33 results about "Hadamard product" patented technology

The invention provides a click rate prediction method based on an attention mechanism. The method comprises the following steps: 1, preprocessing the features of a user, and carrying out One-hot one-hot coding on the features of the same type of users to obtain a high-dimensional sparse feature vector; 2, performing dimension reduction on the high-dimensional sparse feature vector through an embedded vector, and taking the feature vector after dimension reduction as an input vector of a click rate model to be respectively substituted into a compression interaction network and a deep neural network; and 3, performing Hadamard product on the input initial feature vector and the input vector of each hidden layer, taking the obtained result as the input value of the next hidden layer, and enabling the combination of the features to rise by one dimension every other hidden layer. The low-dimensional features, the explicit high-dimensional features and the implicit high-dimensional featuresof the user are comprehensively considered, useful feature combinations are screened through a self-attention mechanism, the prediction efficiency is improved, manual feature extraction is not needed,and the high-dimensional feature combinations can be extracted.

The invention provides a feature extraction method and device based on voice signal time domain and frequency domain, and an echo cancellation method and device, and the method comprises the steps: carrying out the short-time Fourier transform of a voice signal to obtain a time-frequency domain feature, and obtaining an intermediate mapping feature through a multilayer convolutional neural network; then obtaining a time weight vector based on a time domain attention module, expanding the time weight vector to the same dimension as the intermediate mapping feature, then performing Hadamard product to obtain a mapping feature subjected to time domain weighting, and then obtaining a frequency weight vector by using a frequency domain attention module; and expanding the time-weighted mapping features to the dimension same as the time-weighted mapping features, and performing Hadamard product to obtain the final time-domain and frequency-domain weighted mapping features. The time domain attention module and the frequency domain attention module can be easily embedded into the acoustic echo cancellation model based on the convolutional neural network, so that the model adaptively learns the weight of the time domain feature and the frequency domain feature, and the performance of the model is improved.

The invention discloses a two-dimensional DOA estimation method based on an L-shaped array in the technical field of array signal angle estimation. The two-dimensional DOA estimation method comprisesthe following steps: 1, establishing a time domain model of array receiving; 2, decomposing an actual steering vector, and processing the decomposed vector by utilizing a Hadamard product; placing thelast Q elements of the vector processed by the Hadamard product at the tail of the actual steering vector to construct a virtual steering vector; 3, constructing estimation of a received signal autocorrelation matrix according to the virtual steering vector, and carrying out eigenvalue decomposition to obtain estimation of a noise subspace; and 4, constructing a spatial spectrum function, and estimating the DOA of the incident signal according to spectral peak search. The underdetermined problem that the number of signal sources to be estimated is greater than the number of physical array elements under the condition of limited array elements is effectively solved. The complexity and hardware cost of equipment in practical application are reduced, and the signal DOA estimation precision is effectively improved.

The invention relates to a target tracking algorithm based on an object space relation, and the algorithm specifically comprises the following steps: G1, taking the features of an image and the position of a target in the image as the input, capturing the correlation of the targets in two frames of images through employing a five-layer convolution small network, and finally obtaining a feature similarity matrix between the targets; and G2, obtaining the distance between the center points of the targets to serve as a spatial similarity matrix between the targets, performing Hadamard product onthe spatial similarity matrix and the feature similarity matrix, calculating the relevance of the targets, preventing the targets with high apparent feature similarity and relatively far spatial position distance from being wrongly associated, and finally obtaining a target tracking result. The problem that the tracking effect of a target tracking algorithm only utilizing the target apparent characteristics is reduced in a scene with high target apparent characteristic similarity and low spatial characteristic similarity can be solved.

The invention discloses a variable pulse repetition interval SAR imaging method based on compressed sensing, is applied to the technical field of radars, and aims at solving the problems that the existing low oversampling Staggered SAR imaging algorithm is limited by a pulse repetition interval (PRI) slow change mode and the frequency spectrum of a surface target is not sparse, so that high-precision imaging cannot be carried out. According to the method, range cell migration (RCM), non-uniform sampling and echo blocking loss are embedded into an optimization problem by using matrix multiplication of a three-dimensional tensor, a two-dimensional matrix and a Hadamard product respectively, an image without azimuth ambiguity is reconstructed by solving the optimization problem, and the imaging performance reconstructed by the method is not influenced by a low oversampling rate, the PRI slow transformation mode and a target type.

