自适应融合残差网在图像分类中应用研究

针对卷积神经网络存在随着网络深度增加导致优化困难,识别正确率降低、泛化性能差等问题,在Res Net(残差网络)基础上,提出了一种基于softmax全连接自适应门控网络融合模型.该方法在隐层网络深度达到一定层数后,设置多种卷积核尺寸作为独立网络输出,通过softmax全连接门控网络输出各模型选择概率,融合多种卷积尺寸残差网输出作为模型最终输出.实验表明,本文提出的融合残差网络模型更适合于多类别、精细化数据集,与单网络模型相比,在训练集上具有更好的收敛性,在测试集上具有更好的泛化性能.     

The design of the user interface for a large physics experiment

The paper describes the experience in developing and supporting the user interface of RFX, one of the large nuclear fusion experiments of the co-ordinated nuclear fusion experiment programme of the European Community. The aim of this work is to present the problems and some possible solutions when developing user interfaces in a scientific environment, especially in large physics experiments. An overview of the current state of interface technology in such an environment is first provided. The control and data acquisition system of the RFX experiment is then introduced and its user interface described in greater detail. Finally, our experience both in maintaining the system interface and in training its users is described.     

Supporting multiple-input,multiple-output custom functions in configurable processors

Configurable processors have emerged as a promising solution for high performance embedded systems. Many of these processors extend a RISC core with configurable functional units that execute dual-input, single-output (DISO) custom functions. Although studies have shown that supporting multiple-input, multiple-output (MIMO) custom functions can lead to significant speedups, mechanisms to efficiently achieve this have not been adequately addressed. The underlying reason is that a custom function is normally invoked by a single instruction, which usually transfers only two inputs and one output. Attempts to transfer more inputs and outputs in one instruction are impeded by the instruction length and the register file’s R/W ports. This paper proposes a simple extension to transfer multiple inputs and outputs of the custom functions using repeated instructions. While transferring the inputs and outputs may take a few extra cycles, our experiments show that the MIMO extension can still achieve an average 51% increase in speedup compared to a DISO extension and an average 27% increase in speedup compared to a multiple-input, single-output (MISO) extension.     

Modeling of a dynamic dual-input dual-output fast steeringmirror system

A modeling method is proposed for a dynamic fast steering mirror (FSM) system with dual inputs and dual outputs. A physical model of the FSM system is derived based on first principles, describing the dynamics and coupling between the inputs and outputs of the FSM system. The physical model is then represented in a state-space form. Unknown parameters in the state-space model are identified by the subspace identification algorithm, based on the measured input-output data of the FSM system. The accuracy of the state-space model is evaluated by comparing the model estimates with measurements. The variance-accounted-for value of the state-space model is better than 97%, not only for the modeling data but also for the validation data set, indicating high accuracy of the model. Comparison is also made between the proposed dynamic model and the conventional static model, where improvement in model accuracy is clearly observed. The model identified by the proposed method can be used for optimal controller design for closed-loop FSM systems. The modeling method is also applicable to FSM systems with similar structures.     

Identification of the Tennessee Eastman challenge process with subspace methods

The Tennessee Eastman challenge process is a realistic simulation of a chemical process that has been widely used in process control studies. In this case study, several identification methods are examined and used to develop MIMO models that contain seven inputs and ten outputs. ARX and finite impulse response models are identified using reduced-rank regression techniques (PLS and CCR) and state-space models identified with prediction error methods and subspace algorithms. For a variety of reasons, the only successful models are the state-space models produced by two popular subspace algorithms, N4SID and canonical variate analysis (CVA). The CVA model is the most accurate. Important issues for identifying the Tennessee Eastman challenge process and comparisons between the subspace algorithms are also discussed.     

