Search

Abstract Object detection has been widely applied in various fields with the rapid development of deep learning in recent years. However, detecting small objects is still a challenging task because of the limited information in features and the complex background. To further enhance the detection accuracy of small objects, this paper proposes an efficient single-shot detector with weight-based feature fusion (WFFA-SSD). First, a weight-based feature fusion block is designed to adaptively fuse information from several multi-scale feature maps. The feature fusion block can exploit contextual information for feature maps with large resolutions. Then, a context attention block is applied to reinforce the local region in the feature maps. Moreover, a pyramids aggregation block is applied to combine the two feature pyramids to classify and locate target objects. The experimental results demonstrate that the proposed WFFA-SSD achieves higher mean Average Precision (mAP) under the premise of ensuring real-time performance. WFFA-SSD increases the mAP of the car by 4.12% on the test set of the CARPK.

The traditional network security architecture ensures network security by directing traffic through hardware based network security function devices.Since the architecture consists of fixed hardware devices, it leads to a single form of network security area deployment and poor scalability.Besides, the architecture cannot be flexibly adjusted when facing network security events, making it difficult to meet the security needs of future networks.The intelligent service chain system for network security resource pool was based on software-defined network and network function virtualization technologies, which can effectively solve the above problems.Network security functions of virtual form were added based on network function virtualization technology, combined with the existing hardware network elements to build a network security resource pool.In addition, the switching equipment connected to the network security elements can be flexibly controlled based on software-defined network technology.Then a dynamically adjustable network security service chain was built.Network security events were detected based on security log detection and a expert library consisting of security rules.This enabled dynamic and intelligent regulation of the service chain by means of centralized control in the face of network security events.The deployment process of the service chain was mathematically modeled and a heuristic algorithm was designed to realize the optimal deployment of the service chain.By building a prototype system and conducting experiments, the results show that the designed system can detect security events in seconds and automatically adjust the security service chain in minutes when facing security events, and the designed heuristic algorithm can reduce the occupation of virtual resources by 65%.The proposed system is expected to be applied to the network security area at the exit of the campus and data center network, simplifying the operation and maintenance of this area and improving the deployment flexibility of this area.

Abstract Predicting the destination of a moving object is a popular research subject in location‐based services. By predicting destinations, suggestions can be offered to people regarding their trips. At present, there are problems such as data sparsity and long‐term dependence based on historical trajectory prediction methods, which affect the accuracy of prediction. To solve data sparsity problem, this paper has devised an improved minimum description length method, which incorporates weighting parameters and optimizes the partitioning of trajectories with undirected complete graph. Long–short‐term memory is a trajectory‐prediction model that solves the problem of long‐term dependence, but the model tends to have vanishing gradient issues when used to process longer sequences. This is because the hidden layers of long–short‐term memory is largely affected by the lengths of sequences. Using embedded technology, the authors convert trajectory sequences into embedded vector sequences, and thus propose a deep‐learning prediction model, EP‐LSTM (Embedded Processing ‐ Long Short Term Memory), which integrates embedded technology and long–short‐term memory. The authors have conducted a great amount of testing with real data sets, comparing EP‐LSTM with currently available predicting methods. The results have shown that EP‐LSTM not only effectively solves data sparsity and long‐term dependence but also achieves a high degree of prediction accuracy.

Surface roughness is a key indicator of the quality of mechanical products, which can precisely portray the fatigue strength, wear resistance, surface hardness and other properties of the products. The convergence of current machine-learning-based surface roughness prediction methods to local minima may lead to poor model generalization or results that violate existing physical laws. Therefore, this paper combined physical knowledge with deep learning to propose a physics-informed deep learning method (PIDL) for milling surface roughness predictions under the constraints of physical laws. This method introduced physical knowledge in the input phase and training phase of deep learning. Data augmentation was performed on the limited experimental data by constructing surface roughness mechanism models with tolerable accuracy prior to training. In the training, a physically guided loss function was constructed to guide the training process of the model with physical knowledge. Considering the excellent feature extraction capability of convolutional neural networks (CNNs) and gated recurrent units (GRUs) in the spatial and temporal scales, a CNN–GRU model was adopted as the main model for milling surface roughness predictions. Meanwhile, a bi-directional gated recurrent unit and a multi-headed self-attentive mechanism were introduced to enhance data correlation. In this paper, surface roughness prediction experiments were conducted on the open-source datasets S45C and GAMHE 5.0. In comparison with the results of state-of-the-art methods, the proposed model has the highest prediction accuracy on both datasets, and the mean absolute percentage error on the test set was reduced by 3.029% on average compared to the best comparison method. Physical-model-guided machine learning prediction methods may be a future pathway for machine learning evolution.

