114 results about "Positive-definite matrix" patented technology

A method constructs a classifier from training data and detects moving objects in test data using the trained classifier. High-level features are generated from low-level features extracted from training data. The high level features are positive definite matrices on an analytical manifold. A subset of the high-level features is selected, and an intrinsic mean matrix is determined. Each high-level feature is mapped to a feature vector on a tangent space of the analytical manifold using the intrinsic mean matrix. An untrained classifier is trained with the feature vectors to obtain a trained classifier. Test high-level features are similarly generated from test low-level features. The test high-level features are classified using the trained classifier to detect moving objects in the test data.

A memory-centric neural network system and operating method thereof includes: a processing unit; semiconductor memory devices coupled to the processing unit, the semiconductor memory devices contain instructions executed by the processing unit; weight matrixes including a positive weight matrix and a negative weight matrix constructed with rows and columns of memory cells, inputs of the memory cells of a same row are connected to one of Axons, outputs of the memory cells of a same column are connected to one of Neurons; timestamp registers registering timestamps of the Axons and the Neurons; and a lookup table containing adjusting values indexed in accordance with the timestamps, the processing unit updates the weight matrixes in accordance with the adjusting values.

A system includes a rotary device, a rotary absolute position (RAP) sensor generating encoded pairs of voltage signals describing positional data of the rotary device, a host machine, and an algorithm. The algorithm calculates calibration parameters usable to determine an absolute position of the rotary device using the encoded pairs, and is adapted for linearly-mapping an ellipse defined by the encoded pairs to thereby calculate the calibration parameters. A method of calibrating the RAP sensor includes measuring the rotary position as encoded pairs of voltage signals, linearly-mapping an ellipse defined by the encoded pairs to thereby calculate the calibration parameters, and calculating an absolute position of the rotary device using the calibration parameters. The calibration parameters include a positive definite matrix (A) and a center point (q) of the ellipse. The voltage signals may include an encoded sine and cosine of a rotary angle of the rotary device.

The invention provides a satellite lithium ion battery residual life prediction system and a satellite lithium ion battery residual life prediction method based on RVM (relevance vector machine) dynamic reconfiguration, and relates to a lithium ion battery residual life prediction system and a lithium ion battery residual life prediction method. The uncertainty expression of the lithium ion battery predication is realized, and the lithium ion battery residual life prediction method is more applicable to satellite system environment with limited resources. A dynamic reconfiguration module of the prediction system comprises a reconfiguration unit A and a reconfiguration unit B, the reconfiguration unit A and the reconfiguration unit B realize the time sharing multiplex of logic resources of the dynamic reconfiguration module, and the RVM training and predication is realized; and the Gaussian kernel function flowing water calculation is realized by a multistage flowing water segmented linear proximity method and a parallel computing structure, and the computational efficiency is enabled to be fully improved. The inverse calculation of symmetric positive definite matrices is realized by a Cholesky decomposition method, the computing resources consumption is reduced by a multiplying and gradually decreasing device, and the computing delay is reduced. Experiments show that the system and the method have the advantages that FPGA (field programmable gate array) finite computing resources are utilized for realizing the computational accuracy similar to a PC (personal computer) platform, the four-times computing efficiency improvement relative to the PC platform is obtained, and the utilization rate of hardware resources is effectively improved through dynamic reconfiguration strategies.

The invention discloses a convolutional neural network-based human hand image region detection method. The method comprises the following steps of: carrying out feature extraction on an image by utilizing a convolutional neural network, and training a weak classifier; for the image, angles of which are marked, segmenting the image on the basis of the classifier to obtain a plurality of candidate regions; modeling each candidate region by utilizing the convolutional neural network so as to obtain an angle estimation model, and carrying out angle marking to rotate the candidate area to a positive definite attitude; modeling each candidate region again by utilizing the convolutional neural network so as to obtain a classification model; for a test image, firstly segmenting the test image by using the weak classifier so as to obtain candidate regions, and for each candidate region, estimating angle of the candidate region through the angle estimation model and rotating the candidate region to a positive definite attitude; and inputting the candidate region under the positive definite attitude into the classification model to obtain a position and an angle of a human hand in the image. According to the method, the classification precision is improved by adopting convolutional neural network-based coding classification, and by utilizing the angle model, the method has rotation variance and very high human hand region detection precision.