Comparison of input signals in subspace identification of multivariable ill-conditioned systems

Ill-conditioned processes often produce data of low quality for model identification in general, and for subspace identification in particular, because data vectors of different outputs are typically close to collinearity, being aligned in the “strong” direction. One of the solutions suggested in the literature is the use of appropriate input signals, usually called “rotated” inputs, which must excite sufficiently the process in the “weak” direction. In this paper open-loop (uncorrelated and rotated) random signals are compared against inputs generated in closed-loop operation, with the aim of finding the most appropriate ones to be used in multivariable subspace identification of ill-conditioned processes. Two multivariable ill-conditioned processes are investigated and as a result it is found that closed-loop identification gives superior models, both in the sense of lower error in the frequency response and in terms of higher performance when used to build a model predictive control system.     

引入低秩约束先验的深度子空间聚类

大多数子空间聚类算法将高维数据映射到低维子空间时不能较好捕获数据间几何结构.针对上述问题,文中提出引入低秩约束先验的深度子空间聚类算法,兼顾数据全局和局部结构信息.算法结合低秩表示与深度自编码器,利用低秩约束捕获数据全局结构,并将约束神经网络的潜在特征表示为低秩.自编码通过最小化重构误差进行非线性低维子空间映射,保留数据的局部特性.以多元逻辑回归函数作为判别模型,预测子空间分割.整个算法在无监督联合学习框架下进行优化.在5个数据集上的实验验证文中方法的有效性.     

Robust output tracking of constrained perturbed linear systems via model predictive sliding mode control

A robustifying strategy for constrained linear multivariable systems is proposed. A combination of tracking model predictive control with output integral sliding mode techniques is used to completely reject bounded matched perturbations. It can be guaranteed that all constraints on inputs, states, and outputs are satisfied although only output information is used. Finally, real‐world experiments with an unstable plant are presented in order to demonstrate the validity and the effectiveness of the proposed approach.     

Identification of positive real models in subspace identification by using regularization

In time-domain subspace methods for identifying linear-time invariant dynamical systems, the model matrices are typically estimated from least squares, based on estimated Kalman filter state sequences and the observed outputs and/or inputs. It is well known that for an infinite amount of data, this least squares estimate of the system matrices is unbiased, when the system order is correctly estimated. However, for a finite amount of data, the obtained model may not be positive real, in which case the algorithm is not able to identify a valid stochastic model. In this note, positive realness is imposed by adding a regularization term to a least squares cost function in the subspace identification algorithm. The regularization term is the trace of a matrix which involves the dynamic system matrix and the output matrix.     

The kernel common vector method: a novel nonlinear subspace classifier for pattern recognition.

The common vector (CV) method is a linear subspace classifier method which allows one to discriminate between classes of data sets, such as those arising in image and word recognition. This method utilizes subspaces that represent classes during classification. Each subspace is modeled such that common features of all samples in the corresponding class are extracted. To accomplish this goal, the method eliminates features that are in the direction of the eigenvectors corresponding to the nonzero eigenvalues of the covariance matrix of each class. In this paper, we introduce a variation of the CV method, which will be referred to as the modified CV (MCV) method. Then, a novel approach is proposed to apply the MCV method in a nonlinearly mapped higher dimensional feature space. In this approach, all samples are mapped into a higher dimensional feature space using a kernel mapping function, and then, the MCV method is applied in the mapped space. Under certain conditions, each class gives rise to a unique CV, and the method guarantees a 100% recognition rate with respect to the training set data. Moreover, experiments with several test cases also show that the generalization performance of the proposed kernel method is comparable to the generalization performances of other linear subspace classifier methods as well as the kernel-based nonlinear subspace method. While both the MCV method and its kernel counterpart did not outperform the support vector machine (SVM) classifier in most of the reported experiments, the application of our proposed methods is simpler than that of the multiclass SVM classifier. In addition, it is not necessary to adjust any parameters in our approach.     

