32 results about "Kronecker product" patented technology

The invention relates to the technical field of artificial intelligence, and discloses a large model construction method and device, equipment and a medium, and the method comprises the steps: carrying out the partitioning of a Transform weight matrix of a target model, so as to form a plurality of sub-matrixes; performing low-rank approximation on each sub-matrix through two-dimensional Kronecker decomposition to obtain a Kronecker product of the two small matrixes respectively corresponding to each sub-matrix; parameters of all the small matrixes are frozen, and additional parameters corresponding to all the small matrixes are obtained by using LoRA, so that a two-dimensional Kronecker-LoRA compression model is obtained; and taking the target model as a teacher model, and training the two-dimensional Kronecker-LoRA compression model by using the data set to obtain a large model. According to the method, the parameter scale and the resource demand of an existing large model can be reduced, an applicable large model is formed, meanwhile, the performance loss is effectively reduced, the balance between the high compression ratio and the high precision is achieved, dynamic reasoning and the image processing flow are deeply matched, and wide deployment on edge equipment is easy.

The application discloses a scanning radar two-dimensional super-resolution imaging method based on Bayesian sparse learning, first constructs a bearing-pitch two-dimensional scanning radar signal model, then, according to a maximum posteriori criterion under a Bayesian framework, establishes a sparse optimization target function about target scattering and environmental noise, finally, utilizes a conjugate gradient algorithm and Kronecker product properties to accelerate iterative estimation of target scattering and noise power, realizes adaptive sparse two-dimensional super-resolution imaging of the scanning radar. The method solves the problems of high complexity and poor noise robustness of prior art means, compared with prior art two-dimensional super-resolution methods, has lower calculation complexity, is more robust, has excellent noise adaptive capacity, and can realize two-dimensional scanning radar super-resolution imaging under a low signal-to-noise ratio condition.

The application provides a cylindrical array sonar system and a target detection method based on high-resolution beam forming. The method comprises the following steps: converting a beam forming weight vector of a received sound wave signal into a Kronecker product, combining a joint probability density model to construct a negative log-likelihood function, and optimizing a vector in an azimuth dimension and an elevation dimension through a preset direction constraint condition to form a high-resolution beam; performing correlation analysis on echo signals reflected by a target at different azimuth angles and elevation angles and on transmitted sound wave signals at different time delays; constructing a three-dimensional matrix containing a distance, an azimuth angle and an elevation angle according to a correlation analysis result, performing normalization processing on the three-dimensional matrix to eliminate signal intensity differences, and outputting a target detection positioning result. Through the combination of Kronecker integral decomposition and three-dimensional matrix modeling, the application realizes high-resolution joint positioning detection of a target in an azimuth angle, an elevation angle and a distance dimension, effectively eliminates signal intensity differences, and improves positioning accuracy.

The application relates to the technical field of time series prediction, in particular to a data center cooling system return water temperature prediction method and device, wherein the method comprises the following steps: collecting initial operation data of a cooling system in a data center; pre-processing the initial operation data to obtain final operation data meeting certain conditions; inputting the final operation data into a pre-trained time series prediction model to obtain a cooling system return water temperature prediction result, wherein the time series prediction model has a Kronecker product layer, a kernel matrix layer and a multilayer perception layer. Thus, the problems that, in the related art, model parameters are large, blackboxing is caused by expressing nonlinear characteristics through a complex activation function, it is difficult to directly express cross-variable relationships, the effectiveness and interpretability of the model in actual application are limited, and the understanding and trust of a user for a model prediction result are affected are solved.

In accordance with an example aspect of the present disclosure, data is encoded or decoded using a polar code, where a generation matrix of the polar code includes rows of a polar transformation matrix formula (I), where is an N * N-bit inversion matrix, where N = 2n is a non-negative integer, and where formula (II) is an n-fold Kronecker product of formula (III), where N = 2n is a non-negative integer. The index set of the rows forming the polarization transformation matrix of the generating matrix is set A.