The invention provides an RSSI-based indoor positioning method and system. RSSI signals received by at least three positioning base stations are respectively acquired; Gaussian fitting processing is performed on the acquired RSSI signals, high probability RSSI signals meeting the preset probability threshold in the acquired RSSI signals are reserved and effective RSSI signals are acquired; wavelet transformation is performed on the effective RSSI signals, and the denoised effective RSSI signals are acquired; and indoor positioning is performed by using a Taylor algorithm improved by a positive definite correction matrix according to the denoised effective RSSI signals so as to acquire the positioning result. Data screening and denoising are performed by the Gaussian fitting and wavelet transformation processing in the whole process, the change of the measured RSSI value caused by movement of large equipment and personnel in the indoor scene is considered and the positive definite matrix is introduced to correct the error of estimation offset so as to enhance the positioning accuracy.

The invention discloses a positive definite matrix floating point inversion device based on an FPGA and an inversion method thereof. The inversion device comprises a process control module, an operation module and a storage module. The process control module is respectively connected with the operation module and the storage module. The operation module is connected with the storage module. The process control module is used for generating a signal to control the orderly operation of the operation module and the storage module. The operation module is used for carrying out the matrix operation. The storage module is used for caching data of a matrix to be operated and data of a result matrix and providing a system bus access interface. The positive definite matrix floating point inversion device can be conveniently and rapidly achieved on a chip of the FPGA, and on the premise of guaranteeing the precision, the matrix inversion speed is improved.

The invention discloses a data processing device. The data processing device is used for solving a lower triangular matrix L corresponding to an n*n symmetric positive definite matrix R, wherein the R is equal to LDLH, the D is a diagonal matrix, the LH is a conjugate transpose matrix of the L, and the n is an integer larger than or equal to 2. The data processing device comprises a multiplying unit, an accumulator and a summator. The input end of the data processing device is connected with the input end of the multiplying unit, the output end of the multiplying unit is connected with the input end of the accumulator, the output end of the accumulator is connected with the output end of the summator, the input end of the data processing device is also connected with the input end of the summator, the output end of the summator is used for outputting matrix data to a storage device, and the output end of the multiplying unit is also used for outputting matrix data to the storage device. By means of the data processing device, the symmetric positive definite matrix can be decomposed without a large amount of root extraction operation, operation complexity is lowered, and logical operation resources are saved.

The invention relates to an arrival direction self-correction method for a sensor array. According to the method, a noise subspace estimated value is determined; an array phase error of the sensor array is initialized, and an initial value of the signal arrival direction is estimated through utilizing an MUSIC algorithm; based on the calculation value of the signal arrival direction, a Hermitian positive definite matrix is calculated, and the Hermitian positive definite matrix satisfies a first equation; an array element phase error is made to satisfy the first limit condition, and the array phase error is converted to satisfy a second equation; the first equation and the second equation are solved through utilizing a Lagrangian multiplier method to acquire a vector estimated value of the array phase error; a correction value of the signal arrival direction is acquired through utilizing the MUSIC algorithm or a high resolution subspace algorithm; and iteration of the previous steps is repeatedly carried out till iteration stop conditions are satisfied. The method is advantaged in that dependence on special structure characteristics of an array output covariance matrix is avoided, excellent performance is realized under the condition of limited sampling data, and influence of the sensor array error on the arrival direction can be effectively reduced.

The invention discloses a human body behavior recognition method based on kernel sparse coding, and belongs to the technical field of digital image processing. The method comprises the steps: firstly dividing an inputted video into video segments which have a fixed length and are mutually overlapped; secondly extracting the gradient and a light stream characteristic covariance or shape characteristic covariance of each video segment; and thirdly carrying out the dimension reduction of a covariance matrix through employing a symmetric positive definite matrix dimension reduction method. On the basis of the Stein kernel, the method proposes a sparse maximization symmetric positive definite matrix dictionary learning, and proposes a Riemann sparse solver which enables Riemann manifold to be embedded into a kernel Hilbert space. The method is used for the recognition of human body behaviors in a video, is simple in processing, is low in calculation complexity, and is robust for behavior difference, view change and low resolution.

