基于高效Transformer的图像去雾研究

随着Transformer技术发展使其在计算机视觉(CV)领域中大放异彩,其构建全局依赖关系能力是卷积神经网络(CNN)所不具备的,而Transformer对图像尺寸变化敏感与处理图像位置信息方面却不如CNN,因此,近几年基于Transformer与CNN的网络结构在CV领域是研究热点。图像去雾技术可用于车载避让系统、遥感系统以及物联网设备被设为公共领域的热点问题,因此,本文提出基于Transformer的网络结构实现图像去雾。结合Transformer与UNet构建网络框架,针对Self-Attention因二次判别导致高计算复杂度的问题,我们使用聚焦线性注意力替换Softmax函数,将二次计算复杂度降为O(n),并且为了解决使用线性注意力引起性能下降的问题,使用恢复模块增加输出特征的多样性,实现性能的恢复。经实验表明,与现有的去雾霾网络相比,文章提出的方法在公开数据集RESIDE上获得了更好的实验结果以及更愉悦的去雾效果。

基于Transformer-LSTM网络的轴承寿命预测

轴承是旋转机械设备中的重要部件,由于工况、材质、加工方式等原因,轴承寿命时长相差许多。传统的并行或串行神经网络预测方式,对数据集有一定要求。因此,需要一种能够适用于不同数据长短的轴承剩余使用寿命预测网络。为此提出了一种能够预测不同寿命时长的Transformer-LSTM串并行神经网络预测模型。通过将Transformer解码层进行重构,并与长短期记忆时序神经网络(long short-term memory,LSTM)网络结构融合,实现轴承寿命数据的串并行预测处理。试验结果表明Transformer-LSTM神经网络能够精准预测长、中、短不同寿命时长的轴承失效时间,具有较强的模型泛化能力,提升轴承寿命预测精度与模型的泛化能力。

An Improved Robust Model Predictive Control Approach to Systems with Linear Fractional Transformation Perturbations

In this paper, a robust model predictive control approach is proposed for a class of uncertain systems with time-varying, linear fractional transformation perturbations. By adopting a sequence of feedback control laws instead of a single one, the control performance can be improved and the region of attraction can be enlarged compared with the existing model predictive control （MPC） approaches. Moreover, a synthesis approach of MPC is developed to achieve high performance with lower on-line computational burden. The effectiveness of the proposed approach is verified by simulation examples.

Linear Complementary Method for Moving Boundary Problems with Phase Transformation

In this paper, the linear complementary method for moving boundary problems with phase transformation is presented, in which a pair of unknown vectors of heat source with phase transforming and the temperature field can be solved exactly, and a large amount of iterative calculations can be avoided.

基于对抗训练与Transformer的风力发电机故障分类方法

风力发电机故障分类的复杂性和多样性严重影响风能发电效率,传统的人工方法效率低下,准确率较低,已有的深度学习模型在真实环境中易受数据噪声干扰而表现不佳。为提升风力发电机故障分类模型在真实环境下的分类性能与鲁棒性,提出一种基于对抗训练与Transformer的故障分类方法。首先通过引入一维卷积与门控线性单元(GLU)增强注意力机制对局部特征的学习,保留易被忽略的局部信息,提升模型对于局部特征的敏感度。其次结合限制因子约束对抗样本,提高对抗样本产生的准确性。最后在消除错误样本的同时反馈生成过程,使其具备更好的抗干扰能力。实验结果表明,与5种常用的分类模型相比,所提模型分类性能平均提升7.76%,与真实结果之间的误差最小。局部增强的注意力机制和所提的对抗训练方法分别使模型的分类性能平均提升4.51%、4.95%。所提模型在10%~20%噪声环境中仍保持较好性能,增强了其在真实环境中的稳定性。该方法在提高分类准确率的同时使模型具备更强的泛化能力,对于提升风力发电机故障分类性能与鲁棒性具有重要意义。

基于局部Transformer的泰语分词和词性标注联合模型

泰语分词和词性标注任务二者之间存在高关联性,已有研究表明将分词和词性标注任务进行联合学习可以有效提升模型性能,为此,提出了一种针对泰语拼写和构词特点的分词和词性标注联合模型。针对泰语中字符构成音节,音节组成词语的特点,采用局部Transformer网络从音节序列中学习分词特征;考虑到词根和词缀等音节与词性的关联,将用于分词的音节特征融入词语序列特征,缓解未知词的词性标注特征缺失问题。在此基础上,模型采用线性分类层预测分词标签,采用线性条件随机场建模词性序列的依赖关系。在泰语数据集LST20上的试验结果表明,模型分词F1、词性标注微平均F1和宏平均F1分别达到96.33%、97.06%和85.98%,相较基线模型分别提升了0.33%、0.44%和0.12%。

