This PDF file contains the front matter associated with SPIE Proceedings Volume 13176, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

In order to solve the problem of feature point registration and local alignment in parallax image stitching, we proposed an image stitching algorithm based on grid refinement and deviation correction, and optimizes feature points through random sample consensus (RANSAC) and normal distribution theory. Refine the grid according to the distribution of matching points, use the moving direct linear transformation (MDLT) to calculate the local homography matrix and combine with the global optimal similarity matrix to complete the grid distortion, and apply the thin plate spline (TPS) theory to correct the local projection deviation. To achieve better alignment of images in local overlapping areas. The experimental results show that the method in this paper has obvious advantages compared with other advanced algorithms in stitching quality and stitching speed, which proves that the algorithm in this paper has the feasibility of practical application in image mosaic work.

This paper aims to evaluate the economic and safety performance of a ship’s lock head under maintenance conditions. The geometric and mesh descriptions of the lock head are presented for finite element analysis. A multi-objective structural optimization model is established based on the ship lock head volume and maximum principal tensile stress. Thirty-four parameters are distinguished as design variables for optimization. There are three types of constraints considered in simulation calculations. The whale optimization algorithm is then used to deal with the structural optimization problem. Design variables are continuously updated in positions through the exploitation and exploration stages to obtain the optimal solution. By way of example, the optimal design’s volume is reduced by 7.3% compared with the initial design. In particular, the tensile strength of the concrete material is fully utilized.

An artificial intelligence recognition method for multivariate temporal abnormal behavior has been proposed for abnormal behavior detection in electricity safety scenarios. The weighted method based on the target pixel center is used to identify the work clothes, and the computer vision technology in the field of artificial intelligence is used to analyze and process the images, so as to ensure that the workers in the electricity safety scene wear safety helmets. Based on the identification results of work clothes and safety helmets, a "AI+" based electricity safety scene operation control model is constructed. The identified and processed results are integrated to generate abnormal information and alarms, in order to improve the safety of electricity scenes. The experimental results show that the minimum mean square error of the proposed method's recognition results is 4.08, and the minimum miss rate is only 0.8%, indicating that the recognition results of this method are both accurate and comprehensive.

In order to address the challenges of high false alarm rate, slow modeling speed, and poor real-time performance in flame detection algorithms under complex scenes, a lightweight and efficient two-stage video flame detection algorithm was designed.In the first stage, an Adaptive Gaussian Mixture Model (AGMM) is utilized to rapidly build the background model of the video image sequence and extract suspicious candidate regions from the sequence.In the second stage, a Deep Normalization and Convolutional Neural Network (DNCNN) is employed to discriminate the suspicious candidate regions, achieving accurate localization of flames.Compared to traditional classification algorithms, the proposed two-stage video flame detection algorithm effectively overcomes environmental disturbances in complex scenes, accurately and rapidly identifies flames, and demonstrates higher detection rates and adaptability.

Left boundary flight training of trainer aircraft is usually accompanied by higher flight risk, therefore, it is very necessary to study flight risk identification of left boundary flight action. Based on flight training outline and expert decision-making, flight risk of left boundary is divided into three levels: low, medium and high. Considering strong independence of risk identification features, by using naive Bayes algorithm and typical flight parameters such as altitude, velocity and pitch angle and so on, probability distribution results of flight parameters risk evaluation indexes corresponding to different risk levels are obtained to form risk labels. According to probability distribution of risk evaluation index, actual flight risk grade of left boundary flight action is identified by means of maximum likelihood method.

At present, objectives matching has become an essential method in face recognitions, images stitching and autonomous driving. Existing objectives matching process is almost relied on the scale invariant transform, which ignores the flexibility information of ORB (Oriented Fast and Rotated Brief) features. In this work, we propose a novel objective matching method by deploying an improved ORB features extraction module and decrease the computation costs. Specifically, we detect the features in multiple scales of input images and utilize the features from accelerated segment algorithm to extract the ORB features, which can guarantee the uniform distribution and scale invariant characteristic of extracted features. Subsequently, we adopt the similarity algorithm to match the extracted features and novel images to achieve the objective matching. From our extensive experiment and analysis, we can significantly observe that our model achieves the highest matching accuracy with acceptable computation costs through comparing with existing methods.

Multiscale Dense Network (MSDN) is a deep convolutional neural network method proposed to solve the loss of fine features in hyperspectral remote sensing images. Multiscale Dense Network (MSDN) makes full use of the different scale information in the network to realize the feature extraction of HSI in both horizontal and vertical dimensions. Since MSDN maintains all fine scales until the last layer of the network, the computation is large. To address this problem, this paper proposes a hyperspectral image classification method based on lightweight multi-scale dense network (LMSDN). It uses dividing the MSDN network into S blocks along the horizontal dimension, keeping the coarsest (S-i+1) scale only in the ith block, and only down sampling the final output of each block. Based on the LMSDN network, in order to maintain the classification accuracy of the network, the 3D Gabor filter is combined with the LMSDN, and the 3D Gabor filter is first used to extract some internal structural information such as orientation and size of the hyperspectral remote sensing image, which is then sent to the LMSDN network for the classification of HSI. Experiments are conducted on two publicly available datasets, and the experimental results are compared with other methods, which proves that the Gabor-LMSDN network has less computation and shortens the experimental time compared with the MSDN network while guaranteeing the classification accuracy.

Convolutional neural network (CNN) is recognized and widely applied to classify hyperspectral remote sensing images (HSI). For the existing hyperspectral image classification algorithms, the network structure is simple, the generalization performance of the feature extraction model is poor, and the training samples are not uniform, this paper proposes a multi-scale multi-channel convolutional neural network (MMC-CNN) model based on the feature fusion. Firstly, the data is divided into two scales 𝑿_𝟏 × 𝑿_𝟏 pixel module and 𝑿_𝟐 × 𝑿_𝟐 pixel module, and then the two module channels are used again for feature extraction respectively, and three-dimensional convolution is used as the feature extractor in each channel. The experimental results show that the present model has a significant improvement in classification accuracy on commonly used HSI datasets, which is better than previous algorithms, thus verifying the high accuracy of the feature extraction of the present model.

This study introduces a scalable self-attention based contrastive image classification algorithm, aiming to train a backbone network suitable for downstream image classification tasks through contrastive learning. It begins by addressing the challenges of high training overhead and low model accuracy encountered when applying the contrastive learning model represented by Moco V3 to image classification. Subsequently, the Momentum Contrast For Image Classification (MocoIC) algorithm is proposed as a solution to these issues. The MocoIC algorithm comprises the Scalable Shift Window Transformer (SSWT) feature extraction network, feature mapping network, prediction head, and Info NCE loss function. This model exhibits superior algorithmic performance and requires less training overhead compared to the original contrastive learning model.

Artificial intelligence will play an important supporting role in urban traffic congestion management and smart transportation work. In recent years, China has vigorously developed infrastructure construction, and intelligent transportation system equipment such as cameras, microwaves, ultrasound, and electronic police have basically covered urban main roads, highways, and national and provincial roads. On top of a relatively complete transportation infrastructure, a large amount of traffic data is generated through various devices and technologies. With the support of artificial intelligence technology, reasonable driving speeds and traffic flow scales are set, and the duration of signal traffic lights at different road intervals is calculated. This enables efficient passage of vehicles within the corresponding road sections, greatly improving the efficiency of road travel in the entire city.

The prediction result of Remaining Useful Life (RUL) of aero engine determines the timing of engine maintenance according to the condition, which is of great significance to the operation safety of the engine. In order to improve the prediction accuracy of aero-engine residual life, a chaotic genetic algorithm optimization sequential convolutional network (TGA-TCN) based residual life prediction method is proposed. The time dependence relationship of time series data is constructed based on time series convolutional network, and the optimal network structure is constructed by genetic algorithm and hyperparameter design. A turbofan engine degradation dataset (C-MPASS) is used to verify that the prediction accuracy of the model is significantly improved than that of CNN, LSTM, TCN, GA-TCN, etc.

In modern power systems, the condition of electrical equipment is vital for maintaining power grid stability and efficiency. This paper presents the construction of a comprehensive knowledge graph for power quality and electrical equipment, derived from Chinese web sources. This knowledge graph facilitates real-time monitoring, predictive maintenance, and informed decision-making for electrical equipment management. The study covers core technologies, including data acquisition through web crawling using Selenium, entity recognition using BiLSTM-CRF, and relation extraction using R-BERT. The practical architecture includes ontology design, data pre-processing, entity recognition, relation extraction, and knowledge graph construction. SPARQL queries enable complex analysis, and applications support power quality monitoring and equipment health assessment. The knowledge graph serves as a foundation for understanding power quality and equipment status in a Chinese context, emphasizing the relationship between power quality and equipment failure. Ongoing efforts will expand and refine the graph to enhance its value for power quality-based equipment status assessment.

In recent years, the frequency of fog, haze and other bad weather phenomena has increased, and the suspended particles in the fog directly affect people's work in acquiring clear images from outdoors. The degraded image acquired on foggy days, the quality of which is shown as blurred, seriously affecting people's work on the identification and extraction of image feature information. Clear defogging of foggy images has an important research value in the fields of traffic monitoring, military and civilian aerial photography, target tracking, remote sensing satellites and so on. For many existing image defogging methods generally have long time-consuming, low efficiency and prone to distortion and other problems, this paper is based on the mechanism of deep learning technology, the use of conventional convolutional neural network model defogging method to carry out in-depth research and discussion, proposed a multi-scale convolutional neural network based on the improvement of image defogging algorithm.

As a new generation of electrical insulation medium, SF₆ gas is widely used in electrical insulation and arc extinguishing in high-voltage power equipment-Therefore, the accuracy of its concentration measurement is very important, which has an important impact on improving the performance and safety of power equipment. The paper elaborates particle swarm optimization (PSO) algorithm combining the application effect of radial basis function (RBF) network and backpropagation (BP) network in predicting sulfur hexafluoride (SF₆) gas concentration prediction. In order to avoid the factors of insufficient data and obvious data characteristics, a large number of data sets with different trends are randomly generated for model verification. The performance of the two methods is verified experimentally, and the results show that the PSO-RBF method performs better in predicting SF₆ gas concentration, by which the changes of the gas concentration can be predicted more accurately, and shows robustness for prediction under different conditions. In addition, the PSO-RBF method converges faster in the training of temperature compensation model, which improves the efficiency of the prediction model. It has practical application value for the prediction and monitoring of power equipment, and also provides new solutions for other similar gas prediction problems.

Aiming at some practical problems such as low inspection efficiency and high false detection rate of traditional manual transmission lines, we introduce an enhanced YOLOV7-based algorithm for infrared insulator target detection. The algorithm uses K-Means ++ clustering algorithm to regenerate the anchor frame on the self-made infrared insulator data set, which makes the algorithm converge faster. The algorithm substitutes the original YOLOV7 backbone with Efficientformerv2 to simplify the backbone network. Additionally, it eliminates redundant information from the feature map by incorporating the GhostConv module in the feature fusion network, and introducing the attention mechanism called SimAM. WIoU is used as the loss function of the improved model. Experiments on self-made infrared insulator data set show that when compared to the initial YOLOV7 algorithm, the refined algorithm demonstrates notable enhancements across multiple performance indicators. Particularly, the mAP@0.5 metric has seen a substantial improvement of 5.2%, elevating it to an impressive 98.3%. Concurrently, the number of parameters has experienced a significant reduction of 43.5%, dwindling to 20.5 million, and the computational workload has been markedly lowered by 66.2% to 35.0 GFPLOS.

This paper presents a Yolov5s-I algorithm to enhance the helmet and cigarette detection model in power construction. This study focuses on improving the performance of the popular YOLOv5s object detector in detecting safety helmets and cigarette, specifically in Electricity power construction. We investigated how we can modify the YOLOv5s model to make it better at detecting these small objects. We observed that by replacing certain parts of the model and adjusting their connections and settings, we can enhance its ability to detect safety helmets and cigarette. This model show an improvement in detecting smaller objects with a 50% Intersection over Union (IOU) threshold. Our main objective is to provide insights for future research on how to adapt popular detectors like YOLOv5 for specific tasks, such as detecting safety helmets and cigarette smoking in power construction sites. By making these changes, autonomous systems used in such settings can have access to more detailed information, thereby improving their effectiveness and safety.

