Email alert
2022 Vol. 44, No. 11
Display Method:
2022, 44(11): 3709-3720.
doi: 10.11999/JEIT211588
Abstract:
The defect detection technology of power equipment based on video image is one of the key technologies to realize intelligent operation and maintenance. It can solve the problems of intelligent identification of external defects in automatic fault diagnosis, active warning and online maintenance of power equipment. Moreover, it is able to reduce the waste of human resources and greatly improve the reliability of system operation and maintenance, thus making up for the shortcomings of traditional protection maintenance mode and providing technical support for the stable operation of power grid. This paper summarizes current typical defect detection algorithms and image processing technology of transmission and transformation equipment based on video images. Additionally, it analyzes the advantages and disadvantages of traditional image processing methods and deep learning methods in the field of power equipment defect detection. Finally, current algorithm development platforms are summarized, and the future development is predicted.
The defect detection technology of power equipment based on video image is one of the key technologies to realize intelligent operation and maintenance. It can solve the problems of intelligent identification of external defects in automatic fault diagnosis, active warning and online maintenance of power equipment. Moreover, it is able to reduce the waste of human resources and greatly improve the reliability of system operation and maintenance, thus making up for the shortcomings of traditional protection maintenance mode and providing technical support for the stable operation of power grid. This paper summarizes current typical defect detection algorithms and image processing technology of transmission and transformation equipment based on video images. Additionally, it analyzes the advantages and disadvantages of traditional image processing methods and deep learning methods in the field of power equipment defect detection. Finally, current algorithm development platforms are summarized, and the future development is predicted.
2022, 44(11): 3721-3733.
doi: 10.11999/JEIT220629
Abstract:
Digital twin power grid aims to build the digital twin of physical power grid for power grid company using the emerging digital twin technology. The three key characteristics of digital twin power grid are summarized as data knowledge hybrid driven, real-time bidirectional interaction, and the mixture and symbiosis of virtual space and physical space. The standard evaluation criteria of digital twin power grid project is discussed. The typical architecture design of digital twin power grid is reviewed. Based on five-dimension digital twin model, a four-layer general reference architecture including physical part of power grid layer, digital twin data layer, digital space of power grid layer and application layer is proposed. These applications of digital twin power grid in system analysis, state evaluation, data prediction, health maintenance, simulation and modeling and other aspects are concluded. The significance and value of the evolution from digital twin power grid to digital twin Energy Internet and Smart Energy System are discussed. Finally, the existing challenging problems of digital twin power grid are summarized from six aspects: data management, model construction, visualization, information and physical security, standard establishment and ecosystem construction.
Digital twin power grid aims to build the digital twin of physical power grid for power grid company using the emerging digital twin technology. The three key characteristics of digital twin power grid are summarized as data knowledge hybrid driven, real-time bidirectional interaction, and the mixture and symbiosis of virtual space and physical space. The standard evaluation criteria of digital twin power grid project is discussed. The typical architecture design of digital twin power grid is reviewed. Based on five-dimension digital twin model, a four-layer general reference architecture including physical part of power grid layer, digital twin data layer, digital space of power grid layer and application layer is proposed. These applications of digital twin power grid in system analysis, state evaluation, data prediction, health maintenance, simulation and modeling and other aspects are concluded. The significance and value of the evolution from digital twin power grid to digital twin Energy Internet and Smart Energy System are discussed. Finally, the existing challenging problems of digital twin power grid are summarized from six aspects: data management, model construction, visualization, information and physical security, standard establishment and ecosystem construction.
2022, 44(11): 3734-3747.
doi: 10.11999/JEIT210975
Abstract:
Battery pack is an important part of the energy system of electric vehicles. Ensuring its safety is of great significance to the intelligent development of electric vehicles and human life and property. Detecting and guaranteeing the safety of battery pack in the energy system has become a research hotspot in the field of power batteries. Neural network is widely used in battery data detection, but the signal processing method based on correlation coefficient is still widely used in battery short circuit fault, and its implementation scheme often has some problems, such as targeting specific objects, requiring specific environment, and poor performance in general use. Based on this, this paper combines the characteristics of correlation coefficient and neural network, a neural network fault detection algorithm for internal short circuit in battery packs based on Three-channel parallel Bidirectional Gating Recurrent Unit (TBi-GRU) is proposed. Firstly, based on Spearman's rank correlation coefficient, the sliding window is combined with dimensionless and standardized multi-dimensional battery pack operating characteristics. Then, the TBi-GRU neural network is trained by using the extracted operating characteristics of the battery in the normal state. Then, based on the trained TBi-GRU model, the operating characteristics of the battery packs under the internal short circuit state are detected, and the condition of the battery string is detected by combining the prediction results with the dynamic thresholds of each channel. Through simulation analysis of ideal conditions and platform verification of actual environment, it is proved that this method can fully combine the strong robustness of Szpilman's rank correlation coefficient and the strong universality of TBI-GRU neural network to identify accurately the battery pack's internal short circuit fault.
Battery pack is an important part of the energy system of electric vehicles. Ensuring its safety is of great significance to the intelligent development of electric vehicles and human life and property. Detecting and guaranteeing the safety of battery pack in the energy system has become a research hotspot in the field of power batteries. Neural network is widely used in battery data detection, but the signal processing method based on correlation coefficient is still widely used in battery short circuit fault, and its implementation scheme often has some problems, such as targeting specific objects, requiring specific environment, and poor performance in general use. Based on this, this paper combines the characteristics of correlation coefficient and neural network, a neural network fault detection algorithm for internal short circuit in battery packs based on Three-channel parallel Bidirectional Gating Recurrent Unit (TBi-GRU) is proposed. Firstly, based on Spearman's rank correlation coefficient, the sliding window is combined with dimensionless and standardized multi-dimensional battery pack operating characteristics. Then, the TBi-GRU neural network is trained by using the extracted operating characteristics of the battery in the normal state. Then, based on the trained TBi-GRU model, the operating characteristics of the battery packs under the internal short circuit state are detected, and the condition of the battery string is detected by combining the prediction results with the dynamic thresholds of each channel. Through simulation analysis of ideal conditions and platform verification of actual environment, it is proved that this method can fully combine the strong robustness of Szpilman's rank correlation coefficient and the strong universality of TBI-GRU neural network to identify accurately the battery pack's internal short circuit fault.
2022, 44(11): 3748-3756.
doi: 10.11999/JEIT220267
Abstract:
Edge computing has become an effective solution for the Internet Of Things (IOT) and the microservice model divides the IOT application into a group of loosely coupled and interdependent fine-grained microservices. Due to the limit resource of edge nodes and concurrent requests compete for container instances, how to generate an appropriate microservice selection scheme for concurrent requests of complex workflow application in mobile edge computing environment is an important problem to be solved. Therefore, a container based microservice selection architecture is established in this paper firstly, and the service delay model and network resource consumption model are constructed to reduce the average delay and network consumption. Secondly, Microservice Selection algorithm based on Priority mechanism and improved Ant Colony (MS-PAC) based on priority mechanism and improved ant colony algorithm is proposed, which uses the task deadline to assign urgent tasks first to ensure the delay, and uses the pheromone mechanism of ant colony algorithm to find the global optimal solution. Experimentation demonstrates that the proposed algorithm can reduce the average delay and network consumption effectively.
Edge computing has become an effective solution for the Internet Of Things (IOT) and the microservice model divides the IOT application into a group of loosely coupled and interdependent fine-grained microservices. Due to the limit resource of edge nodes and concurrent requests compete for container instances, how to generate an appropriate microservice selection scheme for concurrent requests of complex workflow application in mobile edge computing environment is an important problem to be solved. Therefore, a container based microservice selection architecture is established in this paper firstly, and the service delay model and network resource consumption model are constructed to reduce the average delay and network consumption. Secondly, Microservice Selection algorithm based on Priority mechanism and improved Ant Colony (MS-PAC) based on priority mechanism and improved ant colony algorithm is proposed, which uses the task deadline to assign urgent tasks first to ensure the delay, and uses the pheromone mechanism of ant colony algorithm to find the global optimal solution. Experimentation demonstrates that the proposed algorithm can reduce the average delay and network consumption effectively.
2022, 44(11): 3757-3766.
doi: 10.11999/JEIT220200
Abstract:
The inspection of transmission line fittings is an indispensable part of power grid security situation awareness. Focusing on the fact that the current transmission line component defect classification model cannot handle the problem of unlimited data flow in real situations, a transmission line component and its defect classification method based on adversarial continuous learning is proposed. In this paper, continuous learning technology is introduced into the task of transmission line component defect classification, so that the classification model can continuously learn new classification tasks from the infinite growth of data stream while ensuring the classification accuracy, and reduce the consumption of time and resources. By integrating attention mechanism, the ability of the model to extract subtle features is enhanced, the problem of small difference between classification tasks is solved, and the classification accuracy is improved. Focusing on the problem of sorting unknowability in continual learning tasks, a method of sorting based on discrete degree is proposed to achieve the optimal utilization of continual learning classification model. Finally, experiments are carried out on CIFAR-100 public data set and self built data set, and various performances of the model are analyzed and compared. The results show that the proposed method realizes the sustainable learning of component and defect classification task, and alleviates the problem of catastrophic forgetting. The accuracy of classification is improved by 1.43% and 2.25% respectively by integrating attention mechanism and using L3 loss function. The optimal utilization of continuous learning classification model is realized, which lays a solid foundation for power grid security situational awareness.
