A method and device for identifying defects in power insulators based on deep decision fusion

Abstract

The present invention provides a method and device for identifying defects in power insulators based on deep decision fusion, the method comprising: extracting features from power inspection images, obtaining k The depth feature maps of different levels are k The depth feature maps of different levels are fused to obtain the fused depth feature map F ; Based on the fused deep feature map F , obtain the covariance descriptor feature matrix, principal component analysis feature matrix and maximum noise fraction method feature matrix; input , and into the corresponding support vector machine model respectively, and output the initial recognition result (, ,); use the decision fusion model of the voting mechanism to analyze the initial recognition result (, , ), and output the final insulator defect recognition result. The present invention can realize high-precision recognition of power insulator defects under the condition of a small number of training samples.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power intelligent inspection, and in particular to a method and device for identifying defects of electric insulators based on deep decision fusion. Background Art

[0002] Insulators are an important part of transmission lines. To ensure the stability and safety of the power system, it is essential to regularly check the status of insulators. Traditional manual inspection methods have disadvantages such as low efficiency, high labor intensity, and high safety risks. With the development of artificial intelligence technology, drone intelligent inspection has been widely used in the power industry. It can intelligently identify the status of insulators and detect equipment defects in a timely manner, thereby avoiding power accidents.

[0003] Traditional intelligent recognition methods require enough training samples to build a defect recognition model. However, it is very difficult to collect a large number of defect samples. How to train a high-quality insulator defect recognition model with a small number of defect samples becomes a key issue.

[0004] In summary, there is an urgent need for a power insulator defect identification method and device based on deep decision fusion to solve the problems existing in the prior art. Summary of the invention

[0005] The present invention aims to provide a method for identifying defects in power insulators based on deep decision fusion, aiming to make full use of deep features at different levels and multiple feature extraction methods, fully mine power insulator defect information from different image dimensions, and achieve high-precision identification of power insulator defects under the condition of a small number of training samples. The specific technical scheme is as follows:

[0006] A method for identifying defects in power insulators based on deep decision fusion includes the following steps:

[0007] Step S1: Extract features from the power inspection image to obtain k The depth feature maps of different levels are k The depth feature maps of different levels are fused to obtain the fused depth feature map F ;

[0008] Step S2: Based on the fused depth feature map F , get the covariance descriptor feature matrix , principal component analysis feature matrix and the maximum noise fraction characteristic matrix ;

[0009] Step S3: transform the covariance descriptor feature matrix , principal component analysis feature matrix and the maximum noise fraction characteristic matrix Input them into the corresponding support vector machine model respectively, and output the initial recognition result ( , , );

[0010] in, is the covariance descriptor feature matrix Input the initial recognition result outputted by the corresponding support vector machine model, Principal component analysis feature matrix Input the initial recognition result outputted by the corresponding support vector machine model, is the characteristic matrix of the maximum noise fraction method Input the initial recognition result outputted by the corresponding support vector machine model;

[0011] Step S4: Analyze the initial recognition results using the decision fusion model of the voting mechanism ( , , ), and output the final insulator defect identification result.

[0012] In the above technical solution, preferably, in step S1, the power inspection image A Input to Convolutional Neural Network After feature extraction, we get k Deep feature maps of different levels, where the feature extraction process is expressed as:

[0013] (1),

[0014] in, l is the number of convolutional layers of the convolutional neural network, Indicates l convolutional layers, Represents the first A deep feature map, and is an integer, k is an integer greater than or equal to 3 and k ≤ l .

[0015] The preferred technical solution is as follows: k The depth feature maps of different levels are fused to obtain the fused depth feature map F , specifically:

[0016] (2),

[0017] in, Representatives willk The deep feature maps are fused. F The spatial scale is , is the fused deep feature map F The spatial scale of is the fused deep feature map F The number of channels, Indicates A deep feature map to be fused.