In the traditional security field, user data is collected by different monitoring points, which can not be shared effectively, resulting in the phenomenon of "data island". To this end, this paper designs a privacy-efficient federated learning method for the field of intelligent security. By using a horizontal federated learning framework, under the premise of ensuring that the data does not go out of each region, the distributed monitoring points across the domain are jointly trained to ensure data privacy and security. In order to ensure the security of gradient data transmission in the process of joint modeling, a privacy protection protocol based on additive secret sharing was designed to effectively prevent gradient leakage attacks and enhance the efficiency of calculation. This paper proposes a communication compression strategy based on adaptive bidirectional TOP-K gradient sparsification, which greatly reduces the communication overhead in the joint modeling process while ensuring the model training performance. Our experiments on COCO dataset and Gunknipe-dataset dataset show that our method can not only compress communication overhead and improve training efficiency, but also effectively prevent gradient leakage attacks and ensure the security of transmitted data.

Obstruction is one of the significant sources of errors impacting the quality of weather radar observations. During the radar's operation, it encounters obstructions from terrains and tall buildings. These obstructions interfere with the radar signal's propagation and reception, leading to incomplete or inaccurate meteorological data in certain areas. In this paper, we introduce a novel approach, termed DOR, which initially detects obstruction masks and then recovers these regions using deep learning. Specifically, we incorporate DenseASPP into the final layer of the Unet encoder network, enriching the feature maps post-convolution with comprehensive spatial-semantic information. This design ensures that the encoder extracts highly-correlated contour detail features from the lower layers, achieving precise segmentation and localization of the target. Subsequently, we utilize generative adversarial networks (GAN) to recover the obscured areas. Extensive experiments on real weather radar images demonstrate that our proposed DOR method can effectively detect and recover.

Accurate and fast detection of targets in remote sensing images has always been a challenge due to the small size and diverse scales of the targets in the images. In order to improve the detection accuracy and shorten the detection time, this paper proposes a remote sensing image target detection method based on the CA attention mechanism and the lightweight network Slim-neck, which improves the target detection ability of YOLOv5s network. This paper focuses on enhancing the neck portion of the YOLOv5s. First, to improve the network's capacity for feature learning, multiple CA attention mechanisms are added to Neck. Secondly, the GSConv and VoVGSCSP modules are introduced to replace the last Conv and C3 modules in the neck. These improvements have allowed the network to increase detection speed along with detection precision. Experiments on the dataset DOTA-v1.0 show that the improved YOLOv5s model increases the detection precision by 3.5% and reduces the computational complexity by 2.5%.

Semantic segmentation is a typical task in computer vision. To address the issue of poor performance for small objects in high-resolution images by existing semantic segmentation networks, a method based on a multi-task network is proposed in this paper. This approach jointly learns relevant tasks, enhancing the accuracy of small object segmentation. The methodology of upsampling in semantic segmentation has been refined through the amalgamation of sub-pixel convolution and conventional convolution techniques, ensuring the meticulous preservation of intricate details. Additionally, a sophisticated post-processing module has been introduced, effectively enhancing the accuracy of small object edge segmentation. The empirical results highlight the effectiveness of this approach, significantly enhancing the segmentation quality of small objects in high-resolution images, while simultaneously ensuring real-time computational efficiency.

With the rapid growth of mobile internet and self-media, there is an increasingly urgent need to understand the emotional states of characters in images. Traditional methods primarily depend on facial expressions to infer emotions in characters, which does not sufficiently address issues such as face occlusion or invisibility in self-media images. The algorithms proposed in some recent studies, which utilize contextual context for emotion recognition, do not reach adequate extraction and fusion of contextual cues. This paper puts forth an algorithm for emotion recognition of image characters, which is based on multilevel association relations, to tackle the aforementioned issues. The algorithm examines the association relationships of the target character at the levels of association relationship acquisition, feature extraction, and feature fusion to recognize character emotions. At the level of acquiring association relationships, the algorithm enhances the relationships between scene information through the Anti-Mask of the original image and the semantically segmented image. In addition, it augments the character's U2NET semantically segmented image to obtain association relationships for character interaction in the image. At the feature extraction level, the algorithm inputs the image of the target character and each related image to separate networks for feature extraction. At the feature fusion level, the algorithm uses the Split Attention module to replace the fully connected layer for deep feature fusion. This method effectively leverages the associative relationships in self-media images, enhancing the accuracy of emotion recognition in such images. The experimental results indicate that the average accuracy on the emotion dataset EMOTIC has improved by 3.64 percentage points compared to the baseline algorithm, reaching 31.02%.

Cloud-assisted machine learning and analysis rely on cloud computing for various data processing services. Usually, data providers protect their data’s privacy using encryption before outsourcing, but the resulting encrypted data are known to be of poor utility. To this end, existing works combine differential privacy with public key cryptography to support multiple user analyses and queries of outsourced data. However, with the emergence of quantum computing, the schemes mentioned above based on the traditional hardness assumption will be threatened. This paper proposes quantum-secure outsourcing of privacy-preserving data publishing schemes. In particular, we develop a highly efficient proxy re-encryption scheme based on the ring learning with errors problem. This scheme allows the cloud service providers to return only the encrypted data that satisfies the user’s query without decrypting it, and the encrypted results can be decrypted using the user’s key. Meanwhile, a novel noise addition mechanism is integrated into the proxy re-encryption scheme, enabling cloud service providers to achieve a dataset sanitization procedure that supports the requests from multiple data users and allows providers to go offline after uploading their datasets. Finally, we present a detailed theoretical analysis and report an experimental evaluation of the real dataset.

In previous voxel-based object detection models, the Point Cloud Feature Extraction module (VFE) overlooked the uneven distribution of points within voxels, and as a result, it did not address the issues caused by non-uniform sampling of point clouds. Recognizing this limitation, this paper introduces a new module, VFEPlus, to enhance the feature extraction process. VFEPlus utilizes kernel density estimation and nonlinear transformations to extract the inverse density factors from the point cloud data. This allows for the extraction of point cloud features with density information, thus improving the detection performance that was previously compromised by the non-uniform distribution of point clouds. Experimental comparisons on publicly available datasets demonstrate that this module achieves higher accuracy.

In order to better analyze and process National Day highway travel trajectory data, a method and its calculation method for measuring the similarity between trajectories were proposed. Firstly, a new method for measuring trajectory similarity, namely TGI_GED, was introduced in this paper. Secondly, a graph neural network model was employed to enhance the computational efficiency of TGI_GED, enabling the calculation of TGI_GED between complex trajectories. Finally, this method combined clustering algorithms to further explore the characteristics and patterns of trajectories. Experimental results indicated that the graph neural network model not only significantly improved the computational efficiency of TGI_GED but also allowed for the calculation of TGI_GED between complex trajectories. Additionally, clustering results suggested that TGI_GED was more suitable for measuring the similarity between vehicle trajectories than graph edit distance. TGI_GED facilitated the discovery of latent patterns in trajectory data, offering a fresh perspective and solution for research in the field of vehicle trajectories.

Over the past year, the popular deep learning-based low-dose CT image enhancement methods mostly use the encoder-decoder structure. This structure expands the receptive field of the network by successive downsampling for deeper features. However, continuous downsampling leads to the loss of spatial information in the image, the excessively long distance between downsampling and upsampling also causes reconstruction errors. To solve the problems, this paper proposes a dense transformer network for low-dose CT image enhancement. This network uses dense connection as the basic architecture, and uses Up/Down-Tokenization in the Transformer module instead of regular up/down-sampling operations. In addition, by adding Tokons2Tokon modules between two consecutive multi-headed self-attention modules, improve the global modeling capability of the network and reduce computation. A channel fusion module is designed to replace the MLP in the standard transformer to introducing local information for reconstruction. Experiments on the AAPM LDCT dataset show that the proposed network can retain more texture details while suppressing noise compared to existing networks.

At present, image processing has become an important method in tremendous downstream applications including unmanned driving, facial recognition, medical detection, etc. However, existing processing methods are concentrated on single targets and leak the ability of disposing multiple objectives especially in complex environment. As far as we can concern, current methods have unacceptable classification accuracy in multiple objectives scenario due to the confounding influence from similar pixels information. In this work, we propose a novel framework to identify multiple targets by extracting features from feature space and not rely on the pixel difference information. Initially, we utilize a supervised contrastive learning (S-CTL) to encourage the feature representation of input images. Subsequently, multiple convolutional neural networks are trained to classify the labels of multiple features with attention mechanism. From our extensive experiment analysis and results, we can significantly observe that our model can outperform existing multi-objectives classification methods with satisfied accuracy and reasonable computation costs.

This paper proposes a 3D object detection algorithm that combines dynamic convolution with adaptive pooling as extraction (PAC-PAPRCNN). The algorithm is divided into two stages. The first stage generates 3D feature proposals in a bottom-up manner, and the second stage refines the proposals generated in the first stage. Firstly, in the first stage, a new type of dynamic convolution is used to adaptively learn the location features of points. To classify foreground and background points in the pooling layer, adaptive pooling theory is applied. Meanwhile, the extracted foreground points are used as regression boxes one by one, and then the regression box with the best score is selected by non-maximum suppression. Secondly, in the second stage, based on the first stage, granularity is refined to obtain a 3D regression box of revised specification, thus achieving 3D object detection. Finally, a series of experimental results on the KITTI dataset demonstrate that the proposed method is reasonable and effective.

Network clustering for protein complex identification within protein interaction networks has long been a focal point for researchers in the field of machine learning and data mining. Although existing algorithms take into account the dynamic properties of protein complexes, they ignore the computational dynamics. This paper introduces a novel network clustering algorithm designed for dynamic protein complex detection. This approach integrates gene expression data and static network data to construct a dynamic network construction model and incorporates multiple data sources to formulate a node representation model. It devises a new dynamic clustering model that simulates the splitting, merging, and growth of protein complexes, providing a framework for dynamic network clustering. Experimental results demonstrate that the proposed algorithm outperforms dynamic clustering algorithms and traditional clustering algorithms.

To address the challenge that the automatic extraction of bottom line is prone to interference from external factors, resulting in inaccurate detection and tracking of the bottom line in practical applications, this paper introduces the unsupervised learning expectation maximization algorithm (EM algorithm) into the peak detection and bottom line tracking stages of side-scan sonar images. An automatic extraction method of bottom line based on the EM machine learning algorithm is proposed. In order to evaluate the ability of the proposed method, experiments were conducted in simulated water environment. The results of the experiments show that the application of the EM machine learning method for automatic extraction of bottom line effectively addresses challenges such as occlusion by suspended particles and poor sea conditions. The experimental concludes that the proposed method achieves accurate detection and extraction of bottom line in complex environments, and exhibits promising practical applications.

Image registration is one of the essential pre-processing steps in remote sensing data applications such as change detection, image fusion, and so on. Since SAR images and optical images differ in several ways but have strong complementarity, a new registration approach based on image cross-domain translation is proposed. The image cross-domain translation was realized on the foundation of the diffusion model which had a strong ability for image generation. First, the unconditional diffusion model was trained to generate SAR-like images. Then, a novel condition was introduced to the inverse process to convert the optical image into the corresponding SAR-like image via a generative network. Finally, RIFT is used to extract and match the features between the generated SAR-like image and the actual SAR image to achieve the registration between the original optical image and the actual SAR image. The experimental results show that the diffusion model-based generative network is stable in training and convenient in controlling the conditional parameters, which can better solve the problem of radiometric discrepancy between SAR images and optical images.

Focusing on the issue that global path optimization and real-time obstacle avoidance cannot be combined while employing the A* algorithm or dynamic window technique alone in mobile robot navigation. A suggested mobile robot navigation algorithm combines the dynamic window approach with the enhanced A* algorithm. First, in the improved A* algorithm, the cost function is adaptively adjusted to improve the search efficiency of A* algorithm by introducing environmental information. In order to plan a global path with just key nodes, redundant nodes are removed and necessary path nodes are retained using a key node selection approach. Then, the evaluation function of dynamic window approach combined with key nodes information is constructed. The dynamic window approach is applied to plan the local path with the key nodes as the intermediate target points to improve the smoothness of the path, so as to achieve the global path optimization and real-time obstacle avoidance function. Finally, the viability and validity of the suggested navigation method is verified by demonstrating both simulation and real-world experiments simultaneously.

