Image classification method, device and medium based on graph embedding extreme learning machine

Abstract

The present invention discloses an image classification method, device, and medium based on a graph embedding extreme learning machine. The method comprises the following steps: Step S1: obtaining and preprocessing a data set to obtain a training set and a test set; Step S2: using a multi-layer graph embedding extreme learning machine algorithm to perform model training on the training set to obtain a classification model; and Step S3: inputting the test set into the classification model to obtain an image classification result. In conjunction with the graph embedding method, the present invention proposes a new extreme learning machine autoencoder. By redefining the data feature reconstruction error function in the feature space, the present invention increases feature aggregation capabilities based on label information, extracts local features, and enhances feature representation capabilities. A graph-based penalty term is embedded in the loss function, which reduces intra-class spacing and increases inter-class spacing in the ELM feature space, extracts global feature structure information, and thus finds a more effective discrimination boundary in the feature space.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to an image classification method, device and medium based on a graph embedding extreme learning machine. Background Art

[0002] Currently, wind turbines are primarily concentrated in high-latitude regions and are often installed in hilly and mountainous areas. These harsh operating environments and prolonged exposure to cold and humid conditions make icing on wind turbine blades very likely. Icing not only affects power supply stability but also shortens the turbine's operating life, resulting in economic losses. Wind turbine blades are a key component of wind turbines, making the research of real-time blade icing fault detection methods of great significance.

[0003] Most research on wind turbine blade icing prediction involves three aspects: wind tunnel testing, numerical calculations, and analysis of wind turbine SCADA system data. Simulated wind tunnel experiments predict blade icing by measuring the difference between the wind turbine's theoretical and actual power. Numerous icing detection studies use numerical methods to calculate the basic parameters of ice formation mechanisms and conduct related fault detection. Machine learning-based SCADA data analysis methods offer advantages in efficient and accurate wind blade icing prediction. Although many classification methods based on deep learning and data mining have been proposed for wind turbine blade icing prediction in recent years, these methods remain limited by technical challenges, hindering their predictive capabilities. For example, existing methods generally fail to address the inherent class imbalance and noise generated by a small number of sample data within wind turbine monitoring system SCADA data. At the feature level, existing classification methods based on traditional pattern recognition generally fail to address the weak correlation between data sample features, which results in low decision recognition rates.

[0004] In summary, there is an urgent need for an image classification method, device and medium based on graph embedding extreme learning machine to solve the problems existing in the prior art. Summary of the Invention

[0005] The present invention aims to provide an image classification method, device, and medium based on a graph embedding extreme learning machine, which fully utilizes the feature dimensionality reduction capability of each autoencoder to solve the problems of insufficient feature representation capability and incomplete feature reconstruction of existing autoencoders. The specific technical solution is as follows:

[0006] An image classification method based on graph embedding extreme learning machine, characterized by the following steps:

[0007] Step S1: Obtain the data set and preprocess it to obtain the training set and test set;

[0008] Step S2: Use the multi-layer graph embedding extreme learning machine algorithm to train the model on the training set to obtain a classification model;

[0009] Step S3: input the test set into the classification model to obtain the image classification result;

[0010] Among them, the multi-layer graph embedding extreme learning machine algorithm in step S2 is specifically: constructing multiple graph embedding autoencoders, forming information interaction with the graph embedding autoencoders in a cyclic manner, converting them into a multi-layer graph embedding extreme learning machine, and using a single graph embedding autoencoder as a child node. The data in the training set is output to the next layer of the network through the child node to obtain the weight output of the information interaction, and the multi-layer network is trained using the optimal layer-by-layer method to obtain the final feature output.

