A power transmission line insulator defect detection method, system, device and medium

Abstract

The application discloses a power transmission line insulator defect detection method, system, device and medium. In the application, when a power transmission line insulator image is received, the power transmission line insulator image is preprocessed, a power transmission line insulator sample set is generated, a preset convolutional neural network model is approximated based on an SVD strategy, an initial insulator defect detection model is generated, the initial insulator defect detection model is pruned according to a cross-layer strategy, a target insulator defect detection model is generated, defect detection is performed on a to-be-recognized power transmission line insulator image through the target insulator defect detection model, and a defect detection result is output. The technical problem that the existing target detection model has low recognition efficiency when channel pruning is performed is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of neural networks, and particularly relates to a power transmission line insulator defect detection method, system, device and medium. BACKGROUND

[0002] In recent years, deep learning techniques, especially convolutional neural networks (CNN), have shown excellent performance on a range of artificial intelligence tasks, such as image classification and object detection. However, the high demand for hardware resources hinders the combination of Internet of Things and deep learning, and in some time-sensitive tasks, such as intelligent monitoring, autonomous driving and health monitoring, uploading data to a cloud server will inevitably cause information delay. In order to achieve real-time reaction, it is essential to use edge computing and embedded devices to make processing efficient and hardware friendly.

[0003] At present, the hardware implementation based on CNN is mainly limited by model storage, computing cost and memory occupation. In order to obtain the benefits of actual acceleration on general platforms, it is necessary to ensure the compact structure of processing convolutional layers, and the existing power transmission line insulator defect detection is usually performed by channel pruning on a convolutional neural network to generate an object detection model, but during channel pruning, each filter may be either completely deleted or retained, so that pruning the wrong filter may discard a large number of important weights and cause irreparable damage to the performance of the object detection model, and channel pruning cannot utilize weight-level sparsity, which may result in low recognition efficiency of the object detection model. SUMMARY

[0004] The present application provides a power transmission line insulator defect detection method, system, device and medium, which solves the technical problem that the existing object detection model may be either completely deleted or retained during channel pruning, so that pruning the wrong filter may discard a large number of important weights and cause irreparable damage to the performance of the object detection model, and channel pruning cannot utilize weight-level sparsity, which may result in low recognition efficiency of the object detection model.

[0005] The first aspect of the present application provides a power transmission line insulator defect detection method, comprising:

[0006] When receiving a power transmission line insulator image, performing image preprocessing on the power transmission line insulator image to generate a power transmission line insulator sample set;

[0007] Based on the SVD strategy, performing low-rank approximation on the preset convolutional neural network model to generate an initial insulator defect detection model;

[0008] Taylor pruning is performed on the initial insulator defect detection model according to the cross-layer strategy to generate the target insulator defect detection model;

[0009] The target insulator defect detection model is used to detect defects in the image of the transmission line insulator to be identified, and the defect detection results are output.

[0010] Optionally, the step of generating an initial insulator defect detection model by performing a low-rank approximation on a preset convolutional neural network model based on the SVD strategy includes:

[0011] Singular value decomposition is used to decompose each convolutional layer in the preset convolutional neural network model to generate multiple initial convolutional layers.

[0012] A preset number of maximum singular values ​​are selected to perform dimensionality reduction on each of the initial convolutional layers, generating multiple intermediate convolutional layers;

[0013] Based on the retained singular components, multiplication operations are performed on each of the intermediate convolutional layers to generate multiple target convolutional layers;

[0014] The preset convolutional neural network model containing multiple target convolutional layers is used as the initial insulator defect detection model.

[0015] Optionally, the step of performing Taylor pruning on the initial insulator defect detection model according to a cross-layer strategy to generate the target insulator defect detection model includes:

[0016] The target convolutional layer for Taylor pruning is selected from the initial insulator defect detection model according to the cross-layer strategy;

[0017] A preset number of transmission line insulator images are extracted from the transmission line insulator sample set and input into the initial insulator defect detection model for training. Multiple target feature maps corresponding to the target convolutional layer are generated based on the training results.

[0018] Calculate the pruning loss value between each target feature map and the preset standard feature map;

[0019] Obtain the initial loss value associated with the preset convolutional neural network model;

[0020] The difference between the initial loss value and the trimming loss value is calculated to generate the initial importance data of each filter in the initial insulator defect detection model.

[0021] The initial importance data are normalized to generate corresponding target importance scores;

[0022] The importance scores of the targets within each target convolutional layer are sorted.

[0023] Compare the importance scores of each target with the associated preset standard importance scores;

[0024] If the target importance score is greater than the preset standard importance score, then the filter associated with the target importance score in the target convolutional layer is retained to generate a target insulator defect detection model.

[0025] Optionally, it also includes:

[0026] The initial insulator defect detection model is trained by inputting a preset test sample set, and corresponding test indicators are generated based on the test results. The test indicators include target pruning probability index, target compression ratio, and target acceleration ratio.

[0027] Compare the test indicators with the preset standard indicator conditions;

[0028] If the test index meets the preset standard index conditions, then training is stopped and a target insulator defect detection model is generated.

[0029] If the test index does not meet the preset standard index conditions, the network parameters of the initial insulator defect detection model are adjusted according to the preset gradient.

[0030] The process jumps to the step of performing a low-rank approximation of the preset convolutional neural network model based on the SVD strategy to generate an initial insulator defect detection model, until the test index meets the preset standard index conditions, and then optimizes the target insulator defect detection model.

[0031] Optionally, the step of performing defect detection on the image of the transmission line insulator to be identified using the target insulator defect detection model and outputting the defect detection result includes:

[0032] The target insulator defect detection model is used to detect defects in the image of the transmission line insulator to be identified, and a target identification block diagram is generated.

[0033] Calculate the target overlap between the target identifier diagram and the preset defective insulator diagram;

[0034] Compare the target overlap with a preset overlap threshold;

[0035] If the target overlap is less than the preset overlap threshold, it is determined that the insulator associated with the image of the insulator to be identified has no defects.

[0036] If the target overlap is greater than or equal to the preset overlap threshold, then the defect type associated with the preset defect insulator diagram is taken as the target defect.

