This PDF file contains the front matter associated with SPIE Proceedings Volume 12715, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

The fusion of a low spatial resolution hyperspectral image (LR-HSI) and a high spatial resolution multispectral image (HR-MSI) in the same scene is a common method to get a high spatial resolution hyperspectral image (HR-HSI). For the drawback that the standard nuclear norm regularization treats each singular value equally, this paper proposes a weighted nuclear norm model based on sparse matrix factorization (called WNNS) for hyperspectral image fusion. Specifically, we promote the sparsity of fused images by adding the ℓ₁ norm of coefficients. Furthermore, to preserve important data components, we combine with the weighted nuclear norm regularization, where different weights are given to singular values. To efficiently solve the proposed model, we apply an alternating direction method of multipliers (ADMM). Experiments show that the proposed method has better performances in terms of numerical results and visual effects.

There are many electrical equipment inside ultra-high voltage converter valve hall, including converter transformer, converter transformer bushing, wall bushing and other typical electrical equipment. Among them, there are many insulation media inside the converter valve hall, including gas insulation, liquid insulation and solid insulation media. For layered medium, there are two problems in the ultra-high voltage converter valve hall: 1) The layered medium in the valve hall includes the interface between different insulating mediums such as the epoxy impregnated paper and oil impregnated paper. The layered interface belongs to the macro category, which can be fully described by applying the classical electromagnetic theory. Its transient process can be characterized by the classical lumped parameter differential equation model, and its time constant scale is around 5000s. Its spatial scale is near cm scale. 2) The layered interface of semiconductor devices inside the valve hall converter belongs to the micro category, with spatial scale in the nm scale and the time scale distributed near the ns dimension. Internal charge and potential distribution can be described by partial differential equations, and the parameters of the differential equations are variables with clear physical quantities. In this paper, firstly, the substrate free of impurities of multi gate MOS capacitors is analyzed from the microscopic carrier level, and the minority carrier term and acceptor impurity concentration N_a at any point x near the surface are given. Based on the full mathematical derivation and analysis of the micro particle's sink and sink analytic expressions, the space charge distribution inside the layered insulating medium is obtained, especially at the interface of layered medium. Furthermore, for the electrical equipment inside the converter valve hall, which is subject to the combined action of high voltage and large current, the closed analytical solution of the electro-thermal coupling is derived from the perspective of analytical model through Bessel function, and the quantitative characterization relationship between the leakage current and voltage is obtained. This paper derives full theoretical analytical relationship between the space potential and charge distribution of the layered medium in the UHV converter valve hall from theoretical point of view, which has the good theoretical guidance and the practical application value for operation and maintenance of electrical equipment in the UHV converter valve hall.

Helicopters often encounter obstacles such as towers and high-voltage lines when flying at low altitude, and the safety problem is increasingly prominent. Optical images have high resolution, which provide rich color, texture, edge and other details of the detection object. Infrared images can still maintain the advantage of high detection rate at night or in the environment with poor visibility. Combining the characteristics and advantages of infrared and optical images, this paper designs a dual branch convolution neural network to detect helicopter flying obstacles. For infrared images, a single branch infrared image feature extraction network SBI-Net (Single Branch Infrared image Network) is designed to automatically extract the features of infrared images; For optical images, a single branch optical image feature extraction network SBO-Net (Single Branch Optical image Network) is designed to extract the features of optical images; Finally, the two networks are fused, and a dual branch feature fusion network IODBFF-Net (Dual Branch Feature Fusion Network model based on Infrared and Optical image) is proposed. The experimental results show that compared with infrared single branch network and optical single branch network, the detection accuracy of dual branch convolution neural network is improved by 2.06% and 40.25% respectively.

The Direct Position Determination (DPD) method can obtain the motion state information of the target directly from the signal received by the base station without measuring the positioning parameters in the first place, so it can provide more accurate positioning estimation. In order to solve the problem that the interferometer system’s positioning error increases or even fails to locate when the signal-to-noise ratio is low, a DPD method based on the interferometer system is proposed. Based on the information of time difference and phase difference, a cost function was constructed based on the principle of maximum likelihood estimation. Simulation analysis shows that the proposed positioning method is also effective under the condition of low signal-to-noise ratio.

The main components of light fuels such as gasoline, diesel oil and aviation fuel are hydrocarbons, which are almost non-conductive substances with extremely low conductivity. In the process of processing, transportation, storage and filling of light fuel oil, it is easy to accumulate more electrostatic charges, which will lead to major safety accidents such as fire and explosion when electrostatic sparks are generated. Therefore, in order to prevent the occurrence of serious accidents caused by electrostatic accumulation, the conductivity of light fuel oil should be tested regularly during transportation and filling. When the conductivity value is found to be low, an appropriate amount of antistatic agent should be added in time to ensure the safety range of conductivity specified in the Safety Static Conductivity of Oil Products. At the same time, the electrical conductivity of oil products also reflects the number of impurities in oil products. In particular, aviation fuel has stricter requirements on electrical conductivity in order to ensure the sufficiency and safety of its combustion, so it is of great significance to detect the electrical conductivity of light fuel. Aiming at the detection requirements of light fuel conductivity, based on the principle of conductivity detection, a fuel conductivity monitoring system based on STM32 and 4G communication is designed. With STM32 as the master, the oil conductivity value is obtained by detecting the weak current at both ends of the electrode by using the precision electrometer measurement method, and the data is transmitted to the terminal for monitoring through 4G. After testing, the system has high detection sensitivity, low cost and can realize remote monitoring, which provides safety guarantee for oil transportation and provides reference for the design of oil quality monitoring system.

A traffic sign detection algorithm based on improved YOLOv3 in natural environment is constructed to address the problems of the existing traffic sign detection model with a large number of small-sized traffic sign, low accuracy rate and slow recognition speed for traffic sign with small object. Firstly, model compression is performed on the network structure, using the network pruning method to reduce the model size and improve the operation speed; Secondly, one scale detection structure is added on the basis of YOLOv3 network structure, and four-scale detection structure are used to fully enable the network to obtain more feature information of small traffic sign and effectively improve the detection rate of small object traffic sign; Finally, the K-means++algorithm is used to re-cluster and calculate the size of the bounding box of the object in the TT100K data set to generate a smaller size prior bounding box, which improves the recognition accuracy of the model for small target traffic sign. The experimental results on TT100K traffic sign data set show that the proposed method has better accuracy and recall than the original YOLOv3 model.

As Moore’s law indicates, the number of transistors on a chip doubles every 18 months, which guarantees many resourcedemanding applications can be implemented on these advanced chips. In order to fulfill this purpose, CAD tools should be precise and efficient. In this paper, we dig into FPGAs, which unavoidably require CAD tools to be configured. A new timing database construction method mainly focusing on reformatting the timing models of programmable interconnections and routing wires is proposed to improve routing efficiency for FPGAs. A contrast experiment has been carried out to compare routing efficiency with original and new database. The results of our experiment show that routing with this new database can implement circuits of high quality (1.000× critical path delay) within less time (0.994× original routing time). And it can at most route resource-demanding circuits within 0.787× original routing time.

The frequent occurrences of surge current[1] of memory architecture in FeRAM, and the lack of necessary formula derivation in related researches of FeRAM are main aspects of what this work focuses. Therefore, the surge protection circuit is designed and simulated by this work with solid memory architecture of FeRAM to decrease the surge current and average power consumption creatively. The surge protection circuit is composed of a MOSFET and an RC delay circuit, which forces the generation of the surge current to be slowed down by mandatory precharge of the capacitor. On the other hand, necessary circuit analysis and formula derivation are concluded to predict the DC operating point of sensitive amplifier and build a clearly defined limits of operating voltage and the ratio of the maximum ferroelectric capacitance and bitline capacitance in the read/write operation of FeRAM.

With the rapid development of Internet of Vehicles, computing-intensive and delay-sensitive applications are widely used. Faced with the shortcomings of less computing resources and limited power supply of vehicle mobile terminals, mobile edge technology came into being. Firstly, a multi-terminal single-edge vehicle network model is established to alleviate the terminal pressure by accessing Multi-Access Edge Computing ( MEC ) servers. Aiming at the heterogeneous computing network, this paper constructs a joint optimization problem of task offloading and resource allocation under the constraints of delay and energy consumption. Aiming at the problem that the resources in the offloading decision process exceed the load, the modified chromosome and the improved elite selection strategy are used to optimize the genetic algorithm. The dynamic parameter adjustment strategy is used to optimize the particle swarm optimization algorithm to prevent premature convergence. A two-stage joint optimization algorithm is proposed to solve the problem. It can be seen from the results that the optimized algorithm can find the appropriate optimal solution and effectively reduce the cost compared with GAVECOS and Partial algorithm.

Existing publicly available data of river network generally have low accuracy and even the data of confluence area are inconsistent with the physical characteristics of water flow, which makes it difficult to been directly used for modeling water environments. In this paper, to solve this problem, a semantic modeling method of river network for large-scale 3D virtual water environment is proposed. Firstly, the semantic objects of the river network with hierarchical coupling relationships are constructed by River, Link and Intersection. The constituent areas of the river network are refined, and the topological relationships among different areas of the river network are represented. Secondly, the geometric semantic data of the intersection area are divided into three categories, including intersection boundary, confluence boundary and tributary boundary, based on the influence of the flow dynamics. The Chézy’s formula is introduced to model the tributary boundary, which is significantly impacted by water flow scouring. Finally, the boundary data of various areas in the river network are automatically generated based on these semantic data. The experimental results indicate that high-precision boundary data can be automatically generated for different river network areas, which can improve the realism and reliability of river network virtual simulation. Additionally, the generated boundary data can also be used for large-scale 3D virtual water environment simulation modeling.

