Due to the fact that the current neural network-based traffic flow forecasting method embeds part of the manually designed features,the feature extracted by the network has single function,bad adaptability,poor robustness and inaccurate characterization of the local features.Therefore,this paper proposes a traffic flow forecasting method based on residual Long Short Term Memory(LSTM).The model uses the idea of ensemble learning to train the spatially distributed data into a residual LSTM network.In addition,dimension weighted units are introduced after each LSTM unit to present the interdependencies between different dimensions of modeling features.Experimental results show that the method can realize adaptive modeling analysis of short term traffic flow data.
In order to improve the tracking precision of the unmanned vehicle steering system to the target path,this paper proposes a new unmanned vehicle path tracking method.It introduces the lateral control scheme of unmanned vehicle based on tracking preview point,and gives the dynamic linearized data model of the system and its multi-step prediction equation.The estimation and prediction model of pseudo-gradient vectors is derived by using the method of least squares.Adopting the advantages of model-free adaptive predictive control,it can perform repeated online calculations through rolling optimization strategies,so as to obtain better dynamic performance.Based on the CarSim/Simulink co-simulation platform,the method is verified when the vehicle speed is 5 m/s and 20 m/s.Results show that compared with the MPC scheme based on vehicle dynamics,this proposed method has better tracking performance.
To address the re-identification problem of different individual vehicles with identical types,a new vehicle re-identification algorithm is proposed.According to the component detection algorithm,the window and the vehicle face region with large differences between different vehicles are obtained,and the vehicle features of the detected vehicle window and the vehicle face region are extracted and merged to generate new fusion features.The distance measurement between image features is calculated for classification and recognition.The test is carried out on the public dataset VRID-1 of Sun Yat-sen university and results show that the Rank1 matching rate of the algorithm reaches 66.67%,which is obviously better than the classical traditional feature representation algorithm,thus verifies the feasibility and validity of the algorithm.
Based on the traffic wave theory and VISSIM software simulation analysis,the critical condition of traffic flow that is appropriate for the intersection to set the Contraflow Left-turn Lanes(CLL) is studied.The critical calculation formula of setting left-turn lane is obtained by wave theory analysis and verified by example calculation and VISSIM simulation analysis.Based on the VISSIM software,the difference between the operational indicators before and after the intersection setting the left-turn lane is analyzed.Simulation results show that when the left-turn traffic flow is small that enters the CLL with unsaturated flow rate,it is not advisable to set the CLL.When the traffic condition which is appropriate to set the CLL,the setting of CLL can effectively improve the left-turn capacity and reduce the average delay,schedule time,length of queue,parking time,etc.The setting of CLL may not increase the number of the parking times,which is related to the left-turn traffic flow.
Aiming at the non-stationary characteristics of traffic trajectory big data,the preprocessing method of traffic trajectory big data is studied.According to the multi-resolution analysis characteristics of it,two-dimensional discrete wavelet,it is selected to denoise and compress the traffic trajectory big data.Building a preprocessing platform for traffic trajectory big data and combining with road traffic congestion state judgment criteria to analyze the real-time traffic situation in important sections.Analysis results show that the method can improve the speed of data processing and the analysis accuracy of congested sections.
To address the multi-class Lightill-Whitham-Richards(LWR) traffic flow model on non-uniform roads,a low-dissipation central-upwind scheme is proposed.Based on the 4th-order Central Weighted Essentially Non-Oscillatory(CWENO) reconstruction and the low dissipation central-upwind numerical flux,the dissipative characteristics of the numerical format are optimized by constructing different forms of global smoothing factors and increasing the nonlinear weights corresponding to non-smooth templates.The Runge-Kutta method discretizes the semi-discrete numerical scheme in the time direction to maintain 4th-order precision.The numerical simulation of the lane number change and traffic signal control problem of multi-class LWR traffic flow model on non-uniform roads shows that the scheme has 4th-order solving accuracy and high resolution.