The embodiment of the invention provides a data preprocessing method and device for a knowledge graph. The method comprises the steps of when vector expressions of nodes and connection relationship categories in a knowledge graph are determined, fusing the Hadamard product of a first node vector corresponding to a head node in the triple and a first relationship vector corresponding to a connection relationship category between the head node and a tail node to obtain an intermediate vector; and taking the distance between the intermediate vector and a second node vector corresponding to the tail node as a reference evaluation index, and updating the corresponding node vector and the relationship vector based on the adjustment of the values of the positive and negative samples on the reference evaluation index. According to the method, on the basis of saving parameters, common association relationships such as a symmetric relationship, an opposite relationship and a combined relationship can be expressed at the same time, and the expression capability of the knowledge graph is improved.

The invention discloses a sparse subspace clustering algorithm based on semi-supervision. The sparse subspace clustering algorithm comprises the steps that data prior information is converted into a constraint matrix suitable for a sparse subspace model in the form of point pair constraint; interference of flag-free bits is eliminated in the form of Hadamard product, the state of the coefficient represented by different constraint conditions is also considered and corresponding constraint terms are established; and a semi-supervised sparse subspace model of two hard threshold and soft threshold forms is established by using the constraint terms, and a semi-supervised framework is accordingly established on the sparse subspace clustering algorithm. The clustering accuracy of the sparse subspace algorithm can still be maintained by the algorithm without prior information. Meanwhile, the performance advantages of the sparse subspace clustering algorithm are also absorbed so that the high-dimensional clustering problem containing interference information data can be directly and effectively processed, the clustering performance is ensured to be effectively enhanced under the condition of less known prior information and thus the algorithm applicability can be increased.

The invention provides a beam forming method combined with an antenna array element directional diagram. By utilizing the beam forming method, the performance when the directional diagrams of beam forming array antenna units are inconsistent can be enhanced. According to the technical scheme, the method comprises the following steps: for an array antenna, carrying out directional diagram test on each antenna unit of the array antenna, and establishing an array antenna directional diagram database according to an antenna test point directional diagram; obtaining a fine directional diagram responded by the array antenna in a direction with a higher discretization degree by utilizing an antenna directional diagram database and adopting an interpolation mode of linear interpolation and splineinterpolation; carrying out interpolation on an actually measured antenna directional diagram to obtain a finer directional diagram of the array antenna, utilizing the antenna directional diagram to correct a guide vector of an ideal array antenna, wherein the correction method comprises the following steps: carrying out Hadamard product on the fine directional diagram obtained by interpolation and the guide vector of the ideal array antenna to obtain a corrected guide vector; and then obtaining a beam forming weight vector by using algorithms such as a linear constraint minimum variance LCMVand the like.

The invention relates to the field of direction-of-arrival positioning, and discloses a two-dimensional DOA estimation and channel phase disturbance correction method based on a uniform circular array. According to the two-dimensional DOA estimation and channel phase disturbance correction method, phase disturbance is processed at the cost of sacrificing the array aperture, and the method is a two-stage rank loss method. The two-dimensional DOA estimation and channel phase disturbance correction method comprises the steps of: 1, according to the characteristics of a Bessel function, truncatingan ideal steering vector in a compact uniform circular array mode space domain to construct a dimension reduction vector, and constructing a target function only related to phase disturbance based onthe dimension reduction vector and mode space domain conversion of channel phase disturbance; and 2, decoupling pitch and azimuth parameters in an ideal steering vector in a uniform circular array mode space domain to obtain a dimension reduction vector related to a pitch angle, and constructing a one-dimensional search target function only related to an azimuth angle. The two-dimensional DOA estimation and channel phase disturbance correction method provided by the invention can work online, and does not need an auxiliary array element. Besides, compared with a phase disturbance removing method based on a Hadamard product, the two-dimensional DOA estimation and channel phase disturbance correction method can be applied to a uniform circular array with a larger aperture, only needs a small number of array elements and can be applied to two-dimensional DOA estimation.

