Power distribution network hidden danger identification model compression method, system and equipment based on sparse low-rank decomposition, and medium

Abstract

The invention relates to the technical field of electric power visual inspection, and discloses a sparse low-rank decomposition-based power distribution network hidden danger identification model compression method, system and device and a medium, and the method comprises the steps: carrying out the singular value decomposition of a weight matrix of a full connection layer of a convolutional neural network, and obtaining a low-rank approximation matrix; adopting a neuron screening strategy to calculate a neuron importance score, and identifying unimportant input and output neurons; according to a preset sparse rate and a preset reduction rate, performing sparse processing on rows and columns corresponding to unimportant neurons in the decomposition matrix to realize structured compression; and finally, reconstructing to obtain a compressed weight matrix. According to the method, a controllable sparsification mechanism is introduced on the basis of traditional low-rank decomposition, the compression efficiency can be remarkably improved on the premise that the model recognition precision is basically kept, and the deployment problem of a power distribution network hidden danger recognition model in resource-limited edge equipment is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of power visual inspection technology, and in particular to a method, system, equipment and medium for compressing a distribution network hidden danger identification model based on sparse low-rank decomposition. Background Technology

[0002] In recent years, with the rise of artificial neural networks, especially convolutional neural networks (CNNs), deep learning has become a reliable, efficient, and fast method for identifying risks in power distribution networks. However, the efficient computing power of deep neural networks relies on massive memory requirements and complex designs. Edge devices are typically subject to strict limitations in terms of resources and energy, posing significant challenges to their deployment. Therefore, the compression of deep learning models has become a key research direction for the efficient identification of power distribution network risks in edge computing scenarios.

[0003] Currently, network pruning reduces network weight redundancy by removing unimportant parameters from neurons, thereby lowering network complexity. Low-rank decomposition (LLD) applied to convolutional filters not only accelerates time-consuming convolution operations but also compresses the parameters of the convolution kernel. Both network pruning and LLD are effective compression strategies. Among them, Singular Value Decomposition (SVD) in LLD exhibits strong stability and significant effectiveness in compressing fully connected layer weights. It achieves efficient compression by evaluating the information value of parameters, and the decomposed factor matrix can be efficiently stored on resource-constrained devices. Inspired by pruning and sparsity theory, the importance or contribution of neurons in the layer should be considered during the low-rank approximation process. Based on the above, this study proposes a novel sparse low-rank method (SLR)—by sparsifying the decomposition matrix, lower ranks are retained for unimportant neurons while maintaining the ranks of important neurons, thus improving the compression effect of SVD. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention provides a method, system, device and medium for compressing a distribution network hidden danger identification model based on sparse low-rank decomposition.

[0005] Therefore, the technical problems solved by this invention are: first, the contradiction between the large size and high computational complexity of the model and the limited storage and computing power of edge devices; second, traditional model compression methods (such as standard singular value decomposition) are difficult to balance high compression ratio and model accuracy during the compression process, which often leads to a significant decrease in the performance of the compressed model.

[0006] To solve the above technical problems, the present invention provides the following technical solution: a method for compressing a distribution network hidden danger identification model based on sparse low-rank decomposition, comprising: performing singular value decomposition on the weight matrix of the fully connected layer in a convolutional neural network to obtain three decomposition matrices; The decomposition matrix is ​​truncated using SVD, retaining the k most important singular values ​​to obtain a low-rank approximate representation; A neuron screening strategy was employed to identify unimportant input and output neurons by evaluating their importance scores. Based on the preset sparsity and reduction rates, determine the number of rows and columns to be sparsified and the corresponding reduction rank; The rows and columns corresponding to the selected unimportant neurons in the decomposition matrix are sparsified to compress the model, resulting in a compressed convolutional neural network model.

