• 北大核心期刊(《中文核心期刊要目总览》来源期刊)
  • 中国科技核心期刊(中国科技论文统计源期刊)
  • JST 日本科学技术振兴机构数据库(日)收录期刊

NoticeMore>

Special TopicMore >

Download CenterMore >

Wechat

Articles just accepted have been peer-reviewed and accepted, which are not yet assigned to volumes /issues, but are citable by Digital Object Identifier (DOI).
Display Method:
Articles in press have been peer-reviewed and accepted, which are not yet assigned to volumes /issues, but are citable by Digital Object Identifier (DOI).
Display Method:
Display Method:
Global filter pruning based on multi-source information
XIU Hui, XUE Lixia, WANG Ronggui, YANG Juan
2022, 39(9): 1-10.   doi: 10.19304/J.ISSN1000-7180.2022.0163
Abstract(0) HTML(0) PDF (0)
Abstract:

To solve the problem that the existing neural network pruning methods do not fully evaluate the importance of filters and there are some differences in the importance of cross-layer filters, a global filter pruning algorithm based on multi-source information is proposed, which establishes the connection between features and weights. Firstly, the relative and absolute importance of the filter are evaluated by the correlation between features and the entropy of weights, respectively, according to the characteristics of rich feature information and low influence of weight information on data noise. Then, the filters with different compression ratios in each layer are considered as a whole, their global importance to the model is evaluated, and the least important parts of the model are cut across layers according to the compression requirements. Finally, knowledge distillation is used to restore the accuracy of the model after pruning, and the model can be compressed and fine-tuned independently of other datasets. To verify the applicability of the proposed method, a large number of experiments are carried out on three semantic segmentation datasets for DeepLabV3, DABNet and U-Net networks. Verification is also carried out on the image classification dataset for various depths of ResNet networks. The experimental results show that the importance of filters in a single layer can be evaluated more accurately by using multi-source information, and the loss of key information can be minimized by using global importance to guide cross-layer pruning.

Feature relation dependent network for few-shot learning
LI Jingyu, WANG Ronggui, YANG Juan, XUE Lixia, DONG Bowen
2022, 39(9): 11-19.   doi: 10.19304/J.ISSN1000-7180.2021.0177
Abstract(0) HTML(0) PDF (0)
Abstract:

Few-shot learning aims to build a classifier that recognizes new unseen classes given only a few samples. Existing traditional metric learning methods map the samples to the shared embedding space, and calculate the feature similarity in this space for classification, but only map the features of samples independently while neglecting to observe the whole task. At the same time, the basic prototypes computed in the low-data regime are biased against the expected prototypes, resulting in low generalization on the query set. In view of the above problems, a feature relation dependent network is proposed (FRDN). The feature relation dependent network consists of two modules: Firstly, the relation mining module can fully mine the intra-class and inter-class relations in the task, use it as the self-attention values to adjust the class clusters to obtain a more discriminative task-adaptive embedding spaceand calculate basic prototypes; Then, the bias diminishing module is used to correct the initial prototype to obtain an optimized prototype with higher generalization on the query set, further improve the classification accuracy. On the MiniImagenet dataset, the 1-shot accuracy of the method is 59.17%, and the 5-shot accuracy is74.11%, which are 6.13% and 2.83% higher than that of the traditional metric learning method; on the CUB dataset, increases of 9.3% and 2.74% are reached respectively.

Spiking timing dependent plasticity algorithm with mixed reward-modulated signals
CHEN Yunxiang, FENG Ren, CHEN Yunhua
2022, 39(9): 20-25.   doi: 10.19304/J.ISSN1000-7180.2022.0108
Abstract(0) HTML(0) PDF (0)
Abstract:

In recent years, Spiking Timing-Dependent Plasticity (STDP) rules with physiological basis have been applied more and more in spiking neural networks. The R-STDP (reward-modulated STDP) learning algorithm combining STDP with the reinforcement learning reward modulation embraces great effect on improving the performance of SNN. However, the feedback only reflects on the last layer of spiking deep convolutional neural networks as the R-STDP algorithm works, which means the middle layer cannot get feedback. Inspired by the unsupervised characteristics of the Auto-Encoder, a mix reward-modulatedSTDP (MR-STDP) algorithm with mixed reward/punishment signal was proposed. In this algorithm, the reconstruction layer was added to the middle layer to establish the rewards/punishment signal factor model. The guiding factor signal is the similarity measure of spiking sequences issued by the neurons at the same position of the input layer of the interlayer autoencoder and the reconstruction layer, and it is combined with R-STDP, so that the middle layer can obtain the weight guiding signal. Experiments on MNIST and COVID-19 CT data sets shows that the proposed method achieves higher accuracy than R-STDP, and the efficiency of learning in middle layer is greatly improved.

