小样本下基于改进ACGAN数据增强的X射线矿石图像分类方法

针对工业领域利用深度学习模型对矿石进行在线分类时,矿石样本稀少导致的模型过拟合、分类准确率低的问题,提出一种结合X射线透射成像技术的矿石数据增强分类方法。该方法基于改进辅助生成对抗网络(Enhance-based Classification ACGAN-gp, EC-ACGAN-gp),采用卷积和连续残差块构建判别器和生成器,引入注意力机制捕捉矿石细节特征,生成高质量样本扩充原始数据集,同时使用带梯度惩罚的Wasserstein距离重构判别器的损失函数提高对抗训练的稳定性,避免模式崩溃。通过增加辅助分类器重建样本标签信息,最终实现矿石样本的类别预测。结果表明,该方法能实现矿石品位分类的精准预测,准确率可达89.62%,比现有传统方法提高3.98%。该模型生成的矿石样本泛化性良好,能够显著提高小样本数据集的泛化性,在SVM、LeNet5、VGGNet、ResNet上测试,精度分别提升了2.83%、2.36%、1.89%和3.74%,可进一步用于提升其他分类模型在矿石品位预测方面的性能。

基于CVAE-ACGAN特征生成模型的轴承故障诊断

针对传统故障特征生成模型在学习训练时具备不可控性、单一性和收敛速度慢的缺点从而导致故障诊断模型的诊断效果不理想的问题,通过将条件变分自动编码器(CVAE)模型可结合故障数据的类别属性进行隐含特征提取的优势与辅助分类生成式对抗网络(ACGAN)模型较好的提取类条件特征能力相结合,创新性地提出CVAE-ACGAN特征生成模型。在提高生成特征质量的基础上,兼顾模型的收敛速度和抗干扰能力。以公开轴承数据集作为数据源,与4种特征生成模型的故障诊断效果进行对比。结果表明,CVAE-ACGAN模型可在实际故障诊断中对故障数据集进行有效的扩充,进而提高故障诊断的精确度。

Ascertaining the influences of auxiliary qubits on the Einstein–Podolsky–Rosen steering and its directions

Einstein–Podolsky–Rosen(EPR) steering is an example of nontrivial quantum nonlocality and characteristic in the non-classical world.The directivity(or asymmetry) is a fascinating trait of EPR steering,and it is different from other quantum nonlocalities.Here,we consider the strategy in which two atoms compose a two-qubit X state,and the two atoms are owned by Alice and Bob,respectively.The atom of Alice suffers from a reservoir,and the atom of Bob couples with a bit flip channel.The influences of auxiliary qubits on EPR steering and its directions are revealed by means of the entropy uncertainty relation.The results indicate that EPR steering declines with growing time t when adding fewer auxiliary qubits.The EPR steering behaves as damped oscillation when introducing more auxiliary qubits in the strong coupling regime.In the weak coupling regime,the EPR steering monotonously decreases as t increases when coupling auxiliary qubits.The increases in auxiliary qubits are responsible for the fact that the steerability from Alice to Bob(or from Bob to Alice) can be more effectively revealed.Notably,the introductions of more auxiliary qubits can change the situation that steerability from Alice to Bob is certain to a situation in which steerability from Bob to Alice is certain.

A new adaptive co-site broadband interference cancellation method with auxiliary channel

With the boom of the communication systems on some independent platforms(such as satellites,space stations,airplanes,and vessels),co-site interference is becoming prominent.The adaptive interference cancellation method has been adopted to solve the co-site interference problem.But the broadband interference cancellation performance of traditional Adaptive Co-site Interference Cancellation System(ACICS)with large delay mismatching and antenna sway is relatively poor.This study put forward an Adaptive Co-site Broadband Interference Cancellation System With Two Auxiliary Channels(ACBICS-2A).The system model was established,and the steady state weights and Interference Cancellation Ratio(ICR)were deduced by solving a time-varying differential equation.The relationship of ICR,system gain,modulation factor,interference signal bandwidth and delay mismatching degree was acquired through an in-depth analysis.Compared with traditional adaptive interference cancellation system,the proposed ACBICS-2A can improve broadband interference cancellation ability remarkably with large delay mismatching and antenna sway for the effect of auxiliary channel.The maximum improved ICR is more than 25 dB.Finally,the theoretical and simulation results were verified by experiments.

