ISSN 
2021 Vol. 50, No. 2
2021, 50(2): 161-166.
doi: 10.12178/1001-0548.2020442
Abstract:
The parallel decoding system with multiple sub-decoders has much performance improvement than that of single-decoder system for decoding the block codes. However, the construction of sub-decoders for the parallel decoding implementation is still the challenging problem. To solve this problem, this paper proposes a parallel decoding method with multiple sub-decoders based on BP (belief propagation) algorithm for some specific low-density parity-check (LDPC) codes. This method is particularly effective for decoding the LDPC code generated via primitive polynomial. The method has characteristics as the parity-check matrix used for each sub-decoder depends on a proper cyclic shift of the original parity-check matrix, and the number of cyclic shift depends on the sampling property of the m sequence (which uniquely corresponds to a primitive polynomial). The iteration times of each BP processes in the sub-decoder are set as half of the girth of Tanner graph of the parity-check matrix, thus the affection of short cycles on BP performance would be eliminated. The output extrinsic information for each bit generated by the sub-decoder is processed further by a decoding module to output the candidate codeword, and then the LMS module picks out the maximum likelihood candidate codeword as the output of the decoding system. The simulation results show that the performance of the proposed parallel decoding method with 5 sub-decoders is about 0.4 dB superior to that of the original single-decoder decoding method at the bit error rate of 10−5.
The parallel decoding system with multiple sub-decoders has much performance improvement than that of single-decoder system for decoding the block codes. However, the construction of sub-decoders for the parallel decoding implementation is still the challenging problem. To solve this problem, this paper proposes a parallel decoding method with multiple sub-decoders based on BP (belief propagation) algorithm for some specific low-density parity-check (LDPC) codes. This method is particularly effective for decoding the LDPC code generated via primitive polynomial. The method has characteristics as the parity-check matrix used for each sub-decoder depends on a proper cyclic shift of the original parity-check matrix, and the number of cyclic shift depends on the sampling property of the m sequence (which uniquely corresponds to a primitive polynomial). The iteration times of each BP processes in the sub-decoder are set as half of the girth of Tanner graph of the parity-check matrix, thus the affection of short cycles on BP performance would be eliminated. The output extrinsic information for each bit generated by the sub-decoder is processed further by a decoding module to output the candidate codeword, and then the LMS module picks out the maximum likelihood candidate codeword as the output of the decoding system. The simulation results show that the performance of the proposed parallel decoding method with 5 sub-decoders is about 0.4 dB superior to that of the original single-decoder decoding method at the bit error rate of 10−5.
2021, 50(2): 167-172.
doi: 10.12178/1001-0548.2021037
Abstract:
This paper investigates joint optimization of beampattern synthesis and spectral compatibility of multiple-input multiple-output (MIMO) radar transmit waveform under constant modulus constraint. In complex electromagnetic environments, it is necessary to consider the spectrum compatibility and flexibility in the congested frequency band to avoid poor transmission performance of the communication system in the adjacent frequency band. In this paper, the joint optimization problem is modeled as a weighted mean square error cost function constrained by constant modulus, which is non-convex and non-smooth. In each step of the iterative optimization algorithm, firstly update the matching coefficient of the current step and scale the cost function to construct a majorization function. Next, the majorization function is optimized to obtain a waveform correspond to a smaller cost function value. The performance of the proposed algorithm has a good performance on beampattern synthesis and spectrally compatibility, it also has a small amount of calculation. Simulation and numerical results prove the effectiveness of the proposed method.
This paper investigates joint optimization of beampattern synthesis and spectral compatibility of multiple-input multiple-output (MIMO) radar transmit waveform under constant modulus constraint. In complex electromagnetic environments, it is necessary to consider the spectrum compatibility and flexibility in the congested frequency band to avoid poor transmission performance of the communication system in the adjacent frequency band. In this paper, the joint optimization problem is modeled as a weighted mean square error cost function constrained by constant modulus, which is non-convex and non-smooth. In each step of the iterative optimization algorithm, firstly update the matching coefficient of the current step and scale the cost function to construct a majorization function. Next, the majorization function is optimized to obtain a waveform correspond to a smaller cost function value. The performance of the proposed algorithm has a good performance on beampattern synthesis and spectrally compatibility, it also has a small amount of calculation. Simulation and numerical results prove the effectiveness of the proposed method.
2021, 50(2): 173-179.
doi: 10.12178/1001-0548.2020028
Abstract:
In order to solve the problem of fault node repair in distributed storage system, a construction algorithm of fractional repetition (FR) code is proposed. Specifically, the FR code is constructed directly by Hadamard matrix through simple transformation. Then, the grouping idea is introduced and the 8-order Hadamard matrix is used to construct the grouping FR code, which is more concise and intuitive and can realize the precise non-coding repair of multiple fault nodes in the local repair group. Compared with Reed-Solomon (RS) codes and simple regenerating codes (SRC), theoretical analysis shows that designed FR codes have lower repair locality, repair bandwidth overhead and repair complexity. In addition, this method has high repair efficiency and reduces the repair time of failed nodes.
