The fetal electrocardiogram(FECG) reflects health status of fetus. However, the FECG has not been widely used in clinical practice due to its relatively low signal-to-noise ratio(SNR). How to effectively extract high-quality fetal ECG signal remains a great challenge. In this paper, FECG extraction method utilizing improved LightGBM(light gradient boosting machine) with SNR regularization is proposed. Firstly, the raw maternal abdominal mixed signals are denoised by conventional filtering method. Then, FastICA(fast independent component analysis) is used to separate the maternal electrocardiogram(MECG) estimation and FECG estimation. The residual MECG component in FECG estimation is non-linear transform of MECG and the non-linear transform is fitted by the improved LightGBM with SNR regularization, which is constructed by adding a regularization term, the SNR of extracted FECG, to the objective function of conventional LightGBM. The SNR is estimated using cross correlation. By MECG estimation undergoing the fitted non-linear transform, the residual MECG component in FECG estimation is obtained. At last, the high-quality FECG is extracted by suppressing the estimated MECG component. The real data are adopted to verify the proposed FECG extraction method. The sensitivity, positive predictive value and
The detection of a pilot's brain fatigue state faces two important problems, which are how to generate a brain cognitive map and how to build a brain fatigue detection model. To solve the first problem, this paper uses the isometric azimuth projection to map brain fatigue indicators into a new type of brain power map. To solve the second problem, this work develops a deep latent Dirichlet model(DLDM), which solves the topic detection problem of pilot fatigue state. DLDM expands the probability distribution information contained in the developed brain power map layer by layer through multiple distributions, infers their hierarchical probability distribution characteristics, and gets more effective topic detection accuracy of pilot fatigue state. In order to avoid heuristic assumptions, this paper also proposes an effective stochastic gradient descent inference method with an adaptive learning rate between different layers and topics to more efficiently infer structure parameters of DLDM. The experimental results show that the DLDM can expand the probability distribution information of the brain power map layer by layer, infer richer abstract feature information, and detect brain fatigue cognitive topic. Compared with other brain fatigue detection methods, the classification accuracy of the proposed method can be improved by 2%.
Convolutional neural network involves in high computational complexity and excessive hardware resources, which greatly increases hardware deployment cost of deep learning algorithm. It is a promising scheme to make full use of the information redundancy of sparsity activation between layers can reduce the inference delay and power consumption with low resource overhead and almost lossless network accuracy. To solve low utilization problem of operation module caused by coarse-grained control in sparse convolution neural network accelerator, a sparsity-aware accelerator with flexible parallelism based on FPGA is designed. Convolution operation module is flexibly scheduled based on operation clustering idea,and the parallelism of input channel and output activation is adjusted online.In addition, a parallel propagation mode of input data is designed according to the data consistency during output activated parallel operation. The proposed hardware architecture is implemented on Xilinx VC709. It contains up to 1 024 multiplication and accumulation units and provides 409.6GOP/s peak computing power, and the operation speed is up to 325.8GOP/ s in VGG-16 model, which is equivalent to 794.63GOP/s of accelerator without sparse activation optimization. Its performance is 4.6 times more than that of baseline model.
In the current research of wireless rechargeable sensor networks, the charging planning problem often considers a single goal only, without considering the impact of the energy hole problem on the path planning of the wireless harging equipment(WCE) and the performance of the entire network. Based on the charging time window and penalty function, to maximize the energy utilization of WCE in the network and minimize the node penalty value, a multi-objective path planning model with time window is proposed to optimize the charging time of WCE to nodes and improve the energy utilization of WCE. To solve the multi-objective optimization problem, based on the multi-objective continuous firework algorithm, a multi-objective discrete firework algorithm is proposed, which speeds up the convergence speed and avoids falling into the local optimum through the firework explosion operation. The simulation results show that compared with the traditional algorithm, the Pareto optimal solution obtained by the algorithm proposed in this paper has at least 22.5% improvement in the distribution uniformity and 14.5% improvement in the performance of distribution range.