[0007] As a preferred embodiment of the distribution network hidden danger identification model compression method based on sparse low-rank decomposition described in this invention, the step of performing singular value decomposition includes decomposing the weight matrix into a first orthogonal matrix, a diagonal matrix, and a second orthogonal matrix through SVD. By retaining the k most important singular values ​​in the diagonal matrix, the decomposition matrix is ​​truncated to obtain a low-rank approximation matrix, where k is the rank chosen based on the compression ratio and reconstruction error.

[0008] As a preferred embodiment of the distribution network hidden danger identification model compression method based on sparse low-rank decomposition described in this invention, the neuron screening strategy includes: calculating the importance scores of input neurons and output neurons, and screening out unimportant neurons based on the importance scores; The importance score is calculated based on either the absolute value of the weight, the absolute value of the activation value, or the absolute change in cost.

[0009] As a preferred embodiment of the distribution network hidden danger identification model compression method based on sparse low-rank decomposition described in this invention, the sparsification process includes determining the number of rows and columns used for sparsification based on the reduction rate and the sparsity rate. The rows and columns corresponding to the unimportant input and output neurons in the decomposition matrix are represented by reduced rank, such that the element values ​​in these rows and columns are reduced to zero.

[0010] As a preferred embodiment of the distribution network hidden danger identification model compression method based on sparse low-rank decomposition described in this invention, the importance score is calculated based on the absolute value of the weights. For the input neuron, the importance score is obtained by summing the absolute values ​​of the weights of its received connections. For the output neuron, the importance score is obtained by summing the absolute values ​​of the weights of its sending connections.

[0011] As a preferred embodiment of the distribution network hidden danger identification model compression method based on sparse low-rank decomposition described in this invention, the following is provided: a minimization problem is constructed, the objective function is to minimize the sum of importance scores of the selected rows, and the constraint condition is that the number of selected rows is equal to rm; Construct a minimization problem with the objective function being to minimize the sum of importance scores for the selected columns, and the constraint being that the number of selected columns equals rn; By solving these minimization problems, the rows and columns that need to be sparsified are determined.

[0012] As a preferred embodiment of the distribution network hidden danger identification model compression method based on sparse low-rank decomposition described in this invention, wherein: screening rm row sum in matrix The sparsification process for the rn columns in a matrix is ​​transformed into a minimization problem: selecting a set of rows and columns that satisfy the following conditions. in, and For binary variables, to minimize the objective function, all of them must satisfy the condition that " or The sum of is equal to rm or rn, and each or The constraints "can only take 0 or 1" are given, where i and j are respectively... Matrix row sum The matrix column index; the selected row and column will be represented by reduced rank rk.

[0013] This invention provides a compression system for a distribution network hazard identification model based on sparse low-rank decomposition.

[0014] As a preferred embodiment of the power distribution network hidden danger identification model compression system based on sparse low-rank decomposition described in this invention, it includes a model decomposition module, a neuron screening module, a matrix sparsification module, and a model reconstruction test module. The model decomposition module is responsible for performing singular value decomposition on the weight matrix of the fully connected layer of the trained convolutional neural network. By truncating the matrix, the most important k singular values ​​are retained to obtain a low-rank approximate matrix. The neuron selection module calculates the importance scores of input and output neurons based on the absolute value of weights, the absolute value of activation values, or the absolute change of cost, and selects unimportant neurons according to a preset sparsity rate. The matrix sparsification module determines the number of rows and columns to be sparsified based on the output of the neuron selection module, and performs rank reduction processing on the rows and columns corresponding to unimportant neurons in the decomposition matrix, setting their element values ​​to zero, thereby achieving structured sparsity of the matrix. The model reconstruction test module reconstructs the sparsed decomposition matrix into a compressed weight matrix, replaces the original fully connected layer weights, and outputs the final compressed convolutional neural network model.

[0015] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of a method for compressing a distribution network hazard identification model based on sparse low-rank decomposition.

[0016] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of a method for compressing a distribution network hazard identification model based on sparse low-rank decomposition are implemented.

