53 results about "Matrix factorisation" patented technology

A correlation-aware adaptive codebook compaction system for multi-modal data compression that preserves cross-modal relationships while providing enhanced reconstruction quality. The system analyzes temporal and spatial relationships between different data modalities to generate correlation maps that guide compression decisions. A virtual management layer performs stream characterization and adaptive routing, while a processing pipeline implements primary codebook compression with mismatch handling for novel data blocks. High-entropy data segments receive pre-compression processing before codebook compression. Sequential registration data is processed through matrix factorization and dedicated matrix codebooks. The system continuously monitors data distribution characteristics and automatically retrains codebooks when drift thresholds are exceeded. A neural upsampling subsystem uses correlation information to guide cross-modal enhancement processes through modality-specific networks and attention mechanisms. The unified output includes compressed data streams, correlation maps, synchronization metadata, neural model parameters, and updated codebooks, enabling synchronized reconstruction with preserved cross-modal relationships and enhanced quality through correlation-guided neural upsampling.

The invention relates to the field of environmental monitoring, and particularly discloses a non-negative matrix factorization and adaptive peak recognition fluorescence feature extraction and traceability system, which comprises a spectral data preprocessing module, a spectral data non-negative matrix factorization module, a component number automatic selection module, an adaptive peak recognition module, a feature library construction module and a similarity comparison module. An improved non-negative matrix factorization model is adopted to decompose the three-dimensional fluorescence spectrum matrix of a single sample, and an optimal component number K is automatically determined through multiplicative update rule iterative optimization; the self-adaptive peak identification module carries out selective filtering, accurately extracts the position and intensity of a fluorescence peak through multiple mechanisms, and carries out peak position calibration in a neighborhood; the Hungary algorithm is adopted to carry out characteristic peak matching to calculate the comprehensive similarity between the samples, and rapid and accurate identification of the pollution source is realized. The method has the advantages of high resolution, strong anti-interference capability, low requirement on the number of samples, automation and the like, and is suitable for water quality fingerprint feature extraction of a water sample in a complex environment and real-time source tracing of sewage.

The invention discloses an emotion recognition method based on multi-modal signal fusion, and relates to the technical field of emotion recognition, and the method comprises the steps: obtaining an electroencephalogram signal and a peripheral physiological signal to form a multi-modal signal, and carrying out the preprocessing of down-sampling, baseline correction and band-pass filtering on the multi-modal signal; based on the preprocessed signal, extracting a difference entropy feature and a power spectrum density feature; mapping the extracted features into a standardized space grid to generate a feature tensor with a uniform structure; performing adaptive weighting on the mapped feature tensor by using a frequency band fusion attention mechanism, and generating a frequency band weight through global average pooling and a full connection layer; inputting the weighted features into a full connection layer of the parameterized hypermatrix, and performing feature compression and modeling through matrix decomposition and reconstruction; a multi-task learning framework is adopted, classification results of emotion titer and awakening degree are output at the same time based on compressed features, and multi-task collaboration is optimized through a shared feature layer and a dynamic loss weight.

The invention discloses an unbalanced classification method for electroencephalogram data in epilepsy detection, and relates to the technical field of data optimization and big data processing. According to the method, firstly, covariance matrixes of majority class samples and minority class samples are calculated, linear transformation is achieved through matrix decomposition, and the minority class samples inherit global distribution characteristics of the majority class samples; and then, in the transformed feature space, sorting samples based on mahalanobis distance and performing partition pairing, selecting sample pairs with large difference to generate convex combination synthesis samples, and ensuring sample diversity and boundary consistency. Experimental results show that on a CHB-MIT electroencephalogram data set, the method effectively solves the problems that a traditional oversampling technology is prone to expanding minority class decision boundaries and generated samples are lack of diversity, and the reliability of epileptic seizure detection is remarkably improved.