In order to solve discrete memory access problem of unstructured grid in high performance computing,this paper proposes a general multi-core optimization algorithm according to the architecture features of the heterogeneous multi-core processor SW26010.This algorithm takes the Chinese supercomputer,Sunway TaihuLight,as the platform,and is based on a sorting approach.Based on the principle of mesh generation,generated non-zero elements of the sparse matrix are reordered in O(n) time.An internal mapping method is used to extend or transform the computational vectors,and the fine-grained memory access is transformed into the coarse-grained access without writing conflicts.Multi-core optimization is carried out for the flux calculation in several practical examples.Experimental results show that compared with the serial algorithm on the main core,the proposed algorithm can achieve an average acceleration of more than 10 times.
Aiming at online task allocation problem of spatiotemporal crowdsourcing,a task range adjustment algorithm DMRA and an online task allocation algorithm PAMA based on predictive analysis are proposed.The DMRA algorithm takes task location as the center and dynamically adjusts the range of tasks according to worker density.The PAMA algorithm uses Bayesian classifier to predict the distribution of the next timestamp object based on historical statistical probability.On this basis,the weighted bipartite graph optimal matching algorithm is executed to complete the task allocation.Experimental results show that the combination of DMRA algorithm and PAMA algorithm can improve the total utility of task allocation and reduce the travel cost of workers,and the performance of task allocation is better than that of greedy algorithm and random threshold algorithm.
Fuzzy C-Means(FCM) clustering algorithm can only deal with low-dimensional data and is sensitive to the initial center,without considering the interactions between class centers.For this reason,an improved method of initial center selection is designed based on the idea of conditional fuzzy clustering,replacing the Euclidean distance in the traditional FCM algorithm with the cosine distance.A wHFCLM algorithm is proposed and combined with extended incremental clustering algorithms,spFCM,oFCM and rseFCM,to generate their extended incremental fuzzy clustering algorithms,spHF(c+l)M,oHF(c+l)M and rseHF(c+l)M.Experimental results show that compared with spFCM,oFCM and rseFCM,the extended incremental fuzzy clustering algorithms is less sensitive to the selection of initial centers.It can better handle large-scale sparse high-dimensional data sets,and has better clustering performance under blocks of the appropriate size.
Internet Service Providers(ISP) implement local rerouting by deploying Downstream Path Criterion(DC).To reduce the computational overhead of DC implementation method and balance the relationship between fault protection rate and computational overhead,a DC implementation method DC-iSPF based on incremental Shortest Path First(iSPF) algorithm is proposed.The link cost from computing node to neighboring node is set to 0,and the iSPF algorithm is run on the updated topology to calculate all neighboring nodes that conform to DC.Experimental results show that compared with TBFH algorithm and DMPA algorithm,DC-iSPF method can reduce computational overhead and improve fault protection rate.
Hardware data prefetching technology can effectively improve the memory access performance of processors,but the traditional stream prefetching strategy has the problem of untimely prefetching.Therefore,a double step stream prefetching strategy is proposed,and the corresponding prefetching component structure is designed.The prefetching component automatically detects the fixed step size of the data stream and enlarges the step size to twice of the original one to calculate the prefetching address.Experimental results show that the performance of the processor can be improved by 45% and 57% respectively when SPEC2006 test set integer application and floating-point application are run with the prefetching component.For applications with high Cache Miss rate,the prefetch component can effectively hide the memory access latency.
In order to improve the sensing ability of sensor nodes to acoustic events under environmental noise interference,a fixed time window cumulative sum method is proposed based on the maximum likelihood cumulative algorithm.The influence law of quaternary positioning parameters on the positioning accuracy is analyzed.According to the principle of achieving the maximum circumradius and most appropriate reference angle, the most reasonable node combination is selected from the sensor nodes that sense the acoustic event, and the positioning result is obtained through multiple iterations. Simulation results show that compared with the existing distributed positioning method, the positioning accuracy of this method is improved by about 30%.