As a typical environmental target, the staircase area needs to have the function of detecting and recognizing stairs, whether it is a robot system that can climb stairs independently or a software system that can remind visually impaired people to pay attention to obstacles. A staircase region detection algorithm with dual attention mechanism and bidirectional cross scale feature fusion is proposed to address the issue of false detections and missed detections in current staircase images due to their similar shapes to areas such as tiles, tables, and chairs in reality. Firstly, a dual attention mechanism was introduced into the backbone network of Yolov5s to extract semantic dependencies between different dimensions, eliminate indirect correspondence between channels and weights, and improve detection accuracy by focusing on differences in similar samples. Secondly, a weighted bidirectional feature pyramid network structure is adopted in the feature pyramid to achieve efficient bidirectional cross scale connections and weighted feature fusion, enhancing the network's feature extraction ability. The experimental results show that compared with the original algorithm, the improved algorithm has mAP@0.5 Improved by 1.3%, with higher detection accuracy. In addition, we also conducted experiments on staircase data collected under different lighting conditions to verify the effectiveness of the algorithm.

The transformation of all-optical networks is an important task for communication operators to meet the needs of the new diversified comprehensive information services. The service experience of users was affected by quality of the network directly, network operation and maintenance is an important task of operation. This paper focus on the application of machine learning technology in network operation and maintenance quality analysis; automatically determine fault points by training historical network performance indicators and status data. The method proposed can predict the location of faults, improve the quality and efficiency of PON maintenance and reduce daily maintenance workload. Through practical, the solution proposed has demonstrated feasibility and can provide reference for the construction of intelligent operation and maintenance platforms for network operators.

On the basis of the synchrosqueezing transform time-frequency analysis method, the rotor fault texture feature extraction method based on the SVD order analysis of synchrosqueezing transform time-frequency graph, Which uses the synchrosqueezing transform to aggregate the energy of the short-time Fourier time-frequency graph, and then uses the SVD order analysis method to analyze the separability of the fault texture features of the synchrosqueezing transform time-frequency graph. Through the above method, the energy aggregation and noise suppression effect of synchrosqueezing transform time-frequency graph are remarkable. This method has been verified by the gas turbine model rotor fault simulation test bench, and has a high fault detection rate. Compared with time-frequency analysis fault diagnosis methods such as short-time Fourier transform, continuous wavelet transform, synchrosqueezing transform and Wigner-Ville distribution, the rotor fault texture feature extraction method based on SVD order analysis of synchrosqueezing transform time-frequency graph shows the best performance in fault detection rate and stability.

With the rise of the Internet of Things (IoT), some emerging mobile devices have been widely used such as wireless sensor networks, Radio Frequency Identification (RFID) chips, and smart cards etc. However, their communication security issues in open environments are increasingly prominent. Physical Unclonable Function (PUF) is a new type of "hardware fingerprint" that can authenticate IoT devices in the aspect of hardware. However, the Challenge-Response Pair (CRP) mechanism of PUF is vulnerable to Machine Learning (ML) modeling attacks. Based on this, the paper proposes a Dynamic Adversarial (DA) PUF through modifying the original CRP mechanism of APUF. Experimental results show that the PUF can effectively resist ML modeling attacks, while maintaining good uniformity, uniqueness, and reliability.

Debris flow is a natural geological disaster that frequently occurs in mountainous areas, posing a serious threat to the lives and property of local residents. However, conducting large-scale on-site investigations of debris flows is challenging due to the complex terrain of these areas. To address this issue, a CNN model based on dual-channel feature fusion is proposed to assess the susceptibility of debris flows. First, a dual-channel architecture is constructed, where a 2D CNN extracts the spatial features of the DEM image, and a 3D CNN extracts the spatial-spectral features of the multispectral image. Second, the residual structure is improved for feature extraction, and a CBAM block is added to enhance the network's ability to extract key features of valley images. Then, a fusion module is designed to fully integrate the features of the two channels. Finally, the susceptibility index is calculated based on the similarity score, and the susceptibility assessment results are divided into five levels and verified. The proposed model achieves an accuracy of 78.62% in the valley classification task and shows promising results in assessing the susceptibility of debris flows. Specifically, the proportion of high and moderately susceptible areas is 75.52%, the debris flow ratio is 96.38%, and the frequency ratio precision is 93.45%. These results demonstrate the feasibility of the proposed susceptibility assessment method and highlight its potential as a reference for debris flow prevention and disaster reduction efforts.

The unmanned boat navigates in the complex Marine environment and is prone to random interference. The robust control can deal with the optimization of moment information with distributed uncertainty. Although we do not know the distribution of random interference, we can obtain moment information for random interference. In this paper, we will consider how to apply distributed robust optimal control to deal with the problem of unmanned boat navigation under the condition of uncertain moment information of random interference and then convert the distributed robust optimization problem that is transformed into a second-order cone problem (SOCP).

In order to accurately measure the pose of cross-scale wiring target, a method for measuring the pose of cross-scale wiring target based on field of view space reconstruction algorithm and resolution coupling is proposed. Construct the pose measurement model of cross-scale wiring target, calculate the three-dimensional rotation coefficient under different coordinate axes, and extract the pose measurement characteristics of cross-scale wiring target according to the coefficient; Under the coupling of the field of view space reconstruction algorithm and resolution, the velocity recurrence relationship between internal and external parameters is obtained, and the position of three-dimensional point cloud is calculated, and the position and attitude measurement results of cross-scale wiring target are obtained, and realize the field of view space reconstruction algorithm and the position and attitude measurement of scale wiring target under resolution coupling. The experimental results show that there is no deviation between the measured trajectory and the expected trajectory under the proposed method, and the measurement error is obviously low, which can obtain better measurement results of cross-scale wiring target pose.

The surrounding illumination and temperature conditions of photovoltaic (PV) systems are often non-constant. To achieve a higher power output, it is essential to employ Maximum Power Point Tracking (MPPT) methods, which allow the system to adjust promptly in response to changes in illumination and temperature. Among all MPPT methods, the Incremental Conductance (INC) technique is extensively adopted because of its ease of understanding and implementation. However, the traditional INC MPPT method relies on a fixed duty cycle step size, which restricts the tracking speed and often leads to large oscillations around the MPP. In this article, an improved INC (IMINC) method is put forward to solve these problems. On the P-V characteristic curve of PV modules, by comparing the slope value corresponding to this point and a slope threshold F, this technique automatically determines the next step size according to the actual circuit voltage and current, thus overcoming the limitations of fixed step sizes and enabling faster system tracking. Moreover, the proposed approach uses the steady-state judgment threshold J to expand the range of the MPP, facilitating the system to achieve a higher power output, a faster response time, and an enhanced stability. Subsequently, simulations comparing the performance of the traditional INC and the proposed IMINC methods are conducted, and the results validate the superiority of IMINC.

The rapid increase in the amount of rubbish and the diversity of the substances in people's daily rubbish are two major problems of rubbish disposal at present, and the rubbish problem brings a serious burden to the environment as well as a waste of resources. At present, the main approach to deal with the waste problem is through waste separation. However, today’s traditional manual classification method is still used to screen the rubbish, which is labour-intensive with low sorting efficiency. This paper proposes a deep learning-based intelligent rubbish classification system, which can quickly achieve the positioning and recognition of rubbish in complex backgrounds for classification by incoming pictures and effectively improve the efficiency of rubbish classification. The system uses TensorFlow as the deep learning platform, and the combination of a convolutional neural network and a fully connected neural network is used to achieve the construction of the deep learning model. Meanwhile, Some nodes of the fully connected layer are discarded with a given probability in the training phase to avoid the overfitting of the system to optimise the system, and the system is also optimised by collecting various types of rubbish pictures and building inspection datasets, using data augmentation to improve the robustness of the model, the fast and accurate identification of different types of rubbish is achieved. The experimental results demonstrate that the intelligent rubbish classification system proposed in this paper has a high performance in terms of both accuracy and efficiency.

By studying the basic principle of ICP and the characteristics of low-temperature plasma, the advantages of using a plasma metal cylinder discharge chamber are discussed, and the influence of skin effect on discharge effect is deeply investigated. COMSOL multi-physics finite element software is used to analyze the physical characteristics of the sensitively coupled plasma. The results show that COMSOL has high feasibility and accuracy in numerical simulation of the perceptual-coupled plasma, which provides theoretical and technical support for the application of plasma in material processing, environmental purification, and other fields.

Unmanned Aerial Vehicle (UAV)-assisted mobile edge computing (MEC) technology has garnered extensive attention with the widespread adoption of 5G technology, owing to its flexibility, improved system capacity and computational capabilities, and ability to overcome complex geographical environments. This paper initially elucidates the advantages of UAV-assisted MEC systems in the current communication domain, outlining the definitions of various modules within the UAV-assisted MEC architecture. Subsequently, it summarizes the main application scenarios and key technologies, providing a concise description of the design approach for UAV MEC systems and highlighting the information security issues therein. Finally, the paper discusses the challenges and prospects for future research in this field.

Forest fire seriously affects the safety of the ecological environment, and even threatens people's lives, solving the problem of forest fire is the top priority of forestry work, so how to effectively respond to it has become an important task. The application of UAV technology to forest firefighting and rescue work is of great practical significance in improving rescue efficiency and safety. This study is based on the UAV forest firefighting system, with the theme of inspection aircraft fire intelligence recognition, through the flame smoke recognition technology to inspect the forest. The inspection aircraft uses the Yolov5 algorithm for flame smoke recognition and transmits images of detected open fires to designated personnel and alarms. Using the Pytorch framework and the Yolov5 algorithm, we can easily train a customized dataset and label the flame smoke data. By adjusting the training parameters, we can obtain a highly accurate detection model and finally realize the real-time detection of flame smoke by inspection UAVs.

As navigation systems are increasingly used in social life, people's demands for accuracy in positioning are also rising. However, commonly used technologies such as GPS still have some limitations. For instance, GPS signals can experience attenuation and obstruction indoors, in urban canyons, and in complex environments, which limits their positioning accuracy and robustness. In addition, IMU sensors accumulate errors over time, leading to the consequence of position drift. Therefore, in order to mitigate the drawbacks of using a single sensor, enhance the reliability of positioning systems, and provide people with better services, this paper primarily investigates the fusion algorithms of multimodal sensors. It conducts experiments on various algorithms related to sensor data fusion, and based on experimental results, compares them from different perspectives to determine the optimal algorithm. The studied algorithms include Kalman filtering and machine learning. Finally, a comparison of the noise resistance was conducted for the fusion algorithm with good performance.

Echo signal processing affects the measurement accuracy of ultrasonic Doppler flowmeter, in order to solve this problem, this paper uses Doppler method as the principle of flow measurement to develop an ultrasonic Doppler flowmeter based on STM32. Its main control chip adopts STM32F412VGT6 and Cortex-M4 core, which has powerful digital signal processing capabilities, which is conducive to the processing of echo signals. The hardware system part mainly designs ultrasonic signal transmission circuit, echo receiving circuit, operation amplification circuit, carrier signal generation circuit and mixing circuit. In terms of hardware system design, the transducer is used to transmit ultrasonic waves with a frequency of 1MHz, because the transmitted ultrasonic waves can not all be reflected, so the received echo signal is first amplified by the operation amplification circuit, and then the mixer is used to mix the amplified echo signal and the carrier signal generation circuit with a frequency of 1MHz frequency, and finally the mixed signal is filtered by the low-pass filter, and finally the Doppler shift signal with a lower frequency is obtained. Analog circuitry is validated to meet design requirements, improving accuracy and stability. Pave the way for digital signal processing.

In this paper, a miniaturized 5G quad-band antenna is proposed. The proposed antenna can operate in three 5G communication bands, n41, n78, and n79, as well as 5G Wi-Fi. The antenna adopts the method of etching slits in the radiation patch and the floor to etch different shaped slots respectively to realize the operation in four frequency bands, etching "C" and "U" type slots on the radiation patch to excite the n79 and 5G Wi-Fi bands respectively, and etching two inverted F type slots on the floor to make the antenna work in the n78 band, the low frequency band is excited by the circular radiation patch. Compared with other designs, the antenna has good performance, exhibits multiple frequency bands while maintaining miniaturization, and also has near omnidirectional radiation performance, which offers great potential for practical applications.

