基于改进遗传算法的天波超视距雷达二维阵列稀疏优化设计

针对均匀矩形平面阵在2维阵天波超视距雷达(OTHR)工程应用时阵元数目庞大问题,该文提出利用改进遗传算法进行稀疏优化设计的解决办法。从波束赋形角度出发建立了稀疏矩形平面阵的布阵优化模型;以俯仰波束能分辨OTHR多模传播的波达角(DOA)为原则确定了遗传算法的初始种群;为避免遗传进化中出现早熟和随机漫游现象修正了适应度函数;为实现稀疏率精确控制改进了交叉和变异算子。仿真结果表明,该文算法不仅实现了稀疏率的精确控制,同时提高了优化性能。该文最后对2维稀疏阵列在OTHR工程应用时的可行性进行了分析,指出其应用条件和存在的技术难点,并给出了相应的解决方案。

一种稀疏性增强的人脸识别改进算法

针对在现有稀疏表示分类(SRC)算法中,用l1范数取代l0范数并不能得到有效的稀疏解,提出了一种将lp(0≤p<1)范数和新判别规则完美结合的人脸识别方法。首先通过迭代算法求解lp范数最小化问题,以此代替传统SRC中的l1范数来求解编码系数,得到更稀疏和有效的解。为了从稀疏编码系数中捕捉到更多的差分信息,并兼顾残差反映每一类样本的贡献,用系数和与残差之比这一新判别规则来分类测试样本。在AR人脸数据库的实验结果表明,本算法可得到更稀疏有效的解,且可在一定程度上提高识别率,尤其是在伪装情况下,有较为明显的提高。

基于Z-Score动态压缩的高效联邦学习算法

联邦学习作为一种具有隐私保护的新兴分布式计算范式,在一定程度上保护了用户隐私和数据安全。然而,由于联邦学习系统中客户端与服务器需要频繁地交换模型参数,造成了较大的通信开销。在带宽有限的无线通信场景中,这成为了限制联邦学习发展的主要瓶颈。针对这一问题,提出了一种基于Z-Score的动态稀疏压缩算法。通过引入Z-Score,对局部模型更新进行离群点检测,将重要的更新值视为离群点,从而将其挑选出来。在不需要复杂的排序算法以及原始模型更新的先验知识的情况下,实现模型更新的稀疏化。同时随着通信轮次的增加,根据全局模型的损失值动态地调整稀疏率,从而在保证模型精度的前提下最大程度地减少总通信量。通过实验证明,在I.I.D.数据场景下,该算法与联邦平均(FedAvg)算法相比可以降低95%的通信量,精度损失仅仅为1.6%,与FTTQ算法相比可以降低40%~50%的通信量,精度损失仅为1.29%,证明了该方法在保证模型性能的同时显著降低了通信成本。

基于强化学习和3σ准则的组合剪枝方法

针对结构复杂、参数冗余的深度神经网络无法部署到资源受限的嵌入式系统的问题,受稀疏率对性能影响的启示,提出基于强化学习和3σ准则的组合剪枝方法.根据稀疏率对准确率的影响,确定最佳全局稀疏率,使稀疏率和精度达到较好平衡.在最佳全局稀疏率的指导下,利用强化学习方法自动搜索每层卷积层的最佳剪枝率,根据剪枝率剪去不重要的权重.通过3σ准则确定全连接层每层的权重剪枝阈值,对全连接层进行权重剪枝.通过再训练来恢复模型识别的精度.实验结果表明,所提剪枝方法可以将网络VGG16、ResNet56和ResNet50的参数,分别压缩83.33%、70.1%和80.9%,模型的识别准确率分别降低1.55%、1.98%和1.86%.