The inspection of transmission line fittings is an indispensable part of power grid security situation awareness. Focusing on the fact that the current transmission line component defect classification model cannot handle the problem of unlimited data flow in real situations, a transmission line component and its defect classification method based on adversarial continuous learning is proposed. In this paper, continuous learning technology is introduced into the task of transmission line component defect classification, so that the classification model can continuously learn new classification tasks from the infinite growth of data stream while ensuring the classification accuracy, and reduce the consumption of time and resources. By integrating attention mechanism, the ability of the model to extract subtle features is enhanced, the problem of small difference between classification tasks is solved, and the classification accuracy is improved. Focusing on the problem of sorting unknowability in continual learning tasks, a method of sorting based on discrete degree is proposed to achieve the optimal utilization of continual learning classification model. Finally, experiments are carried out on CIFAR-100 public data set and self built data set, and various performances of the model are analyzed and compared. The results show that the proposed method realizes the sustainable learning of component and defect classification task, and alleviates the problem of catastrophic forgetting. The accuracy of classification is improved by 1.43% and 2.25% respectively by integrating attention mechanism and using L3 loss function. The optimal utilization of continuous learning classification model is realized, which lays a solid foundation for power grid security situational awareness.
2022, 44(11): 3767-3776.
doi: 10.11999/JEIT220491
Abstract:
With the increasing scale of domestic power system and the continuous deepening of the reform of domestic power market and carbon emission market, the reasonable arrangement of the unit maintenance scheduling has a more and more important impact on the reliability of the power system and the profits of generator manufacturers from the power market and the carbon emission market. On the other hand, the scale of integer variables and constraints of unit maintenance optimization problem also increases sharply. Considering the above problems, a unit maintenance optimization model considering power system reliability is proposed. In addition, with a method based on the Bayesian optimization proposed, to the solution progress of the model obtains the best branch scoring factor value to accelerate the branch-and-bound solution process in integer programming, which is more suitable for the application of the large-scale power system maintenance models. Moreover, a unit maintenance coordination mechanism considering the carbon emission cost and the predicted electricity price is advanced, which maximizing the generation profits in both electric energy market and carbon emission market during the annual unit maintenance scheduling with the safe operation of the power system. Finally, the effectiveness and the practicability of the above model are verified on the IEEE-118 bus system.
With the increasing scale of domestic power system and the continuous deepening of the reform of domestic power market and carbon emission market, the reasonable arrangement of the unit maintenance scheduling has a more and more important impact on the reliability of the power system and the profits of generator manufacturers from the power market and the carbon emission market. On the other hand, the scale of integer variables and constraints of unit maintenance optimization problem also increases sharply. Considering the above problems, a unit maintenance optimization model considering power system reliability is proposed. In addition, with a method based on the Bayesian optimization proposed, to the solution progress of the model obtains the best branch scoring factor value to accelerate the branch-and-bound solution process in integer programming, which is more suitable for the application of the large-scale power system maintenance models. Moreover, a unit maintenance coordination mechanism considering the carbon emission cost and the predicted electricity price is advanced, which maximizing the generation profits in both electric energy market and carbon emission market during the annual unit maintenance scheduling with the safe operation of the power system. Finally, the effectiveness and the practicability of the above model are verified on the IEEE-118 bus system.
2022, 44(11): 3777-3787.
doi: 10.11999/JEIT220288
Abstract:
Since status lights and instruments of power switchgear have the characteristics of high density and ectopic image, the detection ability of basic features such as target morphology, chromaticity comparison, and lightweight recognition ability in edge image processing technology presents higher requirements. Therefore, a Ghost-Bifpn-YOLOv5m(GB-YOLOv5m) method is proposed. Specifically, the Bi-directional Feature Pyramid Network (BiFPN) structure is adopted to give different weights to the feature layer to transmit more effective feature information. A detection layer scale is added to improve the detection accuracy of the network for small targets and tackle the complication of small target recognition caused by the high-density layout of status lights. The Ghost-Bottleneck structure is employed to replace the complex Bottleneck structure of the original backbone and realize the lightweight of the model, contributing to the formation of favorable conditions for deploying the model at the edge. Additionally, the transmission characteristics of status lights and instruments for limited samples are expanded with the image enhancement technology, and the high-speed convergence is realized through migration learning. The experimental results of 10 kV switchgear demonstrate that the algorithm has high recognition accuracy for 16 categories of cabinet status lights and instruments, with mean Average Precision (mAP) of 97.3% and fps of 37.533. Compared with the YOLOv5m algorithm, the model size is reduced by 37.04%, and mAP is increased by 10.2%, implying that the proposed method possesses a significantly enhanced detection ability for lamp bodies and table bodies, as well as remarkably improved lightweight recognition efficiency, which has certain practical significance for the real-time verification of the power state of the switchgear and the interaction of digital twins.
Since status lights and instruments of power switchgear have the characteristics of high density and ectopic image, the detection ability of basic features such as target morphology, chromaticity comparison, and lightweight recognition ability in edge image processing technology presents higher requirements. Therefore, a Ghost-Bifpn-YOLOv5m(GB-YOLOv5m) method is proposed. Specifically, the Bi-directional Feature Pyramid Network (BiFPN) structure is adopted to give different weights to the feature layer to transmit more effective feature information. A detection layer scale is added to improve the detection accuracy of the network for small targets and tackle the complication of small target recognition caused by the high-density layout of status lights. The Ghost-Bottleneck structure is employed to replace the complex Bottleneck structure of the original backbone and realize the lightweight of the model, contributing to the formation of favorable conditions for deploying the model at the edge. Additionally, the transmission characteristics of status lights and instruments for limited samples are expanded with the image enhancement technology, and the high-speed convergence is realized through migration learning. The experimental results of 10 kV switchgear demonstrate that the algorithm has high recognition accuracy for 16 categories of cabinet status lights and instruments, with mean Average Precision (mAP) of 97.3% and fps of 37.533. Compared with the YOLOv5m algorithm, the model size is reduced by 37.04%, and mAP is increased by 10.2%, implying that the proposed method possesses a significantly enhanced detection ability for lamp bodies and table bodies, as well as remarkably improved lightweight recognition efficiency, which has certain practical significance for the real-time verification of the power state of the switchgear and the interaction of digital twins.
2022, 44(11): 3788-3795.
doi: 10.11999/JEIT220565
Abstract:
Considering the diversity of power services in smart grid, the reliability requirements of different levels of services are analyzed. A link-failure model is built, and an availability-oriented routing planning algorithm is designed based on the built model. The feasibility of the proposed algorithm, in terms of the network blocking rate and resource utilization, is validated by comparing with the classical link-failure routing algorithms. The classical link-failure routing algorithms ignore the diversity of power services, and process all service requests of the access network indiscriminately. The classical routing planning constraints are relatively simple, which leads to the high blocking rate. Regarding to the different requirements on the network availability from kinds of services, the routing algorithm proposed in this paper adjusts properly the target function and allocates the resource according to the incoming service. Thus, the network blocking rate is significantly reduced. Meanwhile, the network availability and resource utilization are greatly improved.
Considering the diversity of power services in smart grid, the reliability requirements of different levels of services are analyzed. A link-failure model is built, and an availability-oriented routing planning algorithm is designed based on the built model. The feasibility of the proposed algorithm, in terms of the network blocking rate and resource utilization, is validated by comparing with the classical link-failure routing algorithms. The classical link-failure routing algorithms ignore the diversity of power services, and process all service requests of the access network indiscriminately. The classical routing planning constraints are relatively simple, which leads to the high blocking rate. Regarding to the different requirements on the network availability from kinds of services, the routing algorithm proposed in this paper adjusts properly the target function and allocates the resource according to the incoming service. Thus, the network blocking rate is significantly reduced. Meanwhile, the network availability and resource utilization are greatly improved.
2022, 44(11): 3796-3805.
doi: 10.11999/JEIT220344
Abstract:
Focusing on the problems of low time-frequency resolution and large amount of calculation in the traditional S-Transform, a modified S-Transform based on the optimal Bohman window is proposed. In order to extract accurately and quickly the characteristics of all kinds of disturbance signals, this method obtains the optimal time-frequency resolution by controlling directly the window length and carries out time-frequency analysis only for the main frequency points. Firstly, the optimal length parameter is determined according to the proposed evaluation criteria. Secondly, the sampled signal spectrum is obtained through fast Fourier transform, and then the main frequency points are determined by the dynamic measurement fast algorithm based on the maximum envelope; Then the corresponding optimal length parameter is selected according to the frequency band of the main frequency point for calculation and processing; Finally, the time-frequency feature extraction is completed by calculating the time-frequency amplitude vector based on modulus time-frequency matrix. Simulation analysis and experimental results show that the proposed method has higher time-frequency resolution and shorter calculation time than the traditional S-Transform, and is suitable for the accurate and fast extraction of power quality disturbance signal features.