The application discloses a bidirectional discriminative feature alignment hierarchical knowledge fusion method, comprising the following steps: inputting samples into a teacher model to obtain a teacher soft prediction result set, and inputting unlabeled image data into an initial student model to obtain a student model prediction result; extracting the last layer feature and inputting the last layer feature into a common feature extractor to obtain a common feature; through a discriminative centroid clustering strategy, different class centers are far away from each other, and each teacher common feature is close to the same class center; the fuzziness of the teacher soft prediction result is measured by entropy impurity, reliable source domain features and target domain features are constructed, pseudo labels are respectively subjected to Kronecker product with the common features of the source domain and the common features of the target domain for discriminative mapping, and feature alignment is performed through a maximum average difference method; a total loss function is constructed, and the initial student model is trained through the total loss function to obtain a comprehensive student model capable of accurate classification.

The invention provides a memristor-based fixed-point matrix FFT (Fast Fourier Transform) calculation method, which belongs to the technical field of signal processing and image processing, and comprises the following steps of: firstly, constructing a non-linear voltage controlled memristor model, simulating a state conversion behavior in logic operation through a dynamic regulation mechanism, and realizing the function of a basic logic gate; according to the method, an FFT algorithm based on matrix block optimization is designed, FFT calculation is decomposed into small-scale matrix operation units, calculation efficiency is optimized in combination with a Kronecker product and an arrangement matrix, FFT operation with higher point number is expanded in a recursion mode, meanwhile, a full adder and a multiplier are designed based on a memristor, efficient fixed-point number operation is achieved, and finally, the FFT calculation efficiency is improved. The method is applied to image FFT processing, and a processing result is obtained. The scheme of the invention has higher calculation efficiency and lower energy consumption while ensuring the calculation precision, can be widely applied to the fields of image processing, signal analysis, scientific calculation and the like, and has important practical value and application prospect.

This invention relates to the field of traffic safety technology, specifically to an efficient simulation method for quantum systems based on parallel order reduction. The method includes the following steps: receiving the physical parameters of the quantum system and simulation requirements, whereby the physical parameters include electron mass, potential well size, and initial wave packet parameters; and the simulation requirements include simulation physical time and accuracy requirements. Based on the physical parameters, a time-dependent Schrödinger equation describing the dynamic behavior of the quantum system is established, and the Schrödinger equation is rearranged into matrix form using the Kronecker product. This invention utilizes Arnoldi to reduce the order of the matrix-form Schrödinger equation. By reducing the order, the large matrix in the original space is projected into a relatively small subspace, improving the efficiency of the simulation solution. The reduced-order Schrödinger equation is solved using a time-parallel algorithm. This approach overcomes the time step limitation imposed by the CFL condition, ensuring that a stable solution can be obtained with fewer time steps.

The application relates to the technical field of human-computer interaction, in particular to a method and device for predicting an embodied intelligent human-computer interaction system based on motion recognition, wherein the method comprises the following steps: collecting three-dimensional coordinates of a plurality of skeleton key points of a human body and time frames corresponding to the coordinates of each skeleton joint to generate initial skeleton data; preprocessing the skeleton data to obtain final skeleton data meeting a condition; performing skeleton coding on the final skeleton data to extract at least one skeleton feature; and training a prediction model based on the at least one skeleton feature, wherein the prediction model comprises a Kronecker product layer, a Kernel Matrix layer and a multilayer perception machine layer to identify a future motion trajectory. Thus, the problem that, in the prior art, due to a large model parameter quantity, strong data dependency and difficulty in intuitively displaying the relationship between variables, the effectiveness and interpretability of the model in actual application are limited, and the trust of a user in an identification result is affected, is solved.

The application relates to a design method of a robust vector array differential beamformer, belonging to the field of sound signal processing, which comprises the following steps: setting system working parameters; transmitting a detection signal; judging whether a target signal is detected; receiving and preprocessing array signals; reconstructing full-array steering vectors by using Kronecker product operation; selecting an ideal differential beam pattern form, setting target direction distortionless response constraints and non-target direction response constraints; decoupling the system total weight vector into the Kronecker product form of spatial weight and channel weight by using the Kronecker product property, taking the white noise gain maximization as an objective criterion, and solving the optimal channel weight; constructing a weight approximation model based on the Maclaurin series expansion, solving the optimal spatial weight by using the least square method, synthesizing the system total weight vector, and realizing the design of a differential beamformer of any order. The application can significantly improve the robustness of the differential beamformer and effectively suppress spatial white noise.

