Abstract:Intelligent anti-jam communication is a new generation of anti-interference technology combined with artificial in-telligence,and the identification of interference signals is the basis of the technology.It is required to achieve better identification results with lower computational complexity in engineering applications.However,previous research has shown that they cannot bal-ance these two sides.Here,an interference signal identification algorithm based on Extreme Learning Machine(ELM)is reported.Five typical oppressive interference signals are recognized based on ELM which is based on feature extraction.The overall correct identification rate is more than 96%under the condition of 40 neurons in a single hidden layer,and it has certain generalization abil-ity,which objectively promotes the engineering application of this technology.

Abstract:In recent years,with the rapid development of new energy vehicle technology,advanced driver assistance systems put forward higher requirements for automotive braking systems.The linear control dynamic system has the advantages of fast re-sponse speed and easy control,which has become the development direction of the next generation of automotive braking systems.The One-box electronic hydraulic braking system(One-box EHB)can realize the functions of ABS and ESP system because of the integration of wheel cylinder control unit,and has great research value.However,there are not many researches in this field at home and abroad.On the basis of consulting relevant literature,the structure principle and mainstream control method of One-box EHB are introduced,and the advantages and disadvantages of related technologies are analyzed,and the development prospect of hydraulic pressure estimation and control technology of wheel cylinder is put forward.

Abstract:Contextual word embeddings can bring more semantic and grammatical information for named entity recognition(NER),but the level of information only stays between words in one sentence,lacking the usage of morpheme level information of words and the overall theme information of the article.Addressing this issue,one named entity recognition model is proposed with contextual embeddings,in which multi-level context information is carried as a part of an input.In the MPNet training process,doc-ument context word embeddings is obtained via passing a sentence with its surrounding context.The contextual word embeddings and string embeddings are combined as the input of sequence labeling module.Experimental results show that the final model achieves the state-of-the-art results on the datasets,such as CoNLL-2003,CoNLL++and OntoNotes 5.0.

Abstract:In a multi agent system with switched topology,in order to solve the problem of low and high density information weakening consistency,this paper proposes a density based adaptive consistency protocol.This protocol can effectively adapt to the high and low density transitions during the convergence process of the system,and effectively maintain the connectivity of the sys-tem,making the number of clusters generated during the convergence of the system reduce.Subsequently,the stability of the consis-tency protocol is proved by using the Lyapunov function method.Finally,through simulation experiments and comparison with sever-al consistency protocols,it is verified that the designed consistency protocol can effectively enhance consistency.

Abstract:In this paper,a machining feature recognition method based on cluster analysis is given to solve the problem of ma-chining feature recognition such as holes,slots and convex plates on box parts.Based on the topological and geometric characteris-tics of different features in the part model,this method generates the attributed adjacency graphs of the features,and obtains the cluster samples of the features by defining the feature cluster index.Finally,the cluster analysis is carried out through the cluster samples of the standard features and the features,and the features clustered into one class are regarded as the same features,so as to realize the classification and recognition of features.The experimental results show that the proposed method can effectively recog-nize different machining features in box parts.

Abstract:This paper proposes a calibration method for extrinsic parameter calibration of multiple lidar sensors with non-over-lapping field-of-view(FOV)in the shiplock environment.First,this paper establishes a point cloud map of the shiplock environ-ment,introduces an additional lidar to record the data of the shiplock environment,and uses the FAST-LIO algorithm to build a map to obtain a point cloud map.The cloud map coordinate system is converted to the artificially specified absolute coordinate sys-tem of the shiplock.The RANSAC algorithm is used to extract the plane in the point cloud map and calculate the normal vector.The plane is matched to the corresponding plane in the absolute coordinate system of the shiplock,and the Kabsch algorithm is used to calculate the initial value,and the Ceres library is used for nonlinear optimization.Finally,the lidar installed on the shiplock is cali-brated to the point cloud map,and the NDT algorithm is used for matching and calibration,and the external parameters of each li-dar are calculated.

