Abstract:Accurate high-resolution Marine environment data, especially in the area of low-density sea surface survey devices, is crucial for the planning and management of fishery resources, and can improve the accuracy of fishing situation simulation and prediction. In this study, two classical methods, optimal interpolation (IO) and successive correction (SCM), were used to combine Marine environment data from satellite remote sensing and ocean buoys. Four different algorithms were used to evaluate the deviation correction in remote sensing Marine environment data :(1) mean deviation correction, (2) regression equation, (3) distribution transformation, and (4) spatial transformation. The Marine environment data is provided by NMSDC (national Marine science data center), covering the south China sea. The sea surface temperature data collected from January 2009 to December 2018 were statistically analyzed, and the performance of the two data merging techniques was visually examined and qualitatively compared.

Abstract:A unified system model combining target detection and parameter estimation is established by introducing a target existence state variable. The strict definition of radar information， detection information and spatial information of target is given， and it is proved that the radar information is sum of the detection information and the spatial information of known of target states， which theoretically solves the quantitative problem of the detection information and detection information. We derive the theoretical formula of detection information under the condition of target matching and non-matching， propose a random target detection method， and prove the target detection theorem. The target detection theorem shows that the detection information is achievable， otherwise the empirical detection information of any detector is not greater than the detection information. The target detection information theory proposed in this paper breaks the dominance of the Neyman-Pearson criterion and opens up a new direction for the system theory and design method of target detection.

Abstract:Speech spoofing refers to the technology of counterfeiting an illegal speech without the authentication by automatic speaker verification （ASV） system to the speech of a legally authenticated speaker by ASV through recording， text-to-speech， voice conversion and other means， so as to achieve the goal of passing the ASV system. With the development of artificial intelligence and speech anti-spoofing methods， ASV systems have encountered severe challenges in security. It is a hot topic in the field of speech research in recent years to detect the authenticity of the speech input to the ASV system and to prevent spoof speech from passing the verification of ASV to improve the security of the ASV system. The latest research of scholars at home and abroad explores the influence of different speech spoofing methods on ASV system from the perspective of acoustic feature and recognition model， and further studies the corresponding speech anti-spoofing technology， which improves the anti-spoofing ability of ASV systems to a certain extent. This paper summarizes the latest methods of speech spoofing to ASV systems and the latest anti-spoofing methods， focusing on the state-of-the-art research results around the world， and prospects the development direction of speech anti-spoofing technology.

Abstract:It is essential for the planning and management of fishery resources that accurate and high-resolution marine environmental data in the sea area of low-density sea surface measurement devices. Hence， the sea environmental data measured by satellite remote sensing and marine buoys are merged to improve the resolution of sea environmental data at a large scale by the successive correction method （SCM）. The South China Sea surface temperature （SST） data collected from January 2009 to December 2010 are taken as a sample. The SCM algorithm is more versatile with fast calculation speed. The iterative smoothness of the correction can be improved， compared with the traditional optimal interpolation （OI） method.

Abstract:Mass data storage and synchronization are important issues in intelligent manufacturing production lines. The current mainstream method of data synchronization is remote synchronization （RSYNC） algorithm. It uses the method of synchronizing incremental data to reduce the amount of data transmission. The data generated by the intelligent manufacturing line has a deep hierarchical structure and a complex directory structure， thus leading to a long evaluation time during synchronization. Here， a hierarchical Hash numbering algorithm is proposed for synchronization based on data hierarchical numbering of data files， by using a Hash table to record hierarchical information， quickly comparing differential data， and adopting different backup strategies for different types of differential data. Experimental results show that compared with standard RSYNC， the proposed method effectively reduces the amount of data evaluated by RSYNC and synchronization time， which also improves synchronization backup efficiency.