[0017] The beneficial effects of this invention are as follows: A traditional low-rank decomposition method—Singular Value Decomposition (SVD)—is applied to the weight matrix of the fully connected layers of a convolutional neural network. A neuron selection strategy is used to identify input and output neurons that require sparsification. The components in the decomposition matrix corresponding to the selected neurons are then sparsified. Finally, the functionality of the proposed method is tested. Test results show that this method has good model compression performance and can generate convolutional neural network models that balance model accuracy and size, providing a new approach for model compression in convolutional neural networks. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of a method for compressing a distribution network hazard identification model based on sparse low-rank decomposition, provided in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the low-rank decomposition of the distribution network hidden danger identification model compression method based on sparse low-rank decomposition provided in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of low-rank matrix sparsification in a distribution network hidden danger identification model compression method based on sparse low-rank decomposition provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0023] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for compressing a distribution network hazard identification model based on sparse low-rank decomposition, including: S1: Apply singular value decomposition (SVD) to the weight matrix of the fully connected layer of the convolutional neural network to obtain the decomposed matrix.

[0024] S2: Use a neuron screening strategy to select input and output neurons that need to be sparsified.

[0025] S3: Perform sparsification on the components in the decomposition matrix corresponding to the selected neurons to compress the model.

[0026] It should be noted that, addressing the memory and computational resource bottlenecks faced by existing deep learning models when deployed on edge devices, a sparse low-rank decomposition model compression method is proposed by introducing a neuron importance evaluation mechanism to optimize the traditional singular value decomposition method. This method, while maintaining model accuracy, effectively solves the deployment challenge of power distribution network hazard identification models in resource-constrained environments by performing low-rank decomposition on the weight matrix of fully connected layers and applying structured sparse processing to the decomposition matrix based on neuron importance.

[0027] Example 2 is an embodiment of the present invention. Based on the above embodiment, a method for compressing the distribution network hidden danger identification model based on sparse low-rank decomposition is provided.

[0028] Furthermore, in this embodiment, step S1 applies singular value decomposition (SVD) to the weight matrix of the fully connected layer of the convolutional neural network to obtain the decomposed matrix. The specific steps include: The traditional low-rank decomposition method, Singular Value Decomposition (SVD), is applied to the weight matrix of the fully connected layer. By optimizing the inference speed of convolution operations through low-rank decomposition, the weights can be stored more efficiently.

[0029] Convolutional neural networks typically contain a large number of convolutional layers and fully connected layers. Convolutional layers consume significant time during inference, while fully connected layers require substantial storage space. Applying low-rank decomposition to the convolutional kernel weights primarily aims to optimize the inference speed of convolution operations—because multiplying by the decomposed matrix takes less time compared to multiplying by the 3D convolutional kernel weights. In contrast, the proposed method focuses on the decomposition of fully connected layers, with the core objective of optimizing weight storage while also improving computational efficiency through decomposition. In a fully connected layer containing m input neurons and n output neurons, the activation values ​​of the n nodes in this layer... It can be represented as: Where X is the input of this layer. Let be any activation function. A fully connected layer contains a weight matrix. and a bias vector b Each parameter in the weight matrix (satisfy , Each element in the bias vector (satisfy The proposed method processes the weight matrix W after the neural network is fully trained. Using the SVD method, the weight matrix can be decomposed into: in, and It is an orthogonal matrix. It is a diagonal matrix.

[0030] Furthermore, the matrix S obtained from SVD decomposition is always a diagonal matrix containing n non-negative singular values, with the most important singular values ​​arranged in descending order from the top left to the bottom right. Truncating SVD retains the k most important singular values, at which point the resulting matrices U, S, and... They become , and Therefore, the original weight matrix W can be approximately reconstructed as: Decomposition Matrix The diagonal elements (i.e., non-negative singular values) represent k potential concepts associated with the parameters of the weight matrix W. The magnitude of the singular values ​​in the weight matrix W reflects the importance of the k-th latent factor. Each row of the matrix represents the correlation between each input neuron and these k potential concepts; Each column of the matrix represents the relevance of each output neuron to the k latent concepts. Truncated SVD is a variant of SVD that truncates the relevant data. and The columns and rows are then processed, and unimportant zero elements are removed. At this point, the principle for choosing the value of k is: to minimize the loss of accuracy of the compressed test set while achieving a lower compression ratio (cr) and a lower reconstruction error (...). The rank k of the decomposition can be determined by compressing the matrix with different k values, analyzing the decrease in accuracy of the test set corresponding to each k value, and then determining the rank k of the decomposition.

