43 results about "Matrix similarity" patented technology

The embodiment of the invention provides a multi-view clustering method and device, equipment, a storage medium and a program product, and relates to the field of financial science and technology. Obtaining a view data set corresponding to each view; based on the similarity between the feature elements in each view data set, constructing a similar matrix of each view data set; multiplying the similar matrixes of the plurality of view data sets to obtain a consistency matrix, and reconstructing the similar matrix of each view data set based on the consistency matrix to obtain a reconstructed similar matrix; and performing multi-view clustering based on the plurality of reconstructed similar matrixes to obtain a clustering result, the clustering result being used for indicating a clustering label corresponding to each sample. Through the above mode, the complementary information of the consistency matrix is utilized to maintain the stability of the clustering boundary, the misjudgment and missed judgment caused by the missing of the sample information in the view are reduced, and the accuracy of the multi-view clustering result is improved.

The application discloses a kind of passive cross-domain adaptive image classification methods based on multi-modal pre-training model, it is related to transfer learning technical field, first by quantitative estimation target domain data on source model existing category similarity matrix;Then the category similarity matrix of source model is introduced into CLIP (multi-modal pre-training model), directional generation confusion category text feature is used for classification, and the corresponding classification result is obtained;Then the classification result of source model and the classification result of CLIP are weighted, and the output of two models is respectively aligned with the weighted result;Again, the category feature center library of each model is calculated according to the category similarity matrix;Finally, the projection of input image on two feature center libraries is aligned using attention network to realize the alignment of feature space.The application dynamically estimates the similarity degree between categories in downstream task from the angle of model, learns using the complementary information between models, and realizes the fine-grained classification between similar categories, and has certain explainability.

The present invention discloses an image matching method based on quadratic signature, including: obtaining a correspondence set; obtaining a first cost function based on the feature points and the neighborhood points of the feature points in the first relation set and the second feature point set; converting the first cost function into a simplified second cost function; obtaining a quadratic signature distance according to the weight vector between the feature points and the neighborhood points and the similarity matrix; obtaining a first quantization distance according to the consistency of the neighborhood topology of the vector angle and length, and obtaining a second quantization distance according to the consistency measure of the structure by the quadratic signature distance; converting the second cost function into a third cost function based on the first quantization distance and the second quantization distance; converting the third cost function into a fourth cost function, and minimizing the simplified fourth cost function to obtain an optimal correspondence set. The present invention can more accurately measure the similarity of local structures, improve the ability to remove incorrect correspondences, and retain more reliable correspondences.

The invention relates to the technical field of power grid dispatching, in particular to a power distribution network dispatching optimization method and device under participation of an electric vehicle cluster. The method comprises the following steps: collecting related data of a vehicle and a charging point through a monitoring device; the total waiting time of the vehicle is determined according to the time when the vehicle arrives at the charging point, the waiting time after the vehicle arrives at the charging point and the charging time of the vehicle at the charging point; calculating the load unbalance degree of the charging points according to the load difference of the charging points and the variation coefficient; feature vectors are constructed, the similarity of the vehicles is calculated based on the feature vectors, and a similar matrix is constructed; acquiring a cluster based on similar matrix clustering, and acquiring a penalty coefficient and conflict penalty according to similarity; and constructing a fitness function based on the total waiting time, the load unbalance degree and the conflict penalty, thereby determining a power distribution scheduling scheme. The problem of unbalanced load of the power distribution network is solved.

The invention discloses a kernel norm-based robust flexible multi-view subspace clustering method and system. The method comprises the steps of mapping original image data to a high-dimensional regeneration kernel Hilbert space through a kernel technique; learning a consistency kernel through a self-weighted multi-kernel learning strategy; introducing low-rank constraint into self-expression learning of the Hilbert space to obtain a similar matrix; and calculating in the similar matrix according to a k-means algorithm to obtain an image clustering result. According to the method, original image data are mapped to a high-dimensional regeneration kernel Hilbert space by adopting a kernel technology, so that nonlinear data are effectively processed, weights are automatically distributed for different kernels, the importance of different views is effectively considered, the influence of noise is considered, low-rank constraints are added in a similarity matrix, and the image quality is improved. Therefore, an optimal similarity matrix with a clear block diagonal structure is constructed, and a better image clustering result can be obtained.