In order to solve the problem of fault node repair in distributed storage system, a construction algorithm of fractional repetition (FR) code is proposed. Specifically, the FR code is constructed directly by Hadamard matrix through simple transformation. Then, the grouping idea is introduced and the 8-order Hadamard matrix is used to construct the grouping FR code, which is more concise and intuitive and can realize the precise non-coding repair of multiple fault nodes in the local repair group. Compared with Reed-Solomon (RS) codes and simple regenerating codes (SRC), theoretical analysis shows that designed FR codes have lower repair locality, repair bandwidth overhead and repair complexity. In addition, this method has high repair efficiency and reduces the repair time of failed nodes.
2021, 50(2): 180-185.
doi: 10.12178/1001-0548.2019246
Abstract:
Based on TSMC 0.18 μm RF CMOS process, a hybrid digital analog composite structure is implemented to accelerate the locking time of charge-pump phase-locked loop (CP-PLL). The composite structure mainly includes two independent units: dynamic loop bandwidth unit and preset feedback loop. Among them, the control circuits of the two units are all digital circuits, and the layout information is obtained through DC synthesis and ICC automatic layout. Through the comparative analysis under the same CP-PLL parameter environment, the locking times of three schemes including the traditional structure are compared. Under the working power supply of 1.8 V, the optimized locking time is 1.12 μs, which is 76.7% higher than that of the traditional structure; the overall phase noise keeps −103.1dBc/Hz@1MHz in the steady state, which is only 0.3% higher than that of the traditional structure. Therefore, the composite structure can effectively reduce the lock-in time of power on and frequency hopping.
Based on TSMC 0.18 μm RF CMOS process, a hybrid digital analog composite structure is implemented to accelerate the locking time of charge-pump phase-locked loop (CP-PLL). The composite structure mainly includes two independent units: dynamic loop bandwidth unit and preset feedback loop. Among them, the control circuits of the two units are all digital circuits, and the layout information is obtained through DC synthesis and ICC automatic layout. Through the comparative analysis under the same CP-PLL parameter environment, the locking times of three schemes including the traditional structure are compared. Under the working power supply of 1.8 V, the optimized locking time is 1.12 μs, which is 76.7% higher than that of the traditional structure; the overall phase noise keeps −103.1dBc/Hz@1MHz in the steady state, which is only 0.3% higher than that of the traditional structure. Therefore, the composite structure can effectively reduce the lock-in time of power on and frequency hopping.
2021, 50(2): 186-192.
doi: 10.12178/1001-0548.2020117
Abstract:
In this paper, a dynamic estimation and compensation algorithm for matching the error characteristics of multi-sensor system is proposed to solve the engineering application problem of the multi-sensor system error estimation in the non-cooperative target scenario. A real-time estimation model of system errors is established and a dynamic iterative estimation and compensation algorithm is designed, which makes the error characteristics of the original plot and the fusion track to be consistent, overcomes the difficulty of plot-track association in the multi-sensor system, and ensures the correctness of relative system error estimation and compensation. The measured data from multiple scenes of different devices show that this method can improve the accuracy and stability of multi-sensor fusion system, so as to improve the plot quality of multi-sensor fusion system.
In this paper, a dynamic estimation and compensation algorithm for matching the error characteristics of multi-sensor system is proposed to solve the engineering application problem of the multi-sensor system error estimation in the non-cooperative target scenario. A real-time estimation model of system errors is established and a dynamic iterative estimation and compensation algorithm is designed, which makes the error characteristics of the original plot and the fusion track to be consistent, overcomes the difficulty of plot-track association in the multi-sensor system, and ensures the correctness of relative system error estimation and compensation. The measured data from multiple scenes of different devices show that this method can improve the accuracy and stability of multi-sensor fusion system, so as to improve the plot quality of multi-sensor fusion system.
2021, 50(2): 193-198.
doi: 10.12178/1001-0548.2020284
Abstract:
Metamodel has been widely used in complex electrical equipment design field to improve the computational efficiency and shorten the time of simulation optimization. This paper proposes the adaptive infilling method based on significant domain. In this method, new samples are selected through the local sampling model and global sampling model to improve the local exploitation capability and the global exploration capability of metamodel at every iteration. The method is tested with several benchmark numerical problems and gets verified.
Metamodel has been widely used in complex electrical equipment design field to improve the computational efficiency and shorten the time of simulation optimization. This paper proposes the adaptive infilling method based on significant domain. In this method, new samples are selected through the local sampling model and global sampling model to improve the local exploitation capability and the global exploration capability of metamodel at every iteration. The method is tested with several benchmark numerical problems and gets verified.
2021, 50(2): 199-206.
doi: 10.12178/1001-0548.2020420
Abstract:
In the traditional particle swarm optimization (PSO) algorithm, all particles follow some initial parameters to explore themselves. This scheme is easy to lead to premature maturity, and easy to be trapped in the local optimum. To solve the above problems, a particle swarm optimization algorithm based on deep deterministic policy gradient (DDPGPSO) is proposed. The action function and action value function are realized by constructing neural network. The parameters required by the algorithm can be generated dynamically by using the neural network, which reduces the difficulty of manual configuration of the algorithm. The experimental results show that DDPGPSO has a great improvement in convergence speed and optimization accuracy compared with nine similar algorithms.