Identification of stable models in subspace identification by usingregularization

In subspace identification methods, the system matrices are usually estimated by least squares, based on estimated Kalman filter state sequences and the observed inputs and outputs. For a finite number of data points, the estimated system matrix is not guaranteed to be stable, even when the true linear system is known to be stable. In this paper, stability is imposed by using regularization. The regularization term used here is the trace of a matrix which involves the dynamical system matrix and a positive (semi) definite weighting matrix. The amount of regularization can be determined from a generalized eigenvalue problem. The data augmentation method of Chui and Maciejowski (1996) is obtained by using specific choices for the weighting matrix in the regularization term     

LPV model identification for gain scheduling control: An application to rotating stall and surge control problem

We approach the problem of identifying a nonlinear plant by parameterizing its dynamics as a linear parameter varying (LPV) model. The system under consideration is the Moore–Greitzer model which captures surge and stall phenomena in compressors. The control task is formulated as a problem of output regulation at various set points (stable and unstable) of the system under inputs and states constraints. We assume that inputs, outputs and scheduling parameters are measurable. It is worth pointing out that the adopted technique allows for identification of an LPV model's coefficients without the requirements of slow variations amongst set points. An example of combined identification, feedback control design and subsequent validation is presented.     

关于非线性Morgan问题的几点注记

对输入数等于输出数加1情形的非线性Morgan问题,Glumineau&Moog曾给 出一个构造性解.这一结果最近被指出是错误的.文章给出上述结果的一个修正,并构造出 10阶系统,该系统具有唯一的解耦反馈,且解耦后的系统比原系统具有更大的本性阶.     

Target setting with minimum improving costs in data envelopment analysis: A mixed integer linear programming approach

This study addresses a problem called cost‐minimizing target setting in data envelopment analysis (DEA) methodology. The problem is how to make an inefficient decision‐making unit efficient by allocating to it as few organizational resources as possible, assuming that the marginal costs of reducing inputs or increasing outputs are known and available, which is different from previous furthest and closest DEA targets setting methods. In this study, an existed cost minimizing target setting heuristics approach based on input‐oriented model is examined to show that there exist some limitations. This study develops a simple mixed integer linear programming to determine the desired targets on the strongly efficient frontier based on non‐oriented DEA model considering the situation in the presence of known marginal costs of reducing inputs and increasing outputs simultaneously. Some experiments with the simulated datasets are conducted, and results show that the proposed model can obtain more accurate projections with lower costs compared with those from furthest and closest target setting approaches. Besides, the proposed model can be realistic and efficient in solving cost‐minimizing target setting problem.     

The robust control of a servomechanism problem for linear time-invariant multivariable systems

Necessary and sufficient conditions are found for there to exist a robust controller for a linear, time-invariant, multivariable system (plant) so that asymptotic tracking/regulation occurs independent of input disturbances and arbitrary perturbations in the plant parameters of the system. In this problem, the class of plant parameter perturbations allowed is quite large; in particular, any perturbations in the plant data are allowed as long as the resultant closed-loop system remains stable. A complete characterization of all such robust controllers is made. It is shown that any robust controller must consist of two devices 1) a servocompensator and 2) a stabilizing compensator. The servocompensator is a feedback compensator with error input consisting of a number of unstable subsystems (equal to the number of outputs to be regulated) with identical dynamics which depend on the disturbances and reference inputs to the system. The sorvocompensator is a compensator in its own right, quite distinct from an observer and corresponds to a generalization of the integral controller of classical control theory. The sole purpose of the stabilizing compensator is to stabilize the resultant system obtained by applying the servocompensator to the plant. It is shown that there exists a robust controller for "almost all" systems provided that the number of independent plant inputs is not less than the number of independent plant outputs to be regulated, and that the outputs to be regulated are contained in the measurable outputs of the system; if either of these two conditions is not satisfied, there exists no robust controller for the system.     