With the rapid development of autonomous driving technology, intra-vehicle communication plays a key role in efficient inter-vehicle interaction and internal system synergy. However, the complexity and challenge of in-vehicle communication scheduling in complex autonomous driving scenarios are becoming more and more prominent. An efficient scheduling method is required to meet the transmission requirements of different data streams and provide reliable communication services. This thesis proposes an in-vehicle communication scheduling scheme for autonomous driving based on decision tree algorithm. The scheme utilizes to time-sensitive network (TSN) technology traffic scheduling protocols, IEEE 802.1Qav, IEEE 802.1Qbv, respectively, and then by constructing a decision tree model, the method can intelligently make scheduling decisions based on the characteristics of the in-vehicle communication data, and the attributes in the autonomous driving scenario, which include priority, bandwidth limitation, packet loss recovery, real-time requirements, as well as the key attributes such as network topology and resource distribution. The decision tree algorithm can select the appropriate traffic scheduling scheme based on the importance and requirements of different attributes. Going through these key attributes to select the appropriate traffic scheduling aims to provide feasible and efficient scheduling schemes and to explore in depth the potential application of the decision tree algorithm in scheduling of in-vehicle communications for autonomous driving.

With the large number of low-voltage distributed new energy access, it has brought great challenges to the safe and stable operation of the power system, and it is necessary to effectively control the new energy with randomness, volatility and uncertainty in the distribution network. In this paper, based on the research of the site-wide distributed photovoltaic interactive system based on HPLC communication, the problems of data transmission network and communication mode are solved through the site-wide distributed photovoltaic interactive system networking, and the real-time data acquisition and precise command control of site-wide distributed photovoltaic power generation units are realized through the distributed photovoltaic interactive system. The system realizes the "observable, measurable, adjustable and controllable" requirements of the new power system after the high proportion of distributed photovoltaic is connected to the grid, helps the high proportion of distributed photovoltaic to the grid and reasonable consumption, and improves the security and stability of the power system.

With the rapid acceleration of urbanization, the scale of urban distribution networks has experienced significant growth. However, the prevalence of issues related to cable equipment quality disparities, inadequate construction process oversight, unfavorable operating conditions, and other factors has led to a pronounced increase in safety concerns for both individuals and the grid resulting from cable failures. Furthermore, the limitations of conventional online monitoring methods have posed considerable challenges in meeting the monitoring demands of widely distributed and large-scale urban distribution networks. To address these multifaceted challenges, this paper introduces the design of a low-power wireless online monitoring system grounded in the Internet of Things (IoT) paradigm. This system is adept at concurrently monitoring multiple status signals within urban distribution networks. The proposed methodology entails the establishment of a monitoring sensing network for distribution networks, built upon IoT principles. It encompasses the optimization of low-power hardware circuitry, power management strategies, data communication protocols, and software control mechanisms, among other facets. Through the aforementioned research endeavors, the successful realization of a low-power design and application of IoT-based distributed monitoring systems for multiple signals within power distribution networks has been achieved. This accomplishment serves as a robust technical foundation for ensuring the dependable operation of urban power distribution networks.

The Precision Airdrop System (PADS) is widely used in the accurate recovery of spacecrafts and the precision delivery of supplies. The accuracy of control is the core of the PADS achieving the precision airdrop. To enhance control accuracy, an intelligent control technology for enhancing PADS autonomy based on simulation environment is developed. The dynamic model employed to calculate the PADS state is a 6-degree-of-freedom (6-DOF) model, and the reinforcement learning control training method is based on the A-C architecture. Through analysis of simulation results in both windless and windy environments, it is evident that the reinforcement learning control strategy based on the AC architecture outperforms the traditional PID strategy. The analysis highlights the challenges posed by wind in achieving precise control and presents an effective approach for enhancing PADS autonomy through intelligent control.

Addressing the issue of seismic signal classification and recognition for natural earthquakes and artificial blasting, a deep learning model suitable for seismic signal classification and recognition is proposed. In this study, a dataset is created by selecting 14,001 natural earthquake waveforms and 12,870 artificial blasting waveforms from the seismic waveform data provided by the China National Earthquake Science Data Sharing Centre for the years 2010-2016. Based on the DenseNet architecture, an attention mechanism is introduced into the original Dense Block, adaptively enhancing valuable feature channels, thereby improving the model's perception and expression capabilities for important information. The model achieves an accuracy of 96.25% on the test set, surpassing the compared methods. Experimental results demonstrate that the attention-enhanced DenseNet exhibits excellent classification effectiveness and generalization ability when applied to seismic signals.

This paper reviews the research progress of gait recognition and rehabilitation training control based on EEG signals. Firstly, the importance and application scenarios of gait recognition research were introduced, with a focus on discussing gait recognition methods based on EEG signals. Then, the basic characteristics of EEG signals and the main EEG signal analysis methods used for gait recognition were summarized, including time-domain analysis, frequency-domain analysis, and time-frequency analysis. Secondly, this article collected some motion EEG signals and selected corresponding recognition algorithms through data processing, feature extraction, and classifier selection to perform gait recognition on the EEG signals. The results show that the algorithm proposed in this paper has a high recognition rate.

As the scale and complexity of power grid continue to increase, some abnormal events pose a significant threat to the security and stability of the power grid. This paper presents a novel approach based on the Random Forest algorithm for the substation fault detection. The approach is designed to improve the accuracy and efficiency of fault detection of substation equipment. By utilizing the machine learning techniques, specifically taking the decision tree algorithm as the base model, and following the principles of ensemble learning, a Random Forest algorithm is designed from multiple decision trees. The bagging technique is employed to create an ensemble learning model. The data preprocessing phase includes some crucial steps such as feature selection, interpolation preprocessing, normalization preprocessing, and the creations of training and testing sets. Additionally, appropriate evaluation metrics are chosen to assess the performance of our proposed approach. The experiments conducted on substation fault data collected from Suzhou City from January 2023 to August 2023 demonstrate the effectiveness of the approach. The dataset comprises six monitoring signals regarding the substations, with labels indicating the respective types of substation faults. The experiment results indicate that our proposed Random Forest-based approach achieves a extremely high accuracy rate compared to other baseline approaches. This approach can also help to enhance the substation fault detection capabilities and ensure the reliability and stability of power grid system.

With the development of intelligent technology, intelligent tools are widely needed, and automatic following car comes into being in the field of transportation. This paper designs an automatic following car system based on MSP432 from the actual needs. The system is composed of two vehicles, the front car is the lead car, the rear car is the following car, with MSP432 as the control core, mainly composed of L298N drive module, ultrasonic module, Bluetooth module and tracking module four main modules. MSP432 MCU controls the L298N drive module to drive the motor by output PWM signal, and the motor adjusts the wheel speed to realize the car's straight running and steering. The tracking module innovatively uses the combination of infrared sensor and grayscale sensor to achieve accurate path navigation. The two cars use the Bluetooth module to maintain communication, and the ultrasonic module is equipped with real-time automatic ranging to maintain the distance between the two cars and prevent the collision of the two cars. The test car can drive at various speeds, the driving process is stable, there is no collision, and the parking position is accurate. The system can realize automatic following, overtaking, stopping and waiting. It has the advantages of timely response, strong anti-interference ability, simple operation, low cost and high scalability.

The spread of light pollution engulfs the night sky and also has adverse effects on or-ganisms and the environment. To mitigate the light pollution problem, this paper estab-lish-es a comprehensive evaluation model of light pollution degree by optimizing TOPSIS model based on the combination of AHP and EWM. After that, a policy effect maximization model based on 0-1 planning is established and the optimal solution policy is solved.A comprehensive evaluation model of light pollution level was es-tablished. To widely measure the regional light pollution level, six effective indicators were selected to evaluate the light pollution level in different regions, namely: Level of devel-opment, Population, Air quality, Night light intensity, Climate, Geography. Matlab was used to calculate the weights of the selected indicators and then combined with TOPSIS scores to evaluate the regional light pollution.The highlight of this paper is the combination of subjective and objective. Using AHP and EWM to solve the weights makes the model more convincing. Finally, the model was evaluated with R2 and F-tests, and the model fit reached more than 95%.

With the advancement of technology, autonomous driving has become a popular research field. However, designing an autonomous driving system that can handle complex environments and meet real-time requirements remains an open challenge. In this study, we explore how to implement an efficient and reliable autonomous driving system using deep learning techniques. We experimented with a series of deep learning models, including VGG16, Inception series models, ResNet, MobileNetV2 and DenseNet121. By comparing the performance of these models on our dataset, we identified the best performing ones and fine-tuned them to optimize their performance for our task. In addition, to further improve the performance of the model, we introduced a spatial channel attention mechanism. We found that deploying the spatial channel attention mechanism at appropriate locations could significantly improve the performance of the model. However, deploying the spatial channel attention mechanism at all locations might cause the model to focus too much on local features and ignore global features. Overall, our results show that the performance of autonomous driving systems can be effectively improved by improving the network structure and training methods.

With the continuous development of technology, computers have infiltrated people's daily lives and work. Along with this, information security issues have become increasingly serious, leading to frequent incidents of privacy breaches. Without timely and scientifically effective measures for prevention and control, it can bring incalculable economic or psychological losses to individuals. Based on this, this article summarizes the current content of computer network security technology, analyzes the influencing factors, and proposes corresponding preventive measures for the main security issues, aiming to provide reference for relevant professionals.

In order to further improve network security, a network security situation assessment model fuzzy AHP is designed. Currently significantly between using a diffusion network, first proposed a set of line with the current realities of the system, the system comprises three trend indicators: decision-making, rule layer and index layer; and using the best mean clustering and C- (FCM) clustering criteria to resolve because of its ambiguity creates uncertainty error of assessment of the situation assessment and the balance of the data; the finalization of the establishment of two multi-factor assessment model for situational assessments vector. As can be seen from the last simulation results, the establishment of a new model based on the results of a traditional trend of Fuzzy AHP evaluation model is more accurate than some of the smaller weighting factors are also taken into account, so that even smaller standard deviation.

During the infrared zero movement measurement process, affected by the diversity of divisions of the unknown sight being measured, it is difficult for traditional algorithms to maintain stable identification and positioning accuracy. In order to solve this problem, this paper proposes a robust detection method for multi-type sight divisions based on the YOLOv5 target detection model. First, Transformer Integrates Multiple Convolutional Modules (TIMCM) is added to the C3 module of the YOLOv5 backbone network to enhance the network's feature expression ability for multi-type sight divisions and improve the model's global perception and interaction capabilities; secondly, In the Neck part, a jump bidirectional feature map enhancement structure (BiFPN with Diversified Scale, BDS) is designed to use jump connections to take into account feature information of large, medium and small scales and perform weighted fusion to improve the model's accuracy of sight target information of different scales. Extraction accuracy; finally, a high-resolution detector (HRD) is added to the Head part to further improve the detection accuracy of low-resolution targets. Experimental results show that when the improved algorithm identifies and locates conventional infrared sight divisions, the accuracy of false detection or missed detection is reduced, the stability of the algorithm is significantly enhanced, the accuracy reaches 96.5%, the recall rate reaches 97.4%, and the average precision average It reached 97.1%, and it can accurately locate unknown aiming reticles, which shows that the algorithm has strong robustness and generalization, and has certain practical value and research significance in the measurement of zero position movement of infrared sights.

As the complexity of cyber attacks increases, honeypot networks, acting as a proactive security mechanism, have become an important area of research in the field of cybersecurity. Although traditional honeynets offer a degree of defense, their limitations in dealing with complex and persistent cyber attacks, including a lack of flexibility, insufficient allure for attackers, and high maintenance costs, are increasingly exposed. This paper introduces a versatile dynamic honeynet architecture (VDH) to compensate for these deficiencies in existing honeynet technologies. By deploying in parallel with enterprise networks and dynamically generating honeypot nodes, the VDH architecture achieves flexible control and redirection of suspicious traffic, thus enhancing network protection capabilities. The paper further abstracts honeynet attack and defense behaviors into a non-cooperative incomplete information game model, and through precise modeling and analysis of the single-stage attack-defense game process, proposes a novel optimal configuration strategy for dynamic honeynets. Lastly, the paper designs and implements the VDH prototype system based on network function virtualization, and verifies the effectiveness and superiority of the VDH structure and configuration strategy.

Due to safety hazards, many residential buildings have restricted the entry of electric bicycles and chargers into the building. However, the phenomenon of electric bicycles entering the elevator in violation of regulations often occurs. This article develops a machine vision based electric bicycle detection and recognition algorithm, which uses the original image of the electric bicycle as input to the neural network and designs a feature extraction algorithm. This algorithm utilizes the external morphological features of electric bicycles to filter out interference and intelligently analyze and accurately distinguish interfering objects. And combined with pressure sensing technology to determine the precise position of electric bicycles in elevators, complete automatic detection and recognition. Embed the algorithm into the internal camera of the elevator, intelligently monitor the electric bicycles inside the elevator, and achieve detection, analysis, and recognition of the electric bicycles. As long as the camera inside the elevator detects that an electric bicycle has entered the elevator, it will immediately issue a warning of "electric bicycles no enter", and the elevator door will remain opening. Only by pushing the electric bicycle out of the elevator door can it operate normally. This function solves the problem of electric bicycles going upstairs in high-rise buildings from the source, greatly reducing safety hazards.