[0031] In an optional embodiment, the rank k of the decomposition can also be determined by directly presetting the value of k to a fixed proportion of the rank of the original weight matrix.

[0032] In an optional embodiment, determining the rank k of the decomposition can also involve setting a predefined singular value threshold, retaining all components whose singular values ​​exceed the threshold, determining the k value, and avoiding multiple experimental comparisons.

[0033] Furthermore, the truncated matrix can be adjusted based on the importance of the m input neurons and n output neurons. and Further compression. Partial lines and A subset of columns (less important input and output neurons) are represented by a rank less than k (called "reduced rank" rk). In rows and columns represented by reduced rank, the corresponding element values ​​are reduced to zero; this reduction operation is called "sparserization." and When sparsifying matrix elements, it is essential to ensure that the rank of a single neuron is less than k; the L1 norm cannot be used directly. and Sparsity is achieved by setting elements with smaller amplitudes to zero. Because... and Elements with smaller amplitudes may be randomly distributed in various places. The leftmost column of the matrix (corresponding to important latent concepts) or Sparsification of the elements in the top row of a matrix (unstructured sparsification) severely impacts the matrix reconstruction, leading to a decrease in model accuracy. Therefore, vector-level sparsification is more necessary for matrix decomposition.

[0034] Furthermore, the reduced rank rk of unimportant rows and columns is determined by the reduction rate. (satisfy The decision is made using the following formula: The number of sparse rows (rm) and sparse columns (rn) that need to be represented by reduced rank rk are determined by the sparsity ratio. satisfy( The determination and calculation formulas are as follows: and The proposed SLR method assumes that the proportion of unimportant neurons in the input and output neurons is the same, therefore... row sum of a matrix The columns of the matrix are reduced by the same rate. For the matrix... , and After sparsification, the total number of non-zero parameters in these three matrices is This value is less than the total number of non-zero parameters in the truncated SVD. .

[0035] Furthermore, in this embodiment, step S2 employs a neuron screening strategy to select input neurons and output neurons that require sparsification processing. The specific steps include: A neuron screening strategy was used to identify the input and output neurons that needed to be sparsified.

[0036] In an optional embodiment, since the neural network is trained from small random weights, the larger the magnitude of the weights between neurons, the higher the importance of the corresponding connection. Therefore, the importance of the connections associated with the input and output neurons can be assessed by analyzing the absolute values ​​of the weights. The calculation is based on the rows and columns of the weight matrix W, and the formula based on the absolute values ​​of the weights is as follows: in, The importance of the i-th input neuron is represented by summing the absolute values ​​of the weights of its receiving connections. The importance of the j-th output neuron is represented by summing the absolute values ​​of the weights of its sending connections.

[0037] In another optional embodiment, the activation state of neurons in the network reflects whether the neuron has detected a specific feature from the input; that is, the activation value represents the similarity between the input and the neuron's weights. The importance (is) and importance (os) of the input and output neurons can be obtained by summing the absolute values ​​of the activation values ​​of all samples in the training set, as shown in the following formula: in, It contains the activation values ​​of the m input neurons corresponding to p samples; This is the output neuron activation matrix, containing the activation values ​​of n output neurons.

[0038] If removing a neuron does not change the network performance, that neuron is considered to have low importance. The importance (is) and importance (os) of input and output neurons can be determined by... Specific lines or The absolute change in cost value after truncating a specific column using reduced rank rk is measured by the following formula: Where c is the baseline cost of the neural network using the original training weights W on p training samples; let To reconstruct weights using sparse SVD (in The i-th row or The network cost is represented by the reduced rank rk in the j-th column.