The invention relates to the technical field of industrial process soft measurement and quality control, and discloses a paper quality prediction method based on multivariate statistical latent variable fusion and space-time transformation, which comprises the following steps: acquiring space-time sequence data of a multi-source sensor in a papermaking process, constructing a space-time diagram structure reflecting a topological relation of equipment, and preprocessing. Then, multi-view latent variables are extracted through non-negative matrix factorization, independent component analysis and robust principal component analysis, attention fusion is conducted on the latent variables through an LV fusion module, and fusion latent variables are obtained; and inputting the fusion latent variable and original node data into a multi-scale convolution auto-encoder to obtain spatial feature embedding, and inputting the spatial feature embedding and the fusion latent variable into a space-time Transform module together to realize joint modeling of space correlation and time dependence. And finally, outputting a paper quality predicted value through a linear regression module. The method can achieve the accurate prediction of the paper quality under a high-dimensional and multi-noise working condition, and is suitable for online monitoring and modeling updating.

This invention discloses a multi-sensor fusion mapping method and apparatus for degraded scenarios, comprising: a front-end fusion odometry method: receiving multi-source data and obtaining initial pose through coordinate system alignment; performing point cloud distortion removal through IMU integration and extrinsic parameter transformation to generate a priori LiDAR pose; analyzing point pair contribution vectors based on ICP Hessian matrix decomposition to determine pose degradation; for non-degraded frames, constructing LiDAR point-to-surface residuals and RTK pose residuals, updating the state through ESIKF, and outputting keyframe pose, point cloud, and degradation labels; a back-end factor graph optimization method: filtering non-degraded keyframes through a sliding window, optimizing pose through point cloud beam adjustment; performing secondary verification of RTK pose to remove invalid values; constructing odometry, loop closure, and RTK factors, and fusing and optimizing to obtain a globally consistent high-precision point cloud map. This method enhances robustness through degradation detection and ensures global consistency through factor graph fusion, making it suitable for mapping complex scenarios.

A processing device (400) for performing non-negative matrix factorization on measured X-ray powder diffraction profiles comprises a measurement profile acquisition section (410) for acquiring a plurality of measured profiles, a decomposition section (420) for applying non-negative matrix factorization to the measured profiles and for calculating base profiles; an index calculation section (430) for acquiring the base profiles and for calculating indices based on a non-uniformity of the base profiles; a base profile classification section (440) for classifying the base profiles into a plurality of groups based on the indices; and a base profile correction section (450) for performing a correction based on the indices on at least one of the plurality of groups and for calculating a corrected base profile.

The application discloses an unsupervised hyperspectral image super-resolution method based on a matrix decomposition network and belongs to the technical field of hyperspectral image processing. The application is used for processing a hyperspectral image, generating a simulated low spatial resolution hyperspectral image Y and a high spatial resolution multispectral image Z; first, inputting the generated data pair (Y, Z) into a designed auto-encoder network, training iteration to obtain a point spread function and a spectral response function; for a target high spatial resolution hyperspectral image X, the target high spatial resolution hyperspectral image X can be assumed to be a linear combination of a terminal member matrix A and a corresponding abundance matrix S, that is, X = AS, a spectral and spatial degradation model is combined to model, a deep CP decomposition module is designed to calculate A, A and S are iteratively solved, and finally a fusion result is obtained. The application can obtain more rich spectral and spatial features, obtain a better fusion result, and has good performance in practice.

The invention discloses a remote sensing image sub-region rapid classification method based on dual-scale multi-graph regularization non-negative matrix factorization, and provides a dual-scale basis selection strategy and a multi-graph regularization non-negative matrix factorization model for overcoming the defects of complex scene adaptation, small sample learning and classification speed and accuracy balance in the prior art. Unsupervised subregion classification is realized through the steps of data preprocessing, double-scale seed set division, basis matrix construction, double information graph regularization item construction, model optimization solution and the like. According to the method, the classification number does not need to be predefined, the basis matrix representativeness is enhanced through dual-scale basis selection, and the classification precision and robustness are improved by combining a dual-graph regularization item and utilizing a manifold structure and discrimination information of data at the same time; the efficient optimization solution of the non-negative matrix factorization realizes the rapid classification of the sub-regions, and improves the accuracy and efficiency of the classification of the sub-regions of the remote sensing image.