面向5G的LDPC码正则校验矩阵设计研究

3GPP已确定LDPC码为eMBB场景数据业务信道长码块编码方案,正则校验矩阵是LDPC正则编码的关键。正则校验矩阵是一种稀疏矩阵,分解成6个子阵后,既简化了正则校验矩阵的设计,又方便了发射端的数据编码和接收端的数据解码。在讨论设计正则校验矩阵的基本条件后,对基于近似下三角形奇偶校验的正则校验矩阵的编码进行了分析,在此基础上设计了7款正则校验矩阵,并通过仿真全面分析了这些正则校验矩阵的性能及设计中的注意事项,指出了正则校验矩阵的编码译码性能所依靠的重点,为研究LDPC编码提供了重要参考。

落叶松人工林首次间伐期的确定

分析了落叶松人工林首次间伐时间的确定及其影响因素,以多年植林及间伐的数据为基础,给出了有统计意义的有价值的数据,用以指导人工林经营。

Multichannel deconvolution with spatial refl ection regularization

Seismic deconvolution plays an important role in the seismic characterization of thin-layer structures and seismic resolution enhancement.However,the trace-by-trace processing strategy is applied and ignores the spatial connection along seismic traces,which gives the deconvolved result strong ambiguity and poor spatial continuity.To alleviate this issue,we developed a structurally constrained deconvolution algorithm.The proposed method extracts the refl ection structure characterization from the raw seismic data and introduces it to the multichannel deconvolution algorithm as a spatial refl ection regularization.Benefi ting from the introduction of the reflection regularization,the proposed method enhances the stability and spatial continuity of conventional deconvolution methods.Synthetic and field data examples confi rm the correctness and feasibility of the proposed method.

Single frame super-resolution reconstruction based on sparse representation

In order to effectively improve the quality of recovered images, a single frame super-resolution reconstruction method based on sparse representation is proposed. The combination method of local orientation estimation-based image patch clustering and principal component analysis is used to obtain a series of geometric dictionaries of different orientations in the dictionary learning process. Subsequently, the dictionary of the nearest orientation is adaptively assigned to each of the input patches that need to be represented in the sparse coding process. Moreover, the consistency of gradients is further incorporated into the basic framework to make more substantial progress in preserving more fine edges and producing sharper results. Two groups of experiments on different types of natural images indicate that the proposed method outperforms some state-of- the-art counterparts in terms of both numerical indicators and visual quality.

Short-term photovoltaic power prediction using combined K-SVD-OMP and KELM method

For photovoltaic power prediction,a kind of sparse representation modeling method using feature extraction techniques is proposed.Firstly,all these factors affecting the photovoltaic power output are regarded as the input data of the model.Next,the dictionary learning techniques using the K-mean singular value decomposition(K-SVD)algorithm and the orthogonal matching pursuit(OMP)algorithm are used to obtain the corresponding sparse encoding based on all the input data,i.e.the initial dictionary.Then,to build the global prediction model,the sparse coding vectors are used as the input of the model of the kernel extreme learning machine(KELM).Finally,to verify the effectiveness of the combined K-SVD-OMP and KELM method,the proposed method is applied to a instance of the photovoltaic power prediction.Compared with KELM,SVM and ELM under the same conditions,experimental results show that different combined sparse representation methods achieve better prediction results,among which the combined K-SVD-OMP and KELM method shows better prediction results and modeling accuracy.

Comparison between several seismic inversion methods and their application in mountainous coal fields of western China

With the objective of establishing the necessary conditions for 3D seismic data from mountainous areas in western China, we compared the application results of wave impedance technology in the lithology and exploration of coal fields. First, we introduce principles and features of three kinds of inversion methods. i.e., Model-Based Inversion, Constrained Sparse Spike Inversion (CSSI) and Geology-Seismic Feature Inversion. Secondly, these inversion methods are contrasted in their application to 3D seismic data from some coalfields in western China. The main information provided by the research includes: improving the vertical resolution of coal deposit strata, inferring lateral variation of the lithology and predicting coal seams and their roof lithology. Finally, the comparison between the three methods shows that the model-based inversion has the higher resolution, while CSSI inversion has better waveform continuity. The geology-seismic feature inversion requires information from a large number of wells and many types of logging curves of good quality. All three methods can meet the requirements of seismic exploration for lithological exploration in coal fields.