The researches of network representation learning have made many achievements. Since most of the researches ignore the dynamics and heterogeneity of the networks, coupled temporal and spatial structure features can not be distinguished, and rich semantic information of the network cannot be captured well. In this paper, meta-path based dynamic heterogeneous network representation learning method is proposed. Firstly, the neighborhood structures of nodes are divided into different sub-spaces according to their time, then the sequences of all time-weighted meta-paths for each node are sampled. Secondly, the neighborhood information on all time-weighted meta-paths of each node is integrated by a gated recurrent unit network(GRU). Furthermore, a bi-directional gated recurrent unit network(Bi-GRU) with an attention mechanism is used to learn the spatio-temporal contextual information from the merged sequences, and the final node representation will be received. Experiments on real data sets show that our algorithm has greatly improved performance on the downstream network tasks, such as node classification, clustering and visualization. Compared with state-of-the-art baseline methods,the Micro-F1 value has been raised by 1.09%~3.72% averagely on classification tasks, and the ARI value has been increased by 3.23%~14.49% on clustering tasks.
In this paper, an envelope optimization based on adaptive clustering for open-set camera model identification is proposed for the open-set problem of source camera identification, which solved the problem of low detection accuracy of the existing methods in the bad situation with few known camera models. Firstly, the clustering number of each type of camera data is obtained by the elbow method, and k-means clustering is performed with this clustering number as the reference. Then the sub-class data of the camera model are described by the technique of support vector data description, respectively to describe its hypersphere sub-envelope, and the sub-envelope is synthesized into a new hypersphere envelope with more detailed features according to the class of the camera model. Finally, the images from unknown camera models are excluded by the decision rule, and the images from known sources are classified and traceable to achieve source camera identification in the open-set. Experimental results on the two public datasets Dresden and SOCRatES show that the method proposed in this paper has better robustness and scalability. Compared with the existing methods, the three evaluation indicators of KACC, UACC, and OACC and time complexity are superior.
Dialogue state tracking is an important module of task-oriented dialogue system. Previous studies exploited the historical dialogue information by attention-based graph structure simulation, but these methods cannot explicitly take advantage of the structure of the dialogue state. In addition, how to generate complex format dialogue states also brings challenges to research. In this paper, we propose a state memory graph network(SMGN). The network saves historical information through the state memory graph, and uses the graph to interact with the current dialogue. We also implement a complex dialogue state generation method based on state memory graph. Experimental results show that the proposed method improves the joint accuracy by 1.39% on the CrossWOZ dataset and 1.86% on the MultiWOZ dataset.
Research on terahertz waveguide packaging technology and design verification in D-band (110~170GHz) and 220GHz band, respectively. Based on the wire bonding method, D-band LNA module is developed with self-designed amplifier chip. The module measurement shows the peak gain is 10.8dB at 139GHz，the gain higher than 7.8dB from 137GHz to 144GHz, the measured input return loss and output return loss are better than 5 dB and 8.5dB in operating frequencies, respectively. The tendency of packaged curve is same as the on-chip measured and its value is worse than the on-chip measurement about 5dB. A waveguide-to-integrated probe transition structure for terahertz band is proposed and verified in 220GHz. The module measurement shows the return loss is better than 8dB and the insertion loss is better than 3dB during 208GHz to 233GHz and the best performance is achieved at 224GHz with insertion loss is 1.3dB and return loss is 46.4dB. The waveguide-to-integrated probe transition structure provides experience for development of fully integrated terahertz chip.
Based on perfect Gaussian integer sequences and orthogonal matrices, a class of optimal periodic inter-group complementary(IGC) sequence set is constructed on the Gaussian integer set, which can realize the flexible number of complementary codes and complementary code sets within any zero-correlation zone length. By designing 2-order and 3-order kernel orthogonal matrices, splicing them on the diagonal way and filtering with perfect Gaussian integer sequences, a class of orthogonal matrices of arbitrary order is obtained. By using the orthogonal matrices and perfect Gaussian integer sequences of arbitrary length, the periodic inter-group complementary sequence set can be constructed, in which the length of zero correlation zone equals to that of perfect sequence and the set parameters reach the theoretical bound. Compared with the existing literature, the construction results show that the parameters can be optimized without any restriction. IGC sequence sets can be applied to the multi-carrier code division multiple access communication system to reduce adjacent cell interference and increase user capacity.
