environmental data has become possible. To enhance the level of intelligence in environmental monitoring, this paper studies the performance of environmental sensor data operations in data warehouses such as Hive, ClickHouse, and Doris, and constructs an environmental sensor data analysis system based on a real-time data warehouse. The system employs Flink for real-time collection of environmental data, ClickHouse for data storage, Spark for data analysis, and front-end and back-end technologies for data visualization. The system achieves automatic collection, processing, and analysis of environmental data, thereby improving the efficiency and accuracy of environmental monitoring.

With the surge of 5G technology, research on intelligent transportation systems has emerged as a forefront area, with vehicle detection technology at its core. This paper delves into vehicle detection utilizing YOLOv5, analyzing its performance through key metrics like training/validation losses, precision, recall, and mAP. The study highlights YOLOv5's superiority in vehicle detection, balancing swiftness with high accuracy and robustness. This work reinforces deep learning-based vehicle detection and presents valuable insights for future research endeavors.

Data has become an important factor of production, and the collection, storage and value sharing of data have become increasingly important. At present, there is a difficult problem of how to realize data value mining and sharing on the premise of ensuring data security. In order to solve these problems, this article conducts relevant research on the key technologies of the "data+cloud" power ecological data service model based on privacy computing and computing power network technology, and proposes to use privacy computing technology and computing power network to build a "data+cloud" power ecological data service realizes multi-party data joint modeling and analysis, thereby building a safer and more reliable data application solution.

Accompanied by the progress and development of artificial intelligence technology, the surface unmanned ship plays a key role in the unmanned control system, and plays a key role in data acquisition and hydrological survey. In the process of practical application, the stable operation of the unmanned ship will be greatly affected because of the complex operation and the unknown environment, so it is very necessary to predict its bits. Accurately predicting the movement of unmanned vessels is critical for navigation, safety monitoring and mission planning, and can significantly improve their stability and operational efficiency in complex waters. This paper introduces the variational modal decomposition (VMD) algorithm and proposes a combined prediction model (VMD-NGO-LSTM) based on the Northern Goshawk Optimisation Algorithm (NGO) and Bidirectional Long and Short-Term Storage Memory Network (BiLSTM), which is found to be highly accurate by processing the prediction of unmanned boat navigation data in Xiyuan River. It reflects a faster training speed when processing large unmanned boat data, and has a certain reference value for the position adjustment of unmanned boats.

To address the challenges of remote sensing images, including diverse target angles, intricate backgrounds, and overall clutter, we propose a remote sensing target detection method that utilizes attention mechanisms and feature alignment. First, a convolutional triplet attention mechanism is incorporated into the backbone network to improve the capture of contextual dependencies between adjacent regions in the feature map, thereby enhancing multi-scale feature learning from remote sensing images. Next, the detection head employs a rotational feature alignment network to create high-quality rotational anchors and aligned features for classification and coordinate regression. Finally, the Kernelized Focal IoU is used as the regression loss function for rotated bounding boxes, increasing the model's sensitivity to positioning accuracy and minimizing errors. The algorithm presented in this paper was evaluated on the large-scale remote sensing dataset DOTAv1.0, achieving an mAP of 0.786, and the experiments have demonstrated the effectiveness of the algorithm.

With the continuous growth of global energy demand and the urgent need for the utilization of renewable energy, the copyrolysis technology of biomass and coal has been receiving increasing attention. This study, based on the co-pyrolysis data of biomass and coal from the 9th Dimension Cup National College Mathematical Contest in Modeling 2024, uses the multiple linear regression model as a benchmark, combining LASSO regression and genetic algorithms to optimize the tar yield in the co-pyrolysis reaction of biomass and coal. LASSO regression is employed for variable selection and regularization to improve the model’s prediction accuracy and generalization ability. Additionally, genetic algorithms further identify the optimal ratios and the level of Insoluble N-hexane Substances (INS) that maximize tar yield. Through model optimization and the analysis of experimental data, the best mixing ratio is found to be approximately 51.7%, the optimal INS level is about 48.0%, and the predicted maximum tar yield is 0.275. To a certain extent, our model provide theoretical support for improving the energy conversion efficiency of the biomass and coal co-pyrolysis process.