Semantic segmentation of remote sensing image is a key technology in the field of remote sensing image processing, and its segmentation results can be used in land resource management, yield estimation, disaster evaluation and many other aspects. However, the forest land is widely distributed and the tree species are diverse, which brings difficulties to the extraction of forest land. Different from traditional manual investigation, semantic segmentation can quickly extract forest land from remote sensing images. U-Net is a deep codec structure, which has been frequently used for high-precision image segmentation. In this paper, Multi-scale features of different levels of U-Net are used to extract forest land from high-resolution remote sensing images. Multi-scale features can capture features of different scales for fusion, and note the importance of boundary information. A boundary attention module is added to explicitly use boundary information for context aggregation, which makes the extracted boundary effect more remarkable Attention module is designed to enhance the learning of vegetation characteristics, so as to improve the segmentation performance. This study effectively improves the problems of fuzzy boundary of semantic segmentation, large intra-class differences and small inter-class differences, and can quickly extract forest land.

With the integration and development of computer and communication technologies, intelligent communication is playing an increasingly important role in various fields. In response to the intelligent, flexible, and robust communication demands in emergency scenarios, this paper proposes a network for constructing a self-organizing communication network based on Unmanned Aerial Vehicle (UAV). In this network, the network consists of a master node UAV and a branch node UAV, forming a master-slave structure. This UAV communication network can be applied in scenarios like forest wildfires, where the master node UAV dynamically schedules the branch node UAV to achieve coverage and monitoring of ground fire scenes, providing support for forest fire prevention, control, and assessment. Simulation experiments demonstrate that the proposed network exhibits significant effects in ensuring network connectivity, coverage, and mobility efficiency.

In recent years, Artificial Intelligence (AI) has found widespread application in various fields, owing to its efficient data processing capabilities. In Structural Health Monitoring (SHM), AI methods are employed to analyze structural vibration frequency signals, thereby enabling the identification of structural damage. However, due to the existence of numerous methods, further research is essential to investigate the impact of different approaches of damage identification results. In this study, three types of convolutional neural networks (CNNs) were employed to identify structural damage of the ASCE benchmark structure. Additionally, an efficient time-history data processing technique was incorporated into these CNNs. Subsequently, the damage identification results, considering different input data and network structures, were compared. The results reveal the following key findings: (1) The three utilized networks (1D-CNN, 2D-CNN, and PCNN) exhibit high identification accuracy when dealing with relatively simple damage scenarios. (2) The transformation of one-dimensional vector data into frequency-domain signals through the Fast Fourier Transform (FFT), followed by inputting this FFT signal into the 2D-CNN, results in a significant improvement in damage identification accuracy, particularly in more complex damage scenarios. (3) The PCNN_FFT model demonstrates high accuracy when dealing with relatively simple damage scenarios, but its accuracy can be substantially reduced in more complex situations.

In view of the difficulty of machine abnormal sound detection under the condition that abnormal sound samples are difficult to collect, this paper proposes an unsupervised abnormal sound detection model based on a self-coding model, which effectively improves the accuracy of abnormal sound detection under this condition. In this paper, location coding in Transformer is replaced with relational awareness self-attention to improve the representation capability of location coding. Secondly, the relevance scores in multi-head attention are mixed to enhance the understanding of context in the attention matrix. At the same time, Layer Normalization was replaced with Batch Normalization to speed up model training, and improved Transformer was introduced into the encoders and decoders of self-coding models. Finally, the improved self-coding model is used for unsupervised learning of the machine's normal sound to obtain the potential feature distribution of its normal sound. ToyADMOS and MIMII open data sets are used for experiments. Compared with traditional autoencoders and two improved self-coding models, The AUC score of toycar, Toycar, fan, slider and valve machines increased by 2.1%, 1.97%, 3.06%, 0.34% and 2.99%, respectively.

Object detection is an important application of optical satellite remote sensing imagery interpretation. Since the objects of interest, such as aircraft, ships, and vehicles, are small in size with obscure contour and texture, it is difficult for object detection in satellite images. The spatial resolution of aerial images is higher than satellite images, and the object detection model can achieve higher precision. Knowledge distillation has been validated as an effective technique by learning the common features of aerial and satellite images to improve the precision of object detection in satellite images. It means that a teacher model pre-trained on aerial image datasets guides the training of a compact student model on satellite image datasets. However, there are data distribution differences between aerial images and satellite images. The distribution differences may cause the teacher model to give guidance signals that deviate from the ground truth, thus leading to sub-optimization of the student model. In this paper, we proposed a new distillation scheme, termed DC-KD, which updates the teacher model using the predictions of the teacher model that are inconsistent with the ground truth, and the rest are used to guide the training of the student model. We achieved a 3.88% mAP₅₀ improvement on the xView dataset based on the YOLOX-S model.

Traditional power automation sensitive data classification algorithms are prone to generalization errors, resulting in low geometric mean of data classification. Therefore, it is necessary to design a sensitive data classification algorithm for power automation based on trusted privacy computing technology. A power automation sensitive data classification model was constructed using trusted privacy computing technology, and an optimization framework for power automation sensitive data classification was generated, achieving the classification of power automation sensitive data. The experimental results show that the designed power automation sensitive data classification algorithm based on trusted privacy technology has a high geometric mean of data classification, proving that the designed power automation sensitive data classification algorithm has good classification performance, effectiveness, and certain application value.

In order to improve the correction accuracy and stability of power quantity data, reduce errors, and improve the management efficiency and operation quality of power system, a power quantity data correction method based on improved DBSCAN density clustering algorithm was proposed. The principal component method is used to process the collected power quantity data, divide the power quantity data into different time interval levels, collect missing values according to the recorded power consumption data, and filter out the irregular power consumption detection error information in the power system environment. The pre-processing is completed by selecting the Romanjofs reference, and the remaining data is detected by the T-distribution detection method to determine whether it is gross error. The power quantity error data obtained by DBSCAN clustering algorithm is summarized, the distance between classification attribute data is calculated, and the original DBSCAN is extended instead of the European distance, so that it can process classification attribute data, set the correction period of the correction matrix, and constantly control the collection process of power quantity error data by using the positive period parameter. Finally, the correction of power error data is realized. The experimental results show that the power consumption of the proposed method reaches the maximum value of 48.53KWH at the end of the experiment, the average correct rate is about 97%, the correction correct rate is the highest, and the actual correction effect is the best.

When identifying abnormal timing of power metering devices, the identification effect is poor due to the diverse attributes of the original power metering device state data. Therefore, a research on abnormal timing identification algorithm of power metering devices based on variational autoencoder and support vector machine is proposed. A VAE-LSTM- DTW model was constructed with a variational autoencoder as the core, which can be mainly divided into two parts. The reconstruction model is composed of a VAE network improved by LSTM, which is responsible for generating the time series reconstruction data of the power metering device corresponding to the input data. The evaluation model is responsible for comparing the reconstruction effect of the model on the input power metering device status data through DTW and algorithm, and performing anomaly detection accordingly. When identifying abnormal time series, support vector machines are used to match and identify the abnormal features of the operating state data of individual power metering devices. In the test results, the identification accuracy of the design algorithm is stable at above 0.85, the recall rate is stable at above 0.80, and the F1 score is stable at above 0.90.

Considering that anxiety affects people's work and life, accurate detection of anxiety has important application value. Considering the interaction between different channels, a PLI and TD-2DPCA anxiety recognition method based on EEG signal is proposed. In this method, the collected original EEG signal is first preprocessed, the pre-processed signal is decomposed into five rhythmic waves by wavelet change, and the functional connection information of the brain is obtained by phase delay index coupling method for the five signals respectively, and then the anxiety features are extracted by bidirectional and two-dimensional principal component analysis method. Finally, the feature classification is implemented by support vector machine to realize the detection of anxiety. The results show that the proposed method has high recognition accuracy and can identify anxiety information well.

Transmission lines are an important component of the power system, and their stable operation is crucial for ensuring power supply safety. In a strong wind environment, it may cause the tower to tilt, causing transmission line faults, and thus affecting the normal electricity consumption of users. Therefore, a computer vision based method for detecting the inclination of transmission line towers under high wind speed impact is proposed. Preprocess the transmission line tower image and use the random forest algorithm to extract the features of the transmission line tower image. On this basis, based on the Region Proposal Network (RPN) in computer vision, the tilt detection of transmission line towers under high wind speed impact is achieved. The experimental results show that the method proposed in this paper has a good detection effect on the tilt of transmission line towers, and can accurately detect the tilt angle of transmission line towers. When the maximum wind speed is reached, the tilt angle of the tower is the maximum.

With the continuous and rapid development of the technology of the empowered Internet of Things (IoT), multi-service scenarios continue to appear, which poses a new challenge to the delay of large data services in IoT. Aiming at the cloud and edge cooperation for the IoT architecture, this work proposes a cloud and edge collaboration mechanism for the network delay optimization based on software-defined network, which tries to minimize the average delay by introducing a continuously variable cloud and edge task segmentation coefficient. Besides, we transform the delay optimization problem into a convex optimization one, based on which an optimized task segmentation strategy is obtained. The experimental results indicate that the service delay and network congestion can be reduced greatly by reasonable loading of edge node tasks to the cloud center.

Crop type recognition in remote sensing plays a crucial role in agricultural management and resource monitoring. In addressing the challenge of temporal feature extraction from remote sensing time series data, this study introduces a novel temporal extraction module that integrates 1D convolutional residual blocks and temporal attention mechanisms. Our approach effectively captures temporal patterns in crop growth processes and achieves high-precision crop classification in remote sensing imagery. Through experimental validation on the Sentinel2-Agri dataset, our model demonstrates performance improvements in classification accuracy compared to traditional methods, yielding higher accuracy and finer-grained segmentation results. Additionally, ablation studies and exploration of model configurations substantiate the effectiveness and robustness of our approach. This research holds important implications for the advancement of remote sensing-based crop classification.

Aiming at the high computing performance requirements of visual Simultaneous localization and mapping (SLAM) on robots, this article proposes a cloud-based visual SLAM method for low-cost robots which have limited computing performance. In this method, the monocular ORB-SLAM algorithm is segmented, and the task of key point extraction is completed by the robot. The cloud completes pose calculation and map construction through the feature values sent by the robot. Finally, the hardware and software platforms are designed for verification. The experimental results show that the robot based on ESP32-S3 module can meet the computing power requirements of feature extraction, and the cloud can also smoothly locate and build maps through the feature point data from the robot.

With the rapid development of computer vision and image processing, geometric shape recognition has become a highly regarded research area. This study aims to explore a method that combines the FAST feature point recognition algorithm with the Hough transform for fast and accurate recognition of geometric shapes in images, including lines, circles, and rectangles. The FAST feature point recognition algorithm is known for its speed and robustness in effectively detecting key points in images, while the Hough transform allows mapping these key points into parameter space, enabling us to detect and recognize specific geometric shapes. This research focuses on explaining the principles of the method, presenting experimental results, performance evaluations, and discussing future research directions.

The bad text containing variant words seriously harms the health of the network environment. The existing methods for the recognition of bad variant text do not take into account the importance of the phonetic and positional information for the recognition of bad variant text. In this paper, a ChineseBERT-BiGRU variant bad text recognition model is proposed. This model first learns and trains the word vector of the text by inputting the pinyin information, font information and character information of the text, and then combines the position information of the text. Then the word vector of the text is input into BiGRU to learn richer semantic information of the text. Finally, the bad text containing variant words is identified through Softmax classification. The accuracy, accuracy and F1 values of ChineseBERT-BiGRU in the data set in this paper are compared with those of other models.

The aircraft development process involves a large number of simulation and testing tasks, as well as simulation and testing data. Establishing a platform to analyze these data is important to accelerate the iterative optimization of product design. Therefore, this paper proposes a simulation and test data correlation analysis framework for the optimization of the whole process of aircraft design. The framework starts with an analysis of the functions that a data correlation analysis platform should have based on the application requirements for each phase of aircraft design. Then, the architecture of the data association analysis platform is proposed based on the functional analysis, and each key technology supporting the architecture is introduced. Finally, typical data association analysis objectives and their regulated analysis processes are summarized.

