Application of deep dictionary learning based on neighborhood atom enhancement in anomaly detection
Deep dictionary learning with neighborhood atom enhancement addresses the limitations of traditional methods by integrating local and global features through adaptive graph learning, enhancing anomaly detection accuracy and robustness.
Patent Information
- Application Number
- CN202410953220.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-12-27
- Estimated Expiration
- 2044-07-16
AI Technical Summary
Traditional anomaly detection methods struggle with high-dimensional and dynamically changing data, failing to capture local structures and dynamic changes due to limited dictionary structure updates and lack of explicit local relationship modeling.
A method involving deep dictionary learning with neighborhood atom enhancement, utilizing multi-scale feature extraction, adaptive graph learning, and SVM classification to enhance local feature representation and adaptability, incorporating neighborhood manifolds and self-expression relationships to update dictionary atoms.
Enhances anomaly detection accuracy and robustness by dynamically capturing local data relationships and subtle changes, improving feature representation and classification performance.
Smart Images
Figure CN118736322B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to computer vision and deep learning technology, and in particular to an application method of deep dictionary learning based on neighborhood atom enhancement in anomaly detection. Background Art
[0002] Anomaly detection is an important topic in the field of data analysis and machine learning. Its goal is to identify abnormal patterns that do not conform to expected behaviors or are significantly different from normal patterns. This technology is widely used in many application fields such as network security detection, system fault diagnosis, credit card fraud detection, and medical diagnosis. Traditional anomaly detection methods mainly include statistical-based methods, distance-based methods, density-based methods, and machine learning-based methods. However, these methods often have difficulty capturing accurate behavioral information when processing high-dimensional data, complex patterns, and dynamically changing data, and thus fail to identify true outliers or misjudge abnormal behaviors.
[0003] In order to overcome these challenges, dictionary learning technology has been introduced into the field of anomaly detection. Dictionary learning learns a set of dictionary atoms from the data so that the data can be represented by a sparse linear combination of these atoms, thereby improving the sparsity and robustness of the data representation. Anomaly detection methods based on dictionary learning usually capture abnormal patterns in the data by constructing dictionaries and using sparse representations. Traditional anomaly detection combined with dictionary learning performs well in processing high-dimensional data and sparse data, but it still has certain shortcomings in processing local structures and dynamic changes of data. Specifically, traditional dictionary learning methods usually rely on global sparse coding to represent data due to the lack of explicit modeling of local relationships of data. During the entire training process, the update range of the dictionary structure is limited, and it is easy to ignore the local information of the data samples, resulting in insufficient dictionary expression. Therefore, it solves the dilemma of anomaly detection facing high-dimensional data and nonlinear data to a certain extent, but it is still easy to fail to achieve the ideal recognition effect due to the lack of flexibility when facing frequent changes in local structures and large differences. Summary of the invention
[0004] Purpose of the invention: The purpose of the present invention is to solve the deficiencies in the prior art and to provide an application method of deep dictionary learning based on neighborhood atom enhancement in anomaly detection.
[0005] Technical solution: The present invention provides a method for applying deep dictionary learning based on neighborhood atom enhancement in anomaly detection, comprising the following steps:
[0006] Step 1, obtain the sample image I0 to be tested, input it into the constructed deep dictionary learning network for training, and obtain the multi-scale global features of the sample image I0 to be tested; the deep dictionary learning network includes three-layer dictionary networks of different scales, and the global features extracted through learning of the three-layer dictionary network include, for example, color distribution and texture pattern of the sample image;
[0007] Step 2: For the sample image I0 to be tested in step 1, preprocess it, perform normalization, data enhancement and principal component analysis (PCA) to remove redundant information, and obtain the reconstructed original features;
[0008] Step 3: Perform multi-dimensional feature extraction (mean, standard deviation, maximum frequency, etc.) on the reconstructed original features, use the sliding window segmentation technology to mark the dynamically changing data points, and calculate the mean, standard deviation, maximum frequency, frequency domain features, etc. of the data in the window;
[0009] Step 4: For the feature map obtained after segmentation in step 3, use the k-means clustering method to segment the sample data, construct the neighborhood manifold around the sample, and calculate the weight matrix W and Laplace matrix L of the feature map after segmentation;
[0010] Step 5: The weight matrix W of the graph contains the neighborhood manifold structure information. Local modeling is performed inside the neighborhood manifold to obtain the self-expression relationship between data points. The collected self-expression information is used to update the weighted dictionary atoms. The updated dictionary learning network obtains the local features in the sample to be tested.