In order to decrease the interference between small cells in Ultra Dense Network(UDN),this paper a propose an user-centric semi-dynamic clustering method for Coordinated Multiple Point(CoMP) Joint Transmission(JT) scenarios.It divides small stations into non-overlapping clusters,and takes the small stations in the cluster and the small stations located outside the cluster but having large interference with the cluster as the optional service stations of the user.Zero-forcing precoding is used to eliminate interference to users from non-serving base stations in the optional service stations.Users select a number of candidate service stations from the optional service stations service and selected the cluster heads,with the goal of maximizing the sum of throughput of each user select the station clusters for users from the optional service stations.At the meanwhile a suboptimal method for selecting service station clusters for users is given to reduce the complexity.Simulation results show that compared with the existing scheme in the same scenario the proposed scheme improves the system throughput.
In order to solve the problem that topological communities gotten by community structure detection algorithms are not same with the functional communities,a community structure detection algorithm based on Typed-Edge Similarity Clustering(TESC) is proposed.The algorithm clusters nodes with local topological features and heterogeneous information,and constructs similarity matrices between network nodes based on node neighboring edge types to obtain heterogeneous edge information.On the basis of the similarity matrix,the nodes with large similarity are continuously merged by the idea of traditional hierarchical clustering,and then the number of communities is optimized by using the contour coefficients to obtain the final community division result.Selecting 4 real networks with known community structure and 6 artificial synthetic LFR benchmark networks,comparing with GN,Louvain algorithm of homogeneous network and Hete-SPAEM and Hetero-Attractor algorithms of heterogeneous network,the results show that the community structure obtained by the TESC algorithm are more consistent with the actual community of the network.
When Wireless Sensor Network (WSN) performs node positioning in complex mountain environments,sparse node deployment can cause positioning errors.Based on the sparseness degree of nodes,the position of an unknown node can be determined by fusing the three-dimensional Approximate Point-in-triangulation Test(APIT) algorithm and the DV-Hop algorithm.By searching neighboring nodes and connecting them,planes could be formed.Then the algorithm makes perpendiculars from the unknown node to the planes,and the average value of the coordinates of perpendicular feet is the final position of the unknown node.Experimental results show that compared with the APIT algorithm and the DV-Hop algorithm,the algorithm improves the node positioning accuracy.
In order to solve the serious traffic load and network congestion caused by the increasing data traffic to the cellular network,a mobile Data Offloading algorithm based on node Selfishness and Centrality(SCDO) in the opportunistic networks is proposed.The cellular network passe the data directly to the seed node,and the seed node uses the contact brought by the node movement to pass the data to other nodes requesting the data.If a node does not receive data when it arrives at the delay tolerance time,the node can directly download data from the cellular network.This paper compares three methods namcly,the cellular network traffic load without spilt flow,the random network scenario when randomly selecting 10% of the nodes as the seed node,and the cellular network traffic load when the SCDO relecting 10% of the seed nodes.Results show that compared with the cellular scenario and the random scenario,the time delay of SCDO is reduced by about 48% and 30% respectively,and the data traffic volume is increased by about 20% and 12% respectively.
In order to improve the energy efficiency and convergence speed of cognitive user channel and power allocation algorithms in distributed cognitive wireless networks,use the average number of bits per unit of energy as a communication efficiency indicator,and balance user communication quality and system energy consumption,this paper proposes a distributed channel and power allocation algorithm based on multi-Agent cooperative reinforcement learning.The collaborative learning is introduced on the basis of multi-agent independent Q-learning is introduced,and users share Q values and fuse after independent Q-learning.Simulation results show that the algorithm can effectively improve the convergence speed of cognitive users in transmitting power and channel allocation compared with energy efficiency-based independent Q-learning algorithm,independent Q-learning algorithm and random power allocation algorithm.
In order to simplify the access structure and ensure that all integer programming have solutions,a scheme is proposed to apply integer programming directly to the secret sharing of general access structure.By constructing integer programming,the secret is hidden in the solution to the objective function,and the constraint condition is sent to the participants as the secret share.Participants can use their shares to reconstruct the integer programming problem,and get the correct solution to the objective function quickly by the method of solving equations,and then the secret is recovered.Analysis results show that,compared with the scheme using (t,n) threshold,this scheme can achieve all the access structures.It doesn’t need to solve integer programming with the classical method or deduce the maximal unauthorized subsets,which reduce its computational complexity.
