To improve conference management and realize self-served control by users,this paper proposes a specific design of tablet-based control software on video conference,incorporating demands of video conference systems in...To improve conference management and realize self-served control by users,this paper proposes a specific design of tablet-based control software on video conference,incorporating demands of video conference systems in State Grid Corporation of China.The software has been designed and implemented with studies on the system structure and key technologies as the bedrock.With in-depth analysis on various operations’frequencies,streamlined interface,and exquisite designs,the software enables users to independently control regular conferences without on-site professional technicians.Moreover,it meets different demands for different scenarios such as for public conference room and normalized management.展开更多
Traditional inspection cameras determine targets and detect defects by capturing images of their light intensity,but in complex environments,the accuracy of inspection may decrease.Information based on polarization of...Traditional inspection cameras determine targets and detect defects by capturing images of their light intensity,but in complex environments,the accuracy of inspection may decrease.Information based on polarization of light can characterize various features of a material,such as the roughness,texture,and refractive index,thus improving classification and recognition of targets.This paper uses a method based on noise template threshold matching to denoise and preprocess polarized images.It also reports on design of an image fusion algorithm,based on NSCT transform,to fuse light intensity images and polarized images.The results show that the fused image improves both subjective and objective evaluation indicators,relative to the source image,and can better preserve edge information and help to improve the accuracy of target recognition.This study provides a reference for the comprehensive application of multi-dimensional optical information in power inspection.展开更多
Non-orthogonal multiple access(NOMA), featuring high spectrum efficiency, massive connectivity and low latency, holds immense potential to be a novel multi-access technique in fifth-generation(5G) communication. Succe...Non-orthogonal multiple access(NOMA), featuring high spectrum efficiency, massive connectivity and low latency, holds immense potential to be a novel multi-access technique in fifth-generation(5G) communication. Successive interference cancellation(SIC) is proved to be an effective method to detect the NOMA signal by ordering the power of received signals and then decoding them. However, the error accumulation effect referred to as error propagation is an inevitable problem. In this paper,we propose a convolutional neural networks(CNNs) approach to restore the desired signal impaired by the multiple input multiple output(MIMO) channel. Especially in the uplink NOMA scenario,the proposed method can decode multiple users' information in a cluster instantaneously without any traditional communication signal processing steps. Simulation experiments are conducted in the Rayleigh channel and the results demonstrate that the error performance of the proposed learning system outperforms that of the classic SIC detection. Consequently, deep learning has disruptive potential to replace the conventional signal detection method.展开更多
Aiming at the problem that the current encrypted traffic classification methods only use the single network framework such as convolutional neural network(CNN),recurrent neural network(RNN),and stacked autoencoder(SAE...Aiming at the problem that the current encrypted traffic classification methods only use the single network framework such as convolutional neural network(CNN),recurrent neural network(RNN),and stacked autoencoder(SAE),and only construct a shallow network to extract features,which leads to the low accuracy of encrypted traffic classification,an encrypted traffic classification framework based on the fusion of vision transformer and temporal features was proposed.Bottleneck transformer network(BoTNet)was used to extract spatial features and bi-directional long short-term memory(BiLSTM)was used to extract temporal features.After the two sub-networks are parallelized,the feature fusion method of early fusion was used in the framework to perform feature fusion.Finally,the encrypted traffic was identified through the fused features.The experimental results show that the BiLSTM and BoTNet fusion transformer(BTFT)model can enhance the performance of encrypted traffic classification by fusing multi-dimensional features.The accuracy rate of a virtual private network(VPN)and non-VPN binary classification is 99.9%,and the accuracy rate of fine-grained encrypted traffic twelve-classification can also reach 97%.展开更多
Network attacks evolved from single-step and simple attacks to complex multistep attacks.Current methods of multistep attack detection usually match multistep attacks from intrusion detection systems(IDS)alarms based ...Network attacks evolved from single-step and simple attacks to complex multistep attacks.Current methods of multistep attack detection usually match multistep attacks from intrusion detection systems(IDS)alarms based on the correlation between attack steps.However,IDS has false negatives and false positives,which leads to incomplete or incorrect multistep attacks.Association based on simple similarity is difficult to obtain an accurate attack cluster,while association based on prior knowledge such as attack graphs is difficult to guarantee a complete attack knowledge base.To solve the above problems,a heuristic multistep attack scenarios construction method based on the kill chain(HMASCKC)model was proposed.The attack model graph can be obtained from dual data sources and heuristic multistep attack scenarios can be obtained through graph matching.The model graph of the attack and the predicted value of the next attack are obtained by calculating the matching value.And according to the purpose of the multistep attack,the kill chain model is used to define the initial multistep attack model,which is used as the initial graph for graph matching.Experimental results show that HMASCKC model can better fit the multistep attack behavior,the effect has some advantages over the longest common subsequence(LCS)algorithm,which can close to or match the prediction error of judge evaluation of attack intension(JEAN)system.The method can make multistep attack model matching for unknown attacks,so it has some advantages in practical application.展开更多