[0011] Step 6: Construct an adaptive graph learning module, and update the graph of the local modeling of the neighborhood manifold through the adaptive graph adaptation module to further enhance the learned local feature representation; the adaptive graph learning block is responsible for enabling the deep dictionary network to further learn and train the local features of the image, and further enhance the timeliness of the dictionary learning after the neighborhood manifold atom enhancement in anomaly detection by dynamically adjusting the edge weights of the graph, so that the method has dynamic adaptability to the data stream of anomaly detection;
[0012] The adaptive graph learning module is stacked with M learning blocks, each of which can capture the characteristics of local and partial graph structures. Stacking multiple learning blocks gradually integrates these local features into higher-level and more abstract feature representations.
[0013] Step 7, repeat S5 to S6, iteratively optimize the generated dictionary and sparse representation to ensure sufficiently accurate local feature representation; the purpose of steps 4 to 7 in the present invention is to obtain local features;
[0014] Step 8: Integrate the global features obtained in step 1 and the local features obtained in step 7, fuse the feature representations of different layers, and obtain the final feature representation;
[0015] Step 9: Input the fused feature representation into the SVM classifier for training. The SVM distinguishes normal data from abnormal data by finding the maximum margin hyperplane.
[0016] Step 10: According to the screening output classification results of the SVM classifier, identify and mark abnormal data points.
[0017] Furthermore, the deep dictionary learning network in step 1 includes a three-layer dictionary network structure, and the number of dictionary atoms in each layer is 200, 100 and 50 respectively. The three-layer dictionary learns features of different scales; since information of different scales can reveal features of multiple levels of data, the deep dictionary learning network here can better grasp the global features of the sample, and the model optimization problem of dictionary learning is expressed as follows:
[0018]
[0019] Where Y represents sample data, D represents dictionary, and X represents sparse representation. The training objective requires that the matrix two norm of the difference between the original matrix and the product of the decomposed dictionary matrix and the sparse code matrix be minimized when the zero norm of the sparse representation X is as small as possible.
[0020] Furthermore, the specific method of step 2 is:
[0021] Step a: For image I0, in order to eliminate the dimension effect of the data and make the data on the same scale, it is necessary to perform standardization processing to obtain the high-dimensional original feature I1. The standardization processing formula is as follows:
[0022]
[0023] Among them, μ is the mean of the image, σ is the standard deviation of the image, and the high-dimensional original feature I1 has zero mean and unit variance, which is used to accelerate the convergence of the model;
[0024] Step b: Calculate the covariance matrix C of I1, perform eigenvalue decomposition on C, and obtain eigenvalues and eigenvectors:
[0025]
[0026] CV=VΛ
[0027] Among them, m is the number of samples, is the number of rows of the feature matrix I1, Λ is a diagonal matrix containing the eigenvalues of the covariance matrix, V is the eigenvector matrix, each column is an eigenvector, and the eigenvalues are arranged from large to small; first, the inner product of the feature matrix is scaled to generate an unbiased estimate of the sample, and then the eigendecomposition is performed to obtain the eigenvalues and eigenvectors;
[0028] Step c: Project to the principal component space and reconstruct the original feature I2:
[0029] I2=PCA(I2)=I1V k
[0030] Where V k is the first k eigenvectors, select V k As a projection matrix, the original high-dimensional matrix I1 is projected into the low-dimensional principal component space, and I2 is restored to the same scale as the original feature I1.