This invention discloses a water pollution data monitoring method and system based on multi-objective analysis, belonging to the field of water pollution source tracing technology. The method includes: collecting a GIS map of the target watershed and time series data on rainfall, flow, and water quality; hydrologically segmenting the flow series according to the opening and closing status of dams to obtain surface runoff and baseflow components; constructing a two-dimensional water quality observation matrix, setting pulse and delay constraints in conjunction with rainfall and runoff components, and using a non-negative matrix factorization algorithm with ratio constraints for source analysis to obtain a source feature matrix and a source contribution time coefficient matrix; identifying pollution source emission attributes based on the ammonia nitrogen to total phosphorus ratio and the correlation between the source contribution coefficient and rainfall in the source feature matrix, and marking them on the GIS map. This invention integrates dam-controlled hydrological segmentation with matrix factorization based on physicochemical constraints to achieve refined source tracing, improving the accuracy and interpretability of pollution source identification.

The invention relates to the technical field of power customer service and automatic matching, in particular to an automatic matching method for power customer engineering. Comprising the following steps: screening out candidate service units meeting qualification conditions by using a multi-attribute decision analysis method; based on a content-based recommendation algorithm, extracting features of customer engineering, and recommending electric power engineering service units matched with the features; according to a recommendation algorithm based on matrix decomposition, power customer engineering and power engineering service units are expressed as rows and columns of a two-dimensional matrix respectively, and service units with potential matching are recommended by calculating the incidence relation between matrixes; the recommendation results of the two algorithms are fused, other factors are considered, sorting is carried out, and the final recommendation result is comprehensively generated. According to the design, automatic matching between the customer and the electric power engineering service unit can be realized, the matching success rate is improved, the transaction cost is reduced, and the operation efficiency of the electric power market is improved.

The application discloses a hyperspectral anomaly detection method and device based on factor group sparse regularization and belongs to the technical field of remote sensing image processing. The method reduces three-dimensional hyperspectral data into a two-dimensional matrix form, decomposes the matrix into a background part and an anomaly part based on a background dictionary; utilizes a Schatten-p norm to regularize and constrain the low-rank characteristics of the background part, utilizes a 2,1 norm to regularize and constrain the column sparse characteristics of the anomaly part; converts the Schatten-p norm into a factor group sparse regularization form, introduces an auxiliary matrix to replace variables after performing singular value decomposition on a coefficient matrix; iteratively solves the replaced variables by using an alternating direction multiplier method to obtain an anomaly matrix and calculates an anomaly detection value of each pixel. The application avoids a complex calculation process, guarantees detection accuracy and significantly improves detection speed.

The invention discloses a biological entity multivariate association prediction system combining linear and nonlinear fusion matrix factorization, which relates to the technical field of biological entity multivariate association prediction and comprises a multivariate biological entity input module, a cross-modal feature extraction module, a dual-channel fusion matrix factorization module, a dynamic feature fusion device and a multivariate association prediction engine. According to the method, the limitation of a traditional biological entity association prediction method is broken through by fusing linear and nonlinear matrix decomposition technologies, and a dynamic feature fusion mechanism realizes optimal combination of cross-modal features through adaptive weight adjustment, so that the problem of insufficient flexibility of a traditional static fusion strategy is overcome; in addition, the system adopts a three-dimensional tensor modeling technology, a unified prediction framework of multiple types of associations such as gene-disease, drug-target, metabolite-pathway and the like is realized, heterogeneity characteristics of biological associations can be effectively maintained, and compared with the prior art, the analysis capability of a complex biological network is remarkably improved.

The invention discloses an outer limb robot self-adaptive impedance control method and system based on a motor function. According to the method, motion information and EMG / EEG multi-mode signals are collected, a motion intention is predicted through weighted stacking and nonlinear compensation, the weight is updated based on an admittance model, and a planning trajectory is learned iteratively; the motor function index is obtained by decomposing the electromyographic signal through a non-negative matrix, the damping and rigidity are dynamically adjusted through feed-forward-impedance double-loop control, and the system stability is verified through the passivity theory. According to the method, the self-adaptive response of the outer limb robot to the human body motor function change is realized, the accuracy and safety of man-machine interaction are improved, and the method is suitable for outer limb robot application scenes such as rehabilitation training and industrial assistance.