A Weakly-Supervised Crowd Density Estimation Method Based on Two-Stage Linear Feature Calibration

In a crowd density estimation dataset,the annotation of crowd locations is an extremely laborious task,and they are not taken into the evaluation metrics.In this paper,we aim to reduce the annotation cost of crowd datasets,and propose a crowd density estimation method based on weakly-supervised learning,in the absence of crowd position supervision information,which directly reduces the number of crowds by using the number of pedestrians in the image as the supervised information.For this purpose,we design a new training method,which exploits the correlation between global and local image features by incremental learning to train the network.Specifically,we design a parent-child network(PC-Net)focusing on the global and local image respectively,and propose a linear feature calibration structure to train the PC-Net simultaneously,and the child network learns feature transfer factors and feature bias weights,and uses the transfer factors and bias weights to linearly feature calibrate the features extracted from the Parent network,to improve the convergence of the network by using local features hidden in the crowd images.In addition,we use the pyramid vision transformer as the backbone of the PC-Net to extract crowd features at different levels,and design a global-local feature loss function(L2).We combine it with a crowd counting loss(LC)to enhance the sensitivity of the network to crowd features during the training process,which effectively improves the accuracy of crowd density estimation.The experimental results show that the PC-Net significantly reduces the gap between fullysupervised and weakly-supervised crowd density estimation,and outperforms the comparison methods on five datasets of Shanghai Tech Part A,ShanghaiTech Part B,UCF_CC_50,UCF_QNRF and JHU-CROWD++.

THE BLUE AND MINQUE IN GAUSS-MARKOFF MODEL WITH LINEAR TRANSFORMATION OF THE OBSERVABLE VARIABLES

For a singular linear model A = （y, Xβ, σ2 V） and its transformed model MF = （Fy, FXβ, σ 2FVF＇）, where V is nonnegative definite and X can be rank-deficient, the expressions for the differences of the estimates for the vector of FXβ and the variance factor σ2 are given. Moreover, the necessary and sufficient conditions for the equalities of the estimates for the vector of FXβ and the variance factor σ2 are also established. In the meantime, works in Baksalary and Kala （1981） are strengthened and consequences in Puntanen and Nurhonen （1992）, and Puntanen （1996） are extended.

NOVEL LINEAR TRANSFORMATION SWITCHED-CAPACITOR FILTER DESIGN

This paper presents a novel method that is applied to realize the Linear Transformation(LT)Switched-Capacitor Filter(SCF).It adopts the Voltage Control Voltage Source(VCVS)equalized transfor-mation to revise the original LC ladder filter and induce it into 16 basic sections and then extend the princi-ple of the LT in order to fit active and 3 port networks and give out switched-capacitor circuits corre-sponding to the 16 basic sections,which can realize all four kinds of filters——LP,HP,BP,BS filters.De-signed examples are given here.An Nth order filter only requires N amplifiers and the circuit is insensitive toparasitic capacitances.The experimental results of a 3rd order elliptic LP and a 6th order elliptic BP are giv-en and agree with the theory.

A DECOMPOSITION DEPENDED ON DIRECTIONS FOR NONSINGULAR LINEAR TRANSFORMATION

In this paper. we give a decomposition depending on p(1≤p≤n-2) orthonormaldirections assigned for nonsingular linear transformation F on a n-dimension (n≥3)Euclidean space En, and then prove foal there exist q(q=n-p) quasi-Principaldirections.for F depending on the preceding p orthonormal directions. As applicance ofthe preceding result, we derive that there exist at least two orthonormal principaldirections of strain in arbitrary plane of body which is in homogeneous deformation,and strain energy density is.function of 5 real numbers under arbitrary quasi-principalbase.for the preceding nonsingular linear transformation.

Criteria for System of Three Second-Order Ordinary Differential Equations to Be Reduced to a Linear System via Restricted Class of Point Transformation

This paper is devoted to the study of the linearization problem of system of three second-order ordinary differential equations and . The necessary conditions for linearization by general point transformation and are found. The sufficient conditions for linearization by restricted class of point transformation and are obtained. Moreover, the procedure for obtaining the linearizing transformation is provided in explicit forms. Examples demonstrating the procedure of using the linearization theorems are presented.

