Vessel trajectories implied in AIS data are crucial to obtain a good understanding of the maritime traffic situation for shipping safety. Starting from raw AIS data, a trajectory database is created for vessels within surveillance area after parsing, noise reduction, and DBSCAN clustering. With mmsi as the key index, the trajectory for each vessel is extracted ordering by timestamp. To remove the time interval difference between points in trajectories, interpolation and cleaning are carried out on each vessel trajectory to get trajectories with equal time intervals. Through implied motion pattern computation between adjacent points in each trajectory, maneuvering actions can be identified. Then, sailing segments with continuous same maneuvering actions are merged. With sailing segments partition results, critical points are extracted for already known different collision avoidance situations. Trajectory similarity computation for different vessels are computed with our new multi-scale and multi-resolution trajectory matching method. Experiments for the recognition of collision avoidance situations show that the adoption of the matching algorithm with multi-scale and multi-resolution trajectories for different vessel pairs to complete collision avoidance situations analysis is effective and achieves good performance.

With the network model architecture of Google Inception, research is conducted on issues such as the structural design of the model, data preprocessing, tuning of training parameters, computing clusters in a distributed environment, and multi-machine parallel training. According to the performances of different deep neural network models on different data sets, the Google Inception V3 depth network model is used as the prototype to conduct the tuning of training parameters, and the classification of remote sensing images is then realized with this model in the single-machine environment. Furthermore, due to the effectiveness of distributed systems for very large data sets and compute-intensive applications, a data parallel training scheme based on the distributed platform is designed for the convolution neural network model with more complex data form, larger quantity of parameters, and more network levels, after studying the mainstream designs of the distributed machine learning and analyzing the training methods and steps of the convolutional neural network model in a multi-machine environment. It greatly improves the training time of the model, and then the classification of remote sensing images under distributed clusters is realized.

For the micro-blog sentiment analysis problem in big data environments, the "dimension disaster" caused by the continuous increase in text information data brings great challenges to the emotional analysis of micro-blogs. To solve this problem, this paper proposes a fusion of the advantages of three feature dimensionality reduction algorithms, based on the traditional document frequency (DF), mutual information (MI), and chi-square test (CHI). Firstly, the document frequency factor is added to the mutual information (MI) algorithm to solve the problem of low-frequency word defects. Then, the standard score factor is added to the chi-square test (CHI) algorithm to solve the negative correlation problem. Finally, the average value is calculated and the advantages of the three algorithms are fused. An improved Proposed DF-MI-CHI fusion algorithm is proposed. The simulation results show that after using this algorithm to process the micro-blog data, the accuracy of sentiment analysis is improved and maintained at 95%. The recall rate is more than 90%, and the F value is maintained between 92% and 94%. In the % interval, it is higher than other improved algorithms and tends to be stable, which indicates that the algorithm can effectively improve the accuracy and efficiency of micro-blog emotional sentiment analysis when dealing with massive micro-blog text data.

Higher levels of PM2.5 concentration are becoming the leading cause of hazy days in China. However, studies have shown that the variations of PM2.5 involve complicated physical and chemical processes, which make their accurate predictions challenging. Meanwhile, the forecast results from numerical models frequently deviate from observation values. The deep learning method is a good substitute for the prediction of mass time series data in the field of meteorology. In the present study, a framework for PM2.5 concentration prediction is presented based on a three-dimensional convolutional neural network (3DCNN) and long short term memory neural network (LSTM). Using preprocessing, correlation analysis, feature extraction, and transformation, spatiotemporal sequence data was generated. In the spatiotemporal feature extraction phase, 3DCNN was used to extract high-level spatial features, and LSTM was used to extract temporal features. In the prediction phase, full connect (FC) was used to combine spatial and temporal features. To examine the efficacy of the proposed model, the PM2.5 concentration data, meteorological observation data, and grid dataset collected at ten observation stations in the Beijing Meteorological Bureau (BMB) were used. After the performance evaluation was compared with several methods including this proposed model, support vector machine (SVM), and the existing PM2.5 forecast system in BMB, root mean square errors (RMSE) and mean absolute errors (MAE) were chosen as evaluation indicators. The experimental results showed that the proposed model performed the best, the minimum MAE value was 3.24μg/m3, and the minimum RMSE value was 13.56μg/m3 over the ten stations. In addition, the proposed model overcame the underestimation produced by the existing PM2.5 forecast system in BMB and demonstrated superior performance for different time lengths over a 24-hour period. The results also confirmed the effectiveness of the deep learning method in the prediction of PM2.5 concentration.

