Abstract:In order to overcome the shortcoming of particle swarm optimization(PSO)that it is easy to fall into local optimal when dealing with multi-objective optimization problems,and improve the solving stability of the algorithm,a multi-strategy multi-objective particle swarm optimization algorithm(MOPSOMS)is proposed.Firstly,a phased selection strategy of global opti-mal particle based on dominant number is proposed to make the population approach the real Pareto optimal solution faster.Com-bined with adaptive grid technology and roulette strategy,the algorithm can take into account the diversity better.Secondly,the multi-objective Fibonacci mutation strategy and the dynamic population position average strategy are proposed.The mutation is car-ried out according to the Fibonacci principle in a single randomly selected dimension,and the dynamic population position average is used to replace the individual optimal particle to improve the diversity of solution and avoid the algorithm falling into the local opti-mal.The proposed MOPSOMS algorithm is compared with 8 multi-objective optimization algorithms on the ZDT test function set.The experimental results show that MOPSOMS algorithm has better solution stability,and the obtained solution set has better conver-gence and distribution.

Abstract:How to estimate the number of users in online social networks efficiently and accurately is a fundamental problem in the field of network science.The random walk sampling based estimation algorithms estimates the number of users of social net-work by randomly visiting some of the network users and counting the number of repeated visits of each user.However,the relative errors of current estimation algorithms would be quite large.In this paper,it is found that the main reason for the large relative error in the estimation results is the"false collision"caused by low degree nodes in the network.Therefore,this study proposes an error correction algorithm for collision counting based random walk sampling to improve the accuracy of the user number estimation of so-cial networks.Specifically,it corrects the collision error generated by"false collision"caused by low degree nodes.The experimen-tal verification on real social network data sets show that,compared with the traditional random walk-based sampling estimation al-gorithms,the proposed error correction algorithm can effectively reduce the relative error of the estimation.

Abstract:In order to tackle the problem of the container-based microservice scheduling in edge computing,the network delay between microservices,microservices'reliability,and the load balance of physical node clusters in edge computing are modeled to simplify the problem into a multi-objective optimization problem.On this basis,a multi-objective micro-service scheduling algo-rithm based on whale optimization is proposed.The differences in the demands of different microservices on calculate,memory,stor-age are taken as the constraints for microservices selection of edge nodes.A new individual coding method is proposed to solve the problem,and the searching ability of the optimal solution of the algorithm is accelerated by adopting a nonlinear convergence factor.A comparative experiment is conducted in seven scenarios with different user requests.The results show that the algorithm has im-proved network delay,balancing cluster load,and service stability compared with other mainstream algorithms.

Abstract:Vehicle re-identification(Re-ID)is designed to identify the same vehicle with multiple non-overlapping cameras.The existing re-identification studies mainly extract vehicle features in the high information area of the image,but ignore that there may be important information conducive to vehicle re-identification in the low activity area.To solve this problem,this paper propos-es a low active region information mining for vehicle re-identification,LRM-Net,by erasing the high active region,mining the key features in the low active region to enhance the performance of the network.LRM-Net has two branch structures,which are global branch and feature masking branch module.The global branch is used to learn the overall appearance characteristics of the vehicle.Feature masking module is used to learn the distinguishing vehicle features in the low active area.When the vehicle features are very similar,this module can distinguish vehicles effectively.Using two common data sets of VehicleID(large)and VeRi-776,the Rank-1 accuracy of LRM-Net is 90.7%and 95.9%,and the mean accuracy(mAP)is 77.1%and 79.7%,respectively.Qualitative and quantitative experiments show the effectiveness of LRM-Net in mining information in low active regions.

Abstract:The system confrontation simulation test of radar countermeasure requires models of the equipments,and then large-sample tests are carried out based on the mathematical simulation test platform to verify the combat effectiveness of equip-ment.This paper focuses on the multi-granularity modeling and model application of radar equipment,starts from the system archi-tecture of the mathematical simulation platform and test method of radar confrontation mathematical simulation,designs the struc-ture of the multi-granularity radar equipment model,and designs the detail of the signal-level and functional-level core algorithm.The simulation model architecture designed in this paper can provide a basis for the model of radar equipment and provide a refer-ence for the development of simulation platform.