The application discloses a hyperspectral anomaly detection method and device based on factor group sparse regularization and belongs to the technical field of remote sensing image processing. The method reduces three-dimensional hyperspectral data into a two-dimensional matrix form, decomposes the matrix into a background part and an anomaly part based on a background dictionary; utilizes a Schatten-p norm to regularize and constrain the low-rank characteristics of the background part, utilizes a 2,1 norm to regularize and constrain the column sparse characteristics of the anomaly part; converts the Schatten-p norm into a factor group sparse regularization form, introduces an auxiliary matrix to replace variables after performing singular value decomposition on a coefficient matrix; iteratively solves the replaced variables by using an alternating direction multiplier method to obtain an anomaly matrix and calculates an anomaly detection value of each pixel. The application avoids a complex calculation process, guarantees detection accuracy and significantly improves detection speed.

The invention relates to the technical field of gas spectrum detection, and discloses a noise suppression non-negative matrix factorization gas cross interference removal measurement method and system, the noise suppression non-negative matrix factorization gas cross interference removal measurement system comprises a gas supply module, a detection module and a control analysis module, the gas supply module comprises a plurality of gas tanks, a dynamic gas mixing device, a pressure controller and a one-way valve, the plurality of gas tanks are respectively and independently connected with a gas inlet of the dynamic gas mixing device to form a plurality of independent gas paths, and the pressure controller is arranged between a gas outlet of the dynamic gas mixing device and a gas inlet pipeline of the multi-way gas pool; and one-way valves for preventing backflow and cross contamination are respectively arranged on the gas inlet and the gas outlet of the multi-way gas tank. The method can stably separate and quantify the ethylene / ethane near-infrared weak spectrum band with weak absorption and serious absorption spectrum overlapping, has obvious anti-noise and anti-interference capabilities, and is convenient for miniaturization, low power consumption and batch deployment.

This invention relates to the field of image encryption and discloses an optical image asymmetric encryption method based on nonnegative matrix factorization. The encryption method includes the following steps: S1, generating a preliminary encryption result through a first optical transformation based on the original image and a first random phase mask; S2, generating an encrypted result through a second optical transformation based on the preliminary encryption result and a second random phase mask; S3, performing nonnegative matrix factorization on the encrypted result to generate an image matrix and a coefficient matrix; S4, using the coefficient matrix as ciphertext and the image matrix as the private key to complete image encryption. This invention significantly improves the security of optical encryption. It combines nonnegative matrix factorization with double random phase coding to construct an asymmetric encryption mechanism that effectively resists known plaintext, chosen plaintext, and special attacks, while also being robust to cropping and noise interference. The framework is highly scalable and applicable to image and audio data encryption.

The invention discloses an intelligent picture control method and a related device, and relates to the technical field of data analysis, and the method comprises the steps: carrying out the depth feature map and tensor feature extraction based on an image sequence, and carrying out the gait recognition in combination with a gravity center track, and obtaining the target gait data; performing user action prediction based on the target gait data in combination with brain-like semantic feature extraction; carrying out human voice audio separation based on the audio information in combination with non-negative matrix factorization so as to carry out voiceprint feature analysis, and obtaining voiceprint feature information; performing user identity detection based on the voiceprint feature information to obtain user identity information; and carrying out picture zooming analysis and picture switching analysis on the display screen based on the user identity information and the user action prediction information so as to adjust the picture of the display screen. According to the invention, the picture of the user can be tracked in real time, omission of the picture of the user is avoided, and a more ideal picture control effect is achieved.

According to a sequence analysis method for inferring the polyad content of a polymer, which polyad is formed by aligning a plurality of units, the method including determining the number K of variants of the polyads; sequentially ionizing gas components generated by heating each sample of a reference sample and an inference target sample to acquire a data matrix including two-dimensional mass spectra; performing non-negative matrix factorization of the data matrix to factorize the data matrix into the product of a basis spectrum matrix and an intensity distribution matrix; performing non-negative matrix factorization of the intensity distribution matrix to factorize the matrix into the product of a matrix representing the mass proportion of a model polymer formed only from the polyad and a matrix representing a feature vector; defining the feature vector of the model polymer as an end member and determining a K−1 dimensional simplex including all of the feature vectors of samples; and inferring each polyad content ratio by the distance ratio between the end member and the feature vector of the inference target sample, a simple sequence analysis method for polymers is provided.