This study proposes a method for the real-time detection and recognition of targets in streetscape videos. The proposed method is based on separation confidence computation and scale synthesis optimization. First, on the basis of generalization in transfer learning, we combine a fine-tuning method suitable for non-convex optimization and adaptive moment estimation in high-dimensional space. Then, we dynamically adjust the learning rates of parameters on the basis of first and second gradient moment estimations. We establish the framework and implementation steps of the proposed method by organically combining regular term super-parameter generalization and hard-example mining technology. We use the proposed method to detect and recognize targets in streetscape videos with high frame rates and high definition. Furthermore, we experimentally demonstrate that the accuracy and robustness of our proposed method are superior to those of conventional methods.

In this paper, a decision support system (DSS) of model, data, and human-computer interaction is established to evaluate the fire evacuation performance of atriums. Firstly, based on a large amount of literature and consultation with relevant scholars and experts, this paper established an indicator system for the influencing factors of fire evacuation in building atriums. Secondly, this paper used the entropy weight method model and the comprehensive rating model as the core of the model library to calculate the weight of each level indicator system and complete the evaluation work. Then, the back-end database of the data components established by Access (including the basic database, algorithm database, and expert database) was used to store the data tables and data files generated in the system work. Finally, based on the relevant literature and norms, a scientific indicator scoring standard was established. By inputting the indicator score of the evaluated object into the DSS, the final evaluation result was obtained. At the end of this paper, a large shopping center was used as a case study. The DSS was awarded a medium fire evacuation level with a score of 78.1. In order to improve its fire evacuation capability, the DSS gave corresponding decision suggestions.

With the expansion of college enrollment in recent years, the quality of students' learning is beginning to decline. At present, education quality governance has become the internal demand of the reform and development of higher education. Learning performance prediction is an important means to effectively resolve the academic crisis and improve the overall education quality. In this study, firstly, the current status and problems about learning performance prediction were analyzed from the perspective of basic data, evaluation indicators, and prediction methods. Secondly, driven by ten items of basic learning situation data, a learning performance prediction model based on the RBF neural network was established, which included three layers in network topology the input layer, hidden layer, and output layer. The activation functions of the hidden layer and output layer were a Gauss radial basis function and linear function, respectively. The modeling process included three steps forward propagation computing prediction loss, error backward propagation adjusting network parameters, and network optimization determining model hyperparameters. The obtained results showed that the trained model had small relative root mean square error values for both the training data and testing data. When comparing the original observation values and model predicted values, it was observed that most of the sample points were evenly distributed on both sides of the diagonal line of the contrast graph, which indicates that the RBF neural network model employed in this study is promising in learning performance prediction. It is of good reference significance for promoting more accurate and efficient learning performance prediction and improving the efficiency and effectiveness of education quality governance.

The relay selection and power allocation of cooperative MIMO with multiple-antenna terminals are investigated. Based on the classic worst-link-first algorithm, we propose a joint power allocation and relay selection algorithm, which takes maximizing the system energy gain as the optimization purpose. We consider quasi-static, flat fading channels, select relays, and allocate power according to the instantaneous channel state information. Theoretical analysis and computer simulations show that the proposed algorithm can achieve higher average network energy gain than the traditional worst-link-first algorithm, and its computational complexity is much lower.

To resolve the problem of unreasonable resource allocations caused by the continuous arrival of different types of jobs in a hybrid cloud environment, an adaptive job-scheduling algorithm based on queuing theory is proposed. This paper analyses job load types, and the jobs are classified according to the logistic regression method. A resource utility is used to classify the nodes in a private cloud cluster by considering the heterogeneity of the private cloud resources. Based on the job classification and the resource classification, a queuing model is established, and an adaptive genetic algorithm is used to manage the job queue's arrival rate that becomes the basis of the resource allocation. The proposed algorithm is compared with some existed similar algorithms to verify its performance in terms of job response times and throughput.

Credit evaluation is an important issue for investors in the financial field. However, there is a large amount of missing data in the P2P lending platform. To evaluate borrowers' credit from missing data, a credit evaluation Bayesian network model learning algorithm is proposed based on domain knowledge. Specifically, we first give a credit evaluation Bayesian network (CEBN) model to represent the borrowers' attributions and the relationships between attributions, and then we design the CEBN learning algorithm based on domain knowledge. Furthermore, we analyze and discuss the time complexity of the algorithm. Finally, the experimental results demonstrate that the CEBN model has good interpretability, learning performance, and evaluation performance by comparing it with other methods.