Data hiding has been widely investigated as a viable method for multimedia authentication, numerous data hiding method in images have been proposed during the past two decades. Reversible data hiding refers to data hiding applications where the host media can be partially or completely reconstructed. Although such method possesses advantageous characteristics in applications, its development is faced with numerous obstacles. Little effort has been made to develop method for reversible hiding information in binary images, the existing method either have unsatisfactory capacity or require additional information. It is challenging work that have significant potential. In this paper, we propose a novel high-capacity reversible data hiding in binary images. The recovery of original image is conducted with the aid of auxiliary data hidden. The experiments show the fidelity of host images after data hiding is very high, and data hiding capacity is as high as 0.5 per pixel, and recovery rate of original image is almost above 60%.

The lifting parameters of high explosive cloud are important data source item for calculating explosive cloud diffusion and serve as the start for subsequent cloud diffusion calculations. Furthermore, lifting cloud of air explosion and near-surface explosion is significantly different due to the impact of ground reflected shock waves. In this paper, numerical simulation is applied to calculate the floating cloud which is from high explosive explosion. Two numerical examples are presented with comparisons from the lifting cloud of air explosion and the explosion under the effect of ground reflected shock wave. The simulating results show that the ground reflected shock wave makes significant difference to explosive cloud, and accelerates the turbulent motion which make more serious exchange between floating cloud and surrounding medium. Major physics of the cloud lifting from high explosion can be well captured in the simulation, and lift parameters of high explosive cloud can give good prediction for calculating cloud diffusion and settlement.

Virtual reality technology has been widely used in education scenarios. Among them, the immersive virtual multimedia classroom can mimic the physical classroom, so as to facilitate the performance of on-line learning. However, there are still gaps between the virtual and physical classroom, especially the lighting environment which plays an important role in the visual experience of participants. In this paper, we aim to study the relationship between the lighting parameters and the visual comfort. We firstly establish a virtual classroom with adjustable lighting parameters. Then we conduct user cases which mimic the natural and indoor lighting respectively, investigating the effect of the two different lighting environments upon the visual comfort. Particularly, we establish the empirical fitting models from the collected subjective ratings, which can provide the perception threshold and the optimal lighting conditions. The proposed method can benefit the design and implementation of virtual multimedia classroom.

In recent years, there has been a substantial rise in the utilization of fisheye lenses, which offer a wide field-of-view. However, the distortion inherent in these lenses presents a major challenge for intelligent recognition of dense analogs (IRDA) in applications based on convolutional neural network (CNN). To enhance the accuracy of IRDA, we introduce a novel algorithm called Key Point Calibrating and Clustering (KPCC), which is based on an equidistant projection model. Our method can fully mine hard examples and effectively correct their misclassifications predicted by the CNN, thereby significantly improving the accuracy of IRDA.

With the rapid progress of human technology level and artificial intelligence field, gesture recognition technology based on deep learning plays an important role in the development of human-computer interaction. Currently, the accuracy of most target detection networks in recognising gesture actions has reached a relatively desirable level, while the training speed and results of many deep learning models are often easily limited by the computational power of these hardware due to the fact that they are used on some low-computing-power platforms, such as low-end CPUs and GPUs; Secondly, the size and complexity of the model will also have a non-negligible impact on the subsequent deployment phase of the application. Aiming at the series of problems raised above, this paper adopts a gesture recognition detection algorithm based on the improved YOLOv5, replacing part of the convolution module with the Ghost module to reduce the parameters, which in turn transforms C3 into C3Ghost, and Conv adopts a parallel structure and transforms it into GhostConv, which further reduces the amount of computation during the training process and accelerates the inference speed of the model, so as to realise that the model improves the training efficiency of the network under the condition of limited hardware arithmetic power. At the same time, in order to solve the problem of accuracy degradation that may be accompanied by the completion of lightweighting, the CBAM attention mechanism is also added to strengthen the network's ability to extract target features to improve detection accuracy; Then the α-IoU loss function is used instead of the CIoU loss function to make the model converge more quickly during training. Ultimately, experimental comparisons show that compared with the original algorithm, the improved YOLOv5s PRO has 46.2% fewer Parameters, 42.6% smaller Model Size, and 48.1% fewer GFLOPs, which can effectively balance the speed and accuracy of training.

In recent years, human motion recognition based on wireless sensing has been widely used in smart home, virtual reality and other fields. Among them, human motion recognition based on WiFi channel state information (CSI) has received great attention. Although the current large-action recognition method based on CSI has good performance, the recognition accuracy is low when performing fine-grained human gesture recognition, which is due to the difficulty of collecting CSI human gesture perception dataset and the small number of high-quality datasets. Therefore, we convert CSI data into images and augment CSI images using super-resolution image generation methods. At the same time, we propose a SCBAM attention mechanism to further enhance the attention of the model to important features. In the experiment, we performed experimental analysis using the publicly available Widar3 dataset. The results show that the recognition accuracy of our mixed attention model in the same scene, across location, direction and scene reaches 99.73%, 97.41%, 94.57% and 95.98%, which is 0.07%, 0.29%, 0.05% and 0.99% higher than before data enhancement.

Most of the current frameworks for outlier data mining are one-way structures, and the efficiency of data mining is low, which leads to the increase of data mining redundancy. Therefore, this paper proposes the design and analysis of outlier data mining methods for power data middle office business systems based on decision trees. According to the actual mining requirements and standards, the initial outlier data sources are collected, and the multi-level method is adopted to improve the efficiency of data mining, and the multi-level power data system data mining framework is designed. On this basis, the decision tree outlier data mining model is constructed, and the unit grid integration is used to realize outlier data mining. The final test results show that: for the selected five data mining test cycles, the final data mining redundancy is better held 20%, which indicates that the comprehensive decision tree, the designed G power data middle office business system outlier data mining effect has been significantly improved, targeted, mining processing effect is better, and has practical application value.

Time-sensitive network (TSN) guarantees deterministic delivery of traffic by prioritizing it differently. Commonly used traffic scheduling algorithm is time-aware shaper (TAS) combined with credit-based shaper (CBS). Aiming at the problem of excessive reserved bandwidth for stream reservation (SR) traffic during CBS scheduling, a traffic scheduling strategy based on load balancing was proposed. And the reserved bandwidth parameters of the SR traffic were optimized after completing the modeling of its delay problem. This strategy sliced and encapsulated the SR traffic packets into equal-length data frames at the end nodes and made them uniformly transmitted during the sending cycle. It reduced the end-to-end worst response time of the SR traffic due to packet congestion. After that the delay model was solved using simulated annealing algorithm to get the optimal reserved bandwidth. Simulation results showed that the proposed SR traffic scheduling strategy effectively reduced the reserved bandwidth and the delay of traffic transmission inside the switch on the basis of satisfying the traffic schedulability.

We propose SonicScape, a novel CRNN-based model specifically designed to effectively tackle the issue of excessive natural sounds in urban environments. This is achieved by leveraging mel spectrograms. Our model aims to evaluate and mitigate the impact of natural sound proportions on urban environmental noise pollution, with a particular focus on the city of Shenzhen. To ensure accurate assessment, we introduce a novel evaluation metric called NSP (Natural Sound Proportions), which quantifies the contribution of natural sounds to the overall noise levels. To support our research, we have created the Shenzhen Environmental Sound Dataset, which addresses the lack of publicly available datasets by incorporating unique sound characteristics specific to Shenzhen. SonicScape combines the power of CNNs for efficient feature extraction with GRUs for precise temporal modeling, resulting in superior performance on this dataset. In conclusion, we suggest future work to expand the dataset through an ongoing annotation process, explore more robust annotation and computation methods for NSP, and enhance the models by incorporating cutting-edge techniques from the domains of image recognition and time series analysis. This will improve classification and regression tasks based on spectrograms.

Obstacle detection systems serve as a linchpin in ensuring the safety of autonomous vehicles. While contemporary detection techniques grapple with intricate results, diminished efficiency, and substantial computational burdens, our research introduces a groundbreaking method anchored in point cloud registration and target reduction for areas with obstructions. This approach not only facilitates high-precision positioning but also expedites the process of obstacle detection, a cornerstone for the subsequent navigation and obstacle evasion tasks in intelligent vehicles. Initially, we harness the power of NDT registration to achieve meticulous positioning. Subsequent stages involve refining the detection landscape by filtering out non-road obstacles using our innovative target reduction process and then employing advanced clustering techniques to pinpoint obstacles. We've validated our approach using the comprehensive Mulran dataset. Both qualitative and quantitative assessments underscore the method's robustness, reliability, and superior effectiveness. Experimental outcomes compellingly demonstrate that our newly proposed obstacle detection framework considerably elevates both operational efficiency and accuracy, paving the way for safer and smarter autonomous driving.

Publisher's Note: This paper, originally published on 22 May 2024, was withdrawn on 03 June 2024 per author request.

Coal is an important part of the world’s energy structure and gangue is mined along with coal. The extremely high visual similarity between them makes the sorting process more difficult. Existing studies are mainly based on feature extraction or the non-end-to-end convolutional neural networks to classify the target in images. Although some of these methods can have excellent accuracy, have problems with complex feature extraction processes and long image processing time. In this paper, I propose a method for detecting and classifying coal and gangue based on an X-ray map and the YOLOv8 model. The experimental results show that this method greatly improves the speed of image processing and ensures high classification accuracy, which can better solve the problem of real-time detection of coal and gangue. It has a good effect in the real-time detection and classification task.

In the event of a severe accident occurring at a nuclear power plant, it is necessary to promptly diagnose the state of the plant in order to implement appropriate strategies and measures for mitigating potentially severe consequences. This diagnosis is based on factors such as the timing of the initial event, the type and location of the accident, and the response information from the nuclear safety systems. Advancements in information technology (IT) have led to the application of big data and deep learning techniques in nuclear severity accident management. However, there is a scarcity of real-world data on nuclear severity accidents, with only limited information available from three notable incidents: the Three Mile Island nuclear accident in 1979, the Chernobyl nuclear accident in 1986, and the Fukushima nuclear accident in 2011. To address this issue, this paper proposes the use of an integrated nuclear severity accident simulation computer code to generate a substantial amount of deep learning training data. The resulting data and labels are saved into tens of thousands of small files. However, mainstream operating systems such as Linux, Windows, and MacOS, as well as file systems like Hadoop Distributed File System (HDFS), GlusterFS, Ceph, and MinIO, lack sufficient support for managing a large number of small files (LOSF). To overcome this limitation, a method that combines the SeaweedFS cloud storage database with deep learning techniques is proposed for the severity accident management system in Nuclear Power Plants (NPPs). The simulation-generated small files are efficiently transferred to the SeaweedFS cloud storage database using the Application Programming Interface (API) provided by Amazon Simple Storage Service (Amazon S3), ensuring parallel efficiency and optimal storage space utilization. Furthermore, the SeaweedFS database offers enhanced portability through the use of Docker images and containers. TensorFlow is utilized for deep learning training using the data stored in SeaweedFS. Following the training process, a private, scalable, and portable nuclear severity management system is established, integrating Artificial Intelligence (AI) and big data analytics to enable the quick diagnosis of nuclear severity accidents and the assessment of the damage sustained in NPPs.

Video matting aims at accurately separating foreground from videos. Recent video matting researches pursue to eliminate auxiliary inputs. However, due to the limited ability of extracting global correlation features, these methods suffer from performance degradation when dealing with complex scenes or natural background videos. To address this challenge, we propose a video matting method called Video Matting Based on Local-Global Features Fusion (VMBLGFF) which can extract both comprehensive global correlation features and local subtle features. VMBLGFF contains two closely connected networks: a transformer network that utilizes window and global attention mechanisms to obtain global correlation features within and cross windows, and a fusion network that integrates local subtle features into the global correlation features to supplement the local detail information which may be overlooked by the attention mechanisms. VMBLGFF alleviates the issue of limiting global correlation features and has been benchmarked on both synthetic and real datasets, and the results demonstrate that VMBLGFF improves the quality of video matting and exhibits good generalization performance.

The accurate prediction of landslide displacement is an important part of social stability and plays an important role in economic development. Therefore, it is a very important topic for how to accurately predict the changes in landslide displacement. According to the research of previous scholars, this paper proposes a multi-model analysis and prediction method that combines gray system preprocessing, radial basis function neural network model, and Markov chain state matrix to predict and modify landslide displacement. Compared with the single RBF model, RMSE and MAPE can be increased to 449.8mm and 0.38%. Grey-PSO-RBF model RMSE and MAPE respectively 437.5mm and 0.33%.No difference between Grey-Markov-RBF and Grey-PSO-RBF.