Focusing on the problems of low time-frequency resolution and large amount of calculation in the traditional S-Transform, a modified S-Transform based on the optimal Bohman window is proposed. In order to extract accurately and quickly the characteristics of all kinds of disturbance signals, this method obtains the optimal time-frequency resolution by controlling directly the window length and carries out time-frequency analysis only for the main frequency points. Firstly, the optimal length parameter is determined according to the proposed evaluation criteria. Secondly, the sampled signal spectrum is obtained through fast Fourier transform, and then the main frequency points are determined by the dynamic measurement fast algorithm based on the maximum envelope; Then the corresponding optimal length parameter is selected according to the frequency band of the main frequency point for calculation and processing; Finally, the time-frequency feature extraction is completed by calculating the time-frequency amplitude vector based on modulus time-frequency matrix. Simulation analysis and experimental results show that the proposed method has higher time-frequency resolution and shorter calculation time than the traditional S-Transform, and is suitable for the accurate and fast extraction of power quality disturbance signal features.
2022, 44(11): 3806-3814.
doi: 10.11999/JEIT220866
Abstract:
In order to achieve the flexible operation of microgrids, distributed cooperative control is always implemented to manage the distributed renewable energy resources within microgrids due to its better flexibility reliability and scalability. However, the time-triggered traditional distributed control strategies lead to a great waste of the communication resources of distributed generators’ local controller, and hence reduce the efficiency of microgrids’ operation. To this end, a distributed event-triggered control-based frequency secondary control method in microgrids under directed communication network is proposed. By designing an event-triggered mechanism for the active power sharing control under directed communication network and designing a local controller for the frequency restoration control, the secondary control of islanded microgrids can be achieved with greatly reduced communication resources. It is demonstrated by the theoretical analysis that the designed control method does not exist Zeno behavior. The simulation results show the effectiveness and superiority of the proposed control method.
In order to achieve the flexible operation of microgrids, distributed cooperative control is always implemented to manage the distributed renewable energy resources within microgrids due to its better flexibility reliability and scalability. However, the time-triggered traditional distributed control strategies lead to a great waste of the communication resources of distributed generators’ local controller, and hence reduce the efficiency of microgrids’ operation. To this end, a distributed event-triggered control-based frequency secondary control method in microgrids under directed communication network is proposed. By designing an event-triggered mechanism for the active power sharing control under directed communication network and designing a local controller for the frequency restoration control, the secondary control of islanded microgrids can be achieved with greatly reduced communication resources. It is demonstrated by the theoretical analysis that the designed control method does not exist Zeno behavior. The simulation results show the effectiveness and superiority of the proposed control method.
2022, 44(11): 3815-3824.
doi: 10.11999/JEIT220241
Abstract:
Deep learning has been widely applied to the field of indoor personnel detection. However, the traditional convolutional neural networks have a high complexity and require the support of highly computational GPU. It is difficult to accomplish the implementation in the embedded devices. For the above problems, a lightweight network model based on improved YOLOv4-tiny network is proposed for indoor personnel detection. Firstly, an improved Ghost convolution feature extraction module is designed to reduce effectively the model complexity. Simultaneously, to reduce network parameters, a depth-wise separable convolution with channel shuffle mechanism is adopted in this paper. Secondly, a multi-scale dilated convolution module is developed in this paper to obtain more discriminative feature information, which combines the improved dilated space pyramid pooling module and the attention mechanism with location information for effective feature fusion, thereby improving inference accuracy and inference speed, simultaneously. The experiments on multiple datasets and hardware platforms show that the proposed model is superior to the original YOLOv4-tiny network in terms of accuracy, speed, model parameters and volume. Therefore, the proposed model is more suitable for deployment in resource-limited embedded devices.
Deep learning has been widely applied to the field of indoor personnel detection. However, the traditional convolutional neural networks have a high complexity and require the support of highly computational GPU. It is difficult to accomplish the implementation in the embedded devices. For the above problems, a lightweight network model based on improved YOLOv4-tiny network is proposed for indoor personnel detection. Firstly, an improved Ghost convolution feature extraction module is designed to reduce effectively the model complexity. Simultaneously, to reduce network parameters, a depth-wise separable convolution with channel shuffle mechanism is adopted in this paper. Secondly, a multi-scale dilated convolution module is developed in this paper to obtain more discriminative feature information, which combines the improved dilated space pyramid pooling module and the attention mechanism with location information for effective feature fusion, thereby improving inference accuracy and inference speed, simultaneously. The experiments on multiple datasets and hardware platforms show that the proposed model is superior to the original YOLOv4-tiny network in terms of accuracy, speed, model parameters and volume. Therefore, the proposed model is more suitable for deployment in resource-limited embedded devices.
2022, 44(11): 3825-3832.
doi: 10.11999/JEIT220096
Abstract:
Future smart grid will incorporate an increasing number of Distributed Energy Resources (DERs), which have the potential to enhance the system energy efficiency, economics, resilience, and sustainability. However, The DERs, dominated by wind power and photovoltaic power generation, would lead to many problems for a power system with large-scale DERs integrated due to their inherent fluctuation characteristic. Therefore, quantitatively evaluating the fluctuation level of the DERs’ power is of great importance for modern power system. To this end, by defining time window, using envelope and Lebesgue integration theory, the fluctuation quantitative index of DERs’ power—fluctuation rate—is defined by extracting the fluctuations of high frequency information and trends of the DER’s output power time series. The validity of the fluctuation rate for measuring the fluctuation of DERs’ power is validated by testing the fluctuation, smoothing effects of wind power and conducting comparative analysis with prediction error and existing fluctuation index.
Future smart grid will incorporate an increasing number of Distributed Energy Resources (DERs), which have the potential to enhance the system energy efficiency, economics, resilience, and sustainability. However, The DERs, dominated by wind power and photovoltaic power generation, would lead to many problems for a power system with large-scale DERs integrated due to their inherent fluctuation characteristic. Therefore, quantitatively evaluating the fluctuation level of the DERs’ power is of great importance for modern power system. To this end, by defining time window, using envelope and Lebesgue integration theory, the fluctuation quantitative index of DERs’ power—fluctuation rate—is defined by extracting the fluctuations of high frequency information and trends of the DER’s output power time series. The validity of the fluctuation rate for measuring the fluctuation of DERs’ power is validated by testing the fluctuation, smoothing effects of wind power and conducting comparative analysis with prediction error and existing fluctuation index.
2022, 44(11): 3833-3841.
doi: 10.11999/JEIT211136
Abstract:
In a large-scale Weapon-Target Assignment (WTA) problem, the explored solution space becomes enormous due to the curse of dimensionality, and it causes low-efficiency in searching optimization solution. For solving this problem effectively, a WTA optimization approach based on multi-attribute decision-making and Deep Q-Network (DQN) is proposed. Firstly, a threat-assessment model for attacking missiles is built based on the approach of Analytic Hierarchy Process (AHP). Meanwhile, an entropy method, used for evaluating the differences of target attributes, is introduced, to increase objective in computing threat-assessment results. Then, an assignment criterion of maximum intercept probability is designed based on assess results, and a multi-steps WTA decision model is built in DQN frame. A uniform experience sampling strategy is designed, making sure that each target type of assignment experience has the same probability to be selected. Furthermore, for balancing the DQN convergence speed and global optimum, a reward function that combines local and global rewards is designed. Lastly, simulation results shows that the proposed WTA approach has the advantage in solving large-scale WTA problem fast and effectively, compared with the general heuristic approach. Also, it presents the robust performance for WTA scenario elements variation.
In a large-scale Weapon-Target Assignment (WTA) problem, the explored solution space becomes enormous due to the curse of dimensionality, and it causes low-efficiency in searching optimization solution. For solving this problem effectively, a WTA optimization approach based on multi-attribute decision-making and Deep Q-Network (DQN) is proposed. Firstly, a threat-assessment model for attacking missiles is built based on the approach of Analytic Hierarchy Process (AHP). Meanwhile, an entropy method, used for evaluating the differences of target attributes, is introduced, to increase objective in computing threat-assessment results. Then, an assignment criterion of maximum intercept probability is designed based on assess results, and a multi-steps WTA decision model is built in DQN frame. A uniform experience sampling strategy is designed, making sure that each target type of assignment experience has the same probability to be selected. Furthermore, for balancing the DQN convergence speed and global optimum, a reward function that combines local and global rewards is designed. Lastly, simulation results shows that the proposed WTA approach has the advantage in solving large-scale WTA problem fast and effectively, compared with the general heuristic approach. Also, it presents the robust performance for WTA scenario elements variation.
