41 results about "Nonnegative matrix factorisation" patented technology

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 a gas detection precision improving method based on a multi-algorithm fusion framework. A photonic crystal resonant cavity and a tunable band-pass filtering structure are integrated in a Fourier transform spectrometer light path; collecting a wide-spectrum light intensity signal, and constructing a spectral signal database in combination with the enhancement characteristic and the filtering characteristic of the resonant cavity; zero calibration and dynamic baseline deduction operation are carried out, and self-adaptive variational mode decomposition and wavelet transform are adopted to carry out signal denoising on the spectral signals; an environment compensation model is established, and an Arrhenius type correction factor is adopted to suppress water vapor cross interference; constructing an RLS and fuzzy control combined hybrid adaptive filter; separating aliasing spectral signals by adopting a non-negative matrix factorization algorithm; establishing a quantitative relation model; in the online concentration prediction process, the initial concentration value is subjected to recursive optimization by using a hybrid adaptive filter, the deviation is continuously corrected through an environment compensation model, and an accurate concentration value is output. The method can significantly improve the accuracy and stability of multi-gas detection.

This invention relates to the field of resource scheduling, and proposes a task scheduling optimization method and system based on dynamic task profile modeling. The method includes the following steps: real-time collection of task metadata of the target task and computing node operation data of the intelligent computing center, and standardization processing; extraction of discriminative features from the standardized task metadata; input of the discriminative features into the dynamic task profile modeling model to generate a dynamic task profile; wherein the dynamic task profile modeling model is configured with a spatiotemporally coupled tensor model for multi-dimensional representation of the target task, and an online non-negative matrix factorization algorithm for dynamically updating the feature matrix of the discriminative features; based on the standardized computing node operation data and the dynamic task profile, a Pareto optimal solution set is generated based on a multi-objective optimization algorithm, and the task scheduling optimization scheme is output.

The application provides a non-negative matrix factorization community detection method based on latent structure and attribute, and relates to the field of community network detection. The method mainly solves the problem that the existing method does not sufficiently utilize community network information. The method comprises the following steps: firstly, preprocessing community network data, and representing information in the network; calculating an adjacency matrix, an attribute matrix and a latent structure matrix; establishing a community detection model mainly based on direct topological structure and attribute information and supplemented by latent structure information; calculating an iterative updating rule; setting the number of iterations, initializing a community member matrix, two community-community matrices and the latent structure matrix, and adjusting a weight parameter; performing iteration to obtain a target matrix; and finally discovering a community according to the iteration result community member matrix.

A radar system for vital signs monitoring such as at least heart rate or breathing and a method for mitigating random body movement and external interference effects in vital signs monitoring such as at least heart rate or breathing by employing a radar system that is configured for providing in-phase and quadrature signals from reflected and received radar waves are provided. The problem of random body movement effects and external interference affecting vital signs monitoring is at least mitigated by decomposing and filtering a reconstructed displacement signal using a time and frequency analysis technique such as Short-Time Fourier Transform (STFT), followed by a Non-negative Matrix Factorization (NMF) operation. This decomposition allows the identification of time and frequency basis components containing the random body movement interference. Hence, the filtered signal can be reconstructed after removing the random body movement components, thus enabling reliable and robust vital-signs parameter estimation.

The application discloses a coal water-structure state recognition method based on terahertz spectroscopy, and relates to the technical field of material property characterization. The method prepares a characteristic state coal sample containing raw coal, collects a terahertz frequency band transmission spectrum and constructs a difference spectrum; three typical mechanism spectra of water-induced polarization absorption, structure scattering enhancement and cavity polarization nonlinearity are extracted through non-negative matrix decomposition to construct a frequency spectrum-mechanism-structure mapping model; the spectrum of a coal sample to be measured is collected and a difference spectrum is constructed, and after solving an activation weight vector, the difference spectrum is matched with characteristic samples to inverse the water-structure state thereof. The method can realize nondestructive and accurate recognition of the water-structure state of coal, adapt to disaster precursor warning such as water inrush in coal mines, and expand the engineering practical value of terahertz technology.

