Robust Damage Detection and Localization Under Complex Environmental Conditions Using Singular Value Decomposition-based Feature Extraction and One-dimensional Convolutional Neural Network

Ultrasonic guided wave is an attractive monitoring technique for large-scale structures but is vulnerable to changes in environmental and operational conditions(EOC),which are inevitable in the normal inspection of civil and mechanical structures.This paper thus presents a robust guided wave-based method for damage detection and localization under complex environmental conditions by singular value decomposition-based feature extraction and one-dimensional convolutional neural network(1D-CNN).After singular value decomposition-based feature extraction processing,a temporal robust damage index(TRDI)is extracted,and the effect of EOCs is well removed.Hence,even for the signals with a very large temperature-varying range and low signal-to-noise ratios(SNRs),the final damage detection and localization accuracy retain perfect 100%.Verifications are conducted on two different experimental datasets.The first dataset consists of guided wave signals collected from a thin aluminum plate with artificial noises,and the second is a publicly available experimental dataset of guided wave signals acquired on a composite plate with a temperature ranging from 20℃to 60℃.It is demonstrated that the proposed method can detect and localize the damage accurately and rapidly,showing great potential for application in complex and unknown EOC.

Synthesization of high-capacity auto-associative memories using complex-valued neural networks

In this paper, a novel design procedure is proposed for synthesizing high-capacity auto-associative memories based on complex-valued neural networks with real-imaginary-type activation functions and constant delays. Stability criteria dependent on external inputs of neural networks are derived. The designed networks can retrieve the stored patterns by external inputs rather than initial conditions. The derivation can memorize the desired patterns with lower-dimensional neural networks than real-valued neural networks, and eliminate spurious equilibria of complex-valued neural networks. One numerical example is provided to show the effectiveness and superiority of the presented results.

Complex Valued Recurrent Neural Network: From Architecture to Training

Recurrent Neural Networks were invented a long time ago, and dozens of different architectures have been published. In this paper we generalize recurrent architectures to a state space model, and we also generalize the numbers the network can process to the complex domain. We show how to train the recurrent network in the complex valued case, and we present the theorems and procedures to make the training stable. We also show that the complex valued recurrent neural network is a generalization of the real valued counterpart and that it has specific advantages over the latter. We conclude the paper with a discussion of possible applications and scenarios for using these networks.

Multistability of delayed complex-valued recurrent neural networks with discontinuous real-imaginarytype activation functions

In this paper, the multistability issue is discussed for delayed complex-valued recurrent neural networks with discontinuous real-imaginary-type activation functions. Based on a fixed theorem and stability definition, sufficient criteria are established for the existence and stability of multiple equilibria of complex-valued recurrent neural networks. The number of stable equilibria is larger than that of real-valued recurrent neural networks, which can be used to achieve high-capacity associative memories. One numerical example is provided to show the effectiveness and superiority of the presented results.

Learning Dynamics of the Complex-Valued Neural Network in the Neighborhood of Singular Points

In this paper, the singularity and its effect on learning dynamics in the complex-valued neural network are elucidated. It has learned that the linear combination structure in the updating rule of the complex-valued neural network increases the speed of moving away from the singular points, and the complex-valued neural network cannot be easily influenced by the singular points, whereas the learning of the usual real-valued neural network can be attracted in the neighborhood of singular points, which causes a standstill in learning. Simulation results on the learning dynamics of the three-layered real-valued and complex-valued neural networks in the neighborhood of singularities support the analytical results.

Intelligent recognition and information extraction of radar complex jamming based on time-frequency features

In modern war,radar countermeasure is becoming increasingly fierce,and the enemy jamming time and pattern are changing more randomly.It is challenging for the radar to efficiently identify jamming and obtain precise parameter information,particularly in low signal-to-noise ratio(SNR)situations.In this paper,an approach to intelligent recognition and complex jamming parameter estimate based on joint time-frequency distribution features is proposed to address this challenging issue.Firstly,a joint algorithm based on YOLOv5 convolutional neural networks(CNNs)is proposed,which is used to achieve the jamming signal classification and preliminary parameter estimation.Furthermore,an accurate jamming key parameters estimation algorithm is constructed by comprehensively utilizing chi-square statistical test,feature region search,position regression,spectrum interpolation,etc.,which realizes the accurate estimation of jamming carrier frequency,relative delay,Doppler frequency shift,and other parameters.Finally,the approach has improved performance for complex jamming recognition and parameter estimation under low SNR,and the recognition rate can reach 98%under−15 dB SNR,according to simulation and real data verification results.