In the prediction of zinc-binding sites in proteins, there are few real binding-site residues, whereas most residues are non-binding-site residues, resulting in a typical imbalanced classification problem. This paper proposes a novel method, SSWPNN (an ensemble of support vector machine and sample-weighted probabilistic neural network), based on downsampling and an ensemble of different classifiers, in view of the imbalance of zinc-binding sites in proteins. Multiple random downsampling techniques without replacement are performed on the whole set, and the support vector machine is trained as the base classifier on each subset to calculate the weights of samples, while the sample-weighted probabilistic neural network is constructed as a strong classifier for prediction. The experimental results showed that our method is superior to other methods not only in the overall prediction performance for the four types of residues but also in the prediction performance for any type of residue. The results of experimental testing on an independent test set collected by the authors in recent years showed that our method achieved better prediction performance than others not only for the four types of residues overall but also for any one type of residue. In addition, the importance of the features selected by the method is analyzed by reducing certain feature to calculate the scores of the performance index. The source code and datasets are available at http://net.jitsec.cn:88/UploadedImages/SSWPNN.rar.

Critical infrastructures, e.g., electricity generation and dispersal networks, chemical processing plants, and gas distribution, are governed and monitored by supervisory control and data acquisition systems (SCADA). Detecting intrusion is a prevalent area of study for numerous years, and several intrusion detection systems have been suggested in the literature for cyber-physical systems and industrial control system (ICS). In recent years, the viruses seismic net, duqu, and flame against ICS attacks have caused tremendous damage to nuclear facilities and critical infrastructure in some countries. These intensified attacks have sounded the alarm for the security of the ICS in many countries. The challenge in constructing an intrusion detection framework is to deal with unbalanced intrusion datasets, i.e. when one class is signified by a lesser amount of instances (minority class). To this end, we outline an approach to deal with this issue and propose an anomaly detection method for the ICS. Our proposed approach uses a hybrid model that takes advantage of the anticipated and consistent nature of communication patterns that occur among ground devices in ICS setups. First, we applied some preprocessing techniques to standardize and scale the data. Second, the dimensionality reduction algorithms are applied to improve the process of anomaly detection. Third, we employed an edited nearest-neighbor rule algorithm to balance the dataset. Fourth, by using the Bloom filter, a signature database is created by noting the system for a specific period lacking the occurrence of abnormalities. Finally, to detect new attacks, we combined our package contents-level detection with another instance-based learner to make a hybrid method for anomaly detection. The experimental results with a real large-scale dataset generated from a gas pipeline SCADA system show that the proposed approach HML-IDS outperforms the benchmark models with an accuracy rate of 97%.

The zinc ion is the second richest metal ion in organisms. The proteins binding to zinc ions have important biological functions. However, few scholars have integrated the existing tools to predict the zinc-binding sites in proteins. To make up for this gap, this paper combines three well-known prediction tools into an improved model called IBayes_Zinc to predict the zinc-binding sites, and utilizes the advantages of the Bayesian method in handling incomplete or partial missing data. Specifically, the prediction scores of three existing sequence-based prediction tools were adopted, and the missing values were padded, forming an integrated classification tool. Then, the probabilities of positive and negative samples were computed and categorized as the class with higher probabilities. Experiments were conducted on a non-redundant training dataset and an independent testing dataset. The results show that our method surpassed the other three methods by nearly 5–13% in Matthew correlation coefficient (MCC) and outperformed the latter in recall and precision. The research findings promote the detection of zinc-binding sites in protein sequence and the identification of metalloprotein functions.

This paper studies a topology identification problem of complex networks with dynamics on different time scales. Using the adaptive synchronization method, some criteria for a successful estimation are obtained. In particular, by regulating the original network to synchronize with an auxiliary chaotic network, this work further explores a way to avoid the precondition of linear independence. When the adaptive controller fails to achieve the outer synchronization, an impulsive control method is used. In the end, we conclude with three numerical simulations. The results obtained in this paper generalize continuous, discrete with arbitrary time step size and mixed cases.