[0011] Preferably, the graph embedding autoencoder comprises the following steps:

[0012] 1) Get the data in the training set And set the activation function; where X=[x1,...,x N ];x i represents the i-th sample in the data, x i ∈R d , d represents the feature dimension of each sample; T = [t1,…,t N ];t i represents the i-th category in the data, t i ∈R f ; N represents the amount of data; f represents the feature dimension of each category;

[0013] 2) Calculate the reconstruction error weight w ij ; Generate graph embedding matrix S; Initialize weight parameter A, A∈R D×L , D represents the diagonal matrix of error weights, L is the number of hidden layer nodes;

[0014] 3) Calculate the optimized output weight β, the expression is as follows:

[0015] β=[H T DH+(I L +λS) / C] -1 H T WX;

[0016] Among them, I represents the identity matrix, λ represents the balance parameter, C represents the regularization coefficient, and H represents the hidden layer output matrix.

[0017] Preferably, the activation function is the sigmoid activation function g(x), which is expressed as

[0018] Preferably, the reconstruction error weight wij The calculation expression is as follows:

[0019]

[0020] in, Represents based on sample x i The set of all features of category t in the neighborhood, set the number of neighbor points to P, and ω is the median of the square distance of all paired input data.

[0021] Preferably, the expression of the computation graph embedding matrix S is as follows:

[0022] S=S w S b =H(D w -A w )H T H(D b -A b )H T ;

[0023]

[0024] in, Diagonal matrix h represents the feature space output matrix, c represents the data category, and n is the number of input data.

[0025] Preferably, the calculation expression of the hidden layer data matrix is: H=g(A*X+b); wherein b is a bias parameter.

[0026] Preferably, the training method of the multi-layer graph embedding extreme learning machine is specifically: performing cyclic calculations on the data in the dataset from k=1 to k=K based on the graph embedding autoencoder, and taking the output of the Kth information interaction as the final feature output, which is expressed as follows:

[0027] H k =g(G (k-1) (β k ) T );

[0028] Among them, H k represents the characteristic representation of the k-th output matrix; G k =σ(∑f(G k-1 ,G k-2 ), σ represents the scaling parameter, G k represents the input of the kth information interaction or the output of the k-1th information interaction; β k represents the output weight of the kth information interaction; T represents the matrix transpose.

[0029] Preferably, the data set is a wind turbine fault detection data set.

[0030] In addition, the present invention also includes a wind turbine fault detection device based on a graph embedding extreme learning machine, comprising:

[0031] at least one processor;

[0032] and a memory communicatively coupled to the at least one processor;

[0033] Wherein, the memory is used to store computer programs;

[0034] The processor is configured to implement the above-mentioned image classification method when executing the computer program.

[0035] In addition, the present invention also includes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the image classification method as described above is implemented.

[0036] The application of the technical solution of the present invention has the following beneficial effects:

[0037] (1) The image classification method in the present invention combines graph theory with the label information of the wind turbine blade icing fault data to embed the weights into the reconstruction error function of the extreme learning machine autoencoder, and applies the graph embedding matrix in the ELM feature space to maintain the global structural information of the feature data and retain the global similarity between the data, so that the algorithm can find a more effective feature discrimination boundary in the feature space. Different from the existing extreme learning machine, the present invention uses the extreme learning machine autoencoder to reconstruct the feature data, and at the same time considers the relationship between the compressed feature representation and other features based on the graph weight, calculates the similarity between each data feature and other data features in the original feature space of the data, reconstructs the original data in the ELM feature space according to the data label information based on the distance, reconstructs the data error term in the autoencoder, and then adds the graph embedding penalty term in the feature subspace to maximize the separability between classes and reduce the local structure within the class, so as to find a more effective feature discrimination boundary in the feature space.

[0038] (2) This paper proposes a multi-layer graph embedding extreme learning machine (ML-GEELM) structure based on a cyclic structure, which fully utilizes the feature dimensionality reduction capability of each autoencoder. Different from the existing stacked extreme learning machine, this paper uses a graph theory-based method to effectively aggregate features with different weights using multiple layers of output weights, maintaining the similarity between the global data structure after feature dimensionality reduction and the original spatial structure.