[0037] A second aspect of the present invention provides a defect detection system for transmission line insulators, comprising:

[0038] The sample set module is used to perform image preprocessing on the received transmission line insulator image to generate a transmission line insulator sample set.

[0039] The initial insulator defect detection model module is used to generate an initial insulator defect detection model by performing a low-rank approximation on a preset convolutional neural network model based on the SVD strategy.

[0040] The target insulator defect detection model module is used to perform Taylor pruning on the initial insulator defect detection model according to the cross-layer strategy to generate the target insulator defect detection model.

[0041] The defect detection module is used to perform defect detection on the image of the transmission line insulator to be identified using the target insulator defect detection model, and output the defect detection results.

[0042] Optionally, the initial insulator defect detection model module includes:

[0043] The initial convolutional layer submodule is used to decompose each convolutional layer in the preset convolutional neural network model using singular value decomposition to generate multiple initial convolutional layers.

[0044] The intermediate convolutional layer submodule is used to select a preset number of maximum singular values ​​to perform dimensionality reduction operations on each of the initial convolutional layers, thereby generating multiple intermediate convolutional layers;

[0045] The target convolutional layer submodule is used to perform multiplication operations on each of the intermediate convolutional layers based on the retained singular components to generate multiple target convolutional layers;

[0046] The convolutional layer update submodule is used to use the preset convolutional neural network model containing multiple target convolutional layers as the initial insulator defect detection model.

[0047] Optionally, it also includes:

[0048] The test index module is used to train the initial insulator defect detection model by inputting a preset test sample set, and to generate corresponding test indexes based on the test results. The test indexes include target pruning probability index, target compression ratio, and target acceleration ratio.

[0049] The indicator comparison module is used to compare the test indicators with preset standard indicator conditions.

[0050] The first training module is used to stop training and generate a target insulator defect detection model if the test index meets the preset standard index conditions.

[0051] The second training module is used to adjust the network parameters of the initial insulator defect detection model according to a preset gradient if the test index does not meet the preset standard index conditions.

[0052] The jump module is used to jump to execute the step of performing a low-rank approximation of the preset convolutional neural network model based on the SVD strategy to generate an initial insulator defect detection model, until the test index meets the preset standard index condition, and optimize the target insulator defect detection model.

[0053] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the transmission line insulator defect detection method as described in any of the preceding claims.

[0054] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed, it implements the method for detecting defects in transmission line insulators as described in any of the preceding claims.

[0055] As can be seen from the above technical solutions, the present invention has the following advantages:

[0056] In this invention, when a transmission line insulator image is received, image preprocessing is performed to generate a transmission line insulator sample set. Based on the SVD strategy, a low-rank approximation is performed on a preset convolutional neural network model to generate an initial insulator defect detection model. Taylor pruning is then performed on the initial insulator defect detection model according to a cross-layer strategy to generate a target insulator defect detection model. The target insulator defect detection model is used to detect defects in the transmission line insulator image to be identified, and the defect detection result is output. This invention solves the technical problem that existing target detection models may have each filter either completely deleted or retained during channel pruning. Therefore, pruning incorrect filters may discard a large number of important weights and cause irreparable damage to the performance of the target detection model. Moreover, channel pruning cannot utilize the sparsity of weight levels, which leads to low recognition efficiency of the target detection model. This invention achieves acceleration by using low-rank approximation to eliminate redundancy within the filter and converting the filter into a smaller, more compact filter. Channel pruning is applied globally on the approximation network, resulting in a better compressed and accelerated low-rank target insulator defect detection model. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0058] Figure 1 This is a flowchart of the steps of a method for detecting defects in insulators of power transmission lines provided in Embodiment 1 of the present invention;

[0059] Figure 2 This is a schematic diagram of the framework of a target insulator defect detection model provided in Embodiment 1 of the present invention;

[0060] Figure 3 This is a flowchart of the steps in a method for detecting defects in transmission line insulators according to Embodiment 2 of the present invention;

[0061] Figure 4 This is a flowchart of a process for constructing a target insulator defect detection model, provided in Embodiment 2 of the present invention.

[0062] Figure 5 This is a schematic diagram of a smaller filter decomposed in the model of a method for detecting defects in transmission line insulators provided in Embodiment 2 of the present invention;

[0063] Figure 6 This is a structural block diagram of a power transmission line insulator defect detection system provided in Embodiment 3 of the present invention. Detailed Implementation

[0064] This invention provides a method, system, device, and medium for detecting defects in transmission line insulators. It addresses the technical problem that existing target detection models may have each filter either completely deleted or retained during channel pruning. Therefore, pruning incorrect filters may discard a large number of important weights and cause irreparable damage to the performance of the target detection model. Moreover, channel pruning cannot utilize the sparsity of weight levels, which leads to low recognition efficiency of the target detection model.

[0065] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0066] Please see Figure 1 ,Figure 1 The flowchart illustrates the steps of a method for detecting defects in insulators of power transmission lines, as provided in Embodiment 1 of the present invention.

[0067] This invention provides a method for detecting defects in transmission line insulators, comprising:

[0068] Step 101: When the image of the transmission line insulator is received, perform image preprocessing on the image to generate a transmission line insulator sample set.

[0069] The transmission line insulator sample set refers to an image set constructed by acquiring image data of transmission line insulators in multiple scenarios.

[0070] Image preprocessing refers to sorting and cleaning up acquired image data that has abnormal brightness, excessive noise, or is blurry, and then performing image correction, image enhancement, and annotation operations.

[0071] Image correction refers to automatically detecting the tilt direction and tilt angle of an image based on its features, and then correcting the image position based on the detected information.

[0072] Image enhancement refers to enhancing the useful information in an image. It can be a distortion-based process aimed at improving the visual quality of the image, tailored to the specific application. It involves intentionally emphasizing the overall or local characteristics of an image, making an originally blurry image clearer, highlighting certain features of interest, amplifying the differences between features of different objects in the image, and suppressing features of little interest. This improves image quality, enriches information, enhances image interpretation and recognition, and meets the needs of specific analyses.

[0073] In this embodiment of the invention, when a transmission line insulator image is received, the transmission line insulator image set is sorted, data with abnormal brightness, high noise, and blurred images are cleaned, and then image correction, image enhancement, and annotation operations are performed to generate a transmission line insulator sample set.