The utility model provides a kind of power station parameter matrix analysis method, belongs to power station fault diagnosis technical field, solve how to design a kind of power station parameter matrix analysis method, solve the numerous power station parameters, data effective information density is low, lead to the problem that fault cause analysis is difficult, and the massive data stream generated occupies a large amount of storage space, by power station parameter data matrixization representation, constructs parameter data matrix, by matrix operation, the rank of matrix is as quantization index, effectively distinguishes the information density of parameter, the correlation between different parameters can be analyzed by the multi-parameter matrix constructed, not limited by parameter number and time sequence length, and by matrix similarity operation, judge the similarity of unit state, provide quick judgment basis for similar fault in later period;Dimensionality reduction is carried out to operation data matrix using matrix singular value decomposition method, retain the part of large singular value proportion, can extract fault feature, realize compressed storage, save storage space.

The invention provides an electromagnetic environment monitoring optimization point distribution method and system, and relates to the technical field of electromagnetic environment monitoring. In order to solve the problems that in the prior art, the representativeness and the accuracy of selected monitoring point positions are low, and the real environment of a region cannot be reflected, the method comprises the steps that a to-be-monitored region is divided into a plurality of grids, electromagnetic radiation indexes of the grids are obtained, and an original data matrix is constructed; determining a high-optimal index value according to an optimal index method, and performing normalization processing on the original data matrix to obtain a decision matrix; converting the decision matrix to obtain a proximity degree matrix; calculating similarity values among the grids based on the proximity degree matrix, and constructing a similar matrix; classifying the grids based on the similar matrix to obtain multiple groups of classification results; and for each group of classification results, analyzing the consistency of the same type of grids and the difference of different types of grids, determining a final classification result, and determining a final representative grid according to the final classification result. According to the method, the representativeness and the accuracy of selecting the final representative grid are relatively high.

This invention proposes a method and system for ECG clustering based on multi-lead semantic consistency, belonging to the field of ECG signal processing technology. The method includes: constructing a similarity correlation matrix for an ECG signal set using adaptive graph learning; learning the spectral representation of each lead of the ECG signal using spectral clustering on the similarity correlation matrix, whereby spectral clustering utilizes multi-lead shared spectral embedding to measure the correlation between ECG signals, obtained through an optimized minimum edge weight objective function; wherein the optimized minimum edge weight objective function includes the multi-lead shared semantic similarity matrix; after obtaining the spectral embedding matrix, constructing a spectral rotation objective function by introducing an orthogonal rotation factor matrix; merging the minimum edge weight objective function and the spectral rotation objective function into a total objective function, and optimizing the total objective function using the alternating direction multiplier method to obtain the ECG clustering result. This invention significantly improves the clustering quality through the learning of semantically consistent graphs.

The invention discloses a feature representation method for realizing multi-view image alignment based on a dynamic graph. The feature representation method comprises the following steps: 1) initializing a multi-view structure; 2) common dynamic graph learning; 3) constructing a joint optimization model based on dynamic graph regularization; 4) solving a common source representation S and a projection matrix by minimizing projection features of each view; 5) initializing the dynamic similar matrix during the first iteration; 6) expanding a matrix dimension and constructing a coefficient matrix; 7) expanding and converting the expanded target function into a matrix trace form; 8) deducing a gradient expression of a dynamic graph learning objective function; and 9) setting an iteration stop criterion of the gradient descent method, and finally outputting an optimization result. Through combined utilization of the structure information of each feature view, adverse effects caused by insufficient samples or noise interference are reduced, and the image alignment precision in severe construction environments such as dark light is effectively improved.

The present invention discloses a multi-view clustering image segmentation method and system based on embedded approximate learning. The method includes the following steps: obtaining an image multi-view data set to be subjected to image segmentation; inputting the image multi-view data set into a pre-constructed clustering model, iteratively updating the parameters of the clustering model according to an optimization objective, and when the change value of the optimization objective is less than a set threshold, obtaining an optimized clustering model; inputting the image multi-view data set into the optimized clustering model, outputting a clustering result, thereby completing multi-view subspace clustering; and performing image segmentation according to the clustering result to obtain an image segmentation result. By combining self-representation learning, robust principal component analysis technology, and Grassmann manifold space approximate learning, the present invention avoids an increase in computational cost caused by eigenvalue decomposition of a similarity matrix during the optimization process, and improves the efficiency and quality of clustering.