Object detection in road scenes is a crucial component of autonomous driving. Due to the significant variations in target scales, it is prone to false positives, false negatives, and some underperforming devices cannot deploy the state-of-the-art detectors. To address these issues, we propose a lightweight algorithm based on an improved YOLOv8. We simplify the model by using the FasterNet Block from FasterNet to replace the BottleNeck module in YOLOv8's backbone network C2f, reducing parameters and floating-point computations. We also enhance feature extraction by substituting RFCAConv for the downsampling standard convolution in C2f. Additionally, we introduce Wise-IoU to replace the original activation function, directing the network's focus towards anchor boxes of average quality.To promote effective fusion of original, shallow, and deep features, we introduce the BiFPN structure to replace YOLOv8's PAN structure. Furthermore, a small object detection layer is added to the head to handle the drastic scale variations in road scenes. Experimental results on the SODA10M dataset demonstrate that the improved YOLOv8 model achieves a 55.8% mAP@0.5 and a 35.5% mAP@0.5:0.95. The model's parameter count, size, and floating-point computations decrease by 58.3%, 57.0%, and 21.1%, respectively. Analysis of the experimental results confirms that the proposed model is effective and superior, striking a balance between detection accuracy and model lightweightness.

AFDX (Avionics Full Duplex Switched Ethernet) network is a deterministic network widely used for data exchange between aviation subsystems. In order to accurately and efficiently detect and analyze aircraft AFDX network data, a WinPcap-based AFDX network data analyzer design and frame parsing method are proposed to realize real-time capture, display and parsing of AFDX data frames. Experimental verification was carried out using a dedicated AFDX board. The results show that the designed analyzer is able to capture and analyze AFDX data frames in real time, and can be tested for its characteristics.

Contemporary object detection algorithms have achieved significant advancements in various fields. However, existing object detection algorithms suffer from reduced image clarity and increased noise under low visibility conditions, resulting in a decline in algorithm accuracy. To address this issue, we propose a novel Light Source Influence Matrix-YOLO (LSIM-YOLO) framework. we employ the LSIM algorithm to enhance the contrast of objects within the images. Conventional methods for object detection under low visibility conditions tend to overlook latent information within the images. To mitigate this, we introduce the se-ECA (Spatial Excitation and Channel Attention) mechanism to augment spatial information in features, which can enhance feature quality, thereby improving the performance of visual tasks under low visibility. Furthermore, we address the limitations of existing object detection networks in extracting features from low visibility images by replacing traditional strided convolution layers in the CNN architecture with Spatial-to-Depth Convolution (SPD-Conv). This compensates for the deficiency in utilizing fine-grained information during feature extraction. These efforts culminate in the development of the LSIM-YOLO model. Experimental results on the real dataset RTTS demonstrate a notable 3% improvement in average precision (AP), validating the feasibility and effectiveness of our improved detection framework.

With the rise of network informatization in assisting teaching, students in schools use various informatization methods for teaching. Students do not need to go to school to attend classes and participate in experimental learning activities through the Human-Computer Interaction Learning Center and "Remote Electronic Laboratory". The rapid development of online teaching should require the design of simulation tools, software and facilities for course content and experimental environments, but these key factors have become one of the bottlenecks affecting the quality of online teaching. This paper uses the neural network algorithm in machine learning and artificial intelligence to conduct intensive training and learning on the output and reception signal samples of semiconductor devices to obtain virtual visualization simulation results. Fit the simple data of the received signal to a system of two-dimensional linear equations to obtain the different magnitudes of the signals and obtain accurate coordinate positions. With the help of machine learning visualization processing, model changes can be compared, and the output effects are reasonable and very accurate. Compared with traditional semiconductor devices that do not have visual interface operations, virtual simulation in remote laboratories further expands the scope of use. The GUI program of the virtual visual simulation model created in this article makes the simulation of electronic circuit models simpler and more convenient, and has extensive promotion and reference effects.

In order to solve the problem of wall plate deformation of 100,000 cubic meters external floating roof crude oil tank in a commercial reserve oil depot, the whole cycle deformation of the tank wall was detected by 3D laser scanning detection system, and the tank wall structure shape based on high-precision point cloud data was obtained.The full-scale finite element model of the tank wall structure is established by using the general finite element analysis software ANSYS.The 3D scanning inspection data are used as the initial geometric deformation of the tank wall for nonlinear buckling analysis.The safety performance evaluation of the deformed tank wall structure is carried out from the aspects of the tank wall stiffness and stability, and the operation and maintenance suggestions are put forward.

In this paper, the problem of nonstandard samples of cigarette packet found in daily patrol inspection is taken as the starting point. Based on the research of equipment testing principle, we designed a method to obtain the minimum detection limit of the cigarette package appearance detection device. Through a large number of simulation experiments, accurate and reliable experimental data conclusions are obtained, the quantitative evaluation of the detection capability of the cigarette packet appearance detection device is realized, and provided a quantitative reference for the validity verification of the cigarette package appearance detection device.

Addressing the problem of sensitivity in parameter selection within traditional particle swarm optimization (PSO), where varying parameter values can significantly impact path optimization performance, this study replaces PSO with gray wolf optimization (GWO). An improved algorithm, grounded in GWO, is proposed for static three-dimensional (3D) path planning of unmanned aerial vehicles (UAVs). The improved algorithm combines GWO and PSO. By improving the position update formula, the individual gray wolf can adjust its position more flexibly during the search process, and improve the issue that GWO is prone to stuck in local optimum prematurely when searching for optimization. Test and simulation outcomes demonstrate the improved algorithm's superior performance in path planning.

This paper explores the evolution of industrial automation and AI's application in manufacturing, focusing on machine tool automation. It delves into its definition, characteristics, development trends, impact on production efficiency and quality, as well as existing challenges. AI applications in machine tool automation such as predictive maintenance, adaptive control, and intelligent optimization are discussed in detail, showcasing how AI enhances automation through examples. The study also touches upon the new production model combining AI with automation, exploring its implications for businesses and industry growth. In future outlooks, current study issues are highlighted alongside predictions for machine tool automation and AI's future, including potential challenges and solutions. Suggestions for policy makers and business decision makers are offered. Overall, this research underscores the significance of the new production model combining machine tool automation and AI in enhancing efficiency and industry development, despite existing challenges. Research limitations are acknowledged and future directions suggested.

Due to the continuous expansion of the Internet scale, network traffic has experienced an explosive growth, accompanied by increasingly complex structures. Improving the detection accuracy of malicious traffic and efficiently distinguishing different categories of malicious traffic has become an urgent problem to be addressed. Research has shown that hybrid approaches combining CNN and BiLSTM exhibit strong responsiveness and perform well in solving research problems such as video classification, sentiment analysis, and emotion recognition. Therefore, in order to enhance the learning capability and detection performance of IDS, this paper proposes an improved version of an intrusion detection method based on the Transformer and Conv-BiLSTM networks. This model combines the advantages of both modules to improve performance compared to traditional models.

Many of the major casualties caused by ship sailing accidents have proved that life jackets play a vital role in safeguarding the lives of people on board. In this paper, the K-means++ algorithm is used to optimize the target prior frame, and the CBAM attention mechanism is added to the YOLOv7 network to enhance the effective features and weaken the ineffective features in the feature extraction process, to enhance the localization accuracy and detection ability. The BiFPN network is used to enhance the information acquisition degree of the target, and finally, a comparison experiment is conducted with the YOLOv7 model before improvement. The experimental results show that the accuracy of the YOLOv7-CB model after training on the lifejacket dataset reaches 92.3%, the detection rate is 6.4% higher than that of the pre-improvement period, and the detection rate is 8.5% higher, which is good for the lifejacket wearing detection of shipboard personnel in different environments.

Accurate precipitation prediction is crucial for a range of sectors, including agriculture, water resource management, and disaster preparedness. Traditional meteorological models often struggle to capture the complex spatial and temporal patterns associated with precipitation events. To address this gap, this study introduces a groundbreaking approach that combines Transformer and Generative Adversarial Network (GAN) technologies. The objective is to downscale low-resolution (25km) precipitation data to a finer resolution (8km) specifically for the Beijing region in China. Our proposed model enhances the accuracy of precipitation forecasts by leveraging a hybrid architecture that combines the strengths of Transformers and Generative Adversarial Networks (GANs). The model is particularly effective in downscaling low-resolution meteorological data to high-resolution precipitation forecasts. Comparative analyses with existing models like CorrectorGAN and ResDeepD indicate a significant improvement in forecast accuracy, validating the efficacy of our novel approach.

Faced with the large number of images of electrical equipment in electrical drone inspections, the efficiency of human visual recognition is low, and it is difficult to locate high-temperature defects in electrical equipment quickly. Therefore, this paper proposes a method for screening images of high-temperature defects in electrical equipment. Firstly, a target dataset is constructed, and data augmentation is applied to address the issue of imbalanced target samples. Then, improvements are made to the YOLOv5 network by adding a object detection layer to obtain larger feature maps. Finally, the modified YOLOv5 network is trained and compared with methods such as YOLOv5 and YOLOv4 through experimental studies. The experimental results demonstrate that the proposed method can effectively enhance the network's detection capabilities for targets and achieve accurate identification of electrical high-temperature defects, even under conditions with significant infrared interference.

Conventional UAV flight assignment methods mainly use two-dimensional directed acyclic graph to generate flight task integration matrix, which is easily influenced by the time sequence of flight tasks, resulting in long task assignment time. Therefore, it is necessary to design a new UAV flight assignment method for dense equipment in converter stations. That is to say, the UAV flight task assignment model is constructed for the inspection of dense equipment in converter station, and the UAV flight task assignment algorithm is designed, thus realizing the UAV flight task assignment. The experimental results show that the designed UAV flight task assignment method has a good assignment effect, and the task assignment time under different task numbers is short, which is efficient and has certain application value, and has made certain contributions to improving the efficiency and ensuring the safety of UAV inspection.

A convenient and effective modeling method for contact collision force was explored to address the difficulties in modeling and simulation verification of cam pairs. Taking the L-N contact collision force model into consideration, based on the modified hybrid contact collision force model of Bai, a contact collision force model for the Roller-Cam system is established, and it can be used to solve the collision force and collision deformation. Finally, the influence of the initial collision speed and the radius of curvature for the impact point on the collision force and deformation is calculated respectively, which points out that the initial collision speed and the curvature radius of the cam affect the dynamic response of the collision process, furthermore, the initial velocity of collision has a more obvious impact on the collision force and collision deformation of the system.

To promote energy absorption capacity and the efficiency of virtual power plant (VPP) operation and consider the complex and uncertain characteristics of the VPP system, a three-stage optimal dispatch model is proposed for VPPs. On this basis, a VPP three-stage bidding model is established that simultaneously takes into account electric vehicles (EVs) and demand response. The established model has the characteristics of high-dimensional nonlinearity. The proposed scheme has the effect of peak shaving and valley filling, and can enhance the consumption of renewable energy. Eventually, the effectiveness of the constructed model is verified through the simulation model. The simulation results confirm that the proposed optimization scheduling model has the advantages of efficiency and accuracy.

This article introduces an accident case of a 500kV SF₆ current transformer insulation breakdown leading to fault tripping. The on-site testing and disassembly of the faulty current transformer were described in detail, and the discharge path at the time of the fault was analyzed. Using simulation software to calculate the operating overvoltage value of the fault point during the switching of the filter bank, the cause of the current transformer fault is determined based on the disassembly and simulation results, and operation and maintenance suggestions are provided for SF₆ current transformers operating under the same working conditions.

Deep learning is currently a popular development in the field of artificial intelligence, which is based on the analysis of the process of input and processing information by the human brain, and the encoding and implementation of that process by a computer. There has been a tremendous development in deep learning for image and speech recognition, resulting in a number of related application technologies. FPGAs (Field Programmable Logic Gate Arrays), as a semi-custom, reconfigurable hardware device, are increasingly used in the field of artificial intelligence. Intelligent robots, which combine information processing capabilities such as image, text, speech, position and motion, have very promising applications in industrial production and social life. In this paper, the correctness of the functionality of each functional module is confirmed through simulation, on-board verification schemes are designed and data results are statistically and analytically analysed, demonstrating the stability of the circuit system. The FPGA implementation based on the YOLOv3 algorithm was confirmed to be highly usable by connecting the development platform to an intelligent robot and conducting a comparative analysis.