[0074] Step 102: Based on the SVD strategy, perform low-rank approximation on the preset convolutional neural network model to generate an initial insulator defect detection model.

[0075] The Singular Value Decomposition (SVD) strategy is used to decompose a two-dimensional matrix that needs to be reshaped within a pre-defined convolutional neural network model that requires compression.

[0076] Low-rank approximation refers to decomposing a high-dimensional matrix in a pre-defined convolutional neural network model into multiple low-dimensional matrices, thereby reducing the storage space and computational complexity of the matrices.

[0077] In this embodiment of the invention, based on singular value decomposition, the two-dimensional matrix reshaped by each convolutional layer in the preset convolutional neural network model is decomposed to construct an approximate convolutional layer. The preset convolutional neural network model is then updated using the approximate convolutional layer to generate an initial insulator defect detection model.

[0078] Step 103: Perform Taylor pruning on the initial insulator defect detection model according to the cross-layer strategy to generate the target insulator defect detection model.

[0079] The cross-layer strategy refers to performing cross-layer pruning on all convolutional layers within the initial insulator defect detection model.

[0080] The target insulator defect detection model refers to the insulator defect detection model generated after Taylor pruning the initial insulator defect detection model.

[0081] In this embodiment of the invention, the convolutional layers in the initial insulator defect detection model are pruned across layers to remove convolutional layers containing unimportant filters, thereby forming the target insulator defect detection model from the remaining convolutional layers containing effective filters.

[0082] Step 104: Perform defect detection on the image of the transmission line insulator to be identified using the target insulator defect detection model, and output the defect detection results.

[0083] In this embodiment of the invention, when an image of a transmission line insulator to be identified is received, a defect detection model is used to detect defects in the image. The target identifier diagram is parsed, and the target overlap between the target identifier diagram and the preset defect insulator diagram is calculated. If the target overlap is less than the preset overlap threshold, it is determined that the transmission line insulator associated with the image to be identified has no defects. If the target overlap is greater than or equal to the preset overlap threshold, the defect type associated with the preset defect insulator diagram is taken as the target defect, and the defect detection result is generated and displayed.

[0084] In this invention, when a transmission line insulator image is received, image preprocessing is performed to generate a transmission line insulator sample set. Based on the SVD strategy, a low-rank approximation is applied to a preset convolutional neural network model to generate an initial insulator defect detection model. Taylor pruning is then performed on the initial insulator defect detection model according to a cross-layer strategy to generate a target insulator defect detection model. The target insulator defect detection model is then used to detect defects in the transmission line insulator image to be identified, and the defect detection result is output. This invention solves the technical problem that existing target detection models may suffer from issues during channel pruning, where each filter is either completely deleted or retained. Therefore, pruning incorrect filters may discard a large number of important weights, causing irreparable damage to the target detection model's performance. Furthermore, channel pruning cannot utilize weight sparsity, leading to low recognition efficiency. This invention achieves acceleration by using low-rank approximation to eliminate redundancy within filters and converting filters into smaller, more compact filters. Channel pruning is applied globally on the approximation network, resulting in a better compressed and accelerated low-rank target insulator defect detection model. Please refer to [link to relevant documentation]. Figure 2 This framework improves channel pruning through low-rank approximation. It eliminates redundancy at both the weight and filter levels, avoiding aggressive network processing. The entire framework is as follows: Figure 2 As shown.

[0085] Please see Figure 3 , Figure 3 This is a flowchart illustrating the steps of a method for detecting defects in insulators of power transmission lines, as provided in Embodiment 2 of the present invention.

[0086] This invention provides a method for detecting defects in transmission line insulators, comprising:

[0087] Step 201: When the image of the transmission line insulator is received, perform image preprocessing on the image to generate a transmission line insulator sample set.

[0088] In this embodiment of the invention, the specific implementation process of step 201 is similar to that of step 101, and will not be repeated here.

[0089] Step 202: Use singular value decomposition to decompose each convolutional layer in the preset convolutional neural network model to generate multiple initial convolutional layers.

[0090] It is worth mentioning that the low-rank approximation based on SVD achieves a low-rank approximation for each convolutional layer in the preset convolutional neural network model.

[0091] In this embodiment of the invention, singular value decomposition is used to decompose each convolutional layer in the preset convolutional neural network model to generate multiple initial convolutional layers.

[0092] Step 203: Select a preset number of maximum singular values ​​to perform dimensionality reduction on each initial convolutional layer to generate multiple intermediate convolutional layers.

[0093] In this embodiment of the invention, a preset number of maximum singular values ​​are selected to perform dimensionality reduction on each initial convolutional layer, generating multiple intermediate convolutional layers.

[0094] Step 204: Based on the retained singular components, perform multiplication operations on each intermediate convolutional layer to generate multiple target convolutional layers.

[0095] In this embodiment of the invention, based on the retained singular components, a multiplication operation is performed on each intermediate convolutional layer to generate multiple target convolutional layers.

[0096] In the specific implementation of steps 202-204, the weight tensor W∈R corresponding to the preset convolutional neural network model is... N ×d×d×C Reshape into a two-dimensional matrix W∈R Nd×dC Where N and C are the number of output channels and input channels, respectively, d is the spatial dimension of the filter, and the singular values ​​of W are expressed as σ1≥σ2≥…≥σ n ≥0, and then the initial weight matrix W of the convolutional layer is obtained using singular value decomposition, as follows:

[0097] W = USV

[0098] Where S∈R n×n It is a diagonal matrix with singular values, U∈R Nd×n ,V∈R n×dC U and V refer to the left and right eigenvector matrices, respectively.

[0099] It is worth mentioning that for a general matrix, singular value decomposition can be used. A general matrix can be decomposed as follows: U is the left singular vector, V is the right singular vector, and the non-zero values ​​in D are called singular values.

[0100] The approximate W is obtained by truncating S by selecting a predetermined number of the largest singular values ​​from multiple singular values, and the weight matrix of the target convolutional layer is obtained after processing. Specifically as follows:

[0101]

[0102] in Store the k diagonal values ​​of S. Keep k corresponding column vectors of U, Keep the k corresponding row vectors of V, where n represents the number of singular values.