This invention relates to the field of power load decomposition technology, and discloses a non-intrusive power load decomposition method and system based on multimodal feature learning. The method involves: synchronously collecting power parameter data from smart meters, environmental parameter data from environmental sensors, and user equipment usage behavior data to obtain multimodal load monitoring data; extracting cross-modal features through the non-negative matrix decomposition layer of a first equipment status recognition model to obtain a multimodal fusion feature vector; identifying load patterns through the first decomposition layer of the first equipment status recognition model to obtain a first decomposed load matrix; performing clustering optimization through the second decomposition layer of the first equipment status recognition model to obtain a second load decomposition matrix; and performing dynamic fuzzy decision-making based on the second load decomposition matrix to obtain the equipment operating status identification result. This invention overcomes the limitations of traditional methods that rely solely on a single power signal, achieving high-precision and highly robust non-intrusive power load decomposition.

This invention discloses a method for detecting the aging state of asphalt pavement by integrating hyperspectral imagery and deep learning, relating to the field of deep learning technology. The method includes: acquiring and preprocessing hyperspectral images of the target road section to obtain surface reflectance images; dynamically identifying and masking vehicles and their shadows in the images using a spectral angle matching method to extract effective pixel spectral datasets; performing hybrid pixel decomposition on the effective pixel spectral datasets using a non-negative matrix factorization algorithm to obtain asphalt endmember spectra and aggregate endmember abundance maps; inputting the asphalt endmember spectra into a pre-trained oxidation degree assessment model to output asphalt oxidation degree values; overlaying the aggregate endmember abundance map with near-infrared band images and inputting it into a pre-trained exposure rate segmentation network to output a binary mask of aggregate exposure areas and calculate the aggregate exposure rate. This method solves the technical problem that existing technologies cannot quickly and non-destructively detect the aging state of asphalt pavement.

This invention relates to the field of biomedical image processing technology, and discloses a method and system for quantitative analysis of hypothalamic immunofluorescence images. The method includes: performing spectral unmixing on multispectral fluorescence microscopy images based on a nonnegative matrix factorization algorithm to obtain a clean signal distribution map; using Gaussian Laplace filtering and watershed transform to achieve cell detection and segmentation; performing affine and B-spline registration between slice images and standard brain atlases to generate regions of interest masks for neural nuclei; using a local background adaptive correction strategy to perform fluorescence quantification and positive determination; and calculating Pearson correlation coefficient and Manders overlap coefficient to achieve colocalization analysis. The system includes a spectral unmixing module, a cell detection and segmentation module, an atlas registration and region recognition module, a fluorescence intensity quantification module, and a colocalization analysis and statistical output module.

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 present application relates to the technical field of computer vision, and relates to a 3D Gaussian sputtering stylization method, system, device and storage medium; wherein the 3D Gaussian sputtering stylization method comprises the following steps: obtaining a style image and a three-dimensional Gaussian scene to be stylized; performing non-negative matrix decomposition on the style image to extract an initial color basis matrix and an initial coefficient matrix of the style image; constructing a color orthogonal decoding model according to the initial color basis matrix and the initial coefficient matrix, and recalculating color parameters of each Gaussian primitive of the three-dimensional Gaussian scene; performing style transfer optimization on the Gaussian primitive, updating trainable parameters until an iteration termination condition is reached; and determining a stylized three-dimensional Gaussian scene according to updated geometric attribute parameters and color attribute parameters. The present application can solve the problems of color incoordination and dirty color during 3D Gaussian sputtering stylization.

