Robust adaptive radar beamforming based on iterative training sample selection using kurtosis of generalized inner product statistics

In engineering application,there is only one adaptive weights estimated by most of traditional early warning radars for adaptive interference suppression in a pulse reputation interval(PRI).Therefore,if the training samples used to calculate the weight vector does not contain the jamming,then the jamming cannot be removed by adaptive spatial filtering.If the weight vector is constantly updated in the range dimension,the training data may contain target echo signals,resulting in signal cancellation effect.To cope with the situation that the training samples are contaminated by target signal,an iterative training sample selection method based on non-homogeneous detector(NHD)is proposed in this paper for updating the weight vector in entire range dimension.The principle is presented,and the validity is proven by simulation results.

Sampling Methods for Efficient Training of Graph Convolutional Networks:A Survey

Graph convolutional networks(GCNs)have received significant attention from various research fields due to the excellent performance in learning graph representations.Although GCN performs well compared with other methods,it still faces challenges.Training a GCN model for large-scale graphs in a conventional way requires high computation and storage costs.Therefore,motivated by an urgent need in terms of efficiency and scalability in training GCN,sampling methods have been proposed and achieved a significant effect.In this paper,we categorize sampling methods based on the sampling mechanisms and provide a comprehensive survey of sampling methods for efficient training of GCN.To highlight the characteristics and differences of sampling methods,we present a detailed comparison within each category and further give an overall comparative analysis for the sampling methods in all categories.Finally,we discuss some challenges and future research directions of the sampling methods.

Virtual sample generation for model-based prognostics and health management of on-board high-speed train control system

In view of class imbalance in data-driven modeling for Prognostics and Health Management(PHM),existing classification methods may fail in generating effective fault prediction models for the on-board high-speed train control equipment.A virtual sample generation solution based on Generative Adversarial Network(GAN)is proposed to overcome this shortcoming.Aiming at augmenting the sample classes with the imbalanced data problem,the GAN-based virtual sample generation strategy is embedded into the establishment of fault prediction models.Under the PHM framework of the on-board train control system,the virtual sample generation principle and the detailed procedures are presented.With the enhanced class-balancing mechanism and the designed sample augmentation logic,the PHM scheme of the on-board train control equipment has powerful data condition adaptability and can effectively predict the fault probability and life cycle status.Practical data from a specific type of on-board train control system is employed for the validation of the presented solution.The comparative results indicate that GAN-based sample augmentation is capable of achieving a desirable sample balancing level and enhancing the performance of correspondingly derived fault prediction models for the Condition-based Maintenance(CBM)operations.

Semi-supervised LIBS quantitative analysis method based on co-training regression model with selection of effective unlabeled samples

The accuracy of laser-induced breakdown spectroscopy(LIBS) quantitative method is greatly dependent on the amount of certified standard samples used for training. However, in practical applications, only limited standard samples with labeled certified concentrations are available. A novel semi-supervised LIBS quantitative analysis method is proposed, based on co-training regression model with selection of effective unlabeled samples. The main idea of the proposed method is to obtain better regression performance by adding effective unlabeled samples in semisupervised learning. First, effective unlabeled samples are selected according to the testing samples by Euclidean metric. Two original regression models based on least squares support vector machine with different parameters are trained by the labeled samples separately, and then the effective unlabeled samples predicted by the two models are used to enlarge the training dataset based on labeling confidence estimation. The final predictions of the proposed method on the testing samples will be determined by weighted combinations of the predictions of two updated regression models. Chromium concentration analysis experiments of 23 certified standard high-alloy steel samples were carried out, in which 5 samples with labeled concentrations and 11 unlabeled samples were used to train the regression models and the remaining 7 samples were used for testing. With the numbers of effective unlabeled samples increasing, the root mean square error of the proposed method went down from 1.80% to 0.84% and the relative prediction error was reduced from 9.15% to 4.04%.