2022, 44(11): 3842-3849.
doi: 10.11999/JEIT210965
Abstract:
In this article, the unknown and dynamic interference in the wireless communication environment is studied. Jointly considering communication channel access and transmit power control, a fast reinforcement learning strategy is proposed to ensure reliable communication at the transceivers. The interference avoidance problem is firstly modeled as a Markov decision process to lower the transmission power of the system and reduce the number of channel switching while ensuring the communication quality. Subsequently, a Win or Learn Fast Policy Hill-Climbing (WoLF-PHC) learning method is proposed to avoid rapidly interference. Simulation results show that the anti-interference performance and convergence speed of the proposed WoLF-PHC algorithm are superior to the traditional random selection method and Q learning algorithm under different interference situations.
In this article, the unknown and dynamic interference in the wireless communication environment is studied. Jointly considering communication channel access and transmit power control, a fast reinforcement learning strategy is proposed to ensure reliable communication at the transceivers. The interference avoidance problem is firstly modeled as a Markov decision process to lower the transmission power of the system and reduce the number of channel switching while ensuring the communication quality. Subsequently, a Win or Learn Fast Policy Hill-Climbing (WoLF-PHC) learning method is proposed to avoid rapidly interference. Simulation results show that the anti-interference performance and convergence speed of the proposed WoLF-PHC algorithm are superior to the traditional random selection method and Q learning algorithm under different interference situations.
2022, 44(11): 3850-3857.
doi: 10.11999/JEIT210962
Abstract:
With the wide application of Unmanned Aerial Vehicle (UAV), the UAV-assisted Internet of Things (IoT) data collection architecture has expanded IoT’s application scope, which is especially suitable for extreme scenarios like military battlefields or disaster rescue. This paper proposes a UAV trajectory planning algorithm based on Deep Q-Network (DQN) framework for the above scenarios. The proposed algorithm takes the Age of Information (AoI) of collected data in a UAV’s flight cycle as the optimization goal to maintain data freshness. The simulation results show that this algorithm can effectively reduce the average AoI of the collected data. Compared with the random algorithm, the greedy algorithm based on the maximum AoI, the shortest path algorithm and the AoI-based Trajectory Planning (ATP) algorithm, the proposed algorithm can reduce AoI by about 81%, 67%, 56% and 39%, respectively. This paper has realized the efficient and low-latency data collection in the UAV-assisted IoT system.
With the wide application of Unmanned Aerial Vehicle (UAV), the UAV-assisted Internet of Things (IoT) data collection architecture has expanded IoT’s application scope, which is especially suitable for extreme scenarios like military battlefields or disaster rescue. This paper proposes a UAV trajectory planning algorithm based on Deep Q-Network (DQN) framework for the above scenarios. The proposed algorithm takes the Age of Information (AoI) of collected data in a UAV’s flight cycle as the optimization goal to maintain data freshness. The simulation results show that this algorithm can effectively reduce the average AoI of the collected data. Compared with the random algorithm, the greedy algorithm based on the maximum AoI, the shortest path algorithm and the AoI-based Trajectory Planning (ATP) algorithm, the proposed algorithm can reduce AoI by about 81%, 67%, 56% and 39%, respectively. This paper has realized the efficient and low-latency data collection in the UAV-assisted IoT system.
2022, 44(11): 3858-3865.
doi: 10.11999/JEIT210832
Abstract:
For the Unmanned Aerial Vehicle (UAV)-assisted Mobile Edge Computing (MEC) system, it is proposed to balance the energy consumption of UAV and ground equipment by adding a weight factor to the energy consumption of ground equipment, considering that their energy consumption is not in the same order of magnitude. At the same time, to meet the requirements of ground equipment tasks, the weighted energy consumption of UAV and ground equipment are minimized by joint optimization of UAV trajectory and system resource allocation. The formulated problem is highly non-convex, thus an alternating optimization based two-stage resource allocation optimization scheme is proposed to solve it. In the first stage, given the unloading power of the ground equipment, the Successive Convex Approximation (SCA) method is used to solve the UAV trajectory optimization, Central Processing Unit (CPU) frequency resource allocation and unloading time allocation. In the second stage, the unloading power allocation of ground equipment is optimized. Such two-stage alternating and iterative optimization is used to find the sub-optimal solution of the original problem. The effectiveness of the proposed scheme in reducing system energy consumption is verified by simulation results.
For the Unmanned Aerial Vehicle (UAV)-assisted Mobile Edge Computing (MEC) system, it is proposed to balance the energy consumption of UAV and ground equipment by adding a weight factor to the energy consumption of ground equipment, considering that their energy consumption is not in the same order of magnitude. At the same time, to meet the requirements of ground equipment tasks, the weighted energy consumption of UAV and ground equipment are minimized by joint optimization of UAV trajectory and system resource allocation. The formulated problem is highly non-convex, thus an alternating optimization based two-stage resource allocation optimization scheme is proposed to solve it. In the first stage, given the unloading power of the ground equipment, the Successive Convex Approximation (SCA) method is used to solve the UAV trajectory optimization, Central Processing Unit (CPU) frequency resource allocation and unloading time allocation. In the second stage, the unloading power allocation of ground equipment is optimized. Such two-stage alternating and iterative optimization is used to find the sub-optimal solution of the original problem. The effectiveness of the proposed scheme in reducing system energy consumption is verified by simulation results.
2022, 44(11): 3866-3873.
doi: 10.11999/JEIT220663
Abstract:
Recently, Reconfigurable Intelligent Surface (RIS) has attracted a lot of attention from both academia and industry as a new revolutionary technology. With the increase of communication frequency and the RIS elements, the operating conditions of RIS-assisted wireless communication are gradually approaching the near field radiation pattern of antennas, not just the existence of far field radiation in the traditional sense. Considering either far field or near field, the transmission characteristics of RIS-assisted wireless communication can not be portrayed accurately by the channel model, which results in a loss in performance. In order to solve the problem, a hybrid near-far field channel model is established for massive RIS-assisted communication in this paper by introducing a weighting factor. The gain, the loss and the robustness analysis of the hybrid system are derived in this paper. It is indicated that the hybrid model brings significant improvement in the gain of system and the robustness of model according to the simulation results.
Recently, Reconfigurable Intelligent Surface (RIS) has attracted a lot of attention from both academia and industry as a new revolutionary technology. With the increase of communication frequency and the RIS elements, the operating conditions of RIS-assisted wireless communication are gradually approaching the near field radiation pattern of antennas, not just the existence of far field radiation in the traditional sense. Considering either far field or near field, the transmission characteristics of RIS-assisted wireless communication can not be portrayed accurately by the channel model, which results in a loss in performance. In order to solve the problem, a hybrid near-far field channel model is established for massive RIS-assisted communication in this paper by introducing a weighting factor. The gain, the loss and the robustness analysis of the hybrid system are derived in this paper. It is indicated that the hybrid model brings significant improvement in the gain of system and the robustness of model according to the simulation results.
2022, 44(11): 3874-3881.
doi: 10.11999/JEIT220110
Abstract:
In-Band Full Duplex (IBFD) technology can effectively improve the spectral efficiency of wireless communication system, which has attracted extensive attention in recent years. However, the linear and nonlinear self interference caused by simultaneous transmission and reception brings great challenges to IBFD. The traditional nonlinear self interference cancellation is mainly based on polynomial model and Deep Neural Network(DNN). Polynomial-model-based method has the risk of deterioration of self-interference effect caused by model mismatch, while DNN-based method can not deal with the unique characteristics of space-frequency correlation and time correlation of high-dimensional data. Based on Convolution Long short-term memory Deep Neural Network(CLDNN), two network structures for reconstructing self-interference signals, Two-Dimensional CLDNN(2D-CLDNN) and Complex-Value-CLDNN(CV-CLDNN), are designed by introducing three-dimensional tensor in the input layer and setting complex convolution layer structure in the convolution layer, which makes full use of the advantages of local perception and weight sharing of convolutional neural network, so as to learn more abstract low-dimensional features from high-dimensional features, so as to improve the effect of self-interference cancellation. The evaluation results of the data obtained in the actual scene show that, when the memory length M of power amplifier and the multipath length L of self interference channel meet M+L=13, through a total of 60 training epochs, the structure proposed in this paper can achieve at least 26% improvement in nonlinear self-interference cancellation compared to the traditional DNN method, the training period is also significantly reduced.
In-Band Full Duplex (IBFD) technology can effectively improve the spectral efficiency of wireless communication system, which has attracted extensive attention in recent years. However, the linear and nonlinear self interference caused by simultaneous transmission and reception brings great challenges to IBFD. The traditional nonlinear self interference cancellation is mainly based on polynomial model and Deep Neural Network(DNN). Polynomial-model-based method has the risk of deterioration of self-interference effect caused by model mismatch, while DNN-based method can not deal with the unique characteristics of space-frequency correlation and time correlation of high-dimensional data. Based on Convolution Long short-term memory Deep Neural Network(CLDNN), two network structures for reconstructing self-interference signals, Two-Dimensional CLDNN(2D-CLDNN) and Complex-Value-CLDNN(CV-CLDNN), are designed by introducing three-dimensional tensor in the input layer and setting complex convolution layer structure in the convolution layer, which makes full use of the advantages of local perception and weight sharing of convolutional neural network, so as to learn more abstract low-dimensional features from high-dimensional features, so as to improve the effect of self-interference cancellation. The evaluation results of the data obtained in the actual scene show that, when the memory length M of power amplifier and the multipath length L of self interference channel meet M+L=13, through a total of 60 training epochs, the structure proposed in this paper can achieve at least 26% improvement in nonlinear self-interference cancellation compared to the traditional DNN method, the training period is also significantly reduced.