To solve the problem of time synchronization between two stations in an optical two-way time synchronization system, an optical time synchronization scheme based on spread spectrum code ranging is presented. Time difference between sites is measured by spread spectrum code ranging, and link delay is eliminated by two-way method to achieve time synchronization between two sites. The accuracy of time synchronization between stations is affected by the accuracy of time difference measurement. By analyzing the structure and measurement principle of receiver code tracking loop, it is found that the accuracy of time difference measurement is mainly related to the signal system, carrier-noise ratio and receiver parameter settings. The carrier-noise ratio of links and receiver parameter conditions are determined. Through the analysis of different signal systems time difference accuracy, the signal system that can meet the requirements of highprecision time synchronization is selected.

The semantic segmentation task of medical image is to segment the focus, organ or substructure of human body in medical image. It plays an important role in locating and identifying the diseased area and making medical plan. In various medical image segmentation tasks, the U-shaped architecture has achieved great success. Transunet introduces Transformer with global attention mechanism into the U-shaped architecture, which overcomes the inherent limitations of convolution, but because it still continues the original skip connections structure, it will bring the strong noise from features in the shallow network into the high semantic features of the deep network, thus affecting the segmentation accuracy. UTSN-net model based on the combination of Transformer and nonlocal attention mechanism is proposed. UTSN-net uses Transformer to overcome the inherent limitations of convolution, and introduces the skip connections module based on nonlocal attention mechanism into the U-shaped network, which can comprehensively consider the deep features with global context information and the shallow features with accurate high-resolution positioning information to improve the accuracy of segmentation results. Experiments on synapse multi-organ abdominal CT dataset verify that UTSN-net has better semantic segmentation performance.

For the integrated navigation system GNSS/SINS in the process of data fusion, the traditional filtering algorithm does not consider the correlation between the two system and the poor robustness when measurement outliers occur, this paper proposes an iterated split covariance intersection filter algorithm to fuse the data of the two systems. It combines the Split CIF and the Gauss-Newton iteration and separate state covariance matrix into independent parts and dependent parts, and adjusts the posterior state estimation by calculating the Kalman filter gain iteratively during the measurement update process to reduce the error caused by outliers. The simulation show that the Iterated Split CIF based integrated navigation system has higher accuracy and better robustness. Compared with Split CIF and Kalman filter, the east velocity error is reduced by 30% and 35% respectively, and the latitude error is reduced by 22% and 30% respectively. In addition, the position accuracy still remains at a high level when outliers occur, so the algorithm has good robustness.

Due to the continuous and rapid expansion of the scale of the communication network, the deepening of the isomerization of the communication equipment in the network and the impact of network security factors, large-scale communication networks need to have a certain fault-tolerant mechanism to maintain connectivity for a long time. This paper studies faulttolerance from the perspective of structural statistics of communication networks. By studying how to make communication networks remain connected when some nodes fail, we can get the relevant laws, and provide the basis for highly reliable communication network deployment and the robustness of routing protocol design.

According to the principle of digital modulation and demodulation system, the Quadrature Phase Shift Keying (QPSK) modulation and demodulation module is implemented with Verilog HDL language, and the error caused by incoherent demodulation is reduced by compensating the circuit carrier error. In the process of development, intellectual property core (IP) is used to develop the communication system, including bit synchronization technology, baseband signal generator, phase-locked loop, etc., which greatly increases the convenience of development. The reception performance of visible light receiving circuit is improved by combining three-stage amplifier circuit, and the communication and anti-noise performance of visible light communication system are enhanced by using high-order modulation mode with multi-stage photoelectric conversion receiving circuit. Through the FPGA (Field Programmable Gate Array) development board for on-board testing, the whole set of visible light communication system is designed with photoelectric conditioning circuit

Because the memristor is a kind of nonlinear resistance, its resistance coefficient depends on the total charge q flowing through the element, that is, the output voltage of the memristor to the past time integral determines the resistance value of the device, according to this feature of the element can make the resistance value of the memristor realize programmable calculation, and then improve the performance of the circuit parameters. In this paper, we design a programmable memristor circuit by using a bipolar threshold memristor to realize the adjustable resistance value in the circuit, and apply it to the Schmidt trigger circuit. Through the memristor programming circuit to change the threshold setting voltage value of the Schmidt trigger, to achieve the controllable range of the back error, so as to achieve a better waveform integration effect.

This paper is aimed at improving the previous experiments of duffing oscillator detecting weak signals. The influence of phase transition direction, signal phase, frequency and amplitude on Duffing system is analyzed and a new method for detecting weak signal by Duffing Oscillator is proposed. This method detects weak signals by the way of reverse detection, and calculates the threshold through Poincare section, so that the detection is more accurate.

As one of the seven basic units in the International System of Units (SI), the temperature quantity is also the most important physical parameter of the ocean. The development of high-performance temperature sensing technology methods and sensors to achieve high-precision measurement of ocean temperature and compensation of elemental parameters is an important guarantee to improve ocean monitoring technology and promote the development of ocean resources. In this paper, we use mask lithography method combined with magnetron sputtering method to study the effect of microstructure size parameters of platinum film resistors on their temperature sensing performance, such as temperature coefficient, measurement accuracy, etc. This experiment simply and accurately determines the effect of the length and width of the resistor wire on the measurement error and temperature coefficient (TCR) of the platinum film resistor, and achieves the preparation of high precision and high stability platinum film resistors by controlling the structural parameters of the platinum film resistor wire.

In the information society, voice communication is the most common and indispensable in daily life, so the communication security issue is particularly important. Aiming at the security problem of voice communication, a low-cost communication security system based on single chip microcomputer is designed. The system is mainly divided into a confidential sending module and a confidential receiving module, which can realize two functions of ordinary call and confidential call, and realize synchronous confidential communication by using a single chip computer to generate random keys. The test results show that the system can realize the function of communication security, and can realize the module miniaturization, which is convenient to make a high density communication security tool.

To address the problem of low accuracy of traditional low-resolution radar target classification and recognition. In this paper, a low-resolution radar target classification algorithm based on a one-dimensional Densely Connected Convolutional Network (DenseNet) is proposed. The algorithm first directly downscales the Densely Connected Convolutional Network, then takes the original 1D radar target signal as the input for training, uses a segmented loss function for the characteristics of different classes of signals, makes the network use different loss functions in different training stages, and then back-propagates the loss to optimize the weights to improve the recognition effect of the network. The experimental results show that the recognition rate of the proposed method is higher than that of traditional radar target classification methods and simple one-dimensional convolutional neural networks (CNN) for low-spectral radar target classification, especially under low signal-to-noise ratio conditions, which fully demonstrates the effectiveness of the proposed method.

In recent years, the global prevalence of chronic kidney disease (CKD) has been increasing year by year, and heavy metals that are widely distributed in the environment are nephrotoxic, leading to possible kidney damage and affecting human health. Therefore, this study used laboratory heavy metal data from the National Health and Nutrition Examination Survey (NHANES) to select the main heavy metals that affect the kidney by fusing SHAP values and XGBoost algorithm of heavy metal selection method. Later, we combined Odds Ratio (OR) of heavy metals and quartiles of different population risk subgroups to validate the feature selection results. We found that the selected blood lead and urinary cadmium had a strong effect to CKD and the results were statistically significant. the method based on SHAP and XGBoost could discover the possible causal factors in vivo.

Enterprise information contains a large amount of valuable content. Analyzing enterprise information and clarifying the structure of the industry chain in which the enterprise is located can provide assistance in optimizing the structure of the industry chain. For this reason, this article proposes an enterprise classification model that integrates focal loss and ERNIE-textCNN to classify enterprises. The attention mechanism and textCNN are used to extract semantic features at different levels to solve the problem of missing features and contextual semantic relationships in enterprise short text data. To address the imbalance in enterprise data, the loss function in the model is modified to focal loss function. Experimental verification shows that in all samples, the classification accuracy of a small sample category can be improved by 10%.Finally, the enterprise is matched to the industrial chain graph.

Aiming at the problems of poor detection results and low reading accuracy caused by small targets and oblique shooting angles of pointer instruments in complex background environments, a combined method of depth learning, perspective transformation, and Canny edge detection was proposed to perform pointer instrument readings. This recognition method uses an improved YOLO V7 target detection algorithm to detect and extract instruments in complex environments, and then corrects the extracted instruments through perspective transformation. Finally, the Canny edge detection algorithm and Hough transform are used to determine the center and pointer characteristics to obtain pointer readings. Through experimental comparison and verification, this method is more accurate and reliable than traditional methods, with a certain speed. It provides a more accurate and faster method for identifying pointer type instrument readings for subsequent work.

Based on the information provided in Problem D of the 2022 National College Mathematical Modeling Contest, only considering the transmission and reception of meteorological message between the main stations, the success rate of sending and receiving is 100% , set half-time information as one unit, then each main station can transmit up to three units、150 characters (less than the limit of 158) at a time.Using the method of sequential transmission in turn, the first round and the last round are studied separately, the other rounds meet the maximum transmission of cases, the relationship between the minimum value of K and N is modeled, as shown in the formula (1) . As result, it takes at least 4.67(rounded up to 5) minutes to transmit 6 rounds, the information sharing between 9 master stations can be realized. Then, according to the state characteristics of each main station after each round of transmission,the state transfer equations between the main stations are established, such as formulas (8) and (9), and the scheme of meteorological message transmission for each round of main station 1 is obtained.

With the increasing diversification of signal modulation types, the importance of signal modulation recognition is increasing, which is an important part between signal detection and demodulation. It has great applied value in jamming identification, electronic countermeasures, intelligent modem and other fields. Aiming at the improvement of recognition accuracy for some modulation types, a communication signal modulation recognition method based on cyclic spectrum features and bagged decision tree is proposed. The method extracts the cyclic spectrum features of signals and inputs them into the bagged decision tree for model training. Simulation results show that the accuracy of the proposed method reaches 93.8%, which is 39.4% higher than that of the traditional recognition method with high-order cumulants and 22.2% higher than that of the method using the original signal directly.