[0103] It's worth noting that the preset number of maximum singular values ​​refers to the k maximum singular values ​​in the diagonal matrix. Truncation refers to generating a decomposition matrix of a specified dimension using dimensionality reduction. In this way, the initial convolutional layer is divided by two... The weighted low-rank convolutional layers are replaced. With an appropriate rank k, the low-rank approximation achieves a reduction in parameters and FLOPs.

[0104] The retained singular components refer to the singular values ​​remaining after selecting a predetermined number of the largest singular values ​​from all singular values. These retained singular components σ... i ,σ j Multiply by u respectively i and v j To construct an approximate convolutional layer, u i Let v represent the i-th decomposition matrix. j The j-th decomposition matrix is ​​shown below:

[0105]

[0106]

[0107] Where, u′ i After approximation The i-th column vector, v′ j After approximation The i-th column vector, u i express The i-th column vector, v j express The i-th column vector.

[0108] Step 205: Use the preset convolutional neural network model containing multiple target convolutional layers as the initial insulator defect detection model.

[0109] In this embodiment of the invention, after performing low-rank approximation on the convolutional layers in the preset convolutional neural network model, a convolutional neural network model containing the target convolutional layer is obtained as the initial insulator defect detection model.

[0110] It is worth mentioning that the low-rank approximation based on SVD achieves a low-rank approximation for each convolutional layer, thereby accelerating the convolution calculation of transmission line insulator defect data.

[0111] Step 206: Perform Taylor pruning on the initial insulator defect detection model according to the cross-layer strategy to generate the target insulator defect detection model;

[0112] Furthermore, step 206 includes the following sub-steps:

[0113] S11. Select the target convolutional layer for Taylor pruning from the initial insulator defect detection model according to the cross-layer strategy.

[0114] In this embodiment of the invention, the convolutional layers in the initial insulator defect detection model are pruned across layers to remove convolutional layers containing unimportant filters.

[0115] S12. Extract a preset number of transmission line insulator images from the transmission line insulator sample set and input them into the initial insulator defect detection model for training. Generate multiple target feature maps corresponding to the target convolutional layer based on the training results.

[0116] In this embodiment of the invention, a preset number of transmission line insulator images are extracted from the transmission line insulator sample set and input into the initial insulator defect detection model for training. Based on the training results, multiple target feature maps corresponding to the target convolutional layer are generated.

[0117] S13. Calculate the pruning loss value between each target feature map and the preset standard feature map.

[0118] The trimming loss value refers to the loss value between the target feature map output by the initial insulator defect detection model and the preset standard feature map.

[0119] In this embodiment of the invention, the pruning loss value between each target feature map and the preset standard feature map is calculated.

[0120] S14. Obtain the initial loss value associated with the preset convolutional neural network model.

[0121] The initial loss value refers to the loss value between the original feature map output by the preset convolutional neural network model and the preset standard feature map.

[0122] In this embodiment of the invention, the initial loss value associated with a preset convolutional neural network model is obtained.

[0123] S15. Perform a difference calculation between the initial loss value and the trimmed loss value to generate the initial importance data of each filter in the initial insulator defect detection model.

[0124] Initial importance data refers to the difference between the cost function of the preset convolutional neural network model and the loss value of the initial insulator defect detection model.

[0125] In this embodiment of the invention, the difference between the initial loss value and the trimming loss value is calculated to generate the initial importance data of each filter in the initial insulator defect detection model.

[0126] S16. Normalize each initial importance data to generate the corresponding target importance score.

[0127] The target importance score refers to the score used to determine the importance of a filter, and is used to judge whether it is an unimportant filter so that it can be eliminated.

[0128] In this embodiment of the invention, each initial importance data is normalized to generate a corresponding target importance score.

[0129] S17. Sort the target importance scores within each target convolutional layer.

[0130] In this embodiment of the invention, the importance scores of multiple targets within each target convolutional layer are sorted.

[0131] S18. Compare the importance scores of each target with the preset standard importance scores of the associated targets.

[0132] The preset standard importance score refers to the scoring threshold used to determine whether or not to apply the filter.

[0133] In this embodiment of the invention, the target importance score associated with each filter is compared with the preset standard importance score.

[0134] S19. If the target importance score is greater than the preset standard importance score, the filters associated with the target importance score in the target convolutional layer will be retained to generate the target insulator defect detection model.

[0135] In this embodiment of the invention, if the target importance score is greater than the preset standard importance score, the filter associated with the target importance score in the target convolutional layer is retained to generate a target insulator defect detection model.

[0136] In the specific implementation, the first-order term of the Taylor expansion is used to estimate the change in the cost function, and the difference between the cost function of the pruned network and the original network is used as the importance score for each filter. The importance score TE is calculated as the difference between the cost function of the pruned network with parameter hi and the original network:

[0137] Θ TE (h i )=|C(h i =0)-C(h i )|

[0138] 1) Where C(hi=0) represents the cost value of pruning parameter hi, while C(hi) represents the cost of keeping hi.

[0139] For the function f(x), the Taylor expansion at x = a is:

[0140]

[0141] Where f (p)(a) is the p-th derivative of f calculated when x = a, and Rp(x) is the p-th remainder.

[0142] 2)C(h i =0) indicates that the cost value of parameter hi has been removed, which is a first-order Taylor polynomial near hi=0:

[0143]

[0144] In particular, the remainder R1 (hi = 0) can be calculated using the Lagrange form:

[0145]

[0146] Where ξ is a real number between 0 and hi.

[0147] By combining 1) and 2), the target importance score TE of the multivariate output is calculated.

[0148] By substituting equation 2) into equation 1), the target importance score TE is calculated as follows:

[0149]

[0150] In practice, TE is computed for multivariate outputs, such as feature maps:

[0151]

[0152] Where M is the length of the vectorized feature map. It is the q-th feature map in layer l.

[0153] Taylor pruning is performed across layers rather than layer by layer, and the score TE is normalized within each layer, as shown in the following formula:

[0154]

[0155] in, This represents the value obtained by normalizing TE. This represents the calculation of the output of the q-th feature map in the l-th layer. This indicates the calculation of the output of the j-th feature map in the l-th layer.