Density peak clustering algorithm based on improved fruit fly optimization algorithm
YANG Shuangshuang, SHI Hongyan
2022, 39(9): 26-34.   doi: 10.19304/J.ISSN1000-7180.2021.1307
Abstract(0) HTML(0) PDF (0)
Abstract:

The cutoff distance of clustering by fast search and find of density peaks (DPC) requires manual intervention, and the selection of the parameters has great influence on the results of the algorithm. To overcome this problem, a density peak clustering method based on improved fruit fly optimization algorithm is proposed. The population of fruit fly is initialized by the Tent chaotic mapping, and using the characteristics of randomness, ergodicity and regularity of Tent chaotic sequence, the diversity of the initial population and the global exploration ability of the algorithm are enhanced. And the basic fruit fly optimization algorithm is improved by introducing dynamic step factor and Cauchy mutation strategy to enhance its local exploration ability and help the algorithm jump out of the local optimization. The convergence of the improved FOA algorithm is analyzed theoretically by using the convergence criterion of random algorithm. The experimental results of six test functions show that the improved FOA algorithm has faster convergence speed and higher solution accuracy. The improved FOA and DPC algorithm are fused into a new DPC algorithm, using the effective optimization ability of the improved FOA to find the best cutoff distance and realize the final clustering. Experimental results show that the clustering performance of new algorithm under UCI data set and artificial data set are improved, the new algorithm outperforms DPC, FOA-DPC, FADPC, ACS-FSDP with the better performance indexes, and the effect of manually selecting truncation distance parameter is effectively suppressed.

Feature selection method based on neighborhood rough sets and marine predator algorithm
GONG Rong, XIE Ningxin, LI Delun, HE Xuedong
2022, 39(9): 35-45.   doi: 10.19304/J.ISSN1000-7180.2022.0043
Abstract(0) HTML(0) PDF (0)
Abstract:

he feature selection method in rough set model has large computational overhead and can't directly handle continuous data, and Marine Predator Algorithm (MPA) still has some problems, such as slower convergence speed and easy to fall into local optimum. Therefore, a feature selection method based on neighborhood rough set (NRS) and Marine Predator Algorithm is proposed. Firstly, the original algorithm is improved by using opposition-based learning based on Tent Chaotic Map and Gaussian Perturbation strategy to obtain IMPA, and then a transmission mechanism is constructed to form a binary algorithm. Then, a fitness function is developed based on the neighborhood dependence in NRS and the length of feature subset. IMPA is used to iteratively search for the optimal feature subset, and a meta-heuristic feature selection algorithm is designed. Finally, the optimization performance of IMPA on 9 benchmark functions and the classification ability of feature selection algorithm on UCI data set is evaluated. Experimental results show IMPA is significantly better than Particle Swarm Optimization (PSO) algorithm and Salp Swarm Algorithm (SSA) in terms of average value and standard deviation. On UCI datasets, compared with the optimized feature selection algorithms based on rough sets and the optimized feature selection algorithms based on neighborhood rough sets, the average values of classification accuracy of the proposed feature selection method using KNN classifier is improved by 10.28~14.13 percentage points, 2.71~12.11 percentage points respectively. The average values of classification accuracy of the proposed feature selection method using CART classifier is improved by 9.41~13.24 percentage points, 2.90~12.31 percentage points respectively.

Dynamic gesture recognition based on lightweight (2+1)D convolution structure
ZHAO Kang, LI Xiangfeng, LI Gaoyang, ZUO Dunwen
2022, 39(9): 46-54.   doi: 10.19304/J.ISSN1000-7180.2022.0115
Abstract(0) HTML(0) PDF (0)
Abstract:

At present, great progress has been made in dynamic gesture recognition based on convolutional neural network. But neural network model has a large number of parameters, the cost of calculation and memory footprint is high, and it is difficult to apply for the occasion of limited equipment resources. In order to reduce the amount of calculation and parameter, a lightweight (2+1)D convolution structure is proposed. Based on the (2+1)D convolution structure, the 3D convolution is replaced by the 3D depthwise separable convolution. So the computation and parameter number of (2+1)D convolution structure are further reduced under the premise that the dimension of the output vector is unchanged. In order to make up for the deficiency of spatio-temporal features in the representation of dynamic gestures, attention mechanism module that focusing on the extraction of motion features is integrated. Combined with the spatio-temporal features that be extracted by the lightweight (2+1)D convolution structure, it can better represent gestures. Experimental results show that by inserting the attention mechanism module, the recognition accuracy of the model is further improved without increasing too much extra calculation and space cost. On 20BN-jester, EgoGesture and IsoGD datasets, the model based on the above structure achieved the recognition accuracy of 96.62%, 91.83% and 60.1%, respectively. The number of parameters and floating point of operations are 5.05M and 12.81GFLOPs respectively, which greatly reduces the calculation cost and memory footprint. Recognition speed is 70 frames per second in the real-time gesture recognition.

Self curing facial expression recognition based on multi-attention mechanism
WANG Wenxiang, ZHA Cheng, MIN Weidong, LU Zhuoqun, YU Guanghua
2022, 39(9): 55-62.   doi: 10.19304/J.ISSN1000-7180.2022.0029
Abstract(0) HTML(0) PDF (0)
Abstract:

Facial expression recognition technology has important application value and broad application prospects in social life, criminal detectives and other fields. Aiming at the problem of insufficient expression feature extraction in the existing methods, which makes high-dimensional features easy to lose local key information; And the ambiguity of expressions in complex backgrounds leads to weak network generalization. In order to solve these problems, a self curing network under multi-attention mechanism (MASCNet) is proposed. The network will generate multi-scale features with attention weights, and by fusing features of different scales, the ability of the network model to represent local key information at a fine-grained level is improved. The self-attention mechanism module can assign importance weights to the fused features, constrain the proportion of uncertain samples in network training, and improve the generalization ability of the network. The highest recognition accuracy rates of this method on the FER2013 and RAF-DB datasets are 74.21% and 88.74% respectively. Experimental results show that this method can effectively recognize facial expressions and is superior to the existing mainstream methods such as MHBP and AHBRPN.

Video prediction based on attention spatiotemporal decoupling 3D convolution LSTM
HUANG Jingui, HUANG Yiju
2022, 39(9): 63-72.   doi: 10.19304/J.ISSN1000-7180.2022.0023
Abstract(0) HTML(0) PDF (0)
Abstract:

To efficiently extract video spatio-temporal features to improve video prediction accuracy, an attentional spatio-temporal decoupling 3D convolutional LSTM algorithm is proposed. Firstly, the traditional 2D convolutional operation of the internal unit of convolutional LSTM is changed to 3D convolution to additionally extract short-term spatial motion information between video frames; and the correlation of long-term dynamic information between video frames is automatically captured by the attention mechanism. Since the Z-shaped transfer direction of feature information in the convolutional LSTM network in all layers leads to gradient disappearance, for this reason, inter-layer high-speed channels are added to the network structure to optimize the transfer process of video information flow between different inter-layer LSTM units. Meanwhile, temporal and spatial features in the network will interfere with each other to learn redundant functions, resulting in inefficient acquisition of feature information and degradation of network prediction quality, so temporal decoupling operations are added to the loss function to separate the learning of temporal and spatial features. For the data input process in the training encoding phase and the prediction decoding phase, data input resampling is proposed to reduce the differences between the encoder and decoder by using similar and opposite data input strategies in the model training and prediction phases. Experimental results on synthetic datasets as well as human action databases show that the algorithmic model has better performance in spatio-temporal feature extraction.

Design and implementation of spiking neural network based on FPGA
Xiao Yunkai, Zou Chengming
2022, 39(9): 73-79.   doi: 10.19304/J.ISSN1000-7180.2021.1312
Abstract(0) HTML(0) PDF (0)
Abstract:

Existing software simulations for spiking neural networks usually have the disadvantages of low processing speed and high-power consumption, while the hardware implementations have the disadvantages of high development difficulty and poor flexibility. To explore a reasonable implementation of the spiking neural networks, a novel method is proposed in which the network topology is simulated by the software simulation libraries, and the key computations are handed over to the FPGA forparallel computing to meet the requirements of easydevelopment, high flexibility, fast processing speed, and low power consumption. The main work is as follows: The software library and the OpenCL development library are extended, and the key modules of the software library are reconstructed into the FPGA kernels so that the software library can call the FPGA to execute the computing tasks. The experimental results on image classification of MNIST datasets show that the classification accuracy of the proposed scheme is comparable with that of the software simulation on GPU, and the reference power consumption is reduced by about 63.6% at the cost of a slight reduction in processing efficiency.