A GENERALIZED SCALAR AUXILIARY VARIABLE METHOD FOR THE TIME-DEPENDENT GINZBURG-LANDAU EQUATIONS

This paper develops a generalized scalar auxiliary variable(SAV)method for the time-dependent Ginzburg-Landau equations.The backward Euler method is used for discretizing the temporal derivative of the time-dependent Ginzburg-Landau equations.In this method,the system is decoupled and linearized to avoid solving the non-linear equation at each step.The theoretical analysis proves that the generalized SAV method can preserve the maximum bound principle and energy stability,and this is confirmed by the numerical result,and also shows that the numerical algorithm is stable.

Research on uninterruptable power supply technology for auxiliary winding of electric locomotive while passing the neutral section

Purpose–Auxiliary power system is an indispensable part of the train;the auxiliary systems of both electric locomotives and EMUs mainly are powered by one of the two ways,which are either from auxiliary windings of traction transformers or from DC-link voltage of traction converters.Powered by DC-link voltage of traction converters,the auxiliary systems were maintained of uninterruptable power supply with energy from electric braking.Meanwhile,powered by traction transformers,the auxiliary systems were always out of power while passing the neutral section of power supply grid and control system is powered by battery at this time.Design/methodology/approach–Uninterrupted power supply of auxiliary power system powered by auxiliary winding of traction transformer was studied.Failure reasons why previous solutions cannot be realized are analyzed.An uninterruptable power supply scheme for the auxiliary systems powered by auxiliary windings of traction transformers is proposed in this paper.The validity of the proposed scheme is verified by simulation and experimental results and on-site operation of an upgraded HXD3C type locomotive.This scheme is attractive for upgrading practical locomotives with the auxiliary systems powered by auxiliary windings of traction transformers.Findings–This scheme regenerates braking power supplied to auxiliary windings of traction transformers while a locomotive runs in the neutral section of the power supply grid.Control objectives of uninterrupted power supply technology are proposed,which are no overvoltage,no overcurrent and uninterrupted power supply.Originality/value–The control strategies of the scheme ensure both overvoltage free and inrush current free when a locomotive enters or leaves the neutral section.Furthermore,this scheme is cost low by employing updated control strategy of software and add both the two current sensors and two connection wires of hardware.

2024 National Practical Technical Training on Textile Printing and Dyeing Auxiliary Preparation Succeeds

From July 16th to 19th,2024,2024 National Practical Technical Training on Textile Printing and Dyeing Auxiliary Preparation,organized by China Research Institute of Daily Chemical and National Engineering Research Center for Surfactants,and co-organized by the Dyeing and Finishing Technology Research Institute of Zhejiang Fashion Institute of Technology and the High Fastness Plant-based Dye Research Base of China Textile Engineering Society,was held successfully in Shangyu,Zhejiang Province,with 92 representatives from 64 work units from all over China.

CycleGAN、ACGAN在人工智能医疗器械数据增广中的应用

目的探究人工智能医疗器械领域中使用循环生成对抗网络(Cycle-Consistent Generative Adversarial Networks,CycleGAN)和辅助分类生成对抗网络(Auxiliary Classification Generative Adversarial Network,ACGAN)进行数据增广的方法。方法使用CycleGAN和ACGAN分别生成干扰图像和特定领域数据,对图像增加不规律的变换,对原始图像数据进行数据加工或应用生成对抗网络生成该领域所需的图像数据。结果在医学影像数据集上评估了本文提出方法的性能,结果表明,CycleGAN和ACGAN可有效生成逼真的医学影像,从而用于训练机器学习模型。结论该方法解决了人工智能领域图像数据不足的问题,保证了模型对该数据的不可见性,使后期模型评估结果更准确。

Auxiliary Software for Defining the Parameters of the Structural Organization of a Complex System

The developed auxiliary software serves to simplify, standardize and facilitate the software loading of the structural organization of a complex technological system, as well as its further manipulation within the process of solving the considered technological system. Its help can be especially useful in the case of a complex structural organization of a technological system with a large number of different functional elements grouped into several technological subsystems. This paper presents the results of its application for a special complex technological system related to the reference steam block for the combined production of heat and electricity.