With the increasing variety of cloud platforms and the rising demand for cloud resources, how cloud providers allocate their cloud resources has gradually become a trouble some issue. At the same time, under the existing cloud resource allocation strategy, the interests of cloud resource users have not been given much attention, which will lead to the emergence of SLF penalties, a lose-lose outcome. This article addresses the aforementioned issues through an optimization algorithm based on auction mechanisms, which aims to maximize the interests of both parties during the allocation of cloud resources to ensure a win-win outcome. At the same time, decision factors are proposed during the allocation process to ensure the timely completion of user tasks, thereby minimizing the generation of penalties.

Natural Language Processing (NLP) has undergone a remarkable transformation, with representation learning playing a pivotal role in reshaping the field. This review explores the evolution of NLP representation learning, from traditional feature engineering to modern techniques, focusing on distributed representations and deep learning methods. We delve into the significance of word embeddings like Word2Vec and GloVe, which have become instrumental in enhancing text processing. Furthermore, we examine common deep learning architectures, including Multilayer Perceptrons (MLP) and Convolutional Neural Networks (CNNs), highlighting their role in feature extraction from textual data. Additionally, optimization methods crucial for efficient deep neural network training are discussed. This review provides a comprehensive overview of the advancements in NLP architecture, emphasizing the enduring impact of prior research and the promise of future innovations.

The increasing growth of the aviation industry has been accompanied by increasing problems of flight delays. This research aims to predict flight delays by employing machine learning models to enhance the effectiveness of the aviation transportation system. Domestic flight takeoff and landing data from the United States in January 2019 and 2020 were used for the research, and data preprocessing was carried out. Four machine learning models (XGBoost, Random Forest, LightGBM, and Gradient Boosting Tree) were used for model training, followed by an assessment of their respective performances on a designated test set. The outcomes indicate that the XGBoost model outperforms others in accuracy and the AUC of the test set, followed by the LightGBM model, while the Random Forest model shows overfitting on the training set. This research offers a comprehensive analysis and resolution to the issue of flight delay prediction, which helps airlines and passengers to make rational decisions.

In this paper, RNN (Recurrent neural network) model in DL (Deep Learning) is adopted, and a large-scale RNN model is constructed by using historical power load data and other related information. The model is used to forecast the future power load, and applied to the power demand forecasting and marketing service platform. In addition, the selection and optimization methods of model parameters and superparameters are analyzed, which further improves the prediction effect of the model. Through the training and optimization of the model, the prediction accuracy and practicability are improved, and the marketing service efficiency and enterprise income are also improved. Through the experimental verification, this model can get a recall rate of 95.62%, which is better than the traditional CNN model. In addition, after many iterations, the accuracy of this model in power demand forecasting has obvious advantages over the traditional CNN model.

In vehicular edge computing environments, where resources are limited and task latency sensitivity is high, we propose a resource allocation method based on deep reinforcement learning. First, a vehicular edge computing system model is constructed, then an optimal resource allocation problem is derived using a gradient iteration method and KKT conditions. The resource-optimized MADDPG algorithm is employed to replace the original DDPG algorithm to reduce time delays caused by uneven resource allocation. The effectiveness of the algorithm is verified through simulation experiments, and its performance in actual traffic scenarios is analyzed. Simulation results show that our method has superior performance in reducing time delays compared to traditional methods.

Raman spectroscopy technology is widely used in various fields due to its advantages such as non-destructiveness, speed, and high sensitivity. To make effective use of this technology, preprocessing operations including additive noise reduction and baseline correction are usually required. Traditional preprocessing tasks involve the appropriate selection of parameters and methods. To address these challenges, we proposed using a multi-task deep learning network for preprocessing. This network is built on ResNet and can perform baseline correction and noise removal simultaneously. To train the deep learning network, we generate training data using mathematical methods to overcome the problem of data scarcity. We verified the superiority of our method compared to existing preprocessing methods using both simulated and real Raman spectral data.

In this paper, we study the problem of key node identification in social networks and further analyse the influence of nodes in Twitter networks. We introduce a variety of node importance evaluation metrics such as centrality metric, PageRank algorithm, including HITS algorithm and SALSA algorithm, and combine a variety of features such as user interaction frequency, posting timeliness and content quality, which are used to improve the prediction ability of the model. By applying graph neural network (GNN) data analysis tools, graph attention network (GAT) and graph convolutional network (GCN), which capture the complex relationships between nodes, the performance of the model and the reliability of the conclusions are enhanced by these armies.