[0039] In addition to the above-described objects, features and advantages, the present invention has other objects, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0041] Figure 1 is a flowchart of the steps of the image classification method in a preferred embodiment of the present invention;

[0042] Figure 2 1 is a schematic diagram of a framework of an automatic encoder embedded in a graph in a preferred embodiment of the present invention;

[0043] Figure 3 Schematic diagram of the feature subspace of the graph embedding autoencoder in a preferred embodiment of the present invention;

[0044] Figure 4 1 is a schematic diagram of a framework of a multi-layer graph embedding extreme learning machine in a preferred embodiment of the present invention;

[0045] Figure 5 This is a visualization of IRIS’s feature data after dimensionality reduction using the ELM-AE algorithm;

[0046] Figure 6 This is a visualization of the feature data of IRIS after dimensionality reduction using the GEELM algorithm;

[0047] Figure 7 This is a comparison chart between this embodiment and the GDR-ML-ELM algorithm at different feature nodes. DETAILED DESCRIPTION

[0048] The original wind turbine SCADA data has problems such as large data volume, too high data feature dimension, weak feature correlation, resulting in low detection correctness and accuracy. The present invention combines graph embedding theory to propose a graph embedding-based extreme learning machine autoencoder (Graph Embedded Extreme Learning Machine AutoEncoder, GEELM-AE). Starting from the extreme learning machine autoencoder, the present invention addresses the problem that the original error function only focuses on its own data reconstruction, thereby affecting incomplete feature reconstruction and insufficient feature representation ability. Compared with existing algorithms, the present invention has stronger feature representation capabilities and better classification decision-making capabilities in the application of wind turbine blade icing fault detection. The following is a detailed description of the embodiments of the present invention in conjunction with the accompanying drawings, but the present invention can be implemented in a variety of different ways as defined and covered by the claims.

[0049] Example:

[0050] like Figure 1 As shown, this embodiment discloses an image classification method based on graph embedding extreme learning machine, the steps are as follows:

[0051] Step S1: Acquire a data set and preprocess it to obtain a training set and a test set. The data set in this embodiment is a wind turbine fault detection data set;

[0052] Step S2: Using a multi-layer graph embedding extreme learning machine algorithm to train the training set, a classification model is obtained, i.e., a wind turbine blade fault detection model;

[0053] Step S3: inputting the test set into the fan blade fault detection model to obtain a fan blade fault detection result;

[0054] Among them, the multi-layer graph embedding extreme learning machine algorithm in step S2 is specifically: constructing multiple graph embedding autoencoders, forming information interaction with the graph embedding autoencoders in a cyclic manner, converting them into a multi-layer graph embedding extreme learning machine, and using a single graph embedding autoencoder as a child node. The data in the training set is output to the next layer of the network through the child node to obtain the weight output of the information interaction, and the multi-layer network is trained using the optimal layer-by-layer method to obtain the final feature output.

[0055] like Figure 2 As shown, this embodiment is based on a graph embedding framework and combines a supervised method for reconstructing original features based on graph weights and data label information to retain the local feature information and global feature structure information of the original data.

[0056] like Figure 3 As shown in FIG, the main purpose of the graph embedding autoencoder is to cluster similar data together using label information and penalty graph in the feature subspace and increase the distance between classes, thereby finding a more effective discrimination boundary.

[0057] Furthermore, the graph embedding autoencoder specifically includes the following steps:

[0058] 1) Get the data in the training set And set the activation function; where X=[x1,...,x N ];x i represents the i-th sample in the data, x i ∈R d , d represents the feature dimension of each sample; T = [t1,…,t N ];t i represents the i-th category in the data, t i ∈R f ; N represents the amount of data; f represents the feature dimension of each category; the mapping from the hidden layer to the output layer is a linear mapping X = Hβ, where H represents the hidden layer output matrix and β represents the output weight; the activation function is the sigmoid activation function g(x), which is expressed as

[0059] 2) First, to redefine the structural error of the encoder, the error weight matrix W = w is defined by reconstructing the original data. ij (i=1……N,j=1……N), then the reconstruction error function of the graph embedding autoencoder is defined as the distance relationship between the data and each reconstructed data e i :

[0060] e i =∑ j w ij ||x i -h j β|| 2 ;

[0061] The accumulated error J of the autoencoder 1st It can be expressed as:

[0062]

[0063] Calculate the reconstruction error weight w ij , the reconstruction error weight w ij The calculation expression is as follows:

[0064]

[0065] in, Represents based on sample x i The set of all features of category t in the neighborhood, set the number of neighbor points to P, and ω to the median of the square distance of all paired input data;

[0066] Then, the graph embedding matrix is ​​applied to the hidden layer of the autoencoder to enhance the similarity of the local structure and the difference of the global structure of the feature information in the feature space. The loss function J of the hidden layer is 2nd Designed to:

[0067]

[0068] Where S is the graph embedding matrix, so the GEELM-AE joint loss function is as follows:

[0069]

[0070] Among them, the first term is the reconstruction term based on the feature data error of the graph weight matrix, w ijAfter the control samples are feature mapped, the weights of adjacent categories in the feature space will be clustered together, thereby reducing the feature misrepresentation caused by different classifications of the same category, and more effectively discriminating features in the feature space clustering. The second term is the regularization term, which controls the complexity of the model, prevents overfitting, and ensures the stability of the loss function; wherein, C represents the regularization coefficient, and C>0 adjusts the balance between the reconstruction error term and the association regularization term. The third term is the regularization term of the feature data correlation based on the graph embedded in the feature space, which enhances the similarity of the local structure of the feature information in the feature space and the difference of the global structure. λ is the balance parameter of the loss function, and λ>0 is preferred in this embodiment.

[0071] In addition, the objective function J GEELM-AE Take the derivative and let Solve for β:

[0072]

[0073] Where D represents the diagonal matrix of error weights, D = d ii =∑ j w ij .

[0074] Specifically, the expression of the computation graph embedding matrix S is as follows:

[0075] S=S w S b =H(D w -A w )H T H(D b -A b )H T ;

[0076]

[0077] in, Diagonal matrix h represents the feature space output matrix, c represents the data category, and n is the number of input data;

[0078] Then initialize the weight parameter A, A∈R D×L , L is the number of hidden layer nodes;

[0079] 3) Calculate the optimized output weight β, the expression is as follows:

[0080] β=[H T DH+(I L +λS) / C] -1 H T WX;

[0081] Where I represents the identity matrix, λ represents the balance parameter, C represents the regularization coefficient, H represents the hidden layer output matrix, and the calculation expression of the hidden layer data matrix is: H = g(A*X+b); where b is the bias parameter.

[0082] It should be noted that in the field of extreme learning machine autoencoders, traditional multi-layer extreme learning machine algorithms mostly rely on simple direct connections between multi-layer perceptrons to construct deep models, and do not fully utilize the relationship information between feature data brought by the graph structure. This embodiment proposes to form an information interaction-related graph structure in a cyclic manner using a graph embedding autoencoder (GEELM-AE), and then convert it into a stacked multi-layer graph embedded extreme learning machine (ML-GEELM). Each GEELM-AE is treated as a node, and the structural relationship of each part under different weights (the generalization performance of the ELM random weight increase algorithm) is fully identified. The "node" relationship is converted into a multi-layer encoding structure to form a generalized multi-layer extreme learning machine framework.

[0083] like Figure 5 As shown, this embodiment takes a three-layer structure as an example. The training method of the multi-layer graph embedding extreme learning machine is specifically: the data in the dataset is cyclically calculated from k=1 to k=K based on the graph embedding autoencoder, and the output of the Kth information interaction is taken as the final feature output, which is expressed as follows:

[0084] H k =g(G (k-1) (β k ) T );

[0085] Among them, H k represents the characteristic representation of the k-th output matrix; G k =σ(∑f(G k-1 ,G k-2 ), σ represents the scaling parameter, G k represents the input of the kth information interaction or the output of the k-1th information interaction, f(·) is the matrix multiplication; β k represents the output weight of the kth information interaction; T represents the matrix transpose.