Complex Network Classification with Convolutional Neural Network

Classifying large-scale networks into several categories and distinguishing them according to their fine structures is of great importance to several real-life applications.However,most studies on complex networks focus on the properties of a single network and seldom on classification,clustering,and comparison between different networks,in which the network is treated as a whole.Conventional methods can hardly be applied on networks directly due to the non-Euclidean properties of data.In this paper,we propose a novel framework of Complex Network Classifier(CNC)by integrating network embedding and convolutional neural network to tackle the problem of network classification.By training the classifier on synthetic complex network data,we show CNC can not only classify networks with high accuracy and robustness but can also extract the features of the networks automatically.We also compare our CNC with baseline methods on benchmark datasets,which shows that our method performs well on large-scale networks.

复杂场景下自适应注意力机制融合实时语义分割

实现高准确度和低计算负担是卷积神经网络(CNN)实时语义分割面临的严峻挑战。针对复杂城市街道场景目标种类众多、光照变化大等特点,该文设计了一种高效的实时语义分割自适应注意力机制融合网络(AAFNet)分别提取图像空间细节和语义信息,再经过特征融合网络(FFN)获得准确语义图像。AAFNet采用扩展的深度可分离卷积(DDW)可增大语义特征提取感受野,提出自适应平均池化(Avp)和自适应最大池化(Amp)构成自适应注意力机制融合模块(AAFM),可细化目标边缘分割效果并降低小目标的漏分率。最后在复杂城市街道场景Cityscapes和CamVid数据集上分别进行了语义分割实验,所设计的AAFNet以32帧/s(Cityscapes)和52帧/s(CamVid)的推理速度获得73.0%和69.8%的平均分割精度(mIoU),且与扩展的空间注意力网络(DSANet)、多尺度上下文融合网络(MSCFNet)以及轻量级双边非对称残差网络(LBARNet)相比,AAFNet平均分割精度最高。

基于改进两步法采样策略和卷积神经网络的崩塌易发性评价

机器学习在崩塌滑坡泥石流地质灾害易发性分析评价领域已得到广泛的研究性应用,非灾害样本的选取是易发性建模过程中的关键问题,传统随机抽样和手工标注方法可能存在随机性和主观性。将土质崩塌易发性评价视为正例无标记(positive and unlabeled,简称PU)学习,提出了一种结合信息量(information value,简称IV)和间谍技术(Spy)的两步卷积神经网络(convolutional neural networks,简称CNN)框架(ISpy-CNN)。以广州市黄埔区崩塌编录和15类基础环境因子,通过信息量模型筛选出部分低信息量样本;采用间谍技术训练CNN模型,从低信息量样本中识别出具有高置信度的可靠负例划分为非崩塌样本;分别基于该学习框架、传统间谍技术和随机抽样,使用支持向量机(support vector machine,简称SVM)和随机森林(random forest,简称RF)对比验证。结果表明,ISpy-CNN框架在验证集上的准确率、F1值、敏感度和特异度较随机采样分别提升了6.82%,6.82%,6.82%,8.23%,较传统Spy技术分别提升了2.86%,2.89%,2.86%,2.31%;PU学习中第2步采用CNN模型的预测精度高于RF和SVM模型;与传统Spy技术相比,增加相同数量训练样本,ISpy-CNN框架筛选的样本集表现出较高的稳定性、预测精度和增长率。本研究提出的ISpy-CNN框架能更好地辅助选取高质量非灾害样本,且崩塌易发性分区结果更符合实际的崩塌空间分布。

用于低剂量CT图像去噪的多级双树复小波网络

基于卷积神经网络(CNN)的图像去噪方法能有效去除低剂量计算机断层扫描(CT)图像伴随的伪影和噪声,从而确保CT设备输出高质量图像同时降低辐射,这对患者健康和医学诊断具有重要意义。为了进一步提高低剂量CT图像的质量,提出一种小波域去噪网络MDTNet。首先,基于双树复小波变换(DTCWT)构造多级编解码去噪网络,在多个尺度上提取特征以保留更多高频细节;然后,利用扩展的像素重排技术替代卷积上下采样,实现多级输入和特征融合,从而降低计算复杂度;最后,通过大量训练找到最佳的去噪模型,即二级MDTNet配合LeGall滤波器和Qshift_b滤波器,并选择较大尺寸的CT图像作为训练数据。使用AAPM数据集评估MDTNet的性能,实验结果表明,MDTNet能有效去除条纹状伪影和噪声,在定量和定性评估中性能均优于同类型去噪方法。与FWDNet相比,对于1 mm的切片,MDTNet的平均峰值信噪比(PSNR)和结构相似性指数(SSIM)分别提高了0.0887 dB和0.0024;对于3 mm的切片,分别提升了0.1443 dB和0.003。对于单张512×512像素的低剂量CT图像去噪,MDTNet在GPU上仅需0.193 s。MDTNet在保持高效率的同时保留了更多的高频细节,能够为低剂量CT图像去噪提供一种新的框架。

DeepSVDNet:A Deep Learning-Based Approach for Detecting and Classifying Vision-Threatening Diabetic Retinopathy in Retinal Fundus Images