The invention discloses a terminal area approach mode mining method based on ASM-HAC. The method comprises the following steps: acquiring ADS-B data of an airport terminal area; the track features are screened, and the correlation degree between the track features is analyzed in combination with a Pearson's correlation coefficient; obtaining a track distance matrix by using a multi-dimensional dynamic time warping method under weighted Euclidean distance; improving a Gaussian kernel function by using adaptive parameters to obtain an ASM track similar matrix; a Laplacian matrix and a feature gap method are introduced to improve an HAC hierarchical clustering algorithm, and after the improved HAC hierarchical clustering algorithm is applied to an ASM track similar matrix, an optimal track clustering result is obtained; and evaluating the rationality of the optimal clustering number by adopting a contour coefficient. According to the method, the problem that an original hierarchical clustering method excessively depends on manual intervention to obtain the optimal clustering number is solved, the similarity measurement requirement of a track dense region and a track sparse region is balanced, the clustering process better fits the space-time heterogeneity of track distribution, a controller is assisted in mastering the characteristics of different approach modes, and control decision making is facilitated to be completed.

The present invention relates to the field of dynamic parameter identification of nonlinear systems, and particularly to a method for identifying dynamic parameters of nonlinear systems based on high-order dynamic mode decomposition. The scheme includes: obtaining the time series of the nonlinear system through experiments or numerical simulations; determining the optimal time delay of the time series according to the mutual information method; according to the Takens embedding theorem, reconstructing the original time series into a high-order time series matrix through phase space; obtaining the similarity matrix of the high-order time series system matrix; obtaining the high-order dynamic modes of the nonlinear system; judging whether the characteristic spectrum of the high-order dynamic modes is located on or close to the unit circle in the complex plane, if not, returning to the step of reconstructing the original time series into a high-order time series matrix through phase space; if so, obtaining the modal frequency and damping ratio parameters of the nonlinear system. By introducing the mutual information method and the phase space reconstruction theory, the present invention optimizes the analysis process of the existing technology, and the present invention is applicable to the identification of dynamic parameters of nonlinear systems.

The present invention belongs to the field of feature fusion technology, and discloses a method and system for the progressive fusion of multimodal medical data. The present invention uses a progressive fusion method to achieve hierarchical fusion feature vector interaction based on a cross-level attention mechanism, and uses a residual network to alleviate the problem of easy loss of shallow information due to the increase in the number of layers, so as to maximize the hierarchical information contained in the progressive fusion features in a progressive interactive manner. The present invention provides a dynamic graph learning method, which captures the nonlinear interaction relationship of features through graph adaptive learning, reduces the heterogeneity of the similarity matrix, and improves the accuracy of information transmission; in the adaptive learning process, a threshold prediction is performed for the features of each patient to prune the similarity matrix, so that the pruned similarity matrix is ​​more accurate.

This invention discloses an unsupervised crowd counting method, apparatus, and storage medium. The method includes cropping a first input image into image blocks and obtaining coarse-grained text for each image block; inputting the image blocks into a first image encoder and the coarse-grained text into a first text encoder to generate a first similarity matrix; filtering image blocks of a first target category based on the first similarity matrix and a first discriminative category similarity; obtaining fine-grained text for image blocks of a second target category and inputting it into a second text encoder to generate a second similarity matrix; filtering image blocks of a second target category based on the second similarity matrix and a second discriminative category similarity and inputting them into a second image encoder; inputting the counting text into a third text encoder to generate a target similarity matrix; and obtaining the number of people in the image based on the similarity between the target similarity matrix and the counting text. This invention eliminates the need for any manual labeling, significantly reducing annotation costs.

Apparatus (100, 120, 300, 304, 320, 330) and method (200, 350) for reducing the dimensionality of embeddings in a machine learning (ML) system. In some embodiments, original embeddings (124, 206, 378, 394, 404) from a pre-trained deep learning model (204, 322, 370, 390) are extracted for a set of data (202, 302), the original embeddings having an initial dimensionality. A set of projection vectors (128) are initialized (208, 352) with a specified dimensionality smaller than the initial dimensionality. A neural network (122, 304, 324, 340) is used to optimize the projection vectors by minimizing a loss function based on a similarity matrix (132, 134). The set of original embeddings are thereafter projected (210, 214, 360) onto the optimized projection vectors to obtain a set of reduced-dimensional embeddings (126, 216, 306, 388, 396, 406) with the specified dimensionality. A neural network of an ML system (308, 380, 392, 408) is thereafter configured using the reduced-dimensional embeddings, such as by a training operation (364) to duplicate operation of the deep learning model in a smaller latent space. The original embeddings may be single modal or multimodal embeddings.