Infrared target images have low signal-to-noise ratio, blurred edges and missing textures, which make it a great challenge to identify the target and achieve stable tracking in the tracking process. However, ordinary target trackers use feature fusion as a convolutional operation, which is a local matching process that easily leads to the absence of high-level semantic information of the image, and is further limited on infrared images. Inspired by transformer, its attention mechanism can capture global features, as well as contextual relationships between features, and can well establish the association between remote features , long-range dependency and other advantages, we designed transofmer-based infrared target tracker, which is a network that performs feature enhancement and fusion on infrared images by tranformer, and classifies and regresses targets by classification head, and has proved the effectiveness of the method by conducting extensive experiments on challenging benchmarks.

The study of cell nuclei is the starting point of pathological analysis and new drug development in modern medicine [1], and nuclear segmentation is a primary task of nuclear research. This paper proposes an optimization method for nuclear segmentation. It regards the Conditional Generative Adversarial Network (CGAN) [2] as the fundamental segmentation structure, segments the nuclear images by using deep learning Convolutional Neural Network (CNN) [3], and then optimizes and improves the generator, discriminator, and objective function. The experimental results demonstrate that the improved UGAN has superior performance on the semantic segmentation task of the images of pathological tissue slices and can be used as a tool for automatic segmentation of pathological tissue sections.

Syndrome differentiation (SD) is a basic task in TCM (Traditional Chinese Medicine) diagnosis and treatment. TCM syndrome differentiation is very complex and time-consuming. Meanwhile, the accuracy of the results depends on the experience of TCM practitioners. To help TCM practitioners differentiate syndrome more quickly, we propose a syndrome differentiation method of deep learning based on multi-feature fusion. We extracted char, word and POS (Part of Speech) from TCM diagnosis and treatment records. The vector representation of char feature is obtained by ZY-BERT (Zhong Yi BERT), ZY-BERT was pre-trained on large datasets of TCM-SD (TCM Syndrome Differentiation). The vector representation of word and POS is obtained by Word2vec (Word to vector). We construct text graphs of char, word and POS according to context. GCN (Graph Convolutional Networks) is used to extract spatial structure information between multiple features to achieve multi-feature fusion. The experiment was carried out on TCM-SD. The experimental results showed that the accuracy of the proposed method was 81.52%, which was better than the comparison method. This method is helpful in the development of TCM modernization.

In this study, we present a new method for remote detection of mental stress via webcam. The system is based on remote Photoplethysmograph (rPPG) obtained from face video frames of heart rate, breathing rate, and pulse rate variability (PRV). The experiment collected pulse wave data from 14 healthy students with a stress distribution consisting of four phases: Rest, Stroop-Color-Word Test, Mental Arithmetic Task, and Recovery. We combined the stress questionnaire to select data to assess the human autonomic response to stress and recovery, the results showed significant differences in frequency domain characteristics and nonlinear parameters between phases. The average classification accuracy under different stress sources was 80.31%. The results demonstrate the applicability and convenience of the remote stress detection method. It can be used without disturbing a person’s daily life and provides an alternative to traditional contact techniques for those who want to monitor stress levels regularly.

In order to achieve reliable terahertz communication at ultra-high data rates, this paper proposes Turbo coding with S-random interleaver combining circular propagation constraint and circular shift, firstly, the overall structure is optimized mainly by using parallel loops for multi-core computation and reserving memory in advance, secondly, the coding structure is optimized, then S-random interleaver with circular propagation constraint and circular shift is used, and finally, the decoding The structure is optimized. Through simulation experiments, it can be concluded that the proposed Turbo coding method improves the coding operation rate and reduces the BER and transmission delay.

Aiming at the problem that the terahertz wave beam is narrow and directionality is strong, which makes it easy for the enemy to eavesdrop on the data information, and thus the information cannot be transmitted safely, a pseudo-random sequence generation method based on n-LFSR and AV-NL chaotic mapping is proposed. Different algorithms are selected according to the number of LFSRs. At the same time, the absolute value nonlinear chaotic map is generalized. Finally, the generated two pseudo-random sequences are XOR combined to form the final pseudo-random sequence for source encryption. The simulation results show that this method takes into account the advantages of simple design of linear feedback shift register and high chaotic randomness. The generated pseudo-random sequence has both high linear complexity and the advantages of chaotic non-linearity. The high randomness used in source encryption can improve the security of the system.

Emotion recognition from social network texts aims to mine netizens’ subjective emotions, such as stances and emotional tendencies, over an event, which is imperative for monitoring Internet public opinion. Convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) are widely used in the field of text classification. The combination of the two can exploit the CNNs’ feature extraction ability and BiLSTM’s ability to extract context dependency. However, for emotional recognition, it is also necessary to consider some specific words in online social text. Therefore, we constructed a model for emotion recognition based on Internet public opinion in three steps. First, the CNN was used to extract local features of social network texts. Second, context-related global features were extracted using the BiLSTM. Finally, we introduced an attention mechanism to obtain important features. Experiments were conducted using netizens’ comments from a microblog during the COVID-19 epidemic as the dataset. Experimental results showed that the feature vector of the proposed model (i.e., the CNN-BiLSTM-Attention model) contains richer semantic information of texts, which can effectively improve the performance of emotion recognition from Internet public opinions.

The main factors limiting the long-distance application of vortex beam are the low receiving power and the wavefront phase distortion caused by atmosphere turbulence. Coherent beam combining (CBC) technology is an effective way to generating high power vortex beams. However, most common coherent combined vortex (CCV) fiber array is currently based on a single-ring structure with limited output power enhancement. In this paper, a dual-ring fiber array is developed to achieve higher output power and improved stochastic-parallel-gradient-descent (SPGD) correction accuracy. To improve SPGD correction speed, cross-grouping method is used. The results show that CCV beam in dual-ring structure can maintain good intensity distribution and mode distribution after SPGD correction.

For achieving fast and flexible retrieval across heterogeneous modalities, unsupervised is more flexible and easy to use than supervised methods, of which the unsupervised method GAN is the most popular. However, GAN has been suffering from the problems of lack of diversity in generated samples, debugging difficulties and training instability. A cross-modal hashing method based on a diffusion model is proposed in the paper. Specifically: (1) For the first time, the diffusion model is applied to the field of cross-modal retrieval, targeting three modalities for mutual retrieval. (2) The combination of adversarial network GAN and diffusion model improves the sample quality and sample diversity, and ameliorates the problems of complex GAN debugging and unstable training. The effectiveness of the proposed method is demonstrated through experiments on three datasets and comparison with state-of-the-art methods.

The paper focuses on the integrated analysis of the communication control platform and proposes the integration of communication information and its processing through middleware, so that users can find the desired information more accurately and improve the efficiency of maritime communication. The integrated communication and control platform for ship networking makes use of VTS (Vessel Traffic Management System) and AIS (Automatic Identification System) information fusion technology, related network technology, information processing technology, data management technology, etc., to realise the integration of communication automation betweenin the ship and between the ship and the outside world, solving the problems of scattered and confused ship communication information, thus realising efficient data communication and resource sharing, which makes the various systems on the ship highly integrated and integrated, and solves the problem of long communication time and short communication distance.

Distribution of the radial electric and temperature field is an important parameter to characterize the performance of resin impregnated paper (RIP) high voltage direct current (HVDC) bushings. The upper discharge path is air insulated discharge, and its 50% impulse discharge voltage U50 can be calculated by lightning impulse withstand level and operation impulse withstand level, which can be corrected to 50% impulse flash-over voltage under standard atmospheric conditions, and the minimum safety clearance can be calculated according to the air clearance formula under standard meteorological conditions. In the paper, under 6250A current, the maximum temperature of fittings has reached 80 ℃, which exceeds the maximum operating temperature of 90 ℃ of the air medium bare aluminum alloy material in GB/T 11022 and the maximum temperature rise of 50 ℃ in relative environment, far exceeds the serious defect limit of hot spot temperature exceeding 80 ℃ in DL/T 664-2008 Code for Diagnosis and Application of Live Equipment, approaches critical defect limit of 110 ℃, and cannot meet the requirements for long-term operation reliability on site. The design and process are improved and optimized to improve the current carrying temperature rise performance. The results are in good agreement which proving validity of the finite element electro-thermal coupling model. Then the finite element model was used in the cone model which is much closer to the actual condenser structure. Optimization design for the condenser outer contour structure was also conducted to make sure that the hottest-spot temperature is close to 90℃ under rated operating condition, and the maximum radial E-field strength is approximately equal to 3.5kV/mm and distributed uniformly. Finally, the temperature rise and electrical performance test were carried out on RIP HVDC bushing prototype, which can vertify the feasibility of optimized design.

Under the shadow of the epidemic, people's health awareness is increasingly enhanced. Wearable intelligent devices have the functions of "sports health" and "body monitoring" to provide users with comprehensive and professional health management. This paper presents a design scheme of a Bluetooth smart bracelet about sports health and body detection. This electronic bracelet adopts triaxial acceleration sensor ADXL345, which can obtain the dynamic data of human body through single-chip control. It has the advantages of low power consumption, high precision and high sensitivity. It is a new favorite of consumers in the field of electronic products.

In this paper a 1 to 4 time-delay power-divider was provided. The phase and amplitude relation between each port was simulated. The time-delay between the port was 0.016 ns in the x-direction and 0.16 ns in the x-direction. The maximum amplitude difference between each port was 1 dB and the maximum phase difference error was ±6°. The S11 value was lower than -20 dB. The proposed time delay power divider can be applied to VICTS antenna to enhance the instantaneous bandwidth.