Abstract:At present,the non-contact heart rate detection method based on face video has many problems,such as large noise interference,low accuracy and poor robustness.This paper proposes a heart rate detection method for face video based on im-proved DRSN,which can effectively solve the above problems.The improved DRSN network,on the one hand,modifies SENet in the original DRSN network to ECA-Net to effectively avoid the impact of SENet dimensionality reduction on the attention of learning channels,while maintaining the information sharing between channels,so that the channel attention mechanism can better serve the soft threshold function.On the other hand,it replaces the ReLU activation function in the original DRSN network with Leaky ReLU to reduce the occurrence of dead neurons caused by the ReLU activation function.The experimental results show that the MAE de-creases by 18.7%,RMSE decreases by 17.6%,SD decreases by 17.9%,and Pearson correlation coefficient r increases by 4.3%on the VIPL dataset.In PURE data set,MAE decreases by 9.5%,RMSE decreases by 15.2%,SD decreased by 10%,and Pearson cor-relation coefficient r increases by 0.3%.It has been verified that the improved face video heart rate detection method has higher de-tection accuracy and stronger anti-interference ability,effectively improving the accuracy and robustness of face video heart rate de-tection.

Abstract:Aiming at the problem of path selection of underwater robot fish in three-dimensional obstacle avoidance,a path planning method integrating global static and local dynamic obstacle avoidance is proposed.To obtain the global static planning route,and the speed obstacle method is used to judge the relative position of robot fish and dynamic obstacles in real time to obtain the collision area and safety range.The optimal route for robot fish to navigate safely and stably underwater is obtained.The simula-tion experiment and robot fish platform experiment results show that compared with the path planned by the traditional single algo-rithm,the fused method can obtain a better obstacle avoidance path.

Abstract:To address the problems of low recognition efficiency and poor recognition accuracy of current computer vi-sion-based irregular driving behavior analysis models,this paper proposes an improved LRCN deep learning model,which can pro-cess both temporal video and single-frame picture,by replacing the original LSTM module with the GRU module.By using the State-Farm dataset from Kaggle as the training data,the improved LRCN model with fused attention mechanism is applied to form a recog-nition method for irregular driving behaviors.The experimental results demonstrate that the improved framework enjoys advantages compared to the original LRCN model in terms of both accuracy and training speed.

Abstract:Due to the complexity of brain function connectivity and the high dimensionality expression of brain dynamic attri-butions,unsupervised clustering analysis is commonly used to analyze the time-varying network characteristics.However,the tradi-tional unsupervised clustering method is difficult to separate clusters of different sizes and densities and is highly sensitive to outli-ers.To solve this problem,a supervised clustering method is proposed to analyze the changes of brain state.The dynamic state of brain network is obtained by combining minimum intra-class distance and least-square methods.Then,the significant difference of dynamic state between mild cognitive impairment and normal control group is analyzed.The results show that patients with MCI have lower dynamic mobility and range of motion,that is,their non-stationarity is weaker and supervised clustering yields superior clas-sification accuracy to unsupervised clustering,which maximizes the state difference.

Abstract:The paper aims to solve the problems of weakened image spatial information,the low-quality local structure of the output image,and the inapplicability to the research of sketch style transfer(SST)of traditional generative adversarial network(GAN).Firstly,the key technologies adopted here are elaborated and analyzed in detail.Secondly,the convolutional neural network(CNN)with excellent feature extraction ability is introduced into GAN to construct an SST model.Finally,the validity of this model is verified by constructing data sets and designing experiments.The experimental results show that the artificial subjective score ob-tained by the SST model is 25.59%higher than other algorithms.When the index value calculation is used for objective evaluation,the overall quality of images output by the SST model on different data sets is much higher than that of other algorithms.The SST model reported here can still generate sketch-style images with high similarity to the original input images of non-frontal portrait that do not belong to the training set.Besides,each index is better than that of other algorithms.To sum up,the SST model proposed here has excellent image generation ability and broad application prospects in the study of SST.The purpose of the present work is to provide important technical support for the improvement of image SST technology,as well as a reference for the research field of face identification.