Research Progress of Aerodynamic Multi-Objective Optimization on High-Speed Train Nose Shape

The aerodynamic optimization design of high-speed trains(HSTs)is crucial for energy conservation,environmental preservation,operational safety,and speeding up.This study aims to review the current state and progress of the aerodynamic multi-objective optimization of HSTs.First,the study explores the impact of train nose shape parameters on aerodynamic performance.The parameterization methods involved in the aerodynamic multiobjective optimization ofHSTs are summarized and classified as shape-based and disturbance-based parameterizationmethods.Meanwhile,the advantages and limitations of each parameterizationmethod,aswell as the applicable scope,are briefly discussed.In addition,the NSGA-II algorithm,particle swarm optimization algorithm,standard genetic algorithm,and other commonly used multi-objective optimization algorithms and the improvements in the field of aerodynamic optimization for HSTs are summarized.Second,this study investigates the aerodynamic multi-objective optimization technology for HSTs using the surrogate model,focusing on the Kriging surrogate models,neural network,and support vector regression.Moreover,the construction methods of surrogate models are summarized,and the influence of different sample infill criteria on the efficiency ofmulti-objective optimization is analyzed.Meanwhile,advanced aerodynamic optimization methods in the field of aircraft have been briefly introduced to guide research on the aerodynamic optimization of HSTs.Finally,based on the summary of the research progress of the aerodynamicmulti-objective optimization ofHSTs,future research directions are proposed,such as intelligent recognition technology of characteristic parameters,collaborative optimization of multiple operating environments,and sample infill criterion of the surrogate model.

A 3D Matching Method for Organic Training Samples Alignment Based on Surface Curvature Distribution

The fundamental step to get a Statistical Shape Model (SSM) is to align all the training samples to the same spatial modality. In this paper, we propose a new 3D alignment method for organic training samples matching, whose modalities are orientable and surface figures could be recognized. It is a feature based alignment method which matches two models depending on the distribution of surface curvature. According to the affine transformation on 2D Gaussian map, the distances between the corresponding parts on surface could be minimized. We applied our proposed method on 5 cases left lung training samples alignment and 4 cases liver training samples alignment. The experiment results were performed on the left lung training samples and the liver training samples. The availability of proposed method was confirmed.

Method to generate training samples for neural network used in target recognition

Training neural network to recognize targets needs a lot of samples.People usually get these samples in a non-systematic way,which can miss or overemphasize some target information.To improve this situation,a new method based on virtual model and invariant moments was proposed to generate training samples.The method was composed of the following steps：use computer and simulation software to build target object＇s virtual model and then simulate the environment,light condition,camera parameter,etc.;rotate the model by spin and nutation of inclination to get the image sequence by virtual camera;preprocess each image and transfer them into binary image;calculate the invariant moments for each image and get a vectors＇ sequence.The vectors＇ sequence which was proved to be complete became the training samples together with the target outputs.The simulated results showed that the proposed method could be used to recognize the real targets and improve the accuracy of target recognition effectively when the sampling interval was short enough and the circumstance simulation was close enough.

The Single Training Sample Extraction of Visual Evoked Potentials Based on Wavelet Transform

Based on the good localization characteristic of the wavelet transform both in time and frequency domain, a de-noising method based on wavelet transform is presented, which can make the extraction of visual evoked potentials in single training sample from the EEG background noise in favor of studying the changes between the single sample response happen. The information is probably related with the different function, appearance and pathologies of the brain. At the same time this method can also be used to remove those signal’s artifacts that do not appear with EP within the same scope of time or frequency. The traditional Fourier filter can hardly attain the similar result. This method is different from other wavelet de-noising methods in which different criteria are employed in choosing wavelet coefficient. It has a biggest virtue of noting the differences among the single training sample and making use of the characteristics of high time frequency resolution to reduce the effect of interference factors to a maximum extent within the time scope that EP appear. The experiment result proves that this method is not restricted by the signal-to-noise ratio of evoked potential and electroencephalograph (EEG) and even can recognize instantaneous event under the condition of lower signal-to-noise ratio, as well as recognize the samples which evoked evident response more easily. Therefore, more evident average evoked response could be achieved by de-nosing the signals obtained through averaging out the samples that can evoke evident responses than de-nosing the average of original signals. In addition, averaging methodology can dramatically reduce the number of record samples needed, thus avoiding the effect of behavior change during the recording process. This methodology pays attention to the differences among single training sample and also accomplishes the extraction of visual evoked potentials from single trainings sample. As a result, system speed and accuracy could be improved to a great extent if this methodology is applied to brain-computer interface system based on evoked responses.