[0031] Furthermore, step 3 selects a fixed-size window and step size for I2, calculates the mean, standard deviation, maximum value, frequency domain features, etc. of the data in the window, and uses the k-means algorithm to segment and label the feature vectors of I2. The specific operation is as follows: randomly select n initial center points, for each feature vector, first calculate the distance between them and each center point, and iteratively calculate the average value assigned to the center point until convergence. Finally, each vector will be assigned to a fixed center point to complete the segmentation and labeling results;
[0032] The obtained segmentation and annotated labels are used as the basis for the k-means algorithm to segment the samples. The method is:
[0033] Use the label as the cluster center to divide the data points into neighborhoods N of uniform size, each neighborhood N i Select k nearest neighbor data points, for each data point x i , calculate its Euclidean distance to each nearest cluster center, and the weight matrix W is:
[0034]
[0035] Construct the Laplacian matrix L. Each element in the matrix L ij Represents data point x i and x j The similarity between them, L is calculated as follows:
[0036] L=AW
[0037] Where W is the weight matrix, A is a diagonal matrix, and A ii =∑ j W ij .
[0038] The weight matrix W contains the neighborhood manifold structure information. The elements of the weight matrix W are the coefficients of the correlation between the current instance data point and its neighboring data points. Since each data point can be represented by a linear combination of other data points in its neighborhood, the reconstructed dictionary D is combined with the weight matrix W (the elements in the weight matrix W represent the self-expression coefficients). The combined DW has similar feature semantics to the original dictionary D. Specifically, the linear combination of DW and the original dictionary is regarded as an enhancement of the original dictionary. The formula is as follows:
[0039]
[0040] where D is the original dictionary, is the enhanced dictionary, W is the weight matrix, and λ is the regularization parameter. ;
[0041] Next, use the enhanced To replace the original dictionary D to participate in the sparse coding process, add a neighborhood regularization constraint to ensure that the self-expression relationship can effectively update the atoms and use the coordinate gradient descent method (CGD) to update the dictionary in the target model And sparse representation X is alternately optimized and solved:
[0042]
[0043] Furthermore, step 6 updates the graph structure and learns it by dynamically adjusting the edge weights of the graph, so that the learned local feature representation has dynamic adaptability to the data stream of anomaly detection. The adaptive graph learning module is implemented by stacking M learning blocks, each of which can capture the features of local and partial graph structures. The forward propagation process of the graph convolutional network in each learning block is:
[0044]
[0045] in, is the normalized adjacency matrix of the lth layer, defined as for yes After each learning block is completed, the feature representations in all learning blocks are integrated.
[0046] Furthermore, in step 8, feature splicing is used to fuse global features and local features, and feature representations at different levels are spliced together, where L represents the number of network layers:
[0047] H concat =[H (1) ,H (2) ,H (3) ,...H (L) ]
[0048] The fused features are represented by H concat The input is sent to the SVM classifier for training. The penalty coefficient threshold C and the gamma correlation coefficient G are adjusted, and the control variables use different kernel functions to confirm the classification results and changes in the samples to be processed.
[0049] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0050] (1) The present invention introduces the construction of neighborhood manifolds in dictionary learning and uses the rich self-expression relationships in the neighborhood to explore the potential of dictionary learning in the direction of local representation, dynamically capturing the local relationships and subtle changes between data points, which is suitable for the actual scenario of anomaly detection. Through the multi-level feature fusion mechanism, the feature representations of different levels are effectively fused to obtain a more representative final feature representation, and SVM is used to improve the detection performance by maximizing the classification interval. The combination of the two makes the anomaly detection results more reliable and accurate.