2022, 44(11): 3882-3890.
doi: 10.11999/JEIT210853
Abstract:
Video Synthetic Aperture Radar (VideoSAR) can acquire high frame rate image sequence of observation scene. Moving target state can be detected by using the shadow formed by ground moving targets such as vehicles in the image sequence. This method has the advantages of high positioning accuracy, high detection probability and no minimum detectable speed limit. VideoSAR has the special image features,such as sharp change of moving target shadow, low signal-to-noise ratio and Doppler smear interference. Taking full advantage of the spatial and temporal information of frame images, three technologies including VideoSAR data preprocessing, moving target shadow detection and VideoSAR multi-target tracking method are studied. The effectiveness of the proposed method is verified by the real data whole processing results.
Video Synthetic Aperture Radar (VideoSAR) can acquire high frame rate image sequence of observation scene. Moving target state can be detected by using the shadow formed by ground moving targets such as vehicles in the image sequence. This method has the advantages of high positioning accuracy, high detection probability and no minimum detectable speed limit. VideoSAR has the special image features,such as sharp change of moving target shadow, low signal-to-noise ratio and Doppler smear interference. Taking full advantage of the spatial and temporal information of frame images, three technologies including VideoSAR data preprocessing, moving target shadow detection and VideoSAR multi-target tracking method are studied. The effectiveness of the proposed method is verified by the real data whole processing results.
2022, 44(11): 3891-3899.
doi: 10.11999/JEIT210871
Abstract:
Considering solving the problem that the application of conventional feature recognition methods is limited and the depth learning method needs a large amount of labeled data to achieve high recognition performance in radar deception jamming recognition, a domain adaptive radar deception jamming recognition method based on depth residual model is proposed to improve the labeling limit. The attention mechanism is integrated to improve further the recognition accuracy. Firstly, after the time-frequency transformation of the radar received signal, the domain adaptation technology based on the idea of countermeasure network is applied to realize the migration recognition from labeled source domain samples to unlabeled target domain samples. Secondly, through the designed spatial channel attention residual module, the network training focuses on the global spatial features and high response channels of the time-frequency image, so as to ignore the areas with low mobility in the time-frequency image and suppress the generation of negative migration. Experimental results on radar deception jamming data sets in different source and target domains show the feasibility and effectiveness of the proposed method.
Considering solving the problem that the application of conventional feature recognition methods is limited and the depth learning method needs a large amount of labeled data to achieve high recognition performance in radar deception jamming recognition, a domain adaptive radar deception jamming recognition method based on depth residual model is proposed to improve the labeling limit. The attention mechanism is integrated to improve further the recognition accuracy. Firstly, after the time-frequency transformation of the radar received signal, the domain adaptation technology based on the idea of countermeasure network is applied to realize the migration recognition from labeled source domain samples to unlabeled target domain samples. Secondly, through the designed spatial channel attention residual module, the network training focuses on the global spatial features and high response channels of the time-frequency image, so as to ignore the areas with low mobility in the time-frequency image and suppress the generation of negative migration. Experimental results on radar deception jamming data sets in different source and target domains show the feasibility and effectiveness of the proposed method.
2022, 44(11): 3900-3909.
doi: 10.11999/JEIT210982
Abstract:
In the surrounding electromagnetic space, there are a large number of interleaved radar pulse trains consisting of pulse sequences with fixed repetition intervals, such as ship-borne navigation radar signals, airborne pulse Doppler radar signals, etc. These pulse trains exist in the form of time segments, and electronic reconnaissance systems are unable to determine their starting and ending time in advance, which makes it difficult to estimate the repetition intervals and deinterleave the pulses of this kind of radar. This paper first analyzes the negative impact of the short duration of such pulse trains on the performance of traditional pulse deinterleaving methods, and then introduces the idea of sliding time window to weaken this impact. Based on this idea, this paper proposes a high-precision estimator of Pulse Repetition Intervals (PRI) and a pulse deinterleaving method. Simulation results verify the performance of the new method on PRI estimation and pulse deinterleaving.
In the surrounding electromagnetic space, there are a large number of interleaved radar pulse trains consisting of pulse sequences with fixed repetition intervals, such as ship-borne navigation radar signals, airborne pulse Doppler radar signals, etc. These pulse trains exist in the form of time segments, and electronic reconnaissance systems are unable to determine their starting and ending time in advance, which makes it difficult to estimate the repetition intervals and deinterleave the pulses of this kind of radar. This paper first analyzes the negative impact of the short duration of such pulse trains on the performance of traditional pulse deinterleaving methods, and then introduces the idea of sliding time window to weaken this impact. Based on this idea, this paper proposes a high-precision estimator of Pulse Repetition Intervals (PRI) and a pulse deinterleaving method. Simulation results verify the performance of the new method on PRI estimation and pulse deinterleaving.
2022, 44(11): 3910-3916.
doi: 10.11999/JEIT210887
Abstract:
The Three-Dimensional(3D) parameter estimation algorithm of the helicopter with constant speed circular flight from the underwater acoustic data with single hydrophone is proposed. Firstly, the helicopter line spectrum is used as the characteristic of the source, and its 3D Doppler propagation model in two-layer air-water medium is established. Then, the parameters estimation for helicopter in 3D space is derived according to the Doppler frequency shift curve, the sound source motion model and the geometric model of sound propagation. Finally, the effectiveness and accuracy of the algorithm for Doppler signal with the alpha stable noise on single hydrophone is verified.
The Three-Dimensional(3D) parameter estimation algorithm of the helicopter with constant speed circular flight from the underwater acoustic data with single hydrophone is proposed. Firstly, the helicopter line spectrum is used as the characteristic of the source, and its 3D Doppler propagation model in two-layer air-water medium is established. Then, the parameters estimation for helicopter in 3D space is derived according to the Doppler frequency shift curve, the sound source motion model and the geometric model of sound propagation. Finally, the effectiveness and accuracy of the algorithm for Doppler signal with the alpha stable noise on single hydrophone is verified.
2022, 44(11): 3917-3930.
doi: 10.11999/JEIT210848
Abstract:
Considering the problems of the existing acoustic target depth classification methods in shallow water, such as limited frequency range and high signal-to-noise ratio requirements, on the premise of effective ranging results, a novel target depth classification algorithm based on new matching variable is proposed. By analyzing the depth distribution characteristics of the mode cross-correlation items, the target depth classification model is established by using the vertical complex acoustic intensity as matching variable. When the receiving depths are different, although the algorithms all use the vertical complex acoustic intensity as matching variable, the mode cross-correlation items that directly affect the depth classification effect are different. According to the different target depth classification requirements, by specifying receiving depths of dual vector sensors, the matching variable selection of the target depth classification model can be optimized, thereby achieving the improvement of the target depth classification algorithm performance. The simulation results indicate that this method is suitable for targets whose frequency excites three modes, so as to expand frequency range of the algorithm. The algorithm can obtain valuable depth classification results in complex ocean waveguide under low Signal-to-Noise Ratio (SNR= 0 dB).
Considering the problems of the existing acoustic target depth classification methods in shallow water, such as limited frequency range and high signal-to-noise ratio requirements, on the premise of effective ranging results, a novel target depth classification algorithm based on new matching variable is proposed. By analyzing the depth distribution characteristics of the mode cross-correlation items, the target depth classification model is established by using the vertical complex acoustic intensity as matching variable. When the receiving depths are different, although the algorithms all use the vertical complex acoustic intensity as matching variable, the mode cross-correlation items that directly affect the depth classification effect are different. According to the different target depth classification requirements, by specifying receiving depths of dual vector sensors, the matching variable selection of the target depth classification model can be optimized, thereby achieving the improvement of the target depth classification algorithm performance. The simulation results indicate that this method is suitable for targets whose frequency excites three modes, so as to expand frequency range of the algorithm. The algorithm can obtain valuable depth classification results in complex ocean waveguide under low Signal-to-Noise Ratio (SNR= 0 dB).
2022, 44(11): 3931-3940.
doi: 10.11999/JEIT210385
Abstract:
Noise may be generated in the process of image capture, transmission or processing. When the image is affected by a large amount of noise, it is difficult for many pedestrian Re-IDentification(ReID) methods to extract pedestrian features with sufficient expressive ability, which shows poor robustness. This paper focuses on the pedestrian re-identification with low quality image. The dual-domain filtering decomposition is proposed to construct triplet, which is used to train metric learning model. The proposed method mainly consists of two parts. Firstly, the distribution characteristics of different image noise in surveillance videos is analyzed and images are enhanced by dual-domain filtering. Secondly, based on the separation effect of dual-domain filtering, a new triplet is proposed. In the training stage, the original image with the low-frequency component, the noise with high-frequency component generated by the dual-domain filtering and the original image are used as the input triplet. So the noise component can be further suppressed by the network. At the same time, the loss function is optimized, and the triple loss and contrast loss are used in combination. Finally, re-ranking is used to expand the sorting table to improve the accuracy of identification. The average Rank-1 on the noisy Market-1501 and CUHK03 datasets are 78.3% and 21.7%, and the mean Average Precision(mAP) is 66.9% and 20.5%. The accuracy loss of Rank-1 before and after adding noise is only 1.9% and 7.8%, which indicates that the model in this paper shows strong robustness in the case of noise.