Artificial Intelligence(AI)is being increasingly used for diagnosing Vision-Threatening Diabetic Retinopathy(VTDR),which is a leading cause of visual impairment and blindness worldwide.However,previous automated VTDR detection methods have mainly relied on manual feature extraction and classification,leading to errors.This paper proposes a novel VTDR detection and classification model that combines different models through majority voting.Our proposed methodology involves preprocessing,data augmentation,feature extraction,and classification stages.We use a hybrid convolutional neural network-singular value decomposition(CNN-SVD)model for feature extraction and selection and an improved SVM-RBF with a Decision Tree(DT)and K-Nearest Neighbor(KNN)for classification.We tested our model on the IDRiD dataset and achieved an accuracy of 98.06%,a sensitivity of 83.67%,and a specificity of 100%for DR detection and evaluation tests,respectively.Our proposed approach outperforms baseline techniques and provides a more robust and accurate method for VTDR detection.

基于深度复数神经网络的雷达目标DOA估计算法

传统模型驱动的波达方向(direction of arrival, DOA)估计算法性能受限于有限的信号特征、快拍数、信噪比、信杂比等因素,在低信噪比、快拍数少的极端情况下,性能较差。为克服上述问题,提高在极端条件下的估计精度,文章提出基于深度复数神经网络(complex-valued neural networks, CVNN)的单快拍DOA估计算法,构建深度复数神经网络模型,学习原始带噪信号与理想无噪复信号之间的映射关系,进而实现噪声抑制和期望信号特征增强的目的,提高DOA估计精度。仿真实验结果表明,经CVNN增强后,数据的等效信噪比约提高了1 dB,等效快拍数提高了3,该文所提算法相较于已有的多种物理驱动算法而言,具有更高的估计精度和泛化性。

面向时变复数的模糊归零神经网络算法

【目的】为了求解时变复数西尔维斯特方程(Sylvester equation),提出两种全新的复值模糊归零神经网络(fuzzy logic system for zeroing neural network, FLSVZNN)模型。【方法】首先,在两种复数归零神经网络(complex-valued zeroing neural network, CVZNN)模型的基础上,引入模糊逻辑系统(fuzzy logic system, FLS)来控制信号的处理,从而提出两种FLSVZNN模型;然后,利用李亚普诺夫定理(Lyapunov's theorem)来分析模型的稳定性和收敛速度;最后,通过仿真试验来进一步验证FLSVZNN的优越性能。【结果】在求解时变复数西尔维斯特方程时,相比传统的神经网络模型,使用改进的符号双幂(sign-bi-power, SBP)函数来激活的FLSVZNN模型具有更好的收敛性和稳定性,可使误差函数在0.3 s左右收敛至0。【结论】本研究提出的两种FLSVZNN模型能快速求解时变复数西尔维斯特方程,这可为神经网络模型的建立及工程应用提供参考。

复杂背景下基于改进Mask R-CNN的路面裂缝检测算法

裂缝检测对路面养护具有重要意义,深度学习在该领域取得一定成效。然而,在实际应用中,图像中的噪声纹理背景、复杂的裂缝拓扑结构和图像采集设备给裂缝检测带来了一定的挑战。为了提升在复杂场景下的路面裂缝检测精度,提出了一种改进掩码区域卷积神经网络(Mask R-CNN)模型的实例分割算法。使用ConvNeXt-T替代Mask R-CNN的ResNet50框架作为特征生成网络,在自下而上捕获长期依赖的同时保持裂缝特征多样性;设计高维特征提取模块(HFEM)获取高级语义信息,消除背景噪声;引入感受野模块(RFB),扩大感受野,增强多尺度特征信息交互能力。在多结构裂缝图像(MSCI)数据集上进行对比实验,结果表明,提出的改进方法能显著提升Mask R-CNN模型的分割精度,优于经典的Cascade Mask RCNN,最佳模型F1得分84.15%,相较原算法提高了6.29%。在DeepCrack数据集上进行泛化性实验,表现优异。

基于神经网络的HEVC帧间预测方法及其硬件研究

相比于H.264,高效视频编码标准(HEVC)提出了许多新技术,提高了编码性能,但是也显著提高了编码复杂度。本文从硬件实现的角度出发,对已有的帧间CU划分预测神经网络的结构进行了多方面的优化,使其参数减少了70%,加法和乘法运算分别减少了60%、58.2%。并对优化后的卷积神经网络参数采用10位定点数方案进行定点化处理,进一步有效减少硬件资源的开支。对比于HEVC参考软件(HM16.5),优化后网络引起的BD-BR和BD-PSNR平均损失为1.718%和-0.056dB,平均节省35%~52%的编码复杂度,并且定点化处理后引起的性能损失可忽略不计。