Electroencephalograms (EEG) signals can be used to decode hand movement parameters. Most of the researches on decoding hand movement parameters concentrated efforts on low-frequency feature of EEG extracted from timefrequency domain. However, complex brain structure and performance of various frequency bands have been less taken into account for hand movement direction decoding. In this paper, the node strength variability (NSV), a novel feature, was proposed to decode hand movement between horizontal and vertical direction. It is generated from discrepancy of connectivity among electrode nodes in two adjacent periods of brain networks constructed based on the phase locking value (PLV) from EEG signals. Five volunteers participated in our experiments, and totally 600 sets of EEG data were collected. NSV of five distinct frequency bands obtained between preparation and movement execution were applied to classify hand movement direction for each subject through a ten-fold cross-validation support vector machine (SVM). The results indicated that NSV of alpha band has the best effect on distinguishing horizontal and vertical hand movement directions, which provides new ideas for hand movement direction decoding.

As a basic module of integrated circuits, LDO is widely used in digital-to-analog and analog-to-digital conversion circuits. Based on the requirements of these circuits for power management chips, a fast-transient capacitorless low dropout regulator is proposed in this paper. To enhance the transient response, a push-pull buffer is added between the error amplifier and the power transistor. In addition, a transient enhancement circuit is used to effectively suppress the undershoot voltage and overshoot voltage and reduce the recovery time. The capacitorless LDO is verified in TSMC 180nm CMOS process. The simulation results show that when the load current changes from 1mA to 100mA within 1us, the output undershoot voltage is 83mV, and the recovery time is 1.5us; When the load current changes from 100mA to 1mA within 1us, the output overshoot voltage is 80mV and the recovery time is 1.7us. The power supply rejection of the whole system is -60dB at 1kHz and can still reach -17dB at 1MHz.

Time digital conversion technology has been widely used in the field of measurement technology. A high-resolution, widerange Time-to-Digital-Converter (TDC) circuit is proposed based on Multiplier Delay-Locked Loop (MDLL) and Time Amplifier (TA). In this paper, a two-segment combination structure is adopted to obtain a wide range by using the cycle counting method in the high segment, and to take into account the high resolution by using the time amplification technique in the middle and low segments. Finally, a TDC circuit with a range of 500 ps ~ 4 us and a resolution of 18 ps is presented.

In the past decade, the research of wireless sensor network (WSN) has aroused great interest from academia, industry and government. The WSN testbed has become the key of various development projects by conducting more authentic and reliable experiments in capturing the underlying hardware, software, algorithms and dynamics of WSN. These experiments need to solve several problems, such as experimental reliability, resource efficiency, and sustainable maintenance. Due to the problems of wireless sensor networks and the diversity of applications, the wireless sensor network test bench needs to flexibly support various network topologies and be robust to achieve reliable experimental results. This puts forward practical requirements for the reliability of control software. This paper designs a stable state machine to simplify the experiment, remote health diagnosis during operation, automatic replacement of fault nodes and TDMA data log transmission to improve the reliable operation of control software.

In recent years, wind power has become more and more important in the energy component. In order to improve the prediction accuracy of wind farms and help management and scheduling, a multi-site short-term wind power spatiotemporal combination forecasting model based on dynamic graph convolution and graph attention is proposed. Firstly, graph convolution is used to realize neighbor aggregation of temporal features between multiple sites, and the graph attention mechanism is used to enhance its ability to extract spatial features. At the same time, in view of the problem that the traditional model cannot deal with the real-time change of graph node correlation, the adjacency matrix is dynamically constructed according to the correlation coefficient and distance between nodes in the graph convolution process. Finally, the Gated Recurrent Unit is used to process the context information of dynamic graph convolution output to complete the prediction of wind power. The experimental results show that the proposed combined model is optimal in the aspects of prediction accuracy, stability and multi-step prediction performance.

The main variable temperature controller has not achieved an automatic testing function so far, which consumes great labor and capital costs. The innovation of this paper is to achieve fully automatic and intelligent detection of the main variable temperature controller. We construct an intelligent testing equipment remote control system, including a remote monitoring and management system for testing equipment backstage, a mobile APP for testing equipment, an APP for receiving terminals, and an application service for the testing report interface, which realizes the function of remote management and viewing of reports. A fully automatic calibration device for the main substation thermostat is designed, including the inspected thermostat, constant temperature oil tank, calibrator, and graphic conversion device, which realizes the device to automatically complete the testing of substation-related devices. By designing the communication serial port of the device and developing the corresponding device on the cloud, the software and hardware are well integrated. The technical solution promotes the turnover rate and utilization rate of testing equipment and enhances the comprehensive management capability of testing equipment.

Recently, Low Earth Orbit (LEO) satellite constellations with low-latency and high-bandwidth attract extensive research. However, most available studies focused on the field of satellite network routing algorithms, ignoring the impact of topology on the efficiency of inter-satellite networking and the quality of inter-satellite communication. In this paper, we propose a satellite network topology design method based on deep reinforcement learning (DRL), with the goal of reducing the latency of the entire satellite network. To achieve this goal, we first model the satellite network communication scene and formulate the topology optimization problem as a Markov decision process (MDP). Then, we further propose the idea of backbone-point satellites and use DRL to optimize the topology structure. Finally, we conduct extensive experiments on different performances of satellite topology, and we conclude that the network topology constructed in this way can provide lower latency communications than the motif and +Grid topologies, optimized by 8.48% and 42.86% respectively.

The traditional test of the system using graphical interface determines the test results by manual. With the higher accuracy requirements of the test software system in current industrial domain, the traditional manual testing has the problems of high costs, low accuracy and efficiency. Then an automatic test system based on image matching algorithm is proposed to simulate manual automatic test result. Firstly, the architecture of automatic test system is researched, which reduces the test cost. Then, an automatically selected method by template matching algorithm and ORB algorithm for the complexity of the target images, which improves the matching rate and accuracy. Finally, an image similarity calculation algorithm based on perceptual hash algorithm is built, which improves the test accuracy and efficiency twice. The experimental results show that the accuracy achieved to 97%, which meets the accuracy requirements for the industrial field.

Aiming at the problem that Tracking accuracy of Tracking-Learning-Detection (TLD) tracking algorithm decreases when targets are under different light and shade conditions and target scales change, an improved TLD tracking algorithm is proposed. In this paper, Speeded Up Robust Features (SURF) feature point matching method was adopted as the tracking module, and the feature point pairs with low confidence were removed by adding the evaluation of feature point pairs. By introducing Contrast Limited Adaptive Histogram Equalization (CLAHE) into the detection module, a random Circle feature classifier is proposed, and the HOG feature matching method is used to replace the normalized correlation matching method in the nearest neighbor classifier. In addition, the detection range is adjusted adaptively, which reduces the computational complexity and effectively improves the adaptability of the algorithm to multi-scale. Experimental results show that the proposed algorithm can effectively overcome the influence of environmental shading conditions, and has strong robustness to scale changes and high tracking accuracy. Compared with the classical TLD algorithm, the improved algorithm performs better.

The parameter measurement of gas-liquid two-phase flow has important industrial applications and scientific research significance. In order to realize the measurement of gas-liquid two-phase flow parameters, this paper designs a sensor for capacitively coupled contactless conductivity detection (C⁴D) based on series resonance technology. Through the analysis and design of the sensor structure, it is not only possible to combine the designed sensor shielding structure, which can weaken the influence of external interference signals; it is also possible to eliminate the influence of capacitive resistance in the measurement through the series resonance technique; and its input and output characteristics are monotonic, which facilitates the measurement of parameters. In order to verify the theoretical linearity of the input and output characteristics of the dual inductance sensor, conductivity measurement experiments were carried out to obtain the result that the sensor is monotonic in input and output characteristics, but not linear in practice. After the demonstration and experiments, the results show that the developed dual inductance sensor validates well and meets the requirements of the detection.

With the development of deep learning techniques, computer vision techniques have also been significantly improved. Image retrieval is a common technique used to retrieve images of interest from image databases, which can help users find the desired images more quickly. However, traditional image retrieval methods often fail to meet user needs because they often ignore complex scale information, e.g., features may differ at different scales. Therefore, an image retrieval based on a region-attention feature fusion mechanism can overcome this drawback, and it can improve the performance of image retrieval by emphasizing multi-scale features through a region-attention mechanism. In this paper, we propose an image retrieval method based on regional attention based multi-scale feature fusion, which can effectively use multiscale features. The effectiveness of RMFF is demonstrated by conducting experiments on mainstream image retrieval datasets.

In order to improve the accuracy of heart disease prediction models and address the lack of interpretability in traditional machine learning models, this paper proposes a heart disease prediction method based on random forests and SHAP value. This method first preprocesses the dataset by encoding the data, filling in missing values, and removing outliers. It then uses recursive feature elimination and cross-validation to remove irrelevant features and select relevant features for further model training. The results, compared with other methods using accuracy, precision, recall, and F1 score, show that the proposed method outperforms other models. The interpretable model constructed based on SHAP value reflects the effect of feature values on prediction model results and provides a ranking of feature importance. The experimental results show that the method can effectively improve the accuracy of heart disease prediction, and provide a clear interpretation of the model prediction results. It can be an aid in the treatment and prevention of heart disease.