[0018] In the above technical solution, preferably, in step S2, the covariance descriptor feature matrix is ​​obtained. Specifically:

[0019] For the deep feature map F Matrix processing is performed to obtain F 0, and then obtain the covariance descriptor feature matrix according to formula (3) :

[0020] (3),

[0021] in, F The scale size of 0 is , , , for The column vector, T represents the transposed matrix, is the fused deep feature map F The number of channels, is the fused deep feature map F The size of the space.

[0022] In the above technical solution, preferably, in step S2, the principal component analysis feature matrix is ​​obtained Specifically: According to formula (4), the fused depth feature map F Perform principal component analysis and dimensionality reduction:

[0023] (4),

[0024] in, represents the principal component analysis dimensionality reduction operation, is the number of channels after dimensionality reduction.

[0025] In the above technical solution, preferably, in step S2, the maximum noise fraction method characteristic matrix is ​​obtained Specifically: According to formula (5), the fused depth feature map F Perform dimensionality reduction using the maximum noise fraction method:

[0026] (5),

[0027] in, represents the maximum noise fractional dimensionality reduction operation, is the number of channels after dimensionality reduction.

[0028] In the above technical solution, preferably, in step S4: the initial recognition result , and The occurrence times of are counted respectively, and then the final insulator defect identification result is obtained according to formula (6):

[0029] (6),

[0030] in, The initial recognition result The number of occurrences of The initial recognition result The number of occurrences of The initial recognition result The number of occurrences of This is the final insulator defect identification result.

[0031] The present invention also provides a power insulator defect identification device based on deep decision fusion, the identification device includes a memory and a processor, and when the processor runs the computer instructions stored in the memory, it executes the power insulator defect identification method based on deep decision fusion.

[0032] The present invention also provides another power insulator defect identification device based on deep decision fusion, the identification device includes an image acquisition unit and a data calculation and processing unit, the image acquisition unit is used to collect power inspection images of insulators and transmit the collected power inspection images to the data calculation and processing unit, and the data calculation and processing unit performs defect identification on insulators according to the power insulator defect identification method based on deep decision fusion.

[0033] The application of the technical solution of the present invention has the following beneficial effects:

[0034] In the insulator defect recognition method of the present invention, various deep convolution feature maps from shallow to deep layers are first used to mine the depth information in the input image, and then covariance descriptor, principal component analysis and maximum noise score method are applied to construct three deep feature matrices, and features of different depth levels are jointly mined. By fully mining the power insulator defect information from different image dimensions, high-precision recognition of power insulator defects is achieved under the condition of a small number of training samples; therefore, it has great significance and practical value in power intelligent operation and maintenance.

[0035] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0037] Figure 1 It is a flow chart of the power insulator defect identification method based on deep decision fusion of the present invention;

[0038] Figure 2 It is a schematic diagram of the power insulator defect identification method based on deep decision fusion of the present invention;

[0039] Figure 3 (a) is a power inspection image input to the convolutional neural network;

[0040] Figure 3 (b) is the extracted deep feature map with a spatial scale of 56×56×512;

[0041] Figure 3 (c) is the extracted deep feature map with a spatial scale of 28×28×512;

[0042] Figure 3 (d) is the extracted deep feature map with a spatial scale of 14×14×512. DETAILED DESCRIPTION

[0043] In order to facilitate the understanding of the present invention, the present invention will be described more fully below, and preferred embodiments of the present invention are provided. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0045] Example:

[0046] See also Figure 1 and Figure 2 , this embodiment provides a method for identifying defects in power insulators based on deep decision fusion, comprising the following steps:

[0047] Step S1: Extract features from the power inspection image to obtain kThe depth feature maps of different levels are k The depth feature maps of different levels are fused to obtain the fused depth feature map F ;

[0048] Preferably, in this embodiment, a pre-trained convolutional neural network is used to mine deep feature maps at different levels. The convolutional neural network can be regarded as a combination of a series of functions. , the power inspection image A After inputting into the convolutional neural network for feature extraction, k The process of deep feature maps at different levels can be expressed as:

[0049] (1),

[0050] in, l is the number of convolutional layers of the convolutional neural network, Indicates l convolutional layers, Represents the first A deep feature map, By Convolutional layers To extract, and is an integer, k is an integer greater than or equal to 3 and k ≤ l .