In the traditional particle swarm optimization (PSO) algorithm, all particles follow some initial parameters to explore themselves. This scheme is easy to lead to premature maturity, and easy to be trapped in the local optimum. To solve the above problems, a particle swarm optimization algorithm based on deep deterministic policy gradient (DDPGPSO) is proposed. The action function and action value function are realized by constructing neural network. The parameters required by the algorithm can be generated dynamically by using the neural network, which reduces the difficulty of manual configuration of the algorithm. The experimental results show that DDPGPSO has a great improvement in convergence speed and optimization accuracy compared with nine similar algorithms.
2021, 50(2): 207-213.
doi: 10.12178/1001-0548.2020327
Abstract:
The traditional texture synthesis image inpainting approaches can only extract useful information from the damaged image, but cannot deal with the complex structures. In the meanwhile, the deep-learning-based ones usually have long training time and unsatisfactory texture synthesis effects. To solve the problems, this paper proposes an image inpainting approach based on similar image registration. First, a similarity calculation method of damaged image is proposed by using the deep learning features of images, thus the most similar image of the damaged ones in dataset can be found to provide more useful information for the image inpainting process. Second, this paper matches the damaged image with its similar ones and use the homography transform to realize the automatic rough correction of image space position. At last, the texture synthesis effects are improved by using the improved optimal patch searching method and the relative matching criteria, then the image inpainting is performed. Simulation results demonstrate that the approach can obtain more useful information, yield perfect texture synthesis effect, and overcome the shortcomings of the traditional deep-learning-based and texture synthesis approaches. Besides that, the proposed approach can also obtain ideal inpainting effects even for the damaged images with complex textural information and structures.
The traditional texture synthesis image inpainting approaches can only extract useful information from the damaged image, but cannot deal with the complex structures. In the meanwhile, the deep-learning-based ones usually have long training time and unsatisfactory texture synthesis effects. To solve the problems, this paper proposes an image inpainting approach based on similar image registration. First, a similarity calculation method of damaged image is proposed by using the deep learning features of images, thus the most similar image of the damaged ones in dataset can be found to provide more useful information for the image inpainting process. Second, this paper matches the damaged image with its similar ones and use the homography transform to realize the automatic rough correction of image space position. At last, the texture synthesis effects are improved by using the improved optimal patch searching method and the relative matching criteria, then the image inpainting is performed. Simulation results demonstrate that the approach can obtain more useful information, yield perfect texture synthesis effect, and overcome the shortcomings of the traditional deep-learning-based and texture synthesis approaches. Besides that, the proposed approach can also obtain ideal inpainting effects even for the damaged images with complex textural information and structures.
2021, 50(2): 214-224.
doi: 10.12178/1001-0548.2020246
Abstract:
Objective Ultrasound (US) is one of a primary imageological examination and preoperative assessment for breast nodules. However, in the field of ultrasound diagnosis, it relies heavily on the experience of physicians due to the overlapping image expression of benign and malignant breast nodules. Computer-aided medical diagnosis has gradually become a hot spot of current research. In this paper, a heterogeneous multi-branch network (HMBN) is presented for benign and malignant classification of the breast ultrasound images. In HMBN, the image information includes ultrasound images and contrast-enhanced ultrasound (CEUS) images while non-image information includes patients’ age and other six pathological features. On the other hand, a fusion loss function suitable for this heterogeneous multi-branch network is also proposed. This loss function uses the minimum hyperspherical energy (MHE) based on additive angular margin loss to improve the classification accuracy. Experimental results show that on the breast ultrasound data set of 1303 cases collected, the classification accuracy of the proposed heterogeneous multi-branch network is 92.41%, which is 7.11% higher than the average diagnostic accuracy of doctors with five years of experience, and ranks among the best in diagnostic accuracy in comparison with other latest research results. It is proved that the accuracy and robustness of breast diagnosis are greatly improved by incorporating medical knowledge into the optimization process and adding contrast-enhanced ultrasound images and non-image information to the network.
Objective Ultrasound (US) is one of a primary imageological examination and preoperative assessment for breast nodules. However, in the field of ultrasound diagnosis, it relies heavily on the experience of physicians due to the overlapping image expression of benign and malignant breast nodules. Computer-aided medical diagnosis has gradually become a hot spot of current research. In this paper, a heterogeneous multi-branch network (HMBN) is presented for benign and malignant classification of the breast ultrasound images. In HMBN, the image information includes ultrasound images and contrast-enhanced ultrasound (CEUS) images while non-image information includes patients’ age and other six pathological features. On the other hand, a fusion loss function suitable for this heterogeneous multi-branch network is also proposed. This loss function uses the minimum hyperspherical energy (MHE) based on additive angular margin loss to improve the classification accuracy. Experimental results show that on the breast ultrasound data set of 1303 cases collected, the classification accuracy of the proposed heterogeneous multi-branch network is 92.41%, which is 7.11% higher than the average diagnostic accuracy of doctors with five years of experience, and ranks among the best in diagnostic accuracy in comparison with other latest research results. It is proved that the accuracy and robustness of breast diagnosis are greatly improved by incorporating medical knowledge into the optimization process and adding contrast-enhanced ultrasound images and non-image information to the network.