基于ACGAN的图像识别算法

针对基于辅助分类器生成对抗网络(ACGAN)的图像分类算法在训练过程中稳定性低且分类效果差的问题,提出一种改进的图像识别算法CP-ACGAN。对于网络结构,在判别网络的输出层取消样本的真假判别,只输出样本标签的后验估计并引入池化层。对于损失函数,除真实样本的交叉熵损失外,在判别网络中增加生成样本的条件控制标签及后验估计间的交叉熵损失。在此基础上,利用真假样本的交叉熵损失及属性重构生成器和判别器的损失函数。在MNSIT、CIFAR10、CIFAR100数据集上的实验结果表明,与ACGAN算法、CNN算法相比,该算法具有较好的分类效果与稳定性,且分类准确率分别高达99.62%、79.07%、48.03%。

融合句嵌入的VAACGAN多对多语音转换

针对非平行文本条件下语音转换质量不理想、说话人个性相似度不高的问题,提出一种融合句嵌入的变分自编码辅助分类器生成对抗网络(VAACGAN)语音转换方法,在非平行文本条件下,有效实现了高质量的多对多语音转换。辅助分类器生成对抗网络的鉴别器中包含辅助解码器网络,能够在预测频谱特征真假的同时输出训练数据所属的说话人类别,使得生成对抗网络的训练更为稳定且加快其收敛速度。通过训练文本编码器获得句嵌入,将其作为一种语义内容约束融合到模型中,利用句嵌入包含的语义信息增强隐变量表征语音内容的能力,解决隐变量存在的过度正则化效应的问题,有效改善语音合成质量。实验结果表明:所提方法的转换语音平均MCD值较基准模型降低6.67%,平均MOS值提升8.33%,平均ABX值提升11.56%,证明该方法在语音音质和说话人个性相似度方面均有显著提升,实现了高质量的语音转换。

基于ACGAN和模型融合的电机轴承故障诊断方法

电机轴承疲劳试验成本较高和故障数据不足导致利用机器学习等人工智能算法进行故障诊断时效果不佳。另外,单一模型对电机轴承故障诊断的准确率也较低。为解决这两个问题,提出了一种结合辅助分类器生成对抗网络(ACGAN)和模型融合的电机轴承故障诊断方法。首先将采集到的振动数据转换为二维灰度图,对每个灰度图添加标签后输入ACGAN模型,生成大量与原始数据高度拟合的新样本。然后将新样本与原始样本混合,经数据降维后输入由6个基学习器和1个元学习器融合而成的模型中。最后由融合模型输出诊断结果。试验证明,ACGAN和模型融合能有效提高电机轴承故障诊断的准确率。

基于改进ACGAN算法的车道排队车辆估计及其分类

针对传统模型驱动的排队车辆研究中构建概率分布困难、建模繁琐等问题,结合双向长短时记忆(Bi-LSTM)网络和辅助分类器生成对抗网络(ACGAN)的特点,提出一种数据驱动的车道级排队车辆估计算法。该算法无需对交叉口空间关系建模,其生成器采用Bi-LSTM结构,以速度序列为输入,根据速度与排队车辆的时间相关性,生成最小、最大排队车辆。判别器来自ACGAN,在区分真假样本的同时实现排队车辆到拥堵等级标签的分类。同时,为避免网络训练不稳定、梯度消失的问题,舍弃原ACGAN的真假二分类任务,引入Wasserstein散度来衡量真实序列与生成序列的分布距离,并对相应的目标函数进行优化。结果表明,与其他算法相比,该算法在分类准确率方面提高了3.96%~9.62%,同时总体估计误差最小,验证了利用速度估计车道排队车辆的可行性。

基于改进ACGAN的永磁同步电机数据扩张方法

永磁同步电机(permanent magnet synchronous motor,PMSM)的监测数据呈现出非平稳、非线性、多源异构性和价值低密度性等特点,而仿真数据难以准确地模拟电机故障类型和故障程度,使得正常数据与故障数据的样本呈现严重不均衡现象,导致故障诊断的模型训练容易出现过拟合、精度低等问题。本文提出了一种改进辅助分类生成对抗网络(auxiliary classification generation adversarial network,ACGAN),通过对原始样本的分布特性进行学习,实现对PMSM实测故障数据的扩张,为电机的故障诊断和健康评估提供数据基础。首先,针对ACGAN网络收敛性差和梯度易消失或爆炸的问题,使用Wasserstein距离约束生成数据的重建损失,利用梯度惩罚代替权值剪裁对模型进行优化,解决模型训练不稳定问题;其次,剖析数据之间的变化关系和历史变化规律,在生成器中引入循环神经网络提高生成数据质量;最后,利用PMSM匝间短路的故障数据,对比分析ROS、SMOTE、ADASYN及改进ACGAN 4种数据扩张方法对提升故障诊断模型性能的有效性。分析结果表明,与其他数据扩张方法相比,改进ACGAN方法的模型训练较稳定、收敛速度较快,扩张数据质量较高。

