To address the issues of missing distribution information and distribution collapse encountered by deep variational text clustering models in practical applications， a Deep Variational text Clustering Model based on Distribution augmentation （DVCMD） was proposed. In this model， the enhanced latent semantic distributions were integrated into the original latent semantic distribution by enhancing distribution information， so as to improve information completeness and accuracy of the latent distribution. At the same time， a distribution consistency constraint strategy was employed to promote the learning of consistent semantic representations by the model， thereby enhancing the model’s ability to express true information of the data through learned semantic distributions， and thus improving clustering performance. Experimental results show that compared with existing deep clustering models and structural semantic-enhanced clustering models， DVCMD has the Normalized Mutual Information （NMI） metric improved by at least 0.16， 9.01， 2.30， and 2.72 percentage points on the four real-world datasets： Abstract， BBC， Reuters-10k， and BBCSports， respectively， validating the effectiveness of the model.

Aiming at the problem of low data availability caused by existing disturbance mechanisms that do not consider the semantic relationship of location points， a Trajectory Location Privacy protection Mechanism based on Differential Privacy was proposed， namely DP-TLPM. Firstly， the sliding windows were used to extract trajectory dwell points to generate the fuzzy regions， and the regions were sampled using exponential and Laplacian mechanisms. Secondly， a road network matching algorithm was proposed to eliminate possible semantic free location points in the sampled points， and the trajectory was segmented and iteratively matched by using Error Ellipse Matching （EEM）. Finally， a disturbance trajectory was formed based on the matched location points， which was sent to the server by the user. The mechanism was evaluated comprehensively by confusion quality and Root Mean Square Error （RMSE）. Compared with the GeoInd algorithm， the data quality loss of the DP-TLPM is reduced by 24% and the confusion quality of the trajectories is improved by 52%， verifying the effectiveness of DP-TLPM in terms of both privacy protection strength and data quality.

Aiming at the problem of noise arising from user’s unexpected interactions in practical recommendation scenarios and the challenge of capturing short-term demand biases due to the dispersed attention in self-attention mechanism， a model namely FTARec （sequential Recommendation based on hierarchical Filter and Temporal convolution enhanced self-Attention network） was proposed. Firstly， hierarchical filter was used to filter noise in the original data. Then， user embeddings were obtained by combining temporal convolution enhanced self-attention networks with decoupled hybrid location encoding. The deficiencies in modeling short-term dependencies among items were supplemented by enhancing the self-attention network with temporal convolution in this process. Finally， contrastive learning was incorporated to refine user embeddings and predictions were made based on the final user embeddings. Compared to existing sequential recommendation models such as the Self-Attentive Sequential Recommendation （SASRec） and the Filter-enhanced Multi-Layer Perceptron approach for sequential Recommendation （FMLP-Rec）， FTARec achieves higher Hit Rate （HR） and Normalized Discounted Cumulative Gain （NDCG） on three publicly available datasets： Beauty， Clothing， and Sports. Compared with the suboptimal DuoRec， FTARec has the HR@10 increased by 7.91%， 13.27%， 12.84%， and the NDCG@10 increased by 5.52%， 8.33%， 9.88%， respectively， verifying the effectiveness of the proposed model.

With the rapid development of cloud computing technology， the number of data centers have increased significantly， and the subsequent energy consumption problem gradually become one of the research hotspots. Aiming at the problem of server energy consumption optimization， a data center server energy consumption optimization combining eXtreme Gradient Boosting （XGBoost） and Multi-Gated Recurrent Unit （Multi-GRU） （ECOXG） algorithm was proposed. Firstly， the data such as resource occupation information and energy consumption of each component of the servers were collected by the Linux terminal monitoring commands and power consumption meters， and the data were preprocessed to obtain the resource utilization rates. Secondly， the resource utilization rates were constructed in series into a time series in vector form， which was used to train the Multi-GRU load prediction model， and the simulated frequency reduction was performed to the servers according to the prediction results to obtain the load data after frequency reduction. Thirdly， the resource utilization rates of the servers were combined with the energy consumption data at the same time to train the XGBoost energy consumption prediction model. Finally， the load data after frequency reduction were input into the trained XGBoost model， and the energy consumption of the servers after frequency reduction was predicted. Experiments on the actual resource utilization data of 6 physical servers showed that ECOXG algorithm had a Root Mean Square Error （RMSE） reduced by 50.9%， 31.0%， 32.7%， 22.9% compared with Convolutional Neural Network （CNN）， Long Short-Term Memory （LSTM） network， CNN-GRU and CNN-LSTM models， respectively. Meanwhile， compared with LSTM， CNN-GRU and CNN-LSTM models， ECOXG algorithm saved 43.2%， 47.1%， 59.9% training time， respectively. Experimental results show that ECOXG algorithm can provide a theoretical basis for the prediction and optimization of server energy consumption optimization， and it is significantly better than the comparison algorithms in accuracy and operating efficiency. In addition， the power consumption of the server after the simulated frequency reduction is significantly lower than the real power consumption， and the effect of reducing energy consumption is outstanding when the utilization rates of the servers are low.

