MACDCGAN的发电机轴承故障诊断方法

在实际工况中,发电机中传感器采集到的故障样本数据有限,使用基于深度学习的方法进行故障诊断存在过拟合问题导致模型泛化能力较差以及诊断精度不高。为了解决这个问题,采用样本扩充的思路,提出了一种改进的辅助分类器条件深度卷积生成对抗网络(MACDCGAN)的故障诊断方法。通过对采集的一维时序信号进行小波变换增强特征,构建简化结构参数的条件深度卷积生成对抗网络模型生成样本,并在模型中采用Wasserstein距离优化损失函数解决训练过程中存在模式崩塌和梯度消失的缺点;通过添加一个独立的分类器来改进分类模型的兼容性,并在分类器中引入学习率衰减算法增加模型稳定性。试验结果表明,该方法可以有效地提高故障诊断的精度,并且验证了所提模型具有良好的泛化性能。

基于Transformer和GAN的对抗样本生成算法

对抗攻击与防御是计算机安全领域的一个热门研究方向。针对现有基于梯度的对抗样本生成方法可视质量差、基于优化的方法生成效率低的问题,提出基于Transformer和生成对抗网络(GAN)的对抗样本生成算法Trans-GAN。首先利用Transformer强大的视觉表征能力,将其作为重构网络,用于接收干净图像并生成攻击噪声;其次将Transformer重构网络作为生成器,与基于深度卷积网络的鉴别器相结合组成GAN网络架构,提高生成图像的真实性并保证训练的稳定性,同时提出改进的注意力机制Targeted Self-Attention,在训练网络时引入目标标签作为先验知识,指导网络模型学习生成具有特定攻击目标的对抗扰动;最后利用跳转连接将对抗噪声施加在干净样本上,形成对抗样本,攻击目标分类网络。实验结果表明:Trans-GAN算法针对MNIST数据集中2种模型的攻击成功率都达到99.9%以上,针对CIFAR10数据集中2种模型的攻击成功率分别达到96.36%和98.47%,优于目前先进的基于生成式的对抗样本生成方法;相比快速梯度符号法和投影梯度下降法,Trans-GAN算法生成的对抗噪声扰动量更小,形成的对抗样本更加自然,满足人类视觉不易分辨的要求。

融合门控变换机制和GAN的低光照图像增强方法

针对低光照图像增强过程中存在的配对图像数据依赖、细节损失严重和噪声放大问题,提出结合门控通道变换机制和生成对抗网络(GAN)的低光照图像增强方法AGR-GAN,该方法可以在没有低/正常光图像对的情况下进行训练。首先,设计特征提取网络,该网络由多个基于门控通道变换单元的多尺度卷积残差模块构成,以提取输入图像的全局上下文特征和多尺度局部特征信息;然后,在特征融合网络中,采用卷积残差结构将提取的深浅层特征进行充分融合,再引入横向跳跃连接结构,最大程度保留细节特征信息,获得最终的增强图像;最后,引入联合损失函数指导网络训练过程,抑制图像噪声,使增强图像色彩更自然匀称。实验结果表明,该方法在主观视觉分析和客观指标评价方面相较其他算法均具有显著优势,其能有效提高低光照图像的亮度和对比度,减弱图像噪声,增强后的图像更清晰且色彩更真实,峰值信噪比、结构相似度和无参考图像质量评价指标平均可达16.48 dB、0.93和3.37。

GAN-GLS:Generative Lyric Steganography Based on Generative Adversarial Networks

Steganography based on generative adversarial networks(GANs)has become a hot topic among researchers.Due to GANs being unsuitable for text fields with discrete characteristics,researchers have proposed GANbased steganography methods that are less dependent on text.In this paper,we propose a new method of generative lyrics steganography based on GANs,called GAN-GLS.The proposed method uses the GAN model and the largescale lyrics corpus to construct and train a lyrics generator.In this method,the GAN uses a previously generated line of a lyric as the input sentence in order to generate the next line of the lyric.Using a strategy based on the penalty mechanism in training,the GAN model generates non-repetitive and diverse lyrics.The secret information is then processed according to the data characteristics of the generated lyrics in order to hide information.Unlike other text generation-based linguistic steganographic methods,our method changes the way that multiple generated candidate items are selected as the candidate groups in order to encode the conditional probability distribution.The experimental results demonstrate that our method can generate highquality lyrics as stego-texts.Moreover,compared with other similar methods,the proposed method achieves good performance in terms of imperceptibility,embedding rate,effectiveness,extraction success rate and security.

