This PDF file contains the front matter associated with SPIE Proceedings Volume 12799, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Aerodynamic parameters play a key role in determining the trajectory and stability of projectiles. In order to weaken the impact of the projectile movement model, the PSO variant named chaos adaptive mutation PSO with a stated based adaptive velocity limit (CAMPSO-SAVL) is applied to solve the projectile identification equation (based on maximum likelihood criterion), and output the optimal estimation of aerodynamic parameters. Ballistic reconstruction is carried out based on the identification results. The reconstruction results show that the identification results of CAMPSO-SAVL have excellent accuracy and can meet the engineering needs.

The task of crime prediction is to predict the charges of a case based on the given description of the case, which has become a research hot spot. Most of the existing methods use neural networks and machine learning algorithms to predict. The prediction results have high accuracy, but the prediction results are often not well interpretable. In addition, some early traditional methods can explain the prediction results, but their accuracy is often very low. Formal concept analysis is an effective data extraction tool. Therefore, this paper combines formal concept analysis with neural network to improve the accuracy of the algorithm. The experimental results show that our method has higher prediction results and can better explain the prediction results than the original neural network model.

Wind has an important effect on the ballistic characteristics of the uncontrolled projectile. Without obtaining the wind velocity information during the projectile flight, it is difficult to estimate the impact positions of the projectile accurately. To further improve the accuracy of wind speed estimation and trajectory prediction during projectile flight, a novel rotating projectile wind speed identification method based on a multi-objective chaotic adaptive particle swarm optimization algorithm is proposed. Based on the four-degree-of-freedom (4D) kinematic model of the rotating projectile, three aerodynamic parameters coefficients are selected as reference variables, three target functions are defined using the mean-square error (MSE), and two components of the wind velocity along the ground coordinate system are used as decision variables to identify the two wind components. Numerical simulations of the wind speed identification process under variable wind conditions show that the method can effectively and accurately estimate the wind values during the flight of rotating projectiles.

With the enhancement and enrichment of social media, cross network node alignment algorithms are playing an increasingly important role in interest recommendation and security protection. However, the performance of the same user in different online social networks platforms may differ due to the heterogeneous characteristics of the platform or unclear personal characteristics, which correspondingly poses many challenges. Existing cross network node association methods are restricted by the forgery of attributes and the time complexity of processing user generated content, and the overall effect is not ideal. This article mainly proposes a cross network node alignment algorithm based on friend following relationships and username similarity in different networks in online social networks. Using social network relationships to generate model vectors for two networks, and generating transformation matrices based on training sets, calculate confidence based on friend relationships for test samples, and organically combine the similarity of user names between the two networks to make associated user predictions, thereby linking the accounts of entities in the two networks and achieving node alignment. Experimental results indicate the node alignment method can achieve satisfactory performance.

With vehicles becoming a necessity for people's life and work, the vigorous development of intelligent vehicles is of great significance in today's society in order to reduce traffic accidents, lower operating costs and achieve environmentally friendly travel. Motion planning, as the functional core of intelligent driving systems, is crucial to the driving safety of intelligent vehicles. This paper introduces the development status of intelligent vehicles at home and abroad, studies the current research status of mainstream motion planning techniques, analyses the technical challenges encountered in motion planning for intelligent vehicles in urban roads, and illustrates the development goal of establishing efficient and reliable motion planning algorithms in complex scenarios.

Artificial light sources have a profound impact on the night sky, leading to reduced visibility of stars and contributing to light pollution. The environmental, physiological, and psychological consequences of light pollution are severe, including disrupting ecosystems, biological clocks, and endocrine imbalances in humans. Moreover, excessive lighting also wastes energy and exacerbates the energy crisis. This paper collects eight light pollution indicators and proposes a general measurement model for light pollution risk level using an improved PCA clustering algorithm. This paper also proposes intervention strategy optimization model to reduce the risk level, and evaluates the effectiveness of these interventions in Shanghai. Overall, this paper provides insights into the causes and consequences of light pollution, and practical solutions for mitigating its harmful effects.

With the continuous increase in the size and integration of Field Programmable Gate Arrays (FPGAs), the running time of FPGA Electronic Design Automation (EDA) software has become longer and longer, gradually constraining the healthy development of FPGAs. Among the various stages of the FPGA EDA process, the routing stage is one of the most time-consuming, and the routing results directly affect the performance of the implemented circuit. To address this issue, this paper proposes a high-efficiency and high-quality FPGA routing algorithm, EQRouter. EQRouter improves the performance of the routing algorithm by enhancing the maze routing algorithm and the rip-up and re-route strategy in the Versatile Place and Route (VPR) routing algorithm. Experimental results show that, compared to the VPR 7 router, EQRouter achieves a 24.6%, 14%, and 12% reduction in routing time, critical path delay, and circuit bus length, respectively, on 10 MCNC benchmark circuits.

The main work of this paper is to study the error verification problem of radar echo signal DDC algorithm implementation, and DDC has certain innovations in hardware implementation methods. The main design modules of DDC include band-pass filtering, mixing, low-pass filtering, decimation and zeroing. In the design and verification of DDC system, the time-domain dual-frequency chirp echo signal is first simulated, and the DDC processing process is verified by advanced floating-point simulation on the MATLAB platform to verify its feasibility and correctness. Then, through Verilog HDL programming, the FPGA-based DDC implementation is carried out, and the output results of each module are simulated through Medelsim. Finally, the simulation results in each FPGA module are calculated with the simulation results in MATLAB to verify the accuracy of the hardware design.

With the development of the shipping industry, maritime transportation is preferred by more and more people, and ship collision is the first problem to be solved in intelligent navigation. Compared with collision avoidance algorithms for ground vehicles and air vehicles, there are still many problems to be solved in ship collision avoidance algorithms, among which the dynamic nature, safety, and complexity of environments will impact the development of reliable automatic collision avoidance algorithms. By analyzing the latest research on ship collision avoidance, the paper concludes that the current research mainly focuses on the rapidity, stability, and adaptability of various algorithms in complex situations, moreover, some suggestions and research directions are proposed for further research.

As an important raw material in the process of aluminum electrolysis, the quality of prebaked anode is directly related to the technical indexes such as carbon consumption, production stability and aluminum purity of aluminum electrolysis. As an important process in the production of prebaked anodes, the end point of kneading affects the automation of raw anode production process, production efficiency and product qualification rate. In this paper, the collected data of the kneading process are processed and the model is established by using LSTM algorithm in machine learning, which can accurately predict the remaining time of kneading and realize the judgment of the end point of carbon kneading, and provide guidance value to the actual production.

As the input of vehicle suspension system, road excitation directly affects the dynamic response of the system, and then affects the vehicle ride comfort, handling stability and reliability and other indicators. In the field of active suspension control, the road grade recognizer provides road feedforward information to the active suspension control algorithm. For semi-active suspension control, the result of road grade recognition affects the adjustment of parameter weight, and then determines the direction of vehicle ride comfort or handling stability. Therefore, with the development of intelligent vehicles, more and more attention has been paid to the research of pavement grade recognition methods. In this paper, a pavement grade recognition algorithm based on suspension dynamic response is proposed. By extracting the features of sprung mass acceleration signals and inputting them into probabilistic neural network (PNN) for training, a pavement grade recognition classifier is obtained. Finally, through the test set data, it is verified that the classifier has a high accuracy of pavement grade recognition, and can meet the requirements of semi-active suspension control.

In order to solve the problem of misdetection and time-consuming during the data acquisition of the honeycomb optical fiber sensor network, this paper proposes a security information retrieval to acquire the honeycomb optical fiber sensor network method based on fuzzy configuration. Those. A data storage model for a honeycomb fiber optic sensor network is developed. Fuzzy distribution detection is used to optimize the frequency decomposition and decomposition of decision time distributions. A block fuzzy fusion clustering method is used to test group properties and fuzzy data groups, and combined with network region clustering analysis method to discover security information. According to the experiment, the detection performance of the proposed model is more than 92, the accuracy is more than 92. 2, and the longest time is less than 0. 02 seconds, which shows that the detection is improved, accurate and effective.

The genetic simulated annealing algorithm is suitable for optimization problems in complex multidimensional space because of its strong global search capability and the fact that it does not easily fall into local optimal solutions. The simulation results show that the genetic simulated annealing algorithm can effectively improve the pattern optimization performance and reduce the computing time compared with the standard genetic algorithm.

With the development of information technology, the battlefield forms in air defense operations are complex and changeable, the target allocation is subject to many restrictions, and the real-time requirements of operations are higher, which makes the results of traditional air defense operations unsatisfactory. Based on this, an improved genetic algorithm is proposed for air defense target assignment. This algorithm first sets relevant parameters and establishes a fitness function, then randomly generates an initial population, and then uses genetic operators to optimize the population from generation to generation, and finally judges whether the maximum genetic generation is reached, and if so, ends Algorithm, continue to optimize. The experimental results show that the average total benefit value of the improved genetic algorithm is 6.42525, which is significantly better than the traditional genetic algorithm. In addition, compared with the traditional genetic algorithm, the improved genetic algorithm has faster convergence speed and high search efficiency. More importantly, the improved genetic algorithm is more reasonable in target allocation and has better practical value in air defense operations.

With the increasing requirements for describing scene target information in image processing, multi-sensor image fusion has attracted attention and application. Due to the complementary nature of the image characteristics between visible and infrared images, image fusion between the images is possible. This article studies a visible and infrared image fusion algorithm based on NSCT (Non Sampled Contourlet Transform). Based on the original NSCT decomposition, the NSCT decomposition level is adaptively set according to the resolution size of the input source image, ensuring sufficient decomposition for high-resolution images to improve fusion quality, and appropriate decomposition for low-resolution images to reduce decomposition time. For the visible light and infrared images collected by the FLIR AX8 dual band camera, image fusion processing was carried out. Compared with the FLIR AX8 dual band camera, the average gradient index of the fusion image was increased by about 1-2, and the distortion index was reduced by about 20, making it more suitable for observation and subsequent processing.

Ground penetrating radar (GPR) receives the reflected wave from the interface. Therefore, when the data is collected, the process of data collection should be defined in the vertical direction. However, there is a certain deviation between the reflection point and the detection point in the process of practical detection. Therefore, it is only necessary to use the normal plane to pass through the detection point when analyzing, and all the echoes of the reflection point can be recorded. Therefore, in the process of practical detection, the data of GPR should be defined. In general, this detection method is called GPR data migration processing in practice. In this paper, based on the reverse-time migration and F-K migration theory, the original physical model data of tunnel lining cavity is analyzed by MATLAB. By comparison, reverse-time migration obviously raises a higher requirement for velocity model, and the GPR detection data is more suitable for F-K migration processing. In addition, a graphical user interface (GUI) is built to provide users with the development program for user interface, so it is more convenient for non-professional users, who can operate it more easily without the need to know a great number of commands.

Due to the expansion in scale, complexity in structure, and intelligence in functionality of the power system in recent years, the amount of data generated has been increasing daily, leading the power system into a "data explosion" era. Considering the unique properties of electric energy, load forecasting plays an extremely important role in power production. In the current environment, traditional load forecasting methods are difficult to meet the operational requirements of the power system. Therefore, it has become a top priority in the field of load forecasting to effectively establish models for processing large data to improve the accuracy of load forecasting. Based on the current research status both domestically and internationally, this article proposes a method for processing big data by building an artificial neural network using the backpropagation algorithm.

This paper presents a binary orthogonal keying (BOK) incorporating M-ary phase shift keying (MPSK) modulation for Chirp transmission system. The BOK-MPSK symbol applies both BOK and phasing shift keying as modulation schemes, rather than using chirp as a spreading sequence. The system model and receiving algorithm of BOK-MPSK are presented in this paper. A BOK-MPSK signal is recovered by chirp correlation and judgement. The performance of bit error rate for BOK-MPSK has been analysed in contrast with MPSK. Simulations of bit error rate comparison between BOK-MPSK and MPSK are assessed, which are in accordance with theoretical analysis. It indicates that BOK-MPSK has advantageous of demodulation performance as well as bandwidth spectrum efficiency.

At present, the conventional detection methods for abnormal data of power purchase materials mainly use correlation vector machines to extract and reduce the dimensions of data features. Due to less dimensions of data feature extraction, the detection effect is poor. In this regard, a method based on BIRCH clustering algorithm and time series for anomaly detection of power purchase material data is proposed. The data is preprocessed by removing outliers and supplementing missing values, and a time series autoregression model is constructed according to data dimensions to extract the flow characteristics of material data. The determination of abnormal data is realized by using local density threshold. In the experiment, the detection performance of the designed detection method is tested. The final results can prove that the proposed method has a low false detection rate and an ideal detection effect when it is used for abnormal detection of power purchase material data.

This paper presents a KM-DF ranging correction filtering algorithm based on K-means clustering algorithm. The combination of the KM-DF algorithm and SINS technology enables the rapid separation of normal and outlier values of ranging based on the motion state of AUV. Simulation results demonstrate a 12% improvement in ranging accuracy compared to the traditional algorithm.