In Chinese-Vietnamese cross-language news topic discovery task, the Chinese-Vietnamese parallel corpora are rare, it is difficult to train high-quality bilingual word embedding, and the news text is generally long, so that the method of bilingual word embedding is difficult to represent the text well. In order to solve the problems, a Chinese-Vietnamese news topic discovery method based on Cross-Language Neural Topic Model (CL-NTM) was proposed. In the method, the news topic information was used to represent news text, and the bilingual semantic alignment was converted into bilingual topic alignment tasks. Firstly, the neural topic models based on the variational autoencoder were trained in Chinese and Vietnamese respectively to obtain the monolingual abstract representations of the topics. Then, a small-scale parallel corpus was used to map the bilingual topics into the same semantic space. Finally, the K-means method was used to cluster the bilingual topic representations for finding the topics of news event clusters. Experimental results show that, compared with the Improved Chinese-English Latent Dirichlet Allocation model (ICE-LDA), the proposed method increases the Macro-F1 value and topic-coherence by 4 percentage points and 7 percentage points respectively, showing that the proposed method can effectively improve the clustering effect and topic interpretability of news topics.

For the intelligent dialogue system of customer service robots in power supply business halls, a large-scale dataset of power business user intents was constructed. The dataset includes 9 577 user queries and their labeling categories. First, the real voice data collected from the power supply business halls were cleaned, processed and filtered. In order to enable the data to drive the study of deep learning models related to intent classification, the data were labeled and augmented with high quality by the professionals according to the background knowledge of power business. In the labeling process, 35 types of service category labels were defined according to power business. In order to test the practicability and effectiveness of the proposed dataset, several classical models of intent classification were used for experiments, and the obtained intent classification models were put in the dialogue system. The classical Text classification model-Recurrent Convolutional Neural Network (Text-RCNN) was able to achieve 87.1% accuracy on this dataset. Experimental results show that the proposed dataset can effectively drive the research on power business related dialogue systems and improve user satisfaction.

Concerning the problems of large number of parameters, poor real-time performance and low accuracy of traffic sign recognition algorithms based on deep learning, an improved traffic sign recognition algorithm based on YOLO v3 was proposed. First, the depthwise separable convolution was introduced into the feature extraction layer of YOLO v3, as a result, the convolution process was decomposed into depthwise convolution and pointwise convolution to separate intra-channel convolution and inter-channel convolution, thus greatly reducing the number of parameters and the calculation of the algorithm while ensuring a high accuracy. Second, the Mean Square Error (MSE) loss was replaced by the GIoU (Generalized Intersection over Union) loss, which quantified the evaluation criteria as a loss. As a result, the problems of MSE loss such as optimization inconsistency and scale sensitivity were solved. At the same time, the Focal loss was also added to the loss function to solve the problem of severe imbalance between positive and negative samples. By reducing the weight of simple background classes, the new algorithm was more likely to focus on detecting foreground classes. The results of applying the new algorithm to the traffic sign recognition task show that, on the TT100K (Tsinghua-Tencent 100K) dataset, the mean Average Precision (mAP) of the algorithm reaches 89%, which is 6.6 percentage points higher than that of the YOLO v3 algorithm; the number of parameters is only about 1/5 of the original YOLO v3 algorithm, and the Frames Per Second (FPS) is 60% higher than YOLO v3 algorithm. The proposed algorithm improves detection speed and accuracy while reducing the number of model parameters and calculation.