This invention relates to the field of intelligent identification technology for swine diseases, and discloses a method for identifying and making decisions about swine diseases by integrating body temperature and image features. The method separates the independent basis components of each disease in a mixed infection by performing non-negative matrix decomposition on the body temperature time spectrum. It then uses the Hungarian algorithm to establish a cross-time point correspondence between the basis components, and uses an optimal transmission algorithm to associate lesion instances with the body temperature basis components to generate a feature sequence of disease components. Finally, it uses a time-series perceptual graph neural network to perform independent inference on the symptom evolution time-series graph, outputting the disease type identification confidence and disease stage inference confidence for each pathogen. Based on a multi-pathogen joint decision rule base, it generates a coordinated treatment plan.

This invention discloses a method for identifying the physiological state of adult product users based on multi-sensor data acquisition, relating to the fields of intelligent health monitoring and biosignal processing technology. This invention uses non-negative matrix factorization (NMF) technology to decompose multi-channel non-steady-state electromyographic signals collected from the pelvic floor and core muscle groups into multiple muscle coercive elements and their temporal activation coefficients. Each coercive element represents a fixed pattern of muscle cooperating under neural drive, while the activation coefficient reflects the change of this pattern over time. Sexual arousal, as a specific neurophysiological process, induces the activation of coercive elements with specific spatiotemporal patterns. By tracking the activation and evolution of specific coercive patterns, the neural control fingerprint reflecting sexual arousal is extracted from the original signal contaminated by motion noise. This effectively distinguishes between ordinary muscle contractions caused by device use and physical activity and autonomous neuromuscular activities related to sexual responses, overcoming the defects of susceptibility to interference and insufficient specificity.

The invention discloses a 4D wireless spectrum map reconstruction method in a dynamic environment, and the method comprises the steps: carrying out the modeling of missing spectrum data and mixed noise into a four-dimensional tensor model according to the four dimensions of two-dimensional space, frequency and time; in order to reduce the degree of freedom of the tensor model, the model is decoupled into an incomplete spectrum map corresponding to a single radiation source through a non-negative matrix factorization algorithm. Further considering the problem of mixed noise influence and calculation complexity, proposing an optimization function based on a dynamic threshold value and a fast algorithm combining random singular value decomposition and Nesterov accelerated gradient, and solving under a non-negative matrix factorization algorithm framework; and finally, reconstructing the frequency spectrum map corresponding to each radiation source by adopting parallel computing to obtain a complete 4D frequency spectrum map. According to the method, an accurate 4D spectrum map can be obtained based on spatial sparse sampling in a dynamic time-varying environment, and the high-precision spectrum situation sensing requirement in a complex noise environment is met.

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 present application relates to the technical field of image data processing, and more particularly to a hazardous waste packaging integrity detection and early warning method, which comprises the following steps: collecting multispectral image data of the surface of a hazardous waste packaging container, and obtaining reflectance spectral data by preprocessing; screening the reflectance spectral data to locate suspicious areas; using non-negative matrix factorization to spectrally unmix the reflectance multispectral data of the suspicious areas to obtain an abundance map containing a leakage endmember and a container background endmember; extracting the abundance map of the leakage endmember, calculating a risk quantification index thereof, and forming time series data reflecting the development of the leakage; inputting the time series data into a long short-term memory network model for prediction to obtain the future trend of the risk quantification index; and generating an early warning signal according to the future trend. The present application integrates leakage physical prior knowledge into the algorithm model, and solves the problems of inaccurate early-stage weak leakage identification and high false alarm rate of traditional methods.

The present application relates to the technical field of learning style recognition, and provides a learning style recognition method, system, medium and equipment based on deep learning, which comprises the following steps: obtaining learning behavior characteristics of a learner, and obtaining embedded features through preprocessing; gradually decomposing the embedded features into non-negative matrix factorization feature through multi-level non-negative matrix factorization operation; and fusing the embedded features and the non-negative matrix factorization feature through an adaptive attention fusion mechanism to obtain fusion features; wherein the multi-level non-negative matrix factorization operation adopts a multi-level decomposition structure, and the latent factors are increased layer by layer; and predicting the learning style through a deep classification network based on the fusion features. The accuracy of learning style recognition is improved.