The invention discloses a medical image segmentation method and a medical image segmentation system for fully-represented semi-supervised fast spectral clustering. The method comprises the following steps: obtaining a to-be-treated medical image; circling in the medical image by a touch screen; extracting grey level of pixel, a spatial position and Gabor textural characteristics of the whole medical image, and carrying out characteristic normalization and characteristic dimension reduction treatment; carrying out All-In-One form representation onto reference information of a circled area; generating a graph-based relaxed clustering model based on a full-presenting semi-supervising mechanism; rearranging a quadratic term of the clustering model into a novel positive definite matrix; rewriting as a constraint type minimal enclosing ball form; estimating a final solution based on a rapid approximation strategy of the core set minimal enclosing ball; determining actual category number of clustering segmentation by a graph-based clustering indication vector; and dividing clustering indication components into different subsets based on a K means algorithm according to the category number. The system comprises an FPGA (Field Programmable Gate Array) module and an external device. The method and the system disclosed by the invention are simple to operate, good in real-time performance and high in accuracy.

The invention relates to a method for evaluating the comprehensive performance of numerical control machine tools based on an improved pull-apart grade method, belonging to the technical field of performance evaluation of the numerical control machine tool. The linear proportional method is used to standardize the performance data of machine tools. The subjective weights of each index are determined by order relation analysis, and the objective weights of each index are determined by entropy weight method and mean square deviation method. The subjective and objective weights are synthesized based on the principle of vector Pearson correlation coefficient. The final weight coefficients of the indexes are determined by the positive definite matrix corresponding to the weighted index matrices, and the comprehensive evaluation values of the third-level indexes are obtained based on the linear weighted evaluation function. Finally, a similar method is used to calculate the comprehensive evaluation value of the large-scale system layer by layer. The invention is used for comprehensive performance evaluation of various numerical control machine tools and lateral comparison of specific performance of different machine tools, and provides a scientific and operable evaluation method and flow for comprehensive performance evaluation of machine tools.

The invention belongs to the field of integrated circuits, relating to a method for establishing a space correlation model of technical errors in an integrated circuit chip. The method comprises the following steps of: extracting an unknown parameter of a space correlation function by utilizing the maximum likelihood estimation of a multi-test chip, and establishing the space correlation model ofthe errors in the chip; and multiplying a likelihood function of all test chips to obtain a joint likelihood function, solving through maximizing the joint likelihood function to obtain the space correlation function which is determined by a parameter value and can be directly used for circuit analysis design of the technical errors. The influence of pure random part of the error in the chip and the measurement error can be processed in the extraction process of the space correlation function, and the precision of an extracted result is remarkably improved. Determinant logarithms of symmetrical positive definite matrixes in the joint likelihood function are calculated by utilizing an LU (Logic Unit), and the problem of unstable number value occurring in direct calculation is solved.

Method and system for reducing a number of eigenvectors. For example, a computer-implemented method for reducing a number of eigenvectors, the method comprising: obtaining a plurality of to-be-processed matrices; mapping the plurality of to-be-processed matrices to a space of symmetric positive definite matrices to form a Riemannian manifold corresponding to a Riemannian kernel function; obtaining a kernel-function matrix by using at least a principal component analysis to calculate one or more inner products of the mapped plurality of matrices based on at least the Riemannian kernel function; calculating a first group of eigenvectors of the kernel-function matrix, the first group of eigenvectors including a first number of eigenvectors; and selecting one or more eigenvectors from the first group of eigenvectors to obtain a second group of eigenvectors, the second group of eigenvectors including a second number of eigenvectors; wherein the second number is less than the first number.

The invention puts forward an all Bregman divergence-based matrix constant false alarm rate detection method. A technical solution of the method comprises the following steps: a radar echo signal in each range cell is modeled into a Hermite positive definite matrix, all Bregman divergence distance between mid-values of the matrix in each range cell and corresponding matrixes of neighboring cells thereof is calculated, and whether an object exists is determined after the all Bregman divergence distance is compared with a set threshold value. The method disclosed in the invention is simple in principles and small in calculation amount, the method is not sensitive to noise particularly in conditions of large background clutters, and the method is high in detection performance.

The invention discloses an improved Latin hypercube sampling method suitable for non-positive correlation control. The method comprises the steps of S1, obtaining the cumulative distribution function of input variables or a large amount of measured discrete data and a correlation coefficient matrix among the variables; S2, extracting samples using the importance sampling method when the distribution function is known and using the improved Latin hypercube sampling method when the distribution function is unknown, and obtaining a sample matrix; S3, modifying the correlation coefficient matrix based on the positive definite spectral decomposition method; S4, conducting the Cholesky decomposition and correlation transformation on the correlation coefficient matrix of a modified random order matrix; S5, computing an order matrix through the specified correlation coefficient matrix, and determining the final sample matrix according to the sorting of the order matrix; S6, calculating load flow after the sample matrix is brought into nodes, obtaining node voltage and branch power, and calculating the relative error index. According to the improved Latin hypercube sampling method, the problems that the distribution function of the input variables is unknown and the non-positive definite matrix cannot be decomposed are solved, and the application scope of the Latin hypercube sampling method is expanded.