The normal and stable operation of GIS equipment is related to the safety and reliable operation of the power system. The mechanical sound signal generated during the operation of GIS equipment contains a very rich information on the operation status of GIS equipment. Therefore, in order to accurately monitor the operation status of GIS equipment, based on the artificial intelligence voice print recognition technology, the sound signal feature extraction method under different operating status of GIS equipment is studied, so as to support the fault monitoring and judgment of GIS equipment. This paper proposes the voice print feature extraction method of GIS equipment based on wavelet packet decomposition. The experiment shows that the method can accurately extract the voice print features of GIS equipment and provide support for the next step of fault identification of GIS equipment.

One of the important tasks in target identification is infrared small target detection, but infrared small target is stealthy, often has long distance for detection and noise interference, creating a significant obstacle to infrared small target detection tracking. To the end, this paper proposes to combine the use of YOLOv5 and the centriod tracking algorithm for infrared small target recognition tracking, and to enhance the recognition of small targets by optimization of the up-sampling and down-sampling process. By tracking and recognizing 1799 images, the final mAP was improved by 5.84% using the method of this paper.

The core of future energy Internet is new energy as the main body. With the access of large-scale nonlinear photovoltaic (PV) systems, a phenomenon has been found in recent years: when some PV modules are obstructed, it can cause a "mismatch" in the photovoltaic array. Even if very few components are obstructed, the amount of electricity generated will decrease significantly. Moreover, the photovoltaic output characteristics will show multi-peak characteristics. The traditional tracking algorithm can not effectively track the maximum power. Therefore, this paper proposes a new intelligent algorithm: Lions algorithm. The convergence performance of the algorithm and the influence of each parameter on the convergence are analyzed. Compared with particle swarm optimization (PSO) algorithm, the actual optimization ability of lions algorithm is verified. The lions algorithm is applied to partially occlusive PV maximum power point tracking (MPPT) to verify the effect. The algorithm can also be adopted in other fields that need optimization.

For the issues of cable local heating, isolator ceramic disc breakage, lightning arrester and transformer local breakage among substation power equipment fault diagnosis, this study presented a YOLOv5 target detection approach based on deep convolutional neural network to identify and locate power equipment such as insulators, disconnect switch contacts, bushings, wire clips, lightning arresters and their abnormal heating areas. In the testing stage of the YOLOv5 process, the network divides the image in the database into 416*416 pixels. By using the Yolomark tool to annotate the image, the labels are fed into the deep learning convolutional neural network together with the sample set for training, and the ultimate module is generated after several iterations. Finally, the infrared and visible spectral materials of electronic facilities gathered by the operation and monitoring units with infrared thermographers are used for effect testing. Then the fault diagnosis and grading of various power equipment in substation are carried out. The detective findings are evaluated with the quicker regional convolutional neural network algorithm, YOLOv3 algorithm and SSD algorithm for comparison. Through the experimental analysis, the average accuracy of the model is 96.7%, which can achieve the fault diagnosis target of substation power equipment， and thus enhance the daily running, maintenance and management efficient of the power grid.

A real-time detection algorithm based on the improved YOLOv7 (You Only Look Once v7) was proposed to address the issues of slow detection speed, low accuracy, and easy omission of small targets in traffic sign detection algorithms for intelligent transportation and autonomous driving. The algorithm adopted K-means++ algorithm to cluster initial anchor box sizes of the traffic sign target and used Ghost Net network to replace the backbone feature extraction network of the original YOLOv7 algorithm, reducing the network's parameter size. Additionally, the Convolutional Block Attention Module (CBAM) was introduced in the feature fusion network to enhance the network's sensitivity to key features of traffic signs. Furthermore, a small target detection layer with a smaller receptive field was added to improve the accuracy of detecting small target traffic signs. The SIoU (Scylla-Intersection over Union) loss function was introduced to accelerate the model's convergence. Experimental results showed that the improved YOLOv7 algorithm achieved an accuracy of 96%, a recall rate of 96%, and an average precision of 97%. Compared to the original model, there was an improvement in all metrics. In comparison to YOLOv7, the parameter size and computational cost were decreased by 39% and 32%, respectively. Detection speed was 112.72 frames per second. It can enable real-time and accurate traffic sign detection.

Octane value (RON) is a factor which can represent the combustion performance， a decreased RON loss forecast model based on RFE, XGBoost and LightGBM was proposed in this paper. Firstly we used these 3 effective method to filter the feature subset out, each subset included 30 variables, then picked up the most optimized subset which contained the 30 variables. Using of the XGBoost regression method to establish the RON loss forecast model. This model performed well and stable.

Food adulteration driven by economic interests is an important cause of food safety. Camel milk is widely sought for its high nutritional and medicinal value; some businesses adulterate it for profiteering due to its low yield and high price. Traditional adulteration detection methods rely on supervised learning, which is limited by the data of unknown categories in practical application scenarios, and it is difficult to solve the adulteration problem of category imbalance. For the above scenario, this paper proposes a camel milk adulteration detection framework FIAD based on an unsupervised anomaly detection algorithm, which starts from the perspective of anomaly detection and automatically captures and isolates anomalous features in the data through a tree algorithm without manual labeling of data, directly targeting the adulteration identification problem of category imbalance. We tested the discrimination performance of FIAD in a batch of category-imbalanced camel milk adulteration datasets. At 10% and 20% category imbalance, FIAD achieved AUCs of 0.943 and 0.959 and Recall of 0.915 and 0.949, while occupying less memory, better than eight baseline models. The results show that FIAD has excellent comprehensive identification performance and provides a low-cost and high-efficiency identification method for camel milk adulteration identification.

Aiming at the problem that the density peak clustering algorithm (DPC) can not adaptively adjust the cutoff distance and the clustering center needs to be specified manually, an adaptive density peak clustering algorithm combined with sparrow search algorithm (ISSA-ADPC) is proposed. Firstly, the parameter of local density and relative distance is considered comprehensively, and the adaptive selection of cluster centers is realized according to the arrangement order and downward trend of parameter; Secondly, the homogenized distribution Chebyshev chaotic map is used to initialize the sparrow population, and then the adaptive t distribution strategy is introduced to improve the sparrow search algorithm, which avoids the algorithm falling into local optimization. At the same time, the improved sparrow search algorithm is used to realize the adaptive adjustment of cutoff distance; Finally, Compared with the experimental results of K-means, DBSCAN, DPC algorithm on the artificial and UCI datasets, the proposed algorithm has higher accuracy, adjusted rand index and adjusted mutual information values, and has better clustering performance.

When synthesizing panoramic images from fish-eye images, dynamically segmenting the fish-eye image's valid region is crucial. This paper addresses the challenge of segmenting the valid region in fish-eye images caused by lens glare by proposing a segmentation algorithm that combines adaptive threshold edge detection with improved circle fitting. The process begins by performing gradient and direction calculations on the original fish-eye image to extract edge contour information. Subsequently, the edge contours undergo refinement using the non-maximum suppression method. An adaptive dual threshold is generated using the maximum between-class variance method to enable accurate edge detection. The algorithm further initializes and fits the edge points using the least squares method while incorporating a loss function to enhance the circle-fitting algorithm. Finally, an iterative gradient descent method is employed to iteratively solve for the optimal solution and segment the fish-eye image's valid region. Experimental results demonstrate the algorithm's ability to accurately detect contours and effectively segment the valid region, successfully addressing the challenge of extracting the valid region caused by lens glare.

Conventional cloud server dual end vulnerability detection methods mainly use P2P (peer-to-peer) structure to generate dual end vulnerability detection programs, which are vulnerable to the dual identities of the server and the client, resulting in relatively high vulnerability detection errors. Therefore, a new cloud server dual end vulnerability detection method needs to be designed based on heuristic genetic algorithm. That is to say, the heuristic genetic algorithm is used to build a cloud server dual end vulnerability detection model, and the cloud server dual end vulnerability detection middleware is designed, thus realizing the cloud server dual end vulnerability detection. The experimental results show that the designed dual end vulnerability detection method for cloud server based on heuristic genetic algorithm has low relative error of vulnerability detection, which proves that the designed vulnerability detection method has good detection effect, reliability, and certain application value, and has made certain contributions to improving the security of cloud server operation.

Cigarette empty thin head defect detection is an important step to ensure product quality in tobacco factories. To address the current problems about detecting thin head with little content and low accuracy, a cigarette empty thin head defect detection algorithm based on improved YOLOv5s is proposed. Firstly, a convolutional block attention mechanism is introduced between neck and head to emphasize the extraction of empty thin head defect features; then a weighted bidirectional feature pyramid structure is used to improve the neck network and enhance the feature fusion capability of the model; finally, a lightweight module is designed to reduce the computational complexity of the model. The experimental results show that the improved algorithm can effectively detect cigarette hollow-head defects with an improved accuracy of 2.1%, and the model parameters are only 5.35M, which can provide technical support for subsequent cigarette hollow-head defect detection.

The crystalline state detection of penicillin salt solution is performed on the basis of the acquired images of penicillin salt solution, and in order to improve the accuracy of subsequent detection, an image enhancement algorithm for crystalline state detection of penicillin salt solution is proposed for the problem of poor quality of penicillin salt solution images. Firstly, in order to remove the disturbing factors in the penicillin solution image and extract its target solution region, an automatic image segmentation algorithm based on OTSU algorithm and mathematical morphological processing is proposed; secondly, using the characteristics of three independent channels in HSV space, a linear grayscale transformation and mean filtering method is designed for the H component to realize the color adjustment of the target solution region, and the V component is used to enhance the image brightness by Finally, the enhanced target solution region is obtained by combining the components, and the final enhanced image is obtained by spatial transformation and mapping. The experimental results show that after the algorithm processing, the accuracy of crystallization detection of penicillin solution in uncrystallized state and crystallized state reaches 99.32% and 99.40% respectively, which is 30.72% and 9.22% higher than the accuracy before enhancement, which can effectively improve the accuracy of crystallization detection of penicillin solution.

The conventional binary particle swarm optimization algorithm frequently encounters issues of sluggish convergence velocity and a proclivity to become ensnared in local optima. Within this paper, we present a new binary particle swarm optimization algorithm based on the Cauchy mutation operator and inertia weight dynamic adjustment to address these issues. Firstly, the population is initialized using the Tent chaotic mapping, which preserves the diversity of the population. Secondly, the inertia weight is dynamically adjusted to expedite the rate of convergence and enhance the algorithm's local exploitation capability. Lastly, the Cauchy mutation operator is introduced to improve The formula for updating the velocity, enlarging the search range of particles, enhancing algorithm diversity, and effectively avoiding Being entrapped within local optima. Empirical findings from six benchmark test functions showcase the enhanced capability of the proposed algorithm to evade local optima while simultaneously refining the convergence performance.

As one of the three main insulators, glass insulators are widely used in transmission lines. Due to its own characteristics, glass insulators are prone to self explosion during use, which then cause insulator missing and easily lead to power accidents. With the progress of unmanned aerial vehicle inspection and other inspection technologies, reliable identification of missing insulators based on image recognition become very important. This article uses the most advanced YOLO v5 object detection algorithm to identify the missing of glass insulators. The detection rate for insulators is about 97.2%, and the detection rate for missing glass insulators is about 98.6%, with an average mAP@0.5 of 97.9%.

Due to more bad weather conditions such as fog, rain and snow in the road driving environment, resulting in low target recognition accuracy and missed detection problems in the traffic sign detection algorithm in autonomous driving technology, this paper proposed an improved YOLOv7-based traffic sign detection algorithm. Firstly, the MP conv structure of YOLOv7 was improved and the AC mix attention mechanism was embedded to reduce the feature loss caused during the network feature processing and improve the feature expression capability of the model. Secondly combined with the traffic sign image features, the Kmeans++ algorithm was introduced to re-cluster the anchor frame size and the soft-NMS non-maximum suppression algorithm was introduced. Finally, an image data enhancement algorithm was written in Python to simulate the generation of traffic sign images for foggy scenes to enhance the dataset. Experiments were conducted on the Chinese traffic sign dataset CCTSDB, and the results showed that the improved YOLOv7 algorithm achieved 93.8% recognition accuracy, 91.5% recall, 95.7% average accuracy and 99 frames in complex scenes such as fog, and all indicators were improved compared with the mainstream detection algorithms, which can achieve traffic sign detection in bad weather such as fog.

The occurrence of internal defects is inevitable in the casting process. It affects the reliability and durability of products and even gives rise to potential economic losses and security issues. Therefore, it appears essential to identify and eliminate defects in time. Deep learning algorithms can acquire more effective feature representation in complex scenes. Hence, a lightweight modified DeepLabV3+ algorithm is proposed with the following improvements. First, the ImageNet pre-trained MobileNetV2 is adopted as the backbone network to enhance the model feature extraction and improve the detection speed. Second, the atrous spatial pyramid pooling (ASPP) module is constructed for multi-scale feature fusion using atrous separable convolutions with different dilation rates. Third, the decoder is modified to fuse feature maps of different levels. Moreover, efficient channel attention modules are introduced to obtain more effective features and thus improve the defect edge segmentation effect. Last, a hybrid loss function of Focal Loss and Lovaszsoftmax Loss is used to reduce the influence of sample imbalance on model performance. The experimental results show that compared with the classical algorithms, the modified algorithm has higher accuracy and efficiency in the X-ray casting defect segmentation dataset CDSXray with less computational complexity.