[0086] Compared with the traditional autoencoder method, the advantages of the multi-layer graph embedding extreme learning machine algorithm proposed in this embodiment in network training are divided into two parts: (1) the global optimal solution of each layer can be analytically calculated without multiple feedback iteration algorithms. Compared with other multi-layer extreme learning machines, ML-GEELM solves the data dimensionality reduction problem in a supervised manner and can find more discriminative features; (2) through information interaction, the information sent by each node to its neighbors is integrated into the next unit of GEELM-AE, rather than simply transferring data to increase structural errors.

[0087] Simulation experiment analysis:

[0088] To verify the effectiveness of the algorithm in this embodiment, 10 test data sets were tested and simulated. The experimental environment of this embodiment is: Intel Core i5-8265U CPU @ 1.8GHz, 16GB RAM, Matlab 2018a. As shown in Table 1, the data sets and their parameters are YaleB, ORL and UMIST face recognition data sets, where YaleB contains 2414 face image data in 39 categories; ORL contains 40 categories, 10 images per category; UMIST has a total of 20 categories and 564 data. IRIS, WINE and PageBlocks classification are from the UCI machine learning database. USPS and MINST are 10-category handwriting data sets, USPS contains 2007 data, MNIST has a total of 70,000h of data, and the ratio of data set training to test data set is 6:1.

[0089] Table 1 Datasets and their parameters

[0090]

[0091]

[0092] To demonstrate the effectiveness of the GEELM-AE algorithm proposed in this example, experiments were conducted on real datasets (UCI: IRIS, WINE, Page Blocks classification, and IMAGE) and image face recognition datasets (YaleB, ORL, and UMIST). The comparison algorithms included: AE, ELM-AE, GDR-ELM, ELM-SAE, and ELM-AEIF.

[0093] The simulation experiment includes three hyperparameters, namely: regularization coefficient C, balance parameter λ and the number of neighbors p for generating the graph adjacency matrix based on the original data. The comparison algorithms in this experiment include ELM-AE, GDR-ELM and ELM-AEIF, and the number of neurons in their hidden layers is selected from 500, 1000 and 1500. The selection range of regularization coefficient C is C∈[10 -5 ,10 -4 ,...,10 5 ], the selection range of the balance parameter λ is λ∈[10 0 ,10 1 ,...,10 10 ], the selection range of the number of neighbors p is p∈[[N / C] / 4:[N / C] / 5:[N / C]], and other algorithms use their optimal hyperparameters for experiments.

[0094] When testing the single-layer GEELEM algorithm, it is divided into two parts: image classification and real data set classification.

[0095] (1) Face recognition experiment:

[0096] In the experiment, about 60% of the samples in each category were randomly selected as the training set. The experiment was repeated ten times and the average classification accuracy was calculated as the final classification accuracy. The classification accuracy and the corresponding standard deviation (STD) are shown in Table 2.

[0097] Table 2 Performance comparison of different algorithms on face recognition image datasets

[0098]

[0099]

[0100] As can be seen from Table 2, for the classification of face images, compared with other extreme learning machine autoencoder algorithms, the classification effect has been improved to varying degrees. The reason is that when dealing with the problem of data dimensionality reduction, GEELM-AE integrates graph embedding theory into the original error function, while considering the local structural information and global structural information of the data, and more effectively finds classification discriminant features in the feature space. Compared with other algorithms, the GEELM-AE algorithm has a smaller standard deviation for face recognition data sets, a higher recognition rate, and a more stable algorithm. Compared with the classic AE method, the algorithm of this embodiment: on the one hand, GEELM-AE helps to effectively discriminate features through supervised information mining, while the classic AE does not obtain obvious discriminant features; on the other hand, since the AE problem is mostly a non-convex optimization problem, and the iterative solution using the BP algorithm is very easy to fall into the local optimum, the GEELM-AE algorithm parameters are fixed and do not require iteration, thereby avoiding the problem of falling into the local optimum. Compared with the GDRELM algorithm, which is also a graph embedding method, the algorithm of this embodiment adds a graph embedding matrix to the extreme learning machine feature space, reduces the intra-class distance, increases the inter-class distance, and finds better discriminant features. As can be seen from Table 2, the algorithm of this embodiment is superior to the GDR-ELM algorithm in most cases.