Linear Transformation Theory of Quantum Field Operators and Its Applications

We extend the linear quantum transformation theory to the case of quantum field operators. The corresponding general transformation expressions of CPT transformations and gauge field transformations are considered as its applications.

On the Group of Units for the Ring of Linear Transformations

In this paper, we introduce a practical method for obtaining the structure of thegroup of units for the ring of linear transformations of a vector space over an arbitrary field,and we give a further generalization of the result in [3].

ON THE DECOMPOSITION OF COMPLEX VECTOR SPACES AND THE JORDAN CANONICAL FORM OF COMPLEX LINEAR TRANSFORMATIONS

New objects characterizing the structure of complex linear transformations areintroduced. These new objects yield a new result for the decomposition of complexvector spaces relative to complex lrnear transformations and provide all Jordan basesby which the Jordan canonical form is constructed. Accordingly, they can result in thecelebrated Jordan theorem and the third decomposition theorem of space directly. and,moreover, they can give a new deep insight into the exquisite and subtle structure ofthe Jordan form. The latter indicates that the Jordan canonical form of a complexlinear transformation is an invariant structure associated with double arbitrary. choices.

From Translation to Linear and Linear Canonical Transformations

In order to obtain with simplicity the known and new properties of linear canonical transformations (LCTs), we show that any relation between a couple of operators (A,B) having commutator identical to unity, called dual couple in this work, is valuable for any other dual couple, so that from the known translation operator exp(a∂x) one may obtain the explicit form and properties of a category of linear and linear canonical transformations in 2N-phase spaces. Moreover, other forms of LCTs are also obtained in this work as so as the transforms by them of functions by integrations as so as by derivations. In this way, different kinds of LCTs such as Fast Fourier, Fourier, Laplace, Xin Ma and Rhodes, Baker-Campbell-Haussdorf, Bargman transforms are found again.

Linearization of Emden Differential Equation via the Generalized Sundman Transformations

The Emden differential equation is one of the most widely studied and challenging nonlinear dynamics equations in literature. It finds applications in various areas of study such as celestial mechanics, fluid mechanics, Steller structure, isothermal gas spheres, thermionic currents and so on. Because of the importance of the equation, the method of generalized Sundman transformation (GST) as proposed by Nakpim and Meleshko is used for linearizing the Emden differential equation. The Emden differential equation considered here is a modification of the equation given by Berkovic. The results obtained in this paper imply that the Emden equation cannot be linearized by a point transformation. The general solution of the modified Emden equation is also obtained.

Rhotrix Linear Transformation

This paper considers rank of a rhotrix and characterizes its properties, as an extension of ideas to the rhotrix theory rhomboidal arrays, introduced in 2003 as a new paradigm of matrix theory of rectangular arrays. Furthermore, we present the necessary and sufficient condition under which a linear map can be represented over rhotrix.

Classical and Quantum Behavior of Generalized Oscillators in Terms of Linear Canonical Transformations

The quantum mechanical relationships between time-dependent oscillators, Hamilton-Jacobi theory and an invariant operator are clarified by making reference to a system with a generalized oscillator. We introduce a linear transformation in position and momentum, and show that the correspondence between classical and quantum transformations is exactly one-to-one. We found that classical canonical transformations are constructed from quantum unitary transformations as long as we are concerned with linear transformations. We also show the relationship between the invariant operator and a linear transformation.

Numerical Experiments and Structural Transformation of Non-Linear Evolution in Ill-posed Systems

Numerical experiments on non-linear equations of the 1st-and 3rd-order derivatives have been carried out through structural analyses in the phase space according to the numerical instability of ill-posed systems,with changes of initial values and parameters,etc..The results show that the quantitative instability in an ill-posed system may reveal reversed transformation in system evolution by structural representation,and confirm A·Dauglas' theorem that "a non-linear equation does not satisfy the existence of the initial value in a linear well-posed system".

IMPLEMENTATION OF THE HYBRID GMRES ALGORITHM USING HOUSEHOLDER TRANSFORMATIONS

The standard implementation of the hybrid GMRES algorithm for solving large nonsymmetric linear systems involves a Gram-Schmidt process which is a potential source of significant numerical error. An alternative implementation is outlined here in which orthogonalization by Householder transformations replaces the Gram-Schmidt process. Numerical experiments show that the new implementation is more stable.