With the rapid increase in the photovoltaic (PV) plants, the real-time operation and maintenance of photovoltaic power generation equipment is very important. The maintenance and dispatching of decentralized power stations is still one of the key issues affecting the operation safety of photovoltaic power stations. However, most of the photovoltaic power stations in China fail to rationally optimize the utilization of resources and time. The current study puts forward effort implementation via genetic algorithm-based multiconstrained optimization methodology. The proposed study optimally overrides the traditional PV plant operation and maintenance dispatching operations with automation and reliability. The proposed study is also applicable to multiple PV plants, multiple maintainers, multipoint departure, different dispatching conditions, and cost considerations. We propose an MOOA algorithm to solve this issue, and we strongly believe that, by defining a suitable fitness function, the convergence speed and optimization ability can be greatly improved, and this study puts a forward step.

Accurate prediction of the generation capacity of photovoltaic systems is fundamental to ensuring the stability of the grid and to performing scheduling arrangements correctly. In view of the temporal defect and the local minimum problem of back-propagation neural network, a forecasting method of power generation based on long short-term memory-back-propagation is proposed. On this basis, the traditional prediction data set is improved. According to the three traditional methods listed in this article, we propose a fourth method to improve the traditional photovoltaic power station short-term power generation prediction. Compared with the traditional method, the long short-term memory-back-propagation neural network based on the improved data set has a lower prediction error. At the same time, a horizontal comparison with the multiple linear regression and the support vector machine shows that the long short-term memory-back-propagation method has several advantages. Based on the long short-term memory-back-propagation neural network, the short-term forecasting method proposed in this article for generating capacity of photovoltaic power stations will provide a basis for dispatching plan and optimizing operation of power grid.

Zinc is an important trace element, and it can be used in combination with proteins to play an important biological function. In this paper, three types of prediction tools based on sequence were studied for the prediction of zinc-binding sites in proteins, and a novel integrated predictor termed meta-zincPrediction is presented. Multiple linear regressions were used in the proposed approach to integrate the results of the three prediction tools, and the parameters were estimated by the least square method until the optimal model was constructed. Using the Zhao_dataset data set, the area under recall-precision curve (AURPC) of our predictor reached nearly 0.9 and increased by 2%-9% compared with the other three predictors; the other performance indexes were also improved. To further demonstrate the robustness and accuracy of the integrated predictor, we tested it on a non-redundant independent test dataset (CollectedDataset). The AURPC increased by 2%-8%. The other three indexes, including the precision, specificity, and MCC, were increased by 5%-8%, 2%-8%, and 4%-12%, respectively, with a recall of 70%. The prediction ability of the meta-zincPrediction was better than the other three predictors, regardless of whether the zinc-binding sites contained four types of residues or a single residue. Our method can be applied to the recognition of zinc binding sites based on sequence information, but will also be useful for inferring protein function and more propitious for the treatment of some diseases.

How to model user distance from multiple social networks is an important challenge. People often simultaneously appear in multiple social networks that can provide complementary services. Thus, knowledge from different social networks can help overcome the data sparseness problem. However, the knowledge cannot be directly obtained due to that they are from different social networks. To solve this problem, we construct an adaptive model to learn user distance in multiple social networks via combining distance metric learning and boosting technologies. The basic idea of our model is to embed related social networks into a potential feature space, while retaining the topologies of social networks. To get the solution to our model, we formulate it as a convex optimization problem. Moreover, we propose an adaptive user distance measure algorithm whose time complexity is linear with the number of the links. We verify the feasibility and effectiveness of our model on the link prediction problem. Experiments on two real large-scale data sets demonstrate that our method outperforms the compared methods. To the best of our knowledge, the joint learning of metric learning with boosting is first studied in multiple social networks.

Data representation and similarity measurement are two basic aspects of similarity detection in time series data mining. In this paper, we present two novel approaches to perform similarity detection efficiently and effectively. One is composed of a new time series representation model and a corresponding similarity measure, which is called fragment alignment distance (FAD); the other applies dynamic time warping to the representation model of FAD and is called FAD_DTW. The new data representation model is based on the trend information of time series, which can provide a concise yet feature-rich representation of time series. FAD is able to align the segments of time series in linear time, which greatly accelerates the similarity detection process. We extensively compare FAD and FAD_DTW with state-of-the-art time series representation models and similarity measures in classification and clustering frameworks. Experimental results from efficiency and effectiveness validations on various data sets demonstrate that FAD and FAD_DTW can achieve fast and accurate similarity detection. In particular, FAD is much faster than the other methods.