The application relates to the technical field of data analysis, and discloses a data analysis method combining muscle synergy analysis and muscle function network, which comprises the following steps: step S1: collecting surface electromyogram signals of each muscle block during movement; step S2: muscle synergy mode analysis, integrating the collected surface electromyogram signal data by using non-negative matrix decomposition, extracting a muscle synergy mode, and quantifying the quality of a reconstruction matrix by using a variance explanation rate; and step S3: muscle function network analysis, obtaining normalized mutual information between each muscle channel as a nonlinear correlation characteristic parameter of multi-channel surface electromyogram, and constructing a muscle function network. The data analysis method of the application combines muscle synergy analysis and muscle function network based on surface electromyogram signals, uses non-negative matrix decomposition for muscle synergy analysis, quantifies the mutual relationship between channels by using normalized mutual information, and constructs a muscle function network, so that potential and valuable structured and modularized information in electromyogram data can be found.

The embodiment of the invention provides a large-scale matrix QR decomposition single-core serial calculation method, device and equipment and a storage medium, block calculation is performed on a matrix according to the fast memory capacity to realize multi-stage matrix QR decomposition, and compared with a single-core serial QR decomposition method, the method has the advantages that the matrix is always accessed in a rectangular block mode when the matrix is accessed, so that the matrix access efficiency is improved. In the calculation step, the serial QR decomposition step is executed in the diagonal matrix blocks, and the Level-3 BLAS is used when the non-diagonal matrix blocks are updated, so that the calculation complexity and memory access ratio in a single core are improved, and the calculation efficiency of the single core is improved.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for presenting a new machine learning model architecture. In some aspects, the methods include obtaining a training dataset with a plurality of training samples that includes feature variables and output variables. A first matrix is generated using the training dataset which is a sparse representation of the training dataset. Generating the first matrix can include generating a categorical representation of numeric features and an encoded representation of the categorical features. The methods further include generating a second, third and a fourth matrix. Each feature of the first matrix is then represented using a vector that includes a multiple adjustable parameters. The machine learning model can learn by adjusting values of the adjustable parameters using a combination of a loss function the fourth matrix, and the first matrix.

Various embodiments for systems and methods of tracking disaster footprints using social streaming media using nonnegative matrix factorization are disclosed herein. The system extracts a summarization output from historical data and compares the summarization output with incoming data to identify differing or similar topics within the data. The summarization output is projected to adjust a time-dependency of the summarization output to enable a more direct comparison. The system additionally uses the summarization output to encode topic data within historical data to reduce computational overhead.

This invention discloses a recommendation method based on a robust community detection model using ensemble learning, comprising the following steps: Step 1, obtaining a set of preliminary results: repeat... k The improved Louvain algorithm yields a preliminary set of results; Step 2: Generate an enhanced consensus network: Using the improved consensus network creation algorithm, the preliminary set of results is integrated into a consensus network; Step 3: Community detection is performed using nonnegative matrix factorization. First, matrices X and Y are randomly initialized with shapes as follows: N L and L N； Then the matrix is ​​processed through cross-iteration. X sum matrix Y The process involves iteration and convergence to obtain convergence matrices X and Y. Finally, the community ID corresponding to the largest element in the column of each node in convergence matrix Y is output as the community to which that element belongs. Each node yields the final community division, and the final output provides the unified community structure. This invention utilizes ensemble learning to learn relevant knowledge independently, avoiding the problem of unknowable prior knowledge.

The application relates to the technical field of computer data processing, and discloses a portfolio optimization method based on matrix decomposition technology and related equipment; the method comprises the following steps: obtaining portfolio data from an external memory or a local memory; the portfolio data comprises asset names of all assets in a portfolio and historical price data of a preset time period; based on the portfolio data, the average yield of the portfolio and a covariance matrix are calculated; the covariance matrix is subjected to eigenvalue decomposition to obtain a first matrix Q and a second matrix D; an auxiliary variable is constructed, a target function and a constraint condition are constructed based on the auxiliary variable; the target function and the constraint condition are input into an SQP algorithm for solving to obtain a solution of the auxiliary variable; and the weight distribution of each asset in the portfolio is calculated based on the solution of the auxiliary variable. The application can greatly improve the solving efficiency of the model, reduce the memory occupation, save the computer data processing time, and reduce the computer energy consumption.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for presenting a new machine learning model architecture. In some aspects, the methods include obtaining a training dataset with a plurality of training samples that includes feature variables and output variables. A first matrix is generated using the training dataset which is a sparse representation of the training dataset. Generating the first matrix can include generating a categorical representation of numeric features and an encoded representation of the categorical features. The methods further include generating a second, third and a fourth matrix. Each feature of the first matrix is then represented using a vector that includes a multiple adjustable parameters. The machine learning model can learn by adjusting values of the adjustable parameters using a combination of a loss function the fourth matrix, and the first matrix.