[0051] Further, k The depth feature maps of different levels are fused to obtain the fused depth feature map F , specifically:

[0052] (2),

[0053] in, Representatives will k The deep feature maps are fused. F The spatial scale is , is the fused deep feature map F The spatial scale of is the fused deep feature map F The number of channels.

[0054] Specifically, the convolutional neural network model selected in this embodiment is the VGG-16 model, which is used as a multi-layer deep feature extractor. The last three convolutional layers of the model are selected to extract deep feature maps of different levels. The spatial scale distribution of the three deep feature maps obtained is , and , where Figure 3 (a) shows the power inspection image input to the neural network, Figure 3 (b) shows the extracted deep feature map with a spatial scale of 56×56×512, Figure 3 (c) shows the extracted deep feature map with a spatial scale of 28×28×512, and Figure 3 (d) shows the extracted deep feature map with a spatial scale of 14×14×512. The three selected convolutional layers are generally considered to be the key layers that can capture the rich and discriminative features in the input image, thereby realizing the use of various deep convolutional feature maps from shallow to deep layers to mine the depth information in the input image.

[0055] Furthermore, in this embodiment, when performing fusion, a bilinear interpolation method is used to take the depth feature map with the smallest spatial size as a reference, and downsample the depth feature map with a larger spatial size; this process adjusts the size of the depth feature map so that the depth feature maps of different levels can be effectively fused together, thereby improving the model's ability to understand the image and recognition efficiency, and generating a fused depth feature map F .

[0056] Step S2: Based on the fused depth feature map F , get the covariance descriptor feature matrix , principal component analysis feature matrix and the maximum noise fraction characteristic matrix ;

[0057] The depth feature map obtained after processing in step S1 F , which integrates various deep convolution features from shallow to deep layers to mine the depth information in the image. To further improve the representation ability of deep features, this embodiment uses the deep feature map F Further extract the covariance descriptor feature matrix , principal component analysis feature matrix and the maximum noise fraction characteristic matrix , as follows:

[0058] Get the covariance descriptor feature matrix Specifically:

[0059] For the deep feature map F Matrix processing is performed to obtain F 0, where F The scale size of 0 is , and then obtain the covariance descriptor feature matrix according to formula (3) :

[0060] (3),

[0061] in, , , for The column vector, T Represents the transposed matrix.

[0062] Each element in the covariance descriptor matrix represents the correlation between different feature channels, which can effectively fuse features of different dimensions; at the same time, the covariance descriptor contains the averaging operation, which can suppress the influence of noise in the image.

[0063] Get the principal component analysis feature matrix Specifically:

[0064] In order to reduce the computational complexity and suppress the noise in the power image, the fused deep feature map F Perform principal component analysis (PCA) dimensionality reduction:

[0065] (4),

[0066] in, represents the principal component analysis dimensionality reduction operation, is the number of channels after dimensionality reduction, The value range of is [1, 10]. The size is set to 3, that is, the first three principal components are retained, and the first three principal components can reflect The most critical image information in this embodiment is The spatial scale is .

[0067] Get the maximum noise fraction feature matrix Specifically:

[0068] Similarly, in order to reduce the computational complexity and suppress the noise in the power image, the fused deep feature map F Perform dimensionality reduction using the Maximum Noise Fraction (MNF) method:

[0069] (5),

[0070] in, represents the maximum noise fractional dimensionality reduction operation, is the number of channels after dimensionality reduction, The value range of is [1, 10]. Considering the efficiency of subsequent feature processing and model classification accuracy, in this embodiment The size of is set to 10, that is, 10 feature channels are retained, that is, in this embodiment The spatial scale is .