In today's world, convolution neural networks possess tremendous potential and exhibit remarkable recognition rates in computer vision. However, neural network models are often so large that they are not well-suited for mobile devices, making model compression particularly important. Especially in scenarios with limited hardware resources such as the National Grid, model quantification is essential. This paper introduces an INT8 quantization method for convolution neural networks, which transforms the weights and input/output of the convolution neural network into the logarithmic domain through a non-linear formula. This method provides a better representation of the distribution of small values, resulting in less accuracy loss. Meanwhile, model calibration requires only a limited amount of image data, which reduces the time taken for model calibration while ensuring minimal accuracy loss. Experimental results show that our quantization method results in an accuracy loss of less than 1% on ImageNet and CIFAR.

This paper, addressing the lack of interaction between context and modalities and the deficiency in speaker dependency in dialogue sentiment analysis, proposes a multimodal fusion method based on cross-modal attention. This method leverages cross-modal attention to facilitate pairwise interaction of textual, audio, and visual modality information. Additionally, it introduces a Speaker Dependency Encoder (SDE) network that divides historical dialogues and speaker features according to the identity of the speakers. This network divides historical conversations and speaker features according to speaker identity, and uses attention mechanism to explore inter-speaker and intra-speaker dependence from the divided conversations and speaker features respectively. The above dependencies are then spliced to obtain the speaker dependencies at the current moment. A multimodal dialogue sentiment analysis network, Speaker Guide Emotion Network (SGEN), is proposed, which combines the modal fusion network CMA and speaker feature generation network SDE proposed in this paper to undertake the task of multimodal dialogue sentiment analysis.

The failure of electric valves represents a significant safety hazard in industrial systems. Traditional manual detection and regular replacement strategies are insufficient to address this issue. This study proposes an online verification method combining convolutional neural networks (CNN), two-way long short-term memory networks (BiLSTM) and multi-head attention mechanism (MHA), which is capable of identifying weak fault characteristics such as plugging and twitching caused by regulator wear. Transfer learning addresses the issue of data scarcity and enhances the model's adaptability in the target domain. The CNN-BiLSTM-MHA-TL model exhibited high prediction accuracy, with MAE, RMSE, and MAPE values of 0.4451, 0.5722, and 0.0132, respectively. This online verification method offers notable advantages and promising prospects for application in the field of valve health monitoring.

Currently, when performing pixel-level semantic segmentation of crop planting in high-resolution images, it is difficult for deep convolutional neural networks to simultaneously capture global features and local detailed features at multiple scales in space. This can lead to blurred boundary contours between different farmland plots, as well as lower integrity within similar farmland areas. In view of the above research content, a model that integrates Transformer and CNN is proposed to classify and identify crops in the study area, using drone remote sensing images from a competition as the data source. (1) The model adopts a multi-level skip-connected encoder-decoder network architecture. The encoding part of the model uses an improved MobileNetV2 as the front-end feature extractor to extract local detail features, and then inputs the extracted features into Vision Transformer for global feature capture and further processing, thereby capturing details while retaining global context information. (2) The decoding part adopts the design of UNet, extracting features from different levels of the encoding part and directly transferring them to the corresponding levels of the decoding part through skip connections to ensure the retention of detail information, and using upsampling layers to gradually restore the spatial resolution of the image. In an experiment on a public competition dataset, the experimental results show that the MIoU of the network proposed in this paper reaches 85.96%, the PA reaches 92.30%, and the Dice value reaches 0.922, which has the highest segmentation accuracy compared with the comparison network.

With the continuous increase in electricity demand, power grids face challenges such as peak load regulation, frequency adjustments, and power shortages. Demand-side response (DR) has emerged as an effective tool to balance electricity supply and demand by controlling user-side resources. This study proposes a deep learning-based prediction and optimization model to enhance the accuracy and effectiveness of DR. Using a convolutional neural network (CNN) and long short-term memory network (LSTM), we perform detailed load forecasting and extract resource parameters. Dynamic parameter identification enables online extraction and optimal management of these parameters. Experimental results demonstrate significant improvements in load forecasting accuracy and system reliability. Key contributions include integrating deep learning to model complex factors and resource characteristics, developing a CNN-LSTM-based load forecasting model, and creating a dynamic priority control algorithm for smart appliances based on the comfort index (KApp). This research supports stable power system operation and has significant theoretical and practical value.