The invention discloses a density abrupt change interface inversion method and system based on machine learning constraints. The method comprises the following steps: constructing an initial basin interface, randomly generating a disturbance basin interface data set, and carrying out Hadamard product operation on the initial basin interface and the disturbance basin interface data set to obtain a basin interface data set; performing higher function filling on the basin interface data set to obtain a high-resolution density interface model data set, performing forward modeling calculation to obtain a simulated gravity data set, and performing transformation and weighting on the simulated gravity data set to obtain a low-resolution offset density interface model data set; and optimizing an offset model deep learning network and mapping to obtain a high-resolution constraint density interface prior model, constructing a stable nonlinear loss function and carrying out regularization inversion to obtain a high-resolution density interface model. By exploring the input mode and the learning mode of deep learning, research on the density mutation interface inversion method based on machine learning constraint is carried out, and the precision of interface inversion imaging is enhanced.

The invention provides a recommendation method and system based on deep collaborative filtering, and the method comprises the steps: obtaining massive user and project score data from the Internet, and preprocessing the user and project score data; performing embedding operation on the preprocessed data, and completing the linear representation process of the interaction function; sending the complete linear representation of the interaction function into a deep neural network to obtain a complete adaptive interaction function; predicting scores, and generating a recommendation list. In the same process, the Hadamard product and linear feature extraction process is carried out on the implicit vectors of the user and the item, and a more complex interaction function can be learned to a certain extent by adding a hidden layer on the basis of combining linear representation of the user and the item, so the performance of the model is improved, and the user experience is improved.

The invention discloses an online car-hailing demand prediction method based on a convolutional network and a non-local network. Local spatial features can be extracted by using the convolutional neural network, and global spatial features can be extracted by using the non-local network. A convolutional neural network model and a non-local network model are respectively established for online car-hailing historical demand tensors corresponding to a previous moment of a current moment, the same moment of a previous day and the same moment of a previous week, then modeling is performed on time and weather characteristics, label coding is performed on time data and weather data, then one-hot coding is performed, finally, each time period corresponds to one k-dimensional vector, a full-connection neural network is built, the k-dimensional vectors serve as input and are sent into a two-layer full-connection neural network, and the output shape of the k-dimensional vectors is made to be the same as the shape of a space grid. And finally, the models are fused by using a Hadamard product to complete online car-hailing demand prediction. According to the method, the urban online car-hailing demand prediction precision can be remarkably improved, and the calculation efficiency is greatly improved.

The invention discloses a tea sensory quality comparison method and system based on review comments. The method comprises the following steps: presetting comment words and corresponding assignments thereof; respectively setting a term weight vector and a scale vector of the batch according to the weight coefficient of the primitive term and the scale value of the degree adverb contained in the comment of each sample of the batch; obtaining a Hadamard product of the term weight vector and the scale vector as a comment vector; according to the comment vector of each sample in the batch, extracting a maximum and minimum comment vector; calculating the maximum similarity and the minimum similarity according to the comment vector and the maximum and minimum comment vector; calculating a sample score of each sample based on the maximum similarity and the minimum similarity; and judging the sensory quality of each sample according to the score of the sample. By adopting the method, the accuracy and the stability of a tea quality comparison result can be improved.

The invention relates to a tool wear prediction method and device, equipment and a storage medium. The method comprises the following steps: collecting original cutter data, introducing a Hadamard product to improve an attention gate, fusing the improved attention gate with an independent recurrent neural network to obtain a plurality of independent recurrent network basic models with a fused attention mechanism, and stacking the plurality of independent recurrent network basic models to obtain a fused attention mechanism. According to the method, a deep independent circulation network model is constructed, the deep independent circulation network model is combined with a convolutional neural network, and a tool wear prediction model used for predicting a wear result of a tool is constructed according to original tool data. Due to the fact that the Hadamard product is introduced into the deep independent circulation network model in the tool wear prediction model for improvement, original input can be adjusted, the influence of input elements with high importance on the model can be enhanced, elements with low importance can be restrained, and the prediction accuracy of the tool wear result is improved.