With the growth of the intrusion data scale model, irrelevant or redundant features in high-dimensional intrusion detection data leads to slow processing speed of the intrusion detection algorithm, and the consumption of the algorithm in time and space will increase as the feature dimensions increase. In view of good classification performance of the Elephant Herding Optimization (EHO) algorithm in reducing feature redundancy, this paper introduces the EHO algorithm into feature selection for intrusion detection. Since the basic EHO algorithm tends to fall into a local optimum and lacks strong search ability, the classification performance and dimensional reduction ability of the algorithm are severely limited. Therefore, an Improved Elephant Herding Optimization (IEHO) algorithm is proposed in this paper to search the feature space and find the optimal feature subset, so that the feature number is minimized while the classification performance is maximized. As the scale of intrusion data grows, the large amount of redundant information in the intrusion data will cause the improved algorithm to process slowly. Thus, in this case, the improved algorithm is considered to be parallelized to relieve the pressure of single-machine operation. This paper then proposes a Spark-based distributed parallel IEHO algorithm for intrusion detection, and a feature selection method based on this algorithm for intrusion detection is discussed. The feature selection in a distributed environment can improve the running efficiency of the IEHO algorithm, so as to reduce the running time of the algorithm under the premise of ensuring classification accuracy. As for the experimental validation, both UCI and KDD CUP99 datasets are used to verify the feature selection for intrusion detection. Compared with the classical PSO, MFO, and EHO algorithms, the feature selection by the binary IEHO algorithm is improved by 4.16%, 1.42%, and 0.98%, respectively, and the classification performance is also significantly improved. Compared with the stand-alone version of the IEHO algorithm, the classification efficiency of the parallel IEHO algorithm based on Spark for intrusion feature selection is significantly improved, and the acceleration ratio is increased by two orders of magnitude.

Hyper-sphere Support vector machine (SVM) is a widely used machine learning method for multi-class classification problems such as image recognition, text classification, or handwriting recognition. In most cases, only one hyper-sphere optimization problem is computed to solve the problem. However, there are many complex applications with complicated data distributions. In these cases, the computation cost will be increased with unsatisfied classification results if only one support vector machine is adopted as the classification decision rule. To achieve good classification performance, a piecewise combination of the hyper-sphere support vector machine is put forward in this paper based on the analysis of the data sample distribution. First, statistical analysis is adopted for the original data. Then, the k-means cluster algorithm is introduced to compute cluster centers for different classes of the data. For the n classes classification problem, m (m > n) hyper-spheres are computed to solve the objective problems based on the number of data centers. For simple sphere-distribution and locally linearly separable distribution cases, the minimum enclosing and maximum excluding support vector machine and the combination of hyper-sphere support vector machine are defined. Experimental results show that different support vector machines for different data distributions will improve the final classification performance.

In the artificial intelligence field, using computer vision to track an object is an important research topic. Especially when the target reappears after being occluded for a while, it is hard to precisely track the moving target again. Therefore, this paper proposes an anti-occlusion target tracking strategy that can overcome the occluded problem. Firstly, to make the target clearer, we design a moving target detection method using the Gaussian mixture background subtraction method based on the wavelet transform, which removes the high-frequency noise of video images. Then, in the tracking process, altered strategies are taken to cope with different occlusion situations, which include three cases no occlusion, partial occlusion, and severe occlusion. For the first two cases, we use the distance-based Kalman filter method to track the moving target. For the third case, we designed a method that combines the Camshift method with the distance-based Kalman filter method to track moving targets, which is more efficient than only using the distance-based Kalman filter method. According to one of the cases, our program automatically selects the corresponding method. Experimental results show that our strategy can track moving targets accurately whether targets are in occlusion situation or not.

For privilege escalation attacks in the Android system, the detection method of role behavior was proposed based on component features and process algebra. The classification of roles was constructed from the analysis of the privilege escalation attack model. Feature extraction from components includes component permissions, component communication, API calls, and sensitive data flow. Process algebra was used to construct modes of role behavior, and roles of applications were identified through equivalence relation. Finally, the dangerous path was detected in multi-applications, and then applications constituting to privilege escalation attacks were ascertained. The experiment showed that the proposed method can effectively detect privilege escalation attacks, the potential safe hazards in applications were pointed out, and the role of applications was identified.