The application relates to the technical field of artificial intelligence, and discloses a large model construction method, device, equipment and medium, the method comprising the following steps: performing block processing on a Transformer weight matrix of a target model to form a plurality of sub-matrices; performing low-rank approximation on each sub-matrix through two-dimensional Kronecker decomposition to obtain the Kronecker product of two small matrices corresponding to each sub-matrix; freezing the parameters of all the small matrices, using LoRA to obtain additional parameters corresponding to all the small matrices to obtain a two-dimensional Kronecker-LoRA compression model; taking the target model as a teacher model, and using a data set to train the two-dimensional Kronecker-LoRA compression model to obtain a large model. The method can reduce the parameter size and resource demand of an existing large model, form a suitable large model, effectively reduce performance loss, achieve a balance between high compression rate and high precision, adapt the dynamic reasoning and image processing process depth, and facilitate wide deployment on edge devices.

This invention discloses a method and system for smoothing quantization of large models based on learnable rotation matrix Kronecker integral solutions, relating to the field of artificial intelligence model compression technology. The method includes: constructing a search space containing Kronecker integral solutions of different numbers of matrices, with the tensor-matrix i-th modal product as the computational implementation method for each solution; adaptively selecting the optimal rotation matrix hybrid decomposition scheme for the attention modules and feedforward networks of each decoder layer of the large model through a preset scoring function and layer-by-layer greedy heuristic search; training the complete parameters of the decomposition matrix, learnable weights, activation value truncation threshold, and diagonal scaling matrix parameters of the scheme; fusing the extra parameter matrix on the weight side of the linear layer, retaining the extra parameter matrix on the activation value side, and quantizing the model based on the trained quantization parameters to generate a deployable model. This effectively solves the problems of accuracy loss and computational overhead under low-bit quantization of large models, significantly improving computational and storage efficiency while maintaining high accuracy.

The invention discloses a Bayesian neural network prior construction method based on feature tag data, and relates to the technical field of machine learning, and the method comprises the steps: carrying out the standardized preprocessing of input feature tag data, calculating a covariance matrix between features, and adding a regularization term to guarantee the numerical stability; constructing structured prior distribution of a first layer of weight of the network by using the covariance matrix, and fusing feature correlation information into a covariance structure of weight parameters through a Kronecker product; and on the basis, a complete Bayesian neural network probability model is established, posterior distribution approximation is performed by adopting a variational inference method, and finally a prediction result is obtained through sampling. According to the method, the adaptive prior distribution is constructed by using the covariance structure in the feature tag data, so that the complex relevance between the features can be effectively captured, and the structural information is fused into the regularization constraint of the model parameters.

The invention discloses a deep neural network-based dynamic load prediction method for lifting compartments of an overhead passenger ropeway. The method comprises the following steps of 1, collecting and preprocessing operation data of a plurality of lifting compartments in the overhead passenger ropeway; step 2, inputting the operation data into a Times Net network; 3, calculating a numerical difference value between adjacent time points of the initial load predicted value sequence, marking local anchor points, and constructing a cross-lift-compartment anchor point diagram; 4, calculating the disturbance intensity of the anchor point pair based on the two-dimensional lift car position index matrix, and generating a disturbance embedded tensor; 5, performing Kronecker product operation on the embedded representation tensor and the disturbance embedded tensor, and inputting the embedded representation tensor and the disturbance embedded tensor into a prediction head structure; and 6, calculating a numerical error and a trend error between the load predicted value sequence and the real load sequence, and updating the offset value. According to the method, through the TimesNet network and disturbance reconstruction, the prediction precision and robustness of the lift car load are improved.

Parameters of a probabilistic model comprising Gaussian processes are learned. A system (100) is disclosed for learning a set of parameters of a probabilistic model having layers of multiple Gaussian processes (e.g., deep GPs) from a training data set. The set of parameters includes at least induced locations for the multiple Gaussian processes, and parameters of a probability distribution approximating outputs of the multiple Gaussian processes at the multiple induced locations. The probability distribution comprises a multivariate normal probability distribution having a covariance matrix defined by a Kronecker product of a first matrix indicating similarities between the multiple Gaussian processes and a second matrix indicating similarities between the multiple induced locations. A system (200) is also disclosed for determining one or more samples of an output of the probabilistic model for a given input using the set of parameters, e.g., to determine a mean and / or uncertainty estimate for the probabilistic model.

