SAE-based high-voltage transmission line strand breakage defect identification and detection method

Abstract

The invention relates to the technical field of computer vision application, in particular to a high-voltage transmission line broken strand defect identification and detection method based on SAE, and the method comprises the steps: carrying out the unsupervised feature learning of an obtained transmission line image, obtaining the high-level local features, carrying out the feature selection of the high-level local visual features through a correlation analysis algorithm, and carrying out the recognition and detection of the broken strand defect of a high-voltage transmission line. The correlation between the features is reduced; then, a CNN model is utilized to carry out convolution on a to-be-recognized target object in a target field by using a local high-level feature vector learned by the SAE, and global features of the target object are obtained; and finally, the global feature response is sent to a classifier, and classification and identification of the target object are completed. Compared with the prior art, the identification accuracy is improved under the condition that the number of aerial images is limited.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision applications, and in particular to a method for identifying and detecting broken strand defects in high-voltage transmission lines based on SAE. Background Art

[0002] Electricity transportation mainly relies on high-voltage transmission lines. In recent years, the scale of power systems has gradually expanded, the types of line components have become more diverse, and the types of defects have become more complex. The working state of high-voltage lines plays a key role in whether the power system can operate safely and stably. Therefore, regular inspections of high-voltage transmission lines are necessary. Traditional manual patrols are inefficient and waste human resources. They have gradually been replaced by drone patrols. In order to improve the inspection efficiency of transmission lines, it is necessary to promote the intelligent development of inspection work.

[0003] At present, the target object recognition and classification technology based on vision has gradually become the focus of research at home and abroad. The key to target object recognition lies in the extraction of target object features. In the field of high-voltage transmission line defect recognition, statistical feature extraction methods are mainly used, including: HOG method, LBP method, GIST method, color histogram method based on time domain, and FFT, wavelet transform, Gabor transform, Hadamard transform and KL (PCA) transform based on frequency domain. These feature extraction methods mainly rely on prior knowledge to obtain the underlying visual features such as grayscale, color, texture, etc. in the target image. It is difficult for this type of method to obtain the essential feature information of the target image that is effectively expressed, and the recognition and classification accuracy of the target image is not high. Summary of the invention

[0004] The present invention aims to solve the problem that the recognition and classification accuracy of the image recognition method in the prior art is low when applied to high-voltage line detection.

[0005] In order to solve the above technical problems, the present invention provides a method for identifying and detecting broken strand defects in high-voltage transmission lines based on SAE, the method comprising the following steps:

[0006] S1, photographing an unlabeled data set of high-voltage transmission lines through drone inspection, and randomly collecting multiple image sub-blocks from the unlabeled data set, performing image gradient processing on the image sub-blocks using a preset gradient operator to enhance the edge characteristics of the image sub-blocks, and obtaining enhanced image sub-blocks;

[0007] S2, inputting the enhanced image sub-block into a preset SAE neural network to obtain local features of the image sub-block, wherein the preset SAE neural network is trained using a back propagation method;

[0008] S3, performing a convolution operation on the image of the unlabeled dataset according to the local features of the image sub-block to obtain a global feature of the image;

[0009] S4. Input the global features of the image into a preset softmax regression model to identify and classify broken bone defects of the high-voltage transmission line to obtain broken bone defect identification results.

[0010] Furthermore, the preset gradient operator is a Sobel gradient operator.