[0156] For a detailed implementation, please refer to Figure 4 The initial insulator defect detection model is obtained by performing a low-rank approximation on the preset convolutional neural network model. The filter of the initial insulator defect detection model is evaluated, and the filter with a value lower than or equal to the preset standard importance score is selected for pruning. The network parameters of the initial insulator defect detection model are adjusted until the target importance score is greater than the preset standard importance score. Then the pruning is stopped, the final recovery rate is determined, and the target insulator defect detection model is generated.

[0157] Step 207: Perform defect detection on the image of the transmission line insulator to be identified using the target insulator defect detection model, and output the defect detection results.

[0158] Furthermore, step 207 includes the following sub-steps:

[0159] S21. Defect detection is performed on the image of the insulator of the transmission line to be identified using the target insulator defect detection model, and a target identification block diagram is generated.

[0160] The target identification diagram refers to the diagram that marks the identification area on the image of the insulator of the transmission line to be identified by the target insulator defect detection model.

[0161] In this embodiment of the invention, a target insulator defect detection model is used to detect defects in the image of the transmission line insulator to be identified, and a target identification block diagram is generated.

[0162] S22. Calculate the target overlap between the target identification block diagram and the preset defect insulator diagram.

[0163] Preset defect insulator diagram refers to the image of a defect insulator.

[0164] In this embodiment of the invention, the target overlap between the target identification block diagram and the preset defective insulator diagram is calculated.

[0165] S23. Compare the target overlap with the preset overlap threshold.

[0166] A preset overlap threshold is used to determine whether the insulators of the transmission line to be identified have defects.

[0167] In this embodiment of the invention, the target overlap is compared with a preset overlap threshold.

[0168] S24. If the target overlap is less than the preset overlap threshold, it is determined that the insulator of the transmission line to be identified associated with the image of the insulator of the transmission line to be identified has no defects.

[0169] In this embodiment of the invention, if the target overlap is less than a preset overlap threshold, it is determined that the insulator of the transmission line to be identified associated with the image of the insulator to be identified has no defects.

[0170] S25. If the target overlap is greater than or equal to the preset overlap threshold, the defect type associated with the preset defect insulator diagram will be used as the target defect.

[0171] In this embodiment of the invention, if the target overlap is greater than or equal to a preset overlap threshold, the defect type associated with the preset defect insulator diagram is taken as the target defect.

[0172] Furthermore, it also includes:

[0173] A1. The initial insulator defect detection model is trained by inputting a preset test sample set, and corresponding test indicators are generated based on the test results. The test indicators include the target pruning probability index, the target compression ratio, and the target acceleration ratio.

[0174] A2. Compare the test indicators with the preset standard indicator conditions.

[0175] A3. If the test indicators meet the preset standard indicator conditions, then stop training and generate the target insulator defect detection model.

[0176] A4. If the test indicators do not meet the preset standard indicators, adjust the network parameters of the initial insulator defect detection model according to the preset gradient.

[0177] A5. Jump to the step of performing a low-rank approximation of the preset convolutional neural network model based on the SVD strategy to generate an initial insulator defect detection model, until the test index meets the preset standard index conditions, and optimize the target insulator defect detection model.

[0178] It's worth noting that the probability P1 of pruning an important filter is based on the maximum entropy principle; when pruning is misjudged, which erroneous filter to prune can be considered an equally probable event. For a network with m0 effective filters, the remaining m1 filters are called effective filters and unimportant filters, respectively. In each iteration, the channel pruning algorithm attempts to prune the least important filter, but errors are unavoidable. Given a certain pruning criterion, assume the probability of misjudging during the pruning iteration is constant, denoted as p. Note that misjudging means selecting any filter, not just the least important one, including any other irrelevant and important filters.

[0179] According to the maximum entropy principle, when a misclassified filter is pruned, which erroneous filter to prune can be considered an equally probable event. For a network with m0 effective filters and the remaining m1 unimportant filters, the probability P1 of pruning an effective filter is:

[0180]

[0181] Where p represents the probability of misjudgment by the pruning algorithm during the pruning iteration, which is a constant.

[0182] To reduce the granularity of the filters, each filter in the pre-defined convolutional neural network is decomposed into s smaller filters. Since the decomposition method (e.g., low-rank approximation) preserves most of the information, it is assumed that the ratio between important and unimportant filters will not change during this process.

[0183] In practice, pruning critical filters results in an equivalent degree of damage compared to pruning critical filters in the original network.

[0184] The target pruning probability index refers to the probability P of pruning important filters in the initial insulator defect detection model. s The specific calculation formula is as follows:

[0185]

[0186] Where s << m0, therefore p′ s Approximate to The bottom decomposition network represents the probability of pruning effective filters, m0 represents the effective filter network, m1 represents the unimportant filter network, s represents the important filter, and p s This indicates that the probability of misclassification during the pruning algorithm's pruning iterations is constant, such as... Figure 5 As shown, smaller filters mitigate the damage from channel pruning and reduce the probability of serious errors.

[0187] It is worth noting that in each iteration of the transmission line insulator defect data calculation, the channel pruning algorithm attempts to prune the least important filters, but this can be incorrect. Therefore, given a certain pruning criterion, the probability of misjudgment during the pruning iteration is assumed to be constant. This is used to determine which filters are less important and more important.

[0188] The target compression ratio refers to the compression index used to judge the target insulator defect detection model.

[0189] Target compression ratio r c The calculation formula is:

[0190]

[0191] The target acceleration ratio refers to the acceleration index used to judge the target insulator defect detection model.