[0101] (2) Real dataset classification experiments

[0102] In order to further illustrate the feature extraction capability of GEELM and the clustering performance of feature space, the experiment was tested on the UCI real data set and the algorithm’s feature clustering capability for real data was analyzed. The test results are shown in Table 3. In addition, the experiment also gives the visualization effect of the algorithm on the IRIS data set, as shown in Table 3. Figure 5 and Figure 6 As shown, Figure 5 Visualization of IRIS feature data obtained by ELM-AE; Figure 6 This is the visualization effect of IRIS after dimensionality reduction using the GEELM algorithm.

[0103] Table 3 Performance comparison of different algorithms on the UCI real dataset

[0104]

[0105]

[0106] From Table 3 and Figure 7The test results show that, compared with the ELM-AE method, the data feature representation obtained by GEELM clusters data of the same category closer together and data of different categories farther apart, that is, the inter-class distance is larger and the intra-class distance is smaller, which facilitates the acquisition of more effective feature discrimination boundaries. The performance of GEELM and other methods on real datasets is shown in Table 3. Compared with the AE, ELM-AE, ELM-SAE, and ELM-AEIF algorithms, GEELM-AE has better performance in classification accuracy and standard deviation. When the dimension of the feature dimensionality reduction data is lower, GDR-ELM has better performance; when the dimension increases, GEELM-AE shows good performance and has a lower standard deviation.

[0107] (3) Multi-layer autoencoder structure performance experiment:

[0108] This experiment primarily compares the performance of different algorithms using multi-layer autoencoder structures. For ease of experimentation, the algorithms were tested using optimized parameters. The network structure used a three-layer structure, with only the hidden layer neurons varied for comparison. After feature extraction, each algorithm used the ELM classifier for uniform comparison. Twenty trials were performed on each dataset, and the mean and standard deviation were calculated. The test results for MNIST and NORB are shown in Table 4.

[0109] Table 4 Performance comparison of multi-layer autoencoder structures of different algorithms

[0110]

[0111] The experimental results show that compared with other multi-layer network structures based on ELM-AE: (1) The method of this embodiment has a significant improvement in the average accuracy of the test set and a relatively low standard deviation. The advantages of combining multiple GEELM-AEs into a multi-layer structure in an integrated manner effectively utilize the extraction capabilities of different layers for data features, avoiding the information error transmission caused by simple stacking and affecting the performance of a single layer; the classification performance of ML-GEELM is further improved compared with other multi-layer encoder algorithms. (2) Figure 7 The figure shows a comparison between the algorithm of this embodiment and the GDR-ML-ELM algorithm under different feature nodes. When the dimensionality reduction dimension is the same, the algorithm of this embodiment maintains good performance. This is because this embodiment embeds a graph matrix in the feature space to find more effective discriminant features in the feature space. As the multi-layer matrix reconstructs the feature data, the multi-layer graph embedded in the extreme learning machine composed of cyclic nodes has a stronger feature representation capability, and it is easier to obtain effective features compared to other simple stacking algorithms.

[0112] In summary, this paper proposes a new extreme learning machine autoencoder (GEELM-AE) based on a graph embedding method. By redefining the data feature reconstruction error function in the feature space, this method increases feature aggregation capabilities based on label information, extracts local features, and enhances feature representation capabilities. A graph-based penalty term is embedded in the loss function, which reduces intra-class distances and increases inter-class distances in the ELM feature space, extracting information about the global structure of features and thus finding more effective discrimination boundaries in the feature space. Next, a multi-layer graph embedding extreme learning machine (ML-GEELM) deep network structure constructed from multiple GEELM-AEs is proposed. Through information exchange, the output weights of multiple layers are effectively aggregated based on graph theory and output to the next layer of the network structure, maintaining the similarity between the global data structure of the feature dimensionality reduction and the original spatial structure. To verify the effectiveness of the algorithm, this chapter first conducts relevant experiments on a specific real-world dataset. The experimental results show that the proposed algorithm has stronger feature representation capabilities than other algorithms and can better cluster the local structure of feature data, thereby improving classification performance. For the task of wind turbine blade icing detection, it achieves better classification results than the latest deep learning algorithms.