This invention discloses a recommendation method based on personalized differential privacy and gradient perturbation, belonging to the interdisciplinary field of computer recommendation systems and information security. The method uses an implicit feedback matrix to represent user-item interaction data, randomly initializes user and item latent factor matrices, and constructs and optimizes the objective function through Bayesian personalized ranking matrix decomposition to obtain the user latent factor matrix. In privacy-preserving training, the dataset is divided into groups for sampling, and the sampling rate is calculated by combining the privacy budget, failure probability, and iteration count of each group, and a training subset is extracted. An objective function is constructed separately for the matrix, the gradient is clipped using the L2 norm and the threshold is adaptively adjusted, and Gaussian noise is added to achieve gradient perturbation, thus iteratively updating the matrix. Finally, the matrix is ​​shared, and users calculate predicted ratings locally, rank uninterrupted items, and select items to complete the recommendation. This invention effectively improves the accuracy of the recommendation model while achieving privacy protection.

This invention discloses a method and apparatus for parallel matrix factorization tensor rank reduction reconstruction based on real trace coordinates. The method includes: converting non-uniformly sampled pre-stack 5D seismic data in the time-space domain to the frequency-space domain via Fourier transform; determining weight values ​​for mapping coordinates from real trace coordinates to regular grid points using the IDW interpolation operator; reconstructing the spatial four-dimensional data corresponding to each frequency slice: mapping the spatial four-dimensional data corresponding to the frequency slice to regular grid points using weight values; iteratively reconstructing the spatial fourth-order tensor at the regular grid points using a projection gradient descent algorithm based on BB step size and a parallel matrix factorization rank reduction algorithm; and performing a one-dimensional fast Fourier inverse transform on the reconstructed frequency-space domain seismic data with respect to the frequency variable to obtain the reconstructed time-space seismic data. This invention avoids introducing amplitude and phase errors and can efficiently and accurately reconstruct pre-stack 5D non-uniformly sampled seismic data without the need for surface elementization.

The invention discloses a medical image clustering method based on robust low-rank matrix decomposition. The method comprises the following steps: S1, constructing an anchor map; s2, pulse noise on the feature domain is extracted, an original data matrix with noise is decomposed into a matrix product W * H with clustering discrimination information and a noise matrix E, W is a basis matrix, and H is a low-dimensional representation matrix with discrimination information; s3, matrix decomposition of the anchor point matrix: performing matrix decomposition on the anchor point matrix U, and introducing maximum correlation entropy measurement to replace secondary distance measurement in the decomposition process; and S4, obtaining a loss function model, optimizing the loss function model to finally obtain a matrix H with discrimination information, and performing k-means on the matrix H to obtain a clustering result of the medical image. According to the method, the robustness of the model can be kept when pulse noise on a feature domain and outliers on a sample domain are faced, and various kinds of noise can be processed at the same time.

The present disclosure provides methods, systems, and apparatuses including encoding on a computer storage medium a computer program for presenting a new machine learning model architecture. In some aspects, the method includes obtaining a training dataset having a plurality of training samples, the training dataset including feature variables and output variables. A first matrix is generated using the training dataset, the first matrix being a sparse representation of the training dataset. Generating the first matrix can include generating a categorical representation of numerical features and an encoded representation of categorical features. The method further includes generating a second matrix, a third matrix, and a fourth matrix. Each feature of the first matrix is then represented using a vector including a plurality of tunable parameters. The machine learning model can be learned by adjusting values of the tunable parameters using a combination of a loss function, the fourth matrix, and the first matrix.