Multi-objective selection of analog circuit test points based on fault aliasing degree
ZHANG Yongchao, YOU Feng, WANG Weiwei, ZHAO Ruilian, SHANG Ying
2022, 39(9): 80-88.   doi: 10.19304/J.ISSN1000-7180.2022.0128
Abstract(0) HTML(0) PDF (0)
Abstract:

Test node selection methods in testability design of analog circuits usually use a voltage threshold to determine the fuzzy gap of all faults, but in fact, the voltage gap required by different faults is different.And for the voltage threshold setting problem, a clustering-based fault confusion calculation method is proposed to measure the degree of ambiguity between faults.On this basis, a multi-objective measurement point selection method for analogue circuits is designed to balance the number of measurement points and the number of faults isolated in order to obtain the maximum number of faults isolated with the lowest possible number of measurement points; Firstly, wavelet packet transform is used to extract the characteristics of analog signal. Then the aliasing degree between different faults is calculated by clustering; Finally, aiming at the fault isolation degree and the number of test points, the non-dominated sorting genetic algorithm NSGA-Ⅱ is used to search the test point set to realize the test point selection of analog circuit. The experimental results show that compared with the existing measurement point selection methods, this method can obtain the maximum number of fault isolation when the number of measurement points is as small as possible, and realize the optimization of measurement points.

Research on single-phase fault line selection in distribution network based on TCN+Transformer Self-Attention
OUYANG Yong, WAN Dou, GAO Rong, YE Zhiwei
2022, 39(9): 89-97.   doi: 10.19304/J.ISSN1000-7180.2021.1331
Abstract(0) HTML(0) PDF (0)
Abstract:

The single-phase fault routing problem of small-current grounding systems is an important problem in the fault of distribution network power systems. Due to the temporal continuity and excessive data length of power fault data and the fact that existing research work cannot effectively distinguish the characteristics of single-phase ground fault currents with timing. To overcome these problems, this paper proposed a hybrid neural network model based on a self-attentive TCN+Transformer (called TTHNN-SA model). Since the power fault data has a single feature, the use of wavelet transform decomposition and principal component analysis (PCA) methods can increase the number of features in the sample data. Therefore, the TTHNN-SA model used a temporal convolution network (TCN) to extract features by convolution operations on the original fault data and on the fault data decomposed by wavelet transform separately respectively, applied Transformer to extract the features of the fault data processed by PCA. Then the extracted feature matrices of the three models were fused and input to the self-attentive layer, and this layer assigned higher weights to the important features through matrix calculation, which can solve the long-time dependence problem of the model. Finally, the output of the self-attentive layer was pooled by global averaging and then classified using the softmax function. the TTHNN-SA model can learn the current data relationship between different waveform faults more comprehensively, and it had a good effect on the detection of single-phase faults in distribution networks.

New hybrid bird swarm algorithm for solving no-idle flow-shop scheduling problem
YAN Hongchao, TANG Wei, YAO Bin, CHENG Xuehong
2022, 39(9): 98-106.   doi: 10.19304/J.ISSN1000-7180.2022.0125
Abstract(0) HTML(0) PDF (0)
Abstract:

A new hybrid bird swarm algorithm (NHBSA) for solving no-idle flow-shop scheduling problem was proposed to minimize the makespan. Firstly, a Farahmand-Ruiz-Boroojerdian (FRB) heuristic was modified, the modified FRB heuristic and chaotic mapping were combined to ameliorate the quality and diversity of the population in the initialization phase. Secondly, the Smallest-Position-Value (SPV) rule was adopted to perform conversion between continuous position and discrete job permutation to make the algorithm suitable for dealing with discrete scheduling problems. In addition, to improve the convergence accuracy and the ability to avoid getting stuck in local optima of the algorithm, a local search method for the optimal job permutation of population was come up with by drawing on the ideas of variable neighborhood search and iterative greedy algorithm. Computational simulations and comparisons with several meta-heuristic algorithms for NFSP were carried out based on the widely used Taillard benchmark, the results show that the average percentage relative deviation (APRD) and the performance improvement percentage (PIP) obtained by NHBSA were reduced by 71.017% and 4.653%, respectively, under the premise of ensuring good stability.