Facial recognition has become the most common identity authentication technologies. However, problems such as uneven light and occluded faces have increased the hardness of liveness detection. Nevertheless, there are ...Facial recognition has become the most common identity authentication technologies. However, problems such as uneven light and occluded faces have increased the hardness of liveness detection. Nevertheless, there are a few pieces of research on face liveness detection under occlusion conditions. This paper designs a face recognition technique suitable for different degrees of facial occlusion, which employs the facial datasets of near-infrared(NIR) images and visible(VIS) light images to examine the single-modality detection accuracy rate(experimental control group) and the corresponding high-dimensional features through the residual network(ResNet). Based on the idea of data fusion, we propose two feature fusion methods. The two methods extract and fuse the data of one and two convolutional layers from two single-modality detectors respectively. The fusion of high-dimensional features apply a new ResNet to get the dual-modality detection accuracy. And then, a new ResNet is applied to test the accuracy of dual-modality detection. The experimental results show that the dual-modality face liveness detection model improves face live detection accuracy and robustness compared with the single-modality. The fusion of two-layer features from the single-modality detector can also improve face detection accuracy by utilizing the above-mentioned dual-modality detector, and it doesn’t increase the algorithm’s complexity.展开更多
Anomaly detection is an important problem in various research and application fields.Researchers design reliable schemes to provide solutions for effectively detecting anomaly points.Most of the existing anomaly detec...Anomaly detection is an important problem in various research and application fields.Researchers design reliable schemes to provide solutions for effectively detecting anomaly points.Most of the existing anomaly detection schemes are unsupervised methods,such as anomaly detection methods based on density,distance and clustering.In total,unsupervised anomaly detection methods have many limitations.For example,they cannot be well combined with prior knowledge in some anomaly detection tasks.For some nonlinear anomaly detection tasks,the modeling is complex and faces dimensional disasters,which are greatly affected by noise.Sometimes it is difficult to find abnormal events that users are interested in,and users need to customize model parameters before detection.With the wide application of deep learning technology,it has a good modeling ability to solve linear and nonlinear data relationships,but the application of deep learning technology in the field of anomaly detection has many challenges.If we regard exceptions as a supervised problem,exceptions are a few,and we usually face the problem of too few labels.To obtain a model that performs well in the anomaly detection task,it requires a high initial training set.Therefore,to solve the above problems,this paper proposes a supervised learning method with manual participation.We introduce the integrated learning model and train a supervised anomaly detection model with strong stability and high accuracy through active learning technology.In addition,this paper adopts certain strategies to maximize the accuracy of anomaly detection and minimize the cost of manual labeling.In the experimental link,we will show that our method is better than some traditional anomaly detection algorithms.展开更多
Most of the current methods for anomaly detection in time series are unsupervised.However,unsupervised learning assumes the distribution of the data and cannot obtain satisfactory results in some scenarios.In this pap...Most of the current methods for anomaly detection in time series are unsupervised.However,unsupervised learning assumes the distribution of the data and cannot obtain satisfactory results in some scenarios.In this paper,we design a semisupervised time series anomaly detection algorithm based on metric learning.The algorithm model mines the features in the time series from the perspectives of the time domain and frequency domain.Furthermore,we design a loss function for anomaly detection.Different from the two-class loss function,in the scenario of the loss function we designed,the normal data will be clustered and distributed in the embedding space,and the abnormal data will be far from the normal data distribution.Furthermore,we extend our designed metric learning model to a semisupervised learning model,extending the labeled dataset with the unlabeled dataset by setting different confidence levels.We conduct experiments on different public datasets and compare them with commonly used time series anomaly detection algorithms.The results show that our model has a good effect.At the same time the semisupervised setting does improve the accuracy of model detection.展开更多
The rise of big data has brought various challenges and revolutions to many fields.Even though its development in many industries has gradually become perfect or even mature,its application and development in complex ...The rise of big data has brought various challenges and revolutions to many fields.Even though its development in many industries has gradually become perfect or even mature,its application and development in complex industrial scenarios is still in its infancy.We run research on single-dimensional time series point anomaly detection based on unsupervised learning:Unlike periodic time series,aperiodic or weakly periodic time series in industrial scenarios are more common.Considering the need for online real-time monitoring,we need to solve the problem of point anomaly detection of oil chromatographic characteristic gases.Thus,we propose a sliding window-based method for the unsupervised single-dimensional time series point anomaly detection problem called the confidence interval radius slope method(CIRS).CIRS is a fusion of knowledge-driven and data-driven methods to realize online real-time monitoring of possible data quality problems.From the experimental results,CIRS has obtained higher PR values than other unsupervised methods by the subject data.展开更多
文摘To improve conference management and realize self-served control by users,this paper proposes a specific design of tablet-based control software on video conference,incorporating demands of video conference systems in State Grid Corporation of China.The software has been designed and implemented with studies on the system structure and key technologies as the bedrock.With in-depth analysis on various operations’frequencies,streamlined interface,and exquisite designs,the software enables users to independently control regular conferences without on-site professional technicians.Moreover,it meets different demands for different scenarios such as for public conference room and normalized management.