The existing social recommendation models mostly utilize various explicit user-generated information. Although there exist a few studies adopting the implicit relationship between users for social recommendation, however, these studies do not consider the deeper social relationship, nor simultaneously take into account two or more deeper relationships between users from different angles. To this end, we propose a new deeper membership and friendship awareness for social recommendation. Specifically, we first calculate the deeper membership similarity between users utilizing the improved Jaccard similarity coefficient and the deeper friendship similarity between users using the proposed two-hop random walk algorithm. Second, the deeper membership similarity and the deeper friendship similarity are combined in a unified way to form a comprehensive deeper social relation similarity. Third, we adopt the matrix factorization method incorporating the deeper membership and the deeper friendship between users as a regularization term for social recommendation, and the corresponding comprehensive deeper social relationship similarity is regarded as the regularization parameter. Experiments on two real-world datasets demonstrate the superiority of the proposed recommendation model.

Considerable research has been done on the complex stock market, however, there is very little systematic work on the impact of crisis on global stock markets. For filling in these gaps, we propose a complex network method, which analyzes the effects of the 2008 global financial crisis on global main stock index from 2005 to 2010. Firstly, we construct three weighted networks. The physics-derived technique of minimum spanning tree is utilized to investigate the networks of three stages. Regional clustering is found in each network. Secondly, we construct three average threshold networks and find the small-world property in the network before and during the crisis. Finally, the dynamical change of the network community structure is deeply analyzed with different threshold. The result indicates that for large thresholds, the network before and after the crisis has a significant community structure. Though this analysis, it would be helpful to investors for making decisions regarding their portfolios or to regulators for monitoring the key nodes to ensure the overall stability of the global stock market.

Social influence analysis is important for many social network applications, including recommendation and cybersecurity analysis. We observe that the influence of community including multiple users outweighs the individual influence. Existing models focus on the individual influence analysis, but few studies estimate the community influence that is ubiquitous in online social network. A major challenge lies in that researchers need to take into account many factors, such as user influence, social trust, and user relationship, to model community-level influence. In this paper, aiming to assess the community-level influence effectively and accurately, we formulate the problem of modeling community influence and construct a community-level influence analysis model. It first eliminates the zombie fans and then calculates the user influence. Next, it calculates the user final influence by combining the user influence and the willingness of diffusing theme information. Finally, it evaluates the community influence by comprehensively studying the user final influence, social trust, and relationship tightness between intrausers of communities. To handle real-world applications, we propose a community-level influence analysis algorithm called CIAA. Empirical studies on a real-world dataset from Sina Weibo demonstrate the superiority of the proposed model.

With the rapid development of mobile telecommunication, voice call spam has become a growing problem in China. Many mobile phone users have become the victim of spam calls and suffered heavy financial loss. Discovering of call spammers can benefit mobile network operators as well as users. Nowadays, the popular method for the task of mining call spammers has been performed by different applications on smartphones. These applications combine manual and automatic methods to detect spammers. Although the results of these client-based solutions are quite satisfying, it is extremely unfortunate that many people still use feature phones, which can not be equipped with third party applications. In this paper, we propose a server-based solution and take a call log file as an example, to analyze the characteristics of mobile call patterns. A time-based graph model and a simple and effective call log rank (CLRank) algorithm with ranking and classification were proposed to find potential call spammers. Compared with existing methods, our model just uses link information, and thus protects user privacy to the maximum extent. Experimental results show that our proposed model can find spammers from call logs automatically, dynamically, and effectively (with 84.5~91.8% of accuracy) without any manual interventions.

Video and image sensors in wireless multimedia sensor networks (WMSNs) have directed view and limited sensing angle. So the methods to solve target coverage problem for traditional sensor networks, which use circle sensing model, are not suitable for WMSNs. Based on the FoV (field of view) sensing model and FoV disk model proposed, how expected multimedia sensor covers the target is defined by the deflection angle between target and the sensor’s current orientation and the distance between target and the sensor. Then target coverage optimization algorithms based on expected coverage value are presented for single-sensor single-target, multisensor single-target, and single-sensor multitargets problems distinguishingly. Selecting the orientation that sensor rotated to cover every target falling in the FoV disk of that sensor for candidate orientations and using genetic algorithm to multisensor multitargets problem, which has NP-complete complexity, then result in the approximated minimum subset of sensors which covers all the targets in networks. Simulation results show the algorithm’s performance and the effect of number of targets on the resulting subset.