Decentralized control of multirate systems subject to arbitraryexogenous signals

This paper considers the design of decentralized digital control systems where the outputs are measured with different rates or at different times, whereas the inputs are updated all together. Specifically, a solution is given to the problem of asymptotically zeroing the system error when the exogenous signals coincide with the free motions of any unstable system. The proposed controller is constituted by a time-invariant internal model of the exogenous signals and a decentralized periodic stabilizer     

Akin-based Orthogonal Space (AOS): a subspace learning method for face recognition

A projection learning space is an approach to mapping a high-dimensional vector space to a lower dimensional vector space. In this paper, we proposed an algorithm, namely, AOS: Akin based Orthogonal Space. The algorithm is driven with two major targets - (i) to choose most representative image(s) from a group of face images of an individual, (ii) finally to produce a learning space which follows a Gaussian distribution to reduce the influence of grosses like non-Gaussianly distributed data noises, variations in facial expression and illumination. To improve the recognition performance, we proposed another approach i.e. fusion between AOS features and a custom VGG features. We justify the effectiveness of the proposed approaches over five benchmark face datasets using two classifiers. Experimental results show that the proposed learning algorithm has obtained maximum of 92.22% recognition rate, as well deep learning based fusion approch greatly improves the recognition accuracy. The comparative performances demonstrate that the proposed method could significantly outperform other relevant subspace learning methods.

     

Kernel methods for subspace identification of multivariable LPV and bilinear systems

Subspace identification methods for multivariable linear parameter-varying (LPV) and bilinear state-space systems perform computations with data matrices of which the number of rows grows exponentially with the order of the system. Even for relatively low-order systems with only a few inputs and outputs, the amount of memory required to store these data matrices exceeds the limits of what is currently available on the average desktop computer. This severely limits the applicability of the methods. In this paper, we present kernel methods for subspace identification performing computations with kernel matrices that have much smaller dimensions than the data matrices used in the original LPV and bilinear subspace identification methods. We also describe the integration of regularization in these kernel methods and show the relation with least-squares support vector machines. Regularization is an important tool to balance the bias and variance errors. We compare different regularization strategies in a simulation study.     

Determination of model order for NARX models directly from input-output data

Nonlinear auto-regressive models with exogenous inputs (NARX models) have proved to be versatile and useful empirical models for industrial processes. There are a wide variety of identification methods and model structures from which to choose; in all methods, however, key parameters are the model orders, which are the number of past outputs and inputs used in the model. In this paper the methods of Lipschitz numbers and false nearest neighbors are evaluated as a means of estimating the model orders of dynamic, discrete-time NARX models. No specific model structure is assumed and the model orders are estimated directly from input-output data using only geometric concerns and the continuity property. The two methods are applied to several dynamic systems, including realistic process simulations and experimental data from the UCSB pH neutralization process, and the consistency and accuracy of these methods are examined. The usefulness of these methods of model order determination for radial basis function network (RBFN) identification is examined.     

A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

Data envelopment analysis (DEA) is a linear programming based non-parametric technique for evaluating the relative efficiency of homogeneous decision making units (DMUs) on the basis of multiple inputs and multiple outputs. There exist radial and non-radial models in DEA. Radial models only deal with proportional changes of inputs/outputs and neglect the input/output slacks. On the other hand, non-radial models directly deal with the input/output slacks. The slack-based measure (SBM) model is a non-radial model in which the SBM efficiency can be decomposed into radial, scale and mix-efficiency. The mix-efficiency is a measure to estimate how well the set of inputs are used (or outputs are produced) together. The conventional mix-efficiency measure requires crisp data which may not always be available in real world applications. In real world problems, data may be imprecise or fuzzy. In this paper, we propose (i) a concept of fuzzy input mix-efficiency and evaluate the fuzzy input mix-efficiency using α – cut approach, (ii) a fuzzy correlation coefficient method using expected value approach which calculates the expected intervals and expected values of fuzzy correlation coefficients between fuzzy inputs and fuzzy outputs, and (iii) a new method for ranking the DMUs on the basis of fuzzy input mix-efficiency. The proposed approaches are then applied to the State Bank of Patiala in the Punjab state of India with districts as the DMUs.