This invention relates to a visual pose estimation method, system, laser device, and electronic device, comprising: acquiring two different features of a target object; acquiring a connected coordinate system of a transmitting unit and a receiving unit, wherein the transmitting unit emits laser beams to the receiving unit and is located on different features; acquiring an image of laser spots formed by multiple laser beams, and transforming the quantized representation of the laser spots in the image coordinate system into a representation equation in the connected coordinate system through homography transformation; establishing a relative pose calculation model between the connected coordinate systems, proposing an optimization objective, and transforming the optimization objective function into a relational expression of a column vector consisting of a corresponding matrix and the unknown quantity through the Kronecker product; transforming a nonlinear system of equations into a linear system of equations, and calculating the solution of the linear system of equations; calculating the gradient matrix of the objective function with respect to the desired spinor using mathematical tools such as Lie groups and spinors, and obtaining the optimal solution through Newton-Raphson iteration using the solution of the linear system of equations as the initial value.

An information transmission method and a communication device generate a preamble sequence or a physical uplink control channel (PUCCH) sequence using a product of K vectors (such as a Kronecker product), K being an integer greater than or equal to 2, and transmit the preamble according to the preamble sequence, or transmit uplink control information according to the PUCCH sequence. According to the embodiment of the invention, the number of generated preamble sequences or PUCCH sequences can be increased, so that the communication requirement of the system is met, and the communication efficiency is improved. For example, for a preamble sequence, in a random access process, the number of users accessed by using the preamble sequence can be increased. For the PUCCH sequence, the capability and reliability of transmitting the uplink control information by using the PUCCH sequence can be improved.

This application provides a communication method and apparatus that can reduce the computational complexity of detecting information carrying a first bit, improve detection efficiency, and enhance communication performance. The method includes: determining a first signal based on a first sequence corresponding to the first bit information, wherein the first sequence is constructed based on the Kronecker product of a second sequence and a third sequence, the second sequence carrying the first bit of the first bit information and the third sequence carrying the second bit of the first bit information; and transmitting the first signal.

The application discloses a kind of intelligent reflecting surface auxiliary cascading surface array channel nested tensor decomposition estimation methods, first constructs cascading channel model, initialization each end guiding matrix and channel matrix in turn;Zero intelligent reflecting surface phase shift is first placed, respectively constructs the tensor model of transceiver end parallel factor decomposition, combined with three linear alternating least square method and least square method, completes two-dimensional angle estimation of transceiver end;Again control phase shift, through two linear alternating least square method, Khatri-Rao decomposition, complete intelligent reflecting surface two-way angle estimation, finally rely on Kronecker product and Khatri-Rao product model solves beam gain.The application relies on nested parallel factor decomposition, realizes four angles automatic pairing, adapts single, multi-beam scene, reduces high-dimensional calculation amount, suppresses off-grid energy leakage, without signal autocorrelation matrix and complex angle pairing, adapts super large scale surface array, improves channel estimation applicability, supports sensing integration system design.

The invention provides a four-dimensional ultrasonic signal real-time enhancement method based on anisotropic multiphase decomposition and LTI constraint, and belongs to the field of medical image processing. The method comprises the following steps: constructing a space-time Laplacian pyramid and anisotropic multiphase decomposition, constructing a space-time multiphase filtering network based on dimension decoupling, performing LTI constraint solution based on transposer error minimization, performing cross-scale multiphase feature alignment based on Kronecker product, performing multiphase synthesis inverse rearrangement, and performing image reconstruction. The behavior of the system is described by using the pulse response in the signal processing theory, the shift characteristic introduced by multiphase decomposition is eliminated, the LTI strong constraint is introduced in the filter coefficient optimization stage, and the constraint forces convolution kernel parameters in the space-time multiphase filter network to form a specific complementary structure, so that the performance of the system is improved. According to the invention, different multi-phase components can counteract aliasing terms during synthesis, so that checkerboard artifacts are eliminated macroscopically, low-frequency structure direction information is ensured to accurately cover each high-frequency sub-pixel phase, and no phase deviation exists.