The invention discloses an Internet of Things multi-view feature selection method and system based on similar matrix fusion, and relates to the technical field of machine learning, and the method comprises the steps: firstly splicing incomplete data of each view into a wide table according to columns, and constructing a complete matrix of each view through sampling and filling; secondly, clustering each view data for multiple times, and constructing a corresponding similar matrix; stacking all similar matrixes into tensors along a third dimension, introducing adaptive weight fusion under low-rank constraint, learning a unified common similar matrix to perform eigenvalue decomposition, and extracting eigenvectors corresponding to positive eigenvalues to form pseudo tags; in combination with a pace learning strategy, high-confidence samples are screened, and feature selection coefficients and self-paced weights of all views are jointly optimized; and calculating an importance score based on each view feature to obtain an optimal subset. Through similar matrix tensor fusion, adaptive low-rank constraint and pace learning joint optimization, feature selection of incomplete multi-view data of the industrial Internet of Things is realized.

The present invention discloses an incomplete multi-view clustering method and system based on co-regularized spectral clustering, which relates to the technical fields of computer vision and pattern recognition, and includes: obtaining an incomplete multi-view data set and converting each view into a data matrix; performing dimensionality reduction processing on the data matrix to obtain a coefficient matrix; constructing a similarity matrix based on the self-representation characteristics of the coefficient matrix, and further calculating a Laplacian matrix; obtaining a clustering indicator matrix for each view based on the Laplacian matrix; performing kernel alignment on the clustering indicator matrix to obtain a consistent clustering indicator matrix; finally constructing an objective function for incomplete multi-view clustering and optimizing and solving it to obtain an optimal consistent clustering indicator matrix; and inputting the optimal consistent clustering indicator matrix into an existing classification algorithm to obtain the clustering result of the incomplete multi-view. The present invention deeply mines the complementary and consistent information of view data in the case of incomplete multi-views, and improves the clustering effect of incomplete multi-views.

A data processing method includes determining a first similarity matrix based on a query matrix and a key matrix corresponding to a word vector sequence; filtering out a target element whose value meets a condition from each row of the first similarity matrix, and setting all elements other than the target element in each row to zero to obtain a second similarity matrix; normalizing the second similarity matrix to obtain an attention weight matrix; and determining a self-attention matrix corresponding to the word vector sequence based on the attention weight matrix and a value matrix corresponding o the word vector sequence.

The application discloses a time series polarimetric SAR cumulative change detection method based on a similarity matrix, relates to the technical field of remote sensing image change detection, and comprises the following steps: a similarity matrix of all time phase polarimetric SAR images in a time series is constructed; the similarity matrix is linearly transformed, and the maximum eigenvalue of the similarity matrix is calculated; a difference map of cumulative change of the polarimetric SAR images in the time series is calculated according to the maximum eigenvalue; and the cumulative change detection result of the polarimetric SAR images in the time series is obtained by segmenting the difference map. The time series polarimetric SAR cumulative change detection method based on the similarity matrix provided by the application does not repeatedly calculate a difference map, and compared with a traditional detection method of continuously detecting two changes in a time series, the detection precision and operation efficiency are improved.

The invention belongs to the field of point cloud registration, and discloses a point cloud registration method and system, and the method comprises the steps: extracting plane points, carrying out the clustering of the plane points according to the normal vector of each plane point, dividing the planes in a plane group, and distinguishing different parallel plane layers in the same direction; normal vector included angles between different plane groups are calculated, the plane groups forming similar included angles are searched, rotation matrixes are calculated, similar matrixes are combined, planes in the groups are associated, and the planes of the corresponding plane groups correspond to one another; and calculating an optimal translation vector for each candidate rotation matrix and each plane corresponding condition, evaluating the quality of each calculated rotation matrix and translation vector, and selecting an optimal solution. According to the method, automatic and accurate extraction of the parallel plane group is realized under the condition that the plane direction does not need to be manually stipulated; registration is constrained through a plane geometrical relationship, similar planes are effectively distinguished, mismatching caused by too high geometrical similarity is avoided, and correct matching and alignment of parallel plane groups are achieved.