Single Image Super-Resolution by Clustered Sparse Representation and Adaptive Patch Aggregation

A Single Image Super-Resolution (SISR) reconstruction method that uses clustered sparse representation and adaptive patch aggregation is proposed. First, we randomly extract image patch pairs from the training images, and divide these patch pairs into different groups by K-means clustering. Then, we learn an over-complete sub-dictionary pair offline from corresponding group patch pairs. For a given low-resolution patch, we adaptively select one sub-dictionary to reconstruct the high resolution patch online. In addition, non-local self-similarity and steering kernel regression constraints are integrated into patch aggregation to improve the quality of the recovered images. Experiments show that the proposed method is able to realize state-of-the-art performance in terms of both objective evaluation and visual perception.

Finite rate of innovation sparse sampling for a binary frequency-coded ultrasonic signal

To achieve sparse sampling on a coded ultrasonic signal,the finite rate of innovation(FRI)sparse sampling technique is proposed on a binary frequency-coded(BFC)ultrasonic signal.A framework of FRI-based sparse sampling for an ultrasonic signal pulse is presented.Differences between the pulse and the coded ultrasonic signal are analyzed,and a response mathematical model of the coded ultrasonic signal is established.A time-domain transform algorithm,called the high-order moment method,is applied to obtain a pulse stream signal to assist BFC ultrasonic signal sparse sampling.A sampling of the output signal with a uniform interval is then performed after modulating the pulse stream signal by a sampling kernel.FRI-based sparse sampling is performed using a self-made circuit on an aluminum alloy sample.Experimental results show that the sampling rate reduces to 0.5 MHz,which is at least 12.8 MHz in the Nyquist sampling mode.The echo peak amplitude and the time of flight are estimated from the sparse sampling data with maximum errors of 9.324%and 0.031%,respectively.This research can provide a theoretical basis and practical application reference for reducing the sampling rate and data volume in coded ultrasonic testing.

Traffic danger detection by visual attention model of sparse sampling

A method to detect traffic dangers based on visual attention model of sparse sampling was proposed. The hemispherical sparse sampling model was used to decrease the amount of calculation which increases the detection speed. Bayesian probability model and Gaussian kernel function were applied to calculate the saliency of traffic videos. The method of multiscale saliency was used and the final saliency was the average of all scales, which increased the detection rates extraordinarily. The detection results of several typical traffic dangers show that the proposed method has higher detection rates and speed, which meets the requirement of real-time detection of traffic dangers.

基于l^p范数稀疏增强的人脸识别算法

针对稀疏表示算法横纵采用l^1范数以及l^0范数并不能得到有效的稀疏解的问题,提出了一种将l p(0≤p<1)范数的人脸识别方法。首先通过迭代算法求解l^p范数最小化问题,以此代替传统SRC中的l^1范数来求解编码系数,得到更稀疏和有效的解。为了从稀疏编码系数中捕捉到更多的差分信息,并兼顾残差反映每一类样本的贡献,用系数和与残差之比这一新判别规则来分类测试样本。在AR人脸数据库的实验结果表明,本算法可得到更稀疏有效的解,且可在一定程度上提高识别率,尤其在伪装情况下,有较为明显的提高。

Rapid earthquake focal mechanism inversion using high-rate GPS velometers in sparse network

In this study, we demonstrate an approach for inverting earthquake source parameters based on high-rate global positioning system （GPS） velocity seismograms. The velocity records obtained from single-station GPS velocity solutions with broadcast ephemeris are used directly for earthquake source parameter inversion using the Cut and Paste method, without requiring conversion of the velocity records into displacement records. Taking the E1 Mayor-Cucapah earthquake as an example, GPS velocity records from 10 stations with reasonable azimuthal coverage provide earthquake source parameters very close to those from the Global centroid moment tensor （Global CMT） solution. In sparse network tests, robust source parameters with acceptable bias can be achieved with as few as three stations. When the number of stations is reduced to two, the bias in rake angle becomes appreciable, but the magnitude and strike estimations are still robust. The results of this study demonstrate that rapid and reliable estimation of earthquake source parameters can be obtained from GPS velocity data. These parameters could be used for early earthquake warning and shake map construction, because such GPS velocity records can be obtained in real time.