Abstract:Passenger demand prediction is a crucial but challenging task for intelligent transportation system construction.In this paper,a multi-graph attention spatiotemporal prediction model is proposed.LSTM is used to model the context information of temporal dependence in travel records,and then three graphs are used to model multiple correlations of spatial regions,and GAT is used to capture the spatial dependencies between regions.In addition,weather information is integrated into the model to realize the global prediction of travel demand.Finally,a real data set is used to verify the validity of the proposed model.

Abstract:Aiming at the decision problem in which the attribute weights are unknown or partially known and the evaluation values are T-Spherical fuzzy numbers,a COCOSO decision-making method based on a(R,S)-norm T-Spherical fuzzy new fuzzy entropy measure is proposed.Firstly,the(R,S)-norm T-Spherical fuzzy entropy formula is defined,and it is proved that it satisfies the four axiomatic criteria of T-Spherical fuzzy entropy.Secondly,based on the fuzzy entropy to calculate objective attribute weight,the COCOSO method is extended to the T-spherical fuzzy environment,the T-Spherical fuzzy COCOSO decision method is pro-posed.Finally,the influence of the change of parameters R and S on the decision result is analyzed,the feasibility and effectiveness of the decision-making method are illustrated by examples and comparative analysis.

Abstract:Digital twin can monitor and analyze the status of physical entities based on virtual entities through data interac-tion,so as to achieve the purpose of combining the virtual with the real and controlling the real with the virtual.It has a very impor-tant application prospect and has become a research hotspot to promote the digitalization of multiple industries.Therefore,the paper studies the modeling and simulation of the key technology to realize the digital twin function.Firstly,the origin,definition and appli-cation status of digital twins are summarized.Then,the application of modeling technology in digital twins is classified and summa-rized comprehensively,and the characteristics of digital twin simulation and the differences between digital twin simulation and tra-ditional simulation are analyzed.Finally,on the basis of previous research,the application of modeling and simulation in production line simulation based on digital twins is described,and the development direction of modeling and simulation technology in digital twins is prospected.

Abstract:Because multidimensional time series data has high dimensions,it is difficult to achieve ideal mining results.Therefore,a multidimensional time series big data mining method based on artificial bee colony optimization algorithm is proposed.The multi-dimensional time series data is transformed into the form of pattern representation by extreme value segmentation meth-od,and then the multi-dimensional time series data is transformed into the spectrum space by Haar wavelet transform.The frequen-cy and position information of the data are changed to obtain the characteristic coefficient of the data.The feature extraction algo-rithm is used to find out the useful feature coefficients and filter out the useless ones directly.The objective function is constructed for the extracted characteristic coefficients,the local optimization of the objective function is completed based on the artificial bee colony optimization algorithm,the optimal solution is found on the basis of the fitness function,and the multi-dimensional time se-ries data mining is completed.The experimental results show that the data mining effect of the proposed method is good,which can effectively shorten the data mining time and improve the data mining accuracy.

Abstract:Aiming at the problem of insufficient performance of generalized sidelobe canceller(GSC)in low SNR environment and dynamic complex scenes,an improved GSC speech enhancement algorithm based on time-frequency mask is proposed.The al-gorithm uses the depth neural network speech enhancement system based on time-frequency mask to improve the post filtering sys-tem of GSC,uses the GSC improved by normalized minimum mean square error to filter out directional noise,and then uses pure speech time-frequency mask estimation to effectively retain speech components and suppress noise components.Through the percep-tual evaluation of speech quality(PESQ)and spectrogram analysis under different background noises,it is proved that the algo-rithm has advantages in low SNR,spectrogram is closer to pure speech,overall quality improvement of speech is good,and the pro-cessed speech is closer to the target speech.