The indoor localization technology has important applications in many fields, while traditional wireless local area network(WLAN) fingerprint-based localization methods usually rarely consider both the diversity of WLAN received signal strength(RSS) features and the difference of the position resolution of RSS features from different access points(APs), which results in the low localization accuracy and efficiency. To address this problem, this paper proposes a WLAN indoor localization method based on the multi-dimensional fuzzy mapping for the AP optimization. Specifically, in the offline phase, the information gain ratio of the AP and the corresponding offline fuzzy membership degree are calculated according to the multi-dimensional RSS features which are extracted many times, and meanwhile the fuzzy relationship equation is utilized to solve out fuzzy weights of multi-dimensional RSS features. In the online phase, the fuzzy decision matrix is constructed by the multi-dimensional fuzzy mapping to calculate the online fuzzy membership degree of the AP, and then the target location estimation is realized by combining with the K-nearest neighbor(KNN) algorithm. Experimental results show that compared with the traditional AP optimization based localization methods, the localization calculation overhead in the online stage of the proposed method is reduced by up to 4.12 s, and the confidence probability of the positioning error within 4 meters is 91.91%.
Dialogue sarcasm recognition has been a challenging artificial intelligence(AI) research topic, aiming to discover elusive ironic, contemptuous and metaphoric information implied in daily dialogue. From the perspective of emotional logic, most existing works are insufficient to measure the intrinsic uncertainty in emotional expression and understanding. In view of the advantages of quantum probability(QP) in modeling the uncertainty, this paper explores the potential of QP in dialogue sarcasm recognition and proposes a quantum probability inspired network(QPIN). Specially, QPIN consists of a complex-valued embedding layer, a quantum composition layer, a quantum measurement layer and a dense layer. Each utterance is treated as a quantum superposition-like of a set of basis words, using a complex-valued representation. The contextual interaction between adjacent utterances is described as the composition system between a quantum system and its surrounding environment, which is represented by the density matrix. A quantum measurement is performed on the density matrix of each utterance to extract sarcastic features, and thus feeds these features to a dense layer to yield the probabilistic outcomes. Extensive experiments are conducted on two benchmark datasets, and the results show that our model outperforms the state-of-the-art baselines, with accuracy scores enhanced by 5.2% and 2.38%, respectively.
We have to face the challenge of missing small targets and severe edge noise in semantic segmentation. The existing semantic segmentation algorithms that enhance feature representation and optimize spatial details have difficulty to accurately segment the small targets and edges as the algorithms insufficiently gain detail information from tiny targets and semantic edges. This paper presents a tiny target feature enhancement algorithm for semantic segmentation. Specifically, a pixel spatial attention module(PAM) is designed to obtain strong semantic information from low-level pixel space. Semantic category information including edge features and tiny target features are obtained by modeling mask, respectively. A special loss function is designed for model training and the features gained by the model are fused with the features obtained from above way. Through edge feature enhancement, inner contour noise reduction, the segmentation performance of tiny target is improved while other segmentation categories are not degraded. Experimental results on Cityscapes, VOC2012, ADE20K and Camvid show that the proposed algorithm performance has been significantly improved by 2% in comparison with other state-of-the-art algorithms in the same scene.