[0192] Target acceleration ratio r a The calculation formula is:

[0193]

[0194] Where `origParm` represents the original parameters, `coMP23004952pressedParm` represents the compressed parameters, `origFLOPs` represents the original floating-point operations, and `acceleratedFLOPs` represents the accelerated floating-point operations. For a convolutional layer with parameters, it contains d 2 NC parameters and 2HW(d) 2(C+1)N floating-point operations. When the low-rank approximation truncates k singular values, the compression ratio and acceleration ratio of the new structure are calculated as follows:

[0195]

[0196]

[0197] Channel pruning is improved with the help of low-rank approximation, and redundancy at the weight level and filter level is eliminated in turn. In this way, low-rank approximation channel pruning can avoid processing the network in an aggressive manner. The steps are as follows: 1) Implement the low-rank approximation method to reduce redundancy within the filters. The approximation network consists of smaller filters, which improves the flexibility of subsequent channel pruning by reducing parameters. 2) Based on the previous step 3, smaller filters will help channel pruning generate a more efficient network and limit performance degradation. Therefore, global compression benefits are obtained by eliminating residual redundancy (especially in deeper layers).

[0198] Since the resolution of feature maps typically decreases as layers deepen, computation in shallow layers is more sensitive to parameter reduction. Based on the equation relating parameters and computational cost, it can be deduced that the floating-point operations of a convolutional layer are proportional to the resolution of the output feature map, as shown in the following formula:

[0199] Param = (d 2 C+1)N

[0200] FLOPs = 2HW(d 2 C+1)N

[0201] Where H, W, and N are the height, width, and number of channels of the output feature map, C is the number of channels of the input feature map, and d is the spatial dimension of the filter.

[0202] Existing work shows that channel pruning can prune more filters at deeper layers because deeper layers produce relatively low importance scores under a global pruning metric. While channel pruning cannot achieve effective speedup at shallow layers compared to deeper layers, it is possible to achieve satisfactory speedup benefits by limiting smaller ranks at shallow layers with the help of a low-rank approximation. In this way, the speedup and compression benefits of the low-rank approximation are maximized.

[0203] It's worth noting that in low-rank approximations, preserving the rank determines the benefits of acceleration and induced performance degradation. However, it's impossible to evaluate network performance using every possible rank. Therefore, rank selection requires an indicator. When the indicator approaches a certain value, preserving the rank promises to achieve a good trade-off between the acceleration and performance of the approximate network.

[0204] In practice, the spectral norm has been found to exhibit excellent robustness. For the same approximation matrix ~Wk, the spectral norm of the approximation error is calculated as follows: W∈R m×n (m>n) indicates that there are n singular values ​​σ1≥…≥σ1. n A weight matrix ≥ 0, when approximating a matrix When retaining the k largest singular values, the index α based on the spectral norm is... The normalized error between W and W, especially the approximate matrix spectral norm The spectral norm of the original matrix W is ||W|| spec =σ1, calculated as follows:

[0205]

[0206]

[0207] The index α based on the spectral norm has a non-maximum suppression effect on noise components. Once discarded components with large singular values ​​are considered unimportant to the network, other components with smaller singular values ​​can be reasonably ignored.

[0208] Notably, the CIFAR-10 image dataset was also selected for training and testing. CIFAR-10 is a commonly used image dataset in the field of deep learning. The CIFAR-10 dataset contains 60,000 32×32 color images across 10 classes, with 50,000 training images and 10,000 test images. The CIFAR-10 dataset is divided into 5 training blocks and 1 test block, each containing 10,000 images. The test block contains 1,000 images randomly selected from each class, and the training blocks contain these images in a random order. To demonstrate that the framework of this invention can effectively improve the performance of channel pruning and low-rank approximation, further showcasing our advantages in hardware deployment, experiments were conducted on the CIFAR-10, CIFAR-100, and ImageNet datasets using a VGG-like network. Furthermore, different compression algorithms were applied to various networks and datasets to validate the IDC (Integrated Data Center), and the efficiency of LAP was evaluated with the help of the IDC. Similarly, low-rank approximation and channel pruning are applied to defect detection in power grid transmission line insulator data samples to demonstrate the acceleration and compression performance of convolutional layers.

[0209] In practical applications, the compression efficiency of the target insulator defect detection model is determined by calculating the IDC.

[0210] First, a low-rank approximation method is implemented to reduce redundancy within the filters. Therefore, the approximation network consists of smaller filters, and significant speedup is achieved in this step, particularly in shallower layers, by reducing the number of parameters.

[0211] Secondly, channel pruning is performed on the approximate network in the first step. Based on previous analysis, smaller filters will help channel pruning generate a more efficient network and limit performance degradation. This step is used to obtain global compression benefits by eliminating residual redundancy (especially in deeper layers).

[0212] Assume there is a function between the performance degradation of a compressed network and the compression ratio, where degradation D is defined as the relative decrease in network performance in a specific metric, such as accuracy in classification tasks and mean squared error in regression tasks. For an uncompressed network, the compression ratio r is set to 1, and the corresponding degradation is zero, as shown below:

[0213] D(r=1)=0

[0214] As the compression ratio increases, performance degradation also increases, and the rate of degradation accelerates. Ideally, and until it approaches a certain threshold, D(r) can be considered a convex function and is smooth. Once compression is applied at a rate greater than a certain threshold, the network loses its ability. Smoothness ensures that the function D(r) can be fitted by a low-order polynomial.

[0215] Since the function is smooth and convex, its performance degradation compression ratio relationship r can be fitted by a mordanted polynomial curve:

[0216] D(r)=a m r m +…+a1r+a0

[0217] The area under the fitted curve D(r) represents the cumulative performance degradation. IDC is derived by dividing the integral area over a specified range.

[0218]

[0219] Where r l and r h These are the lower and upper bounds of the integration range, respectively. IDC represents the average performance degradation caused by the compression algorithm within a specified compression ratio range. Under the same comparison conditions, the smaller the IDC value, the better the compression algorithm.

[0220] In this invention, when a transmission line insulator image is received, image preprocessing is performed to generate a transmission line insulator sample set. Based on the SVD strategy, a low-rank approximation is performed on a preset convolutional neural network model to generate an initial insulator defect detection model. Taylor pruning is then performed on the initial insulator defect detection model according to a cross-layer strategy to generate a target insulator defect detection model. The target insulator defect detection model is used to detect defects in the transmission line insulator image to be identified, and the defect detection result is output. This invention solves the technical problem that existing target detection models may have each filter either completely deleted or retained during channel pruning. Therefore, pruning incorrect filters may discard a large number of important weights and cause irreparable damage to the performance of the target detection model. Moreover, channel pruning cannot utilize the sparsity of weight levels, which leads to low recognition efficiency of the target detection model. This invention achieves acceleration by using low-rank approximation to eliminate redundancy within the filter and converting the filter into a smaller, more compact filter. Channel pruning is applied globally on the approximation network, resulting in a better compressed and accelerated low-rank target insulator defect detection model.