Fault Current Identification of DC Traction Feeder Based on Optimized VMD and Sample Entropy

A current identification method based on optimized variational mode decomposition(VMD)and sample entropy(SampEn)is proposed in order to solve the problem that the main protection of the urban rail transit DC feeder cannot distinguish between train charging current and remote short circuit current.This method uses the principle of energy difference to optimize the optimal mode decomposition number k of VMD;the optimal VMD for DC feeder current is decomposed into the intrinsic modal function(IMF)of different frequency bands.The sample entropy algorithm is used to perform feature extraction of each IMF,and then the eigenvalues of the intrinsic modal function of each frequency band of the current signal can be obtained.The recognition feature vector is input into the support vector machine model based on Bayesian hyperparameter optimization for training.After a large number of experimental data are verified,it is found that the optimal VMD_SampEn algorithm to identify the train charging current and remote short circuit current is more accurate than other algorithms.Thus,the algorithm based on optimized VMD_SampEn has certain engineering application value in the fault current identification of the DC traction feeder.

基于样本增强的列车卫星定位伪距欺骗检测方法

基于卫星导航的列车自主定位是列车控制系统等铁路关键装备的重要技术方向。然而,列车卫星定位面临诸多挑战,除信号可视性问题和多径效应之外,来自系统外部的蓄意欺骗等干扰攻击,会对定位功能及性能产生直接威胁。为此,本文以基于全球导航卫星系统(GNSS)的列车定位面临的伪距欺骗这一典型干扰模式为对象,研究并提出一种基于样本增强的伪距欺骗主动检测方法。该方法运用Wasserstein生成式对抗网络(WGAN)解决受欺骗干扰样本数据不均衡问题,利用扩充的数据集训练检测模型,并引入自注意力(SA)机制优化来自不同接收机输入特征之间的相对位置关系,采用生成式对抗学习思想形成一套完整的列车卫星定位伪距欺骗干扰检测方案。由列车卫星定位欺骗干扰注入测试结果可知,提出的方法能够充分运用生成式对抗网络思想解决受欺骗样本的典型受限问题,融合自注意力机制所得检测性能显著优于载噪比检测和代表性机器学习算法等常规检测方案;对建模样本未覆盖特征具备良好的适应能力,具有更优的检测精度和鲁棒性,在多个伪距欺骗干扰模式数据集上测试所得F1分数均超过0.99。该方法在欺骗干扰检测性能方面的优势能够为众多卫星导航系统铁路应用提供有力支撑,为有效防范卫星定位在信息安全层面的攻击入侵提供了有利条件。

基于动态阈值和差异性检验的自训练算法

针对自训练算法在迭代训练分类器的过程中存在难以有效选取高置信度样本以及误标记样本错误累积的问题,本文提出了基于动态阈值和差异性检验的自训练算法。引入样本的局部离群因子,据此剔除有标签样本中的离群点以及分类标注无标签样本,依据标注分批次处理无标签样本,以使模型更易选取到高置信度的无标签样本;根据新增伪标签样本的数量和对比隶属度的变化,设计一种动态隶属度阈值函数,提升高置信度样本的质量;定义密集距离度量样本间的差异性,分别计算伪标签样本与同类和不同类样本之间的密集距离之和,从而找出不确定度高的伪标签样本,并将此类样本并入下轮训练的无标签样本集中,缓解误标记样本错误累积的问题。实验结果表明,该算法在12个UCI基准数据集上均取得理想效果。