This invention discloses a differentiated load profile generation method based on load response characteristic clustering, relating to the field of power system operation and dispatching technology. It addresses the problem of inaccurate load profile generation in existing methods. The method includes the following steps: acquiring historical load data and extracting electricity consumption patterns from the historical load data to obtain electricity consumption pattern curves; acquiring response characteristic parameters; decomposing the daily load curves of each node using a non-negative matrix factorization method based on the electricity consumption pattern curves to obtain the load composition coefficients of each node; constructing multi-dimensional feature vectors and clustering the nodes based on these feature vectors to divide them into multiple node groups; and for each node group, constructing a load profile optimization model and generating a load profile curve adapted to that node group by solving the model. This invention obtains accurate load profiles by constructing a complete process scheme for load characteristic analysis, user clustering, and profile generation.

The application discloses a crawler behavior detection method and device, equipment and a storage medium, and relates to the technical field of data leakage prevention. The method comprises the following steps: obtaining original data of user access behavior; processing the original data according to a preset data processing rule to obtain standard data, and performing vectorization processing on the standard data based on a preset vectorization rule to obtain a behavior feature vector; obtaining a corresponding normalized similarity matrix based on the behavior feature vector, and performing sparse non-negative matrix factorization to obtain an initial crawler behavior screening result; and using a preset BiLSTM+Attention network to predict the initial crawler behavior screening result to obtain a crawler behavior prediction result. In this way, the application can obtain a behavior feature vector corresponding to data based on a preset vectorization rule, and can extract deep access behavior information. The problem of large access traffic and no data label can be solved by sparse non-negative matrix factorization, and then secondary prediction is performed through the preset BiLSTM+Attention network, so that the result accuracy can be improved.

The application discloses a spatial domain identification method and device for biological spatial transcriptome slices, comprising the following steps: constructing a spatial neighbor cell network according to the spatial position information of cells in a noisy spatial transcriptome slice; constructing a gene expression cell network according to the gene expression data in the preprocessed noisy spatial transcriptome slice; performing joint feature decomposition on the spatial neighbor cell network and the gene expression cell network through non-negative matrix decomposition to obtain spatial features and gene expression features; constructing a cell affinity graph according to the spatial features and the gene expression features; and performing unsupervised clustering of cells in the noisy spatial transcriptome slice based on the leiden algorithm according to the cell affinity graph to realize spatial domain identification of the noisy spatial transcriptome slice. The application converts the analysis and spatial domain identification of the noisy spatial transcriptome slice into a network clustering problem, further reduces the influence of noise existing in the gene expression data from a global level, and realizes accurate spatial domain identification.

The invention relates to the technical field of computer vision, and discloses an image forgery detection model interpretability analysis method and system, and the method comprises the steps: extracting the middle layer features of an image forgery detection model to be interpreted, and generating a semantic feature map based on the non-negative matrix decomposition of sparsity constraint; performing feature importance pre-screening on the semantic feature map to obtain an important feature map; for each important feature map, positioning a high activation area and extracting an image block, and analyzing a dominant frequency component and a bandwidth of the image block corresponding to the high activation area; constructing a band elimination filter in the frequency domain of the original image, generating a disturbance image, and endowing each feature with an initial weight based on a decision distance; and performing grouping fusion on the features according to the symbols of the initial weights, and respectively generating final visual saliency evidence graphs which support counterfeiting and reality. According to the invention, through combination of the non-negative matrix factorization and the frequency domain disturbance shielding method, the interpretability of the image forgery detection model can be effectively improved.