[0011] Furthermore, the step of training the preset SAE neural network using the back propagation method in step S2 specifically includes:

[0012] Add constraints to the response of the hidden layer unit of the automatic encoder of the preset SAE neural network to whiten the local features of the image sub-block, and the response vector a of the hidden layer unit of the i-th dimension of the preset SAE neural network (i) satisfy:

[0013] a (i) =σ(Wx (i) +b1)=σ(W SAE W white x (i) +b1);

[0014] The size of the i-th enhanced image sub-block is defined as n×n, and it is sorted according to the R, G, and B components to obtain a vector x(i) of m=n×n×3 dimensions, and the input vector x′(i) after gradient processing is W white x (i) , W white represents the m×m dimensional whitening transformation coefficient matrix, W SAE is the input weight connecting the hidden layer unit and each enhanced image sub-block after whitening, b1 represents the input bias, σ() is the activation function, W=W SAE W white x is the overall weight coefficient after whitening, representing the mapping relationship between the hidden layer and the original data;

[0015] Restored sample of the output of the preset SAE neural network after data reconstruction satisfy:

[0016]

[0017] in, is the i-th restored sample, is the output weight, b2 is the output bias;

[0018] The back propagation method is used to find the minimum cost function for the preset SAE neural network, and the minimum cost function satisfies:

[0019]

[0020] Where N is the number of unlabeled training samples, λ is the weight decay coefficient, β is the sparse penalty weight; s is the number of hidden layer units, ρ is the target sparse value, is the average response of all training samples on the jth hidden layer unit;

[0021] The average response calculated by the input weight in the enhanced image sub-block is the local feature a of the image sub-block corresponding to the enhanced image sub-block. T , which satisfies:

[0022] a T =σ(W s x T +b 1S );

[0023] Among them, W S represents the overall weight including whitening processing, b 1S represents the input bias, x T represents the enhanced image sub-block under the target domain.

[0024] Furthermore, step S3 is specifically as follows:

[0025] The K local features of the image sub-blocks obtained from the enhanced image sub-blocks of n×n dimensions are convolved pixel by pixel with a step size of 1 on the image of the unlabeled data set of size l×l to obtain K (l-n+1)×(l-n+1) dimensional global feature patterns, where K is the number of hidden layer units of the preset SAE neural network.

[0026] Furthermore, in step S3, the local features of each of the image sub-blocks are split according to three color channels, and are respectively convolved pixel by pixel with a step size of 1 with the image of the unlabeled data set, and the feature patterns obtained from the three color channel data are summed to obtain the global feature pattern.

[0027] Furthermore, step S3 further includes:

[0028] Average pooling is used to transform the (l-n+1)×(l-n+1)-dimensional global feature pattern into p×p-dimensional.

[0029] Furthermore, before step S4, the method further includes the following steps:

[0030] The preset softmax regression model is trained using the unlabeled dataset.

[0031] The beneficial effect achieved by the present invention is that a method for detecting broken strand defects in high-voltage transmission lines based on transfer learning convolutional sparse autoencoder (SAE) is proposed. The method obtains high-level local features by unsupervised feature learning of the obtained transmission line images, and selects high-level local visual features by using a correlation analysis algorithm to reduce the correlation between features; then, the local high-level feature vector learned by SAE is convolved with the target object to be identified in the target field by using a CNN model to obtain its global features; finally, the global feature response is sent to a classifier to complete the classification and identification of the target object. Compared with the related art, the recognition accuracy of the present invention is improved when the number of aerial images is limited. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic flow chart of the steps of a method for identifying and detecting broken strand defects in high-voltage transmission lines based on SAE provided in an embodiment of the present invention;

[0033] Figure 2 is a schematic diagram of a 3*3 image template provided by an embodiment of the present invention;

[0034] Figure 3 is a schematic diagram of processing of a Sobel operator provided by an embodiment of the present invention;

[0035] Figure 4 is a schematic diagram of processing an enhanced image sub-block provided by an embodiment of the present invention;

[0036] Figure 5 It is an overall structural block diagram of global feature extraction and classification provided by an embodiment of the present invention;

[0037] Figure 6 Schematic diagram of the training process of the preset softmax regression model provided by an embodiment of the present invention;

[0038] Figure 7 is a schematic diagram of a partial data set of a comparative example provided in an embodiment of the present invention;

[0039] Figure 8 It is a partial flow chart of a comparative example provided in an embodiment of the present invention;