The invention belongs to the field of signal detection, in particular to a radar target detection method based on full KL divergence. The main steps are that the radar echo data in each range unit ismodeled as a Hermite positive definite matrix, the full KL divergence distance between the values of the matrix of each range unit and the corresponding matrixes of the surrounding units is calculatedso as to obtain a one-dimensional range profile; and the magnitude between the amplitude value corresponding to each range unit and the detection threshold is compared so as to determine existence ofthe target. The sample data in each range unit is modeled as the Hermite positive definite matrix, and the magnitude of the elements in the matrix represents the correlation intensity between the sample data and the Doppler information of the sample data. The matrix model can avoid the loss of detection performance caused by the spectrum leakage of Fast Fourier Transform. The radar target detection method based on full KL divergence has low computational complexity, simple detection principle and good detection performance.

The invention discloses a complex field blind source separation method. A complex filed target matrix system is built, and real symmetrization is carried out to obtain a reconstructed target matrix system formed by a real-value target matrix, the complex field combined diagonalization problem is converted into the real field combined diagonalization problem to solve the complex field blind source separation problem; compared with other algorithms that are also suitable for the complex filed, the method doesn't restrain a diagonalization target matrix to be combined into a hermitian symmetric matrix or a positive definite hermitian matrix and is wide in application; an alternative least square iterative algorithm based on combined diagonalization least square cost functions is adopted, and the structural characteristics of the target matrix system formed by the real-value target matrix are fully used to realize the combined diagonalization of a new target matrix system; The cost functions are solved by the alternative least square iterative algorithm, the estimate values of a mixed matrix are obtained, the blind source separation is realized, and the simulation results verify that the method provided is high in convergence precision.

The invention discloses a dimension reduction processing method of high-dimensional vibration signals. According to the method, a neighbor matrix of the signals is obtained by calculating the Euclidean distance between the signals, a reconfiguration weight matrix of the signals is obtained according to the neighbor matrix of the signals and by utilizing sparse constraint conditions, finally the vibration signals after dimension reduction are obtained by utilizing the reconfiguration weight matrix of the signals, and the dimension reduction process is simple. In the process of obtaining the reconfiguration weight matrix of the signals by utilizing the sparse constraint conditions, L1 norms are introduced into the sparse constraint conditions, so that the reconfiguration weight matrix has good sparseness. Accordingly, influence of noise points is effectively eliminated, anti-noise capacity is improved, and robustness of the method is ensured. Final obtaining of the vibration signals after dimension reduction is to solve a feature vector of the sparse, symmetrical and semi-positive definite matrix, and therefore calculation complexity of the method can be lowered.

The invention belongs to the intelligent transportation field, provides a vehicle distributed controller gain solving method and device of a homogeneous vehicle queue and solves a distributed controller gain through relatively small calculation amount. The method comprises steps that a vehicle following error convergence coefficient delta of the homogeneous vehicle queue is determined; a symmetric positive definite matrix P for making inequality equations described as the specifications set up is solved; according to a formula described as the specifications, the vehicle distributed controller gain of the homogeneous vehicle queue is calculated. The method is advantaged in that on one hand, excellent decoupling of design of the vehicle distributed controller gain K and a specific communication topology structure of the homogeneous vehicle queue is realized, and excellent universality is realized, on the other hand, the calculation amount of the formula described as the specifications and the inequality equations described as the specifications is small, and the method is quite suitable for engineering practices.

The invention discloses a multi-target positioning outer approximation nearly convex optimization algorithm on the basis of arrival time. The multi-target positioning outer approximation nearly convex optimization algorithm includes building original problem models; relaxing original problems by the aid of relaxation methods to obtain a mixed integer convex optimization problem model; constructing novel appropriate constraints for semi-positive definite matrixes; solving coordinate values of various to-be-positioned targets by the aid of outer approximation nearness algorithm sub-models, continuous convex optimization problem models and OAA (open application architecture) algorithms. The multi-target positioning outer approximation nearly convex optimization algorithm has the advantages that complicated requirements of existing methods on layout of base stations and regions where the targets are located can be omitted by outer nearness approximation algorithms, and the global optimal solution can be assuredly converged without initial estimation points.