Abstract:Under the environment of big data， with the rapid expansion of the online advertising industry， the online advertising calculation has attracted more and more attention. Computational advertising aims at placing ads on a specific audience， performs data analysis and calculation based on the advertising environment and user characteristics， and selects the best matching ad from the candidate ad library. The core issue is the calculation of click conversion rate prediction for online advertising， which selects the ads with the highest probability of users clicking. The accurate prediction of advertisement click conversion rate is related to benefits of publishers， advertisers and users. Based on the advertising data provided by the TrackMaster platform， this study analyzes user information features， advertising information features， context features and statistical features from the perspective of feature engineering. The larger effects on the advertising click conversion characteristics are excavated out. Layered advertisement click conversion rate prediction model is constructed and trained. The LightGBM algorithm model is adopted to obtain the important feature ranking of the ad click conversion rate. The experimental results indicate that when the feature selection threshold is 0.95， the number of feature choices is 19， and the number of trees is 100， the area under receiver operating characteristic （ROC） curve （AUC） value of the model is the maximum， and the logarithmic loss function value of the model is about 0.136 8. The model has the optimal effect. The prediction model and the result of feature ranking are helpful for the enterprise to make the optimal advertising strategy.

Abstract:The x-vector system maps a variable-length speech to a fixed-dimensional speaker embeddings via neural networks， and performs well in text-independent speaker verification. Here， it is applied to the text-dependent speaker verification and different x-vectors are extracted according to different contents in one sentence. In model selection， deep residual network （DRN） is used to obtain more discriminative x-vector. For a sentence with multiple words， word-dependent DRNs are trained to extract word-dependent x-vectors， which are separately fed to different backend classifiers. Finally， multiple scores are fused to obtain the final verification results. Experiments on Part Ⅲ of the RSR2015 dataset show that the proposed method can achieve equal error rate （EER） reduction of 15.34% and 19.7% for male and female， respectively.

Abstract:Most of the existing cross-domain sentiment classification methods only take advantage of the migration feature from a single source domain to a target domain， without fully considering connections between target domain instances and different source domains. To solve this problem， this paper proposes an unsupervised multiple-source cross-domain sentiment classification model. First， the base classifier is trained by using the migration feature of a single source domain to a target domain， and different base classifiers are weighted. Then， the ensemble consistency of different base classifiers on the target domain instance prediction is taken as the objective function， and the objective function is optimized to obtain the weights of different base classifiers. Finally， the weighted base classifier is used to obtain the sentiment classification results of the target domain. The model is tested on Amazon's product review data set and Skytrax data set， and is compared with six baseline models. Experimental results show that compared with the baseline model， the classification performance of the proposed method is significantly improved in eight different target domains.

Abstract:In order to solve the problem that the traditional high resolution spectral estimation methods estimate the direction of arrival （DOA） of far field sound source with large computation and inaccurate estimation of coherent signals， a two-dimensional DOA estimation method based on circular microphone array and fourth-order cumulant is proposed. The algorithm combines the advantages of circular array location without dead angle and matrix virtual extension to obtain more information of sound source location. Firstly， the uniform circular array （UCA） is virtualized into 2K+1 uniform linear arrays （ULAs） by means of pattern space transformation， and the virtual linear array is divided into L subarrays by space smoothing technology. Then the fourth order cumulative construction method is used to extract the effective matrix information and remove the redundant data. Finally， the azimuth and pitch angles of the acoustic source signal are obtained by the MUSIC-like algorithm. The simulation results show that the proposed algorithm can achieve high-precision estimation of far field coherent signals compared with traditional high-resolution spectral estimation algorithms when the signal-to-noise （SNR） ratio is low. At the same time， the algorithm also has lower root mean square error performance and can effectively reduce the running time.

Abstract:Electronic voice disguise refers to hiding the identity of a speaker by voice changing equipment or voice processing software. The restoration of disguised voice refers to changing it back to its original version， which is of great significance for speaker identification. This paper first models the restoration of disguised voices as the estimation of disguising factors in both frequency and time domains. The estimation of disguising factor is made by automatic speaker verification using i-vector. Symmetric transformation is proposed to improve the performance on parameter estimation. By virtue of the noise robustness of i-vector， the proposed method improves the estimation accuracy of the disguising factor in the real noise-containing scene， thereby improving noise robust restoration effect of electronic disguised voice. Evaluation results on noisy speech library VoxCeleb1 of the trained model on clean speech library TIMIT demonstrated good performance of the approach by reducing the error rate from 9.19% to 4.49%. The quality of the restored voice is also improved in the aspects of automatic speaker verification and auditory perception.