Abstract:Spatial transcriptomics is a mature technique for analyzing histological changes in tissues with complex gene expres-sion.Identifying genes,that display spatial expression patterns,is an important first step in characterizing the spatial transcriptomic landscape of complex tissues.In this paper,the kernel function of the statistical method SPARK is modified to identify the spatial ex-pression patterns of genes in the data generated by various spatially resolved transcriptomic techniques.Two sets of simulated data are given,the SPARK2 method produces higher power than existing methods,such as low false positives and high true positives.In addition,by analyzing three published spatially resolved transcriptomic datasets,the SPARK2 approach is found to be more power-ful than existing methods,such as identifying more SE(spatial expression)genes and revealing more biological findings.Therefore,improving the kernel function can increase the number of genes identified with spatial expression patterns,while also identifying genes of biological significance.

Abstract:Edge detection is a basic problem of computer vision and image processing,and its detection effect is inseparable from the quality of the image.However,in the process of image signal acquisition,it is inevitable that noise and other influencing factors will be mixed.Canny's algorithm has attracted a lot of attention because of its good accuracy in smoothing images and remov-ing noise.However,in the traditional Canny edge detection method,Gaussian filters are mostly based on the same direction,and the utilization of information in the edge direction is low.In order to enhance the utilization of edge information and the smoothing ef-fect of Gaussian filtering,anisotropic Gaussian filters are usually used to improve the filtering effect,but at the same time,the com-plexity of their calculations increases.To solve the above problems,an improved method for image denoising based on recursive Gaussian filter is proposed.In this method,the two-dimensional filtering of the anisotropic Gaussian filter is decomposed into two one-dimensional filters,and a forward filter is performed first,on the basis of which the backward filtering is performed,and the re-sults of each filter are updated iteratively.Experimental results show that the proposed recursive Gaussian filter not only reduces the complexity of the original algorithm,but also significantly improves the edge detection effect applied to the Canny algorithm.

Abstract:Aiming at the problems of limited time-frequency resource constraints and high-concurrency in the cloud call cen-ter,an automatic scheduling algorithm for communication lines based on deep reinforcement learning is proposed.Firstly,a homoge-neous multi-agent system is modeled by using a Markov decision process,and a composite reward function is designed.Subsequent-ly,from the perspective of sequential decision-making,a multiple self-attention mechanism is designed,and three algorithms based on the Actor-Critic architecture are used to train the model.The experimental results show that,the model can automatically output the joint scheduling strategy,which has excellent stability and effectiveness.

Abstract:Image super resolution reconstruction technology can improve image identification ability and accuracy,with the continuous development of deep learning,because of its excellent fitting ability,the generated adversarial network shows a good po-tential in the field of image super resolution reconstruction rate.The SRGAN algorithm network model is improved,a light-weight network model based on attention mechanism is proposed.The spatial channel attention mechanism is introduced to improve the fea-ture extraction ability of the generator module,so that the network can obtain the high frequency feature information of the image adaptively,and further improve the performance of the generated network.At the same time,the number of network layers of the dis-criminator module is compressed,and more efficient convolution method and activation function are adopted to accelerate the con-vergence of the model and improve the quality of the reconstructed image.The effectiveness of the improved method is verified by the test results of the public data set,the improved model parameters are greatly reduced,the texture details are restored more clear-ly,and the image visual effect is better.

Abstract:Plastic bottles can produce defective products at the mouth of the bottle during the actual production process,such as worn,broken,dirty,etc.Such defective bottle tops cannot be sealed,which greatly affects the sealing of the bottle and thus en-dangers the health of the user.Machine vision-based methods can efficiently and automatically achieve defect detection,and edge detection is a very important part of bottle top defect detection.To this end,an improved Holistically-nested Edge Detection(HED)method is proposed to be applied to the edge detection of plastic bottle tops.In this paper,different edge detection algorithms are used for bottle mouth defect detection,and the performance is evaluated based on the subjective evaluation index F-measure value and the PR curve is plotted.The results show that the detection effect value of the improved HED network has a significant improve-ment over the original HED network.The paper then performs edge detection on the test images and finds that the improved HED network extracts edges with better completeness than the traditional edge algorithm and can meet the subsequent defect detection re-quirements.