[0113] In addition, this embodiment also includes a wind turbine fault detection device based on a graph embedding extreme learning machine, including:

[0114] at least one processor;

[0115] and a memory communicatively coupled to the at least one processor;

[0116] Wherein, the memory is used to store computer programs;

[0117] The processor is configured to implement the above-mentioned image classification method when executing the computer program.

[0118] In addition, this embodiment also includes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the image classification method described above is implemented.

[0119] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. An image classification method based on graph embedding extreme learning machine, characterized in that: Here are the steps: Step S1: Obtain the data set and preprocess it to obtain the training set and test set; Step S2: Use the multi-layer graph embedding extreme learning machine algorithm to train the model on the training set to obtain a classification model; Step S3: input the test set into the classification model to obtain the image classification result; The multi-layer graph embedding extreme learning machine algorithm in step S2 specifically comprises: constructing multiple graph embedding autoencoders, forming information interaction with the graph embedding autoencoders in a cyclic manner, converting them into a multi-layer graph embedding extreme learning machine, using a single graph embedding autoencoder as a child node, outputting the data in the training set to the next layer of the network through the child node, obtaining the weight output of the information interaction, and training the multi-layer network using a layer-by-layer optimal method to obtain the final feature output; The graph embedding autoencoder includes the following steps: 1) Get the data in the training set And set the activation function; where X=[x1,...,x N ];x i represents the i-th sample in the data, x i ∈R d , d represents the feature dimension of each sample; T = [t1,…,t N ];t i represents the i-th category in the data, t i ∈R f ; N represents the amount of data; f represents the feature dimension of each category; 2) Calculate the reconstruction error weight w ij ; Generate graph embedding matrix S; Initialize weight parameter A, A∈R D×L , D represents the diagonal matrix of error weights, L is the number of hidden layer nodes; 3) Calculate the optimized output weight β, the expression is as follows: β=[H T DH+(I L +λS) / C] -1 H T WX; Where I represents the identity matrix, λ represents the balance parameter, C represents the regularization coefficient, and H represents the hidden layer output matrix; The reconstruction error weight w ij The calculation expression is as follows: in, Represents based on sample x i The set of all features of category t in the neighborhood, set the number of neighbor points to P, and ω to the median of the square distance of all paired input data; The training method of the multi-layer graph embedding extreme learning machine is as follows: the data in the dataset are cyclically calculated from k=1 to k=K based on the graph embedding autoencoder, and the output of the K-th information interaction is taken as the final feature output, which is expressed as follows: H k =g(G (k-1) (b k ) T ); Among them, H k represents the characteristic representation of the k-th output matrix; G k =σ(∑f(G k-1 ,G k-2 ), σ represents the scaling parameter, G k represents the input of the kth information interaction or the output of the k-1th information interaction; β k represents the output weight of the kth information interaction; T represents the matrix transpose.

2. The image classification method according to claim 1, characterized in that The activation function is the sigmoid activation function g(x), which is expressed as 3. The image classification method according to claim 2, characterized in that The expression of the computation graph embedding matrix S is as follows: S=S w S b =H(D w -A w )H T H(D b -A b )H T ; in, Diagonal matrix h represents the feature space output matrix, c represents the data category, and n is the number of input data.

4. The image classification method according to claim 3, characterized in that The calculation expression of the hidden layer data matrix is: H = g(A*X+b); where b is the bias parameter.

5. The image classification method according to claim 1, wherein: The data set is a wind turbine fault detection data set.

6. An image classification device based on graph embedding extreme learning machine, characterized in that include: at least one processor; and a memory communicatively coupled to the at least one processor; Wherein, the memory is used to store computer programs; The processor is configured to implement the image classification method according to any one of claims 1 to 5 when executing the computer program.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the image classification method according to any one of claims 1 to 5 is implemented.

Images