[0039] Furthermore, in this embodiment, step S3 performs sparsification on the components in the decomposition matrix corresponding to the selected neurons to compress the model. Specific steps include: For matrix and Perform sparsification. Screening. rm row sum in matrix The process of sparsifying the rn columns in a matrix can be transformed into a minimization problem, selecting a set of rows and columns that satisfy the following conditions: in, and For binary variables, to minimize the objective function, all of them must satisfy the condition that " or The sum of is equal to rm or rn, and each or The constraint condition is that the values ​​can only be 0 or 1. i and j are respectively... Matrix row sum The matrix column index; the selected row and column will be represented by reduced rank rk.

[0040] In an optional embodiment, the sparsification operation can also be performed by sorting the input and output neurons in ascending order based on their importance scores, and directly selecting the top-ranked rm (input neurons) and rn (output neurons) neurons as the sparsification targets.

[0041] In another alternative embodiment, the sparsification operation can also preset an importance score threshold, and automatically select all neurons with importance scores below the threshold as targets for sparsification.

[0042] Performance tests were conducted on the proposed method. The results show that the proposed method can achieve a compression ratio reduction of up to 3.6 times when compressing SVD decomposition results, while causing only a slight or no loss of accuracy. This means that, without significantly affecting model performance, the method can save 3.6 times the storage space required for SVD decomposition. This demonstrates that the proposed method can effectively compress CNN models, has high model compression efficiency, and can generate better CNN models.

[0043] Example 3 is the third embodiment of the present invention, which differs from the previous two embodiments in that: This embodiment also provides a distribution network hidden danger identification model compression system based on sparse low-rank decomposition, including: a model decomposition module, a neuron screening module, a matrix sparsification module, and a model reconstruction test module; The model decomposition module is responsible for performing singular value decomposition on the weight matrix of the fully connected layer of the trained convolutional neural network. By truncating the matrix, the most important k singular values ​​are retained to obtain a low-rank approximate matrix. The neuron selection module calculates the importance scores of input and output neurons based on the absolute value of weights, the absolute value of activations, or the absolute change of cost, and selects out unimportant neurons according to a preset sparsity rate. The matrix sparsification module determines the number of rows and columns to be sparsified based on the output of the neuron selection module, and performs rank reduction processing on the rows and columns corresponding to unimportant neurons in the decomposition matrix, setting their element values ​​to zero to achieve structured sparsity of the matrix. The model reconstruction test module reconstructs the sparsed decomposition matrix into a compressed weight matrix, replaces the original fully connected layer weights, and outputs the final compressed convolutional neural network model.

[0044] This embodiment also provides an electronic device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the power distribution network hidden danger identification model compression method based on sparse low-rank decomposition proposed in the above embodiment.

[0045] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the distribution network hidden danger identification model compression method based on sparse low-rank decomposition proposed in the above embodiment.

[0046] The storage medium proposed in this embodiment and the compression method for the distribution network hidden danger identification model based on sparse low-rank decomposition proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0047] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0048] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A compression method for distribution network hidden danger identification model based on sparse low-rank decomposition, characterized in that: include, Singular value decomposition is performed on the weight matrix of the fully connected layer in the convolutional neural network to obtain three decomposition matrices. The decomposition matrix is ​​truncated using SVD, retaining the k most important singular values ​​to obtain a low-rank approximate representation; A neuron screening strategy was employed to identify unimportant input and output neurons by evaluating their importance scores. Based on the preset sparsity and reduction rates, determine the number of rows and columns to be sparsified and the corresponding reduction rank; The rows and columns corresponding to the selected unimportant neurons in the decomposition matrix are sparsified to compress the model, resulting in a compressed convolutional neural network model.