Automated Video Generation of Moving Digits from Text Using Deep Deconvolutional Generative Adversarial Network

Generating realistic and synthetic video from text is a highly challenging task due to the multitude of issues involved,including digit deformation,noise interference between frames,blurred output,and the need for temporal coherence across frames.In this paper,we propose a novel approach for generating coherent videos of moving digits from textual input using a Deep Deconvolutional Generative Adversarial Network(DD-GAN).The DDGAN comprises a Deep Deconvolutional Neural Network(DDNN)as a Generator(G)and a modified Deep Convolutional Neural Network(DCNN)as a Discriminator(D)to ensure temporal coherence between adjacent frames.The proposed research involves several steps.First,the input text is fed into a Long Short Term Memory(LSTM)based text encoder and then smoothed using Conditioning Augmentation(CA)techniques to enhance the effectiveness of the Generator(G).Next,using a DDNN to generate video frames by incorporating enhanced text and random noise and modifying a DCNN to act as a Discriminator(D),effectively distinguishing between generated and real videos.This research evaluates the quality of the generated videos using standard metrics like Inception Score(IS),Fréchet Inception Distance(FID),Fréchet Inception Distance for video(FID2vid),and Generative Adversarial Metric(GAM),along with a human study based on realism,coherence,and relevance.By conducting experiments on Single-Digit Bouncing MNIST GIFs(SBMG),Two-Digit Bouncing MNIST GIFs(TBMG),and a custom dataset of essential mathematics videos with related text,this research demonstrates significant improvements in both metrics and human study results,confirming the effectiveness of DD-GAN.This research also took the exciting challenge of generating preschool math videos from text,handling complex structures,digits,and symbols,and achieving successful results.The proposed research demonstrates promising results for generating coherent videos from textual input.

Feature-Based Augmentation in Sarcasm Detection Using Reverse Generative Adversarial Network

Sarcasm detection in text data is an increasingly vital area of research due to the prevalence of sarcastic content in online communication.This study addresses challenges associated with small datasets and class imbalances in sarcasm detection by employing comprehensive data pre-processing and Generative Adversial Network(GAN)based augmentation on diverse datasets,including iSarcasm,SemEval-18,and Ghosh.This research offers a novel pipeline for augmenting sarcasm data with Reverse Generative Adversarial Network(RGAN).The proposed RGAN method works by inverting labels between original and synthetic data during the training process.This inversion of labels provides feedback to the generator for generating high-quality data closely resembling the original distribution.Notably,the proposed RGAN model exhibits performance on par with standard GAN,showcasing its robust efficacy in augmenting text data.The exploration of various datasets highlights the nuanced impact of augmentation on model performance,with cautionary insights into maintaining a delicate balance between synthetic and original data.The methodological framework encompasses comprehensive data pre-processing and GAN-based augmentation,with a meticulous comparison against Natural Language Processing Augmentation(NLPAug)as an alternative augmentation technique.Overall,the F1-score of our proposed technique outperforms that of the synonym replacement augmentation technique using NLPAug.The increase in F1-score in experiments using RGAN ranged from 0.066%to 1.054%,and the use of standard GAN resulted in a 2.88%increase in F1-score.The proposed RGAN model outperformed the NLPAug method and demonstrated comparable performance to standard GAN,emphasizing its efficacy in text data augmentation.