Noise may be generated in the process of image capture, transmission or processing. When the image is affected by a large amount of noise, it is difficult for many pedestrian Re-IDentification(ReID) methods to extract pedestrian features with sufficient expressive ability, which shows poor robustness. This paper focuses on the pedestrian re-identification with low quality image. The dual-domain filtering decomposition is proposed to construct triplet, which is used to train metric learning model. The proposed method mainly consists of two parts. Firstly, the distribution characteristics of different image noise in surveillance videos is analyzed and images are enhanced by dual-domain filtering. Secondly, based on the separation effect of dual-domain filtering, a new triplet is proposed. In the training stage, the original image with the low-frequency component, the noise with high-frequency component generated by the dual-domain filtering and the original image are used as the input triplet. So the noise component can be further suppressed by the network. At the same time, the loss function is optimized, and the triple loss and contrast loss are used in combination. Finally, re-ranking is used to expand the sorting table to improve the accuracy of identification. The average Rank-1 on the noisy Market-1501 and CUHK03 datasets are 78.3% and 21.7%, and the mean Average Precision(mAP) is 66.9% and 20.5%. The accuracy loss of Rank-1 before and after adding noise is only 1.9% and 7.8%, which indicates that the model in this paper shows strong robustness in the case of noise.
2022, 44(11): 3941-3950.
doi: 10.11999/JEIT210984
Abstract:
Path planning is a step to generate a feasible path for a robot to track along. Locations of the robot are supposed to lie on or at least nearby the planned path, which can thus generate important constraints for robot localization. In this paper, a model, called Path-Induced Location Probability Map (PI-LPM), to exploit such constraint on robot localization is proposed. The proposed PI-LPM model is a Probability Density Function (PDF) over the entire map with the probability to describe the likelihood that the robot is located. The PDF is generated from all the points representing the path by applying the Kernel Density Estimation (KDE) method with each point as a sampling point. Based on the PI-LPM model, a Robot Localization from Planned Path Constraints (RL-PPC) method to enhance robot localization is proposed. In this method, particle filter is applied to fuse the develop PI-LPM model and existing localization methods, where the probability from PI-LPM is an important factor to assign weights to the particles. The proposed method is validated with both simulation and real data. In the experiment, the proposed PI-LPM model is integrated into both GPS and LiDAR based localization systems. Experimental results demonstrate that the RL-PPC method can effectively improve the over-all performance of robot localization.
Path planning is a step to generate a feasible path for a robot to track along. Locations of the robot are supposed to lie on or at least nearby the planned path, which can thus generate important constraints for robot localization. In this paper, a model, called Path-Induced Location Probability Map (PI-LPM), to exploit such constraint on robot localization is proposed. The proposed PI-LPM model is a Probability Density Function (PDF) over the entire map with the probability to describe the likelihood that the robot is located. The PDF is generated from all the points representing the path by applying the Kernel Density Estimation (KDE) method with each point as a sampling point. Based on the PI-LPM model, a Robot Localization from Planned Path Constraints (RL-PPC) method to enhance robot localization is proposed. In this method, particle filter is applied to fuse the develop PI-LPM model and existing localization methods, where the probability from PI-LPM is an important factor to assign weights to the particles. The proposed method is validated with both simulation and real data. In the experiment, the proposed PI-LPM model is integrated into both GPS and LiDAR based localization systems. Experimental results demonstrate that the RL-PPC method can effectively improve the over-all performance of robot localization.
2022, 44(11): 3951-3959.
doi: 10.11999/JEIT210967
Abstract:
Depth completion is a task of estimating dense depth maps from sparse measurements under the guidance of dense RGB images. Dense depth map is critical for object detection, autonomous driving and scene reconstruction. Hence, depth completion of road scenes is a hot research topic at present. In this paper, a novel multi-stage multi-scale lightweight encoder-decoder network for depth completion is proposed. Specifically, the network consists of two sub-encoder-decoder branches named color-guided branch and fine-completion branch. At the encoders, a lightweight multiscale convolution module with channel random mixing is proposed to extract better image features while controlling the parameter amount. At the decoders, a channel-aware mechanism is devised to focus on the important features. Moreover, multi-scale features from the decoder of color-guided branch are fused into the encoder of fine-completion branch to achieve multi-stage multi-scale guidance. Furthermore, an efficient multi-loss strategy is developed for depth completion from coarse to fine in the training process. Experiments demonstrate that the proposed model is relatively lightweight and can achieve superior performance compared with other state-of-the-art methods.
Depth completion is a task of estimating dense depth maps from sparse measurements under the guidance of dense RGB images. Dense depth map is critical for object detection, autonomous driving and scene reconstruction. Hence, depth completion of road scenes is a hot research topic at present. In this paper, a novel multi-stage multi-scale lightweight encoder-decoder network for depth completion is proposed. Specifically, the network consists of two sub-encoder-decoder branches named color-guided branch and fine-completion branch. At the encoders, a lightweight multiscale convolution module with channel random mixing is proposed to extract better image features while controlling the parameter amount. At the decoders, a channel-aware mechanism is devised to focus on the important features. Moreover, multi-scale features from the decoder of color-guided branch are fused into the encoder of fine-completion branch to achieve multi-stage multi-scale guidance. Furthermore, an efficient multi-loss strategy is developed for depth completion from coarse to fine in the training process. Experiments demonstrate that the proposed model is relatively lightweight and can achieve superior performance compared with other state-of-the-art methods.
2022, 44(11): 3960-3966.
doi: 10.11999/JEIT210844
Abstract:
Blind Source Separation (BSS) aims to separate the source signals from the mixed observations without any information about the mixing process and the source signals, which is a major area in the signal processing field. In Underdetermined Blind Source Separation (UBSS), the number of observed signals is less than the number of source signals, and thus UBSS is much closer to reality than the determined/overdetermined BSS. However, the observations are always disturbed by noise, deteriorating the performance of traditional underdetermined blind source separation based on second-order statistics and signal sparsity. Taking the advantage of third-order statistics in dealing with symmetric noise, a novel mixing matrix estimation method based on the third-order statistics of the observations is proposed. Considering the autocorrelations of the sources, a sequence of third-order statistics of the observations corresponding to multiple delays are calculated and stacked into a fourth-order tensor. Then the mixing matrix is estimated via the canonical polyadic decomposition of the fourth-order tensor. Furthermore, the generalized Gaussian distribution is employed to characterize the sources and the expectation-maximum algorithm is utilized to recover the sources. The results from 1000 Monte Carlo experiments demonstrate that the proposed method is robust to the noise. The proposed method archives the normalized mean square error of –20.35 dB and the mean absolute correlation coefficient between the recovered sources and the real ones of 0.84 when the signal to noise ratios equal to 15 dB for the cases with 3×4 mixing matrices. Simulation results demonstrate that the proposed algorithm yields superior performances in comparing with state-of-the-art underdetermined blind source separation methods.
Blind Source Separation (BSS) aims to separate the source signals from the mixed observations without any information about the mixing process and the source signals, which is a major area in the signal processing field. In Underdetermined Blind Source Separation (UBSS), the number of observed signals is less than the number of source signals, and thus UBSS is much closer to reality than the determined/overdetermined BSS. However, the observations are always disturbed by noise, deteriorating the performance of traditional underdetermined blind source separation based on second-order statistics and signal sparsity. Taking the advantage of third-order statistics in dealing with symmetric noise, a novel mixing matrix estimation method based on the third-order statistics of the observations is proposed. Considering the autocorrelations of the sources, a sequence of third-order statistics of the observations corresponding to multiple delays are calculated and stacked into a fourth-order tensor. Then the mixing matrix is estimated via the canonical polyadic decomposition of the fourth-order tensor. Furthermore, the generalized Gaussian distribution is employed to characterize the sources and the expectation-maximum algorithm is utilized to recover the sources. The results from 1000 Monte Carlo experiments demonstrate that the proposed method is robust to the noise. The proposed method archives the normalized mean square error of –20.35 dB and the mean absolute correlation coefficient between the recovered sources and the real ones of 0.84 when the signal to noise ratios equal to 15 dB for the cases with 3×4 mixing matrices. Simulation results demonstrate that the proposed algorithm yields superior performances in comparing with state-of-the-art underdetermined blind source separation methods.