[0040] Fig. 9 is a schematic diagram of local features learned from a strand-broken defect image by a comparison algorithm provided in an embodiment of the present invention;

[0041] Fig.10 It is a schematic diagram of local features learned from a broken strand defect image by the SAE-based high-voltage transmission line broken strand defect recognition and detection method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0043] Please refer to Figure 1 , Figure 1 1 is a schematic flow chart of a method for identifying and detecting broken strand defects in high-voltage transmission lines based on SAE provided in an embodiment of the present invention. The method comprises the following steps:

[0044] S1. An unlabeled data set of high-voltage transmission lines is photographed through drone inspection, and a plurality of image sub-blocks are randomly collected from the unlabeled data set. A preset gradient operator is used to perform image gradient processing on the image sub-blocks to enhance the edge characteristics of the image sub-blocks, thereby obtaining enhanced image sub-blocks.

[0045] The preset gradient operator is a Sobel gradient operator.

[0046] Specifically, the image is stored in the form of a matrix, and it is not possible to differentiate a straight line or curve in mathematical theory. The differentiation of an image is equivalent to the differentiation of a plane or a curved surface. For the operation of the image, a template is used to perform a convolution operation on the original image to achieve the desired effect. Therefore, obtaining the gradient of an image is to convolve the original image with the template. The embodiment of the present invention adopts the Sobel gradient operator. The Sobel operator is mainly used for edge detection. It is a discrete difference operator used to calculate the approximate grayscale of the image brightness function. Using this operator at any point in the image will generate a corresponding grayscale vector or its normal vector. The defect characteristics of broken strands in the transmission line meet the characteristics of this operator. After edge detection processing, the broken strand characteristics will be easier to identify.

[0047] The Sobel operator detects edges based on the phenomenon that the weighted difference in grayscale between the upper and lower and left and right neighboring points of a pixel reaches an extreme value at the edge. It has a smoothing effect on noise and provides more accurate edge direction information. The derivation process of the Sobel gradient operator: use a matrix to represent the operator; traverse the image, and use the 3*3 matrix centered on each pixel as the data to be processed. Multiply the operator matrix with the matrix to be processed, and sum the resulting matrix, which is used as the new value of the center point of the matrix. Figure 2 In the 3*3 template of the image shown, the horizontal and vertical gradients are defined. x :

[0048]

[0049] Vertical direction g y for:

[0050]

[0051] Using a weight of 2 on the center coefficient, the Sobel operator can better suppress (smooth) noise compared to other operators in the prior art (such as the Prewitt operator). A schematic diagram of the processing process of the Sobel operator is shown in Figure 3 shown.

[0052] The two sets of 3*3 matrices in the horizontal and vertical directions of the Sobel convolution operator are convolved with the image to obtain the approximate horizontal and vertical brightness differences. x and G y Respectively represent the grayscale values ​​of the image after horizontal and vertical edge detection, and the formulas are as follows:

[0053]

[0054]

[0055] The specific calculation is as follows:

[0056] G x =[f(x+1,y-1)+2f(x,y-1)+f(x+1,y-1)]-[f(x-1,y+1)+2f(x,y+1)+f(x+1,y+1)] (5);

[0058] G y =[f(x-1,y-1)+2f(x,y-1)+f(x+1,y-1)]-[f(x-1,y+1)+2f(x,y+1)+f(x+1,y+1)] (6);

[0060] Where f(a,b) represents the grayscale value of the image point (a,b). The horizontal and vertical grayscale values ​​of each pixel in the image are combined by the following formula to calculate the grayscale of the point:

[0061]

[0062] Often, an approximation without square root is used for efficiency:

[0063] G=G x +G y (8);

[0064] The specific effects after treatment are as follows: Figure 4 At the same time, the processed image data is used as the data set for subsequent processing.

[0065] S2. Input the enhanced image sub-block into a preset SAE neural network to obtain local features of the image sub-block, wherein the preset SAE neural network is trained using a back propagation method.