2021, 50(2): 225-230.
doi: 10.12178/1001-0548.2020281
Abstract:
Feature-based model retrieval is one of the important research directions in the field of computer vision. It includes two aspects: feature extraction and model retrieval. The robustness of features plays a decisive role in model retrieval algorithm. In order to solve the problem of low efficiency of local features in existing algorithms, a feature fusion based model retrieval algorithm for the fragment of terracotta warriors is proposed. Aiming at the 3D point cloud model of the terracotta warriors fragments, the curvature and normal angle of the points on the fragment point cloud model are calculated and fused weighted firstly. Then, the feature matching algorithm is constructed based on the fusion feature, and the 3D fragment model retrieval is realized by matching the fusion feature. In the experiment, 1036 fragments of 50 terracotta warriors are retrieved. The results show that the algorithm can effectively improve the retrieval accuracy of fragments and avoid the algorithm falling into local extremum. Therefore, the 3D model retrieval algorithm based on feature fusion is an effective method to retrieve the fragments of terracotta warriors.
Feature-based model retrieval is one of the important research directions in the field of computer vision. It includes two aspects: feature extraction and model retrieval. The robustness of features plays a decisive role in model retrieval algorithm. In order to solve the problem of low efficiency of local features in existing algorithms, a feature fusion based model retrieval algorithm for the fragment of terracotta warriors is proposed. Aiming at the 3D point cloud model of the terracotta warriors fragments, the curvature and normal angle of the points on the fragment point cloud model are calculated and fused weighted firstly. Then, the feature matching algorithm is constructed based on the fusion feature, and the 3D fragment model retrieval is realized by matching the fusion feature. In the experiment, 1036 fragments of 50 terracotta warriors are retrieved. The results show that the algorithm can effectively improve the retrieval accuracy of fragments and avoid the algorithm falling into local extremum. Therefore, the 3D model retrieval algorithm based on feature fusion is an effective method to retrieve the fragments of terracotta warriors.
2021, 50(2): 231-235.
doi: 10.12178/1001-0548.2020319
Abstract:
With a large number of video surveillance and camera networks, face recognition of continuous video frames in unrestricted scenes is becoming more and more attractive. Most of the traditional face recognition methods for continuous video frames have the problem of fluctuating recognition results and intensive computing resources. In this paper, an efficient 3D decomposition convolution is designed, which can effectively reduce the computational consumption of video face recognition and improve the recognition accuracy. Finally, we also propose a temporal pyramid network to further effectively mine complementary information between frames to improve the recognition accuracy. The performance has been tested on YTF and PaSC datasets.
With a large number of video surveillance and camera networks, face recognition of continuous video frames in unrestricted scenes is becoming more and more attractive. Most of the traditional face recognition methods for continuous video frames have the problem of fluctuating recognition results and intensive computing resources. In this paper, an efficient 3D decomposition convolution is designed, which can effectively reduce the computational consumption of video face recognition and improve the recognition accuracy. Finally, we also propose a temporal pyramid network to further effectively mine complementary information between frames to improve the recognition accuracy. The performance has been tested on YTF and PaSC datasets.
2021, 50(2): 236-242.
doi: 10.12178/1001-0548.2020325
Abstract:
This paper proposes a deep learning-based user QoE prediction network (UQPN)). In this work, the current user's QoE is predicted and modeled based on the current video playback states, and UQPN is used to replace the existing reward functions, in this way the generated ABR algorithm can make bitrate decisions more in line with user requirements. Experiments and the comparison with the existing reward functions show that he correlation coefficient of UQPN prediction and user QoE is higher, and the algorithm using UQPN as reinforcement learning reward can improve user QoE by at least 20%.
This paper proposes a deep learning-based user QoE prediction network (UQPN)). In this work, the current user's QoE is predicted and modeled based on the current video playback states, and UQPN is used to replace the existing reward functions, in this way the generated ABR algorithm can make bitrate decisions more in line with user requirements. Experiments and the comparison with the existing reward functions show that he correlation coefficient of UQPN prediction and user QoE is higher, and the algorithm using UQPN as reinforcement learning reward can improve user QoE by at least 20%.