Abstract:The development and malicious use of Deepfake videos pose a huge threat to personal,social,and national securi-ty.Therefore,research on Deepfake video detection technology is of great significance.At present,there are problems such as loss of video spatiotemporal information and single feature form in the feature extraction process of detection technology at this stage.This paper proposes a multi-scale feature fusion 3D convolutional network 3D MSFFCN.By improving the multi-scale convolutional attention network and multi-frequency feature fusion module the network feature extraction ability is improved.Finally,the pro-posed algorithm is trained and tested on the data set DFDC,and the results of the evaluation index accuracy and AUC value are 87.50%and 0.967.By comparing with other models and using the Captum tool to analyze the interpretability of the proposed meth-od,it is proved that the method proposed in this paper has a good effect.

Abstract:The traditional Census transform stereo matching algorithm is easy to be affected by the gray value of the center pix-el,which leads to mismatching.In order to further improve the matching accuracy,an improved Census transform stereo matching algorithm is proposed.The δ,which describes the contrast value of local pixels,is added to the traditional Census transform as the evaluation criterion for selecting central pixels,so that the Census value of central pixels contains more information in the window,and the final matching cost function is obtained by combining the AD transform,which effectively improves the matching accuracy and robustness of the algorithm.The cost aggregation is carried out by cross domain algorithm and the optimal parallax is calculated by winner-take-all algorithm.In the parallax optimization stage,the left and right consistency check,iterative local voting and par-allax filling are used to complete the parallax optimization,and the final parallax map is obtained to further improve the accuracy of the algorithm.The proposed algorithm is used to conduct experiments on data sets on the Middleburry platform.The experimental re-sults show that the average mismatching rate of the proposed algorithm is 6.41%,which is lower than that of other improved Census transform algorithms by 3.32%and 4.49%respectively,thus effectively reducing the mismatching rate and achieving better parallax effect.

Abstract:The accuracy detection method of spatio-temporal data is usually used to solve the problem of data scheme optimi-zation or accuracy detection in the case of high-speed spatio-temporal data influx.In the spatiotemporal heterogeneous environ-ment,it is difficult to ensure the quality of detection by traditional methods when solving similar problems due to the influence of fac-tors such as data object dispersion,uncertainty of timing relationship,and spatial variability of coupling relationship.This paper combines the advantages of spatial weighting,graph convolution neural network and time convolution network,and proposes a meth-od that can solve the problem of spatio-temporal data accuracy detection.Firstly,the spatio-temporal data is modeled to generate a time-varying graph group with spatio-temporal characteristics.Secondly,the graph convolution neural network and time convolu-tion network are fused to extract and analyse the spatio-temporal characteristics of the data,and the calculation method of the weight matrix is optimized to improve the detection accuracy.Finally,in the real application scenario test,a series of comparative experiments are designed to verify that this method has obvious advantages in detection accuracy.

Abstract:Family financial planning can help families create a financial plan to take control of their finances,improve life-style and achieve goals.In the past,professional financial planners did this work,which is expensive,time consuming,and with a lot of manual intervention.Currently,automated financial planning relies on linear programming,which has poor explainability and does not conform to the probabilistic and parallel characteristics of financial action.Therefore a method of family financial planning based on probabilistic planning is proposed.This paper models the family financial planning domain and describes it using the rela-tional dynamic influence diagram language(RDDL).Secondly,the generated RDDL domain files and the RDDL problem files for each family are solved using a probabilistic planner.The concept of satisfaction is also introduced in the modelling to establish a re-ward function oriented to family satisfiability.Experiments show that the method is more explainability,better quality and more effi-cient than linear programming and commercial methods.At the same time the use of satisfiability evaluation to simulate the probabi-listic and parallel nature of the financial domain can obtain better action sequences.