The invention discloses an SMV model-based Hadamard product fast DOA estimation method using tail optimization. The method comprises the steps of: estimating a received data DOA by using a DOA estimation algorithm to obtain a first estimation result and an array output result; performing normalization processing on the first estimation result by using a normalization model to obtain an initial solution; obtaining a position sorting result according to a sorting result of the amplitudes of the first estimation result; obtaining a diagonal matrix according to the position sorting result and an all-1 array; obtaining a first parameter and a second parameter according to the array output result, the initial solution and the diagonal matrix; based on an estimation result calculation formula, obtaining a second estimation result according to the first parameter and the second parameter; and obtaining a final estimation result according to the relationship between the indexes of iteration sequence numbers and the number of array elements, the first estimation result and the second estimation result. According to the method, the tail of a received signal is optimized by using the Hadamard product parameterization product principle, and energy is concentrated to a target peak value to a great extent, so that an estimated angle is more accurate.

The invention discloses a text mining method for heterogeneous graph conversion based on meta-structure learning, which comprises the following steps of: aiming at text data, extracting information in a text to construct a heterogeneous information network graph; acquiring a meta-path through a graph conversion layer to capture a relationship between nodes; extracting a meta-graph structure by establishing a channel type Hadamard product module, so that multiple interaction conditions existing between nodes at the same time are captured; generating node embedding for the extracted meta-structure containing the meta-path and the meta-graph by using a graph convolutional network; and embedding and mining a downstream text by using the obtained nodes. The method can be suitable for a complex text recognition environment, loss of semantic information is effectively avoided, and rich and complete semantic information can be obtained.

The invention discloses a blade multi-view-field point cloud registration method based on overlapping features and local distance constraints, and the method comprises the steps: step 100, obtaining the multi-view-field point cloud data of a blade, solving an overlapping region of the point cloud data between two adjacent view fields, and constructing a data set X and a data set Y; step 200, establishing an overlapping feature constraint matrix FMN; step 300, establishing a local distance constraint matrix DMN; step 400, on the basis of the Hadamard product of the matrix, a member probability matrix PMN between the data set X and the data set Y is constructed through an overlapping feature constraint matrix FMN and a local distance constraint matrix DMN, PMN = FMNDMN, and PMN is the Hadamard product; and step 500, calculating a rotation matrix and a translation matrix between the data set X and the data set Y based on an EM algorithm, and splicing a complete blade section contour through the calculated rotation matrix and the translation matrix. The invention provides a blade multi-view-field point cloud data registration method which does not depend on the motion precision of a detection system so as to achieve the purpose of improving the blade detection precision and reliability.

The invention provides a fast coding method based on image texture and gradient features, and relates to the field of multimedia videos. The method comprises the steps of dividing a CU and calculating texture complexity Th and Tv35 of a current coding CU in the horizontal direction and the vertical direction of a CU block, and specifically comprises the following steps: 1, calculating the maximum value of Th and Tv as Tmax and the minimum value as Tmin. The rapid coding method based on the image texture and the gradient features comprises the steps: acquiring a horizontal gradient value Gh after a hadamard product is obtained through a horizontal sobel similar operator and a CU pixel brightness value, and acquiring a vertical gradient value Gv is obtained after an element sum after a hadamard product is obtained through a vertical sobel similar operator and the CU pixel brightness value; calculating the projection value SH of the gradient vector at different intra-frame prediction angles theta, and determining the range of an intra-frame prediction model, thereby solving the problems that a large amount of operation complexity is introduced in the intra-frame prediction process of the existing HEVC, and how to effectively reduce the operation amount of an encoder becomes an urgent problem at present.