To solve the problems of a large number of parameters, low detection accuracy, slow detection speed of human flow target detection model, this paper proposes a YOLOv5 human flow target detection model based on GhostNetv2. The convolution in the first layer of CBS is retained to replace the remaining convolutions with Ghost Conv, and the C3Ghostv2 module is constructed to replace the original CSP structure, reducing the number of parameters, reducing the calculation cost and improving the calculation speed. Finally, the Deep-Sort algorithm tracks and realizes real-time statistics of people flow. The experimental results indicate that the accuracy of the improved YOLOv5 model is 2.4 percentage points higher than that of the original algorithm, the parameter quantity is compressed by 28 %, and the detection speed has also increased.

To address the current problems of low accuracy and poor precision of anti-vibration hammer line clip gap detection, the thesis proposes an improved Harris corner point algorithm for anti-vibration hammer line clip gap detection. Firstly, the DeepLabV3+ semantic segmentation algorithm is used to pre-process the anti-vibration hammer images, and the traditional Harris algorithm is used to obtain the pixel-level corner points as the central corner points, which are continuously approximated by the least squares iterative calculation, and finally the sub-pixel-level corner point coordinates are calculated to greatly improve the accuracy of gap detection. Finally, an adaptive thresholding algorithm is used to find the best threshold value for each image, which improves the accuracy of detection. The experimental results show that the average offset of coordinates using the improved Harris corner point detection algorithm is 0.2 pixels, which is 0.6 pixels lower than the average offset of the Harris corner point detection algorithm and greatly improves the detection accuracy. Hariis corner point detection using the adaptive thresholding algorithm can greatly reduce the number of pseudo corner points and improve the accuracy of detection compared to the traditional manual adjustment of the threshold.

Edge is the important feature in images, which has important applications in image processing, image analysis and image understanding. Accordingly, Edge detection is the basic problem in image processing and computer vision, and is one of the important research content in image processing. In this paper, the edge detection principle of phase stretch transform is analyzed, and the mathematical expression of s-type phase filter with linear or sublinear group delay is derived. Three new phase filters are designed and good edge detection results are obtained. The phase stretch transform based image edge detection is extended to the fractional-order Fourier transform domain, and the simulation results show that the phase stretch transform in the fractional-order Fourier transform domain can also be used for image edge detection and feature description.

With the rapid development of urban transportation, taxi trajectory data has become an important resource. However, processing large-scale taxi trajectory data requires addressing issues such as error analysis and processing, and traditional data processing techniques are no longer sufficient to meet these requirements. Therefore, in order to efficiently process these massive amounts of data, this paper proposes a preprocessing model for massive taxi trajectory data based on Spark distributed computing. This model designs a dataset based on the elastic distributed dataset RDD, which is used to record the structural information of taxi trajectory data. At the same time, the model provides SQL statements for manipulating and managing taxi trajectory data and returns the processed results in the form of a dataframe. By utilizing Spark's parallel computing power and distributed storage characteristics, this model can effectively solve the preprocessing problem of massive taxi trajectory data. The experimental results show that the model has high reliability and processing efficiency in processing massive taxi trajectory data. It lays a solid foundation for subsequent trajectory data mining and analysis.

The least-square method is widely used in the fast phase unwrapping of optical three-dimensional measurement due to its high-speed and comprehensive characteristics. In order to solve the problem of measurement error propagation in the phase unwrapping algorithm based on least-square method, an improved eight-direction least-square phase unwrapping algorithm is proposed in this paper. This algorithm not only researches the gradient difference between the target phase and its eight-neighbor phases, improving measurement accuracy and reducing measurement errors. The superiority and robustness of this algorithm have been demonstrated through simulation data and measurement of industrial bearing roller end faces

Blockchain is a distributed system. In order to achieve the consistency of data and operation, consensus algorithm needs to be adopted. In the practical application, there are the log replication consensus algorithm raft and state machine replication consensus algorithm PBFT in the consortium blockchain. This paper makes a comparative study based on the principles and characteristics of the two algorithms, and makes a comparative study on the performance with experiments in FISCO BCOS. The results show that the PBFT consensus algorithm consumes more CPU resources than raft due to the complexity of its interaction, and the memory resources occupied by the network using raft consensus algorithm are more than those occupied by the network using PBFT consensus algorithm.

The purpose of this paper is to study the reading recognition algorithm of pointer meters in unmanned pump stations based on YOLOv5. The algorithm includes two key stages: dial area detection, pointer recognition and reading conversion. In the stage of dial area detection, YOLOv5 algorithm is used to accurately detect the dial area in the pump station instrument image. In the stage of pointer recognition and reading conversion, Canny edge detection, Hough line detection and angle method are combined to convert the detected pointer information into instrument reading. Through experimental verification, the algorithm presented in this paper shows excellent performance in the reading recognition task of pointer meters in unmanned pump stations. The experimental results show that the algorithm can detect the dial area of pump station instruments in real time and with high accuracy, and successfully recognize and convert pointer information into readable instrument readings. This provides a feasible solution for the automatic management of the operation of the unmanned pump station.

In the digital image defog algorithm, restoring a fog-free image relies on calculating transmittance and atmospheric light intensity. However, existing algorithms with division calculations are unstable in low-transmittance scenarios, amplifying noise and degrading image quality. Estimating atmospheric light direction and mode inaccurately also affects the final image. To address these issues, this paper proposes an iterative algorithm. It transforms the imaging model to avoid division calculations and selects the top 0.1% brightest pixels to determine atmospheric light intensity accurately. Experimental results show significant improvements in color fidelity, image clarity, and visual effect. The algorithm achieves 5% increase in color reduction, 13% increase in average gradient, and 16% increase in dark channel prior ratio.

In this paper, we show an autocorrelation phase retrieval function (APRF) of light field to improve the rendering quality of novel views for light field rendering (LFR) technology. The phase of light field signal carries more abundant scene information, such as for scenes with occlusion, non-Lambertian, complex geometry and color texture. We apply the autocorrelation theory to convert the phase retrieval of the light field signal into the phase retrieval of the light field signal autocorrelation function, so as to facilitate the phase retrieval in the frequency domain. Therefore, we first derive the autocorrelation function of the light field signal and the influence of scene attributes. Based on the autocorrelation function of the light field, the phase retrieval of the light field signal is performed in the frequency domain for multiple iterations. By the phase retrieval optimization of the light field, the estimation and recovery of the scene information carried by the phase are realized, and the rendering quality of multi-view images is optimized. Additionally, to verify the effectiveness of this method, we compare the APRF with other reconstruction algorithms. The experimental results show the effectiveness of APRF in improving the reconstruction quality.

Versatile Video Coding (VVC) introduces more coding tools to improve compression efficiency compared to its predecessor High-Efficiency Video Coding (HEVC). Intra Block Copy (IBC), a coding tool for screen content, has attracted a lot of attention. IBC uses the hash-based search algorithm that significantly improves the compression of screen content. However, the corresponding calculation is complex and intensive. To solve this problem, this paper optimizes the computing process of CRC in IBC and then analyzes the relationship between the area and throughput of a lookup table. Next, a CRC hardware architecture for motion estimation of IBC is proposed. To achieve high throughput, it adopts 16 bits parallel scheme. The hardware implementation was synthesized using GF 28nm process. The measured throughput can reach 4K@60fps at 527MHz, with a gate count of 47.9k and SRAM consumption of 256 KB. The proposed hardware can meet the requirement of real-time video coding.

Remote sensing product production is an important task in the field of remote sensing engineering. Compared with the production process of remote sensing products, the implementation process of remote sensing product production is a multi-task serial and highly concurrent process that integrates a large number of remote sensing digital image processing algorithms and various processes. Remote sensing product workflows transform complex computing tasks into processes that can be described, managed, and automated. In the workflow, algorithms are the smallest execution unit of tasks, and their execution time has a significant impact on the performance and efficiency of the entire workflow. Therefore, accurate prediction of algorithm execution time is crucial for optimizing workflow performance. This paper proposes and designs an algorithm execution time predictor for remote sensing product workflows. Firstly, feature extraction algorithms are used to extract features from algorithm execution logs, and the prediction model suitable for this scenario is determined by analyzing the characteristics of operators in remote sensing product workflows. Then, the model is trained using historical data sets, and the test set results are compared with the actual execution time to verify the accuracy and practicality of the model. Experimental results show that the proposed algorithm execution time predictor outperforms the baseline model in terms of prediction accuracy and time efficiency

Proteins secreted by various cells and tissues into different body fluids can signal several physiological disorders. However, the degree of complexity in different body fluids and the presence of a large number of proteins in fluids can make studying them with existing proteomics techniques complex and result in large discrepancies between experimental studies. To address this, we developed a deep learning framework called SecBert that identifies secreted proteins in two human body fluids. SecBert uses a sequence-based approach with end-to-end automatic feature extraction for protein classification. Our results show that SecBert performs well, achieving an average area under the ROC curve of 0.94-0.95 on each fluid test dataset.

Images form the basis of human vision and are an important source of information for both human perception and machine pattern recognition. Since the development of the image quality evaluation field, a large number of image quality evaluation algorithms have emerged. In recent years, deep learning has become a hot area and has also been applied to the image domain. This paper focuses on reference-free image quality evaluation and reviews the no-reference image quality assessment based on deep learning. Firstly, the classification of IQA, technical indexes for IQA algorithm evaluation, and several public IQA databases available online are introduced. Then, several deep learning models applied to NR-IQA are discussed, compared, and evaluated in detail. Finally, an outlook on future research is provided.

Aiming at the data security management and control requirements in big data sharing and utilization scenarios, a trusted computing based inter agent interaction architecture is designed, which utilizes integrity metrics to achieve trust chain transfer within and between nodes, ensuring the security and reliability of the nodes themselves and their interactions. Aiming at the privacy protection problem in multi agent sharing scenarios, a remote authentication scheme based on Merkle tree is proposed. The Merkle tree is constructed based on information fragmentation and sequence as the basis for remote authentication, which effectively protects the privacy of information without tampering with its structure and content during transmission.

Aspect Sentiment Triple Extraction (ASTE) is a new subtask for fine-grained sentiment analysis aimed to extract the triples of aspect, opinion and sentiment terms from sentences. Previous work has extracted the target triple in an end-to-end approach, but such rely significantly on the interactions between aspect words and opinion words. Thus, they do not handle well the aspect and opinion terms containing multiple words. Recently, the span-level based models have performed well on ASTE tasks by utilizing the prediction of all possible spans. In this paper, we propose a span-level dual-channel interactive learning network for extracting target triples which uses all possible spans in a sentence as input and dual extraction of target terms using a target-aware attention mechanism to complement each other. Our model makes some contributions to both ASTE as well as ATE and OTE tasks. Experimental results show that the framework achieves better performance in extracting triples.

Aiming at the problem that micro-expression (ME) participants have their own expressions that interfere with microexpression recognition (MER), a dual-stream high-dimensional features implicit augmentation for MER is proposed. The innovative fusion method of adaptive face mask and Action Units (AUs) is used to reduce the interference of the participants' own expressions, and combined with optical flow information to form more focused high-dimensional spatiotemporal information. Enhance high-dimensional spatio-temporal features through an improved implicit semantic data enhancement algorithm to alleviate the problem of small sample data. Using high-dimensional spatiotemporal information to focus the network model on ME active areas. Using a two-stream hybrid convolutional network to extract features from temporal and spatial domain information respectively to improve the representation ability of the network. The experimental results show that under the evaluation criteria of MEGC2019, UF1 and UAR are improved by 5.5% and 6.8% respectively compared with the best method.

Rough sets is a granular computing model mainly used for knowledge discovery. How to quantify the uncertainty, information content, and relationships between knowledge is an important aspect of rough set research. Although many scholars have proposed various knowledge measure models, it is still meaningful to study this issue from different perspectives. This paper constructs a new knowledge measure model. Firstly, we propose a new concept of knowledge distance. Unlike existing knowledge distance models, it no longer requires intersection and union operations on sets in knowledge, the calculation process of computing knowledge distance was simplified. Then, we studied its properties and found that it can effectively characterize the essential features of knowledge in approximate spaces. Finally, we discuss its relationship with information measure and knowledge granularity, proving that it is both an information measure and a knowledge granularity measure. These research results will help people more conveniently and intuitively measure and understand the uncertainty and information content of knowledge in approximate space.

Molecular property prediction can be applied to discover new drugs, which has attracted significant attention from both chemists and machine learning researchers. However, the existing methods for predicting molecular properties have the following limitations. The molecular representions learned from 2D structures will lose some important information contained in the molecular representions learned from 1D SMILES strings and vice versa; Second, the few-shot learning for molecular property prediction ignores the great differences between properties of different moleculars. To address the above problems, we propose a molecular property prediction model based on few-shot learning, called FewGS in this paper. We present a molecular graph-string combination method to fuse SMILES and graph represention, which can exploit a small amount of available molecular information to capture the hidden feature of moleculars. To promote the few-shot learning for molecular property prediction, we also propose a sample space transformation strategy to effectively eliminate bias between meta-train data and meta-test data. In addition, we construct a loss function based on graph-string combination, and adjust the weights through a self-attentive mechanism to achieve accurate prediction. We evaluate FewGS on molecular datasets and experimental results show that FewGS achieves state-of-the-art performance.