Network reliability is an essential issue of complex networks; the reliability of complex networks plays an important role in the performance in the research process. At the same time, the number of connected nodes in a complex network is a main measure of the complex network. Due to the randomness of complex networks, we define one new degree sequence and the entropy of the complex network, and we then study the entropy of the network as a new measure for the network reliability. The features of entropy are studied in complex networks, and entropy is analyzed in two representative complex network models, the random network model and scale-free network model. The degree distributions functions in the random network model and scale-free network model have significantly different characteristics, the Poisson distribution and Power-law distribution. Furthermore, we study the entropy features under two nodes fault models, random failures and deliberate attacks. We discuss the entropy of the random network model and scale-free network model in two fault modes with the fault intensity gradually increasing from 0 to 1.0. Then, we study the relation between the average degree distribution and the entropy of the network when the fault intensity is 0.3. The results show that the entropy of the network is reasonable to measure the network reliability similar to the number of connected nodes in the network. The purpose of the research is to provide a new way to study network reliability.

In order to study the influence of design variables on the resonance reliability of axle box bearings of electric multiple units (EMU), two methods of resonance reliability analysis of axle box bearings are proposed based on fuzzy reliability and probabilistic reliability. The randomness of the elastic modulus, Poisson's ratio, density, and speed of axle box bearings and the fuzziness of the resonance criteria are considered. The fuzzy reliability theory is introduced to analyze the resonance reliability of axle box bearings, and the resonance reliability equation is deduced and combined with the theoretical analysis of bearing vibration. Then, the resonance reliability of axle box bearings based on fuzzy theory is obtained. Based on this, the axle box bearing of EMU is parameterized by APDL language. The resonance reliability analysis of axle box bearings is carried out by using the Monte Carlo simulation (MCS) method based on the frequency interference model. The results show that the discreteness of the design variable and speed of axle box bearings have a negative influence on the resonance reliability of bearings, greatly increase the probability of resonance, and reduce the reliability of bearings. By comparing the results of resonance reliability analysis obtained by the probabilistic design method, the accuracy and practicability of the resonance fuzzy reliability analysis are verified, which provides a theoretical reference for the anti-resonance design of axle box bearings.

For the problem of T-S fault tree modeling, the probability of failure of the bottom event is not easy to obtain. Firstly, the index is graded, the expert's scoring matrix is obtained according to the scoring standard, the index weight is obtained by applying the entropy method, and the expert weight is obtained by integrating the experts. Secondly, the fuzzy possibility of different fault states of the bottom event is given by experts according to experience and data, and the fuzzy probability of the bottom event is obtained according to the triangular fuzzy number operation rule. The rule execution possibility fuzzy probability of T-S containing triangular fuzzy numbers is derived. Finally, the analysis is carried out in combination with the screw side screw failure of the explosive.

It is significant to forecast the short-term wind speed for the safety of wind turbine blades and the optimization of power grid dispatching. Firstly, the local comparison method is established to forecast the mean wind speed. Secondly, the universal generating function (UGF) is used to express the wind speed as a multi-state random variable, state probability allocation and the state probability matrix are used to obtain the risk state probability, and equal dimension filling is used to update the information. Then, the maximum wind speed is calculated based on the mean wind speed and risk state probability. Thirdly, local comparison is used for error forecasting, and the forecasting errors are used to correct the forecasting wind speeds. Finally, the mean circular tube is constructed, and the mean wind speed, maximum wind speed, risk state probability, and average relative error are displayed in the combined mean circular tube together.

The mission reliability of space telemetry tracking and command (TT&C) systems directly determines whether the space mission can be carried out successfully. In actual situations, some missions of TT&C systems may just be required to be completed before a given time, rather than be performed in a specified time window. Taking advantage of this feature, mission failures in the originally arranged time period could then be re-executed later according to the rescheduling algorithm. The resources to be used and the time to be executed are two kinds of uncertainties that exist in rescheduling. This paper treats them as dynamic redundancy. The existing evaluation methods are insufficient to evaluate the mission reliability of system under this condition. This paper proposes a heuristic rule based on the system operation-principle to obtain the basic scheduling plan of the mission and then designs a reactive scheduling strategy to deal with mission failures during the execution process. Based on this scheduling architecture, a basic scheduling plan combined with a fault handling strategy can be obtained. A Monte Carlo method is used to simulate the components' failure and repair, and a discrete event simulation process is used to simulate the possible execution options during the execution process. A collection of altered execution options combined with the execution result of the mission can be gathered after a huge amount of simulation. The approximate mission reliability of the system can then be assessed through a statistic reference to this collection.