The classical Feature Pyramid Network often results in the neglect of detail information due to information downsampling. Furthermore, the context dependency of the input area is homogeneous, which limits the accuracy of image segmentation. To solve the above issues, this article proposes a novel Feature Pyramid Network by incorporating feature complementarity and a linear attention mechanism. Reweighing single-scale features in grid fusion, the network utilizes cross-scale complementary knowledge to decrease the neglect of local details in the image. Additionally, a linear attention mechanism with a differentiable linear unit kernel function is leveraged to enhance the long-range pixel associations across the global image scope. This mechanism adaptively allocates attention weights with a lightweight structure, especially emphasizing critical information such as object boundaries for the enhancement of the proposed model's accuracy and robustness. After testing, the overall accuracy of segmentation is up to 92.7% and 92.9%, respectively, which proves that the proposed method can significantly improve the segmentation performance.

This paper discusses the importance and challenges of thin film uniformity inspection during wafer fabrication, and proposes a new solar wafer thin film uniformity inspection method that combines the improved DeepLab v3+ model and infrared thermal imaging technology. The images of wafer surface temperature distribution after deposition in PECVD process recorded by FLIR T860 equipment and image preprocessing are utilized to constitute the infrared imaging dataset of the wafer surface; based on the original model of DeepLab v3+, the backbone network is replaced with MobileNet v3, and the BAM attention mechanism is added to enhance the computational efficiency and feature extraction capability. The improved model significantly reduces computational complexity and memory consumption. Experiments are conducted with the wafer surface temperature distribution images recorded by FLIR T860 device as the data source, and the experiments show that the improved DeepLab v3+ model outperforms the original DeepLab v3+ model in all evaluation indexes, and the accuracy, MIoU, and detection time reach 91.2%, 85.3% and 25ms/image, respectively.

Cryo-electron microscopy (cryo-EM) has become a crucial tool for determining the structures of proteins and macromolecular complexes. However, the extremely low signal-to-noise ratio (SNR) and large image scale(approximately 4K) of electron microscopy images pose significant challenges in selecting hundreds of thousands of particles from these images for reconstructing high-resolution protein three-dimensional structures. Existing particle picking methods still struggle to meet the requirements of such research. We present a deep neural network model that amalgamates convolutional neural network (CNN) and graph convolutional networks (GCN). The model initially employs CNN for preliminary particle picking, modeling protein particles as the center points of protein particle bounding boxes and leveraging keypoint detection to detect as many particles as possible. Subsequently, the detected particle boxes are cropped from the original electron microscopy images and used as inputs for the GCN for further classification, aiming to distinguish between the background and protein particles to achieve higher precision. Experimental evaluations were conducted on various datasets, and the results indicate that it can acquire accurate results.

In order to improve the accuracy and interpretability of catalyst active site identification, multi-scale graph neural network was used to study the application of graph convolutional network and attention mechanism in capturing catalyst surface structural features. The effect of fusion of different scale information on the model performance is analyzed. The results show that the multi-scale graph neural network outperforms the traditional method in all evaluation indexes, especially the F1 score, which shows its potential in catalyst design and optimization.

The accelerated construction of smart grids has imposed higher demands on the lifecycle management of power grid projects. Addressing the issues of low efficiency, high costs, and data inconsistency in the design, construction, and operation phases of current power grid projects, this paper proposes an optimized method based on Domestic BIM Technology. This method integrates Principal Component Analysis (PCA) and Monte Carlo simulation. By utilizing PCA to extract key indicators and best practices, and consolidating design, construction, and operational data throughout the entire lifecycle, the consistency and accuracy of the data are ensured, and construction processes are optimized. Furthermore, Monte Carlo simulation is employed to simulate and optimize different load distribution schemes. The methods discussed in this paper significantly enhance the management efficiency and quality across all stages of the lifecycle management of power grid projects, providing robust technical support and assurance for the intelligent management of power grid projects.

In this paper, an extended access control mechanism is proposed for controlled sharing of data after data flow in complex network environment, which provides more secure, efficient and personalized data access methods, ensuring that users can flexibly obtain data that meet their requirements. The proposed control mechanism is divided into two categories: constraint control and propagation control. Among them, constraint control solves the problem of access authorization of data before access request by the access request entity, and propagation control is used for extended control of data after data leave the data center. The proposed mechanism realizes direct and indirect access control of data, and takes the whole life control of electronic invoices as an example to show the implementation method of the proposed mechanism.

With the global climate change and the intensification of human activities, the frequency and the degree of damage of flood disasters have significantly increased. In order to prevent and control flood and waterlogging effectively, it is very important to obtain water area information timely and accurately. Remote sensing technology can provide a wide range of real-time water monitoring data, but in the complex water boundary identification and change detection, the traditional methods still have limitations. In this paper, a method of water remote sensing image extraction based on improved Mask-RCNN network is proposed, and its application in flood disaster prevention and control is explored. By introducing the attention mechanism of CBAM, the Mask-RCNN network is optimized, and the precision and efficiency of water area recognition are greatly improved. The experimental results show that the improved model is superior to the traditional U-Net model in mIoU, Recall and Precision, which provides technical support for flood disaster monitoring and early warning.