基金supported by the project“Research on enhancement and recognition technology of industrial video in power grid production under all-weather environment based on multi-dimensional optical feature fusion and pulse calculation(5700-202325308A-1-1-ZN)”of the State Grid Corporation of China.
文摘Traditional inspection cameras determine targets and detect defects by capturing images of their light intensity,but in complex environments,the accuracy of inspection may decrease.Information based on polarization of light can characterize various features of a material,such as the roughness,texture,and refractive index,thus improving classification and recognition of targets.This paper uses a method based on noise template threshold matching to denoise and preprocess polarized images.It also reports on design of an image fusion algorithm,based on NSCT transform,to fuse light intensity images and polarized images.The results show that the fused image improves both subjective and objective evaluation indicators,relative to the source image,and can better preserve edge information and help to improve the accuracy of target recognition.This study provides a reference for the comprehensive application of multi-dimensional optical information in power inspection.
基金supported by the National Natural Science Foundation of China (61471021)。
文摘Non-orthogonal multiple access(NOMA), featuring high spectrum efficiency, massive connectivity and low latency, holds immense potential to be a novel multi-access technique in fifth-generation(5G) communication. Successive interference cancellation(SIC) is proved to be an effective method to detect the NOMA signal by ordering the power of received signals and then decoding them. However, the error accumulation effect referred to as error propagation is an inevitable problem. In this paper,we propose a convolutional neural networks(CNNs) approach to restore the desired signal impaired by the multiple input multiple output(MIMO) channel. Especially in the uplink NOMA scenario,the proposed method can decode multiple users' information in a cluster instantaneously without any traditional communication signal processing steps. Simulation experiments are conducted in the Rayleigh channel and the results demonstrate that the error performance of the proposed learning system outperforms that of the classic SIC detection. Consequently, deep learning has disruptive potential to replace the conventional signal detection method.
基金supported by the Science and Technology Project of the Headquarters of State Grid Corporation of China(5700-202152186A-0-0-00)。
文摘Aiming at the problem that the current encrypted traffic classification methods only use the single network framework such as convolutional neural network(CNN),recurrent neural network(RNN),and stacked autoencoder(SAE),and only construct a shallow network to extract features,which leads to the low accuracy of encrypted traffic classification,an encrypted traffic classification framework based on the fusion of vision transformer and temporal features was proposed.Bottleneck transformer network(BoTNet)was used to extract spatial features and bi-directional long short-term memory(BiLSTM)was used to extract temporal features.After the two sub-networks are parallelized,the feature fusion method of early fusion was used in the framework to perform feature fusion.Finally,the encrypted traffic was identified through the fused features.The experimental results show that the BiLSTM and BoTNet fusion transformer(BTFT)model can enhance the performance of encrypted traffic classification by fusing multi-dimensional features.The accuracy rate of a virtual private network(VPN)and non-VPN binary classification is 99.9%,and the accuracy rate of fine-grained encrypted traffic twelve-classification can also reach 97%.
基金supported by the Science and Technology Project of the Headquarters of State Grid Corporation of China(5700-202152186A-0-0-00)。
文摘Network attacks evolved from single-step and simple attacks to complex multistep attacks.Current methods of multistep attack detection usually match multistep attacks from intrusion detection systems(IDS)alarms based on the correlation between attack steps.However,IDS has false negatives and false positives,which leads to incomplete or incorrect multistep attacks.Association based on simple similarity is difficult to obtain an accurate attack cluster,while association based on prior knowledge such as attack graphs is difficult to guarantee a complete attack knowledge base.To solve the above problems,a heuristic multistep attack scenarios construction method based on the kill chain(HMASCKC)model was proposed.The attack model graph can be obtained from dual data sources and heuristic multistep attack scenarios can be obtained through graph matching.The model graph of the attack and the predicted value of the next attack are obtained by calculating the matching value.And according to the purpose of the multistep attack,the kill chain model is used to define the initial multistep attack model,which is used as the initial graph for graph matching.Experimental results show that HMASCKC model can better fit the multistep attack behavior,the effect has some advantages over the longest common subsequence(LCS)algorithm,which can close to or match the prediction error of judge evaluation of attack intension(JEAN)system.The method can make multistep attack model matching for unknown attacks,so it has some advantages in practical application.