Objective: The aim of this study is to develop a classification model based on word attributes and address the classification problem of word difficulty. Methods: We constructed word indexes for difficulty classification based on structure and usage habits, and then trained the SVM classification model using the constructed database. The optimal penalty coefficient C was determined using Monte Carlo simulation with the dichotomous method to optimize the iterative process. Finally, we evaluated the model objectively using five-fold cross-validation and model comparison. Results: The experimental results showed that the highest classification accuracy of 98.4% and 82% for the training and prediction sets, respectively, was achieved when the optimal penalty coefficients were calculated to be (944, 273, 756, 1064, 671). The average classification accuracy obtained by five-fold cross-validation was 0.892. Comparison with other classification models indicated that this model improved the classification accuracy. In summary, the M-SVM model showed better robustness in solving the multi-classification problem, to some extent avoiding the overfitting phenomenon and demonstrating a better classification effect.

To address the issue of interference caused by power line channel noise on time transmission signals, a power line time transmission technology based on OFDM modulation is proposed to use linear frequency modulation (LFM) and direct sequence spread spectrum (DS) techniques for its subcarriers. The simulation results show that the power line time transfer technology using LFM-OFDM and DS-OFDM modulation can effectively reduce the bit error rate of the received signal. At the same time, compared to DS-OFDM modulation technology, the power line time transfer technology using LFMOFDM modulation has a lower bit error rate of the received signal.

The traditional Direction of Arrival (DOA) estimation algorithms are based on model parameters, which depends on the accuracy of the array model. When the array model has errors, the matching between the model and the data will fail, which affects the estimation performance to some extent. Therefore, this paper constructs the nonlinear relationship between the received signal and its spatial spectrum through the neural network framework, and uses the data-driven of deep learning to realize the DOA estimation. The neural network consists of an autoencoder network and multiple parallel 1-D VGG networks to achieve spatial spectrum estimation of the angle region. The simulation results show that the DOA estimation method proposed in this paper has good generalization ability, and also show good robustness under array error conditions.

To address issues in bull face detection, such as unsatisfactory inspection results and fragile inspection devices, we conducted a comparative study on existing representative deep network models (e.g. Mask R-CNN, YOLO, SSD, etc.) based on big data analysis discrepancy theory. The experimental results show that Mask R-CNN has the most efficient comprehensive analysis, although its detection rate is relatively slow. YOLO has a higher detection rate but less efficient comprehensive analysis compared to Mask R-CNN. SSD has the lowest comprehensive efficiency among the models.

Aiming at the low efficiency of fault identification in the process of engine fault identification, an embedded identification system of motor vehicle engine fault based on deep learning is designed. A hardware system consisting of sensor, core processor and communicator is established. The sensor module contains qualitative sensor and quantitative sensor. The qualitative sensor adopts LEI sensor, the quantitative sensor adopts TDK sensor, and the core processor is used as the central processor. The insulation fault information of three bands of high, medium and low frequency is decomposed, the feature vector is determined, and the fault information extraction program is established to complete the design of embedded identification system for motor vehicle engine fault. The experimental results show that the designed system has the best denoising effect, retains the overall trend of the original signal to a greater extent, and presents the detailed information in the signal well, which improves the fault identification accuracy and has stronger fault identification performance.

Tor, as an anonymous communication network tool, presents challenges in terms of identification due to its anonymous and multi-encrypted characteristics. The advent of deep learning has provided new solutions for identifying this type of traffic. However, in practical scenarios, sharing large amounts of traffic data is often not feasible. To address this issue, this paper proposes a Tor traffic identification scheme based on federated learning. By processing distributed data on federated learning clients, we can extract bidirectional temporal information from the traffic and perform model training. Utilizing federated learning, we can aggregate model parameters and achieve distributed identification and classification of Tor traffic. Experimental evaluation was conducted using a publicly available dataset to validate the effectiveness of our proposed approach.

In this paper, the specific approaches, advantages and disadvantages of radar stealth measures are summarized from the aspects of contour stealth technology, structural and coated radar absorbing materials, electromagnetic metamaterials and other aspects. The characteristics of frequency selective surface technology, metamaterial technology and other technologies, the mechanism of action to achieve stealth and the achievements in stealth field that have been developed in recent years are emphasized.

Despite the significant progress achieved by visual Transformers, there are still some limitations that need to be addressed in human pose estimation. Firstly, Transformer lacks CNN’s inductive bias and local feature attention capabilities, which require extensive training data and iterations to achieve satisfactory results. Therefore, we propose a hybrid network that combines convolutional and Transformer. Besides, to address the recognition of human body images at different scales, we established a Transformer pyramid structure, which achieves recognition of human body images at different scales through progressive reduction of the input resolution. Specifically, our algorithm achieves an accuracy of 77.3% with a computational complexity of 19.6 GFLOPs. Compared to traditional direct regression methods, our algorithm considerably enhances detection accuracy while reducing the training complexity and significantly increasing the detection speed compared to traditional Transformer methods.

In order to improve the efficiency of bullet marks inspection and meet the requirements of timeliness and accuracy of bullet marks inspection, it is urgent to make use of the powerful performance of artificial intelligence to develop a deep learing-based gun rifle-comparing system that can achieve accurate identification, rapid inspection and scientific qualitative. To achieve rapid combination and rapid tracing, to provide strong clues and technical support for rapid case detection. This paper introduces the development background of traditional bullet mark inspection, summarizes the existing automatic recognition technology of bullet mark inspection and its shortcomings, analyzes the application status of deep learning in bullet mark inspection, and finally prospects the application of deep learning in rifling comparison technology of bullet.

In recent years, intelligent fault diagnosis methods based on deep learning have achieved remarkable results in both theoretical research and engineering applications, and rolling bearings are one of the core components of rolling mills, and accurate fault diagnosis of rolling mill bearings can effectively guarantee the safe operation and production efficiency of rolling mill equipment. The current deep learning-based fault diagnosis methods are usually unstable in the training process and the models are difficult to converge, resulting in strong randomness in engineering applications. Optimized algorithms are therefore key to the training of deep learning models. Different optimized algorithms have different effects on the stability of training. The choice of learning rate is also crucial, as a large learning rate may lead to unstable training. In addition, regular techniques can also be used to improve the stability of training. This paper proposes a deep fault diagnosis method for rolling mill bearings based on an improved optimized algorithm, which improves the training efficiency, the stability of the model output results and the robustness of the model with respect to parameter changes while ensuring the diagnostic accuracy of the model.

In image recognition, Faster R-CNN detection algorithm based on R-CNN has received more and more attention in recent years. Compared with traditional image processing algorithms, this type of algorithm can extract deeper pattern features in complex embroidery images and improve the robustness and recognition accuracy of the algorithm. In order to extract deeper pattern features in embroidery images, this paper uses Faster R-CNN networks to recognize different types of embroidery patterns, and uses ResNet50 and VGG16 as Faster R-CNN feature extraction networks respectively for comparison. The experimental results show that the Faster R-CNN network based on ResNet50 has better image recognition effect after pre-training.

With the rapid development of machine learning methods, deep learning models based on big data and computer arithmetic are widely used in various scenarios, including target detection and image recognition. In order to detect surface defects of the power machinery more rapidly and accurately, this paper investigates the performance of various convolutional neural network models in the task of detecting surface defects, identifies relatively excellent detection models based on the confusion matrix, and summarizes the more appropriate detection scenarios for different models. The results indicate that EfficientNet has good comprehensive performance in detection tasks.

Infantile hemangioma (IH) and venous malformation (VM) are common vascular diseases in childhood, and they both affect the growth and development of infants and children. Because of the complexity of IH and VM ultrasound images and/or insufficient clinical experience, they are often misdiagnosed, and then patients cannot get good treatment. To solve this problem, this paper develops an intelligent diagnosis system for vascular diseases based on deep learning. The deep learning networks have been trained for each vascular disease with numerous ultrasound image data. The special model for each category of vascular disease is then set up and can be applied to recognize IH and VM on input ultrasound image data. The development and application of this system can both reduce the workload to clinicians and improve the accuracy of diagnosis.

To facilitate the processing of data pertaining to defects and faults in the power grid regulation system, as well as knowledge graph-based support to help regulators in their decision-making, a knowledge graph construction method employing a fuzzy matching algorithm has been devised, the outcomes are presented in the form of a visual knowledge graph. Initially, a fuzzy matching algorithm is devised to to match the defect with the defect-fault database. Subsequently, the Neo4j database is utilized to construct a knowledge graph that provides guidance on fault locations. Lastly, by conducting a case analysis of the data extracted from a specific region's distribution network, experimental results demonstrate the method's capability to accurately match target data. Moreover, the visual knowledge graph constructed has proven effective in enhancing regulators' decision-making efficiency.

Histopathological images are one of the common diagnostic tools used in the medical field. However, the complex cellular and tissue structures inside colorectal cancer histopathological images make it difficult for physicians to perform efficient cancer diagnosis and quantitative analysis. Therefore, this study proposed a multi-task learning-based pathology image classification method Multi-TransPathoNet (A transfer learning-based neural network for pathology image classification with multi-task learning, Multi-TransPathoNet) for detection and visualization output of colorectal cancer histopathology images, thus helping to improve medical efficiency. To verify the practical application value of this method, this study used the paracancerous, highly differentiated, and high-moderately differentiated histopathological images which are more difficult to distinguish in colorectal cancer as the dataset for classification test, and our method obtained 92% accuracy. Besides, we compared with the single-task learning models Resnet50 and Vgg19 respectively, which also have good performance in image classification tasks, to comprehensively evaluate the proposed model’s performance.

This paper aims to research and develop a facial expression recognition algorithm based on convolutional neural networks (CNN) and apply it to a real-time facial expression recognition system. Firstly, we introduce the Local Binary Pattern (LBP) and CNN as commonly used facial recognition algorithms, discussing their characteristics and application areas. We then describe the experimental process in detail, conducting emotion recognition experiments using the FER2013 dataset and performing data preprocessing and augmentation steps. We design a CNN-based model architecture and set the parameters of the model. In the results and analysis section, we compare the performance of our model with other baseline models and discuss possible directions for improvement. Finally, we conclude that our algorithm achieves high accuracy in the experiments and provide prospects for future research and applications.

Optimized data analysis methods can enhance teaching system efficiency. The optimization algorithm has spontaneous recognition of high-dimensional data sparsity in the density peak clustering information. This study proposed a deeplearning algorithm in the density peak clustering analysis of the IUR teaching system, which has the advantages of fast calculation speed and dynamic data update. The research is based on the k-means clustering method to select IUR teaching system loads, and the pheromone parameter is used as an influencing factor in the clustering process. The proposed algorithm can better implement teaching data analysis across systems. In addition, the improved algorithm has low complexity and strong operability. The innovative elements of the IUR teaching systems such as renewable subjects, R&D resource input, and innovative teaching management can be controlled collaboratively. In the experimental analysis, the proposed algorithm can further strengthen the quantitative analysis of the IUR teaching system configuration, improve the refinement method for the effective control and distribution of innovative teaching systems, and fully exploit the data resources for the combination of innovative applications and teaching systems development.

Big data technology is affecting people's lives, leading to an increasing desire for diversity in books. The traditional recommendation algorithms based on data mining techniques can no longer satisfy users to find their favourite books quickly and precisely. Therefore, considering the shortage of traditional book recommendation algorithms in terms of data usage and implementation effect, this paper proposes a hybrid recommendation algorithm that incorporates content and improved item-based collaborative filtering. Firstly, the algorithm uses an improved item-based collaborative filtering algorithm to make recommendations, and the improved collaborative filtering algorithm introduces a common rating weight function on the basis of the traditional algorithm to solve the problem of data sparsity. When the items are less similar, the content-based recommendation algorithm is used instead to make recommendations. In this paper, a cascading approach is undertaken to fuse content with an improved collaborative filtering algorithm, which improves the problem of cold start and data sparseness to a certain extent. Experiments show that the algorithm proposed in this paper can efficiently enhance the accuracy and recall rate of recommendation results, and has excellent recommendation effect.

Information technology has been developing rapidly in recent years, the application of computer aided design (CAD) in architectural design is more and more. Based on the computer-aided design method, designers can parameterize the design scheme in the software, and use a series of variables to control and transform the design scheme, which brings great convenience for the architect to optimize the process of architectural scheme. But the building performance optimization method based on intelligent algorithm is mainly used in scientific research and is difficult to put into practical engineering. Therefore, for the purpose of solving the above problems, this paper carries on the mathematical modeling to the building performance prediction problem which takes the building spatial parameters as the optimization variables. Based on BP neural network, this paper builds an accurate and efficient building performance prediction platform to assist designers to make decisions. The simulation results show that the prediction and optimization proposed here can give accurate and reliable optimal solutions, and the optimal construction scheme has strong rationality and high engineering application value.

Regional location of plant diseases helps to develop targeted prevention and control strategies to improve crop production efficiency and reduce agricultural losses. Recently, some important advances have been made in computer vision-based methods. However, most of the current mainstream methods are based on target detection or rotation detection, which always include environmental noise in the results of locating disease areas. In this paper, we revisit the task from the perspective of segmentation tasks and propose a feature reinforcement module. Specifically, we effectively focus on the shortcomings of previous methods where neighbourhood features cannot interact by looking at the shortcomings of the Feature Pyramid Network (FPN) in the feature extraction process through multi-scale features repetitive sampling. In addition, we compared the dataset annotation methods under different annotations such as target detection, rotation detection, and segmentation, and the visualisation well demonstrates the need to use segmentation methods. In the final experimental results, our method is proved to improve mAP by 1.3% and mIOU by 1% over state-of-the-art methods. A large number of methods have demonstrated the superiority and reliability of our method.

