95 results about "Z eigenvalue" patented technology

The deep Convolutional Neural Networks (CNN) has vast amount of parameters, especially in the Fully Connected (FC) layers, which has become a bottleneck for real-time sensing where processing latency is high due to computational cost. In this invention, we propose to optimize the FC layers in CNN for real-time sensing via making it much slimmer. We derive a CNN Design and Optimization Theorem for FC layers from information theory point of view. The optimization criteria is eigenvalues-based, so we apply Singular Value Decomposition (SVD) to find the maximal eigenvalues and QR to identify the corresponding columns in FC layer. Further, we propose Efficient Weights for CNN Design Theorem, and show that weights with colored Gaussian are much more efficient than those with white Gaussian. We evaluate our optimization approach to AlexNet and apply the slimmer CNN to ImageNet classification. Testing results show our approach performs much better than random dropout.

A method for computing conformal parameterizations is revealed. First discrete conformal maps are reviewed for computing a generalized eigenvalue problem (GEP) arising from spectral conformal parameterization. Then nonequivalence deflation and null-space free compression techniques are applied to transform the GEP to a small-scaled compressed and deflated standard eigenvalue problem (CDSEP). Lastly a skew-Hamiltonian isotropic Lanczos algorithm (SHILA) is used to solve the CDSEP. Numerical experiments and comparisons are presented to show that the present method compute the conformal parameterization accurately and efficiently.

The invention discloses a SOD-IRK based time-delay power system stability determination method. The SOD-IRK based time-delay power system stability determination method comprises: establishing a time-delay power system model; using different implicit Runge-Kutta methods to discretize solution operators to obtain a discretized matrix of the solution operators; transforming an eigenvalue problem ofinfinite dimension into an eigenvalue problem of finite dimension; using an implicit Arnoldi algorithm to calculate an approximate value of the eigenvalue having the maximum modulus, of the discretized matrix of the solution operators; using a rotation-amplifying method to perform preprocessing during calculation, and performing sparse implementation by using the properties of a Kronecker product;transforming the approximate value of the eigenvalue having the maximum modulus, of the discretized matrix into the approximate eigenvalue of the time-delay power system model according to the spectral mapping relationship; using the Newton iteration method to correct the approximate eigenvalues to obtain the exact eigenvalues of the time-delay power system; and determining the stability of the time-delay power system according to the magnitude of the exact eigenvalue damping ratio.

The invention relates to a soft-input soft-out (SISO) minimum mean squared error (MMSE) iteration receiving method based on eigenvalue decomposition, which is characterized by comprising the following steps of: first, figuring out an equivalent channel matrix and an equivalent transmission correlation matrix by channel estimation and precoding of codebook information; carrying out eigenvalue decomposition on the equivalent transmission correlation matrix to acquire an eigenvalue and an eigenvector; inputting the equivalent channel matrix, the eigenvalue, the eigenvector and received signals into an SISO-MMSE detector, wherein the detector and an SISO decoder iteratively work by utilizing soft information input from the detector and the SISO detector as priori information; and finally, outputting a bit judgment by the decoder after the predefined iterations are reached. In the method of the invention, eigenvalue decomposition is introduced to transform the matrix inversion operation in each SISO-MMSE iteration process into division operation, thus the implementation complexity of the system is effectively reduced.

The invention discloses a hyperspectral image recognition method and device based on spatial spectrum group covariance features and a computer readable storage medium. The method comprises the following steps: acquiring a hyperspectral image to be identified, mapping the hyperspectral image to be identified to a Riemann space, and calculating the Riemann distance between data points in the Riemann space; and on the basis of a Riemann cutting space and a local linear measurement criterion, performing adaptive neighborhood calculation on each data point; projecting the data points in the Riemann space to a Riemann local tangent space according to the self-adaptive neighborhood information of each data point; obtaining the low-dimensional image features of the to-be-recognized hyperspectral image by performing eigenvalue decomposition on the reconstruction weight matrix obtained in the linear reconstruction process, and recognizing the surface features of the hyperspectral image based on the low-dimensional image features, so that the surface feature recognition accuracy of the hyperspectral image is effectively improved.