2021, 50(2): 243-248.
doi: 10.12178/1001-0548.2020314
Abstract:
Research shows that there is redundancy in the output feature maps of ResNet bottleneck structure. Such redundancy ensures a comprehensive understanding of the input data, but generates the redundant information which consumes additional computational resources. And the proportion of redundant information is very large when processing small category classification tasks. To solve this problem, a new dimension increasing structure is designed to improve the bottleneck structure by residual-like structure and concatenation operation. This structure is called residual concatenation (RC). The RC can not only reduce the amount of calculation and parameters of bottleneck structure, but also enhance the gradient transmission of back-propagation to improve the accuracy. In this work, the RC is combined with multiple residual networks and image classification experiments are performed on multiple datasets. The results show that the RC-based bottleneck structure can reduce the consumption and improve the accuracy of classification tasks while processing small category classification tasks.
Research shows that there is redundancy in the output feature maps of ResNet bottleneck structure. Such redundancy ensures a comprehensive understanding of the input data, but generates the redundant information which consumes additional computational resources. And the proportion of redundant information is very large when processing small category classification tasks. To solve this problem, a new dimension increasing structure is designed to improve the bottleneck structure by residual-like structure and concatenation operation. This structure is called residual concatenation (RC). The RC can not only reduce the amount of calculation and parameters of bottleneck structure, but also enhance the gradient transmission of back-propagation to improve the accuracy. In this work, the RC is combined with multiple residual networks and image classification experiments are performed on multiple datasets. The results show that the RC-based bottleneck structure can reduce the consumption and improve the accuracy of classification tasks while processing small category classification tasks.
2021, 50(2): 249-255, 293.
doi: 10.12178/1001-0548.2020218
Abstract:
In order to optimize Chinese-Tibetan neural machine translation (NMT) based on attention mechanism, this paper proposes a Tibetan byte-pair encoding algorithm with maximum byte threshold to improve the original byte-pair encoding algorithm. By collecting one million Chinese-Tibetan sentence pairs and dictionaries with 200, 000 Chinese-Tibetan names and places, we train the Chinese-Tibetan NMT model using attention mechanism. Our model has a better translation result in named entity compared with commercial using of Chinese-Tibetan online translation and it achieves 36.84 in bilingual evaluation understudy (BLEU) score. Our work has already deployed in Chinese-Tibetan machine translation system web which will promote the spread and application of Chinese-Tibetan NMT system.
In order to optimize Chinese-Tibetan neural machine translation (NMT) based on attention mechanism, this paper proposes a Tibetan byte-pair encoding algorithm with maximum byte threshold to improve the original byte-pair encoding algorithm. By collecting one million Chinese-Tibetan sentence pairs and dictionaries with 200, 000 Chinese-Tibetan names and places, we train the Chinese-Tibetan NMT model using attention mechanism. Our model has a better translation result in named entity compared with commercial using of Chinese-Tibetan online translation and it achieves 36.84 in bilingual evaluation understudy (BLEU) score. Our work has already deployed in Chinese-Tibetan machine translation system web which will promote the spread and application of Chinese-Tibetan NMT system.
2021, 50(2): 256-260.
doi: 10.12178/1001-0548.2020426
Abstract:
This paper analyzes myopia examination data at home and abroad. Statistics show that the incidence of myopia in Chinese adolescents far exceeds the international adolescents. 8 to 12 years old is a period when the number of myopia is increasing rapidly. About 20% of non-myopia students turn into myopia students every year in this period. The age of 10 to 14 is a dangerous period of suffering from high myopia. Time for outdoor activities and parents’ myopia have the greatest impact on the occurrence of myopia, higher than that of the time spent on computer and the time spent on watching TV. This paper uses five ensemble learning methods to predict people’s future vision. Considering the robustness and accuracy, the random forest model has the best prediction effect. The prediction accuracy of myopia is 92.8% in the case of 70% training set and 30% test set.
This paper analyzes myopia examination data at home and abroad. Statistics show that the incidence of myopia in Chinese adolescents far exceeds the international adolescents. 8 to 12 years old is a period when the number of myopia is increasing rapidly. About 20% of non-myopia students turn into myopia students every year in this period. The age of 10 to 14 is a dangerous period of suffering from high myopia. Time for outdoor activities and parents’ myopia have the greatest impact on the occurrence of myopia, higher than that of the time spent on computer and the time spent on watching TV. This paper uses five ensemble learning methods to predict people’s future vision. Considering the robustness and accuracy, the random forest model has the best prediction effect. The prediction accuracy of myopia is 92.8% in the case of 70% training set and 30% test set.
2021, 50(2): 261-266.
doi: 10.12178/1001-0548.2019191
Abstract:
In order to build temporal features of flight trajectory accurately, the Gaussian Process (GP) is applied to predict the future flight trajectory. Meanwhile, considering the non-linearity characteristics of the aircraft during the high maneuverability motion, the GP is combined with the deep belief network to formulate the deep GP which is applied to flight trajectory prediction. Based on the predicted trajectory, the probabilistic flight conflict detection based on deep GP is proposed and implemented in this paper. The Monte Carlo simulation and Markov Chain Monte Carlo sampling are proposed to compute the conflict probability for the proposed conflict detection method. Deep GP based flight trajectory approach can not only predict the nominal trajectory for aircraft, but also estimate the probabilistic distribution of the confidence interval for the predicted positions, which lays as solid data foundation for the conflict detection task. Experimental results on real data show that the proposed deep GP based trajectory prediction model can obtain higher accuracy and stability than that of baselines. In addition, by applying the predicted trajectory to the conflict detection algorithm, we can achieve the task with lower false alarm and longer warning time.