Abstract:The feature of the slice of ambiguity function main ridge can better reflect the structural essential differences between signals， and it is a feasible parameter to solve the current complex system radar emitter signal sorting problem. The fast and intelligent search for the slice of ambiguity function main ridge is an important issue to increase the practicability of its feature of the slice. In this paper， an improved self-adaptive grey wolf optimization （GWO） combining uniform initialization strategy and improved nonlinear convergence factor was proposed to search the main ridge slice of ambiguity function of six typical radar emitter signals and extract the feature of slices， which were compared with the exhaustive method and standard GWO.The experimental results showed that the average time consumption of the proposed method was only 1.49 s when searching AFMR slice and extracting feature. Compared with the exhaustive method and the standard GWO， the efficiency was improved by 75.7% and 19.0%， respectively， with better timeliness. In a fixed SNR environment， when the SNR was not less than 0 dB， the average clustering accuracy of the extracted feature values was 96.4%； and in a dynamic SNR environment of 0—20 dB， the average clustering accuracy can reach 95.2%， with good accuracy， anti-noise performance and strong intra-class aggregation and inter-class separation ability， which proves the feasibility and effectiveness of the proposed method.

Abstract:For indoor environment， the WiFi signal strength is susceptible to external interference. Due to its instability， the accuracy of matching in the fingerprint database is low and the positioning accuracy is not high. An optimization algorithm based on indoor fingerprint positioning is proposed. This algorithm optimizes the fingerprint database and matching algorithm. Database optimization uses limiting and moving average filtering for pre-processing. According to the indoor environment， assign the ID of the area to which the sampling point belongs to build a multidimensional fingerprint database. The matching algorithm is optimized to classify the points to be located according to the support vector machine （SVM） and obtain the corresponding area ids. The Euclidean distance， Manhattan distance and Chebyshev distance are combined to obtain a position estimate. Finally， combined with the pedestrian dead reckoning （PDR） algorithm， the obtained step size and heading angle are subjected to particle filtering to achieve positioning. The proposed algorithm improves the positioning accuracy by 13.92%.

Abstract:Blind source separation （BSS） algorithm is utilized in extracting and recovering source signals from mixed signals. Among many different BSS algorithms， the principal skewness analysis （PSA） is a BSS algorithm taking third-order statistics as the objective function. One of its advantages is that its calculation speed is faster than the conventional BSS algorithms. However， because of the serial calculation method， error accumulation exists in the calculation process. In order to solve this problem， this paper proposes an algorithm called parallel PSA. In this algorithm， parallel calculation is used instead of serial calculation， and the corresponding directions of each independent component can be estimated simutaneously， so the problem of error accumulation is avoided. Simulation results prove that， compared with the PSA algorithm， the parallel PSA algorithm not only maintains the fast calculation speed， but also improves the estimation accuracy of each source signal.

Abstract:LEACH protocol is a classical algorithm for distributed cluster networks in wireless sensor networks.But the problems such as uneven energy consumption and premature death of nodes are more serious.An improved LEACH-EDP protocol based on partition and energy-distance factor is proposed in this paper.The threshold function of cluster head election is optimized by residual energy and distance correction coefficient.A partitioning policy is used in the protocol.The weights of various gain parameters are adjusted according to the region.The simulation results show that compared with the traditional LEACH protocol， the death node of LEACH-EDP protocol was delayed by 79.5%， and the network death time was delayed by 57.4%.

Abstract:The Q & A system is a kind of system which can answer user’s questions with accurate and natural language. Some improvement measures have been tried for “named entity recognition”. Aiming at the time- and labor-consuming problem of traditional one-way template matching， this paper proposes a lattice bi-directional structure of long short-term memory （Lattice Bi-LSTM）network， which solves the problems of improper sentence processing and dependence on the result of word segmentation in named entity recognition. Compared with the unidirectional structure， the bi-directional structure can make better use of sentence information and make the output more robust， thus capturing semantic information more accurately. To solve the problem of non-linear coupling of similarity between entities in traditional methods， a method is proposed to link “similar” entities to the knowledge base accurately by using periodic kernel function. The two improved methods are verified by experiments， whose results show that they have significant improvement effects compared with the classical method.