Wordle is a popular game, and it is necessary to analyze the massive amount of data to help improve the game. For the prediction problem of the number of reports, the three-times exponential smoothing prediction model with better prediction results was selected. This study collected the relationship between word attributes and the percentage of the number of reports in the difficult mode. After data preprocessing, the RMAE test and 𝑅² test were performed on the prediction effects of the two models; For the problem of predicting the percentage of attempts (1, 2, 3, 4, 5, 6, X), a BP neural network model optimized using a genetic algorithm was built and the results showed that the confidence level of the model was 0.742, which tested by prediction of the percentage distribution of words corresponding to the number of attempts per day; For the word difficulty prediction grading problem, this study built a hierarchical clustering model and solved the optimal number of clusters K=4 for word classification according to the elbow rule, and obtained the classification results using the improved BP neural network model to train the data. The results showed that the accuracy of the model was 0.831. This study can improve Wordle, enhance the user's sense of game experience, and attract more people to participate in the game.

We learn feature representations on unlabeled task sequences. In order to make the disparity between unsupervised representational learning and continuous learning narrow, interpolate examples of current and previous tasks to build new examples. On this basis, we suggest a innovative knowledge extraction method which takes into account feature location and distance function of the extraction. We use the position before ReLU as the distillation point and design a new margin ReLU function. This allows for the centralization of useful information in the middle of the network and further performance improvement, minimizing the forgetting of past activities while increasing learning of new ones.

Based on the actual demand of the public for 3D content generation technology and targeting the problems existing in traditional 2D to 3D image conversion methods, we propose a 2D to 3D pedestrian detection algorithm based on deep learning. The algorithm uses object segmentation to first extract the human body's 2D outline from the input image. It then combines shallow features of the image to use a fully connected layer to map latent vectors to low-dimensional mesh vertex feature vectors, ultimately generating a triangle mesh for the 3D model. By alternately using SpiralConv and upsampling operations, the 3D mesh of the pedestrian can be generated more accurately. Experimental results show that the algorithm can accurately construct a 3D model of a pedestrian and has real-time feedback capability. The algorithm is suitable for real-time monitoring scenarios such as action recognition. Compared to traditional 2D to 3D image conversion methods, the algorithm uses deep learning to automatically learn features, which has higher accuracy and transferability. Additionally, the algorithm extracts 2D contour information of the human body using object segmentation, which improves the accuracy of generating 3D meshes. We believe that this algorithm has an important role to play in the development of 3D content generation technology.

Landslide spatial prediction is of great significance for disaster prevention and mitigation. Traditional landslide spatial prediction methods rely on the weights of characteristic factors subjectively set by domain experts, and do not consider the correlation between historical landslide cases. This paper combines the knowledge graph link prediction task with the landslide space prediction task, and proposes a landslide space prediction method based on landslide cases knowledge graph and RGCN (Relational Graph Convolutional Network). First, the conceptual modeling of landslide cases is realized by analyzing the domain knowledge of landslides, and the storage of historical landslide cases data is carried out with the graph database as the carrier; secondly, the feature information of the neighborhood nodes and edges of entities in the knowledge graph is aggregated by RGCN, and the graph structure information in the landslide cases knowledge graph is effectively used; finally, the specific cases of loess landslide are selected as the reasoning object to verify the proposed method. The results demonstrate the effectiveness and accuracy of the method in landslide spatial prediction tasks. At the same time, it also provides a new idea for the pre-disaster early warning and hidden danger investigation of landslide disasters.

Nonnegative matrix factorization (NMF) is a powerful method for feature extraction, offering explanatory and dimensionality reduction. Alternatively, combining NMF with a neural network requires iterative optimization of an objective function, followed by constructing a specialized neural network based on the derived formula. The interpretability and universality of this approach are limited. To address these issues, this paper introduces a novel model called FCNMFN, which leverages a fully connected neural network to implement NMF. In this model, each layer of the fully connected neural network corresponds to the transpose of the base matrix, the coefficient matrix, and the sample matrix of NMF. This design ensures strong interpretability while achieving nonnegative matrix factorization. To demonstrate the effectiveness of the proposed model, we apply it to emotion recognition using the DEAP dataset. Experimental results confirm its efficacy and showcase its potential in accurately identifying and analyzing emotions.

With the development of deep learning, lip recognition technology has made great progress in English, but there is a certain gap in Chinese in both data set richness and recognition accuracy. By analyzing the visual characteristics of Chinese pronunciation, this paper puts forward the pinyin sequence of picture frames in order to avoid the ambiguity of Chinese visual expression. In order to verify the validity of pinyin sequence of picture frames, a pinyin sequence recognition model combining 3D+2D convolutional neural network with Bi-ConvLSTM and a Chinese character prediction model PPTC are proposed. Using the pinyin sequence of picture frames as the medium, the picture frames are converted into Chinese sentences. Through experiments, the superiority of the model is proved, which provides a benchmark for future work.

The iterative algorithm used in traditional topology optimization methods takes a long time. With the increase of the degree of freedom of the finite element mesh, the calculation time will also exhibit exponential growth, resulting in the problem of dimension explosion. To solve this problem, this paper adopts a fast topology optimization method based on Convolutional Neural Network (CNN). By comparing the optimized results and the computational costs between the traditional SIMP method and the Level Set method, this paper proposes to use the fuzzy images obtained from the initial iteration of the SIMP method as training dataset and uses CNN to train them. Based on the fast prediction of the training model, reasonable structural topology can be obtained with a small number of iterations. Based on the classical 88 lines of MATLAB code and the discrete level set MATLAB code, the differences between the traditional iterative optimization algorithm and the CNN-based optimization algorithm are compared by classical cantilever beam examples used as benchmarks. In terms of computational efficiency, these examples are iterated up to 247 times by using the SIMP method and 74 times by using the Level Set method. These two methods took a maximum of 20.72 seconds and 29.36 seconds, respectively. As an improvement, the CNN prediction mode only iterated 10 times with a calculation time of 1.02 seconds. Meanwhile, the model prediction time was negligible. Compared with the traditional iterative modes, the optimization method coupling with SIMP and CNN can significantly improve the computational efficiency.

In order to enhance the security of encrypted images, a new quantum image encryption algorithm is proposed. This algorithm combines Chen's hyper-chaotic system with Double-random Phase Encoding (DRPE). The pixel values of the original quantum image are diffused using DRPE, while the pixel positions are scrambled through Cyclic Translation (CT). The hyper-chaotic sequences generated by Chen's hyper-chaotic system control the scrambling and diffusion operations. Numerical simulations and theoretical analyses have shown that this algorithm is highly effective in resisting illegal attacks. The encrypted image generated by this algorithm exhibits lower pixel correlation and more evenly distributed histogram. Overall, the proposed algorithm represents a robust and secure approach for image encryption

With the continuous development of the Internet+ education concept and the rapid advancement of online, offline, and blended teaching technologies, the existing methods of evaluating students based on teachers' subjective opinions are insufficiently objective and lack a solid foundation. To improve this situation, we propose a classroom student behavior detection model based on coordinate attention and fused residual networks, along with its talent evaluation system. The classroom student behavior detection model, based on coordinate attention and fused residual networks, effectively combines the coordinate attention mechanism and residual networks. It highlights the relevant information in student image samples by effectively utilizing channel and spatial information. The talent evaluation system accumulates and analyzes classroom data through the first application terminal, forming a learning quality report that judges the quality of talent training and provides type information of talent training. Additionally, the system receives human resource information data through the second data terminal and performs employment classification based on talent information data, obtaining occupational tendency data. Therefore, this thesis will help enhance the reliability and accuracy of human resource judgment and increase the compatibility between human resources and careers.

Deep neural networks of training require a large quantity of effectively labeled datasets, and models trained with insufficient samples will cause a variety of problems, such as overfitting and poor generalization ability. Few-shot learning has been proposed to settle the problem, and the methods introduced in some papers have achieved impressive results. However, most of the methods have poor robustness due to the under-utilization of the global information and local information of the image. For solving the above problems, an improved classification network is proposed in this paper, which introduces two modifications. On the one hand, Convolutional Block Attention Module (CBAM) is embedded in the network for extracting global features, thus the model can quickly suit various tasks. On the other hand, the image is randomly divided into tiles and represented as a set of local features, the Earth Mover’s Distance (EMD) is employed to calculate the distance between embedded sets of two images. Finally, the weighted sum of these two parts’ classification errors, is computed as the total classification error of the proposed model. Experiments are conducted on two benchmark datasets: mini-ImageNet and tiered-ImageNet, and the results demonstrate that our proposed method achieves state-of-theart(SOTA) results in the field of few-shot image classification.

Shortwave signal type recognition plays a very important role in the field of non-cooperative communication, but because of the low signal-to-noise ratio of the shortwave channel, it is often very difficult to identify the shortwave signal type. In this paper, the difference between signal and noise in the wavelet transform is utilized, and the method of correlation denoising is adopted to process the wavelet transform sequence. The correlation denoising wavelet and is taken as the signal characteristics so as to effectively reduce the influence of noise on the signal. Through the experimental performance analysis of 6 kinds of 8PSK signals, the correlation denoising wavelet and feature extraction method proposed in this paper can effectively reduce the impact of noise, and still be effective under the condition of low SNR. Finally, by using the features proposed in this paper as the input of the deep learning network, the recognition accuracy of 98.5% can be achieved when the SNR is 6dB.

In this paper,wavelet transform of vibration signal of CNC machine tool is performed based on Python programming language.wavelet transform function in PyWavelets library,the signal is decomposed into sub-signals of different frequencies.The experimental results show that the vibration signal of CNC machine tool has multiple frequency components,and the noise reduction effect is obvious after wavelet transform of the By implementing wavelet analysis method,EMD[1],and EEMD through Python programming,the three feature extraction methods are compared,and it is concluded that wavelet transform can be more convenient for signal processing and analysis,and the analysis method and parameters Therefore,the wavelet analysis method of vibration signal of CNC machine tool based Therefore,the wavelet analysis method of vibration signal of CNC machine tool based on Python has high practicality and scalability,which can provide important support for the optimization design and machining accuracy improvement of machine tools.

The method of trajectory distribution fusion based on Gaussian mixture model (GMM) is suitable for narrowband target tracking system. Aiming at the problems of inaccurate, ambiguous and narrowband tracking of wide-band tracking results, Aiming at the problems of inaccurate, ambiguous and narrowband tracking of Wide-band tracking results, an algorithm based on hybrid Gaussian model is proposed for automatic narrowband target tracking. An algorithm based on hybrid Gaussian model is proposed for automatic narrowband target tracking. Target azimuth distribution is regarded as mixed Gaussian model, Target Azimuth distribution is regarded as mixed Gaussian model, and the expectation maximization algorithm is used to estimate the parameters of the mixed Gaussian model. And the expectation maximization algorithm is used to estimate the parameters of the mixed Gaussian model. The Gaussian mixture model is used to cluster target azimuth. Target weighting method is used to fuse the target azimuth, and get clear and stable target tracking results.

Visual Simultaneous Localization and Mapping(vSLAM) is considered the foundation technology for augmented reality and mobile robotics, and a number of impressive systems have been developed over the decades. However, some issues are still not well addressed, such as poor positioning accuracy and even tracking loss in low texture and dynamic environments. We propose a robust semantic visual SLAM for indoor low-texture and dynamic environments. It uses deep learning image feature point extraction method to replace the traditional feature point extraction method and uses case segmentation network to provide image semantic information combined with motion consistency check to eliminate the influence of dynamic objects. Finally, we verify the accuracy and robustness of the proposed system on the TUM RGB-D dataset. The results show that the proposed system has higher accuracy than ORB-SLAM2 and other semantic SLAM.

Android is currently one of the most popular mobile operating systems in the world. The popularity and flexibility of the Android platform makes it a prime target for malicious attackers. Android malware with malicious charges and privacy theft poses a serious threat to users' security. With the rapid update of Android apps and the growing number of malware types, existing Android app stores all lack comprehensive malware detection methods and prevention mechanisms. Previous studies have shown that the detection methods based on feature codes are easily bypassed by deformed malicious codes and rely on the feature code database. However, the detection method based on machine learning can deal with different varieties of malicious code to some extent, but this method needs to transform the application program into feature vector through feature engineering, and then build a classifier to detect and classify the malware quickly. In the field of machine learning, the feature information selected by traditional feature selection methods suffers from low contribution value and high redundancy, which leads to the low accuracy of detection models. The classical genetic algorithm also has disadvantages such as slow convergence speed, poor local search ability, and easily influenced by parameters. Therefore, this paper proposes an improved hybrid adaptive genetic algorithm (HAGA) as feature selection and combines with a machine learning classifier, and then uses the accuracy of the classifier as an evaluation criterion to provide an effective method for Android malware detection.