[0051] (2) The present invention ensures that the learned local features can effectively and flexibly detect anomalies accurately by constructing an adaptive graph learning module, obtains a more representative final feature representation, and enhances the ability of feature representation. On the one hand, sparse coding is used to capture subtle changes in data, and on the other hand, the detection performance is improved by maximizing the classification interval, making the anomaly detection results more reliable and accurate. The model's ability to recognize abnormal patterns is improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 It is the overall flow chart of the present invention;
[0053] Figure 2 It is a schematic diagram of the overall network structure of the present invention;
[0054] Figure 3 This is a diagram of anomaly detection results obtained in an embodiment of the present invention. DETAILED DESCRIPTION
[0055] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.
[0056] In order to overcome the defects of the prior art, the present invention finds that there is an obvious correlation between the surrounding data points in the structure of the sample neighbor manifold. Experiments have proved that this correlation can effectively help us capture the dynamic structure of local data. The present invention represents this correlation by constructing a neighborhood manifold and calculating the self-expression relationship. At the same time, the weight of the self-expression relationship is used to update the dictionary atoms, and the learned local features are further enhanced by combining the adaptive graph learning technology, so that the generated dictionary not only takes into account the information of the local structure, but also retains the information of the original dictionary learning from the global structure perspective, which has the advantage of being applied to anomaly detection. This innovative method provides better accuracy and robustness for anomaly detection.
[0057] like Figure 1 As shown, the application method of deep dictionary learning based on neighborhood atom enhancement in anomaly detection of the present invention is characterized by comprising the following steps:
[0058] Step 1: obtain the sample image I0 to be tested, input it into the constructed deep dictionary learning network for training, and obtain the multi-scale global features of the sample image I0 to be tested; the deep dictionary learning network includes three-layer dictionary networks of different scales, and the features of different scales are extracted through the three-layer dictionary network learning;
[0059] Step 2: For the sample image I0 to be tested in step 1, preprocess it, perform normalization, data enhancement and principal component analysis PCA to remove redundant information, and obtain the reconstructed original features;
[0060] Step 3: Perform multi-dimensional feature extraction on the reconstructed original features, and use the sliding window segmentation method to mark the dynamically changing data points, and calculate the mean feature, standard deviation feature, maximum feature and frequency domain feature of the data in the sliding window;
[0061] Step 4: For the feature map obtained after segmentation in step 3, use the k-means clustering method to segment the sample data, construct the neighborhood manifold around the sample, and calculate the weight matrix W and Laplace matrix L of the feature map after segmentation;
[0062] Step 5: The weight matrix W of the graph contains the neighborhood manifold structure information. Local modeling is performed inside the neighborhood manifold to obtain the self-expression relationship between data points. The collected self-expression information is used to update the weighted dictionary atoms. The updated dictionary learning network obtains the local features in the sample to be tested.
[0063] Step 6: construct an adaptive graph learning module, and update the graph of the local modeling of the neighborhood manifold through the adaptive graph adaptation module to enhance the local features obtained in step 5; the adaptive graph learning module is stacked with M learning blocks, each of which captures the features of the local and partial graph structure, and the stacked M learning blocks gradually integrate these local features into a higher-level and more abstract feature representation;
[0064] Step 7: Repeat steps 5 and 6 to iteratively optimize the generated dictionary and sparse representation to ensure sufficiently accurate local features.
[0065] Step 8: Integrate the global features obtained in step 1 and the local features obtained in step 7, fuse the feature representations of different layers, and obtain the final feature representation;
[0066] Step 9: Input the fused feature representation into the SVM classifier for training. The SVM distinguishes normal data from abnormal data by finding the maximum margin hyperplane.
[0067] Step 10: According to the screening output classification results of the SVM classifier, identify and mark abnormal data points.
[0068] In step 1 of this embodiment, the number of atoms in the three-layer dictionary network of the deep dictionary learning network is 200, 100, and 50 respectively;
[0069] The model optimization problem of dictionary learning is expressed as follows:
[0070]
[0071] Among them, Y represents sample data, D represents dictionary, and X represents sparse representation.