[0071] Step S3: transform the covariance descriptor feature matrix , principal component analysis feature matrix and the maximum noise fraction characteristic matrix Input them into the corresponding support vector machine model (SVM) respectively, and output the initial recognition result ( , , );

[0072] in, is the covariance descriptor feature matrix Input the initial recognition result outputted by the corresponding support vector machine model, Principal component analysis feature matrix Input the initial recognition result outputted by the corresponding support vector machine model, is the characteristic matrix of the maximum noise fraction method Input the initial recognition result outputted by the corresponding support vector machine model;

[0073] For more technical details about the support vector machine model, please refer to the prior art "Melgani F, Bruzzone L. Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on Geoscience and Remote Sensing, 2004, 42(8):1778–1790". The support vector machine model is not further described and explained in this embodiment.

[0074] According to step S1 and step S2, feature extraction is performed on the training image and the test image respectively, so as to obtain the training set and the test set. In this embodiment, the linear kernel function is used to train the SVM, and the output result is the category of the insulator defect. The experimental data set has two types of labels, including 3950 images of normal suspension insulators and 1350 images of damaged suspension insulators. 80% of the samples are randomly selected for model training, and 20% of the samples are used for model testing. Three feature matrices , , The corresponding SVM classifiers are used to output the respective insulator defect labels. , , (i.e. the initial recognition result). Among them, ={ }, ={ }, ={ }, To output the label of the normal insulator category (i.e., the recognition result), To output the label of the defective insulator category (i.e., the recognition result), different deep feature maps can mine the deep features of the image from different model angles, so the output results have highly complementary characteristics.

[0075] Step S4: Analyze the initial recognition results using the decision fusion model of the voting mechanism ( , , ), and output the final insulator defect identification result.

[0076] It is known to those skilled in the art that due to the complexity and variability of insulator targets, the recognition results based on a single classification model may not be able to fully mine the deep information of power images. There is strong complementary information in the initial recognition results of the three SVM classification models, which is due to the fact that the initial recognition results of the three SVM classification models are generated based on the same training data and test data; each SVM classification model has its own unique feature extraction capability to mine its own statistical features from power images.

[0077] Therefore, in order to effectively mine the complementary statistical information in the initial recognition results, this embodiment adopts the decision-level fusion technology (i.e., through majority voting) to fusion the three initial recognition results. , , Analyze the initial recognition results , , The number of occurrences is counted separately and defined as , their characteristic size is , is the number of test samples. When the votes of a recognition result belong to the majority vote, the corresponding confidence is the highest. Therefore, the recognition result corresponding to the highest confidence is selected as the defect recognition result. The final insulator defect recognition result is expressed as:

[0078] (6),

[0079] in, The initial recognition result The number of occurrences of The initial recognition result The number of occurrences of The initial recognition result The number of occurrences of This is the final insulator defect identification result.

[0080] This embodiment also provides a power insulator defect identification device based on deep decision fusion, the identification device includes a memory and a processor, and when the processor runs the computer instructions stored in the memory, it executes the power insulator defect identification method based on deep decision fusion.

[0081] In some embodiments, the identification device includes an image acquisition unit and a data calculation and processing unit. The image acquisition unit is used to acquire power inspection images of insulators and transmit the acquired power inspection images to the data calculation and processing unit. The data calculation and processing unit performs defect identification on insulators according to the power insulator defect identification method based on deep decision fusion.