A Sketch-Based Generation Model for Diverse Ceramic Tile Images Using Generative Adversarial Network

Ceramic tiles are one of the most indispensable materials for interior decoration.The ceramic patterns can’t match the design requirements in terms of diversity and interactivity due to their natural textures.In this paper,we propose a sketch-based generation method for generating diverse ceramic tile images based on a hand-drawn sketches using Generative Adversarial Network(GAN).The generated tile images can be tailored to meet the specific needs of the user for the tile textures.The proposed method consists of four steps.Firstly,a dataset of ceramic tile images with diverse distributions is created and then pre-trained based on GAN.Secondly,for each ceramic tile image in the dataset,the corresponding sketch image is generated and then the mapping relationship between the images is trained based on a sketch extraction network using ResNet Block and jump connection to improve the quality of the generated sketches.Thirdly,the sketch style is redefined according to the characteristics of the ceramic tile images and then double cross-domain adversarial loss functions are employed to guide the ceramic tile generation network for fitting in the direction of the sketch style and to improve the training speed.Finally,we apply hidden space perturbation and interpolation for further enriching the output textures style and satisfying the concept of“one style with multiple faces”.We conduct the training process of the proposed generation network on 2583 ceramic tile images dataset.To measure the generative diversity and quality,we use Frechet Inception Distance(FID)and Blind/Referenceless Image Spatial Quality Evaluator(BRISQUE)metrics.The experimental results prove that the proposed model greatly enhances the generation results of the ceramic tile images,with FID of 32.47 and BRISQUE of 28.44.

面向舰船目标检测的SAR图像数据PCGAN生成方法

针对现有合成孔径雷达(SAR)图像数据生成方法大多无法同时生成舰船图像及其检测标签的问题,面向SAR舰船图像生成及目标检测任务,构建基于位置信息的条件生成对抗网络(PCGAN).首先,提出将舰船位置信息作为约束条件用于限制生成图像中舰船的位置,并将其作为舰船图像的检测标签;随后,引入Wasserstein距离稳定PCGAN的训练过程;最后,利用生成的SAR舰船图像及对应检测标签完成YOLOv3网络的端到端训练,实现舰船数据增强与目标检测的协同学习,进而获得更耦合目标检测实际应用的多样性数据.在HRSID(high resolution SAR image dataset)数据集上的实验结果表明,PCGAN方法能生成清晰、鲁棒的SAR舰船数据,舰船检测准确度最高提升1.01%,验证了所提出方法的有效性.

融合情感分析和GAN-TrellisNet的股价预测方法

将时序深度神经网络应用于股票价格预测,已成为量化金融领域的重要研究方向。时序神经网络具有很好的序列数据捕捉能力和学习记忆能力,在股票预测上有一定适用性。但是现有的模型大多存在预测准确度不高、模型结构复杂导致训练时间较长等问题.为了解决以上问题,提出了一种基于情感分析和GAN-TrellisNet的股价预测方法。提出了一个基于LSTM-CNN的情感分析模型,用于分析爬虫获取的主流金融论坛股票评论,并获得股票情感指数。为了提高预测准确度,将情感指数和百度搜索指数加入股票交易数据中作为训练集,提出了一个基于TrellisNet和CNN的改进型GAN股价预测模型,利用TrellisNet生成器的卷积特性来捕捉数据的局部特征,选取特征提取能力较强的CNN作为判别器来区别预测结果和真实股价。通过选取10只代表性股票和三种大盘指数的不同时段数据进行算法验证,结果表明,与ConvLSTM和GAN-LSTM预测模型相比,GAN-TrellisNet模型能有效缩短训练时间,提高预测准确率。