In recent years, deep learning has made remarkable strides, significantly propelling the field forward. However, traditional CNNs still face limitations in handling multi-scale information and capturing global features. To address these issues, we propose the Multi-Layer Self-Attention Fusion Pyramid Network (MSAPN). MSAPN combines ConvNeXt, feature pyramids, and self-attention modules to enhance the capability to integrate multiscale features and model global dependencies. Our network demonstrates superior performance across multiple public datasets, showcasing significant improvements in classification accuracy.

The images have their own defects such as high noise and poor texture, a multi-input fusion technique has been developed for enhancing underwater images, addressing issues of degradation and significant interference. By analyzing and comparing the performance of white balance, histogram equalization, gamma correction and underwater dark channel prior methods, a new image enhancement algorithm has been developed through the integration of the best features from various algorithms. The proposed algorithm is evaluated against traditional methods for enhancing underwater images, and the image is analyzed by image effect and image quality evaluation index. The outcomes demonstrate that the suggested algorithm significantly enhances the quality of underwater images.

Computed Tomography is a crucial diagnostic tool in medicine but exposes patients to harmful radiation. Low-dose CT reduces radiation but leads to noisy images. Super-resolution technology can reconstruct high-resolution images from low-resolution scans, enhancing image quality while reducing radiation. This study proposes a novel deep learning model, CACTSR, which integrates VMamba and Transformer technologies with Mixed Attention Blocks and Cross Attention Blocks to enhance feature utilization and facilitate cross-window information interaction. Experimental results on the QING LUNG dataset demonstrate that CACTSR surpasses existing methods in terms of image quality metrics, generating images with crisp edges and abundant details. This innovative approach effectively mitigates block artifacts and enhances image quality, providing a powerful solution for reducing radiation doses in clinical CT imaging while maintaining diagnostic accuracy.

Evaluating the aesthetics of hard pen calligraphy requires capturing both fine details and overall composition. To address this, we introduce MSRF-Net, a model that integrates a MultiScale Fusion Block (MSFB) and an Adaptive Reweighting Block (ARB) within a U-Net architecture, enhanced by meta-learning. MSRF-Net effectively captures and reweights multiscale features, improving classification accuracy. Experiments on a dataset of 19,867 calligraphy samples demonstrate that MSRF-Net surpasses state-of-the-art models, offering a robust solution for aesthetic evaluation. Future work will explore its application to other perceptual evaluation tasks.

To overcome challenges related to occlusion and varying lighting conditions in cherry detection, we propose an enhanced YOLOv8-based model for cherry recognition. We have modified the YOLOv8 model’s ‘neck’ section by incorporating a BiFPN-based pyramidal network, which enhances feature fusion across various levels, thereby improving the model’s ability to identify partially obscured cherries. Additionally, to enhance adaptability to diverse lighting and to better detect small targets, we utilize an AFPN to refine the detection head, thereby improving the model’s performance across varying lighting scenarios. For training and validation purposes, we created a custom dataset of cherries that includes instances of occlusion and complex lighting. Our experimental findings indicate that the enhanced algorithm outperforms the original YOLOv8 network, with improvements in accuracy, recall, and mean Average Precision (mAP) by 4.2%, 2.6%, and 2.3%, respectively. Notably, the improved model demonstrates greater robustness in handling occluded cherries and complex lighting situations. The model achieves high recognition accuracy and robustness, offering valuable insights for future research in cherry recognition.

To address the challenge of foreign impurity (namely NTRM, non-tobacco related material) detection in the industrial tobacco production line, such as the difficulty of manual identification of NTRM and visual fatigue, we design an online foreign impurity detection system, which is mainly based on compressed sensing, combined with deep space-spectrum fusion network, utilizing CNN-based deep learning algorithms for real-time online detection of NTRM in tobacco leaves. And then we provide the detection results and NTRM location information to workers via a human-machine interface, to assist in the manual NTRM removal process. Experimental results show that for the selected NTRM targets in this study, the average detection accuracy is as high as 85.4%, with a missed detection rate as low as 5.5%, demonstrating effective assistance in manual NTRM removal.