Dynamic geometry software is widely applied to geometric constraint constructions because it is dynamic and intuitive. Aiming at a problem that the data structures in the field of dynamic geometry lack reusable abstract descriptions,a design method of the dynamic geometric software domain model was proposed. Firstly,the basic context boundaries were identified and outlined by means of domain analysis. Then,a dynamic geometry software core domain model was designed through the domain model. Finally,the dynamic geometry software was decoupled in both vertical and horizontal dimensions during the architecture modeling process. Experimental results show that the dynamic geometry software developed by using the design method of the proposed domain model can correctly deal with the graphic degradation situation at a critical position. The domain knowledge expressed by the model is applicable to 2D and 3D dynamic geometry software at the same time,and can design the layout and interaction for different devices respectively,thus a high-level reuse of the domain knowledge is achieved.

As the existing dynamic programming algorithm cannot quickly solve Discounted {0-1} Knapsack Problem (D{0-1}KP), based on the idea of dynamic programming and combined with New Greedy Repair Optimization Algorithm (NGROA) and core algorithm, a Greedy Core Acceleration Dynamic Programming (GCADP) algorithm was proposed with the acceleration of the problem solving by reducing the problem scale. Firstly, the incomplete item was obtained based on the greedy solution of the problem by NGROA. Then, the radius and range of fuzzy core interval were found by calculation. Finally, Basic Dynamic Programming (BDP) algorithm was used to solve the items in the fuzzy core interval and the items in the same item set. The experimental results show that GCADP algorithm is suitable for solving D{0-1}KP. Meanwhile, the average solution speed of GCADP improves by 76.24% and 75.07% respectively compared with that of BDP algorithm and FirEGA (First Elitist reservation strategy Genetic Algorithm).

Aiming at difficult multi-target identification recognition and low localization accuracy in mobile cellular networks, a multi-objective automatic identification and localization method was presented based on cellular network structure to improve the detection efficiency of target number and the localization accuracy of each target. Firstly, multi-target existence was detected through the analysis of the result variance of multiple positioning in the monitoring area. Secondly, cluster analysis on locating points was conducted by k-means unsupervised learning in this study. As it is difficult to find an optimal cluster number for k-means algorithm, a k-value fission algorithm based on beam resolution was proposed to determine the k value, and then the cluster centers were determined. Finally, to enhance the signal-to-noise ratio of received signals, the beam directions were determined according to cluster centers. Then, each target was respectively positioned by Time Difference Of Arrival (TDOA) algorithm by the different beam direction signals received by the linear constrained narrow-band beam former. The simulation results show that, compared to other TDOA and Probability Hypothesis Density (PHD) filter algorithms in recent references, the presented multi-objective automatic identification and localization method for solving multi-target localization problems can improve the signal-to-noise ratio of the received signals by about 10 dB, the Cramer-Mero lower bound of the delay estimation error can be reduced by 67%, and the relative accuracy of the positioning accuracy can be increased more than 10 percentage points. Meanwhile, the proposed algorithm is simple and effective, is relatively independent in each positioning, has a linear time complexity, and is relatively stable.

Focusing on the issue of speech recognition in multi-noise environment, a new hierarchical speech recognition model considering environmental noise as the context of speech recognition was proposed. The proposed model was composed of two layers of noisy speech classification model and acoustic model under specific noise environment. The difference between training data and test data was reduced by noisy speech classification model, which eliminated the limitation of noise stability required in feature space research and solved the disadvantage of low recognition rate caused by traditional multi-type training under certain noise environment. Furthermore, a Deep Neural Network (DNN) was used for modeling of acoustic model, which could further enhance the ability of acoustic model to distinguish noise and speech, and the noise robustness of speech recognition in model space was improved. In the experiment, the proposed model was compared with the benchmark model obtained by multi-type training. The experimental results show that, the proposed hierarchical speech recognition model has relatively reduced the Word Error Rate (WER) by 20.3% compared with the traditional benchmark model. The proposed hierarchical speech recognition model is helpful to enhance the noise robustness of speech recognition.

Focusing on that the existing attribute reduction algorithms are not suitable for dealing with the incomplete data with both numerical attributes and symbolic attributes, an extented incomplete neighborhood rough set model was proposed. Firstly, the distance between the missing attribute values was defined to deal with incomplete data with mixed attributes by considering the probability distribution of the attribute values. Secondly, the concept of neighborhood mixed entropy was defined to evaluate the quality of attribute reduction and the relevant property theorem was proved. An attribute reduction algorithm for incomplete neighborhood rough set based on neighborhood mixed entropy was constructed. Finally, seven sets of data were selected from the UCI dataset for experimentation, and the algorithms was compared with the Attribute Reduction of Dependency (ARD), the Attribute Reduction of neighborhood Conditional Entropy (ARCE) and the Attribute Reduction of Neighborhood Combination Measure (ARNCM) algorithm respectively. The theoretical analysis and the experimental results show that compared to ARD, ARCE, ARNCM algorithms, the proposed algorithm reduces the attributes by about 1, 7, 0 respectively, and improves the classification accuracy by about 2.5 percentage points, 2.1 percentage points, 0.8 percentage points respectively. The proposed algorithm not only has less reducted attributes, but also has higher classification accuracy.