The invention discloses a bearing fault diagnosis method based on a cross-modal manifold density topological space, solves the problems that abnormal values are sensitive and non-linear modes of different density regions are difficult to capture under the condition of high dimension and small sample size, and effectively improves the accuracy of fault diagnosis. The specific implementation process is as follows: (1) constructing a relative neighborhood order module through Euclidean space, effectively identifying outliers with an asymmetric neighborhood relationship, constructing a cross-modal robust similar matrix in combination with a shared neighbor topological structure and local density distribution, and reducing the influence of noise and outliers; (2) constructing an objective function of a cross-modal manifold density topology model by adopting a divide-and-conquer strategy and considering local distribution characteristics of the data, and performing theoretical derivation on the objective function to obtain an analytical solution in a cross-modal manifold density topology projection direction; and (3) directly obtaining cross-modal manifold density topology fault features with good discrimination from the randomly selected bearing fault test sample through the cross-modal manifold density topology projection direction, and inputting the fault features into the classifier to obtain a final bearing fault diagnosis result. Compared with the prior art, the bearing fault diagnosis method is more accurate and robust.

The present invention discloses an unsupervised feature selection method, apparatus, device, and storage medium, which relate to the field of dimensionality reduction in machine learning, and solve the problem that the finally selected features in feature selection are not the optimal features due to the influence of bad features such as noise and redundancy contained in the original data. The unsupervised feature selection method specifically includes: obtaining the initial feature data of each of the n target objects, determining the first similarity between any two target objects, and constructing a first similarity matrix; constructing a second similarity matrix based on the first similarity matrix, performing iterative learning on the second similarity matrix to obtain a target similarity matrix and a corresponding target projection matrix, and determining target feature data according to the initial feature data and the target projection matrix. The present invention integrates the two processes of constructing a similarity matrix and feature selection into a unified framework, and can obtain an optimal similarity matrix.

The application discloses a multi-label feature selection method based on sparse learning coupled mutual information, and comprises the following steps: inputting a feature matrix X, a label matrix Y and hyperparameters alpha, beta, gamma and delta, selecting a feature number k, and initializing a label correlation matrix Z and a feature correlation matrix W; calculating a diagonal matrix A and a similarity matrix S according to the feature matrix X, and calculating a graph Laplacian similarity matrix L of the feature matrix X x ; updating the label correlation matrix Z and the feature correlation matrix W through a target function, iterating n times, and obtaining an updated label correlation matrix Z n and an updated feature correlation matrix W n after the target function reaches a convergence condition; and obtaining the selected k features according to the 2-norm of W n . The application adopts the multi-label feature selection method based on sparse learning coupled mutual information, and the effectiveness of the method is proved by experiments; and the method simultaneously solves the suboptimal solution problem caused by a random initialization strategy in the sparse learning method.

The application discloses an answer selection method considering the space-time dependent relationship of questions and answers, and the steps include: 1. constructing the question and answer data and performing data preprocessing; 2. using a BERT model to obtain a word-level similarity matrix of the question and answer pair; 3. splicing the similarity matrices of multiple answers under the same question thread to obtain a space-time tensor of the question and answer pair; and 4. using a ConvLSTM model to predict the matching degree of the question and answer pair. The application uses the BERT to obtain the word-level similarity matrix of the question and answer pair with implicit semantic correlation, constructs the space-time tensor of the question and answer pair based on the word-level similarity matrix, and learns the space-time dependent relationship information in the question and answer data through the ConvLSTM model, so that the selection of the answer to the question is finally realized, and thus the best answer with the highest matching degree to the question can be accurately recommended.

The present invention discloses a map matching method and device based on trajectory topology, comprising the steps of S1: based on road network data, using three types of nodes, namely, entrances and exits, hubs, and connections, to model urban space and construct a city topology map; S2: using an original vehicle GPS record sequence to obtain a set of candidate points corresponding to the vehicle trajectory topology; S3: calculating a spatial similarity function between candidate points based on the original vehicle GPS record sequence and the city topology map to obtain a weighted candidate map; S4: establishing and calculating a static similarity matrix based on the weighted candidate map to obtain a weighted similarity matrix; S5: calculating a local optimal path, counting votes for all candidate points on the local optimal path, and connecting the candidate points with the highest number of votes to obtain a trajectory topology; and S6: performing map matching through the trajectory topology. The present invention uses the city topology map as a carrier for expressing trajectory data, simplifies the expression of the trajectory, and improves the accuracy and efficiency of matching the trajectory data with the map.