2022, 44(11): 3967-3976.
doi: 10.11999/JEIT210868
Abstract:
In order to improve the accuracy of the single image dehazing method and the detail visibility of its dehazing results, a single image dehazing method based on multi-scale features combined with detail recovery is proposed. Firstly, according to the distribution characteristics and imaging principles of haze in images, the multi-scale feature extraction module and the multi-scale feature fusion module are designed to extract effectively the haze-related multi-scale features in the hazy image and perform nonlinear weighted fusion. Secondly, the end-to-end dehazing network based on the designed multi-scale feature extraction module and multi-scale feature fusion module are constructed, and the preliminary dehazing results are obtained by using this network. Then, a detail recovery network based on image blocking is constructed to extract detail information. Finally, the detail information extracted from the detail recovery network is fused with the preliminary dehazing results obtained from the dehazing network to obtain the final clear dehazed image, which can enhance the visual effect of the dehazing images. The experimental results show that compared with the existing representative image dehazing methods, the proposed method can effectively remove the haze in the synthetic images and the real-world images, and the detailed information of the dehazing results is kept.
In order to improve the accuracy of the single image dehazing method and the detail visibility of its dehazing results, a single image dehazing method based on multi-scale features combined with detail recovery is proposed. Firstly, according to the distribution characteristics and imaging principles of haze in images, the multi-scale feature extraction module and the multi-scale feature fusion module are designed to extract effectively the haze-related multi-scale features in the hazy image and perform nonlinear weighted fusion. Secondly, the end-to-end dehazing network based on the designed multi-scale feature extraction module and multi-scale feature fusion module are constructed, and the preliminary dehazing results are obtained by using this network. Then, a detail recovery network based on image blocking is constructed to extract detail information. Finally, the detail information extracted from the detail recovery network is fused with the preliminary dehazing results obtained from the dehazing network to obtain the final clear dehazed image, which can enhance the visual effect of the dehazing images. The experimental results show that compared with the existing representative image dehazing methods, the proposed method can effectively remove the haze in the synthetic images and the real-world images, and the detailed information of the dehazing results is kept.
2022, 44(11): 3977-3986.
doi: 10.11999/JEIT210823
Abstract:
The integration of view-based 3D model classification and deep learning can effectively improve the classification accuracy. However, current methods consider that the views from different viewpoints of 3D model with same category belong to the same category and ignore the view differences, which makes it difficult for the classifier to learn a reasonable classification surface. To solve this problem, a 3D model classification method based on deep neural network is proposed. The multiple viewpoint groups are set evenly around the 3D model in this method, and the view classifier for each viewpoint group is trained for fully mining the deep information of the 3D model in different viewpoint groups. These classifiers share a feature extraction network, but have their own classification network. In order to extract the discriminative view features, the attention mechanism is added to the feature extraction network; In order to model the views of the non-viewpoint group, additional classes are added to the classification network. In the classification stage, a view selection strategy is first proposed, which can use a small number of views to classify the 3D model and improve classification efficiency. Then a classification strategy is proposed to achieve reliable 3D model classification through classification view. Experimental results on ModelNet10 and ModelNet40 show that the classification accuracy can reach up to 93.6% and 91.0% with only 3 views.
The integration of view-based 3D model classification and deep learning can effectively improve the classification accuracy. However, current methods consider that the views from different viewpoints of 3D model with same category belong to the same category and ignore the view differences, which makes it difficult for the classifier to learn a reasonable classification surface. To solve this problem, a 3D model classification method based on deep neural network is proposed. The multiple viewpoint groups are set evenly around the 3D model in this method, and the view classifier for each viewpoint group is trained for fully mining the deep information of the 3D model in different viewpoint groups. These classifiers share a feature extraction network, but have their own classification network. In order to extract the discriminative view features, the attention mechanism is added to the feature extraction network; In order to model the views of the non-viewpoint group, additional classes are added to the classification network. In the classification stage, a view selection strategy is first proposed, which can use a small number of views to classify the 3D model and improve classification efficiency. Then a classification strategy is proposed to achieve reliable 3D model classification through classification view. Experimental results on ModelNet10 and ModelNet40 show that the classification accuracy can reach up to 93.6% and 91.0% with only 3 views.
2022, 44(11): 3987-3997.
doi: 10.11999/JEIT210854
Abstract:
In the process of target extraction in low Depth Of Field (DOF) image, it is easy to get incomplete target extraction or the background is mistakenly recognized as a target. A low DOF image target extraction method based on Singular Value Difference (SVD) measurement is proposed. Firstly, Gaussian blur is applied to the low DOF image. Taking the current pixel as the center, the image blocks at the same position on the image before and after blur are intercepted by using the sliding window, and singular value decomposition is carried out. Then, the difference feature vector between the two singular values is constructed. Based on this vector, the difference measurement operator is defined to calculate the characteristic intensity value of the corresponding pixel. The feature salient map is obtained by pixel by pixel processing, and the threshold processing is carried out to realize the effective extraction of low DOF image targets. A large number of low DOF images are processed, and compared with several existing methods, the maximum F measure can be increased by 54%, and the average absolute error can be reduced by 76%～87%. The proposed method can completely extract the target and effectively remove the background, and has strong reliability.
In the process of target extraction in low Depth Of Field (DOF) image, it is easy to get incomplete target extraction or the background is mistakenly recognized as a target. A low DOF image target extraction method based on Singular Value Difference (SVD) measurement is proposed. Firstly, Gaussian blur is applied to the low DOF image. Taking the current pixel as the center, the image blocks at the same position on the image before and after blur are intercepted by using the sliding window, and singular value decomposition is carried out. Then, the difference feature vector between the two singular values is constructed. Based on this vector, the difference measurement operator is defined to calculate the characteristic intensity value of the corresponding pixel. The feature salient map is obtained by pixel by pixel processing, and the threshold processing is carried out to realize the effective extraction of low DOF image targets. A large number of low DOF images are processed, and compared with several existing methods, the maximum F measure can be increased by 54%, and the average absolute error can be reduced by 76%～87%. The proposed method can completely extract the target and effectively remove the background, and has strong reliability.
2022, 44(11): 3998-4007.
doi: 10.11999/JEIT210897
Abstract:
Early screening of dysphagia is an important means to reduce the incidence of dysphagia, and accurate identification of Swallowing Events (SE) is a key step in the screening and treatment of dysphagia. Impedance PharyngoGraphy (IPG) is a new non-invasive SE detection method, but the existing IPG technique only detects the impedance amplitude and ignores the equally important phase information. Aiming to comprehensively extracting and intelligently recognizing SE, a Complex Impedance PharyngoGraphy (CIPG) detection method based on integer-period digital lock-in amplifying principle is proposed, and a CIPG measurement system is designed based on FPGA to continuously record the complex impedance (amplitude and phase) information during swallowing process, and an SE intelligent recognition algorithm based on Continuous Wavelet Transform (CWT) and GoogLeNet is designed. A five-SE recognition experiment including drinking water, dry swallowing, eating bread, eating yogurt and coughing is designed. The experimental results show that the SE recognition accuracy is 86.1% when only using impedance amplitude information, and 95.7% when using both impedance amplitude and phase information. The latter SE recognition accuracy is higher than that of other algorithms. This study confirms the effectiveness and superiority of CIPG technology and SE intelligent recognition algorithm, and lays a theoretical and technical foundation for further developing an early screening method of dysphagia based on CIPG.
Early screening of dysphagia is an important means to reduce the incidence of dysphagia, and accurate identification of Swallowing Events (SE) is a key step in the screening and treatment of dysphagia. Impedance PharyngoGraphy (IPG) is a new non-invasive SE detection method, but the existing IPG technique only detects the impedance amplitude and ignores the equally important phase information. Aiming to comprehensively extracting and intelligently recognizing SE, a Complex Impedance PharyngoGraphy (CIPG) detection method based on integer-period digital lock-in amplifying principle is proposed, and a CIPG measurement system is designed based on FPGA to continuously record the complex impedance (amplitude and phase) information during swallowing process, and an SE intelligent recognition algorithm based on Continuous Wavelet Transform (CWT) and GoogLeNet is designed. A five-SE recognition experiment including drinking water, dry swallowing, eating bread, eating yogurt and coughing is designed. The experimental results show that the SE recognition accuracy is 86.1% when only using impedance amplitude information, and 95.7% when using both impedance amplitude and phase information. The latter SE recognition accuracy is higher than that of other algorithms. This study confirms the effectiveness and superiority of CIPG technology and SE intelligent recognition algorithm, and lays a theoretical and technical foundation for further developing an early screening method of dysphagia based on CIPG.