The invention belongs to the field of text information processing, and provides a text processing method and system based on a parallel zero-redundancy long short-term memory network, wherein the method comprises the steps of obtaining to-be-processed text data and converting the to-be-processed text data into a word embedding vector form; according to the number of words contained in the to-be-processed text data, adaptively calculating a context window coverage range of each word in the to-be-processed text data; in a parallel zero-redundancy long-short-term memory network, compressing all word embedding vectors within the coverage range of the context window to form a local attention vector matrix, and performing parallel calculation through Hadamard product matrix multiplication to obtain local context vectors corresponding to all word embedding vectors; and processing the local context vector corresponding to the to-be-processed text data through the classification network model to obtain a text classification or labeling result.

The invention discloses a sea clutter analysis method and system applied to a conformal array radar. The method comprises the following steps: constructing a sea clutter geometric model and obtaining aerial carrier parameters; according to the sea clutter geometric model, calculating a sea clutter beam pointing unit vector and a first sea clutter airspace steering vector; calculating an array element gain vector according to the sea clutter beam pointing unit vector, and solving a Hadamard product of the array element gain vector and the first sea clutter airspace steering vector to obtain a second sea clutter airspace steering vector; calculating a sea clutter time domain steering vector according to the sea clutter beam pointing unit vector and the aerial carrier parameters, and solving a Kronecker product of the sea clutter time domain steering vector and the second sea clutter space domain steering vector to obtain a sea clutter space-time steering vector; obtaining a sea clutter sampling covariance matrix according to the sea clutter space-time steering vector; and obtaining a sea clutter analysis result according to the sea clutter sampling covariance matrix. According to the sea clutter analysis method, the sea clutter influence is considered, and the accurate sea clutter analysis result is obtained.

The invention provides a consistency checking method and device for a scoring matrix. The method is used for improving the consistency checking efficiency. Comprising the steps that after a scoring matrix sent by a user is received, whether the scoring matrix passes consistency checking or not is judged through a first preset algorithm; if the scoring matrix does not pass the consistency check, dividing the scoring matrix by an intermediate matrix to obtain a disturbance matrix, each element in the intermediate matrix being obtained based on each element in the scoring matrix; the disturbance matrix and the zero disturbance matrix are subjected to weighted summation, a first matrix is obtained, the row number and the column number of the zero disturbance matrix are the same as those of the disturbance matrix, and all elements of the zero disturbance matrix are the same set value; and determining a Hadamard product between the first matrix and the intermediate matrix as a second matrix, determining the second matrix as a scoring matrix, and returning to the step of judging whether the scoring matrix passes the consistency check by using the preset algorithm until the scoring matrix passes the consistency check.

The invention discloses a high-dynamic GNSS interference suppression method based on a null-slot widening technology for a dual-polarized antenna array, and belongs to the technical field of polarization-sensitive uniform linear array signal processing. When the present invention performs tapering processing on the autocorrelation matrix of the received signal, the tapering matrix adopted is the hadamard product of the spatial domain tapering matrix and the polarization domain tapering matrix, wherein the polarization domain tapering matrix is ​​based on the polarization domain zero Then, based on the tapered autocorrelation matrix, the output signal power is minimized under unconstrained conditions to obtain the weight vector, and then beamforming is performed to obtain the output signal after interference suppression. The present invention has the advantage of being able to effectively suppress interference under the condition of null notch mismatch; compared with the traditional scalar array to achieve high dynamic GNSS interference suppression through the null notch widening method, it has the ability to distinguish the interference of different polarization modes in the same airspace position, The advantage of more degrees of freedom.

The invention discloses a blade multi-field point cloud registration method based on overlapping features and local distance constraints, including step 100: obtaining multi-field point cloud data of a blade, and solving point clouds between two adjacent fields of view overlap area of ​​data and construct data sets X and Y; step 200: establish overlapping feature constraint matrix F MN ; Step 300: Establish a local distance constraint matrix D MN ; Step 400: Based on the Hadamard product of the matrix, constrain the matrix F by overlapping features MN and the local distance constraint matrix D MN Construct the membership probability matrix P between dataset X and dataset Y MN , P MN =F MN D. MN , is the Hadamard product; Step 500: Calculate the rotation matrix and translation matrix between the data sets X and Y based on the EM algorithm, and splicing the complete blade cross-sectional profile through the calculated rotation matrix and translation matrix. The invention proposes a blade multi-field point cloud data registration method that does not depend on the motion accuracy of the detection system, so as to achieve the purpose of improving the detection accuracy and reliability of the blade.