[0072] The specific method of step 2 of this embodiment is:
[0073] Step a: Standardize the image I0 to obtain the high-dimensional original feature I1. The standardization formula is as follows:
[0074] Among them, μ is the mean of the image, σ is the standard deviation of the image;
[0075] Step b: Calculate the covariance matrix C of I1, perform eigenvalue decomposition on C, and obtain eigenvalues and eigenvectors: CV = VΛ;
[0076] Among them, m is the number of samples, is the number of rows of the feature matrix I1, Λ is a diagonal matrix containing the eigenvalues of the covariance matrix, V is the eigenvector matrix, each column is an eigenvector, and the eigenvalues are arranged from large to small;
[0077] Step c: Project to the principal component space and reconstruct the original feature I2: I2 = PCA (I2) = I1V k ;
[0078] Where V k is the first k eigenvectors, select V k As a projection matrix, the high-dimensional original feature I1 is projected into the low-dimensional principal component space, and I2 is restored to the same scale as the original feature I1.
[0079] In step 4 of this embodiment, the k-means algorithm is used to segment and label the feature vectors of I2. The specific operation is as follows: n initial center points are randomly selected. For each feature vector, the distance between them and each center point is first calculated. The average value assigned to the center point is iteratively calculated until convergence. Finally, each vector is assigned to a fixed center point to complete the segmentation and labeling results.
[0080] Next, the k-means algorithm is used to divide the data points into N neighborhoods of uniform size using the obtained labels as cluster centers. i Select k nearest neighbor data points, for each data point x i , calculate its Euclidean distance to each nearest cluster center, and the weight matrix W is:
[0081] Construct the Laplacian matrix L. Each element in the matrix L ij Represents data point x i and x j The similarity between them, L is calculated as follows: L = AW;
[0082] Where W is the weight matrix, A is a diagonal matrix, and A ii =∑ j W ij .
[0083] The specific method of step 5 of this embodiment is as follows: the elements in the weight matrix W are the coefficients of the correlation between the current instance data point and its neighboring data points, each data point is represented by a linear combination of other data points in its neighborhood, and the reconstructed dictionary D is combined with the weight matrix W, and the combined DW has similar feature semantics to the original dictionary D;
[0084] DW is linearly combined with the original dictionary to enhance the formula:
[0085]
[0086] where D is the original dictionary, is the enhanced dictionary, W is the weight matrix, and λ is the regularization parameter;
[0087] With enhanced To replace the original dictionary D to participate in the sparse coding process, add a neighborhood regularization constraint to ensure that the self-expression relationship can effectively update the atoms and use the coordinate gradient descent method (CGD) to update the dictionary in the target model And sparse representation X is alternately optimized and solved:
[0088]
[0089] The forward propagation process of the graph convolutional network in each learning block in the adaptive graph learning module of this embodiment is:
[0090]
[0091] in, is the normalized adjacency matrix of the lth layer, defined as for yes After each learning block is completed, the feature representations in all learning blocks are integrated.
[0092] In step 8 of this embodiment, feature splicing is used to fuse global features and local features, and feature representations at different levels are spliced together. L represents the number of network layers:
[0093] H concat =[H (1) ,H (2) ,H (3) ,...H (L) ]
[0094] The fused features are represented by H concat Input into the SVM classifier for training, and adjust the penalty coefficient threshold C and gamma correlation coefficient G, and control the variables to use different kernel functions to confirm the classification results and changes in samples to be processed. The specific parameter configuration is shown in Table 1:
[0095] Table 1 SVM vector machine parameter configuration
[0096] SVM Group 1 SVM Group 2 SVM Group 3 SVM Group 4 C 0.38 1.2 0.6 0.25 G 0.0253 2.3 0.0304 0.0201 Kernel Function RBF Liner Polynomial Sigmoid
[0097] like Figure 2 As shown in Figure 1, three layers of deep dictionary networks of different sizes are used to obtain the global features of the image. At the same time, a series of operations such as image standardization, principal component analysis, and multidimensional feature extraction are performed on the image. The k-means algorithm is used to construct the neighborhood manifold and calculate the weight matrix W. ijThe dictionary atoms are updated. The adaptive graph learning block is responsible for integrating the local features obtained by the updated deep dictionary network and the global features obtained previously. The SVM vector machine is responsible for training and distinguishing the concatenated features and filtering and outputting the final anomaly detection results.