Dot matrix invisible code is widely used in anti-counterfeiting and traceability of goods, dot matrix invisible code is a kind of technology that "disappears" in the package decoration. This kind of code will neither destroy the overall effect of packaging decoration nor affect the function of barcode, it provide technical support for anti-counterfeit traceability means, which is difficult to be seen by the naked eye and needs to be read under special lighting conditions, merchants and consumers can access product information by identifying the code, it often has the problem of poor contrast due to light intensity, shooting angle and other reasons. The image enhancement technology is used to improve the image quality and lay the foundation for the subsequent work. This paper proposes an improved γ-CLAHE image enhancement algorithm for dot matrix invisible code, which converts the image into the color space with color and brightness separation, performs histogram equalization enhancement processing on the brightness components, combines with Gamma correction, so that the enhanced image quality is significantly improved. In this study, the CLAHE algorithm and the improved algorithms on LAB, HSV and YCrCb color spaces are compared separately, the experimental results show that the improved algorithm is much more effective than the CLAHE algorithm, and the improved algorithm in YCrCb color space is more suitable for image enhancement of dot matrix invisible codes than other color and bright separation color spaces, It has obvious superiority in indicators such as information entropy, mean gradient, standard deviation, etc., and can effectively improve the contrast of low-quality dot matrix invisible codes, at the same time, this study provide ideas for similar invisible code image enhancement.

Computer simulation technology has changed the traditional process of comparison and selection of architectural design schemes. Unlike the traditional process, which relies more on the experience of architects, computer algorithms can realize the calculation of quantitative indicators for complex problems and provide objective basis for scheme comparison and selection. Combined with the application of engineering projects, this article expounds the application of computer-aided design in architectural design through the method of calculating visibility percent of quantitative indicators to judge the visibility of the interior display space design in small shopping centres by multi-software simulation.

When using the non-contact method to collect the image of the back of the hand, in view of the problems that the different degrees of finger opening and closing affect the segmentation of the finger image, the area of interest of the knuckles of the knuckles is difficult to locate, and the image is easily affected by light, a knuckles of the knuckles of the knuckles based on morphological operations are proposed. Region of interest positioning method. In this method, a combination of opening and closing operations in morphological operations is used to construct an image mask of the back of the hand, and then the mask is used to shield the area of the back of the hand, and five images of the back of the finger are separated and angle-corrected. For the corrected image, the top-hat transformation is performed first, and then the method of combining the sliding window algorithm with image information entropy is used to locate the knuckle pattern area of the dorsum of fingers. Finally, LBP and SIFT are used for feature extraction, and support vector machine (SVM) is used as a classifier, and experimental verification is carried out on the self-built back of hand image database. The experimental results show that the proposed method can eliminate the influence of illumination and environmental noise, and can accurately and quickly locate and extract the region of interest of the dorsal knuckle pattern on the back of the hand image.

This paper takes the toll station system as the research object, and based on the analysis of the current situation and demand of domestic toll collection business, defines the intelligent toll station for the purpose of credit creation and unmanned toll station construction, and designs the overall scheme architecture of intelligent toll station. Based on super-fusion structure technology, disaster tolerance technology, health insight technology, intelligent anti-theft network monitoring technology, self-service payment terminal is studied. Based on virtualization, Docker, micro-service architecture, distributed cluster and other technologies, this paper designs and studies cloud-based systems (cloud toll hall), and develops software and hardware systems of intelligent toll station. The key technology of intelligent toll station is applied in Guangdong province and evaluated according to the needs of intelligent toll station. In order to ensure the application effect of the intelligent toll collection system, the evaluation method of the application effect is studied. The evaluation index system of the application effect of the intelligent toll collection system at the station level, including 3 first-level indicators and 13 second-level indicators, is constructed through on-site investigation, discussion, expert consultation and other forms, considering toll station passage, service and resource utilization. The toll collection system and effect evaluation method are applied and validated in Guangdong Province..

In recent years, end-to-end automatic speech recognition (ASR) based on deep neural networks become popular because of its simple pipeline and excellent performance. However, there exists a main mismatch between its training and testing that might lead to performance degradation: in the training stage, existing method use the maximum likelihood criterion which aims to maximize log-likelihood of the training data, while in the testing stage the performance is evaluated by word error rate (WER), not log-likelihood. In this paper, we propose an alternative method based on reinforcement learning to make the goals of training and testing more consistent. Viewing speech recognition as a sequence decision process, the encoder-decoder based neural network is used as the policy function. The encoder is a pre-trained speech representation model (Wav2vec2.0), which generates the environment state encoding. The decoder is trained using a policy gradient algorithm based on a mix reward function which reflects both the word error rate and language model score. Experimental results on the LibriSpeech corpus show that our proposed method achieves 4% relative improvements over the baseline with a language model in terms of WER.

Image features are essential components for physical detection, classification of objectives and downstream tasks. Specifically, the image features can be utilized to automatically detect the characteristics of images and realize physical information mapping into information domains. However, existing image features are concentrated on the decrease the contrast, which can reduce the influence of lights. Another extraction process converts images into digital images and utilize digital information techniques to obtain the features. In this paper, we utilize the machine learning model to extract the features of images with enough training iterations. Initially, we utilize the CIFAR-10 data set, which contains the 10 categories of physical objectives and simulate as the training set. Indeed, the establish machine learning model is utilize to train through inputting the 80% of total data set. After training process, the output of machine learning mode can obtain the features of any physical images. Finally, we compare our proposed model with existing image features extraction methods and utilize 20% data to evaluate our model. From our extensive experimental results, we can conclude that our established model can effectively achieve the image features extraction with higher extraction accuracy and acceptable computation time through comparing with traditional mathematical analysis methods.

This article presents a machine learning-based method for detecting micro-sleep. The method is simple, efficient, and can be applied in practical scenarios without the need for large-scale equipment such as servers. We recorded the physiological characteristics of 16 young adults in a driving simulation laboratory, mainly consisting of electroencephalogram (EEG) and driver behaviour videos, and used machine learning to detect micro-sleep events. We compared different machine learning algorithms (SVM, KNN, ANN) and ultimately adopted a combination of ANN and SVM algorithms (pre-processing small-scale data), which reduced the recognition error rate from an initial 4.5% to 0.2%. This combination accelerated the recognition speed and improved the accuracy, making it a practical approach.

In a multi-cell and multi-user cellular system, the communication quality of users is seriously affected due to the existence of inter-cell interference (ICI) and inter-user interference (IUI).As an interference management technology, interference alignment (IA) can not only eliminate interference, but also greatly improve the degree of freedom of the system, that is, improve the system capacity.This paper proposes a three-cell multi-user grouping interference alignment algorithm,which first groups the users of the three cells,converts the complex interference channel into a simple communication channel,and then eliminates the inter-user interference by transmitting beamforming matrix, and eliminates inter-cell interference by designing the receiving beamforming matrix.When calculating the transmitting beamforming matrix, the multi-matrix connected at the transmitter and solved by the eigenvector.When calculating the receiving beamforming matrix, the receiving beamforming matrix is obtained by using the inverse property of the matrix.Compared with the traditional method of solving the transmitting beamforming matrix and the receiving beamforming matrix, the complexity of the algorithm is greatly reduced.And the simulation shows that the algorithm has a better sum rate than the traditional interference alignment algorithm under the same signal-to-noise ratio (SNR).

In this paper, a non-stationary detection method based on the artificial intelligence algorithm XGBoost is proposed for the detection of the U-value of the vacuum glass. By analyzing the heat transfer characteristics of vacuum glass and considering the detection efficiency, the features are selected as hot end temperature, ambient temperature, and characteristic temperature change rate. In this paper, the training effect of a model is measured comprehensively by the scores of MAE, MSE, and R2. Three models, KNN, GBDT, and XGBoost, are used to train the dataset and compare the prediction results. After the comparison, XGBoost has the best prediction effect. Finally, the fitted model is validated by 5*2 nested cross-loop, and the analysis results show that the fitted model has better stability, which greatly enhances the credibility of the model. After a series of experiments, it is known that the small sample of non-stationary method and multiple interference problems can all be solved by XGBoost algorithm with certain stability, which can provide ideas for further industrialized testing.

In this paper, a new method is proposed for the intelligent prediction of the thermal insulation performance of vacuum glass, i.e., the use of cascade forest algorithm to detect the heat transfer coefficient (U-value) of vacuum glass. By constructing different intelligent algorithm models, random forest, extreme random forest and cascade forest algorithms are used. By evaluating the proposed method using mean absolute error (MAE), mean square error (MSE) and R-squared value, the cascade forest was evaluated with values of 0.0401, 0.0035 and 0.9896, respectively, and the predicted value curve was very close to the true value curve, so it was concluded that the cascade forest algorithm was superior to the random forest and extreme random forest algorithms in predicting the heat transfer coefficient of vacuum glass. In order to avoid the risk of overfitting, k-fold cross-validation was also added to each random forest in the cascade forest during the training process, and the accuracy of the cross-validated data was improved by 1% as shown by the data. It is known from the experimental results that the algorithm with cascade forest gives a new idea for the work of fast detection of heat transfer characteristics of vacuum glass based on small samples.

The low-orbit satellite constellation is an important part of a space-ground integrated network. MEC servers can be deployed to realize on-orbit processing by offloading tasks to satellites. On the other hand, NDN natively provides mobility support for consumers, so integrating NDN into the satellite network can solve the difficult mobility management problems of IP. However, input parameters in current offloading algorithms are default obtained from users, and they rarely consider the existence of inter-satellite links, which are not suitable for NDN scenarios. Therefore, we propose an architecture, studies the task offloading problem in LEO satellites based on NDN and proposes an optimization algorithm. Simulation results show that the algorithm finally converges and has better performance than the other two algorithms.

Due to its efficiency, energy-saving, and abundant data reuse, systolic array has been a popular choice for Convolutional Neural Network (CNN) accelerators. Dataflow of the systolic array defines computation mapping strategy and memory access and it is one of the most important design points of accelerators. Most conventional accelerator designs choose a single dataflow and optimize around it. This may influence the Processing Element (PE) utilization rate and cause waste of computing resources and energy. This work introduces a self-paced method to alleviate this problem. We analyse and quantify the PE utilization rate related to the three basic dataflows and build a model called PEU-sim to explore workload-oriented flexible dataflow. Experiments show by combining three dataflows, we are able to raise more than 10% of PE utilization rate for most neural networks and we get the highest of 12.4% for MobileNet.