基金supported by the Science and Technology Project of State Grid Corporation of China(SGHEXT00YJJS1900050)。
文摘Facial recognition has become the most common identity authentication technologies. However, problems such as uneven light and occluded faces have increased the hardness of liveness detection. Nevertheless, there are a few pieces of research on face liveness detection under occlusion conditions. This paper designs a face recognition technique suitable for different degrees of facial occlusion, which employs the facial datasets of near-infrared(NIR) images and visible(VIS) light images to examine the single-modality detection accuracy rate(experimental control group) and the corresponding high-dimensional features through the residual network(ResNet). Based on the idea of data fusion, we propose two feature fusion methods. The two methods extract and fuse the data of one and two convolutional layers from two single-modality detectors respectively. The fusion of high-dimensional features apply a new ResNet to get the dual-modality detection accuracy. And then, a new ResNet is applied to test the accuracy of dual-modality detection. The experimental results show that the dual-modality face liveness detection model improves face live detection accuracy and robustness compared with the single-modality. The fusion of two-layer features from the single-modality detector can also improve face detection accuracy by utilizing the above-mentioned dual-modality detector, and it doesn’t increase the algorithm’s complexity.
基金supported by the State Grid Research Project“Study on Intelligent Analysis Technology of Abnormal Power Data Quality based on Rule Mining” (5700-202119176A-0-0-00).
文摘Anomaly detection is an important problem in various research and application fields.Researchers design reliable schemes to provide solutions for effectively detecting anomaly points.Most of the existing anomaly detection schemes are unsupervised methods,such as anomaly detection methods based on density,distance and clustering.In total,unsupervised anomaly detection methods have many limitations.For example,they cannot be well combined with prior knowledge in some anomaly detection tasks.For some nonlinear anomaly detection tasks,the modeling is complex and faces dimensional disasters,which are greatly affected by noise.Sometimes it is difficult to find abnormal events that users are interested in,and users need to customize model parameters before detection.With the wide application of deep learning technology,it has a good modeling ability to solve linear and nonlinear data relationships,but the application of deep learning technology in the field of anomaly detection has many challenges.If we regard exceptions as a supervised problem,exceptions are a few,and we usually face the problem of too few labels.To obtain a model that performs well in the anomaly detection task,it requires a high initial training set.Therefore,to solve the above problems,this paper proposes a supervised learning method with manual participation.We introduce the integrated learning model and train a supervised anomaly detection model with strong stability and high accuracy through active learning technology.In addition,this paper adopts certain strategies to maximize the accuracy of anomaly detection and minimize the cost of manual labeling.In the experimental link,we will show that our method is better than some traditional anomaly detection algorithms.
基金supported by State Grid Research Project“Study on Intelligent Analysis Technology of Abnormal Power Data Quality based on Rule Mining” (5700-202119176A-0-0-00).
文摘Most of the current methods for anomaly detection in time series are unsupervised.However,unsupervised learning assumes the distribution of the data and cannot obtain satisfactory results in some scenarios.In this paper,we design a semisupervised time series anomaly detection algorithm based on metric learning.The algorithm model mines the features in the time series from the perspectives of the time domain and frequency domain.Furthermore,we design a loss function for anomaly detection.Different from the two-class loss function,in the scenario of the loss function we designed,the normal data will be clustered and distributed in the embedding space,and the abnormal data will be far from the normal data distribution.Furthermore,we extend our designed metric learning model to a semisupervised learning model,extending the labeled dataset with the unlabeled dataset by setting different confidence levels.We conduct experiments on different public datasets and compare them with commonly used time series anomaly detection algorithms.The results show that our model has a good effect.At the same time the semisupervised setting does improve the accuracy of model detection.
基金supported by State Grid Research Project“Study on Intelligent Analysis Technology of Abnormal Power Data Quality based on Rule Mining” (5700-202119176A-0-0-00).
文摘The rise of big data has brought various challenges and revolutions to many fields.Even though its development in many industries has gradually become perfect or even mature,its application and development in complex industrial scenarios is still in its infancy.We run research on single-dimensional time series point anomaly detection based on unsupervised learning:Unlike periodic time series,aperiodic or weakly periodic time series in industrial scenarios are more common.Considering the need for online real-time monitoring,we need to solve the problem of point anomaly detection of oil chromatographic characteristic gases.Thus,we propose a sliding window-based method for the unsupervised single-dimensional time series point anomaly detection problem called the confidence interval radius slope method(CIRS).CIRS is a fusion of knowledge-driven and data-driven methods to realize online real-time monitoring of possible data quality problems.From the experimental results,CIRS has obtained higher PR values than other unsupervised methods by the subject data.