The invention discloses a time-varying coupling unmanned system multi-formation cooperative control method for communication constraints. Firstly, a discrete time dynamic model and a controller input model of the networked unmanned multi-formation system are established, and formation division and formation reference trajectory definition are carried out at the same time. Secondly, under the condition that the communication bandwidth is limited, a synchronization error dynamic model is constructed, a hierarchical bit rate constraint relation of a system layer, a formation layer and a node layer is established, a uniform quantization coding-decoding mechanism is introduced to quantize a node state, and augmented modeling of an error system is realized by using a Kronecker product. And then, designing a multi-formation cooperative control strategy with fault-tolerant capability, solving controller gain through a linear matrix inequality, and ensuring bounded convergence of synchronization errors. Furthermore, a dynamic bit rate allocation mechanism based on a phototropic growth optimization algorithm is designed, and optimal configuration of communication resources is realized, so that the collaborative consistency and robustness of a multi-formation system under communication limited and complex working conditions are improved.

An information transmission method and a communication apparatus are disclosed. A preamble sequence or a physical uplink control channel (PUCCH) sequence is generated by using a product (for example, a Kronecker product) of K vectors, where K is an integer greater than or equal to 2, and a preamble is sent based on the preamble sequence, or uplink control information is sent based on the PUCCH sequence. A quantity of generated preamble sequences or PUCCH sequences can be increased, to satisfy a communication requirement of a system, and improve communication efficiency. For example, for the preamble sequence, a quantity of users using the preamble sequence for access can be increased in a random access process. For the PUCCH sequence, a capability and reliability of transmitting the uplink control information through the PUCCH sequence.

A transmitter device for employing tensor-based multiple access is disclosed. The transmitter device includes a user group index extractor module which extracts from the input sequence of bits a subsequence of bits corresponding to a user group index. The transmitter device further includes a forward error correction encoder which receives an input of a remaining sequence of bits after the subsequence of bits corresponding to the user group index has been extracted and generates a coded sequence of bits. Further, the transmitter device includes a splitter which receives output of the forward error correction encoder. The transmitter device further includes a plurality of subconstellation vector mapper units and a quasi-orthogonal subconstellation vector mapper unit. The transmitter device further includes a rank one tensor unit which generates a symbols vector by calculating the Kronecker product of the plurality of modulated subconstellation symbol vectors and the quasi- orthogonal subconstellation symbol vector.

The invention provides a direction of arrival estimation method based on an accelerated near-end gradient method, and the method comprises the following steps: S1, based on the virtual domain characteristics of a received signal, in combination with the mathematical properties of a Kronecker product, carrying out vectorization processing, carrying out form transformation, and eliminating the aperture information of a redundant array by using a dimension reduction matrix; s2, constructing a corresponding atom set, determining an atom norm expression, converting the atom norm expression into a positive semidefinite programming problem by adopting a convex relaxation technology, introducing a regularization constraint mechanism, and solving by adopting an accelerated near-end gradient algorithm to obtain a receiving signal after sparse reconstruction and a corresponding Hermitian-toeplitz matrix; and S3, on the basis of the Hermitian-toeplitz matrix, performing eigenvalue decomposition operation, extracting a noise subspace, and applying a spectral peak search algorithm to realize accurate estimation of the direction of arrival. According to the method, the complete array aperture information of the received signal in the virtual domain is fully mined and utilized, so that the calculation complexity is remarkably reduced, and the better estimation precision performance is realized.

This invention provides a signal azimuth estimation method and device based on high-order cumulants and auxiliary array elements, comprising: S1: receiving N far-field signals from M spatial sensor arrays, and obtaining the array manifold formed by the sensor arrays for the N signals; S2: performing Fourier transform on the received array signals to obtain a signal model; S3: expanding the array based on the signal model to obtain the fourth-order cumulants of the received signals in the model; S4: obtaining the direction vector of the expanded signal, and generating a construction matrix of the fourth-order cumulants based on the Kronecker product, and transforming the matrix; S5: performing eigenvalue decomposition on the fourth-order cumulants to obtain a noise subspace, and adjusting the row order of the noise subspace according to the element order of the array vectors to obtain an adjusted noise subspace, and estimating the source azimuth angle according to the subspace orthogonality principle. This solves the problem of correcting azimuth-dependent array errors.