Dance performance provides an effective way to express human emotions. Compared with voice and facial expression, it is clear that the regularity of posture change is hard to capture. However, changes of posture can make emotional expressions more vivid in nature, which do well in the analysis of emotional states. In this paper, we propose a hybrid feature for emotion classification in dance performance by means of a well-designed feature fusion network. The novelty of this method is the joint prediction of discrete emotions and continuous emotions, which improves the accuracy and robustness of emotion recognition methods. Finally, we conduct experiments on the official dataset, and the effectiveness of our proposed method is clearly demonstrated through quantitative and qualitative experiments.

To address the problem that some source domain samples that are difficult to transfer disturb the target domain data distribution due to the difference in transfer value between different motor imagery electroencephalogram (MI-EEG) sample data, and that the model has poor feature extraction and classification performance when adapting to different motor imagery datasets, this paper improves the conditional domain adversarial network (CDAN) method introduced by domain generalization technology, and proposes a conditional domain adaptation network based on sample weight (SW-CDAN) method. This method makes the entropy output by the domain discriminator as the sample weight, which is used to adjust the classification loss during the model training process, so that the model can extract transferable features from the common features of the data, thereby enhancing the model’s category prediction ability and model generalization ability. The experimental results show that the SW-CDAN method can effectively improve the classification performance and model generalization ability of motor imagery EEG signals, so that even when facing a small amount of motor imagery EEG signals with low effective components, it can still maintain a high classification accuracy. The SWCDAN method achieves relatively high classification accuracy on BCI Competition IV 2a dataset, which is about 1.87% higher than CDAN method respectively.

More and more people are choosing two-wheeled motor cars as their primary mode of transportation as society develops quickly. As a result, concerns about rider safety have grown over time. Using a helmet may successfully protect the rider's head, lower their chance of suffering a head injury, and lessen their exposure to traffic accidents. The electric bike riders in the images without helmets were found using detection and recognition techniques, and their license plate data was gathered. This technique can successfully identify motorcyclists who disobey the law and give the appropriate authorities crucial information to develop more exact safety standards and enforcement actions. Monitoring the level of helmet use among drivers of motor vehicles is therefore crucial. This study compares and contrasts the key technologies and provides an overview of the current helmet-wearing monitoring systems. The rider helmet wearing monitoring system is first introduced, along with its use cases and importance. The approaches are then distilled into two groups in accordance with the detection techniques' underlying operating principles. The benefits and drawbacks of the two common approaches are then compared, along with potential adaptations. The prospect and summary of the rider helmet wearing monitoring system's use and research directions follows.

As the Internet continues to evolve, network attacks are becoming increasingly complex, and malicious traffic is taking on more diverse forms. A key factor in the correct identification of whether traffic is benign or malicious is the method used to extract traffic features. While researchers have investigated the extraction of robust features from raw traffic data, most have focused solely on individual packets and have not considered the semantic representations of bytes in the same location across different packets. In this paper, we propose a novel method for extracting aligned byte stream features, which involves splitting packets into several parts to extract aligned bytes. Based on this method, we propose a deep-level network-based flow feature aggregation approach for detecting malicious traffic, which uses multiple LSTMs to aggregate flow features and employs deep-level networks for detection. Extensive experimental evaluations demonstrate the effectiveness of the proposed method.

The previous selection and layout principles of fire stations in China were based on the principle that fire teams could reach the edge of their jurisdiction within 5 minutes in urban areas. This method did not consider the risk differences between regions and belonged to single-objective decision-making. This paper proposes a multi-objective fire rescue station location algorithm based on risk classification. The model takes the maximum weighted coverage rate as the objective function and the minimum coverage rate of the four risk levels as the constraint conditions. During the solving process, the final solution is selected as the one with the fewest number of fire stations among all feasible solutions that satisfy the constraint conditions. This model can ensure maximum coverage while achieving the minimum number of fire stations, and the constraint conditions can determine the minimum coverage rate of the solution, realizing multi-objective decision-making for fire station layout. This model has been applied in Xuhui District, Shanghai, and has shown good conformity to the actual situation.

Due to the popularity of commercial WiFi devices prompting the use of WiFi devices to achieve indoor human activity recognition(HAR) has become possible, HAR plays an important role in human-computer interaction, virtual reality and meta-universe, human walking, running and other activities will have an impact on the propagation of WiFi signals, by analyzing the information of the affected channel state, WiFi-based HAR can be achieved. In this paper, human activity states of three environments are classified, firstly by Hampel filtering to remove the anomaly noise, after that by DWT to further reduce the noise interference, then by extracting the statistical features in the time-frequency domain, and finally by random forest classification to complete the classification recognition. Through experimental verification, the method proposed in this paper can achieve 92% and 91% recognition rate in laboratory scene and corridor scene, which is better than the traditional algorithm recognition rate. It can also achieve 89% accurate recognition rate in non-line-of-sight(NLOS) environment, which also confirms the robustness of the proposed algorithm to the environment.

In order to solve the problems of increasing energy demand, environmental pollution and low efficiency of energy conversion and utilization, a research on optimal scheduling strategy of regional integrated energy system based on an improved genetic algorithm is proposed. First, the optimal dispatch model of the regional integrated energy system is established, and the objective function is decomposed and iteratively solved by the column constraint generation algorithm. Integer coding is used for pipe diameter coding, and simulated annealing penalty function is used to transform constraints. The global optimization capability of the improved genetic algorithm and the local search capability of simulated annealing are complementary and integrated to realize the optimal scheduling of the regional comprehensive energy system. The experimental results show that the operating cost is 55416 CNY without considering the changing working conditions of the equipment; the operating cost when considering the changing working conditions of the equipment is 49388 CNY, which verifies that the established optimal scheduling model is efficient and feasible.

To address the problem that traditional malicious program detection algorithms require a lot of specialized domain knowledge and convolutional neural networks for detecting malicious programs that are difficult to capture global features and long-range dependencies, we propose a malicious program detection method based on Transformer architecture and innovatively use the program assembly opcode frequency to construct a co-occurrence matrix, which in turn generates images as model inputs. Using the multi-headed attention mechanism of Transformer, we extract the malicious program opcode invocation pattern implied behind the co-occurrence matrix image and thus achieve the detection of malicious programs. The method can adapt to PE files of different sizes with fewer parameters and outperforms ViT-B/16 with an accuracy of 0.9887, precision of 0.9873, and F1 score of 0.9849.

The image processing technology is widely used in digital signal processing, biomedical engineering, and industrial automation control. In image processing, image enhancement and restoration techniques are used in many imaging applications, such as astrophotography, medical imaging camera, film production, and infrared imaging. At the same time, image enhancement and restoration technology based on both deep learning and neural network is a major research hotspot in the area of image processing. This paper reviews the related research of image enhancement and restoration technology. In terms of image enhancement, the working principle of its technology based on spatial domain and frequency domain and contrast enhancement technology, including color image contrast enhancement, are introduced. In image restoration, median filtering technology and the Lucy-Richardson (LR) algorithm are introduced, including the adaptive median filtering algorithm and the improved LR algorithm based on the Gaussian low-pass filter. Finally, the merits and demerits of these technologies are discussed, and the development prospect of image enhancement and restoration technology has prospected.

In response to the current online education platform cannot well recommend personalized content to users, this study is based on user preferences, by building a user preference algorithm model for user survey and recommending courses to users according to the model guidance. A user preference modeling that incorporates long and short-term memory networks and self-attentive algorithm models is designed. Considering the influence of time interval between user behaviors on user preferences, the LSTM algorithm model is improved by using time gates and cosine functions to obtain user long-term preferences from user historical behavior sequences, and then user short-term interests are obtained by modeling user current session behavior sequences by introducing temporal location encoding information between user behaviors in the self-attentive algorithm model. The experimental results show that the user satisfaction is good, which indicates that this user preference algorithm model has some validity.

A small model of wireless charging system for an electric moped is designed and made. Wireless induction charging transceiver and adaptive coupling scheme (maximum frequency tracking) are applied to study how to achieve maximum efficiency. MSP-EXP430G2 single chip microcomputer is used as the main controller to realize the functions of frequency regulation, current control and voltage control. When the trolley reaches the specified position, it detects that the transmitting coil stops moving. When the electric energy is full, the trolley starts automatically when the transmitting coil stops working. The system includes transmitting module, receiving module, voltage regulator module, monitoring and control module, infrared tracking sensor and capacitor module. The system based on the MCU wireless control module circuit, through the software programming to realize the wireless charging and discharging mode automatic switch, the car detects that the transmitter coil stops working, charging is completed, the car starts, fixed point driving, automatic tracking, liquid crystal display and other functions. The design of the charging equipment has the characteristics of flexible use, convenient, reliable work, high power utilization rate, wireless charging location accurate. After testing, transmission efficiency up to 90%, has a wide range of application value.

With the rapid development of automobiles towards electrification and intelligence, 5nm Qualcomm 8295 flagship chip has been gradually applied to traditional and new energy vehicles, which brings higher requirements for high-quality signal transmission. In this study, the signal quality evaluation methods of the insertion loss to crosstalk ratio and the integrated crosstalk noise are explored. In addition, two kinds of differential via optimization schemes with the signal-ground ratio of 1:2 and 2:3 are proposed, respectively. The results show that the performance of the new differential via structure is improved significantly, which provides a reliable solution for high-quality signal transmission.

A moving sound signal separation method based on the sparse representation of Doppler atoms is proposed to address the interference problem of wheel rail contact noise in the process of collecting train bearing signals. Using a single microphone to collect sound signals, first Doppler modulation is performed on dictionary atoms; Then perform sparse matching decomposition on the acoustic signal to obtain the sparse signal; Finally, the signal reconstruction theory is used to separate the sound source and obtain the bearing signal. Through experimental signals, it has been proven that under the influence of Doppler effect, the proposed method can effectively extract the target sound source and has good in band noise cancellation effect.

One of the main reasons for train derailment is the failure of wheel bearings, and the track-side acoustic detection system (TADS) is one of the effective means to achieve online health monitoring. TADS is based on the basic principle that the acoustic signals emitted by the wheel bearings during the running of the train contain a wealth of diagnostic-related information. The challenge of this technology is how to accurately extract information reflecting the health condition of the wheel bearing from the sound signal. One of the main influence factors is the Doppler effect. To solve this problem, A new "Doppler Single Transient Model (DSTM)" and the "one-by-one stripping" strategy is developed to handle with the periodic fluctuation signal with Doppler distortion in this paper. It consists of three steps: (1) Use the Doppler Periodic Transient Model (DPTM) roughly estimate the parameters of the transient model; (2) The DSTM is used to extract transients from the signal one by one. (3) The optimization algorithms is used to accurately identify parameters. Through the proposed DSTM and the transient extraction strategy, the transients can be matched more accurately.

The fractional delay filters are widely used in beamforming of broadband signals; however, conventional digital fractional delay filters are associated with high computational complexity. The Farrow structure fractional delay filter features a multi-stage sub-filter parallel structure, where the fractional delay is not affected by the sub-filter coefficients, thereby minimizing computational complexity. Nevertheless, the conventional Farrow structure is plagued by considerable passband fluctuations and high-frequency errors. To tackle these concerns, this article proposes a Farrow structure fractional delay filter design approach based on the maximum flatness method. This technique employs the maximum flatness method to design Farrow structure sub-filters, thereby reducing high-frequency errors and passband fluctuations compared to the conventional Farrow structure.

In order to reduce the fuel cost during the whole flight when the starting point coordinates and the terminal point coordinates are given, the vertical guidance mode is first designed for climb phase, cruise phase and descent phase, and based on this, the flight performance database of each phase is established for a certain type of aircraft. Then, combining the advantages of genetic algorithm and pattern search algorithm, a vertical profile optimization algorithm based on GA-pattern search algorithm is proposed. After that, based on the flight performance database, this algorithm is used to find the least fuel vertical profile. Through Matlab Simulation, it is verified that the flight path optimization effect is good, the aircraft not only reached the designated point along the optimized flight path, but also achieved the least fuel target.

This discourse delineates the formulation and execution of a cutting-edge apparatus explicitly designed for the assessment of modulation depth. This avant-garde instrument excels in enumerating and illustrating the diverse attributes of the modulating signals emanating from a source, discerning their unique modulation techniques^[1], and engendering demodulated signals. Rigorous empirical evaluation substantiates the device's preeminent capacity to accurately enumerate and discern amplitude modulation (AM) and frequency modulation (FM) for single-tone voltages, in addition to high-frequency voltages. Consequently, this mechanism exhibits extraordinary performance and considerable versatility.

In order to achieve the objectives of rationalizing the layout of emergency material centre sites and distributing emergency materials intelligently, this paper establishes an emergency material centre site selection model, and then designs and introduces an improved particle swarm algorithm to optimally solve the model. The algorithm improves the inertia weight of the algorithm to make it change adaptively according to the number of iterations of the algorithm. And in the learning factor part, an improved second-order oscillation method is proposed to further improve the global search capability of the algorithm. Finally, simulation experiments comparing the standard algorithm optimisation results show the rationality and validity of the siting model as well as the feasibility and efficiency of the designed algorithm.