The invention provides a fuzzy control method based on a chaotic system. According to the method, a piecewise Lyapunov function method is adopted for design first, namely a Lyapunov function is selected for each subsystem, and an aggregate Lyapunov function of a system is a weighted sum of the functions of all the subsystems. Difficulty in solving a common positive definite matrix in a parallel distribution compensation (PDC) method is avoided, conservatism of the design is further reduced, and the range of the solution is widened. The design of a fuzzy observer solves the problem that the state of the system can not be measured directly. Through simulation of the chaotic system, effectiveness of the method is testified. The scheme is also suitable for nonlinear system control of other types, can be used for reference and has practical application value.

A low-voltage active distribution network congestion management method based on node marginal price, Firstly, according to the basic principle of positive semidefinite programming algorithm, The nonlinear constraint in the low voltage active distribution network congestion management model based on intelligent soft switching is transformed into the positive semi-definite constraint by the positivesemi-definite relaxation technique, The original non-convex nonlinear model is transformed into a positive semi-definite programming model, and the symmetric component method is used to reduce the mutual impedance coupling between phases, so that the numerical stability of positive semi-definite relaxation can be improved by increasing the numerical difference of the elements in the positive semi-definite matrix. Then the sensitivity factors of branch active power, branch reactive power, node voltage variation and network loss to node injection power are calculated based on the linearizationapproximation principle. Based on the sensitivity factor, a node marginal pricing model is established. The invention can effectively reduce the difficulty of problem solving and ensure the calculation accuracy while improving the calculation speed. It can effectively solve the congestion management problem of low-voltage active distribution network.

The present invention discloses an FPGA implementation method for Cholesky decomposition of a positive definite matrix. The method is mainly provided with: a top layer control module, for communication and control between modules; a data preprocessing module, for decomposing a positive definite matrix into two matrices for calculation operations in a matrix calculation module; and the matrix calculation module, for calculating two matrices obtained by the data preprocessing module to obtain a final Cholesky decomposition calculation result. Beneficial effects of the method disclosed by the present invention are that: by using the traditional hardware to directly implement Cholesky decomposition of the positive definite matrix, the algorithm is complex, the occupied area is large, and the resource consumption is more, and by using the rotation characteristic of the CORDIC algorithm to implement the Cholesky decomposition of the positive definite matrix, the implementation is simple, only the bit operation is needed, the resource consumption is less, and the computational complexity and the area of the gate circuit are effectively reduced.

The invention discloses an interference alignment method for minimizing interference power and dimension. The method comprises the steps: 1, randomly generating an interference inhibition matrix, and setting an iteration number iter; 2, taking the nuclear norm of an interference mapping matrix as a target function, letting a desired signal matrix be a Hermitian positive definite matrix, and taking a condition that a minimum characteristic value is greater than or equal to 100 as a constraint condition, and solving a pre-coding matrix; 3, taking a received interference signal as the whole body, and solving an interference covariance matrix; 4, letting a characteristic vector corresponding to the minimum characteristic value of the interference covariance matrix be a column vector of the interference inhibition matrix; 5, judging whether all iterations are completed or not: carrying out the orthogonalization of the pre-coding matrix and the interference inhibition matrix if all iterations are completed, or else, returning to step 2. The method can improve the quality of received signals, enlarges the system capacity, and improves the utilization rate of frequency spectrum.

The invention discloses a multi-target positioning external impending approximation convex optimization algorithm based on a time difference of arrival. According to the algorithm, first, an original question model is constructed; next, a relaxation method is used to relax the original question model into a convex question to obtain a first external impending approximation algorithm submodel; new proper constraint is constructed according to a positive semi-definite matrix, and then the first external impending approximation algorithm submodel is solved to obtain an integer solution; the obtained integer solution is substituted into a second external impending approximation algorithm submodel, initial rough coordinate values of all targets to be positioned are solved, and the coordinate values are used as starting points to calculate precise positions of all the targets; and then whether an optimal solution is obtained is judged. Through the algorithm, the external impending approximation algorithm is utilized, complicated requirements of existing methods on a layout of a base station and a region where a to-be-detected target is located are avoided, it can be ensured that a global optimal solution is obtained through convergence, and an initial estimation point is not needed.