In order to build temporal features of flight trajectory accurately, the Gaussian Process (GP) is applied to predict the future flight trajectory. Meanwhile, considering the non-linearity characteristics of the aircraft during the high maneuverability motion, the GP is combined with the deep belief network to formulate the deep GP which is applied to flight trajectory prediction. Based on the predicted trajectory, the probabilistic flight conflict detection based on deep GP is proposed and implemented in this paper. The Monte Carlo simulation and Markov Chain Monte Carlo sampling are proposed to compute the conflict probability for the proposed conflict detection method. Deep GP based flight trajectory approach can not only predict the nominal trajectory for aircraft, but also estimate the probabilistic distribution of the confidence interval for the predicted positions, which lays as solid data foundation for the conflict detection task. Experimental results on real data show that the proposed deep GP based trajectory prediction model can obtain higher accuracy and stability than that of baselines. In addition, by applying the predicted trajectory to the conflict detection algorithm, we can achieve the task with lower false alarm and longer warning time.
2021, 50(2): 267-271, 293.
doi: 10.12178/1001-0548.2019288
Abstract:
A photonic crystal waveguide optical switch based on the phase-change materials VO2 is designed. Through plating the SU-8 cladding material onto the waveguide to compensate the temperature effect of the silicon waveguide, the temperature dependence of the waveguide switch can be reduced. An extinction ratio of approximately 9.5 dB is achieved at 1 591 nm with a thermal actuation. In addition, a new cascaded grating is designed to optimize the switch structure and achieves an extinction ratio of 27.46 dB.
A photonic crystal waveguide optical switch based on the phase-change materials VO2 is designed. Through plating the SU-8 cladding material onto the waveguide to compensate the temperature effect of the silicon waveguide, the temperature dependence of the waveguide switch can be reduced. An extinction ratio of approximately 9.5 dB is achieved at 1 591 nm with a thermal actuation. In addition, a new cascaded grating is designed to optimize the switch structure and achieves an extinction ratio of 27.46 dB.
Research on High-Precision Optical Fiber Axial Strain Sensor Combined with Optoelectronic Oscillator
2021, 50(2): 272-275.
doi: 10.12178/1001-0548.2019292
Abstract:
A high-precision optical fiber axial strain sensing system combined with optoelectronic oscillator (OEO), using a section of standard single-mode fiber as the sensing head, is proposed and experimentally demonstrated. In the proposed scheme, the mode selection in the OEO cavity is achieved via a single-passband microwave photonic filter which is realized by an amplified spontaneous emission (ASE) light source, a dual-output Mach-Zehnder electro-optic intensity modulator (DOMZM), and a Mach-Zehnder interferometer (MZI) involving two sections of optical fiber with different length. Through applying axial strain to either arm of the MZI, the length difference between the two arms of the MZI varies. Hence, the frequency of the generated microwave signal in the OEO varies, which is used for axial strain sensing. In the experiment, when the axial strain is applied to the longer fiber, the sensitivity and the linearity are measured to be 4.67121×10−4 GHz/με and 0.99895, respectively. When the axial strain is applied to the shorter fiber, the sensitivity and the linearity are measured to be −4.48388×10−4 GHz/με and 0.99841, respectively.
A high-precision optical fiber axial strain sensing system combined with optoelectronic oscillator (OEO), using a section of standard single-mode fiber as the sensing head, is proposed and experimentally demonstrated. In the proposed scheme, the mode selection in the OEO cavity is achieved via a single-passband microwave photonic filter which is realized by an amplified spontaneous emission (ASE) light source, a dual-output Mach-Zehnder electro-optic intensity modulator (DOMZM), and a Mach-Zehnder interferometer (MZI) involving two sections of optical fiber with different length. Through applying axial strain to either arm of the MZI, the length difference between the two arms of the MZI varies. Hence, the frequency of the generated microwave signal in the OEO varies, which is used for axial strain sensing. In the experiment, when the axial strain is applied to the longer fiber, the sensitivity and the linearity are measured to be 4.67121×10−4 GHz/με and 0.99895, respectively. When the axial strain is applied to the shorter fiber, the sensitivity and the linearity are measured to be −4.48388×10−4 GHz/με and 0.99841, respectively.
2021, 50(2): 276-284.
doi: 10.12178/1001-0548.2020241
Abstract:
Most of existing methods of link prediction are based on the similarity of network structures and the weight of edges, but they do not effectively use temporal information of forming the weight of edges. The behavior synchronization of two nodes is often caused by the link relationship between them, so the behavior synchronization of nodes has been widely used in many researches of network structure reconstruction to conjecture whether there is a link relationship between any pair of nodes. In this study, we attempt to introduce node synchronization information into the field of link prediction, and propose a novel link prediction algorithm which integrates the synchronization index of node behaviors with network topological similarity. By analyzing and comparing two types of six real-life network data, the proposed method can effectively improve the accuracy of link prediction. Compared with the existing methods, the performance of precision can increase by 15.3% to 68.2%. This study not only finds the joint influence of local structure similarity and behavior synchronization index on link prediction, but also reveals intrinsic structures and dynamic characteristics of different types of real-life weighted networks.