2022, 44(11): 4008-4017.
doi: 10.11999/JEIT211524
Abstract:
Cross-border e-commerce products recommendation has become one of the emerging researching topics in the field of e-commerce. Due to the diversity and complexity of e-commerce product information, the “user-item” correlation matrix is extremely sparse and the cold start problem is prominent. As a result, the traditional collaborative filtering model seems to be malfunctional. Meanwhile, the improved recommendation model based on collaborative filtering or matrix factorization only considers the explicit and implicit feedback information of the users to the products, while ignoring the graph structure information composed of users and items, so that the recommendation performance is difficult to meet the requirements of the platform and users. To tackle these issues, a recommender system of cross-border e-commerce based on heterogeneous graph neural network, named Heterogeneous Graph Neural network Recommender system (HGNR), is proposed in this paper. The model has two significant advantages: (1) the three-part graph is used as input, and high-quality information dissemination and aggregation are carried out on heterogeneous graphs through Graph Convolutional neural Network (GCN); (2) high-quality user and product representation vectors can be obtained, and realize the modeling of the complex interaction between users and products. Experimental results on real cross-border e-commerce order data sets show that HGNR not only owns the superior performance, but also can effectively improve the recommendation accuracy of cold-start users. Compared with nine baseline methods for recommendation, HGNR achieves improvements of at least 3.33%, 0.91%, and 0.54% on evaluation metrics of HitRate@10, Item-coverage@10 and MRR@10.
Cross-border e-commerce products recommendation has become one of the emerging researching topics in the field of e-commerce. Due to the diversity and complexity of e-commerce product information, the “user-item” correlation matrix is extremely sparse and the cold start problem is prominent. As a result, the traditional collaborative filtering model seems to be malfunctional. Meanwhile, the improved recommendation model based on collaborative filtering or matrix factorization only considers the explicit and implicit feedback information of the users to the products, while ignoring the graph structure information composed of users and items, so that the recommendation performance is difficult to meet the requirements of the platform and users. To tackle these issues, a recommender system of cross-border e-commerce based on heterogeneous graph neural network, named Heterogeneous Graph Neural network Recommender system (HGNR), is proposed in this paper. The model has two significant advantages: (1) the three-part graph is used as input, and high-quality information dissemination and aggregation are carried out on heterogeneous graphs through Graph Convolutional neural Network (GCN); (2) high-quality user and product representation vectors can be obtained, and realize the modeling of the complex interaction between users and products. Experimental results on real cross-border e-commerce order data sets show that HGNR not only owns the superior performance, but also can effectively improve the recommendation accuracy of cold-start users. Compared with nine baseline methods for recommendation, HGNR achieves improvements of at least 3.33%, 0.91%, and 0.54% on evaluation metrics of HitRate@10, Item-coverage@10 and MRR@10.
2022, 44(11): 4018-4024.
doi: 10.11999/JEIT210979
Abstract:
Due to good correlation and orthogonal properties, Linear Complementary Dual (LCD) codes over the finite fields can be used to defend against channel attacks. As a very important class of codes in coding theory, self-orthogonal codes can be used to construct quantum error-correcting codes. In this paper, LCD codes over the finite field F3 are studied. By selecting appropriate defining sets and using the conditions for linear codes over the finite field F3 to be LCD codes or self-orthogonal codes, four kinds of ternary LCD codes and some self-orthogonal codes are constructed. And the dual codes of these four kinds of liner codes are also studied and some ternary optimal linear codes are obtained.
Due to good correlation and orthogonal properties, Linear Complementary Dual (LCD) codes over the finite fields can be used to defend against channel attacks. As a very important class of codes in coding theory, self-orthogonal codes can be used to construct quantum error-correcting codes. In this paper, LCD codes over the finite field F3 are studied. By selecting appropriate defining sets and using the conditions for linear codes over the finite field F3 to be LCD codes or self-orthogonal codes, four kinds of ternary LCD codes and some self-orthogonal codes are constructed. And the dual codes of these four kinds of liner codes are also studied and some ternary optimal linear codes are obtained.
2022, 44(11): 4025-4033.
doi: 10.11999/JEIT210829
Abstract:
Considering the problem of effective repair of minimum bandwidth regenerating codes, a construction algorithm of Fractional Repetition (FR) codes based on difference set matrix is proposed. The orthogonal array is constructed by using the difference set matrix and Kronecker sum. According to the orthogonal array, each row of the same element is taken as the coding blocks of the node to obtain the corresponding FR codes. As a result, the constructed FR codes can be divided into multiple parallel classes, and at the repetition of the data blocks and the storage capacity of the node can be adjusted. The simulation results show that compared with the traditional Reed-Solomon (RS) codes and Simple Regenerating Codes (SRC), the constructed FR codes have better performance in terms of repair complexity, repair bandwidth overhead, and repair locality. Although the repair selectivity is a table-based repair scheme, the selectivity can still reach high.
Considering the problem of effective repair of minimum bandwidth regenerating codes, a construction algorithm of Fractional Repetition (FR) codes based on difference set matrix is proposed. The orthogonal array is constructed by using the difference set matrix and Kronecker sum. According to the orthogonal array, each row of the same element is taken as the coding blocks of the node to obtain the corresponding FR codes. As a result, the constructed FR codes can be divided into multiple parallel classes, and at the repetition of the data blocks and the storage capacity of the node can be adjusted. The simulation results show that compared with the traditional Reed-Solomon (RS) codes and Simple Regenerating Codes (SRC), the constructed FR codes have better performance in terms of repair complexity, repair bandwidth overhead, and repair locality. Although the repair selectivity is a table-based repair scheme, the selectivity can still reach high.
2022, 44(11): 4034-4040.
doi: 10.11999/JEIT210881
Abstract:
Based on the support of binary sequences and low correlation sequence sets, a new framework for constructing periodic quasi-complementary sequence sets is proposed. Based on this framework, three classes of asymptotically optimal and asymptotically almost optimal periodic quasi-complementary sequence sets are proposed by using the optimal quaternary sequence family A, family D and Luke sequence set, respectively. In addition, the parameters of sequence set are determined by the binary sequence and the low correlation sequence set. Compared with the traditional complete complementary sequence set, the quasi-complementary sequence set includes more sequences, which can support more users in multi-carrier spread spectrum communication system.
Based on the support of binary sequences and low correlation sequence sets, a new framework for constructing periodic quasi-complementary sequence sets is proposed. Based on this framework, three classes of asymptotically optimal and asymptotically almost optimal periodic quasi-complementary sequence sets are proposed by using the optimal quaternary sequence family A, family D and Luke sequence set, respectively. In addition, the parameters of sequence set are determined by the binary sequence and the low correlation sequence set. Compared with the traditional complete complementary sequence set, the quasi-complementary sequence set includes more sequences, which can support more users in multi-carrier spread spectrum communication system.
2022, 44(11): 4041-4057.
doi: 10.11999/JEIT210896
Abstract:
With the arrival of the “computing power era”, large-scale data need to go back and forth between the memory and the processor. However, the demand for frequent access can not be achieved in the traditional Von Neumann architecture which separates the computing and storage. The Von Neumann bottleneck and the “storage wall” in the traditional computing architecture have been broken with the birth of Computing In-Memory (CIM) Static Random-Access Memory technique. Thereby, for the “computing power era” it has revolutionary significance. Due to Static Random-Access Memory (SRAM) reads data fast and has better compatibility with advanced logic technology. Therefore, the attention of scholars at domestic and international has been attracted by SRAM-based CIM technology. The application of SRAM-based CIM technology is summarized, including machine learning, coding, encryption and decryption algorithm. The various circuit structures to realize the operation function and various quantization techniques based on Analog-to-Digital Conversion (ADC) are summarized and compared in this paper. In addition, the problems and challenges of the existing CIM architectures are analyzed. Then some existing solution strategies for those issues also are presented. Finally, the technique of SRAM-based CIM is prospected from different aspects.
With the arrival of the “computing power era”, large-scale data need to go back and forth between the memory and the processor. However, the demand for frequent access can not be achieved in the traditional Von Neumann architecture which separates the computing and storage. The Von Neumann bottleneck and the “storage wall” in the traditional computing architecture have been broken with the birth of Computing In-Memory (CIM) Static Random-Access Memory technique. Thereby, for the “computing power era” it has revolutionary significance. Due to Static Random-Access Memory (SRAM) reads data fast and has better compatibility with advanced logic technology. Therefore, the attention of scholars at domestic and international has been attracted by SRAM-based CIM technology. The application of SRAM-based CIM technology is summarized, including machine learning, coding, encryption and decryption algorithm. The various circuit structures to realize the operation function and various quantization techniques based on Analog-to-Digital Conversion (ADC) are summarized and compared in this paper. In addition, the problems and challenges of the existing CIM architectures are analyzed. Then some existing solution strategies for those issues also are presented. Finally, the technique of SRAM-based CIM is prospected from different aspects.
2022, 44(11): 4058-4074.
doi: 10.11999/JEIT210886
Abstract:
Radio Frequency Integrated Circuits (RFICs) show poor robustness to static non-ideal factors introduced by process deviations, device mismatches, device nonlinearities, and dynamic non-ideal factors introduced by temperature changes, gain changes, and input/output frequency changes. The key factors that affect the performance of RFICs are excavated deeply, and typical calibration algorithms are summarized to provide theoretical support for the design of high-performance RFICs.
Radio Frequency Integrated Circuits (RFICs) show poor robustness to static non-ideal factors introduced by process deviations, device mismatches, device nonlinearities, and dynamic non-ideal factors introduced by temperature changes, gain changes, and input/output frequency changes. The key factors that affect the performance of RFICs are excavated deeply, and typical calibration algorithms are summarized to provide theoretical support for the design of high-performance RFICs.