基于GAN改进的红外光与可见光图像融合算法研究

针对夜晚户外场景下,传统的单一鉴别器生成对抗网络(GAN)容易忽略红外光的亮度信息和边缘信息的问题,提出一种基于注意力机制与双鉴别器的红外光与可见光图像融合算法。首先,为了有针对性地获得红外光图像的目标信息和可见光图像的背景纹理信息,在生成器网络中引入通道注意力机制;其次,使用双鉴别器的生成对抗网络,并设计一种新的鉴别器输入,在提高训练稳定性的同时更好地保留源图像信息;最后,损失函数设置为对抗损失、结构相似性损失和梯度损失,以约束鉴别器使其生成细节信息丰富的融合图像。在TNO数据集上的实验结果表明,所提算法得到的融合图像梯度变化更明显、边缘更加清晰,更符合人眼视觉效果。

基于SA/WGAN的新能源场景生成方法

随着新能源渗透率逐年提高,其出力的随机性与波动特性难以准确预测,给电力系统的运行、规划和调度提出了严峻的挑战,因此新能源的不确定性建模受到越来越多的关注。为了更有效地获得新能源出力场景中的时序特征,提出了一种基于数据驱动的新能源场景生成方法,通过采用SA/WGAN模型,把自注意力机制和带有梯度惩罚的生成对抗网络判别器结合,构建基于两种模型结合的深度学习模型,有效突显新能源出力场景中时序特性,增强场景生成中非线性拟合能力。算例结果表明,所提模型的新能源生成场景相较于原始WGAN和WGAN-LSTM的场景生成结果,可以有效提高精准度,同时具备了WGAN-GP训练结果稳定和SA计算速度快的优势,更高效地生成与真实新能源场景分布接近的场景。

结合多注意力及IE-GAN的面部属性编辑方法

面部属性编辑主要有两个目的:①将图像从源域转换到目标域,对指定的目标属性(如性别、年龄、发色等)进行修改;②只改变与目标属性相关的人脸区域,保留目标属性之外其他属性的细节。而现有的人脸属性编辑方法不可避免地会对与目标属性无关区域进行更改。因此,提出一种基于IE-GAN和多注意力机制的面部属性编辑方法MAIE-GAN,引入补充注意连接(Complementary attention connection,CAC)的概念连接编解码器,解决了由直接跳跃连接引起的信息冗余问题,将自注意力机制作为生成器中卷积层的补充,使其能更好定位目标属性,限定属性变换区域,此外,还利用互补注意特征的概念,实现目标属性无关区域的更好保留。通过与现有方法的对比分析表明,该方法在属性定位和图像质量方面优于现有的方法。

MTTSNet:Military time-sensitive targets stealth network via real-time mask generation

The automatic stealth task of military time-sensitive targets plays a crucial role in maintaining national military security and mastering battlefield dynamics in military applications.We propose a novel Military Time-sensitive Targets Stealth Network via Real-time Mask Generation(MTTSNet).According to our knowledge,this is the first technology to automatically remove military targets in real-time from videos.The critical steps of MTTSNet are as follows:First,we designed a real-time mask generation network based on the encoder-decoder framework,combined with the domain expansion structure,to effectively extract mask images.Specifically,the ASPP structure in the encoder could achieve advanced semantic feature fusion.The decoder stacked high-dimensional information with low-dimensional information to obtain an effective mask layer.Subsequently,the domain expansion module guided the adaptive expansion of mask images.Second,a context adversarial generation network based on gated convolution was constructed to achieve background restoration of mask positions in the original image.In addition,our method worked in an end-to-end manner.A particular semantic segmentation dataset for military time-sensitive targets has been constructed,called the Military Time-sensitive Target Masking Dataset(MTMD).The MTMD dataset experiment successfully demonstrated that this method could create a mask that completely occludes the target and that the target could be hidden in real time using this mask.We demonstrated the concealment performance of our proposed method by comparing it to a number of well-known and highly optimized baselines.