基于二值神经网络的辐射源信号识别方法

针对用于辐射源信号识别的神经网络存在参数冗余、运算量庞大等问题,提出一种基于二值神经网络的辐射源信号识别方法。该方法指出利用卷积层效用值衡量神经网络卷积层的重要性,根据卷积层效用值的大小,将重要的卷积层保留为实值,其余卷积层进行二值化处理。实验结果表明,在信噪比大于-9 dB时,采用该方法得到的二值神经网络的信号识别准确率相比于实值卷积神经网络降低了0.5%,而网络参数内存大小降低了83.4%,网络运算次数降低了83.8%,网络运算复杂度降低了85.8%,易于部署在各种硬件平台上。

基于复数生成对抗网络的5G OFDM信道估计方法

准确的信道估计能够显著地降低误码率(bit error rate,BER),提高无线通信效率和质量,是5G OFDM通信系统接收机设计的关键环节之一。基于最小二乘(least square,LS)法和基于最小均方差(minimum mean square error,MMSE)的信道估计方法利用系统稀疏性计算信道响应矩阵,但LS算法计算精度较低,而MMSE算法计算量过大。为提升估计精度,业界设计了基于深度学习的信道估计方法。然而,现有的深度学习方法将复数矩阵拆分成实部和虚部,没有充分提取信道中的复数特征,造成估计的信道响应矩阵出现失真。为此,设计了一种基于复数的生成对抗网络模型,充分提取信号的复数特征,从而更准确地估计5G新空口(new radio,NR)标准的物理下行链路共享信道(physical downlink shared channel,PDSCH)的信道响应矩阵。为了验证所提方法的有效性,将所提方法分别与LS算法、实际信道估计、超分辨率神经网络、残差神经网络信道估计算法进行了对比分析。结果表明,当估计的信道响应矩阵与真实矩阵之间的均方差达到0.01时,采用所提方法实现的无线通信系统的信噪比高于现有方法5 dB左右。

基于注意力机制和CNN-GRU模型的脱硫系统pH值预测

针对火电厂石灰石-石膏烟气湿法脱硫系统中所面临的浆液pH值测量不准确、时间长的问题,提出了一种基于注意力机制和CNN-GRU模型来预测吸收塔内浆液pH值。首先,对火电厂监测系统(SIS)数据库中的数据进行预处理,然后使用相关性分析来确定它们之间的关联性。接下来可使用注意力机制(ATT)来自适应分配与pH值相关联的输入数据的权重,并根据权重大小来区分强弱特征变量,以此来解决预测精度低和不准确的问题。其后利用卷积神经网络(CNN)来二次提取和降维这些特征数据,并对送入门控循环神经单元网络(GRU)中的数据进行优化,可大大加快神经网络训练进程,并且能够更准确地处理复杂的动态脱硫变化。对某电厂2×350 MW机组运行数据进行测试,并通过与其他主流算法对比得出所建pH值预测模型具备更高的精确度和稳定性。最后结合模型预测控制(MPC),验证了该模型的实用性。

悬垂式玻璃绝缘子覆冰冰凌识别与评估方法研究

由于悬垂式玻璃绝缘子覆冰冰凌环境较为复杂、影响因素过多、特征难以提取,导致悬垂式玻璃绝缘子覆冰冰凌识别精准度较低。为此,提出了基于图像分析的悬垂式玻璃绝缘子覆冰冰凌识别与评估方法。采用领域平均法,对采集到的绝缘子覆冰冰凌图像进行平滑去噪处理。按照绝缘子形态参数,建立卷积神经网络层次结构。通过镜面成像原理,捕捉冰凌成像特征。按照现场环境参数,设定成像尺寸。通过换算,识别悬垂式玻璃绝缘子覆冰冰凌。根据历史数据模拟冰凌的风险指标,设置权重参数。通过权重比对,完成评估。试验结果表明,所提方法最低识别误差率为7%、评估均方根误差为0.15。应用该方法的图像更为清晰、细节保留效果更好。该方法能够有效实现悬垂式玻璃绝缘子覆冰冰凌识别,且识别与评估的精准度均较高。

基于卷积神经网络的红外与可见光图像融合方法

针对当前红外与可见光图像融合方法存在融合效果差、效率低等问题,为了获得更优的红外与可见光图像融合结果,提出了基于卷积神经网络的红外与可见光图像融合方法。首先收集待融合红外与可见光图像,采用Retinex算法对图像进行增强操作,提升图像亮度和细节信息,然后采用卷积神经网络提取图像融合特征,并设计红外与可见光图像融合规则,根据规则得到图像融合结果,最后应用多个数据集进行红外与可见光图像融合性能测试实验,结果表明,卷积神经网络的图像融合后整体视觉效果好,重要细节信息丰富,熵和平均梯度的值超过了6,融合时间低于1 s,整体性能优于其它红外与可见光图像对比融合方法。