This invention discloses an image specular highlight removal method based on weakly supervised learning, comprising the following steps: First, the highlight region of the image is decomposed using a sparse non-negative matrix, and a highlight-free image is cropped from the non-highlight region as training data; then, the training data is input into three joint training modules connected end-to-end to perform highlight generation, highlight removal, and image reconstruction tasks respectively, and a recurrent generative adversarial network (RGAN) architecture is used to train the network and generate the final highlight-free image. This image specular highlight removal method based on weakly supervised learning completes highlight removal by jointly training the highlight generation, highlight removal, and reconstruction modules, using a RGAN architecture. During training, the loss function of subsequent modules is fed back to the preceding module, enabling training to be completed using only the highlight image and achieving good highlight removal results. Compared with traditional algorithms and existing weakly supervised learning methods, it has the advantages of simple operation and excellent highlight removal effect.

A processing apparatus for performing non-negative matrix factorization to measured profiles of X-ray powder diffraction includes a measured profile acquiring section for acquiring a plurality of measured profiles; a decomposition section for applying non-negative matrix factorization to the measured profiles and calculating base profiles; an index calculating section for acquiring the base profiles and calculating indexes based on unevenness of the base profiles; a base profile classifying section for classifying the base profiles into a plurality of groups based on the indexes; and a base profile correcting section for performing correction based on the indexes on at least one of the plurality of groups and calculating a corrected base profile.

The invention relates to the technical field of anomaly detection of multi-auditory-angle data, in particular to an acoustic sensing array data anomaly detection method based on multi-auditory-angle self-adaptive sparse semi-nonnegative matrix factorization, which comprises the following steps: S1, assuming that a piece of normal array data acquired by an acoustic sensing array consisting of nv acoustic sensors represents vth auditory-angle data, 1 < = v < = nv, the dimension is 1 * m, and training an anomaly detection network W based on multi-auditory-angle sparse semi-nonnegative matrix factorization and Gaussian distribution estimation by using batch normal array data to obtain an anomaly detection model M; s2, the to-be-detected array data y is input into the anomaly detection model M, an anomaly detection result R is obtained, the to-be-detected array data collected by the sensing array is represented, the dimension of the to-be-detected array data is 1 * m1, the to-be-detected data collected by the vth sensor is represented, and v is larger than or equal to 1 and smaller than or equal to nv. According to the method, the problems of multi-auditory-angle information fusion and information redundancy in multi-auditory-angle data processing of a current acoustic detection method can be effectively solved.

The invention discloses a multi-sample joint deconvolution method applied to a GC-MS technology, and relates to the field of chemical analysis, and the method comprises the following steps: carrying out time alignment on a GC-MS data matrix of each sample; performing non-negative matrix factorization on the aligned matrix of each sample so as to obtain a shared mass spectrum matrix under the initial cycle index under the candidate fraction, a peak shape function parameter matrix of each sample and a specific elution curve matrix of each sample; performing alternate iterative optimization on each matrix under the initial cycle index to obtain a final matrix under the candidate group score; performing component sorting on the final matrix; obtaining index values corresponding to the candidate group scores according to the final matrix; and determining an optimal group score according to the comprehensive score of each group score, and outputting a deconvolution result. According to the method, multi-sample information can be integrated, the analysis stability under the overlapping peak condition is improved, and collaborative optimization of retention time and deconvolution is realized.

The invention provides an operational amplifier conducted electromagnetic interference analysis method and device, a medium and a program. The method comprises the steps that firstly, conducted electromagnetic interference mixed time domain signals of an operational amplifier are collected; converting the time domain signal into a frequency domain signal; substituting the frequency domain signal as an input matrix into a conducted electromagnetic interference mixed frequency domain signal non-negative decomposition model, introducing a time sequence sparse method through spectrum smoothing, separating the conducted electromagnetic interference mixed frequency domain signal, and obtaining each component frequency domain signal of the operational amplifier; the mixed frequency domain signal non-negative decomposition model is constructed based on regularization; and finally, based on each component frequency domain signal, reconstructing a corresponding conducted electromagnetic interference component time domain signal. According to the method, each component frequency domain signal corresponding to each noise source is separated from the mixed frequency domain signal based on the non-negative matrix factorization method, and the defects that a traditional separation method is complex in process and low in precision are overcome.