The invention belongs to the field of artificial intelligence research, and provides a high-order logic knowledge graph representation learning method based on structural features. The method comprises the following steps: firstly, performing high-order logic relation feature representation on data, extracting node neighborhood structure motif degree matrixes in a graph, and taking an entity motif degree matrix in each relation as a high-order logic relation feature; secondly, performing entity attribute feature and high-order logic relation feature representation on the graph convolutional network; and finally, carrying out feature polymerization by using three different polymerization methods of Hadamard product, summation and series connection. According to the method, two knowledge graph representation learning features are fused, sample data are learned from different angles and then are used for knowledge graph downstream tasks such as entity classification and link prediction, a high-order logic knowledge representation learning method with high precision is obtained, and meanwhile, the defect that an existing knowledge representation learning method is low in complexity is overcome. The method solves the problems of insufficient expression capability, large expression loss, high calculation cost, poor interpretability and the like of high-order logic relationships such as inverse relationships and composite relationships.

The invention provides an efficient fuzzy searchable encryption method based on Word2vec and ASPE. The method is mainly applied to the field of secret state information fuzzy search. According to the algorithm, firstly, synonym correlation coefficients are generated by means of Word2vec, then the coefficients are combined with a keyword matrix, and a trap door matrix of a user is obtained through an expanded ASPE algorithm. And performing Hadamard product operation on a file index matrix generated by using the expanded ASPE algorithm and a trap door matrix to obtain a file correlation sequence, and finally, decrypting the file by a user through a key to obtain a final plaintext. Experimental simulation shows that the algorithm not only can effectively hide plaintext information and protect forward privacy and backward privacy of a user, but also greatly improves the defects that a traditional algorithm is large in calculation amount and occupies a large amount of storage space, and greatly improves the fuzzy search efficiency.

The invention relates to a method for operating a slave station in a communication network comprising a master station, the method comprising: generating a precoding matrix obtained from the Hadamard product of the alphabet modification matrix and the original precoding matrix, Wherein the original precoding matrix includes complex coefficients of equal amplitude; and a precoding report representing the precoding matrix is ​​sent to the main station.

The invention discloses an Otsu thresholding method with a mask. The method comprises the following steps: 1, determining the number of masks and the number of permeation regions in the masks, making a template with the same size according to an original image, and filling a shielding region and a permeation region in the template with a gray scale shielding factor and a permeation factor respectively; 2, taking the Hadamard product of the image and the mask as an extracted target area; 3, calculating the maximum between-class variance in the target region to obtain a threshold value of the thresholding operation of the target region; and 4, carrying out threshold binarization processing on the threshold obtained in the step 3. According to the method, firstly, the image is masked and permeated through the personalized mask design, then, the target area is extracted by using the Hadamard product, and finally, the threshold is determined by only taking the mask area as an object and adopting the maximum between-class variance calculation theory, so that the principle is simple, and the practicability is relatively high.

The invention discloses a ring code-based multivariate low-density check code design method for a satellite communication link, belongs to the technical field of encoding and decoding, and aims to solve the problem that hardware implementation conditions for multivariate LDPC encoding and decoding are very harsh due to the fact that the satellite communication link has very high propagation loss and is tense in link budget. The multivariate low-density check code design method based on the ring code is efficient and capable of saving hardware resources. The method specifically comprises the steps of firstly determining a basis matrix of a QC-LDPC code, then obtaining a finite field element corresponding to the basis matrix, finally expanding the basis matrix and a finite field element matrix, and finally performing Hadamard product operation to obtain a check matrix of the multivariate QC-LDPC code. According to the method, the construction method of the multivariate LDPC code is simplified, the decoding performance of the LDPC code is effectively improved at the same time, and the method is suitable for an environment in which hardware implementation conditions of multivariate LDPC encoding and decoding of a satellite communication link are very harsh.