[0082] In the insulator defect recognition method of the present embodiment, various deep convolution feature maps from shallow to deep layers are first used to mine the depth information in the input image, and then the covariance descriptor, principal component analysis and maximum noise score method are applied to construct three deep feature matrices, and the features of different depth levels are jointly mined. By fully mining the power insulator defect information from different image dimensions, high-precision recognition of power insulator defects is achieved under the condition of a small number of training samples; therefore, it has great significance and practical value in power intelligent operation and maintenance.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying defects in power insulators based on deep decision fusion, characterized in that: The following steps are involved: Step S1: Extract features from the power inspection image to obtain k The depth feature maps of different levels are k The depth feature maps of different levels are fused to obtain the fused depth feature map F ; Step S2: Based on the fused depth feature map F , get the covariance descriptor feature matrix , principal component analysis feature matrix and the maximum noise fraction characteristic matrix ; Step S3: transform the covariance descriptor feature matrix , principal component analysis feature matrix and the maximum noise fraction characteristic matrix Input them into the corresponding support vector machine model respectively, and output the initial recognition result ( , , ); in, is the covariance descriptor feature matrix Input the initial recognition result outputted by the corresponding support vector machine model, Principal component analysis feature matrix Input the initial recognition result outputted by the corresponding support vector machine model, is the characteristic matrix of the maximum noise fraction method Input the initial recognition result outputted by the corresponding support vector machine model; Step S4: Analyze the initial recognition results using the decision fusion model of the voting mechanism ( , , ), and output the final insulator defect identification result.

2. The power insulator defect identification method based on deep decision fusion according to claim 1 is characterized in that: In step S1, the power inspection image A Input to Convolutional Neural Network After feature extraction, we get k Deep feature maps of different levels, where the feature extraction process is expressed as: （1）， in, l is the number of convolutional layers of the convolutional neural network, Indicates l convolutional layers, Represents the first A deep feature map, and is an integer, k is an integer greater than or equal to 3 and k ≤ l .

3. The power insulator defect identification method based on deep decision fusion according to claim 2 is characterized in that: Will k The depth feature maps of different levels are fused to obtain the fused depth feature map F , specifically: （2）， in, Representatives will k The deep feature maps are fused. F The spatial scale is , is the fused deep feature map F The spatial scale of is the fused deep feature map F The number of channels, Indicates A deep feature map to be fused.

4. The power insulator defect identification method based on deep decision fusion according to claim 1 is characterized in that: In step S2, the covariance descriptor feature matrix is ​​obtained Specifically: For the deep feature map F Matrix processing is performed to obtain F 0, and then obtain the covariance descriptor feature matrix according to formula (3) : （3）， in, F The scale size of 0 is , , , for The column vector, T represents the transposed matrix, is the fused deep feature map F The number of channels, is the fused deep feature map F The size of the space.

5. The power insulator defect identification method based on deep decision fusion according to claim 1 is characterized in that: In step S2, the principal component analysis feature matrix is ​​obtained Specifically: According to formula (4), the fused depth feature map F Perform principal component analysis and dimensionality reduction: （4）， in, represents the principal component analysis dimensionality reduction operation, is the number of channels after dimensionality reduction.

6. The power insulator defect identification method based on deep decision fusion according to claim 1 is characterized in that: In step S2, the maximum noise fraction method feature matrix is ​​obtained Specifically: According to formula (5), the fused depth feature map F Perform dimensionality reduction using the maximum noise fraction method: （5）， in, represents the maximum noise fractional dimensionality reduction operation, is the number of channels after dimensionality reduction.

7. The power insulator defect identification method based on deep decision fusion according to claim 1 is characterized in that: In step S4: the initial recognition result , and The occurrence times of are counted respectively, and then the final insulator defect identification result is obtained according to formula (6): （6）， in, The initial recognition result The number of occurrences of The initial recognition result The number of occurrences of The initial recognition result The number of occurrences of This is the final insulator defect identification result.

8. A power insulator defect identification device based on deep decision fusion, characterized in that: The identification device includes a memory and a processor. When the processor runs the computer instructions stored in the memory, it executes the power insulator defect identification method based on deep decision fusion as described in any one of claims 1 to 7.

9. A power insulator defect identification device based on deep decision fusion, characterized in that: The recognition device includes an image acquisition unit and a data calculation and processing unit. The image acquisition unit is used to acquire power inspection images of insulators and transmit the acquired power inspection images to the data calculation and processing unit. The data calculation and processing unit performs defect identification on insulators according to the power insulator defect identification method based on deep decision fusion as described in any one of claims 1 to 7.

Images