R-ate is an important bilinear mapping in the Identity-Based Cryptography(IBC) algorithm of China state cryptography standard SM9.Its computational performance is very important to the application of SM9 cryptosystem.To improve the computational efficiency of R-ate bilinear pairing,a fast computational algorithm is proposed.By analyzing the computation process of R-ate bilinear pairing on BN curves and the involved principle of inverse operations,the order in which isomorphic mapping takes effect in computation is changed,and most of the inverse operations are transferred from the large feature domain to small feature domain to reduce the computation loss of inverse element solution.The system parameters of SM9 are taken as an example to carry out experiments.Results show that the running time of the proposed algorithm is only 1.8×105 ms.
In order to improve the efficiency of intrusion detection,it is necessary to extract the features of data to reduce the data dimensions.This paper proposes a network intrusion detection method by combining Information Gain(IG) and Principal Components Analysis(PCA).The method uses IG to extract the attribute features with strong classification ability,uses PCA to reduce the dimension of the feature data,and uses Naive Bayes method for classification and detection.The test results of the data set KDDCUP99 show that the detection rate of the method is 94.5%,which is much higher than those of PCA-LDA,FPCA,and KPCA methods.
The multipath effect and time-varying characteristics of wireless signals cause large fluctuations in the measured values based on the Received Signal Strength Indication(RSSI),resulting in large errors in WLAN authentication and attack location based on RSSI location fingerprinting.Therefore,the network access authentication and attack detection location scheme of the Channel State Information (CSI) location fingerprinting are proposed.The Orthogonal Frequency Division Multiplexing(OFDM) technology is used to obtain fine-grained CSI,and the location signal characteristics are described.The K-means optimized initial clustering point algorithm is used for data processing to enhance the differences between location information,and a CSI-based location map is constructed.The CSI location fingerprinting is used to authenticate the user accessing the WLAN,and the attack detection and location are performed on the user who fails the authentication.In the communication standard test of IEEE 802.11n,the correct positioning rate of the scheme is as high as 98.12%.
In order to improve the security and reliability of military Ad Hoc network,this paper proposes a hierarchical network architecture and designs a Shortened Random-based Signcryption (SRS) scheme by using a random number instead of timestamp in Direct Key Exchange Using Timestamp Signcryption (DKEUTS) scheme.According to the security requirements,communication and computing performances of different nodes,this paper chooses the DKEUTS scheme or the SRS schemes with different signcryption parameters.By discriminating whether the network node is a cut-vertex,the network connectivity is predicted.The information of network connectivity can be used to reconstruct network topology and guarantee network reliability.Performance analysis results show that the proposed hierarchical network architecture and security protocol can be applied in node non-homogeneous networks.
In the certificateless cryptography system,there is no authentication relationship between the public key and the holder,which may cause a problem that a malicious user replaces the user’s public key.To This end,this paper improves the definition of certificateless signature,and proposes a provably security certificateless short signature scheme.The security of the scheme is based on the Inv-CDH problem,and the complete security proof is given under the stochastic oracle model,and it is proved that the scheme is anti-existence forgery in the adaptive selection message attack under the new adversary.This scheme is implemented by C language,and its performance is analyzed and compared with those of the classical short signature schemes and the certificateless short signature schemes in recent years.Results show that in the signature phase,the scheme only needs one time of multiple-point operation;in the verification phase,the scheme requires two times of multiple-point operation and two times of bilinear pairing operation.The proposed scheme has a short signature length and a high operation efficiency.