2. The method for compressing the distribution network hidden danger identification model based on sparse low-rank decomposition as described in claim 1, characterized in that: The singular value decomposition includes decomposing the weight matrix into a first orthogonal matrix, a diagonal matrix, and a second orthogonal matrix using SVD. By retaining the k most important singular values ​​in the diagonal matrix, the decomposition matrix is ​​truncated to obtain a low-rank approximation matrix, where k is the rank chosen based on the compression ratio and reconstruction error.

3. The method for compressing the distribution network hidden danger identification model based on sparse low-rank decomposition as described in claim 2, characterized in that: The neuron selection strategy includes calculating the importance scores of input neurons and output neurons, and selecting unimportant neurons based on the importance scores. The importance score is calculated based on either the absolute value of the weight, the absolute value of the activation value, or the absolute change in cost.

4. The method for compressing the distribution network hidden danger identification model based on sparse low-rank decomposition as described in claim 3, characterized in that: The sparsification process includes determining the number of rows and columns to be used for sparsification based on the reduction rate and the sparsity rate; The rows and columns corresponding to the unimportant input and output neurons in the decomposition matrix are represented by reduced rank, such that the element values ​​in these rows and columns are reduced to zero.

5. The method for compressing the distribution network hidden danger identification model based on sparse low-rank decomposition as described in claim 4, characterized in that: Importance scores are calculated based on the absolute values ​​of the weights. For an input neuron, the importance score is obtained by summing the absolute values ​​of the weights of its receiving connections. For the output neuron, the importance score is obtained by summing the absolute values ​​of the weights of its sending connections.

6. The method for compressing the distribution network hidden danger identification model based on sparse low-rank decomposition as described in claim 5, characterized in that: Construct a minimization problem with the objective function being to minimize the sum of importance scores for the selected rows, and the constraint being that the number of selected rows equals rm; Construct a minimization problem with the objective function being to minimize the sum of importance scores for the selected columns, and the constraint being that the number of selected columns equals rn; By solving these minimization problems, the rows and columns that need to be sparsified are determined.

7. The method for compressing the distribution network hidden danger identification model based on sparse low-rank decomposition as described in claim 6, characterized in that: filter rm row sum in matrix The sparsification process for the rn columns in a matrix is ​​transformed into a minimization problem: selecting a set of rows and columns that satisfy the following conditions. in, and For binary variables, to minimize the objective function, all of them must satisfy the condition " or The sum of is equal to rm or rn, and each or The constraints "can only take 0 or 1" are given, where i and j are respectively... Matrix row sum The matrix column index; the selected row and column will be represented by reduced rank rk.

8. A distribution network hazard identification model compression system based on sparse low-rank decomposition, wherein the method for compressing the distribution network hazard identification model based on sparse low-rank decomposition as described in any one of claims 1 to 7 is characterized in that, include: The module includes a model decomposition module, a neuron selection module, a matrix sparsification module, and a model reconstruction testing module. The model decomposition module is responsible for performing singular value decomposition on the weight matrix of the fully connected layer of the trained convolutional neural network. By truncating the matrix, the most important k singular values ​​are retained to obtain a low-rank approximate matrix. The neuron selection module calculates the importance scores of input and output neurons based on the absolute value of weights, the absolute value of activation values, or the absolute change of cost, and selects unimportant neurons according to a preset sparsity rate. The matrix sparsification module determines the number of rows and columns to be sparsified based on the output of the neuron selection module, and performs rank reduction processing on the rows and columns corresponding to unimportant neurons in the decomposition matrix, setting their element values ​​to zero, thereby achieving structured sparsity of the matrix. The model reconstruction test module reconstructs the sparsed decomposition matrix into a compressed weight matrix, replaces the original fully connected layer weights, and outputs the final compressed convolutional neural network model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for compressing the distribution network hidden danger identification model based on sparse low-rank decomposition as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for compressing the distribution network hidden danger identification model based on sparse low-rank decomposition as described in any one of claims 1 to 7.

Images