[0221] Please see Figure 6 , Figure 6 This is a structural block diagram of a power transmission line insulator defect detection system provided in Embodiment 3 of the present invention.

[0222] This invention provides a defect detection system for power transmission line insulators, comprising:

[0223] The sample set module 301 is used to perform image preprocessing on the transmission line insulator image when a transmission line insulator image is received, and generate a transmission line insulator sample set.

[0224] The initial insulator defect detection model module 302 is used to generate an initial insulator defect detection model by performing a low-rank approximation on a preset convolutional neural network model based on the SVD strategy.

[0225] The target insulator defect detection model module 303 is used to perform Taylor pruning on the initial insulator defect detection model according to the cross-layer strategy to generate the target insulator defect detection model.

[0226] The defect detection module 304 is used to perform defect detection on the image of the transmission line insulator to be identified through the target insulator defect detection model and output the defect detection results.

[0227] Furthermore, the initial insulator defect detection model module 302 includes:

[0228] The initial convolutional layer submodule is used to decompose each convolutional layer in the preset convolutional neural network model using singular value decomposition to generate multiple initial convolutional layers.

[0229] The intermediate convolutional layer submodule is used to select a preset number of maximum singular values ​​to perform dimensionality reduction operations on each initial convolutional layer, generating multiple intermediate convolutional layers.

[0230] The target convolutional layer submodule is used to perform multiplication operations on each intermediate convolutional layer based on the retained singular components, generating multiple target convolutional layers.

[0231] The convolutional layer update submodule is used to use a pre-defined convolutional neural network model containing multiple target convolutional layers as the initial insulator defect detection model.

[0232] Furthermore, the target insulator defect detection model module 303 includes:

[0233] The selection submodule is used to select the target convolutional layer for Taylor pruning from the initial insulator defect detection model according to the cross-layer strategy.

[0234] The target feature map submodule is used to extract a preset number of transmission line insulator images from the transmission line insulator sample set and input them into the initial insulator defect detection model for training. Based on the training results, multiple target feature maps corresponding to the target convolutional layer are generated.

[0235] The trimming loss value submodule is used to calculate the trimming loss value between each target feature map and the preset standard feature map.

[0236] The initial loss value submodule is used to obtain the initial loss value associated with the preset convolutional neural network model.

[0237] The initial importance data submodule is used to perform difference calculation between the initial loss value and the trimmed loss value to generate the corresponding initial importance data of each filter in the initial insulator defect detection model.

[0238] The normalization submodule is used to normalize each initial importance data and generate the corresponding target importance score.

[0239] The sorting submodule is used to sort the target importance scores within each target convolutional layer.

[0240] The importance score comparison submodule is used to compare the importance score of each target with the associated preset standard importance score.

[0241] The filter processing submodule is used to retain the filters associated with the target importance score in the target convolutional layer if the target importance score is greater than the preset standard importance score, and to generate a target insulator defect detection model.

[0242] Furthermore, it also includes:

[0243] The test index module is used to train the initial insulator defect detection model by inputting a preset test sample set, and to generate corresponding test indexes based on the test results. The test indexes include the target pruning probability index, the target compression ratio, and the target acceleration ratio.

[0244] The indicator comparison module is used to compare test indicators with preset standard indicator conditions.

[0245] The first training module is used to stop training and generate a target insulator defect detection model if the test indicators meet the preset standard indicator conditions.

[0246] The second training module is used to adjust the network parameters of the initial insulator defect detection model according to the preset gradient if the test indicators do not meet the preset standard indicator conditions.

[0247] The jump module is used to jump to the execution of the steps based on the SVD strategy to perform low-rank approximation on the preset convolutional neural network model and generate the initial insulator defect detection model until the test index meets the preset standard index conditions and optimizes the target insulator defect detection model.

[0248] Furthermore, the defect detection module 304 includes:

[0249] The target identification diagram submodule is used to perform defect detection on the image of the transmission line insulator to be identified using the target insulator defect detection model, and generate a target identification diagram.

[0250] The target overlap submodule is used to calculate the target overlap between the target identifier diagram and the preset defective insulator diagram;

[0251] The overlap comparison submodule is used to compare the target overlap with a preset overlap threshold.

[0252] The first defect data output submodule is used to determine that the insulator associated with the image of the insulator of the transmission line to be identified has no defects if the target overlap is less than the preset overlap threshold.

[0253] The second defect data output submodule is used to take the defect type associated with the preset defect insulator diagram as the target defect if the target overlap is greater than or equal to the preset overlap threshold.

[0254] In this invention, when a transmission line insulator image is received, image preprocessing is performed to generate a transmission line insulator sample set. Based on the SVD strategy, a low-rank approximation is performed on a preset convolutional neural network model to generate an initial insulator defect detection model. Taylor pruning is then performed on the initial insulator defect detection model according to a cross-layer strategy to generate a target insulator defect detection model. The target insulator defect detection model is used to detect defects in the transmission line insulator image to be identified, and the defect detection result is output. This invention solves the technical problem that existing target detection models may have each filter either completely deleted or retained during channel pruning. Therefore, pruning incorrect filters may discard a large number of important weights and cause irreparable damage to the performance of the target detection model. Moreover, channel pruning cannot utilize the sparsity of weight levels, which leads to low recognition efficiency of the target detection model. This invention achieves acceleration by using low-rank approximation to eliminate redundancy within the filter and converting the filter into a smaller, more compact filter. Channel pruning is applied globally on the approximation network, resulting in a better compressed and accelerated low-rank target insulator defect detection model.

[0255] An electronic device according to an embodiment of the present invention includes: a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs a transmission line insulator defect detection method as described in any of the above embodiments.

[0256] The memory can be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. The memory has storage space for program code used to perform any of the method steps described above. For example, the storage space for program code may include individual program codes for implementing the various steps in the methods described above. This program code can be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, compact discs (CDs), memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above.