基于改进ACGAN样本增强的变压器故障诊断技术

基于深度学习算法的故障诊断需要足量的样本作为训练数据集。变压器故障数据匮乏将导致故障诊断准确率较低。对此,提出了一种基于辅助分类GAN的故障诊断方法。该方法引入自我注意机制,提取故障样本的全局特性,以提高生成样本的质量;并加入梯度惩罚,以提高模型收敛速度和训练稳定性。运用该方法对失衡样本进行增强扩充,并在变压器振动试验数据集上进行验证。仿真结果表明,该方法能够有效改善数据不平衡带来的影响,增强扩充后的故障诊断准确率提高了3.4%。

基于ACGAN的功率变换器参数性故障诊断方法

提出了一种基于辅助分类生成对抗网络的功率变换器参数性故障智能诊断方法。首先采集功率变换器的测点电压与支路电流信号,提取信号的时域特征,构成故障特征向量。采用对抗学习机制训练生成器和判别器,由ACGAN中生成器构造与真实故障特征分布近似的伪数据,从而将伪数据与真实数据同时用于训练判别器,判别器通过判别真伪数据来训练生成器。以Buck变换器为例,验证了所提出的故障诊断方法的可行性,结果表明ACGAN故障诊断方法相对于传统神经网络具有更高的故障诊断率与更优的泛化性能。

基于优化ACGAN-GBDT的个人信用风险评估模型研究

针对个人信用数据不平衡、类间重叠、类型多样性等特点,运用优化的辅助分类器生成对抗网络(ACGAN)、梯度提升决策树(GBDT)分别进行数据过采样、学习与分类,在此基础上构建个人信用风险评估模型。依据金融及大数据相关竞赛平台提供的两个信贷数据集进行实证,从AUC、G-mean、Recall等指标出发考量模型的性能。结果显示,模型使用新的过采样技术生成的样本与原始样本非常接近,对违约样本及总样本的识别性能均优于对照模型。

基于ACGAN和迁移学习的骨显像分类方法

由于骨显像存在样本数量有限、类别不平衡的问题,导致骨显像分类存在较大困难。为提升骨显像的分类准确率,本文提出了一种基于结合辅助分类器的生成对抗网络(ACGAN)数据生成和迁移学习的骨显像分类方法。首先,为解决骨显像类别不平衡的问题,设计了一种MU-ACGAN模型。该模型以U-Net为生成器框架,同时结合密集残差连接和通道-空间注意力机制结构来提升骨显像细节特征生成,判别器通过密集残差注意力卷积块提取骨显像特征进行判别;然后,结合传统数据增强方式进一步扩充数据量;最后,设计了一种多尺度卷积神经网络提取骨显像不同尺度的特征,提升分类效果。在模型训练过程中,采用两阶段迁移学习方式,优化模型的初始化参数、解决过拟合的问题。实验结果表明,本文提出方法分类准确率达到了85.71%,有效缓解了小样本骨显像数据集分类准确率不高的问题。

Combining sum-difference and auxiliary beams for adaptive monopulse in jamming

Only in the presence of sidelobe jamming （SLJ）, can the conventional adaptive monopulse technique null the jamming effectively and maintain the monopulse angle estimation accuracy simultaneously. While mainlobe jamming （MLJ） exists, the mainlobe of adaptive pattern will subject to serious distortion, which results in a failure of detecting and tracking targets by monopulse technique. Therefore, a monopulse angle estimation algorithm based on combining sum-difference beam and auxiliary beam is presented. This algorithm utilizes both high gain difference beams and high gain auxiliary beams for cancelling the mainlobe jammer and multiple sidelobe jammers （SLJs） while keeping an adap- tive monopulse ratio. Theoretical analysis and simulation results indicate that the serious invalidation of monopulse technique in MLJ and SLJs scenarios is resolved well, which improves the monopulse angle accuracy greatly. Furthermore, the proposed algorithm is of simple implementation and low computational complexity.

The extended auxiliary the KdV equation with equation method for variable coefficients

This paper applies an extended auxiliary equation method to obtain exact solutions of the KdV equation with variable coefficients. As a result, solitary wave solutions, trigonometric function solutions, rational function solutions, Jacobi elliptic doubly periodic wave solutions, and nonsymmetrical kink solution are obtained. It is shown that the extended auxiliary equation method, with the help of a computer symbolic computation system, is reliable and effective in finding exact solutions of variable coefficient nonlinear evolution equations in mathematical physics.