外部注意力增强语义交互的阅读理解模型

针对传统抽取式阅读理解模型未充分考虑问答样本之间潜在相关性的问题,通过RoBERTa对问题与段落进行编码,利用外部注意力Exatt增强语义交互层特征获取能力,提出外部注意力增强语义交互的阅读理解模型,捕获问题与段落中蕴涵的语义特征和不同问答样本之间的潜在相关性。实验结果表明,在CMRC2018和构建的电力安规问答数据集上,在评价指标EM和F1两方面,该方法较基线模型分别最高提高了0.737%和2.556%。

基于高分辨率遥感影像的居民地分类方法研究

随着我国城镇化进程的不断深入,大量乡村居民地已经转化为城镇居民地,为了度量二者实时的转化情况,该文提出一种基于监督分类机制的居民地分类方法,该方法首先采用边缘特征及高斯函数量化影像上的局部特征,然后构建5种城镇及乡村居民地分类规则,其次创建训练样本对各类规则进行学习,最后通过城镇及乡村测试样本验证该文方法的精度。实验表明,该文方法可以对高分辨率遥感影像城镇及乡村居民地进行初级分类,为“城镇化”进程提供一个新的衡量指标。

考虑训练样本分布不均衡的超短期风电功率概率预测

提出一种考虑训练样本分布不均衡的超短期风电概率预测方法。首先构建深度信念混合密度网络,通过深度信念网络独特的预训练和微调机制提取输入变量的隐特征,利用Beta混合概率分布的有界性准确表征风电预测功率的概率分布,实现隐特征与预测功率概率分布参数之间的非线性映射;然后引入训练样本分布平滑策略,其中特征分布平滑技术用于校准输入特征,标签分布平滑技术用于对各样本误差赋予差异化权重,从输入和输出两方面改善训练样本分布不均衡现象对预测结果的不利影响。实际算例结果表明,与常见风电功率概率预测模型相比,所提模型在点预测和概率预测方面均能获得较高的预测精度,尤其能有效提高低密度样本区域的预测精度。

谛听:面向鲁棒分布外样本检测的半监督对抗训练方法

检测训练集分布之外的分布外(out-of-distribution,OOD)样本对于深度神经网络(deep neural network,DNN)分类器在开放环境的部署至关重要.检测OOD样本可以视为一种二分类问题,即把输入样本分类为“分布内(in-distribution,ID)”类或“分布外”类.进一步地,检测器自身还可能遭受到恶意的对抗攻击而被再次绕过.这些带有恶意扰动的OOD样本称为对抗OOD样本.构建鲁棒的OOD检测器以检测对抗OOD样本是一项更具挑战性的任务.为习得可分离且对恶意扰动鲁棒的表示,现有方法往往利用辅助的干净OOD样本邻域内的对抗OOD样本来训练DNN.然而,由于辅助的OOD训练集与原ID训练集的分布差异,训练对抗OOD样本无法足够有效地使分布内决策边界对对抗扰动真正鲁棒.从ID样本的邻域内生成的对抗ID样本拥有与原ID样本近乎一样的语义信息,是一种离分布内区域更近的OOD样本,对提升分布内边界对对抗扰动的鲁棒性很有效.基于此,提出一种半监督的对抗训练方法——谛听,来构建鲁棒的OOD检测器,用以同时检测干净OOD样本和对抗OOD样本.谛听将对抗ID样本视为一种辅助的“近OOD”样本,并将其与其他辅助的干净OOD样本和对抗OOD样本联合训练DNN,以提升OOD检测的鲁棒性.实验结果表明,谛听在检测由强攻击生成的对抗OOD样本上具有显著的优势,同时在原分类主任务及检测干净OOD样本上保持先进的性能.

面向多模态模型训练的高效样本检索技术

深度学习中,多模态模型的训练通常需要大量高质量不同类型的标注数据,如图像、文本、音频等.然而,获取大规模的多模态标注数据是一项具有挑战性和昂贵的任务.为了解决这一问题,主动学习作为一种有效的学习范式被广泛应用,能够通过有针对性地选择最有信息价值的样本进行标注,从而降低标注成本并提高模型性能.现有的主动学习方法往往面临着低效的数据扫描和数据位置调整问题,当索引需要进行大范围的更新时,会带来巨大的维护代价.为解决这些问题,提出了一种面向多模态模型训练的高效样本检索技术So-CBI.该方法通过感知模型训练类间边界点,精确评估样本对模型的价值;设计了半有序的高效样本索引,通过结合数据排序信息和部分有序性,降低了索引维护代价和时间开销.在多组多模态数据集上通过与传统主动学习训练方法实验对比,验证了So-CBI方法在主动学习下的训练样本检索问题上的有效性.