In the era of big data, the social security data becomes more diverse and its amount increases rapidly, which challenges the analysis and decision of social security events significantly. How to accurately categorize the attack behavior in a short time and support the analysis and decision making of social security events becomes an urgent problem needed to be solved in the field of national and cyberspace security. Aiming at the behavior of aggression in social security events, a new Distributed Neural Network Classification (DNNC) algorithm was proposed based on the Spark platform. The DNNC algorithm was used to analyze the related features of the attack behavior categories, and the features were used as the input of the neural network. Then the function relationship between the individual features and attack categories were established, and a neural network classification model was generated to classify the attack categories of social security events. Experimental results on the data provided by the global terrorism database show that the proposed algorithm can improve the average accuracy by 15.90 percentage points compared with the decision tree classification, and by 8.60 percentage points compared with the ensemble decision tree classification, only decreases the accuracy on part attack type.

An improved Particle Swarm Optimization (PSO) algorithm based on Hamming distance was proposed to solve the discrete problems. The basic idea and process of traditional PSO was retained, and a new speed representation based on Hamming distance was defined. Meanwhile, in order to make the algorithm be more efficient and avoid the iterative process falling into the local optimum, new operators named 2-opt and 3-opt were designed, and the random greedy rule was also used to improve the quality of the solution and speed up the convergence. At the later period of the algorithm, in order to increase the global search ability of the particles in the whole solution space, a part of particles was regenerated to re-explore the solution space. Finally, a number of TSP standard examples were used to verify the effectiveness of the proposed algorithm. The experimental results show that the proposed algorithm can find the historical optimal solution for small scale TSP; for large-scale TSP, for example, the city number is more than 100, satisfactory solutions can also be found, and the deviations between the known and the optimal solutions are small, usually within 5%.

Concerning the construction problem of attribute subset sequence in multi-granulation rough set model, a construction method based on the distance between attributes was proposed. Firstly, the concept of the distance between attributes in information system was introduced. Secondly, the quantitative calculation formula was given, which was then used to compute the distance between the attributes. Finally, according to the distance between the attributes, the neighborhood attribute set of each attribute was obtained, and then the attribute subset sequence was constructed. The experimental results show that the proposed method is more accurate for each object class of the experiment than the random constructional attribute subset sequence.

Since Scale Invariant Local Ternary Pattern (SILTP) background modeling algorithm is of high complexity and slow computing speed, which is not suitable for real-time video processing, a new method named Uniform Scale Invariant Local Ternary Pattern (USILTP) background modeling algorithm was proposed. Firstly, the feature of USILTP was extracted by regulating the frequency of SILTP coding jump in order to reduce the feature dimension of SILTP. Secondly, a USILTP background modeling parallel algorithm based on Intel core graphics (Intel HD) and Open Computing Language technology (OpenCL) was designed and implemented to further accelerate USILTP background modeling algorithm. Finally, the foreground result of USILTP background modeling algorithm was optimized by combing multiple color channel models. The experimental result shows that the proposed algorithm can be applied to process 320×240 resolution video at a rate of 98 frame/s on the Intel HD 4600, which is 4 times faster than that of SILTP background modeling algorithm. In terms of foreground detection, the performance of the proposed algorithm is improved by 2.1% compared with SILTP background modeling algorithm on the public dataset.

Aiming at the problem of low efficiency of tampering detection and accuracy of location, a homologous video copy-move tampering detection and recovering method based on Geometric Mean Decomposition (GMD) and Structural SIMilarity (SSIM) was proposed. Firstly, the videos were translated into grayscale image sequences. Then, the geometric mean decomposition was adopted as a feature and a block-based search strategy was put forward to locate the starting frame of the duplicated sequences. In addition, SSIM was first extended to measure the similarity between two frames of a video. The starting frame of duplicated sequences was rechecked by using the structural similarity. Since the value of similarity between duplicated frames is higher than that between the normal inter-frames, a coarse-to-fine method based on SSIM was put forward to locate the tail frame. Finally, the video was recovered. In comparison with other classical algorithms, the experimental results show that the proposed method can not only achieve detection of copy-move forgery but also accurately detect and localize duplicated clips in different kinds of videos. Besides, the method has a great improvement in terms of precision, recall and computation time.