[0098] Example
[0099] To further verify the technical benefits of the present invention, this embodiment is tested on the MVTec AD dataset, which belongs to the image collection category of industrial product surface defects, and contains a variety of different types of surface defects, such as scratches, cracks, holes, etc., as abnormal data points for detecting the performance of the anomaly detection algorithm. The present invention displays the distribution of abnormal and normal data points on the two-dimensional image in the form of a scatter plot. Due to the large overall length, some illustrations are intercepted to show the actual effect. Figure 3 The specific distribution of normal data points and abnormal data points is clearly displayed. The more obvious the separation between abnormal data points and normal data points is, the better the anomaly detection effect is; conversely, if there is a large amount of overlap between abnormal data points and normal data points, the anomaly detection effect is poor.
Claims
1. An application method of deep dictionary learning based on neighborhood atom enhancement in anomaly detection, characterized in that: The following steps are involved: Step 1: obtain the sample image I0 to be tested, input it into the constructed deep dictionary learning network for training, and obtain the multi-scale global features of the sample image I0 to be tested; the deep dictionary learning network includes three-layer dictionary networks of different scales, and the features of different scales are extracted through the three-layer dictionary network learning; Step 2: For the sample image I0 to be tested in step 1, preprocess it, perform normalization, data enhancement and principal component analysis PCA to remove redundant information, and obtain the reconstructed original features; Step 3: Perform multi-dimensional feature extraction on the reconstructed original features, and use the sliding window segmentation method to mark the dynamically changing data points, and calculate the mean feature, standard deviation feature, maximum feature and frequency domain feature of the data in the sliding window; Step 4: For the feature map obtained after segmentation in step 3, use the k-means clustering method to segment the sample data, construct the neighborhood manifold around the sample, and calculate the weight matrix W and Laplace matrix L of the feature map after segmentation; Step 5: The weight matrix W of the graph contains the neighborhood manifold structure information. Local modeling is performed inside the neighborhood manifold to obtain the self-expression information between data points. The collected self-expression information is used to update the weighted dictionary atoms. The updated dictionary learning network obtains the local features in the sample to be tested. The self-expression information between data points refers to the coefficient of the degree of correlation between the current instance data point and its neighboring data points; Step 6: construct an adaptive graph learning module, and update the graph of the local modeling of the neighborhood manifold through the adaptive graph learning module to enhance the local features obtained in step 5; the adaptive graph learning module is stacked with M learning blocks, each of which captures local features and features of a part of the graph structure, and the stacked M learning blocks gradually integrate these local features into a higher-level and more abstract feature representation; Step 7: Repeat steps 5 and 6 to iteratively optimize the generated dictionary and sparse representation to ensure sufficiently accurate local features. Step 8: Integrate the global features obtained in step 1 and the local features obtained in step 7, fuse the feature representations of different layers, and obtain the final feature representation; Step 9: Input the fused feature representation into the SVM classifier for training. The SVM distinguishes normal data from abnormal data by finding the maximum margin hyperplane. Step 10: According to the screening output classification results of the SVM classifier, identify and mark abnormal data points.
2. The method for applying deep dictionary learning based on neighborhood atom enhancement in anomaly detection according to claim 1, characterized in that: In the step 1, the number of atoms in the three-layer dictionary network of the deep dictionary learning network is 200, 100 and 50 respectively; The model optimization problem of dictionary learning is expressed as follows: Among them, Y represents sample data, D represents dictionary, and X represents sparse representation.