With the rapid development of deep learning-based neural network models, traditional general-purpose chips can hardly meet the needs of large-scale neural network computing tasks, so artificial intelligence chips come out along with the trend. In the application area of artificial intelligence chips, many computational tasks are related to multiplication, and multipliers are used very frequently. Due to the complexity and optimizability of multipliers themselves, their approximate design can effectively reduce the power consumption and improve the computational power of artificial intelligence chips. In this paper, we design two approximate multipliers based on Karatsuba algorithm and radix-8 booth multiplication, which reduce the power consumption by more than 45% and can be well adapted to various different neural network models to replace the exact multiplication unit of artificial intelligence chips.

In this paper, a novel approach for improving the output cartoon-style effect of low-resolution face images in AnimeGAN is proposed, which is an useful and effective method to generate high-quality cartoon-style face images. This new approach I proposed combines generative adversarial networks (GAN), AnimeGAN and SRGAN. The previous existing methods do not give satisfactory results on processing low-resolution images. The cartoon-style images generated from the LR images have many significant visual issues. For example, the output cartoon-style faces have some unreasonable and weird shadows and wrinkles, which is unreal and far from the effect of original images. In this paper, I introduce a new method of using SRGAN to increase the resolution of the original LR images in order to improve the output effect in AnimeGAN. At last, the experimental results show that my combined method has good output from LR images and improves the cartoon-style faces effect as well as the performance of AnimeGAN.

Seed is the key link to ensure the yield and quality of crops, and seed quality plays a vital role in ensuring agricultural production. For seed quality testing, traditional methods have problems such as irreversible destruction, strong professionalism, complex process, long cycle, high cost, and inability to detect seeds online and real-time, which are difficult to meet the requirements of modern agriculture for non-destructive, rapid and efficient production of seeds. Spectral imaging technology has been widely used in the non-destructive and rapid detection of agricultural products, food, etc., with its advantages of fast, efficient, nondestructive and simple. The research progress of three spectral imaging technologies in seed quality detection was reviewed from the perspective of multispectral image recognition, hyperspectral image recognition and near-infrared spectral imaging. The basic principles, key technologies, implementation methods and application research results of the three spectral imaging technologies were compared, analyzed and summarized. The advantages and problems of spectral imaging technology detection and identification in crop seed quality detection were pointed out, In view of these problems, the future research directions are given to improve the understanding of agricultural information detection process using spectral imaging technology and promote the development of agricultural seed sources and other key agricultural core technologies.

Two-wheeled tour balance car is a complex system model which contains a variety of variables with no definite parameters. It is characterized by nonlinearity and strong coupling. The balance car can be used to detect the advantages and disadvantages of control methods. It has broad application prospects. In this paper, an inverted pendulum model is constructed to simulate the balance conditions of the car. Kalman filtering method is used to optimize the inclined angle of the car. The control parameters are adjusted by PID to achieve the balanced operation . A new image acquisition scheme based on TSL1401 linear CCD is adopted. The STM32 processor processes heartbeats to control the speed of the left and right motors for autonomous line patrol. After testing, CCD line performance is better than infrared, with forward-looking, low jitter, relatively fast and other advantages.

High-precision point cloud map is critical for Level 3 and above autonomous vehicles to operate effectively. When autonomous vehicles only use LiDAR to construct the point cloud map, drift deviation will occur in the horizontal and vertical directions, which affects the quality of the constructed map. To solve this issue, this paper proposed a Simultaneous Localization and Mapping (SLAM) framework for outdoor autonomous vehicles based on LiDAR. Firstly, the ground constraint was constructed by the extracted ground point cloud and the standard ground normal vector. Secondly, the preliminary pose was generated by the feature point cloud pose constraint and the ground constraint combined with the uniformly accelerated motion model. Finally, the preliminary pose and similar scene pose transformation detected in loop closure detection were added into the pose graph, and an optimized pose was obtained to construct the point cloud map. The proposed method was evaluated on the MulRan dataset and real scene. The results showed that the proposed method can achieve the positioning deviation within 0.86m and 3.1m respectively in the 3.3km and 7.7km motion range. It was superior to the comparison algorithm in local accuracy and global consistency.

With the advent of the "Internet +" era, network viruses, hacker attacks, and personal data breaches have gradually become the cybersecurity issues that people are concerned about. Blockchain technology has the advantages such as decentralization, immutability, and smart contracts. It can be used in network infrastructure protection to prevent computers from being hacked illegally, ensuring the integrity and security of users' personal data. In this paper, we propose a DDoS attack detection method based on the attention mechanism of LSTM that involves using a deep learning model to analyze network traffic data and identify DDoS attacks, named A-LSTM. The attention mechanism in LSTM helps to focus on the most important features in the data, which are critical for accurate DDoS attack detection. This method has the potential to improve the accuracy of DDoS attack detection compared to traditional methods by leveraging the powerful pattern recognition capabilities of deep learning models.

With the rapid development of cloud computing technology and various applications, enterprises have also built their own cloud platforms.As each information system gradually goes to the cloud and monitoring data increases, the security protection pressure of the cloud platform increases accordingly, resulting in weak data analysis ability and low alarm failure. Computing performance has become a key problem that restricts the security monitoring and analysis of power network under cloud platform. Aiming at the performance bottleneck of security monitoring under cloud level, this paper proposes a method to improve the efficiency of big data monitoring and analysis. This method collects the log data of all kinds of security equipment and security system in the network, and uses the real-time processing framework of Flink and ODPS data processing service on the cloud to quickly analyze the big data. Experimental results show that compared with traditional self-built Hadoop platform and real-time computing Storm, this method consumes less resources and has a faster processing speed for the same log volume.

In order to improve the self-service convenience of campus teachers and students payment, and enhance the ability of campus smart services. This article proposes a smart campus payment service system based on Microservice architecture, which mainly relies on the basic capabilities of the Spring Cloud framework and Docker containers to build the entire Microservice platform. According to the principle of micro-services, the system is divided into micro-services according to functional modules, so as to reduce mutual calls between micro-services, achieve low coupling and high cohesion, improve the stability and scalability of the system, and make the payment process configurable and payment data visualization.

The soil bioremediation process of coking sites is complex, the site environment is harsh, and the project period is long. Compared with the fields of water and air pollution monitoring, the informatization level of soil bioremediation project is low, and it is urgent to improve the digitalization and intelligence. Through the design of an online monitoring and electronic inspection system for the bioremediation process of coke contaminated soil and the development of intelligent early warning software, a study of information-specific technologies and data models for coke contamination remediation has been conducted. This paper focuses on three core elements of this field, including multidimensional data collection technologies such as Internet of Things and image recognition, big data processing technologies realized by relying on communication modules and cloud platform databases, and the construction of a neural network computational model for the soil bioremediation process. The information system has been tried out in the pilot process of soil bioremediation, realizing information management functions such as monitoring the operation status of sensors, inspection management, equipment's own status management, online monitoring and alarming of soil bioremediation parameters, and trend prediction of future soil parameters, forming a new generation of intelligent supervision system for soil bioremediation sites.

Network traffic classification plays an important role in network resource management and security. The application of encryption techniques and the rapid increase in the size of network traffic have placed higher demands on traffic classification. In this paper, we design multi-headed attention (MHA) and deep metric learning (DML) in our model for network traffic classification. In addition, MHA-DML also extracts more subtle and highly differentiated features through the improved triplet measurement loss. Experimental results demonstrate that the model achieves the best classification on all three publicly available web traffic datasets. The MHA-DML guarantees detection accuracy even when facing a classification task with many categories.

This paper proposes a multi-group end-to-end path optimization method based on genetic algorithm(MEEPOGA). Under the condition of meeting the bandwidth requirements and delay requirements of data transmission, in a network with limited link capacity and given delay, MEEPOGA arranges data transmission paths for multiple groups of source nodes to destination nodes. These paths achieve the goal of minimizing overall cost while avoiding link congestion. Considering that the genetic algorithm can provide stable and efficient search in the complex problem space, we solve the above optimization problem by making appropriate improvements to the genetic algorithm. It mainly includes modification of encoding strategy, fitness function and genetic operator. At the same time, we conducted comparative experiments with other algorithms. The optimization method proposed in this paper is mainly divided into two steps: First, MEEPOGA finds a set of possible solutions for each pair of source nodes and destination nodes under the conditions of bandwidth and delay. Then the combination evaluation is carried out through the genetic algorithm to find the optimal solution. For evaluation on a collection of paths, an objective-based penalty function is proposed. Simulation experiments show that our algorithm has good performance.

Nowadays, Wordle became almost everyone's current obsession. To study the reason for Wordle’s explosion, look for the secret behind Wordle. It is beneficial to develop a forecasting model to measure the fluctuations and distributions of the results based on time series and words. In the text used the context processing of words in text sequences in natural language processing to analogize that the same rule can be used for the composition and structure of words, so as to establish a percentage prediction model for the number of attempts of players with the character mechanism of letter position and structure in words. The error uncertainty of the model is evaluated by the MAPE error value. Through the analysis of the MAPE value, the error of the model to the predicted value is about 1.92%, so it is confident that the model can complete the prediction task with an error not exceeding 1.92%. Through this model, Predicting the result of the word "EERIE" as (2.16, 10.90 14.06, 24.49, 25.79, 14.41, 3.45).