Multi task intelligent legal judgment method based on Bert model
FAN Aman, WANG Yanchuan
2022, 39(9): 107-114.   doi: 10.19304/J.ISSN1000-7180.2022.0217
Abstract(0) HTML(0) PDF (0)
Abstract:

In recent years, artificial intelligence technology represented by deep learning and natural language processing has made great breakthroughs, which has promoted the development of judicial intelligence. Deep neural network technology has been widely used in the legal intelligence. Based on the Bert (bidirectional encoder representation from transformers) model, this paper uses natural language processing technology to learn the legal facts to realize the intelligent analysis in legal cases. Through the self-attention mechanism in BERT module, further understand the semantic information of the legal text from multiple feature dimensions to complete multiple judicial intelligent prediction functions such as crime prediction, legal clause recommendation and sentence prediction. With the multi-task learning, the correlation between various tasks is analyzed, which improves the extraction ability of legal text features, so that our model can achieve better generalization effect. Comparative experiments using public data sets verify the excellent performance of the method in the paper.

A noise reduction method for communication signals of SSMD based on Lagrange multipliers and SSA
LUO Min, ZHANG Jiashu
2022, 39(9): 115-124.   doi: 10.19304/J.ISSN1000-7180.2022.0122
Abstract(0) HTML(0) PDF (0)
Abstract:

Aiming at the difficulty of analysis and identification of communication signals under strong noise background, this paper proposes a joint denoising method of Symplectic Singular Mode Decomposition based on Lagrange multiplier (vSSMD) and Singular Spectrum Analysis (SSA). Considering that the random variation of noise makes the power spectral density method to calculate the embedding dimension with large error, this paper introduces the Monte Carlo idea to determine the embedding dimension. When the noise is large, vSSMD enhances useful components and suppresses noise components by constructing a Lagrangian multiplier matrix, and then adopts the SSA method to remove the weak noise in the reconstructed signal of vSSMD. The denoising effect of the vSSMD-SSA algorithm is compared with SSA and vSSMD methods. When the signal-to-noise ratio is -14dB, the signal-to-noise ratio of the vSSMD-SSA algorithm is increased by 4.49dB compared with the traditional algorithm SSA, and the mean square error is increased by 38.25%. The experimental results show that under the low signal-to-noise environment ratio, the vSSMD-SSA algorithm The denoising effect is the best. The vSSMD-SSA algorithm is used to denoise the UAV communication signal, and the noise reduction effect is the most obvious.

Implementation of a capacitive digital isolator based on OOK modulation
LIAO Kun, YANG Junzhong, JU Ling, YANG Sen, XIAO Zhiming
2022, 39(9): 125-132.   doi: 10.19304/J.ISSN1000-7180.2022.0113
Abstract(0) HTML(0) PDF (0)
Abstract:

To solve the problems of high current consumption of high speed capacitive digital isolators in "low speed" application, a low static power consumption fully differential digital isolator based on OOK modulation is designed based on TSMC 180nm BCD process. The proposed structure realizes the modulation of the input signal through three sets of switching signals generated in the logic control circuit of the transmitter and carrier signals generated by the oscillator module. When the transmission signal frequency changes, the midpoint potential bias circuit based on the cross-conductor ring structure can stabilize the DC voltage of the differential signal near VDD/2, so as to effectively avoid the error code caused by dc level attenuation at the receiving end. After amplification by a preamplifier, the receiver signal is demodulated by a dual threshold comparator.PVT simulation shows that the maximum transmission rate of 10Mbps can be achieved in the range of 3~5.5V input voltage, and the typical transmission delay is 13ns. The typical transmission delay is 13ns. The typical static power consumption is only 1.3 mA, and the typical dynamic power consumption is 4mA and 4.8mA at 1Mbps and 10Mbps respectively. This design supports multi-channel expansion, which can further reduce the average power consumption of single channel by sharing the internal oscillator and bias module. In addition, the isolator can decode correctly even at the highest 10Mbps input PRBS (Pseudo-random Binary Sequence) code, which proves that this structure has strong transmission robustness.

Found in 1972
Monthly

Supervisor:
Xi'an Institute of Microelectronics Technology

Sponsor:
China Aerospace Science and Technology Corporation

ISSN 1000-7180

CN 61-1123/TN