The invention discloses a coal-mine gas concentration measurement method. ''A great number'' is identified by adopting a KPCA (Kernel Principal Components Analysis) algorithm; two mixed kernel functions are constructed firstly; a kernel matrix is constructed by utilizing a vector method; and an eigenvector of the kernel matrix is calculated by utilizing KPCA. The algorithm has relatively high identification rate and relatively high calculation speed. According to the algorithm, a KPCA process on a training set is converted into a PCA process in which coordinates of all kernel training samples under a group of standard orthogonal bases are data sets through the group of the standard orthogonal bases of subspaces formed by the training samples in a characteristic space, and characteristics of the training samples are extracted, so that nonlinear characteristics of training data can be effectively captured; and therefore, the algorithm is widely interested and applied in mode identification and regressive analysis. In a solving process of the KPCA, an eigenvalue is required for decomposing an M*M kernel matrix (M represents the number of training samples), and only a kernel function between the samples and a group of formed Kidd samples needs to be calculated during sample characteristic extraction, and an experimental result verifies that the algorithm is valid.

The invention discloses a two-dimensional direction-finding estimation method based on polynomial rooting in a uniform area array. The method comprises the following steps: firstly, obtaining a covariance matrix from received signals of the uniform area array; carrying out eigenvalue decomposition on the covariance matrix to obtain a signal subspace and a noise subspace, and determining a rootingpolynomial according to the orthogonal relation between the direction matrix and the noise subspace; finally, solving the root of the polynomial, completing parameter pairing, and completing two-dimensional angle parameter estimation. According to the method, the complexity and the angle estimation performance can be fully balanced, and the limitation that in a traditional two-dimensional angle estimation method, the angle estimation performance is good but the complexity is high or the complexity is low but the angle estimation performance is common is broken through; moreover, high-resolution two-dimensional DOA estimation can be realized, the angle estimation performance is superior to that of a 2D-PM algorithm and a 2D-ESPRIT algorithm, the angle estimation performance is basically consistent with that of a 2D-MUSIC algorithm, and the algorithm complexity is far lower than that of the 2D-MUSIC algorithm.

The invention relates to eigenvalue decomposition precoding matrix index selection. A method and system for selecting precoding matrix index are herein disclosed. The method includes determining a precoder and candidate beams, selecting base beams based on a correlation power between the determined precoder and determined candidate beams, and estimating amplitude coefficients and cophase coefficients based on a correlation between the determined precoder and the selected base beams.

The invention discloses a face recognition method based on neighbor preserving canonical correlation analysis, and the method comprises the following steps: 1, inputting a face training data set X belonging to Rm*N, Y belonging to Rn*N, and calculating neighbor weight reconstruction matrixes Ux and Uy of an image through neighbor preserving learning; 2, searching two groups of projection vectors wx and wy by adopting canonical correlation analysis, introducing neighbor preserving into a canonical correlation analysis framework by adopting an optimization method, and calculating projection matrixes Wx and Wy by utilizing generalized eigenvalue decomposition; 3, performing low-dimensional projection and fusion on the test face image by adopting two feature fusion strategies; and 4, applyingthe fused features to face recognition by using a nearest neighbor classifier. According to the invention, a face proximity weight reconstruction matrix is learned through neighbor preserving; neighbor preserving is introduced into a typical correlation analysis framework by using an optimization method, and label information of human faces is utilized, so that the extracted human face features not only maximize the correlation between different human faces, but also maintain the neighborhood structure of the human faces as much as possible, and the human face recognition capability and stability are improved.

The invention discloses a local sensitivity discriminant analysis method for a defect thermal image of a composite material. The method comprises the following steps: 1) acquiring a defect thermal imaging sequence set; 2) setting initial parameters and calculating a nearest neighbor distance; 3) constructing a nearest neighborhood graph to obtain a weight matrix; 4) calculating a Laplace operatorto construct an optimal target function; 5) decomposing the eigenvalue to obtain low-dimensional embedding; and 6) defect image reconstruction and evaluation. According to the method, the features areextracted from the simulation data, the local sensitivity discrimination model of the defect thermal image of the composite material is established, the training model is evaluated, the LSDT method better processes the information, uneven backgrounds, defect features and measurement noise can be separated to a large extent, and therefore defect recognition is better facilitated.