融合多模态生成和情景训练的环境无关手势识别

无线通信和感知技术的发展促进了WiFi感知的产生与发展.依据人体及其行为对WiFi信号传播的影响,通过模式匹配可以实现基于WiFi的手势识别、活动识别、定位等感知任务.但是WiFi信号对环境具有较大依赖性,目标人员或周围环境的变化会导致已经建立的感知模型失效.为了解决这个问题,现有方案通常采用半监督或无监督域适应方法.但在实际应用中,无法预先获得新环境中的数据.因此,需要一种无需新环境数据,即可自动泛化到新环境的方法.为了实现这个目标,本文提出一种基于多模态样本生成和情景训练的环境无关手势识别方法.该方法采用若干源域的数据建立手势识别模型,能够在目标域没有任何数据的情况下,泛化到目标域中.实验结果表明,该方法在目标域无数据的情况下,对新用户和新环境的手势识别正确率均超过80%,高于业界现有水平.

基于生成式预训练语言模型的学者画像构建研究

大数据时代,互联网中以多源异构、非结构化形式存在的学者信息在实体抽取时伴有属性混淆、长实体等问题,严重影响学者画像构建的精准度。与此同时,学者属性实体抽取模型作为学者画像构建过程中的关键模型,在实际应用方面还存在较高的技术门槛,这对学者画像的应用推广造成一定阻碍。为此,在开放资源的基础上,通过引导句建模、自回归生成方式、训练语料微调等构建一种基于生成式预训练语言模型的属性实体抽取框架,并从模型整体效果、实体类别抽取效果、主要影响因素实例分析、样例微调影响分析4个方面对该方法进行验证分析。与对比模型相比,所提出的方法在12类学者属性实体上均达到最优效果,其综合F1值为99.34%,不仅能够较好地识别区分相互混淆的属性实体,对“研究方向”这一典型长属性实体的抽取准确率还提升了6.11%,为学者画像的工程化应用提供了更快捷、有效的方法支撑。

基于Tri-training的入侵检测算法

半监督的双协同训练要求划分出的2个数据向量相互独立,不符合真实的网络入侵检测数据特征。为此,提出一种基于三协同训练(Tri-training)的入侵检测算法。使用大量未标记数据,通过3个分类器对检测结果进行循环迭代训练,避免交叉验证。仿真实验表明,在少量样本情况下,该算法的检测准确度比SVM Co-training算法提高了2.1%,并且随着循环次数的增加,其性能优势更加明显。

基于双向对比训练的中文故事结尾生成模型

中文故事结尾生成(SEG)是自然语言处理中的下游任务之一。基于全错误结尾的CLSEG(Contrastive Learning of Story Ending Generation)在故事的一致性方面表现较好。然而,由于错误结尾中也包含与原结尾文本相同的内容,仅使用错误结尾的对比训练会导致生成文本中原结尾正确的主要部分被剥离。因此,在CLSEG基础上增加正向结尾增强训练,以保留对比训练中损失的正确部分;同时,通过正向结尾的引入,使生成的结尾具有更强的多样性和关联性。基于双向对比训练的中文故事结尾生成模型包含两个主要部分:1)多结尾采样,通过不同的模型方法获取正向增强的结尾和反向对比的错误结尾;2)对比训练,在训练过程中修改损失函数,使生成的结尾接近正向结尾,远离错误结尾。在公开的故事数据集OutGen上的实验结果表明,相较于GPT2. ft和深层逐层隐变量融合(Della)等模型,所提模型的BERTScore、METEOR等指标均取得了较优的结果,生成的结尾具有更强的多样性和关联性。