Brain segmentation, a crucial research area in medical image processing, plays a vital role in assisting doctors with diagnosing and treating various neurological disorders. The brain consists of multiple tissues and structures, resulting in differential representation across different imaging modalities. This makes it challenging to achieve effective segmentation of image information within a single modality. Therefore, we propose the multimodal brain segmentation method Selective Modality Segmentation Network (SEL-Net), which combines information from multiple image modalities and adopts a method based on a spatial attention mechanism that can integrate multi-modal image information, adaptively adjusts the weights of each modality using a spatial attention mechanism, and optimizes it through different loss function ratios. Different loss function ratios can balance the influence of different modalities, thereby improving the performance and robustness of the model. In the BraTs2019 dataset, SEL-Net proves its effectiveness and superiority, and SEL-Net achieves better results in the metrics of Dice coefficient, sensitivity and specificity. This indicates that SEL-Net has great potential in multimodal brain segmentation, which can provide more accurate and reliable results for medical diagnosis and treatment.

Image matching is an important step in the process of power inspection image stitching and 3D reconstruction, which directly affects the effect of power line image stitching and 3D reconstruction, and is widely used in power line inspection. Traditional power inspection image matching applications are mostly implemented on X86 architecture, but ARM architecture has gradually entered the server field in recent years, showing high efficiency and energy-saving advantages, so it is necessary to transplant power inspection image matching applications to the server platform of ARM architecture. Since the application of image matching is very computation-based, in the process of porting the algorithm to the ARM platform, this paper adopts the Gaussian pyramid algorithm to accelerate the application of image matching through the search method from coarse to fine.

Most of the existing ensemble feature selection algorithms directly assign different feature selection algorithms the same weights or simply treat accuracy as weight and then vote for each feature. However, these research methods either ignored or did not fully consider the fitness of the feature selecttion algorithm to the data set. To deal with this challenge, we propose a new ensemble feature selection approach with adaptive weights. In detail, we use the softmax function to dynamically adjust the weight of the base feature selector according to its fitness to dataset in the k-round training process. We name it ensemble feature selection based on softmax function (EFS-BSF) algorithm. We demonstrate the superiority of the EFS-BSF approach over previous methods through mathematical analysis and experiments on multiple data sets.

Specific emitter identification (SEI) identifies targets mainly by unintentional modulation of the signal. However, due to the high energy of the primary signal, once the primary signal changes, the recognition becomes less effective or even impossible using a feature database that is not updated. In this paper, we propose to use a mutual information improved variable mode decomposition (VMD) algorithm to suppress the primary signal phase of the transmitter. Furthermore, we simulate the feature extraction of the unintentional phase modulation of the transmitter signal and use support vector machine (SVM) for individual identification. The simulation results show that the algorithm improves the recognition rate by about 6% (0 dB) compared to the retained primary signal. The results demonstrate that our proposed phase suppression technique improves the adaptability and accuracy of individual identification of transmitters.

The proposed video stabilization algorithm for head-mounted mobile platforms addresses the issue of video shake. In the feature detection and extraction stage, the Super point feature extraction algorithm is used to extract feature points. This algorithm provides a more comprehensive set of feature points compared to traditional feature detection methods, enabling accurate motion estimation.In the motion smoothing stage, the camera path of a specific camera is preprocessed using Kalman filtering to remove significant shaking factors. Then, the processed path is used as a baseline to simultaneously smooth all camera paths through spatio-temporal optimization, aiming to make the camera paths as similar as possible and reduce distortion. Finally, the optimal path is selected from the refined paths to address the deviation among multiple cameras.Experimental results have demonstrated that this method achieves comparable stability to other methods while significantly reducing the required processing time.

In order to solve the problem of on-line detection of the secondary use of the box packing machine in the sealing section, the smart code scanner is used to scan the barcode or two-dimensional code of the box with its own brand logo, based on "unlimited frame decoding" and "hot The "map positioning algorithm" technology is triggered by the origin phase signal of the YP11 cartoning machine, and an audible and visual alarm is added to remind the operator to check whether the carton is used correctly. By modifying the man-machine interface, when switching brands, you can directly use the luggage change button designed by the man-machine interface to quickly change the brand settings. The test was conducted on the company's 2# cartoning machine as the research object. The results showed that: after the improvement, the detection rate of the mixed use of secondary cases was 100%, realizing real-time monitoring of the case cases of multi-specification secondary utilization, which can be applied to All kinds of materials, barcodes, and two-dimensional code position of the leather products, realize maintenance-free and adjustment-free. This technology provides a total support for the quality control used in cases.

Accurate classification of network traffic is the research basis for tasks such as network management, service quality optimization, abnormal traffic detection, etc. Aiming at the problems of incomplete feature extraction of existing methods, a VPN encrypted traffic identification model based on improved residual network is proposed. The model takes the conversation composed of bidirectional flows as the analysis object, first uses the multi-layer convolutional network to extract the spatial structure features of the conversation, then introduces the improved residual unit to increase the width of the backbone network, and uses 1x1 and 3x3 different sizes of convolution kernels on the same level to obtain different scale feature information, and improve the model’s feature representation ability. Finally, the extracted features are expanded into one-dimensional vectors and Softmax funciton is used to classify the encrypted traffic. The Experimental results on public VPN-non VPN dataset show that the proposed model performs better than traditional identification methods, and compared with the well-known methods, the accuracy, precision and recall metrics are improved by 3.32%, 2.65% and 3.79%, respectively.

Image annotation of medical scenes is expensive, then weak supervision methods have emerged. Therefore, this paper proposes a pixel affinity adaptive expansion model for medical image label generation in skin scenes. The method is divided into two stages. In the first stage, initial pseudo label is generated based on the characteristics of the skin image using the proposed self-adaptation expansion module . In the second stage, the generated initial pseudo label uses the characteristics of the pixel pair to distinguish the foreground background and class boundaries to generate the final label. Finally, the generated tags are used for fully supervised segmentation. Experiments were conducted on the ISIC 2016 and ISIC 2017 datasets, respectively. The MIoU of the proposed method was 63.24% and 50.43% on the two datasets, while the Dice-score was 76.13% and 67.79%, the accuracy rate was 81.28% and 64.79%, and the recall rate was 73.54% and 71.03%, which is higher than most existing weakly supervised semantic segmentation algorithms, and achieved Sota (State of the arts) on the two skin datasets.

The expert set theory plays a significant role in decision-making. Since it contains the viewpoints of various specialists, it can effectively transmit the views of all of them. In this study, the idea of the expert matrices of GFSES is presented first, then follows the definition of operation approaches and characteristics as the complement, addition, and subtraction operations of these expert matrices. The study also demonstrates how it may be used to solve decision-making issues.

Saliency detection of light field images is a key technology in applications such as visual tracking, target detection and image compression. Existing light field saliency target detection tends to ignore the complementarity of cross-mode light field data, inevitably introducing redundant information and leading to blurred salient images. Even in similar or confusing scenes, there are problems such as incomplete detection objects and difficult background suppression. To this end, this paper proposes an image saliency detection network based on self-selective cross-modal feature fusion. Firstly, hierarchical features are extracted from the backbone network, and each modal feature is optimized based on the attention mechanism using the spatial alignment and channel rescaling modules, and then the two modal features are fused to obtain a more accurate saliency map guided by edge information. Experimental results on the latest light field dataset show that this method outperforms the comparison method both quantitatively and qualitatively.

In recent years, speech synthesis technology has become increasingly advanced, leading to a proliferation of forged audio content on the internet, which poses significant threat to individuals and society. Many studies have utilized a range of deep learning-based techniques to differentiate fake audio content, but the features used in these studies are often limited in their rich and generalizable characteristics. In this paper, we propose a novel fake voice detection technology that utilizes the wav2vec2 model for feature extraction along with a custom-designed residual-based detection module to augment the detection of fake audio content with greater accuracy and precision. Additionally, we incorporate a data augmentation method to improve the performance of the model and enhance its ability to generalize. We trained our model on the ASVspoof2019 dataset and evaluated it on the LA and DF datasets of the ASVspoof2021 dataset. Supplementary experiments demonstrated that our approach achieved state-of-the-art detection performance and illustrated its effectiveness and applicability.

This paper presents an radar blockage recognition approach along with blindness prevention measures utilizing intelligent computing methods to address the challenges in radar chip-based autonomous driving systems. A process and algorithm are initially created for detection, enabling real-time and precise identification. The adaptive blockage performance controller and the automatic diagnosis and recovery strategy are then proposed as part of the adaptive blockage prevention strategy, which enables the system to quickly return to its normal operating state by intelligently adjusting parameters when the sensor blockage occurs. The conventional linear threshold methods are replaced with nonlinear machine learning to deal with complex performance variations, with the aim of lowering false positives and negatives through real-time optimization. A comparative experiment is conducted to show the effectiveness.

In view of the requirement that coal mining enterprises should establish the work system to evaluate the gas extraction meets standards, this paper adopts the deep learning technology——long short term memory (LSTM). The proposed method, firstly processes the outlier and missing value of the coal mine monitoring data, selects the key indicators that affect the extraction volume. Then uses LSTM network model to train and predict. Finally, based on the predicted gas extraction volume, determine whether it meets the standard and estimate the number of days required to consistently meet the standard. LSTM, RNN and GM are applied to the estimation of gas extraction volume and standard on on-site process data. The experimental results show that the constructed LSTM model has better performance estimation than RNN, GM, and can perform more accurate gas extraction compliance estimation, ensuring the safety and efficient production of coal mines.

How to extract and analyze feature information from significantly different image data, and achieve precise positioning and classification of massive images, is the core issue of image classification research. The dataset in this article contained 10 different types of images, each with approximately 5000 images, with image sizes of 100 × 100, 200 × 200, and 300 × 300. In order to deal with large-scale data, this paper used adaptive biogeographical optimizer and traditional gradient descent methods to optimize multi-layer depth Convolutional neural network (CNN), and designed single-layer adaptive biogeographical network, multi-layer adaptive biogeographical network, single-layer gradient descent network and multilayer gradient descent network for experiments. All classifiers used the softmax function for classification. The accuracy of multi-layer network image classification under the algorithm in this article was 96.6%. The accuracy and efficiency of image classification can be improved by using Biogeography optimization algorithm to optimize the depth CNN.

A reservoir scheduling optimization algorithm is formed by improving the particle swarm algorithm to solve the problem of slow speed of intelligent algorithm in reservoir optimization, which makes the particle swarm algorithm obtain the optimal solution of the scheduling process faster and more accurately in reservoir optimization and scheduling. The improved particle swarm algorithm uses the fast convergence feature to obtain the preliminary solution of the scheduling process. According to the month-by-month water supply plan of the South-North Water Transfer Central Line and the layout map of important water conservancy projects in Jiangsu, each water node is divided and the corresponding node diagram is drawn. Combining the engineering and construction of each controlling reservoir in Jiangsu Province, an optimized reservoir scheduling model for the current and planned years of the South-North Water Transfer Central Line is established. The results show that the improved particle swarm algorithm has some application space in reservoir optimization and scheduling

This patent proposes a dynamic aggregation method of distributed resources based on cloud edge collaboration, and proposes load peak shaving performance indicators and load frequency modulation performance indicators to represent the difference in load response performance. On this basis, it proposes to use the improved spectral clustering algorithm of genetic algorithm to achieve dynamic aggregation of distributed energy and achieve intelligent aggregation of distributed resources in multiple stations.

In this paper, firstly, the luff cylinder speed of QL2000 boom bucket wheel reclaimer is calculated in detail by the method of drawing and analysis. Secondly, in order to improve The convenience of processing, the upper metal structure is improved and checked. Finally, in order to improve the stability of the equipment operation, the idler bracket was improved. Practice has proved that these improvements reduce the cost of design and manufacture, improve the reliability of equipment operation, and provide reliable reference experience for the design and manufacture of similar equipment.

Detecting 3D objects from LiDAR points is significant for the environmental perception of robotic systems. Some pillarbased 3D object detectors solely use 2D convolutions as feature encoders, which occupy fewer computation resources but sacrifice model accuracy. To activate the potential performance of pillar-based feature representation manners, we propose HCT-Det, a novel hybrid CNN-Transformer architecture for 3D object detection from point clouds. Motivated by the structure re-parameterization technique and vision transformer (ViT) framework, we redesign the 2D backbone and further introduce the Rep-VGG block and multi-head self-attention (MHSA) mechanism to enrich the scale diversity of feature representation. We perform ablation experiments on the KITTI vision benchmarks to highlight the superiority of our HCT-Det. The evaluation results show that our model outperforms PointPillars baseline, yielding an accuracy of 79.08 moderate AP_3D on the car category at a speed of 57.46 FPS on the NVIDIA Tesla P40 platform. Without bells and whistles, our HCT-Det can achieve a reasonable trade-off between accuracy and speed.

Based on UWB signal measurements and depth learning models, research on single base station moving target precise positioning algorithms has been carried out to address the need for precise positioning of moving targets in complex environments such as non visual range, electromagnetic interference, dust, humidity, and water in coal mines. The calculation model of signal transmission time based on bilateral two-way interaction model is analyzed, and the calculation method of signal distance term parameter vector and signal angle term parameter vector is constructed. The information vector of UWB signal is formed by combining distance term parameter vector and angle term parameter vector; A tag position coordinate calculation model based on depth learning model is constructed, and combined with geometric relationship calculation, the two-dimensional coordinate calculation of mine moving targets is completed. Experimental analysis shows that the proposed positioning algorithm has a minimum root mean square error of 9.22 cm and a maximum root mean square error of 18.87 cm in positioning coordinates, which has a good positioning effect and provides a new idea for accurate positioning of mine moving targets.