3. The method for applying deep dictionary learning based on neighborhood atom enhancement in anomaly detection according to claim 1, characterized in that: The specific method of step 2 is: Step a: Standardize the image I0 to obtain the high-dimensional original feature I1. The standardization formula is as follows: Among them, μ is the mean of the image, σ is the standard deviation of the image; Step b: Calculate the covariance matrix C of I1, perform eigenvalue decomposition on C, and obtain eigenvalues and eigenvectors: CV = VΛ; Among them, m is the number of samples, is the number of rows of the feature matrix I1, Λ is a diagonal matrix containing the eigenvalues of the covariance matrix, V is the eigenvector matrix, each column is an eigenvector, and the eigenvalues are arranged from large to small; Step c: Project to the principal component space and reconstruct the original feature I2: I2 = PCA (I2) = I1V k ; Where V k is the first k eigenvectors, select V k As a projection matrix, the high-dimensional original feature I1 is projected into the low-dimensional principal component space, and I2 is restored to the same scale as the original feature I1.
4. The method for applying deep dictionary learning based on neighborhood atom enhancement in anomaly detection according to claim 1, characterized in that: The step 4 uses the k-means algorithm to segment and label the feature vectors of I2. The specific operation is as follows: randomly select n initial center points, for each feature vector, first calculate the distance between them and each center point, and iteratively calculate the average value assigned to the center point until convergence. Finally, each vector will be assigned to a fixed center point to complete the segmentation and labeling results; Next, the k-means algorithm is used to divide the data points into N neighborhoods of uniform size using the obtained labels as cluster centers. i Select k nearest neighbor data points, for each data point x i , calculate its Euclidean distance to each nearest cluster center, and the weight matrix W is: Construct the Laplacian matrix L. Each element in the matrix L ij Represents data point x i and x j The similarity between them, L is calculated as follows: L = AW; Where W is the weight matrix, A is a diagonal matrix, A ii =∑ j W ij .
5. The method for applying deep dictionary learning based on neighborhood atom enhancement in anomaly detection according to claim 1, characterized in that: The specific method of step 5 is as follows: the elements in the weight matrix W are the coefficients of the correlation between the current instance data point and its neighboring data points, each data point is represented by a linear combination of other data points in its neighborhood, and the reconstructed dictionary D is combined with the weight matrix W, and the combined DW has similar feature semantics to the original dictionary D; DW is linearly combined with the original dictionary to enhance the formula: where D is the original dictionary, is the enhanced dictionary, W is the weight matrix, and λ is the regularization parameter; With enhanced To replace the original dictionary D to participate in the sparse coding process, add a neighborhood regularization constraint to ensure that the self-expression relationship can effectively update the atoms and use the coordinate gradient descent method (CGD) to update the dictionary in the target model And sparse representation X is alternately optimized and solved:
6. The method for applying deep dictionary learning based on neighborhood atom enhancement in anomaly detection according to claim 1, characterized in that: The forward propagation process of the graph convolutional network in each learning block in the adaptive graph learning module is: in, is the normalized adjacency matrix of the lth layer, defined as for yes After each learning block is completed, the feature representations in all learning blocks are integrated.
7. The method for applying deep dictionary learning based on neighborhood atom enhancement in anomaly detection according to claim 1, characterized in that: When the step 8 fuses the global features and the local features, feature splicing is used to splice the feature representations at different levels, where L represents the number of network layers: H concat =[H (1) ,H (2) ,H (3) ,...H (L) ] The fused features are represented by H concat The data is input into the SVM classifier for training. The penalty coefficient threshold P and the gamma correlation coefficient G are adjusted, and the control variables adopt different kernel functions to confirm the classification results and changes in the samples to be processed.