Abstract:The behavior of users spreading rumors in social networks is related to their own interests.This paper introduces the penalty function and risk assessment strategy,and constructs a rumor propagation model based on the idea of evolutionary game.Considering that in the process of rumor spreading,users in social networks will learn and imitate to maximize their own interests,this paper uses risk assessment strategy to describe the process of users changing their decisions,and studies the impact of punish-ment mechanism and risk assessment strategy on rumor spreading.On this basis,numerical simulation is carried out on scale-free network and small-world network respectively.The experimental results show that both individual risk assessment and collective risk assessment can effectively reduce the spread of rumors.By increasing the penalty coefficient or reducing the risk threshold,the effect of risk assessment strategy on curbing the spread of rumors can be improved.

Abstract:Aiming at the problem of uncertainty caused by a small number of labeled samples in few-shot classification,a se-ries of methods of image operation and a method of combining spatial information and frequency domain information in training are proposed.Firstly,RGB images are converted into YCbCr images in the image preprocessing stage,and then discrete cosine trans-form(DCT)is used to generate frequency domain information to preprocess images and ensure the integrity of information.Second-ly,the feature vectors are preprocessed,and the latent space transformation algorithm is used to modify the class distribution of each class so that it tends to Gaussian distribution,and the optimal transmission algorithm(Sinkhorn)is used to estimate the opti-mal transmission to realize the initial distribution tends to Gaussian distribution and improve the accuracy of classification.Finally,the network is trained with the combination of spatial information and frequency domain information to make full use of the image in-formation to improve the training effect.Experimental results show that in the 5-way 1-shot classification task of Mini-ImageNet,CIFAR-FS and CUB 200 datasets,the classification accuracy reaches 85.24%,88.21%and 94.28%.The classification accuracy of and 5-way 5-shot classification task reaches 90.52%,91.18%and 95.88%.

Abstract:Aiming at the problems of weak generalization,image denoising and feature extraction of existing countermeasure defense methods,an attack defense model DAU-NET-GAN is proposed,which integrates improved U-NET and generated counter-measure network.By introducing channel attention mechanism and non-local mean filter in U-NET downsampling process,the in-fluence of small perturbations on the image can be reduced as much as possible,the ability of the model is improved to extract the key features of the image,and soft attention mechanism is added to the jump connection part of the generation network,so as to avoid feature redundancy and realize the reconstruction of the adversus-sample.The GCE loss function is used to replace the cross entropy loss function in the training process,which improves the robustness of the model to the opposing samples.The experimental results show that the proposed model can effectively defend against the counter samples generated by various attacks,and the de-fense success rate on MNIST and CIFAR10 data sets can reach 98.96%and 83.88%,which has good general defense effect.

Abstract:Aiming at people's demand for disease prevention and treatment,and in order to overcome the problems of medical information fragmentation,this paper constructs a medical field question answering system based on knowledge graph.Firstly,this paper completes the construction process of the medical field knowledge graph step by step from knowledge extraction,knowledge fusion and knowledge storage.In the Neo4j database,hierarchical interconnection and semantic processing capabilities are realized in the form of graphic structure,and the systematicness and relevance of knowledge are intuitively displayed.Then,a knowledge graph question and answering method is given,the Bert-BiLSTM-CRF model is used to recognize the entities of natural language questions,and the entity link algorithm is designed to generate the question candidate triples in the knowledge graph as the answer result,which realizes the retrieval and utilization of the knowledge graph of the medical field,and the constructed system can accu-rately answer the related problems of disease prevention and treatment.

Abstract:To address the drawbacks of poor solution accuracy,the tendency to fall into local optimum and slow convergence of the differential evolution algorithm,a differential evolution algorithm BetaDE based on the beta distribution is proposed.BetaDE takes advantage of the beta distribution's distribution in the 0～1 interval and uses the adaptive generation of the learning factor F in the differential evolution algorithm to determine the value of the individual adaptation to the values of the control parameters,thus balancing the diversity of the algorithm as well as the convergence performance,while avoiding falling into local optima.Simulation experiments are carried out on standard test functions,which show that BetaDE outperforms other algorithms in terms of solution ac-curacy and convergence speed.