Due to the downward closure property over uncertain data, existing solutions of mining all the frequent itemsets may lead an exponential number of results. In order to obtain a reasonable result set with small size, frequent closed itemsets discovering over uncertain data were studied, and a new algorithm called Normal Approximation-based Probabilistic Frequent Closed Itemsets Mining (NA-PFCIM) was proposed. The new method regarded the itemset mining process as a probability distribution function, and mined frequent itemsets by using the normal distribution model which supports large databases and can extract frequent itemsets with a high degree of accuracy. Then, the algorithm adopted the depth-first search strategy to obtain all probabilistic frequent closed itemsets, so as to reduce the search space and avoid redundant computation. Two probabilistic pruning techniques including superset pruning and subset pruning were also used in this method. Finally, the effectiveness and efficiency of the proposed methods were verified by comparing with the Possion distribution based algorithm called A-PFCIM. The experimental results show that NA-PFCIM can decrease the number of extending itemsets and reduce the complexity of calculation, it has better performance than the compared algorithm.

After analyzing and comparing the existing security situation assessment methods, a network security situation assessment method was proposed based on time dimension, which focused on the necessity of using different methods for short-term and long-term assessment respectively. Based on the alarm information which came from security device such as firewall and Intrusion Detection Systems (IDS), the whole short-term situation was got according to the score of destination host. Combining the result of short-term assessment and static index, the weight of long-term assessment system was determined by entropy method. The proposed assessment method divides network security situation into short-term and long-term, and makes up for the lack of setting situation assessment boundaries in terms.

Some Chinese time expressions consisting of "cardinal+time unit word" may be time point expressions or time slot expressions in different context. An approach of classification of Chinese time expressions based on dependency parsing was proposed, for the purpose of automatic classification of Chinese time expressions. First some syntactic constraints of Chinese time expressions in sentences were found with the help of dependency parsing. Then some computable dependency rules were extracted from those syntactic constraints. Finally the classification of Chinese time expressions was executed using dependency rules. The experimental results show that in this approach the precision, recall, F-Measure of the confirmation are 82.3%, 88.1%, 85.1%; and the precision, recall, F-Measure of the classification are 77.1%, 82.5%, 79.7%.

To solve the problem that the traditional Multi-Input-Multi-Output (MIMO) Support Vector Machine (SVM) generally ignore the dependency among all outputs, a new MIMO SVM algorithm based on principal curve was proposed in this paper. Following the assumption that the model parameters of all outputs locate on a manifold, this paper firstly constructed a manifold regularization based on the Multi-dimensional Support Vector Regression (M-SVR), where the regularizer was the squared distance from the output parameters to the principal curve through the middle of all parameters' set. Secondly, considering the non-convexity of this regularization, this paper introduced an alternative optimization method to calculate the model parameters and principal curve in turn until convergence. The experiments on simulated data and real-life dynamic load identification data were conducted, and the results show that the proposed algorithm performs better than M-SVR and SVM based separate modeling method in terms of prediction precision and numerical stability.

Concerning the Virtual Machine (VM) migration problems in Wide Area Network (WAN), a live migration model was proposed. The link state between nodes was continuously detected, and the migration time of disk, memory, CPU status and network were optimized. The disk cycle synchronization, unidirectional tunnel and virtual machine localization were implemented. The experimental results show that the model reduces amount of migration data and shortens redirection path in WAN. The total time and pause time are close to the manner of shared storage under simulated conditions.

A management model of virtual machines was proposed based on Peer-to-Peer (P2P) structure and its prototype system was implemented. Host nodes were organized in P2P structure and resource discovery was achieved by multicast. Live migration algorithm of virtual machines was proposed and live migrations between nodes were automatically triggered. The requests of cloud computing user were mapped to the host by elected root node, and the on-demand operations of creating, deleting, and stopping a virtual machine were achieved. The experimental results show that: the model has the features of rapid convergence, low bandwidth utilization and high availability. Load balance of cloud computing resources can be achieved.

This paper analysed the functions of all parts of the USB transfer subsystem in real-time human motion capture device, and introduced its detailed designing methods of this subsystem. The USB transfer subsystem can transfer exact data in real time while the device is running and collecting human motion data.