Aimed at the unsatisfactory effect of security in mobile cloud computing, we propose a novel image encryption algorithm by using the Improved Fruit Fly Optimization Algorithm -- Artificial Bee Colony (hereafter referred to as IFOA-ABC). First, the feature of task scheduling function in mobile cloud computing is described to set up the task scheduling function; second, the population of fruit fly optimization algorithm is initialized by using orthogonal array and quantitative techniques, and the boundary of fruit fly optimization algorithm is processed, the step size in search is adjusted dynamically, and the artificial bee colony algorithm is used to help fruit fly optimization algorithm develop the global optimal solution; in the simulation experiment, the IFOA-ABC algorithm has obvious advantages over other intelligent algorithms in the comparison of four indicators of mobile cloud task scheduling, which means that the proposed algorithm can effectively improve the privacy and efficiency of cloud computing scheduling.

High precision harmonic analysis in electric power systems is the precondition to evaluate the power quality of power grids and to control the harmonic of power grids. The fast Fourier transform (FFT) algorithm has shortcomings such as spectrum leakage and the fence effect in harmonic analysis, resulting in lower accuracy. Thus, a power harmonic analysis method based on atomic decomposition combined with the read-coded quantum evolutionary algorithm (AD-RQEA) is proposed. The core of AD-RQEA is to construct an atom library according to the characteristics of harmonic signals and to optimize feature parameters of atoms using AD-RQEA. Finally, optimal matching atoms are adaptively chosen to reconstruct voltage signals. The comparison tests show that the proposed method has high accuracy, and the effectiveness and practicability of this method are verified.

Aiming at the unbalanced classification problem of remote sensing ship image datasets in ship target classification and the problem that the traditional decision tree classification algorithm needs to rely on artificial construction features to realize classification, a weighted deep neural decision forest is proposed. This method combines deep learning with resampling. The results show that the method can achieve a better classification accuracy than the traditional decision tree on unbalanced classification of ship target.

This paper explores trajectory tracking control for underactuated Autonomous Underwater Vehicles (AUV). The controller of dynamics based on backstepping technology is proposed. Firstly, the dynamics and kinematic model of AUV are designed. Secondly, the cascade method is applied to decompose the tracking error system into position tracking error subsystem and heading angle tracking error subsystem. Finally, backstepping technology is used to design the controller. Two virtual feedback variables are constructed, an AUV trajectory tracking controller is designed, and then the stability of the system is proven by the Lyapunov method. The results demonstrate that the proposed control schemes can make tracking errors converge uniformly.

Due to the high dependency on software, the measurement of its performance has become vital. It is a common practice in the software industry to test products exhaustively before release so that the maximum number of faults is detected and removed. Fault detection and the removal rate are governed by many factors such as changes in testing environment, testing strategies, skills, efficiency, etc. The point at which a change in the fault detection rate occurs is known as the change point. Due to the increasing demand of good quality software in a short span and to remain in the market competition, firms are providing warranties on their products to assure reliability. The defects reported during the operational phase are fixed by providing patches. However, delivering patches after release demands extra effort and resources, which is costly and hence not economical for the firms. Considering the above factors, an effort-based cost model with change point and warranty is proposed in this work to determine the optimum release and patch time of a software by minimizing the overall cost. A numerical illustration is provided to validate the proposed cost model.

Based on the reduction and redundancy methods, the reliability performance of the improved general series-parallel system is considered, assuming the connected components are identically independent and follow the general exponential lifetime model. To extend previous studies, the shape parameter is modified to obtain the reliability equivalence factors of the hot and cold duplications. A hybrid of the hot and cold duplication methods is also considered. Numerical results from a practical example are investigated to illustrate the derived theoretical results of the overall study.

In order to estimate the instantaneous reliability during a periodic inspection of a lifting wire rope, analytical methods based on experimental results have been developed. Our approach is to study the influence of number of broken strands in the outer layer on the mechanical behavior of the wire rope. To do this, we have evaluated the damage of the wire rope through a static damage model used to predict the lifetime of the steel wire rope. Then, we estimate the reliability by the Weibull law. Finally, we linked the two parameters (damage and reliability) across the life fraction to find the correspondent reliability at each stage of the damage to conduct a predictive maintenance and change the wire rope at the right time.