Abstract:In order to solve the problems of memory and classification accuracy of classical convolution neural network in im-age classification,a double observation image classification algorithm model based on quantum convolution neural network is pro-posed.Firstly,the average pool down sampling strategy is used to reduce the dimension of each image,and a highly expressive strong entanglement parameterized quantum circuit is designed to replace the traditional convolution layer to extract the key features of the input image information.In addition,the double observation measurement strategy is used to obtain sufficient hidden informa-tion from the quantum system.The experimental results show that the proposed algorithm is superior to the classical convolution neu-ral network and quantum neural network,and performs better on MNIST multi classification dataset,especially in the data subset{0,1},with an accuracy rate of 100%.

Abstract:Considering the complexity and exploratory nature of scientific research in the professional field of ships,research-ers cannot screen out the suitable APP quickly for the current research objectives from various simulation computing software(APP)in the field of ship overall performance prediction,evaluation and optimization.An intelligent recommendation model of APP for ship overall performance based on deep learning is proposed.Multi-dimensional feature vectors of researchers'data and ship overall performance APP are constructed,personnel features and APP features are extracted through deep neural network,and the cosine similarity between features is calculated.And the fusion information such as APP score,APP computing task statistics and APP ac-cess statistics is used as the similarity measure standard to train the model and generate the recommendation model.The results show that the recommendation model can actively recommend accurate apps for researchers,help them find appropriate ship perfor-mance apps,and realize personalized intelligent recommendation of apps quickly.Compared with the other two traditional methods,this model performs well in terms of accuracy,error,and F1 score.

Abstract:This paper addresses the problem that the problem that traditional defogging methods do not recover the colour of the image to a high degree and can easily lead to artefacts and block effects in the transition edge areas and depth of field areas.In this paper,an image fusion defogging algorithm combining colour space and fog line dark channel prior is proposed.Firstly,based on the rough transmittance values estimated by the dak channel method,the transmittance valuations of several fog lines are correct-ed and adaptively adjusted to obtain a defogged image with low saturation and dark colours using the improved atmospheric light val-ues and transmittance valuations.Secondly,the original fogged image is converted to HSI colour space and the S and I components are enhanced using the MSRCR algorithm to obtain a second image with high saturation and high luminance.Finally,the two images are linearly fused by using weighted averaging to obtain the recovered image.The experimental results show that the proposed new method can effectively avoid the halo effect and colour distortion while better preserving the colour characteristics and detail informa-tion of the original fogged images meanwhile,the mean values of peak signal-to-noise ratio,structural similarity and information en-tropy of the method in five groups of images are 16.9699,0.8381 and 7.5203,respectively,which are all better than the other four comparison algorithms and have certain feasibility.

Abstract:Large language models(LLMs),due to their large number of parameters and the need to handle long contextual in-formation,face significant challenges in deployment and optimization in edge-constrained computing environments.By summariz-ing the main optimization methods in areas such as shared storage space,activation function optimization,attention mechanism opti-mization,relative position encoding optimization,automatic mixed-precision training,and quantization techniques.It analyzes the practical results of large model optimization on the ROCm open-source framework and HIP programming model,and their impact on model performance.By summarizing the characteristics of these technologies and exploring them within the open-source frame-work,it provides new insights for significantly improving computational and storage efficiency,as well as the deployment of domes-tic applications,in edge computing environments.

Abstract:An output feedback controller based on the robust integral of the sgn of the error(RISE)method is proposed for the three-phase inverter with matched and unmatched disturbances,where the needs of performance optimization and cost reduction are taken into account.Firstly,the matched and unmatched disturbances are converted into equivalent lumped disturbances by es-tablishing the system error dynamic model,on this basis,the controller is designed.Secondly,a robust state observer with anti-dis-turbance performance is constructed to estimate the measured state signal.Thirdly,a feedforward compensation term is integrated to improve the voltage tracking precision and an integral robust term is employed to handle lumped disturbances.Finally,the asymptot-ic convergence of observational error and tracking error are derived according to Lyapunov stability analysis,furtherly,the feasibili-ty and effectiveness of the proposed strategy are demonstrated by simulation results.