Most of existing methods of link prediction are based on the similarity of network structures and the weight of edges, but they do not effectively use temporal information of forming the weight of edges. The behavior synchronization of two nodes is often caused by the link relationship between them, so the behavior synchronization of nodes has been widely used in many researches of network structure reconstruction to conjecture whether there is a link relationship between any pair of nodes. In this study, we attempt to introduce node synchronization information into the field of link prediction, and propose a novel link prediction algorithm which integrates the synchronization index of node behaviors with network topological similarity. By analyzing and comparing two types of six real-life network data, the proposed method can effectively improve the accuracy of link prediction. Compared with the existing methods, the performance of precision can increase by 15.3% to 68.2%. This study not only finds the joint influence of local structure similarity and behavior synchronization index on link prediction, but also reveals intrinsic structures and dynamic characteristics of different types of real-life weighted networks.
2021, 50(2): 285-293.
doi: 10.12178/1001-0548.2020355
Abstract:
It is of great significance to correctly evaluate the importance of news in national newspapers and magazines for better understanding the changes of national policies. In this paper, we take People’s Daily as an example, extract news published in 1946−2008, and construct news network by using their content-based similarities. In the view of complex network, news has higher similarities with others, making it be closely connected and larger nodal centrality in news network. In respect to this, we propose an H-PageRank ranking algorithm by introducing the H-index to improve the PageRank ranking algorithm. In the experiment, all news in People’s Daily is divided into four stages according to their styles and editions in different governing times, which is respectively used to construct news networks based on representation learning. The experimental results show that 1) the topologies of four news networks all have a general properties of complex network, including the high clustering coefficients, positive assortativity coefficients and approximately power-law degree distributions; 2) each news network presnets a mostly similar AUC calculated by the global rank score of the front-page news according to diverse nodal centralities, however the precision, recall and F1-score calculated by the Top-N evaluating model according to the H-PageRank centrality are optimal, which validate the efficiency of local ranking news according to the H-PageRank centrality; 3) the precision of each news network is significantly superior to the theoretical baselines even when the ranking list is restricted into different length, which suggests the roubustness of evaluating model.
It is of great significance to correctly evaluate the importance of news in national newspapers and magazines for better understanding the changes of national policies. In this paper, we take People’s Daily as an example, extract news published in 1946−2008, and construct news network by using their content-based similarities. In the view of complex network, news has higher similarities with others, making it be closely connected and larger nodal centrality in news network. In respect to this, we propose an H-PageRank ranking algorithm by introducing the H-index to improve the PageRank ranking algorithm. In the experiment, all news in People’s Daily is divided into four stages according to their styles and editions in different governing times, which is respectively used to construct news networks based on representation learning. The experimental results show that 1) the topologies of four news networks all have a general properties of complex network, including the high clustering coefficients, positive assortativity coefficients and approximately power-law degree distributions; 2) each news network presnets a mostly similar AUC calculated by the global rank score of the front-page news according to diverse nodal centralities, however the precision, recall and F1-score calculated by the Top-N evaluating model according to the H-PageRank centrality are optimal, which validate the efficiency of local ranking news according to the H-PageRank centrality; 3) the precision of each news network is significantly superior to the theoretical baselines even when the ranking list is restricted into different length, which suggests the roubustness of evaluating model.
2021, 50(2): 294-302.
doi: 10.12178/1001-0548.2020084
Abstract:
In the knowledge generation community, the network relationship in the real world does not match the traditional classic model because of some artificial evolution characteristics. In this paper, the parameter distribution and time-varying characteristics of the network are introduced into the network, and the evolution process of the network is analyzed dynamically, which provides an objective reference for the construction of the network. Based on the quantitative analysis of label network parameters, this paper proposes a knowledge label connection model based on Barbési-Albert (BA) model, named batch cross-linking BA model, to simulate the complexity of the time-varying parameter series in generation process of the knowledge label network. The simulation results of the model are in good agreement with the actual knowledge label network. This model is helpful for us to understand the complex network in the real world. At the same time, it is more reasonable to use time-varying analysis method to analyze the dynamic trend of the network than static analysis method.
In the knowledge generation community, the network relationship in the real world does not match the traditional classic model because of some artificial evolution characteristics. In this paper, the parameter distribution and time-varying characteristics of the network are introduced into the network, and the evolution process of the network is analyzed dynamically, which provides an objective reference for the construction of the network. Based on the quantitative analysis of label network parameters, this paper proposes a knowledge label connection model based on Barbési-Albert (BA) model, named batch cross-linking BA model, to simulate the complexity of the time-varying parameter series in generation process of the knowledge label network. The simulation results of the model are in good agreement with the actual knowledge label network. This model is helpful for us to understand the complex network in the real world. At the same time, it is more reasonable to use time-varying analysis method to analyze the dynamic trend of the network than static analysis method.