Due to the complexity of pedestrian interaction and the variability of the surrounding environment, pedestrian trajectory prediction is still a challenging task. However, when modeling pedestrian interaction based on graph structure, there are some problems, such as small sensing field of the network, symmetrical interaction between pedestrians, and fixed graph structure that can not adapt to scene changes, which lead to a large deviation of the predicted trajectory from the real trajectory. To solve these problems, a pedestrian trajectory prediction method based on global adaptive directed graph is proposed. Global feature updating(GFU) and global feature selection(GFS) are designed to improve the perception range in spatial and temporal domain respectively and get global interaction features. A directed feature graph is constructed to define the asymmetric interaction between pedestrians and improve the directionality of network modeling. An adaptive graph model is established to flexibly adjust the relationship between pedestrians, reduce redundant connections and enhance the adaptive ability of the graph. The experimental results on ETH and UCY datasets show that comparing with the optimal value, the average displacement error is reduced by 14% and the final displacement error is reduced by 3%.
In this paper, the dual coupled polar(DC-Polar) code is proposed and its lower bound is derived. Based on this, DC-Polar code combined with bit-interleaved coded modulation(BICM) is presented for high-order constellations. To overcome the rate matching problem, a new coded modulation scheme called DC-Polar two-layer coding(DC-Polar-TLC) is proposed. The advantage of this scheme is that the code length can be adjusted flexibly without applying puncturing or shortening methods. It is shown by numerical simulations that the proposed DC-Polar-TLC scheme has good performance. The presented scheme can achieve about 0.6 dB performance gains compared to the bit-interleaved coded modulation scheme with CRC-aided polar code for the case of 64QAM.
Social network alignment aims to identify social accounts belonging to the same natural person from different social networks. Most of the existing related researches focus on the alignment of static social networks. However, social networks are dynamically evolving. We observe that dynamics can reveal more discriminative patterns and thus can benefit social network alignment. This phenomenon motivates us to rethink this issue in dynamic scenarios. Therefore, we propose to leverage the dynamics of social networks and design a deep learning architecture to address the dynamic social network alignment problem, termed as DeepDSA. Specifically, we first design a deep sequence model to capture the dynamics of social network structure and attributes respectively. For each social network, we merged binary dynamics by maintaining the correlation between structure and attributes of the same user to obtain the original comprehensive embeddings. We finally perform spatial transformation learning in a semi-supervised manner, and project the original embedding of each network into a target subspace in which a natural person is uniquely represented. We conduct extensive experiments on real-world datasets and demonstrate the proposed DeepDSA achieves 10% improvement of precision against the current mainstream algorithm.
Aiming at the two-side assembly line balancing problem of type-Ⅱ in manufacturing, this paper establishes a model whose primary and secondary optimization objectives are cycle time and smoothing index, and proposes an enhanced estimation of distribution algorithm(EEDA) for it. In the initialization, an adaptive strategy is devised to generate the initial cycle time for improving the quality of initial solutions. In the global search, the probability model is used to learn the information of solution, and sample the probability model to generate a new solution. In the local search, a strategy suitable for primary and secondary objectives is developed to further execute a thorough search in promising regions. Meanwhile, by analyzing the characteristics of the problem, a fast judgment method that can determine whether the solutions are feasible is proposed. Simulation experiments verify the effectiveness of the EEDA and the importance of primary and secondary objectives.
A* algorithm is one of the important and commonly used algorithms in path planning of robots. In large map with complex terrain, large-scale node expansion is needed in A* algorithm to find a feasible optimization path in the case of invisibility between path points. This leads to a sharp increase in the demand for storage space and a decrease in running efficiency of the algorithm. To solve low-level path planning problems based on CBS(Conflict-Based Search) framework, an A* path planning algorithm with low node expansion is proposed. Triangulation method and fixed obstacle processing method are introduced to obtain an optimized path that is visible to adjacent points. Moreover, a segmentation strategy is proposed based on integrated visibility optimization to plan the segmented path according to the adjacent visible points with dynamic collision processing ability. Simulation experiments on standard map data sets under complex terrain show that the path length of the proposed algorithm in this paper is 98.1%~102.2% as long as that of A* algorithm, compared with which, the amount of node expansion reduces by 85.4% and the running time of algorithm reduces by 58.1%.