Self-thinning Rules at Chinese Fir(Cunninghamia lanceolata) Plantations——Based on a Permanent Density Trial in Southern China

Data selection and methods for fitting coefficients were considered to test the self-thinning law. TheChinese fir (Cunninghamia lanceolata) in even-aged pure stands with 26 years of observation data were applied tofit Reineke's (1933) empirically derived stand density rule (No∝d^-1.605, N = numbers of stems, d= mean diameter),Yoda's (1963) self-thinning law based on Euclidian geometry (v ∝ N^-3/2, v= tree volume), and West, Brown andEnquist's (1997, 1999)(WBE) fractal geometry (w ∝ d^-8/3). OLS, RMA and SFF algorithms provided observedself-thinning exponents with the seven mortality rate intervals (2%--80%, 5%--80%, 10%- 80%, 15%--80%,20%- 80%, 25%--80% and 30%- 80%), which were tested against the exponents, and expected by the rules con-sidered. Hope for a consistent allometry law that ignores species-specific morphologic allometric and scale differ-ences faded. Exponents a of N ∝ d^α, were significantly different from -1.605 and -2, not expected by Euclidianfractal geometry;exponents β of w ∝ N^β varied around Yoda's self-thinning slope - 3/2, but was significantly differentfrom - 4/3;exponent Y of w ∝ d^γ tended to neither 8/3 nor 3.

An efficient prediction framework for multi-parametric yield analysis under parameter variations

Due to continuous process scaling, process, voltage, and temperature （PVT） parameter variations have become one of the most problematic issues in circuit design. The resulting correlations among performance metrics lead to a significant parametric yield loss. Previous algorithms on parametric yield prediction are limited to predicting a single-parametric yield or performing balanced optimization for several single-parametric yields. Consequently, these methods fail to predict the multiparametric yield that optimizes multiple performance metrics simultaneously, which may result in significant accuracy loss. In this paper we suggest an efficient multi-parametric yield prediction framework, in which multiple performance metrics are considered as simultaneous constraint conditions for parametric yield prediction, to maintain the correlations among metrics. First, the framework models the performance metrics in terms of PVT parameter variations by using the adaptive elastic net （AEN） method. Then the parametric yield for a single performance metric can be predicted through the computation of the cumulative distribution function （CDF） based on the multiplication theorem and the Markov chain Monte Carlo （MCMC） method. Finally, a copula-based parametric yield prediction procedure has been developed to solve the multi-parametric yield prediction problem, and to generate an accurate yield estimate. Experimental results demonstrate that the proposed multi-parametric yield prediction framework is able to provide the designer with either an accurate value for parametric yield under specific performance limits, or a multi-parametric yield surface under all ranges of performance limits.

A two-level method for sparse time-frequency representation of multiscale data Dedicated to Professor LI TaTsien on the Occasion of His 80th Birthday

Based on the recently developed data-driven time-frequency analysis(Hou and Shi, 2013), we propose a two-level method to look for the sparse time-frequency decomposition of multiscale data. In the two-level method, we first run a local algorithm to get a good approximation of the instantaneous frequency. We then pass this instantaneous frequency to the global algorithm to get an accurate global intrinsic mode function(IMF)and instantaneous frequency. The two-level method alleviates the difficulty of the mode mixing to some extent.We also present a method to reduce the end effects.