2021, 50(2): 303-310.
doi: 10.12178/1001-0548.2020252
Abstract:
This paper studies the community structure of international trade products and explores the evolutionary rules of product trade from a perspective of complex networks. First, the international trade network of products and the international total trade network are constructed by using the product trade data every five years from 1995 to 2015. Then the product distance is measured with network similarity measure, and the product network is constructed by combining the minimum spanning tree and threshold setting. Finally, the weighted extremum optimization algorithm is used to classify the products and study the evolution rules of the product cluster. The research concludes that although the total international trade is the sum of products, there are significant differences in the product trade relations between countries. As time goes by, international trade is becoming closer, and the focus of product trade between countries keeps changing, from heavy industry to agriculture, then light industry, finally agriculture. The community division of the 5-year product network shows consistent results except for 1995. But within the community, the tightness of the product cluster has been increasing
This paper studies the community structure of international trade products and explores the evolutionary rules of product trade from a perspective of complex networks. First, the international trade network of products and the international total trade network are constructed by using the product trade data every five years from 1995 to 2015. Then the product distance is measured with network similarity measure, and the product network is constructed by combining the minimum spanning tree and threshold setting. Finally, the weighted extremum optimization algorithm is used to classify the products and study the evolution rules of the product cluster. The research concludes that although the total international trade is the sum of products, there are significant differences in the product trade relations between countries. As time goes by, international trade is becoming closer, and the focus of product trade between countries keeps changing, from heavy industry to agriculture, then light industry, finally agriculture. The community division of the 5-year product network shows consistent results except for 1995. But within the community, the tightness of the product cluster has been increasing
2021, 50(2): 311-316.
doi: 10.12178/1001-0548.2020370
Abstract:
Long Short-Term Memory (LSTM) neural network benefits from its ability to capture long-term dependencies and has shown excellent performance in many practical applications. Multivariable LSTM data-driven prediction model is constructed in this paper to predict Australian forest fires by multivariable input. Firstly, the multivariate LSTM prediction model is used to predict the maximum daily temperature, and the results are compared with those of the back-propagation (BP) neural network and Autoregressive Integrated Moving Average model (ARIMA) prediction model. The results show that the BP neural network with the related variables as input cannot consider the time-series variation law, and the prediction error is the largest. ARIMA with single temperature as input makes corresponding prediction according to time series change, and the prediction effect is good. Multivariable LSTM prediction model comprehensively considers the interaction of many factors, and combines the time series dependence, the prediction effect is the best. Finally, the multivariable LSTM prediction model is used to predict whether a node is on fire, and the prediction results are in good agreement with the actual value. Overall, the multivariable LSTM prediction model is reliable in predicting the Australian fires.
Long Short-Term Memory (LSTM) neural network benefits from its ability to capture long-term dependencies and has shown excellent performance in many practical applications. Multivariable LSTM data-driven prediction model is constructed in this paper to predict Australian forest fires by multivariable input. Firstly, the multivariate LSTM prediction model is used to predict the maximum daily temperature, and the results are compared with those of the back-propagation (BP) neural network and Autoregressive Integrated Moving Average model (ARIMA) prediction model. The results show that the BP neural network with the related variables as input cannot consider the time-series variation law, and the prediction error is the largest. ARIMA with single temperature as input makes corresponding prediction according to time series change, and the prediction effect is good. Multivariable LSTM prediction model comprehensively considers the interaction of many factors, and combines the time series dependence, the prediction effect is the best. Finally, the multivariable LSTM prediction model is used to predict whether a node is on fire, and the prediction results are in good agreement with the actual value. Overall, the multivariable LSTM prediction model is reliable in predicting the Australian fires.
2021, 50(2): 317-320.
doi: 10.12178/1001-0548.2020032
Abstract:
In recent years, the mental health of college students has been the focus of educators. There is no large-scale sample to analyze the impact of students' mental health on their social behavior. By analyzing 4 955 anonymous students' card swiping and psychological evaluation data of a university, we construct a social network among students based on their meal card swiping data. Furthermore, based on the data of SCL-90 evaluation scale, we evaluate the depression of students. Through analysis, we find that the degree of depression of students dramatically affects their social network structure. Students without obvious depression share meals with more students (speculating higher social activity), and students with obvious depression prefer to share meals with fewer students (inferring lower social activity).
In recent years, the mental health of college students has been the focus of educators. There is no large-scale sample to analyze the impact of students' mental health on their social behavior. By analyzing 4 955 anonymous students' card swiping and psychological evaluation data of a university, we construct a social network among students based on their meal card swiping data. Furthermore, based on the data of SCL-90 evaluation scale, we evaluate the depression of students. Through analysis, we find that the degree of depression of students dramatically affects their social network structure. Students without obvious depression share meals with more students (speculating higher social activity), and students with obvious depression prefer to share meals with fewer students (inferring lower social activity).