To improve the efficiency of community detection in large-scale network, a streaming-based overlapping community detection algorithm(SOCD) is proposed. SOCD processes the edges in a streaming fashion and handles one edge at a time, and each edge is processed only once. The algorithm discovers the community based on the degree of node, the contribution of node to the community, and the change of the number of connected edges between communities. Extensive experiments on synthetic networks and real large-scale networks show that SOCD has great advantages in time consumptions and memory footprints. It is more than 10 times faster than traditional methods and has strong robustness. It can efficiently and accurately mine the overlapping community structures in the network under linear time and space complexity.
In the many-objective evolutionary algorithm, recombination operators are usually used to generate high-quality offspring to guide the population search. Previous studies have shown that using similar individuals to reorganize can improve the quality of individual offspring. Since the self-organizing maping(SOM) network can maintain the original topological relationship of the population individuals and obtain the similar information of the individuals through clustering, this paper proposes a many-objective evolutionary algorithm based on SOM clustering and adaptive operator selection(MaOEA-SCAOS). First, the proposed method use self-organizing mapping network to classify the population, extract individual data structure information, and use similarity to build a neighborhood mating pool. Then the method select the adaptive operator based on the individual dominance information in the class to improve the search and convergence performance. Finally, the environmental selection strategy is adopted to manage the diversity of the population to ensure that the population is evenly distributed in the Pareto front. The experimental simulation results show that the SOM clustering and adaptive operator selection(SCAOS) method proposed in this paper has strong competitiveness while dealing with many-objective optimization problems, and the overall performance index is better than other methods.
Business process prediction can effectively facilitate enterprises to control processes and deliver high-quality services. As one of the core tasks of process prediction, remaining time prediction has been widely concerned by scholars. Currently, traditional long short-term memory(LSTM) neural networks have been used to predict the remaining time of business process instances. However, due to the lack of parallelism and limited modeling ability of LSTM in processing sequence data,the accuracy of prediction has further room to improve. In this paper, the remaining time prediction method based on bidirectional quasi-recurrent neural network with attention is proposed. Firstly, this method uses the bidirectional quasi-recurrent neural network to build the prediction model, and adds the attention mechanism to the model enhances the characteristic information of the bidirectional quasi-recurrent neural network output.Secondly, a training iteration strategy based on different length trace prefixes is designed, which solves the problem of the difference in the number of trace prefixes of different lengths. Finally, event representation learning method is proposed, to achieve vectors representation of similarity to the same traces and frequent events, improves the accuracy of the remaining time prediction. Experiments on five public event log datasets show this method has improved the accuracy of prediction by an average of nearly 15%, and the average training time is reduced by about 26%, compared with the existing methods.
To realize the rapid detection and evaluation of dural hematoma, this study is based on the near infra-red spectroscopy(NIRS) and Monte Carlo algorithm to establish a simulation of 5 layers of brain tissue suitable for the positive problem. Source-detector sensitivity(SDS) parameters are defined to construct a forward function matrix based on the power-law attenuation model and the exponential model. These parameters are given by the number of transmitted photons from 10 groups of hematoma thickness models via the controlled variable method. The mathematical model reconstruction of the data obtained from the forward problem is the realization of the inverse problem. The 6 groups of hematoma thicknesses are used for reverse theoretical calculations. Compared with the reference value, the thickness of hematoma is less than 0.7 cm, and the average absolute error of the two function models is less than 3.6%. The thickness is equal to 0.75 cm, the average absolute error of the power-law attenuation model is less than 4.3%, and the error is 6.3886% less than the exponential law attenuation model. The results show that the method is feasible for accurately determining whether hematoma is contained and predicting the thickness of hematoma. The detection sensitivity and detection distance have a logarithmic correlation. The power-law attenuation model is closer to the reference value than the exponential law attenuation model, and the build effect is better.