Exploration of the Relation between Input Noise and Generated Image in Generative Adversarial Networks

In this paper,we propose a hybrid model aiming to map the input noise vector to the label of the generated image by the generative adversarial network(GAN).This model mainly consists of a pre-trained deep convolution generative adversarial network(DCGAN)and a classifier.By using the model,we visualize the distribution of two-dimensional input noise,leading to a specific type of the generated image after each training epoch of GAN.The visualization reveals the distribution feature of the input noise vector and the performance of the generator.With this feature,we try to build a guided generator(GG)with the ability to produce a fake image we need.Two methods are proposed to build GG.One is the most significant noise(MSN)method,and the other utilizes labeled noise.The MSN method can generate images precisely but with less variations.In contrast,the labeled noise method has more variations but is slightly less stable.Finally,we propose a criterion to measure the performance of the generator,which can be used as a loss function to effectively train the network.

应用XCT断层扫描技术和GAN深度学习模型的多孔介质微观结构定量研究

物质在土壤中的迁移转化行为是研究地下水动力学的核心问题。定量化表征多孔介质拓扑结构并计算分析相关表面形态学参数,为研究复杂介质内物质运移微观机理提供了重要的数据基础和参数。应用微米级X射线计算机断层扫描(XCT)和图像处理技术,结合前沿的机器学习算法重建和定量分析多孔介质微观结构,可快速、批量创建高分辨率的复杂多孔介质研究样本。首先采用XCT技术,提取石英砂和散装土壤两类典型多孔介质的微观孔隙结构;而后基于生成对抗神经网络(GAN)模型重构复杂多孔介质的微观空间结构,与XCT扫描图像进行交叉验证;最后,计算获取Minkowski形态学参数,并基于多孔介质微观结构开展计算流体力学(CFD)数值模拟,计算石英砂和土壤多孔介质内的流动特征和渗透率。结果表明:1)GAN生成的合成数据与原始数据符合KS同分布,说明GAN能够成功合成与原始图像结构空间分布模式一致的图像;2)Minkowski宏观参数评价误差的较小,KS同分布结果表明,多孔介质样本的结构异质性会在一定程度上影响GAN模型的计算精度和效率;3)Open FOAM模拟计算得到的渗透率结果表明,GAN模型生成的多孔介质图像与原始图像具有一致的统计特征和物理性质。综上,综合运用前沿的XCT扫描、图像处理技术和机器学习算法,构建了土壤微观结构重建和定量分析模型,并结合多孔介质形态学和计算流体力学方法对模型进行了验证和分析。该研究为多孔介质微观结构研究提供了新技术和新方法,为进一步研究复杂多孔介质内溶质运移提供了科技支撑。

注意力融合双流特征的局部GAN生成人脸检测算法

为解决现有局部生成式对抗网络(GAN)生成人脸检测算法在检测经过后处理的图像时性能严重下降的问题,提出一种注意力融合双流特征的局部GAN生成人脸检测算法.该算法利用双流网络分别从RGB颜色空间和YCbCr颜色空间中提取鲁棒特征,并引入注意力特征融合模块在不同网络层上融合双流特征以获得更鲁棒的特征.同时采用多层次特征融合决策提高网络对局部生成区域特征的提取和辨别能力.实验结果表明,所提算法的鲁棒性优于现有算法,尤其是针对JPEG压缩和双边滤波后处理.在FFHQ+规则子集上与次优算法相比,该算法在3种强度的JPEG压缩和双边滤波上的平均准确率分别提高了1.88%和2.64%;在FFHQ+不规则子集上与次优算法相比,该算法在3种强度的JPEG压缩和双边滤波上的平均准确率分别提高了2.85%和1.60%.

基于GAN的直扩信号生成算法

将深度学习模型应用至电子干扰技术来生成干扰信号具有重要的现实意义。将生成对抗网络(generative adversarial network,GAN)应用于信号生成领域,对电磁扩频信号频谱数据的分布进行深度学习,并生成与其相干的干扰信号。在实验中GAN的生成器和判别器互相博弈训练,通过自适应矩估计(adaptive moment estimation,Adam)进行优化,最终训练出良好的模型,可以生成所需信号。实验结果表明,基于GAN的信号生成算法生成的数据分布已基本具备真实数据分布普遍具有的特点,对同一信噪比的电磁频谱数据进行深度学习后,生成数据能够较为准确地学习到不同信噪比电磁频谱数据的不同特点。