The acoustic channel in ice-covered water largely limits the transmission capacity and performance of the communication signals. Spread spectrum systems have stable performance and strong resistance to multipath interference, which can effectively cope with the interference caused by the ocean environment on signal propagation. Nevertheless, traditional direct-sequence spread spectrum systems have a low communication rate. This paper focuses on the research of an M-ary spread spectrum communication system in ice-covered water based on energy monitoring technology. It explores a robust underwater acoustic communication system that can exponentially improve information transmission capacity and bandwidth utilization. The communication encoding mode is validated in an ice-covered underwater acoustic outfield experiment, achieving zero error code transmission. The experiment verifies the feasibility of the communication system and algorithm.

Unmanned aerial vehicle (UAV) swarm has sparked heated debate around the world in recent years. There are various stages to the battle of UAV swarm. This paper focuses on the attack task assignment after being cast or launched over the battlefield. In order to describe the process of task assignment for UAV swarm, a model introducing multiple constraints, such as limited detection range and imperfect communications, is established first. Then, inspired by contract network algorithm, we proposed a mechanism including competition for local revenue and cooperation for global revenue. And the premise of this mechanism is that UAV swarm is distributed, whose task is assigned by itself, not any other UAV. Finally, the advantages of our mechanism are illustrated by the contrast of different methods on the same issue.

At present, large-scale object classification and recognition tasks inevitably encounter problems of training efficiency and model accuracy. The solution of this problem highly depends on the definition of loss function. A better loss function can get a more accurate model with fewer epoches, which undoubtedly helps to save computility and time cost and improve the security of the model. Nowadays, the task of indoor object recognition plays an important role in the fields of computer vision aided blind people, home robots and so on. However, the task of indoor object recognition is plagued by the above problems due to its huge category of recognition objects. Starting with the task of indoor object recognition, this paper innovatively uses EIOU loss function and YOLOv5 deep learning convolutional neural network in this field, which improves training efficiency and recognition accuracy. Some original bounding box regression loss functions of YOLO series (such as GIOU, DIOU, CIOU) have defects in that the prediction box coincides with the center point of the truth value box, and the horizontal vertical ratio is the same. The paper uses the EIOU loss function to solve this problem. This paper focuses on the network structure of YOLOv5, the defects of YOLOv5's native loss function, the calculation method and advantages of EIOU loss function, and compares the performance gap between EIOU and CIOU. Finally, the EIOU loss function is used to complete the indoor article identification task.

The water consumption of urban users is complex, and there are great differences between residential water and industrial wastewater. Real-time monitoring of water quality is an important task of urban water management. Among the water quality parameters, dissolved oxygen concentration (DO) is an important index to evaluate the quality of water. Therefore, the accuracy of dissolved oxygen prediction results is very important for the management of water bodies in sewage treatment. In this paper, a model based on DBSCAN-CNN-LSTM is proposed to solve the problems of inaccurate prediction accuracy and poor data convergence in the existing water quality prediction model in the sewage treatment process. Firstly, the obtained water quality data are preprocessed to obtain water quality factors with high correlation with DO.Secondly, the input data are predicted and output by establishing a hybrid model based on DBSCAN-CNN-GRU. The DBSCAN algorithm can cluster the sewage data and reduce the influence of local minima on the prediction results, so as to improve the prediction accuracy and data convergence. Compared with the traditional neural network models GRU, LSTM and CNN-LSTM, the predicted values is closer to the real values.

The environment sensing module is a prerequisite for autonomous driving. The selection of on-board sensors and a good 3D object detection algorithm determine the performance of the environment sensing module. The LIDAR point cloud 3D object semi-automatic annotation tool uses the detection algorithm in the open source framework MMDtection3D to complete the pre-labeling, and then enters the joint point cloud image labeling tool 3D-BAT to complete the manual correction after the data processing module, and the output labeling results can be supplied to the deep learning based 3D object detection algorithm training after the format conversion. The self-built data acquisition vehicle equipped with 6 cameras and 1 LIDAR collected more than 4000 frames of point cloud data, and the semi-automatic annotation tool completed the annotation of more than 2000 frames of point cloud data.

By comparing the commonly used 95 type zirconia ceramics and 99 type alumina ceramics in engineering, acoustic emission technology was used to monitor the stress failure process of the two types of ceramics. Corresponding acoustic emission signals were collected and analyzed. The differences in acoustic emission signal counting rate, signal time-frequency domain, and acoustic emission energy were obtained for the damage evolution process of the two types of ceramics. After analysis, it was found that the acoustic emission activity intensity of 95 type zirconia ceramics was much higher than that of 99 type alumina ceramics in the experiment. At the same time, it was found that the acoustic emission signal frequency of these two types of ceramics would decrease when approaching the fracture state. This can effectively determine the state of these two types of ceramics.

Aiming at the problems of large amount of heterogeneous Industrial data, high fault concealment, complex feature engineering of traditional methods, an anomaly detection method combined with Bi-directional long-short term memory, variational autoencoder and whale optimization algorithm based on cloud-edge collaboration. By integrating the output of each detection models with different dimensions through the residual weight matrix, to obtain the comprehensive residual value, and compare with the residual threshold for anomaly detection. Through the experiments on SKAB and TEP datasets, the results verify the effectiveness and general adaptability of the proposed method, and the anomaly detection accuracy is higher than a single detection model and existing anomaly algorithms such as CNN and LSTM-AE.

Identification of basal ganglia (BG) network plays an important role in the construction of closed-loop deep brain stimulation (DBS) system. This work proposes an identification algorithm based on pseudo-random binary sequence (PRBS) and recurrent radial basis function neural network (RRBFNN) for basal ganglia network modeling. The PRBS modulated DBS signal is taken as the input signal of BG network. Then, the relationship of input DBS signal and beta power of BG network is modeled by a RRBFNN. The numerical experiments show that the PRBS modulated DBS signal meet the requirements of BG network identification. What’s more, the results indicate that the proposed RRBFNN can effectively identify the dynamics of BG network.

Aiming at a point-to-point trajectory planning problem for industrial robots, we propose a method for time optimal optimization of trajectories in the robot joint space by employing the honey badger algorithm. Interpolation points at each joint are first acquired using robotic kinematics; Then, to enable the robot to smoothly complete the work task, a 4-5-4 piecewise polynomial is employed to interpolate its trajectory; Finally, we use the honey badger algorithm to optimize the interpolation time of the segmented polynomials to achieve the goal of optimal trajectory planning for the robot's time. The simulation through MATLAB software to obtain the change information of angular displacement, angular velocity and angular acceleration of each joint of the robot, under the restriction conditions that meet the kinematic parameters of the robot, the robot's motion time is shortened by 43.17% compared to that before optimization, which verifies the effectiveness of the honey badger algorithm in the optimization of robotic tracks.

Due to the high delay and large inertia characteristics, there are issues with inefficient regulation and energy waste in district heating systems. Improving the predictive ability of regional heating load can help improve the accuracy of thermal network control. In order to improve the accuracy of heating load prediction, this paper proposes a new heating load prediction algorithm, STL-IEA-LSTM, which combines LSTM model with information entropy attention mechanism to separately model the data modeling of heating load after STL algorithm decomposition, so as to enhance the prediction ability of heating load. The experimental evaluation conducted on the operating data of a heating station in Shijiazhuang during the heating season demonstrated that the STL-IEA-LSTM combination model constructed in this study achieved a MAPE of 0.7031%, indicating its superior performance in predicting heating load. Comparative analysis with other models, including STL-LSTM, entropy-based LSTM, classic LSTM, XGBoost, and MLP, further confirmed the effectiveness of the proposed method in achieving high prediction accuracy.

It is challenging to estimate ocean surface currents from along-track interferometric (ATI) data due to the complexity of the ocean environment. In order to complete the process from modeling the ocean surface signals to extracting the ocean surface currents, this research integrates dynamic sea surface modeling imagery with an iterative technique to inverse the currents. To complete the simulation analysis of time-varying ocean surface imaging, ocean surface backscatter coefficients and ocean surface echoes are constructed using the wave spectrum model as the basis. The ATI data from the simulation modeling are inverted using the iterative inversion algorithm of the Synthetic Aperture Radar (SAR) ocean surface currents field, and the accuracy of the inverse ocean surface currents field obtained reaches 0.1 m/s, proving the viability of the whole system.

In an intelligent completion system, the surface control part not only needs to provide power to the downhole equipment, but also needs to communicate with the downhole equipment. In this paper, a power line carrier communication system based on intelligent well completion is designed. In this system, the surface controller uses a single-core cable to transmit surface commands to the downhole by coupling them to the single-core cable, and the downhole equipment then couples ground commands from the single-core cable, which realizes communication with the downhole equipment and provides power to the downhole equipment to meet the power supply and communication requirements of the complex downhole situation.

Convolutional neural networks (CNNs) have been gradually applied to aerospace equipment in recent years. However, CNN requires a significant amount of hardware resources, while parameters need to be updated for different tasks in space. We propose an FPGA architecture for CNN training through a multi-module collaborative design strategy. On the basis of the intra-layer parallelism, we analyze the data storage and the operations to deduce forward and backward propagation on embedded devices. Based on the experimental results, the working frequency can reach 250 MHz and the performance is 15.9 GOPS. The resource usage and efficiency of this design are superior to those of other similar hardware acceleration platforms.

With the development of deep learning in diagnosing autism spectrum disorder (ASD), multi-site resting-state functional magnetic resonance imaging (rs-fMRI) images have made excellent progress. However, there is a heterogeneous problem between the multi-site data caused by inconsistent data distribution. Existing domain adaptation methods cannot deal with the issue that the target domain data is unavailable during the training stage. To address this problem, we propose a deep domain generalization method via low-rank category constraint (DDGLCC) for multi-site ASD identification. The main idea is capturing the category discriminative information through the domain-specific networks and gaining the consistently shared information through the domain-invariant network. A novel category-based low-rank constraint strategy is used to align two types of networks. In the test stage, the well-trained domain-invariant network is applied to the unseen target domain data. Whether the results on different deep structure experiments or different lowrank constraints experiments, the proposed DDGLCC method achieves the best performance.

The synergy with energy storage devices in the current microgrid hierarchical zoning optimization planning is low, resulting in poor quality optimization planning results. To this end, a microgrid hierarchical zoning optimisation planning algorithm with collaborative scheduling of multiple energy storage devices is proposed. Analysis of weak levels and regions. Distributed arrangement of energy storage devices in the microgrid. Dynamic control of energy storage device parameters and hierarchical partitioning of multiple energy storage devices for collaborative scheduling. Generation of a microgrid hierarchical partitioned optimal planning algorithm. The experiments show that the average index entropy value of the algorithm is 0.4857, which can effectively improve the quality of the microgrid hierarchical zoning optimization planning results and has high practical value.

Single-phase inverter output AC220V/50Hz voltage for the EMU, power supply for the passenger room socket. Traditional single-phase inverter usually adopts voltage RMS control. When passengers use non-controlled rectified nonlinear load such as mobile phone and laptop charger, the output voltage waveform distortion of single-phase inverter is serious, and the low-frequency voltage harmonic content is high. In this paper, an optimization scheme of the control algorithm is proposed. The voltage and current double closed-loop + proportional integral quasi-resonant control method is adopted to improve the dynamic performance of the system and reduce the harmonic content of the low-frequency voltage. By setting up the hardware environment and comparing the voltage and current waveforms before and after the optimization of the control algorithm, the effectiveness of the algorithm is verified.

With the increasing number of vehicles on the road, traffic congestion and safety issues have become more and more prominent. Super-resolution image recognition technology is a effective solution that can be used for image analysis and recognition in large-scale vehicle-to-vehicle communication. This paper introduces the basic principles and application scenarios of super-resolution image recognition technology and elaborates on its application in large-scale vehicle-to-vehicle communication. Experimental results show that this technology can effectively improve the accuracy and efficiency of image recognition and has broad application prospects.

The current conventional security authentication methods for power monitoring system operation and maintenance scenarios mainly verify the user's identity by means of certificate authentication or password authentication, which leads to poor recognition accuracy due to the lack of effective extraction of key user information. For that, the security authentication method of power monitoring system operation and maintenance scene based on face recognition and fingerprint recognition is proposed. The biometric information of staffs in power system operation and maintenance scenes is obtained through face detection and fingerprint morphological feature extraction. The security authentication protocol is also designed, and the data interaction behavior is encrypted in the form of session key, and the security authentication model is constructed on this basis. In the experiment, the designed security authentication method is tested for recognition effect. The final results can prove that when the proposed method is used to identify user personnel information in the power monitoring system, the identification accuracy is high and has a more desirable authentication effect.

This paper proposes a hierarchical path-planning algorithm based on a combination of graphical search algorithms and optimization methods. In the global path planning layer, by using the Hybrid A* algorithm, we can quickly obtain an optimal path that can avoid all static obstacles on the map. In the local path planning layer, the global path is optimized by numerical nonlinear numerical optimization to generate a feasible path that satisfies the safety constraints. By processing global path planning and local path planning separately, the computational complexity of path planning can be effectively reduced, and the efficiency and accuracy of the path planning can be improved. Secondly, the hybrid planning algorithm can generate high-quality paths with both safety and flexibility. We simulate and verify the algorithm, and the results show that the method has practical applications in autonomous driving.