The present invention discloses a method and device for video object segmentation using implicit motion compensation, which belongs to the field of video processing technology and is used to solve the technical problems of slow motion compensation network calculation, content redundancy, and low accuracy in unsupervised video object segmentation in existing video processing. The method comprises: extracting consecutive frames from a video to be processed, and extracting features from the consecutive frames using a shared feature encoder to obtain embedded features; calculating a similarity matrix on the embedded features to obtain a similarity matrix; normalizing the similarity matrix to obtain attention-enhanced features; skipping the embedded features of each layer to connect to a higher layer for prediction guidance to generate a final feature; aligning the final feature of the central frame with the final feature of each adjacent frame to obtain an aligned feature; fusing the aligned features to obtain a fused feature; performing segmentation prediction on the fused feature to obtain a predicted segmentation mask, so as to segment the video object using the predicted segmentation mask.

The invention relates to the field of nonlinear system dynamic parameter identification, in particular to a nonlinear system dynamic parameter identification method based on high-order dynamic modal decomposition. The scheme comprises the following steps: obtaining a time sequence of a nonlinear system through test or numerical simulation; determining the optimal time delay amount of the time sequence according to a mutual information method; reconstructing an original time sequence into a high-order time sequence matrix through a phase space according to the Tarkens embedding theorem; obtaining a similar matrix of the high-order time sequence system matrix; obtaining a high-order dynamic mode of the nonlinear system; judging whether the high-order dynamic modal characteristic spectrum is located in or close to a unit circle in a complex plane, and if not, returning to the step of reconstructing the original time sequence into the high-order time sequence matrix through the phase space; and if yes, obtaining modal frequency and damping ratio parameters of the nonlinear system. By introducing a mutual information method and a phase-space reconstruction theory, the analysis process in the prior art is optimized, and the method is suitable for nonlinear system dynamic parameter identification.

The application discloses a spoon defect detection method and system, electronic equipment and computer storage medium, and relates to the field of machine learning. The method comprises the following steps: acquiring a spoon image; constructing a spatial self-representation tensor according to the spoon image; constructing a multi-view clustering model based on rank approximation and sparse constraint according to the spatial self-representation tensor; solving a target function of the multi-view clustering model to obtain a similarity matrix of multi-view data; and performing clustering by using a spectral clustering algorithm according to the similarity matrix to obtain a spoon defect detection result. The application can improve clustering performance.

The invention provides a raster data spectral clustering method and system integrating spatial constraints. The method comprises the following steps: acquiring different attribute layers of raster data to be classified; determining a clustering number based on the attribute layer, clustering the raster data through a preset algorithm, and calculating a similarity threshold; constructing a clustering network and generating an initial similar matrix based on the similarity threshold according to the grid attributes and the spatial adjacency relationship; reducing a similar graph constructed by the initial similar matrix and removing abnormal nodes to obtain a pruned similar matrix; and inputting the trimmed similar matrix as a parameter into a preset spectral clustering algorithm for clustering to obtain a clustering result. On the basis of an original spectral clustering method, when a similar graph between data nodes is created, spatial constraint is applied to obtain a similarity threshold value, edges with small similarity are deleted to reduce the spectral clustering operand, the defect that an existing spectral clustering method cannot consider spatial continuity is overcome, and the spectral clustering efficiency is improved. And the spatial continuity and the attribute similarity of the raster data clustering result are considered.

The application provides an electromagnetic environment monitoring optimization distribution method and system, and relates to the technical field of electromagnetic environment monitoring. In view of the low representativeness and precision of the selected monitoring points in the prior art, and the problem that the real environment of the region cannot be reflected, the application divides a region to be monitored into multiple grids, obtains electromagnetic radiation indexes of the grids, and constructs an original data matrix; the high optimal index value is determined according to the optimal index method, the original data matrix is normalized to obtain a decision matrix; the decision matrix is converted to obtain a closeness matrix; the similarity values between the grids are calculated based on the closeness matrix to construct a similarity matrix; the grids are classified based on the similarity matrix to obtain multiple sets of classification results; for each set of classification results, the consistency of the same type of grids and the difference of different types of grids are analyzed to determine a final classification result, and the final representative grid is determined according to the final classification result. The representativeness and precision of the selected final representative grid are high.