Nowadays, in the process of program development, it is unavoidable to use third-party code repositories, but the safety of third-party code is not guaranteed. Therefore, the research of software composition analysis (SCA) is very important. In this paper, we first proposed the value dependency analysis technology for feature extraction and entity recognition of code. From the level of functions, classes, files, packages, and projects, the normalized extraction of the intermediate representation of open-source code. In view of the insufficient accuracy of the existing models based on text, identification, syntax, and other analysis technologies, the value dependency analysis technology is used for the first time to build a more accurate normalized model of control flow and data flow, effectively improving the accuracy of feature extraction and anti-aliasing.

In the existing public face datasets, the horizontal frontal and left-right rotation poses are the majority, and the models trained by them can not meet the requirements of face detection in the overlooking situation. Aiming at this phenomenon, the Tilt-angle face dataset TFD is cited and further expanded, and the Tilt-angle face dataset TFD-B is manually collected. The RetinaFace algorithm is adopted to carry out multiple face detection experiments. Typical experiment A shows that compared with WiderFace, the average detection precision of TFD+TFD-B as training set is improved by 4.81% when looking down at 15°, 9.87% when looking down at 30°, 10.56% when looking down at 45°,12.63% when looking down at 60°, and 15.62% when looking down at 75°, which indicates that TFD+TFD-B can effectively improve the precision of face detection in the overlooking situation. At the same time, the experiments carried out further show that expanding the training dataset can improve the precision of face detection. TFD+TFD-B can be obtained at https://github.com/huang1204510135/DFD.

With the development of deep learning technology, image-based object detection algorithms have been widely used in oilfield safety behavior regulation. However, the accuracy of identifying safety warning bands in oilfields is low, mainly due to the extreme aspect ratios. To solve the above problems, this paper proposes an improved YOLOv5-based method for detecting rotating targets of oilfield warning bands. By adding an additional angle prediction task to the original object detection framework and using a cyclic smooth labelling algorithm to transform the angle regression problem into a classification problem, the horizontal and predicted angle decoders can be combined to obtain the rotation bounding box of the target. This provides a more accurate spatial position representation of the warning band target, making it easier for the network to extract strong discriminative features of the target. Compared with traditional object detection algorithms annotated with horizontal bounding boxes, the rotation bounding box annotated object detection algorithm proposed in this paper significantly improves the recognition performance of safety warning bands and meets practical application requirements.

How to predict the students’ learning trend, according to the current situation of students' learning, is meaningful for the management of students' learning process. This paper presents a study early-warning system based on machine learning, using the method of machine learning with the data related to the scores of students’ professional courses, retaking percentage, graduation rate and bachelor's degree data of three classes of School of Information Engineering of Wuhan Business University, which establish a prediction model of student learning trend evaluation (whether there is a risk of not obtaining the degree). It provides reference for college students' study management and can improve the quality and efficiency of teaching management.

Spectrum is a scarce and non-renewable resource. This paper proposes an improved adaptive spectrum access mechanism for unlicensed millimetre waves, which is based on the problem of low spectrum utilisation due to the high signalling overhead of the receiver-assisted Listen Before Talk (LBT) mechanism. By introducing an adaptive duty cycle to the receiver-assisted LBT frame structure, the system parameters can be adaptively adjusted to ensure that New Radio-based Access to Unlicensed Spectrum (NR-U) and Wireless Gigabit (WiGig) The two systems can coexist in a fair and harmonious manner. In addition, the proposed algorithm further increases the total network throughput of the NR-U system by combining spectrum resource allocation and transmission power, and improves the spectrum efficiency in both licensed and unlicensed bands. Simulation results show that the proposed mechanism is significantly better than the receiver-assisted LBT ones.

In this paper, we propose an image feature extraction method based on 3D chaotic attractor, and carry out face video recognition. By adjusting the auxiliary function, the feature point set is located on a plane in the 3D coordinate system. The experiment shows that the feature point set on the plane can extract face features better and recognize faces more efficiently. This method is faster and has higher recognition rate than the method that uses trigonometric function to iteratively generate image feature point set to recognize face in video.

In meta-learning based small-sample HRRP recognition, HRRP data is one-dimensional, and the amount of extractable features is not as much as that of multidimensional data, so it is necessary to splice the one-dimensional data into twodimensional data to improve the recognition rate. This paper strives to reconceptualize the features among HRRP data from a two-dimensional perspective, and proposes a low noise sensitivity feature extraction based on wavelet decomposition and a low-frequency wavelet coefficient splicing method in descending order by scale to make it more applicable to the recognition of small sample targets containing noisy data. The HRRP with noise was decomposed by wavelet packet, and the lowest frequency wavelet coefficient with low noise sensitivity was extracted by wavelet packet sub-band energy and cosine similarity, and then spliced in descending order of scale, combined with the original data to form two-dimensional data, and trained with neural networks. The experiments show that the proposed method has obvious advantages in recognition accuracy, dependence on the number of samples and feature extraction ability.

Federated learning can effectively alleviate the data privacy problem of the participants, but the parameters or gradients passed in the model training may still leak the private data of the participants. Worse, aggregation server may return fake aggregation results. Existing solutions either use complex cryptographic primitives such as zero-knowledge proofs, or require interaction among participants, causing them high computation or communication overhead. Therefore, this paper proposes a secure and verifiable federated learning (SVFL) scheme. Specifically, SVFL performs privacy protection by introducing noise that can be offset during the aggregation process, and utilizes linear homomorphic hash to verify the correctness of the aggregation results. Compared with existing schemes, SVFL hardly loses accuracy due to the introduced security mechanism, and has low computation and communication overhead. Experimental results show that the performance of SVFL is almost consistent with the original federated learning without any protection, which makes SVFL applicable to edge devices. The computation and communication overhead of SVFL does not increase with the number of participants, which makes SVFL applicable to high-concurrency scenarios.

Since Qinghai is located in the high-altitude Qinghai-Tibet Plateau region, the geomorphological types are complex and diverse, and the distribution of ground precipitation observation stations is sparse, improving the accuracy of precipitation data is critical for studying regional ecological change over time. In the paper, we study and construct a multi-source precipitation data fusion model based on neural networks, which consists of back propagation neural network (BPNN) and long short-term memory network (LSTM). The global precipitation measurement (GPM), fifth generation ECMWF atmospheric reanalysis (ERA5), digital elevation model (DEM), and normalized difference vegetation index (NDVI) data are selected as feature data and ground observation station data as label data for model training. The results show that the fused data generated by the BP-LSTM model reduces the root mean square error to 2.48mm and the overall relative bias to 0.25% compared with the original GPM, which is better than ERA5 on data accuracy. The precipitation event capture capability is improved, which is very close to the ERA5 data with strong precipitation event capture capability, and the probability of detection, false alarm rate, and missing event rate are 0.95, 0.53, and 0.04 respectively. Finally, the regional precipitation data is generated by the fusion model with resolution of 0.01°, 1h. The model proposed in the paper incorporates topographic factors and seasonal characteristics to solve the temporal and spatial correlation of precipitation data in Qinghai Province improve the accuracy of precipitation data, and provide reliable data support for the study of regional hydro-ecological spatial and temporal variation patterns.

Combined navigation system is a navigation and positioning system composed of inertial navigation system and BeiDou satellite navigation system. Most of the navigation system models in combined navigation are nonlinear, but the traditional Kalman filtering algorithm is not well applied to nonlinear equations, and the Unscented Kalman filtering algorithm and Extended Kalman filtering algorithm which can be applied to nonlinear equations are constant in the fusion process of noise, so it will cause filtering divergence. In this paper, on the basis of Unscented Kalman filtering algorithm proposed will introduce the square root traceless Kalman filter algorithm, the algorithm through QR decomposition and Cholesk decomposition, the Sage-Husa algorithm combined with Square Root Unscented Kalman Filter algorithm, directly calculate the state error covariance matrix prediction and estimation of the square root factor, maintain the stability of the filtering, through practice proved that compared to Kalman filtering .The Nonlinear adaptive regression square root Kalman filter filter has a good navigation and positioning function, as the filtering is more convergent and the position accuracy can be within 5m, the speed error can be between 0.5m/s-1m/s. Compared with KF algorithm, the position error is increased by about 75%, and the speed error is increased by about 50%.

In the actual operation process, some of the power system bad data identification methods have the problem of low accuracy, for this reason, a deep learning-based power system bad data identification method is designed to improve this defect. The data is collected from power system users, the phase deviation caused by non-integer sampling is reduced by high sampling rate, the measurement signal period is obtained, the operational state of the distribution network is evaluated based on deep learning, the state vector is calculated, the maximum standard residual value is found, the location of the bad data is obtained, and the bad data identification method is designed. Experimental results: The mean accuracy of the power system bad data identification method in the paper is: 78.26%, which indicates that the designed power system bad data identification method performs better after fully integrating the deep learning.

In order to solve the problems of relatively backward agricultural irrigation technology, low utilization rate of water resources and serious waste, with raspberry pie as the main control chip, combined with all kinds of intelligent sensors, solenoid valve external control equipment, big data analysis and intelligent algorithm, an intelligent water-saving irrigation control system based on BP neural network is designed. The system takes raspberry pie as the main control module, according to the real-time weather forecast data of the current area obtained through WI-FI, and combined with the real-time monitoring of soil moisture by the soil moisture information collection module, the BP neural network model is used to intelligently judge whether irrigation is needed and automatically output the amount of water needed for irrigation. The test results of the system show that the predicted value of crop water demand based on BP neural network almost coincides with the expected value, indicating that the system can effectively and accurately predict crop water demand, and all kinds of modules controlled by raspberry pie can be combined organically, and the system is easy to operate, stable and reliable, and can provide optimal precision irrigation for crops. The system ensures the best growth of crops with the minimum amount of irrigation water, provides targeting and refinement for modern agriculture, effectively improves the utilization rate of water resources, has strong practicability, and has strong application value in agricultural production.