The present application provides a communication method and apparatus, capable of reducing the computational complexity in detecting carried first bit information, improving the detection efficiency, and improving the communication performance. The method comprises: determining a first signal on the basis of a first sequence corresponding to first bit information, wherein the first sequence is constructed on the basis of a Kronecker product of a second sequence and a third sequence, the second sequence carries a first bit in the first bit information, and the third sequence carries a second bit in the first bit information; and sending the first signal.

The application discloses a dynamic load prediction method for a hanging passenger cableway based on a deep neural network, which comprises the following steps: step one, collecting and preprocessing the running data of multiple hanging cars in the hanging passenger cableway; step two, inputting the running data into a TimesNet network; step three, calculating the numerical difference between adjacent time points of an initial load prediction value sequence, marking local anchor points, and constructing a cross-hanging car anchor point graph; step four, calculating the disturbance intensity of an anchor point pair based on a two-dimensional hanging car position index matrix, and generating a disturbance embedding tensor; step five, performing a Kronecker product operation on the embedding representation tensor and the disturbance embedding tensor, and inputting them into a prediction head structure; and step six, calculating the numerical error and trend error between the load prediction value sequence and the real load sequence, and updating the bias value. The application improves the load prediction accuracy and robustness of the hanging car through the TimesNet network and disturbance reconstruction.

The application discloses a frequency domain adaptive filtering method for identifying a low-rank acoustic system, which extends the nearest Kronecker product (NKP) to the frequency domain, and establishes an NKP-based frequency domain recursive least square (NKP-FRLS) algorithm for identifying a time-varying acoustic system. The NKP is used to decompose an adaptive filter with a length of L into two groups of sub-filters with lengths of L1 and L2, and a signal model and a recursive least square cost function are established, the input signal spectrum matrix, the sub-filter error vector, the power spectrum matrix and the Kalman filter gain matrix are calculated in sequence, and on this basis, the sub-filter vector is calculated to obtain the coefficient vector of the modeling filter, so that the calculation amount of the adaptive filter is reduced, and a Gaussian noise robust adaptive algorithm is obtained. Experiments show that the application is superior to the traditional frequency domain recursive least square (FRLS) algorithm in convergence performance and calculation efficiency.

This invention provides a fixed-point matrix FFT calculation method based on memristors, belonging to the fields of signal processing and image processing technology. First, a nonlinear voltage-controlled memristor model is constructed, simulating state transition behavior in logic operations through a dynamic adjustment mechanism to realize the function of basic logic gates. Second, an FFT algorithm based on matrix block optimization is designed, decomposing the FFT calculation into small-scale matrix operation units and optimizing computational efficiency by combining Kronecker product and permutation matrix. This is then recursively extended to higher-point FFT operations. Simultaneously, a full adder and multiplier are designed based on memristors to achieve efficient fixed-point operations. Finally, the method is applied to image FFT processing to obtain the processing results. This invention achieves higher computational efficiency and lower energy consumption while maintaining computational accuracy, and can be widely applied in image processing, signal analysis, scientific computing, and other fields, possessing significant practical value and application prospects.

The invention discloses an improved two-dimensional data local differential privacy distribution estimation method based on an expectation maximization algorithm. A terminal device obtains two-dimensional numeric private data of a user, maps the two-dimensional numeric private data to a predefined two-dimensional grid domain, generates disturbance data by applying a two-dimensional disturbance mechanism meeting local differential privacy, and sends the disturbance data to an aggregation server. The aggregation server constructs a non-uniform grid based on the edge distribution characteristics of the noise data to compress dimensions, constructs a compressed two-dimensional transition probability matrix by using the Kronecker product of the one-dimensional transition matrix, and performs dimension reduction preprocessing on the data; and then, an improved expectation maximization iterative algorithm is executed, a maximum entropy regularization strategy is applied in a maximization step, a log-likelihood objective function containing a Shannon entropy regularization term is constructed, and a new distribution estimation vector is calculated by using a nonlinear updating formula based on a Lambert W function or power law scaling. According to the method, the problem of dimension disaster of two-dimensional distribution estimation is effectively solved, the over-fitting phenomenon under sparse data is inhibited through maximum entropy regularization, and the calculation efficiency and accuracy of two-dimensional numerical value distribution reconstruction are remarkably improved.