To improve the efficiency and accuracy of community detection,a random parallel local search algorithm is proposed.The complex system is represented by graph model structure,and vertices are divided into clusters.The greedy stochastic adaptive search process and path reconnection process are constructed to solve the module maximization problem of weighted graph.A {0,1} matrix class feature is introduced and the distance function of clustering is defined,so that the neighborhood search of vertices can be carried out to achieve high-precision community detection and recognition.Experimental results show that the F1 value and NMI index value of the proposed algorithm are both high.
To extract the content from a Web page accurately,an algorithm based on Support Vector Machine(SVM) and gravity radius model of DOM is proposed.Extract the node of text block from Web pages by means of SVM.Use the links information from its page and the node above to calculate the gravity radius,and utilize gravity radius model of DOM to accurately extract content again.The process of corresponding formula derivation and hyper parameters selection are presented in this paper.Experimental results show that compared with statistical extraction,FFT extraction and other algorithm,the proposed algorithm has higher accuracy and efficiency as well as better generalization ability.
Sentence alignment is a process mapping sentences in the source text to their counterparts in the target text.Within the framework of neural network,this paper proposes a sentence alignment method,on the basis that the aligned source sentence and target sentence pair contains a large number of aligned words.The Gated Relevance Network (GRN) is used to capture the semantic interaction between the source sentence and the target sentence pair,and the semantic interaction is used to determine whether the source sentence and the target sentence are aligned.The alignment evaluation of non-monotonic text shows that the F1 value of the method reaches 93.8%,which effectively improves the accuracy of sentence alignment.
The removal under-sampling method does not consider much the influence of data distribution changes on the classification results when the unbalanced data sets are classified,an improved under-sampling method based on clustering fusion and redundancy removal is proposed.The clustering algorithm is used to obtain the clustering centers of the high-density distribution regions of most samples.Most of the samples are divided into different subsets.The redundancy coefficients of each subset are calculated to de-redundantly delete most of the samples.After under-sampling the data sets with different balance rates,the cost-sensitive attribute hybrid strategy multi-decision tree prediction model is used for classification.Experimental results show that the proposed method can enable the prediction model to improve the positive rate of a few samples and the G-mean value of the prediction model under the premise of ensuring that the classification accuracy of the unbalanced data sets is not reduced.
The possible deviations between user ratings and user preferences in collaborative filtering algorithms result in reduced recommendation accuracy.For this problem,a user preference extraction algorithm based on attribution theory is proposed.User preferences are extracted based on the consensus,distinctiveness,positive and negative preference information of user behaviors.Merge preference similarity and rating similarity to get a better nearest neighbor set,and calculate a user’s predicted rating for unrated items.Experiments are carried on Movies Lens-1M dataset and the results show that under the merging condition of 10% preference similarity and 60% rating similarity,the algorithm achieves the highest recommendation accuracy,which is better than the traditional collaborative filtering algorithm and other improved algorithms,such as HU-FCF、BM/CPT-V.
The method of cross-media retrieval mostly maps the original features of two modalities to the common subspace,and performs cross-media retrieval in the subspace,ignoring the selection of discriminant features and the relationship between modalities.Therefore,a new cross-modal retrieval method based on coupled dictionary learning and graph regularization is proposed.A uniform sparse representation is generated for different modalities by associating and jointly updating dictionaries of different modalities.The sparse representations of the different modalities are then projected into the common subspace defined by the class label information to perform cross-modal matching while applying 21 norm terms to the projection matrix to select the correlation and discriminative features of the feature space.On this basis,the regularization term of the graph is used to preserve the inter-modal and intra-modal similar relationship.Experimental results show that compared with the Canonical Correlation Analysis(CCA) method,the method has higher accuracy in cross-media retrieval.
In order to make full use of the historical knowledge and improve the accuracy of rating prediction,a recommendation model based on Lifelong Machine Learning(LML) is proposed to mine both user ratings and comments.The model accumulates knowledge from previous tasks and utilizes it in future tasks to help improve the rating prediction accuracy.Experimental results on real datasets show that compared with models without LML ability,the mean square error of the predicted ratings of this model is reduced by 5.4‰,and with the accumulation of knowledge,its error is continuely dropped.The accuracy of topic word classification results is improved.