Directional modulation (DM) based on phased array (PA) can realize angle-dependent secure transmission. In this paper, frequency agile array (FAA) based DM technique is proposed to achieve range-angle-dependent secure transmission. Different from the conventional frequency diverse array (FDA), whose frequency offsets applied to the array is fixed, FAA can achieve a distortionless constellation at the target location and randomly distorted constellations at other locations by changing the frequency offsets at the symbol rate. Two frequency offset selection schemes are presented. The first scheme randomly selects the frequency offset applied to each element from a given set, and the received signal is randomly distorted both in amplitude and phase except the target location. The second scheme selects the optimal frequency offsets with lower sidelobe based on ant colony algorithm (ACO). Further, the sidelobe level is relaxed appropriately to seek multiple near optimal solutions on the basis of the optimal frequency offsets. The simulation results show that the proposed method generates higher bit error rate (BER) at non-target locations and narrower information beamwidth near the target location, which provides better secure transmission performance compared with the conventional FDA.

In this paper, a directional polarization modulation method based on a time-modulated vector array is proposed for secure communication. The polarization state of the transmitted signal is used to represent the communication information to realize the directional polarization modulation in the desired direction. The vector antenna is divided into two sub-arrays. Each sub-array is controlled by a different time sequence to turn on the antenna. The correlation between the horizontal and vertical components of the polarization signal is destroyed. The communication direction is not affected, but the desired polarization state cannot be synthesized in the non-communication direction. The simulation results show that the desired polarization state is received in the communication direction, the polarization information of the signal in the non-communication direction is disturbed, and the information leakage from the time-modulated array due to sideband radiation is avoided.

In this paper, a neural network time delay prediction method based on phase space reconstruction is presented. This method reconstructs one-dimensional chaotic time series in phase space according to the internal law through phase space reconstruction, and uses BP neural network algorithm to predict the time delay. Simulation experiments show that this method has good prediction performance.

In order to obtain high-precision DEM in the worldwide, the InSAR system of two satellite formation working in the mode of one transmitter and two receivers must maintain the interferometric phase accuracy between two satellites. The internal calibration technology is used to extract and compensate the phase of InSAR system channel to ensure the phase accuracy. Based on the composition and principle of the formation satellite InSAR system, considering the composition of the system phase error, and according to the phase synchronization design method of the phase synchronization transmission and reception links, this paper designs the internal calibration link, carries out the test, and finally gives the implementation suggestions for the internal calibration project.

In geological exploration, the texture and context information generated by different rocks on the inner wall of the drilling hole is of great significance to understand the geological condition. Drilling camera technology can obtain digital drilling video and hole wall images. For a long time, the method of manual interpretation is the main method of studying hole wall images. This method has problems such as poor real-time performance and low work efficiency .In this paper, the processing algorithm is designed based on the imaging characteristics of hole wall image and the method of digital image processing, including image preprocessing, image expansion and image stitching. First, the radius of the area is determined according to the image features, and then the concentric circle expansion method is used to obtain the rectangular expansion image. Finally, the SIFT feature point is used to complete the accurate matching with the Ransac algorithm, and use weighted average method to complete image fusion. The entire algorithm can obtain a panoramic expansion image of the hole wall area, laying the foundation for the quantitative analysis of drilling data.

As two modes of direct contact between robots and external environment, visual and tactile play a critical role in improving robot perception ability. In the real environment, it is difficult for the robot to achieve high accuracy when classifying objects only by a single mode (visual or tactile). In order to improve the classification accuracy of robots, a novel visual-tactile fusion method is proposed in this paper. Firstly, the ResNet18 is selected as the backbone network to extract visual features. To improve the accuracy of object localization and recognition in the visual network, the Position-Channel Attention Mechanism (PCAM) block is added after conv3 and conv4 of ResNet18. Then, the four-layer one-dimensional convolutional neural network is used to extract tactile features, and the extracted tactile features are fused with visual features at the feature layer. Finally, the experimental results demonstrate that compared with the existing methods, on the self-made dataset VHAC-52, the proposed method has improved the AUC and ACC by 1.60% and 1.47%, respectively.

Increasing attention has been paid to air quality monitoring with the development of the industry, and indoor air quality (IAQ) may increase the concentration of harmful substances like carbon monoxide (CO), nitrogen dioxide (NO₂) in the air, which could increase the possibility of the respiratory and cardiovascular diseases. In this paper, the portable and cost-effective monitoring device is designed to detect the concentration of the harmful gas. With this device, the realtime concentration curve is plotted, and the two-stage IIR filter is designed to decrease the potential interference. The results show that the SNR is higher than that without filer in five different gas and meanwhile concentration curve has less burr and is clearer.

Nowadays,more and more enterprises have begun to treat it as the core part of the security infrastructure and apply automation to help solve the problem of “security, cost and efficiency” difficult to balance in enterprise security operations. But the traditional network security is protected based on a concept of stacking security devices, many types of security devices have their security capabilities overlapped. This paper discusses how the atomic device control strategy can be used to standardize the management of network security devices, guide the planning of device deployment and implement automatic security emergency response on various SOAR platforms. For a certain enterprise, its internal network security devices are limited in types and the overall workload is acceptable.

Currently, tunnel shotcrete operations are mainly completed through remote control of the shotcrete machine by operators, and the process is not automated or intelligent enough. To address this issue, we propose an automatic trajectory planning algorithm for tunnel shotcrete robots. The algorithm takes the cumulative thickness model of the shotcrete and information on the tunnel surface thickness as inputs, and discretizes the shotcrete process by rasterizing the tunnel surface. The algorithm's optimization objective is the shotcrete thickness, and we propose an improved gradient descent method to optimize the shotcrete velocity while optimizing the shotcrete path. To improve the algorithm's performance in multi-round shotcrete, we introduce the shotcrete time continuation strategy and the shotcrete path drift strategy. Simulation results demonstrate that the algorithm offers a shorter solution time and guarantees the smoothness of the tunnel surface.

This paper analyzes the order printing process of large Internet trading days, uses multi-agent technology to model the micro printing system, and studies the key technologies such as allocation strategy, status monitoring and fault recovery.^[1] At the same time, the design of intelligent micro printing system with high concurrent continuous printing capability ensures the stability of the micro printing system.^[2] Then, the independent intelligent control components of a single intelligent micro printer are analyzed, and an intelligent control unit for parallel printing is proposed. The unit effectively improves the ability, reliability and reliability of parallel printing, as well as the flexibility of a single printer. Then, based on the design of a single intelligent microprinter, various services in the system are abstracted into different agents, and an intelligent microprinter cluster system is built using multi-agent technology, so that the intelligent microprinter system has high concurrent continuous printing capability under the distributed control architecture.

The stock market is usually regarded as a barometer of the economy, while the stock index can reflect the ups and downs, as well as trend changes of the stock market, to a certain extent. In recent years, the long short-term memory neural network model (LSTM model) has been widely used in the forecasting of stock prices due to its effectiveness. Nonetheless, few studies have focused on the forecasting ability of the LSTM model based on stock-index prices, with the effectiveness of this field still needing to be further explored. Against this background, this paper first constructs and designs the LSTM model of deep learning. Secondly, through the Min-Max normalization method to the data of three kinds of China A-share stock market indexes collected by Python, this paper carries out algorithm training for the LSTM model. Furthermore, based on the cleaned data, this paper conducts an empirical analysis of the price forecasting ability of the LSTM model, thus testing the accuracy of the LSTM model forecasting through the difference between the predicted and the true price curves. In closing, the paper draws relevant conclusions and puts forward targeted recommendations for improvement. Regarding research significance, the greatest contribution of this paper is to improve the stock-index price forecasting system and the research related to the defect system of the LSTM model.

Currently, DeepLab is unable to utilize multiscale feature information at multiple levels, and there are often problems such as blurred segmentation boundaries, unclear detail extraction, and incorrect segmentation.This article optimizes the DeepLabv3 plus model The backbone network has been converted to a lightweight MobileNetV2 network. In Atrous Spatial Pyramid Pooling (ASPP), stripe pooling has been used to replace global average pooling, and the original hole ratio combination of 6, 12, and 18 has been changed to 3, 7, 9, and 17. A branch with R=3 has been added, as well as the use of dense connections. The improved ASPP has the advantage of higher acceptability. The experiment shows that the average intersection ratio of the improved DeepLabv3 plus model on the dataset is 69.71%, and the average pixel accuracy is 79.45%. Compared with the original network model, the improved average intersection ratio is increased by 3.2%. Using the above improved methods has improved the performance of DeepLabv3 plus, enabling more detailed information to be obtained, and improving the resolution of the model.

With the rapid development of computer technology and Internet technology, the information age has come. The combination of computer technology and sports is one of the most popular research fields. This paper mainly completes the construction of football player number dataset and identification of football players through the jersey number. Firstly, the dataset is constructed based on player detection and number region detection. Then in the player identification task, we uses part of the backbone network of YOLOv5 model as the feature extraction module of the player identification network. Moreover, SPD-Conv module is added to improve the network recognition performance under the condition of small size target and low resolution. A series of experiments were also done to verify the performance of our proposed model. Finally, the recognition accuracy of our proposed model reached 92.75%.

Eye tracking technology can show how people focus their attention and emotionally react to their surroundings. In this study, wearable eye tracker was used to conduct eye movement experiments in realistic environment. For signal processing of the data, a finite impulse response (FIR) filter was chosen, and an eye movement data set was created. First, 26 features were chosen by a machine learning algorithm for emotion recognition, and the average rate of recognition on GDBT was 71.1%. 22 noteworthy correlation features were chosen after Spearman and emotion state were used for correlation analysis. GDBT has a recognition rate of 74.61%, while XGBoost has a recognition rate of 75.63%. The experimental results prove the validity of our data set and provide data support for the next research.