This invention discloses a method for imputing missing values ​​in scRNA-seq data based on nonnegative matrix factorization, belonging to the technical field of deep learning and data imputation methods. This invention solves the problem of low accuracy in imputing missing values ​​in single-cell RNA sequencing data using existing methods. The main technical solution of this invention is as follows: Step 1: Filter cells based on gene data to obtain RNA sequencing data of the remaining cells after filtering; normalize, screen genes, and perform logarithmic transformation on the RNA sequencing data of the remaining cells after filtering to generate matrix X; Step 2: Decompose matrix X to obtain a feature matrix and a coefficient matrix; Step 3: Construct an input matrix for an autoencoder based on the feature matrix, use the input matrix as the input of the autoencoder, and output the imputed feature matrix through the autoencoder; Step 4: Perform decomposition, logarithmic reduction, and inverse normalization on the imputed feature matrix in sequence to obtain the imputed result. This invention can be applied to RNA sequencing data imputation.

The application relates to the technical field of miRNA-disease, in particular to a miRNA-disease correlation prediction method based on heterogeneous subnetwork fusion multi-source features, which comprises the following steps: constructing a disease similarity network, a miRNA similarity network and a miRNA-disease interaction network, extracting a disease topological structure feature vector and a miRNA topological structure feature vector by using RWR; performing non-negative matrix decomposition on a disease-miRNA correlation matrix; fusing the disease topological structure feature and a miRNA low-rank space feature; fusing the miRNA topological structure feature and a disease low-rank space feature; obtaining a disease and miRNA heterogeneous subnetwork; and reconstructing the disease-miRNA correlation after node embedding by using GCN on the miRNA heterogeneous subnetwork and the disease heterogeneous subnetwork respectively. The application solves the problem that the existing method can only obtain features from similarity data before graph embedding and cannot comprehensively reflect diseases or miRNAs.

The invention discloses a hyperspectral image domain adaptive method and device based on domain feature decoupling and conditional diffusion. The method comprises the following steps: acquiring a hyperspectral image to be classified; inputting the hyperspectral image into a classification network, and enabling the classification network to execute the following steps: extracting an initial feature map of the hyperspectral image by using a domain specific encoder; inputting the initial feature map into a feature decomposition module, and obtaining domain specific features through deep non-negative matrix decomposition; inputting the domain specific feature, the initial feature map and the random time step into a feature diffusion sub-network to obtain a domain-invariant high-dimensional semantic feature; and mapping the high-dimensional semantic features into feature vectors by using a feature head, and classifying the hyperspectral images based on a nearest neighbor principle by calculating Euclidean distances between the feature vectors and category prototypes of all categories. According to the method, through domain feature decoupling and a conditional diffusion mechanism, the classification precision in a cross-domain few-sample scene is remarkably improved, meanwhile, the model robustness and generalization ability are enhanced, and the implementation process is simplified.

This invention discloses an integrated method for identifying and functionally annotating metaprograms in tumor single-cell transcriptomes, belonging to the fields of bioinformatics, tumor biology, and single-cell transcriptome sequencing. This invention aims to address the problems of unstable metaprogram identification, significant technical artifact interference, and fragmented functional annotation in existing technologies. Based on multi-rank nonnegative matrix factorization, it systematically extracts co-expression modules at different scales from the tumor single-cell expression matrix. Robustness screening ensures the consistency and reliability of the obtained metaprograms under multiple factorizations and cross-sample conditions. Furthermore, it performs systematic functional analysis of the metaprograms to reveal the multi-dimensional functional states of tumor cells. This method improves the stability and biological interpretability of tumor metaprogram identification, providing a new technical means for tumor heterogeneity research and the discovery of precision therapeutic targets.