The invention discloses a two-dimensional direction-finding estimation method based on polynomial rooting in a co-prime area array, and the method comprises the following steps: firstly, obtaining a covariance matrix through a received signal of the co-prime area array; secondly, performing eigenvalue decomposition on the covariance matrix to obtain a signal subspace and a noise subspace, and rebuilding a spectral function by using the relationship between a co-prime area array and the FCPA; then, determining a rooting polynomial according to the orthogonal relation between the direction matrix and the noise subspace; finally, solving the root of the polynomial, finishing parameter pairing, and finishing two-dimensional angle parameter estimation. The invention has the advantages that the complexity and the angle estimation performance can be fully balanced, and the limitation that in a traditional two-dimensional angle estimation method, the angle estimation performance is good but the complexity is high or the calculation complexity is low but the angle estimation performance is poor is broken through. According to the algorithm, high-resolution two-dimensional DOA estimation can be realized, and the problem of influence of initial estimation precision on subsequent estimation is effectively solved.

The invention belongs to the technical field of quantum machine learning, and discloses a selectable accurate quantum principal component analysis method and application, wherein based on a quantum singular value threshold decomposition algorithm, compared with traditional quantum principal component analysis, the flexibility is high, main components, namely main characteristic value characteristic vectors, are output by controlling thresholds, and all components, namely all eigenvalue eigenvectors, can be output, and compared with the previous improved algorithm, the algorithm reduces the number of quantum gates in the parallel direction, and the result is more accurate. According to the invention, the accurate quantum principal component analysis algorithm selected by the invention mainly comprises seven steps of inputting a covariance matrix quantum state, extracting a characteristic value through phase estimation, converting the characteristic value, performing controlled overturning, performing inverse transformation, measuring, extracting and screening the characteristic value through phase estimation, and finally outputting the characteristic value greater than a given threshold value and a corresponding characteristic vector; and the method can be used as a subroutine of other algorithms in the field of quantum machine learning, and the execution efficiency of the wholealgorithm is improved.

The invention discloses a thin-wall section characteristic deformation identification method, which comprises the following steps of: 1) constructing a thin-wall structure high-order model by considering section deformation, and deriving a control differential equation of a thin-wall structure by adopting a Hamiltonian principle; 2) solving the generalized eigenvalue of the control differential equation and a corresponding eigenvector by using a finite element method; 3) performing order reduction approximation processing on the feature matrix to identify the axial change mode of the basis function; and 4) orthogonally decomposing the feature vector into a component which is collinear with the axial change mode of the primary function to obtain a proportional relation between the componentand the axial change mode, and multiplying the original deformation mode by the corresponding proportionality coefficient to generate a new deformation mode. According to the method, numerical valueimplementation can be carried out in a simple and visual mode, the derived deformation mode has definite hierarchy and physical interpretability, the dynamic behavior of the thin-wall structure can betruly reflected, the calculation efficiency is greatly improved, and the calculation cost is reduced.

The invention discloses a method for recognizing chaotic time series based on the random matrix theory. The method comprises the specific steps of A, randomly selecting a time series to be recognized, and constructing original matrix X according to the time series; B, normalizing the matrix X, converting it into a non-Hermite matrix, shown in the description, having a row vector average of 0 and a variance of 1; C, calculating the non-Hermite matrix to obtain its singular value equivalent matrix X, and calculating the singular value equivalent matrix X to obtain a standard matrix product shown in the description; D, calculating according to the standard matrix product to obtain its characteristic values, and calculating an average spectral radius of the characteristic values; E, recognizing the time series as a noise time series or a chaotic time series according to the single ring law in the random matrix theory. Short chaotic time series or noise-containing chaotic time series can be recognized through the method, and subsequent research is thus facilitated; the method has a simple processing procedure and is good in universality and robustness.