To reduce the cost of tracking moving targets by robots and enhance the accuracy of target recognition and tracking, a vision-based tracking robot is designed. The onboard camera captures image data, which is initially processed by the ESP32. After processing, the data is transmitted to a cloud environment via a WiFi module. In the cloud environment, a Siamese neural network algorithm is employed to analyze and process the images, and then it sends motion control commands back to the robot, enabling remote control.

The study assesses road safety performance in the Southeast Asian region by using the CRITIC-TOPSIS-Kmeans model. First, the weight of each indicator is obtained by CRITIC. Then, the obtained weights are embedded into the TOPSIS model. Furthermore, the TOPSIS sores are put into the K-means unsupervised machine learning model. The proposed CRITIC-TOPSIS-Kmeans model is utilized to rank and group the road safety performance of Southeast Asian countries. Finally, radar and bar plots are utilized to deconstruct indicators and TOPSIS scores, providing valuable references for policymakers. Overall, the development of the CRITIC-TOPSIS-Kmeans model not only establishes a fresh foundation for evaluating road safety achievements but also offers potential solutions for other MCDM challenges.

A method for analyzing the morphological characteristics of wear particles in marine cylinder lubrication is proposed. The method uses a BP neural network recognition algorithm to identify these particles. The process involves correcting, filtering, binarizing, and processing typical wear particle images through morphological operations. Statistical and Fourier feature parameters extract size and shape characteristics of the wear images, while gray co-occurrence matrix extracts texture feature parameters. Ultimately, the method establishes a wear recognition system based on the BP neural network. Based on the findings, using a sharpened filter and automatic threshold binarization can enhance the clarity of the texture and boundary of the grinding particle. Additionally, performing open and close operations can eliminate any remaining holes and broken points in the image morphology processing, resulting in a smoother edge for the grinding particle. Furthermore, the BP neural network model is capable of automatically identifying the form of cylinder lubrication particles, determining the wear mechanism, and adjusting the cylinder oil filling rate by utilizing the detection results of cylinder residual oil.

Carbon fiber composite materials (CFRP) are difficult to machine due to their anisotropy. PCD tools are widely used in CFRP drilling, and changes in axial force affect the quality of drilling. This article uses a combination of experimental and finite element techniques to study the axial force of PCD tool drilling CFRP. Through experiments, the minimum value of axial force under cutting parameters of 6500rpm spindle speed and 120mm/min feed rate is 124.1N; Based on Deform-3D, a constitutive model of the material was established, and the maximum axial force and torque values were obtained when the PCD tool was about to drill out the CFRP plate. The correctness of the finite element model was verified through experimental data.

Based on the size data of the human body of Chinese male pilots, this article uses the standard model library of Poser software to adjust and modify the dimensions of various parts, and establishes a three-dimensional human model that conforms to the size of Chinese male pilots. Then, based on the basic geometric dimensions of the aircraft cockpit and the ejection seat model, three-dimensional software is used to establish the spatial layout of the ejection seat and cockpit control mechanism, and a static comfort simulation platform for the ejection seat is obtained. Based on the principles of human factors engineering, the layout simulation analysis of the control mechanism in the simulation platform was conducted to obtain the ejection seat control layout that meets the comfort criteria for human riding, improving the static comfort of the ejection seat. This method can be used for rapid design and optimization of cockpit layout, and has important engineering practical significance.

As the number of electrical loads in multi-electric aircraft, the load characteristics and constraints become increasingly complex, making it difficult to achieve effective load management through a single management objective. Firstly, a typical multi-electric aircraft distribution network is analysed to determine the priority relationship of load management. And then, according to the load power demand and load characteristics, the load management objective is abstracted as a multi-objective mixed integer programming problem, and a multi-objective load system management method is designed. Finally, the multi-objective load system management method is implemented using the CPLEX OPL modelling language. The simulation results show that the method can effectively achieve load management, meet the pre-defined requirements of multiple objectives and improve the economy of load management.

The avionics is an important component of the avionics system , and its condition is closely related to the health state of the aircraft. However, nowadays the complexity of avionics systems makes it more difficult to fault diagnose the electronic equipment in avionics systems. In order to solve this problem, this paper discusses the research of common faults in avionics, outlines the difficulties of fault diagnosis of avionics, summarizes the development and research status of fault diagnosis of avionics, and introduces the key technologies of fault diagnosis of avionics from three perspectives: data preprocessing, anomaly detection, and fault location. This paper examines the issues surrounding fault diagnosis of avionics and look ahead the future development direction in this field.

Rolling bearing is one of the important parts of rotating machinery, the monitoring and timely diagnosis of the running state of bearing is of great significance to ensure the safe operation of equipment. Aiming at the problems of multiple parameters and low recognition efficiency of existing intelligent diagnosis models, a rolling bearing fault diagnosis method based on Deep linear discriminant analysis (Deep LDA) was proposed combining the advantages of linear discriminant analysis and one-dimensional convolutional neural network that can effectively extract discriminant features and Deep feature information. After the one-dimensional original vibration signal is preprocessed, one-dimensional convolutional neural network is directly used to obtain the depth nonlinear characteristics of rolling bearings to avoid information loss when the original signal is converted into frequency domain signals. Using LDA, the model can acquire the ability to learn the potential representation of linear separable, and then extract the discriminant fault feature information from nonlinear fault information, and make the extracted sample feature obtain the best separation effect in the sample space, to achieve accurate classification of rolling bearing fault types. In order to verify the classification effect of Deep LDA, two sets of rolling bearing fault data sets were used to verify the proposed model. Experimental results and comparative analysis show that the proposed model can accurately identify different running states of rolling bearings, and has higher recognition accuracy and diagnosis efficiency.

Through the process of 3D reconstruction, the port’s layout and infrastructure can be presented in a visually immersive manner, allowing for a comprehensive understanding of its structure and distinctive features by relevant personnel. This also helps to optimize container storage and stowage schedules, while providing invaluable data support for port management and operations. Furthermore, it elevates the level of informatization and intelligence within the port, thereby playing a pivotal role in the advancement of intelligent shipping. Presently, traditional 3D reconstruction techniques and existing modeling software pose challenges such as time-consuming processes and low efficiency when reconstructing port scenes. In this study, we employ a multi-view shooting video of the port as input and introduce an enhanced Neural Radiation Field (NeRF) technology for 3D port reconstruction. This approach yields a high-fidelity 3D model of the port, which is then compared to conventional 3D reconstruction methods. The improved NeRF technique achieves superior reconstruction results within approximately 2 minutes, boasting a realistic effect and offering distinct advantages.

To reduce cross-coupling effects and enhance the dynamic response of the Single-Inductor Dual-Output (SIDO) Buck converter, this paper introduces an active disturbance rejection control method based on a cascade extended state observer. It starts with establishing a small-signal model and deriving the transfer function of the SIDO Buck converter. Then, it designs self-disturbance decoupling controllers and series-connected state observers, followed by multi-objective optimization of controller parameters using the particle swarm algorithm. Finally, the control approach is compared with Proportional-Integral-Derivative (PID) control through simulations on a MATLAB platform. The results demonstrate that the proposed controller effectively minimizes cross-coupling effects between the converter outputs and enhances disturbance rejection and robustness.

Avionics equipment is composed of multiple layers and complex structure, so it is difficult to research progress based on fault mechanism, and the amount of effective fault data of the model is insufficient, and it is difficult for the general fault diagnosis algorithm to train fault data. In order to realize the fault diagnosis of avionics equipment, machine learning is applied to the fault diagnosis of avionics equipment. Research samples are selected from ground operation simulation data, and an algorithm based on the combination of feature selection and extended isolation forest is proposed to detect and categorize typical faults of electronic modules. It can be well applied to the actual fault detection of avionics equipment . It can meet the requirements of lightweight applications and has practical engineering value.

To achieve the recognition and positioning functions of indoor mobile robots under limited computing power conditions, a method based on color recognition for robot recognition and positioning is proposed. The global image of the robot working in the field is collected by a camera outside the field, and the position of the robot is obtained through computer vision processing. Then, the robot is controlled to move according to this information. Experiments conducted within a 2m × 1m area have shown that the maximum error during robot operation is 7.8cm/m, with an average error of 7.0cm/m. The maximum error of the steering angle is 16.6°, with an average error of 7.7°.

In order to prevent head injuries caused by falling objects or collision with objects when entering construction sites or production workshops, it is stipulated that helmets should be worn when entering the relevant occasions. However, due to the lack of safety supervision system in some construction sites and workshops and the personnel's sense of chance, it is difficult to make all the employees wear safety helmets, which will lay a hidden safety hazard. Therefore, it is of great significance to detect whether the employees wear helmets or not. Aiming at the problems of low accuracy, complex model and large computation of many target detection algorithms for detecting wearing helmets, this paper proposes a high-precision lightweight helmet target detection algorithm based on YOLOv8. The use of deformable convolutional (DCNv3) to replace the residual structure of the original backbone network C2f module effectively improves the limitations of fixed convolutional computation, reduces the number of model parameters and enhances the ability to detect the target modelling, and improves the effect of target detection for helmets and other targets. Moreover, a novel efficient multi-scale (EMA) attention mechanism is introduced, which reduces the computational resources required without reducing the channel dimensions, and improves the feature image pixel-level attention, so that the model pays more attention to the target information of helmets. The model is validated on a publicly available dataset, and the mAP value of the detection accuracy is improved by 2.1% compared to the original YOLOv8n, and the complexity and computation of the model are also reduced. It is also compared with other state-of-the-art target detection algorithms under the same conditions, and the results show that the DE-YOLOv8 algorithm proposed in this paper performs better.

Compared to conventional observations, Unmanned Aerial Vehicle (UAV) observations often result in a large number of small objects that are fuzzy and lack sufficient feature detail. In this paper, a Dynamic Meta Tandem Transformer (DMTT) network is proposed to solve the small objects problem under the view of the UAV field. We design an Adaptive Cross-dimensional Attention (ACA) to dynamically focus on important feature information in multiple dimensions. a Tandem Pooling Layer (TPL) to reduce the number of parameters and maintain high accuracy. Based on ACA and TPL, Attention Meta Feature Fusion (AMFF) and Adaptive Tandem Transformer (ATT) are proposed. AMMF can capture fine-grained spatial information to address the lack of information on small objects. ATT can aggregate global contextual information to enhance the position of small objects. Extensive experiments on the Visdrone dataset, DMTT improves by 1.3% on the AP metrics compared to the state-of-the-art model. On the UAVDT dataset, DMTT improves by 2.3% on the AP metrics compared to the state-of-the-art model. Meanwhile, the FPS metrics of DMTT on Visdrone and UAVDT reach 30.1 and 30.5, respectively. The results show that DMTT has both accuracy and speed. Moreover, DMTT improves by 0.3% (AP 46.9%) on the MS-COCO dataset, which is encouraging and competitive.

Human Body-part Joint Detection (HBJD) has begun to attract research interest in recent years. However, current HBJD methods are usually hard to be applied in real applications due to their complexity. In our work, we concentrate on improving the efficiency of the HBJD and propose the YOLO-HBJD based on YOLOv5-Nano. Specifically, we devise the Feature Holding Down-sampling Module (FHDM) to preserve features of small body parts while reducing computational complexity. In addition, we propose the Context Cross Attention Module (CCAM) to make the YOLO-HBJD focus more on features related to the HBJD. Experiments on the public dataset illustrate that the YOLO-HBJD achieves the best detection performance compared to the comparison methods while reducing parameters and computational complexity by about 90%.

In the integrated coal supply chain, inventory management at downstream power plants is of significant importance. Precise inventory management in power plants ensures their normal operation and provides sufficient capacity to respond to emergency demands, simultaneously, it ensures the coal production capacity and transportation capabilities upstream are not wasted, maintaining a balance between supply and demand, optimizing the supply chain processes, and saving costs. Coal consumption forecasting is a crucial prerequisite for effective coal inventory management in power plants. In recent years, machine learning algorithms have demonstrated strong capabilities in forecasting. However, for some small and medium-sized power plants that are not as fully developed in terms of information technology, they lack a high level of data accumulation, using time series methods for forecasting is a more appropriate choice for such power plants. This paper investigates the prediction of coal consumption in power plants based on Holt-Winters, LSTM, and the ensemble time series algorithm FBProphet. The experimental analysis is carried using historical data of a thermal power plant and the result shows that the FBProphet model has higher accuracy for the prediction of coal consumption. Subsequently, based on the prediction model, this paper also proposes an easily implementable inventory control method, especially suitable for small and medium-sized power plants within the integrated coal supply chain, which provides more guidance for the inventory control methods of power plants.