[0257] This invention provides a computer-readable storage medium storing a computer program thereon, which, when executed, implements a method for detecting defects in transmission line insulators as described in any embodiment of this invention.

[0258] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0259] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0260] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0261] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0262] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0263] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting defects in transmission line insulators, characterized in that, include: When a transmission line insulator image is received, image preprocessing is performed on the transmission line insulator image to generate a transmission line insulator sample set; Based on the SVD strategy, a low-rank approximation is performed on the preset convolutional neural network model to generate an initial insulator defect detection model. Taylor pruning is performed on the initial insulator defect detection model according to the cross-layer strategy to generate the target insulator defect detection model; The step of performing Taylor pruning on the initial insulator defect detection model according to the cross-layer strategy to generate the target insulator defect detection model includes: According to the cross-layer strategy, the target convolutional layer for Taylor pruning is selected from the initial insulator defect detection model; A preset number of transmission line insulator images are extracted from the transmission line insulator sample set and input into the initial insulator defect detection model for training. Multiple target feature maps corresponding to the target convolutional layer are generated based on the training results. Calculate the pruning loss value between each target feature map and the preset standard feature map; Obtain the initial loss value associated with the preset convolutional neural network model; The difference between the initial loss value and the trimming loss value is calculated to generate the initial importance data of each filter in the initial insulator defect detection model. The initial importance data are normalized to generate corresponding target importance scores; The importance scores of the targets within each target convolutional layer are sorted. Compare the importance scores of each target with the associated preset standard importance scores; If the target importance score is greater than the preset standard importance score, then the filter associated with the target importance score in the target convolutional layer is retained to generate a target insulator defect detection model; The target insulator defect detection model is used to detect defects in the image of the transmission line insulator to be identified, and the defect detection results are output.

2. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The step of generating an initial insulator defect detection model by performing a low-rank approximation on a preset convolutional neural network model based on the SVD strategy includes: Singular value decomposition is used to decompose each convolutional layer in the preset convolutional neural network model to generate multiple initial convolutional layers. A preset number of maximum singular values ​​are selected to perform dimensionality reduction on each of the initial convolutional layers, generating multiple intermediate convolutional layers; Based on the retained singular components, multiplication operations are performed on each of the intermediate convolutional layers to generate multiple target convolutional layers; The preset convolutional neural network model containing multiple target convolutional layers is used as the initial insulator defect detection model.

3. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, Also includes: The initial insulator defect detection model is trained by inputting a preset test sample set, and corresponding test indicators are generated based on the test results. The test indicators include target pruning probability index, target compression ratio, and target acceleration ratio. Compare the test indicators with the preset standard indicator conditions; If the test index meets the preset standard index conditions, then training is stopped and a target insulator defect detection model is generated. If the test index does not meet the preset standard index conditions, the network parameters of the initial insulator defect detection model are adjusted according to the preset gradient. The process jumps to the step of performing a low-rank approximation of the preset convolutional neural network model based on the SVD strategy to generate an initial insulator defect detection model, until the test index meets the preset standard index conditions, and then optimizes the target insulator defect detection model.

4. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The step of performing defect detection on the image of the transmission line insulator to be identified using the target insulator defect detection model and outputting the defect detection result includes: The target insulator defect detection model is used to detect defects in the image of the transmission line insulator to be identified, and a target identification block diagram is generated. Calculate the target overlap between the target identifier diagram and the preset defective insulator diagram; Compare the target overlap with a preset overlap threshold; If the target overlap is less than the preset overlap threshold, it is determined that the insulator associated with the image of the insulator to be identified has no defects. If the target overlap is greater than or equal to the preset overlap threshold, then the defect type associated with the preset defect insulator diagram is taken as the target defect.

5. A defect detection system for transmission line insulators, characterized in that, The transmission line insulator defect detection system is used to implement the transmission line insulator defect detection method as described in any one of claims 1-4, and the transmission line insulator defect detection system includes: The sample set module is used to perform image preprocessing on the received transmission line insulator image to generate a transmission line insulator sample set. The initial insulator defect detection model module is used to generate an initial insulator defect detection model by performing a low-rank approximation on a preset convolutional neural network model based on the SVD strategy. The target insulator defect detection model module is used to perform Taylor pruning on the initial insulator defect detection model according to the cross-layer strategy to generate the target insulator defect detection model. The defect detection module is used to perform defect detection on the image of the transmission line insulator to be identified using the target insulator defect detection model, and output the defect detection results.

6. The transmission line insulator defect detection system according to claim 5, characterized in that, The initial insulator defect detection model module includes: The initial convolutional layer submodule is used to decompose each convolutional layer in the preset convolutional neural network model using singular value decomposition to generate multiple initial convolutional layers. The intermediate convolutional layer submodule is used to select a preset number of maximum singular values ​​to perform dimensionality reduction operations on each of the initial convolutional layers, thereby generating multiple intermediate convolutional layers; The target convolutional layer submodule is used to perform multiplication operations on each of the intermediate convolutional layers based on the retained singular components to generate multiple target convolutional layers; The convolutional layer update submodule is used to use the preset convolutional neural network model containing multiple target convolutional layers as the initial insulator defect detection model.

7. The transmission line insulator defect detection system according to claim 5, characterized in that, Also includes: The test index module is used to train the initial insulator defect detection model by inputting a preset test sample set, and to generate corresponding test indexes based on the test results. The test indexes include target pruning probability index, target compression ratio, and target acceleration ratio. The indicator comparison module is used to compare the test indicators with preset standard indicator conditions. The first training module is used to stop training and generate a target insulator defect detection model if the test index meets the preset standard index conditions. The second training module is used to adjust the network parameters of the initial insulator defect detection model according to a preset gradient if the test index does not meet the preset standard index conditions. The jump module is used to jump to execute the step of performing a low-rank approximation of the preset convolutional neural network model based on the SVD strategy to generate an initial insulator defect detection model, until the test index meets the preset standard index condition, and optimize the target insulator defect detection model.

8. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the method for detecting defects in transmission line insulators as described in any one of claims 1-4.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the method for detecting defects in transmission line insulators as described in any one of claims 1-4.

Images