Semiconductor science and information devices is one of the most important fields of science and technology, which is becoming increasingly important in the international competition environment. The national natural science foundation of China(NSFC) supports basic and applied basic research in semiconductor science and information devices, gradually forming the funding content consisting of free exploration research projects, guidance-guided research projects, talents projects, and national major scientific instrument development projects. We analyze the application and funding status of various types of projects under the F04 code of semiconductor science and information devices during the 13th to 14th Five-Year Plan period (2016-2021). Through the hot word cloud analysis of funded project topics over these years, the research directions and hot topics of the semiconductor field are shown. Based on the outstanding research achievements in semiconductor science and information devices from the 13th to 14th Five-Year Plan, the effects of the funding are demonstrated. This article is expected to show the characteristics of the funding of the Division of Information Science to semiconductor science and information devices in recent years, and to provide reference for researchers in domestic research institutes, enterprises and institutions to understand the research hot spots, future development direction and path in this field.
Expression action unit(AU) recognition based on deep learning is a hot topic in the fields of computer vision and affective computing. Each AU describes a facial local expression action, and the combinations of AUs can quantitatively represent any expression. Current AU recognition mainly faces three challenging factors, scarcity of labels, difficulty of feature capture, and imbalance of labels. On this basis, this paper categorizes the existing researches into transfer learning based, region learning based, and relation learning based methods, and comments and summarizes each category of representative methods. Finally, this paper compares and analyzes different methods, and further discusses the future research directions of AU recognition.
Group behavior recognition is currently a research hotspot in the field of computer vision, and has a wide range of applications in intelligent security monitoring, social role understanding, and sports video analysis. This article mainly reviews group behavior recognition algorithms based on deep learning framework. Firstly, by judging “whether a method including group member interaction relationship modeling”, it can be classified as “group behavior recognition without interaction relationship modeling(GBRWIR)” or “group behavior recognition based on interaction relationship description(GBRBIR)”. Secondly, because GBRWIR mainly focuses on how to design “calculation and purification of overall spatiotemporal characteristics of a group behavior sequence”, this article summarizes it as the following three typical algorithms, which are “multi-stream spatiotemporal feature calculation fusion”, “individual/group multi-level spatiotemporal feature calculation and merging”, and “group behavior spatiotemporal feature purification based on attention mechanism” respectively. Thirdly, for GBRWIR algorithms, depending on its different descriptions of interaction relationship, it can be summarized respectively as “based on group member global interaction relationship modeling”, “based on group division and subgroup interaction modeling”， and “modeling of interactions between core members”. Then, the data sets related to group behavior recognition are introduced, and the test performances of different recognition methods in each data set are compared and summarized. Finally, several challenging issues and future research directions are discussed, which respectively are the duality of group behavior category definition, the difficulty of interactive relationship modeling, the weakly supervised labeling and self-learning of group behavior recognition, and the changes of viewpoint and the comprehensive utilization of scene information.
The distribution of traditional spatiotemporal chaotic system is relatively concentrated, and the uniformity of its generating sequence is poor. In this paper, a new spatiotemporal chaotic system with perturbed one-way coupled map lattice is constructed based on elementary cellular automata. The numerical simulation results of the distribution diagram and phase diagram of the system show that the perturbed system can improve the uniformity of the original system and increase the dynamic complexity of the system. A dynamic S-box generation algorithm is designed based on the homogenized disturbed spatiotemporal chaotic system, and the dynamic S-box is generated according to the dynamic update strategy. The statistical analysis of nonlinearity, strict avalanche criterion and differential uniformity of the S-box generated by the algorithm is carried out. The results show that the dynamic S-box generated by the homogenized disturbed spatiotemporal chaotic system is more secure.
With focus on blind estimation of the pseudo-noise(PN) sequence of a non-periodic long code direct sequence spread spectrum(NPLC-DSSS) signal, a blind estimation approach of PN sequence and synchronous position of information code is proposed based on mean similarity. In this method, the information code library is constructed, the average similarity is used to synchronize the information codes, and eigenvalue decomposition is used to estimate the PN sequence. Simulation results show that compared with the existing algorithms under the same condition, the proposed algorithm not only improves estimation performance by 1 dB, but also can estimate synchronous position of information code and PN sequence jointly.