Abstract:A new physical resource utilization thresholds management strategy called RUT-MS for cloud data centers is proposed in this paper. In RUT-MS， the virtual machine migration process is divided into five steps： Overloaded host status detection， virtual machine selection， the first virtual machine placement， under-loaded host status detection and the second virtual machine placement phases. In overloading host detection， RUT-MS uses an iterative weighted linear regression method to determine two utilization thresholds so that the performance degradation is avoided. Maximum power reduction policy （MPR） is adopted in virtual machine selection phase. A vector magnitude squared of multiple dimension resources is used to decide the low resource utilization thresholds and switch them to a power saving mode. Finally， RUT-MS is evaluated using CloudSim with real-world workload data. Simulation results show that RUT-MS can reduce the energy cost incurred on the system due to migration. The reduced number of virtual machine migrations and improved energy-efficiency are also obtained in our test. Moreover， the service level agreement （SLA） violation and SLA for energy united metrics are also good.

Abstract:To quickly and accurately identify network attacks in a multi-dimensional environment with diversified attack forms and massive intrusion data， an intrusion detection model combining Fisher-PCA feature extraction and deep learning is proposed. Firstly， the Fisher feature selection algorithm selects important features to form feature subsets. Then the dimension of the feature subsets is reduced based on principal component analysis （PCA） and the feature set with strong classification ability is extracted. A new deep neural network （DNN） is constructed to identify and classify network attack data and normal data. Experimental results on KDD99 dataset show that compared with the traditional artificial neural network（ANN） and support vector machine（SVM） algorithms， the accuracy of this intrusion detection algorithm can be improved by 12.63% and 6.77%， respectively， and the false alarm rate is reduced from 2.31% and 1.96% to 0.28%. Compared with DBN4 and PCA-CNN algorithms， its accuracy and detection rate are basically the same， while the false alarm rate is lower.

Abstract:An emotional analysis method based on LDA thematic model and intuitionistic fuzzy TOPSIS for agricultural product online reviews was proposed. The method uses the affective dictionary to analyze the emotional tendency of online comments and calculate the positive emotional value of agricultural products. The LDA thematic model is used to calculate the weight of each attribute， and the comprehensive evaluation value of agricultural products is calculated with the intuitive fuzzy TOPSIS method. SPSS statistical analysis software was used to verify the validity. The results show that the comprehensive evaluation value has a significant positive correlation with monthly sales volume and positive emotional value， which indicates that the method is reasonable and provides a new idea for mining emotional information in the online evaluation of agricultural products.

Abstract:An algorithm based on mutual information （MI） is proposed for the time resource optimal allocation of phased array radar （PAR）， which focuses on multi-target tracking under active suppression jamming. The MI criterion is used as the quantitative indicator of target tracking， and the MI expression of the radar echo and the path gain matrix which contains the time variable is derived. Then the time resource optimal allocation model is established， and the improved genetic algorithm is used to solve the model. The simulation results show that the algorithm can maximize the number of effectively tracked targets under active suppression jamming， and the overall tracking performance of the effective tracking targets is improved.

Abstract:The multi-attribute decision making problem in the relationship between attribute values under Pythagorean fuzzy number information environment is researched. Firstly， the Pythagorean fuzzy number operations based on t-norm and t-conorm are defined. Then， the Heronian mean is integrated into the construction process of aggregation operator. The three characteristic properties of Pythagorean fuzzy Heronian mean （PFHM） and its commonly used special cases are also discussed. In addition， a new improved Pythagorean fuzzy decision making model is constructed， which considers the relationship between input attribute values and improves the scope of decision-making. Finally， an example of multi-attribute decision making is provided to verify the rationality and effectiveness of the proposed model.

Abstract:In order to solve the problem that the multi-screen video display system cannot adapt to the multi-screen transmission of ultra-high-definition （UHD） video according to the number of split screens， a heterogeneous multi-core video streaming transmission method is proposed. First， the embedded system based on ARM is used to realize multi-task processing and real-time monitoring. Then FPGA realizes hardware acceleration of receiving， converting， processing and split-screen output display of video stream. The system can configure the working parameters of FPGA according to the input signal in real time to realize the split screen display of UHD video stream with variable resolution and number of split screens. Finally， the system is tested using the Zynq UltraScale+MPSOC XCZU7EV heterogeneous multi-core processor，and results show that the multi-channel split-screen mosaic of 4K video has no obvious dislocation and consistent synchronization， which better meets the requirements of video multi-channel split-screen display.