DR-GAN:一种无监督学习的探地雷达杂波抑制方法

探地雷达(GPR)是一种基于电磁波的地下无损探测技术,广泛应用于市政工程、交通、军事等领域。在数据采集过程中,由于发射天线和接收天线之间的耦合、起伏地面的散射以及地下随机媒质的复杂性等原因,采集得到的GPR B-scan回波中通常存在杂波,杂波严重影响了地下目标的检测和特征提取。该文提出一种用于GPR B-scan图像杂波抑制的解纠缠表示生成对抗网络(DR-GAN),设计了目标特征编码器和杂波特征编码器用来提取GPR B-scan图像中的目标特征和杂波特征,设计了杂波抑制生成器用来获取杂波抑制后的GPR B-scan图像。与现有的基于监督学习的GPR杂波抑制方法相比,该方法在网络训练时不需要成对的匹配数据,可以更好地应用于实测GPR图像的杂波抑制。在仿真和实测GPR数据上的实验结果表明,DR-GAN这一无监督学习网络具有更好的杂波抑制性能。对石英砂中埋设的钢筋进行数据采集,运用DR-GAN对含杂波的实测数据进行处理,处理结果的改善系数(IF)指标较现有的鲁棒非负矩阵分解(RNMF)方法提高了17.85 dB。

用于脑电数据增强和情绪识别的自注意力GAN

针对脑电信号(electroencephalogram,EEG)情绪识别中数据稀缺及由此导致的情感分类精度不高的问题,提出了一个引入自注意力机制的条件Wasserstein生成对抗网络(SA-cWGAN),通过自注意力模块从训练数据学习长时上下文相关的全局特征,采用Wasserstein距离和梯度惩罚的Lipschitz约束对网络的损失函数进行优化,进而生成高质量的EEG数据对原有训练集进行增强。所提方法分别在DEAP和SEED数据集上进行了大量的二分类和三分类对比实验,生成了与EEG训练数据分布接近的微分熵(DE)和功率谱密度(PSD)特征,以此来增强EEG训练数据集,采用SVM分类器对增强后的EEG特征进行情绪分类。实验结果表明,在DEAP数据集上的唤醒度和效价维度下,增强后的DE、PSD特征较原有DE、PSD特征二分类准确率分别提高了16.63、17.55个百分点和6.48、8.34个百分点;在SEED数据集下,三分类准确率分别提高了4.64、5.18个百分点,证明所提方法生成的特征具有良好的鲁棒性,也表明通过对GAN网络引入自注意力机制生成的特征增强原有训练数据集能够有效提高EEG情绪识别的准确率和稳定性。

基于特征优化和GAN的红外与可见光图像融合算法

针对红外与可见光融合图像存在纹理细节不丰富、对比度较低及目标信息损失等问题,提出了一种基于特征优化和生成对抗网络的图像融合算法。首先,设计一种自适应特征优化模块以增强原始图像纹理细节及对比度;然后,为使融合图像保留更多的多模态信息,将生成对抗网络引入到融合框架中。在生成器模型中,考虑到红外与可见光图像成像机理差异,构建了双支路特征提取网络,并设计多尺度密集连接模块以提取异源图像丰富的特征信息;其次,在融合层构造通道和空间注意力模型以增强局部特征之间联系,减小融合图像中目标信息损失;最后,为使融合结果尽可能保留可见光纹理细节的同时又能够较好突出红外目标,构造了双判别器网络结构。为验证所提算法优势,在TNO数据集上进行实验,并与6种经典融合算法进行主观和客观比较。实验结果表明,所提算法无论在主观还是客观评价上均具有明显优势,生成的融合图像纹理细节更为丰富、边缘及目标更加清晰且具有更好的对比度,客观评价指标信息熵、空间频率、相关熵、视觉保真度和梯度信息分别提高了16.11%、65.46%、7.96%、42.67%和33.24%。