[0066] SAE (transfer learning convolutional sparse autoencoder) adds constraints to the response of the hidden layer units of the autoencoder, so that most neurons of the hidden layer units are in an "inhibited state" and only a small number of neurons are in an "excited state", thereby improving the algorithm's ability to learn data features. The embodiment of the present invention adopts a zero-phase component analysis (ZCA) method to whiten the source domain image sub-blocks learned by SAE, further strengthening the edge information of the image sub-blocks.

[0067] Define the size of the i-th enhanced image sub-block as n×n, sort it according to the R, G, B components, and obtain a vector x(i) of m=n×n×3 dimensions. The input vector x′(i) after gradient processing is W white x (i) , W white represents an m×m dimensional whitening transformation coefficient matrix. The step of using the back propagation method to train the preset SAE neural network in step S2 specifically includes:

[0068] Add constraints to the response of the hidden layer unit of the automatic encoder of the preset SAE neural network to whiten the local features of the image sub-block, and the response vector a of the hidden layer unit of the i-th dimension of the preset SAE neural network (i) satisfy:

[0069] a (i) =σ(Wx (i) +b1)=σ(W SAE W white x (i) +b1) (9);

[0070] Among them, W SAE is the input weight connecting the hidden layer unit and each enhanced image sub-block after whitening, b1 represents the input bias, σ() is the activation function, W=W SAE W white x is the overall weight coefficient after whitening, representing the mapping relationship between the hidden layer and the original data;

[0071] After whitening, the input data will exceed the range of [0, 1]. Therefore, the activation function σ() is no longer needed to map the output of SAE during data reconstruction.

[0072] Restored sample of the output of the preset SAE neural network after data reconstruction satisfy:

[0073]

[0074] in, is the i-th restored sample, is the output weight, b2 is the output bias;

[0075] In order to prevent overfitting and maintain the sparsity of the hidden layer response, it is necessary to add a weight decay term and a sparsity penalty term to the cost function. The embodiment of the present invention uses a back propagation method to find the minimum cost function for the preset SAE neural network, and the minimum cost function satisfies:

[0076]

[0077] Where N is the number of unlabeled training samples, λ is the weight decay coefficient, β is the sparse penalty weight; s is the number of hidden layer units, ρ is the target sparse value, is the average response of all training samples on the jth hidden layer unit;

[0078] The average response calculated by the input weight in the enhanced image sub-block is the local feature a of the image sub-block corresponding to the enhanced image sub-block. T , which satisfies:

[0079] a T =σ(W s x T +b 1S ) (12);

[0080] Among them, W S represents the overall weight including whitening processing, b 1S represents the input bias, x T represents the enhanced image sub-block under the target domain.

[0081] S3. Perform a convolution operation on the image of the unlabeled dataset according to the local features of the image sub-block to obtain global features of the image.

[0082] In order to obtain the global features of the broken strand defects in the UAV inspection of high-voltage transmission line images, firstly, the n×n-dimensional image sub-block local features learned from the source domain database using the SAE algorithm are convolved on the high-voltage transmission line image to obtain the feature response of the high-voltage transmission line image; then, these feature responses are combined into the global features of the image and used for subsequent classification training. The overall structural block diagram of global feature extraction and classification is shown in the figure. Figure 5 shown.

[0083] Step S3 is specifically as follows:

[0084] The K local features of the image sub-blocks obtained from the enhanced image sub-blocks of n×n dimensions are convolved pixel by pixel with a step size of 1 on the image of the unlabeled data set of size l×l to obtain K (l-n+1)×(l-n+1) dimensional global feature patterns, where K is the number of hidden layer units of the preset SAE neural network.

[0085] In step S3, each local feature of the image sub-block is split according to the three color channels, and each local feature is convolved with the image of the unlabeled data set with a step size of 1, and the feature patterns obtained from the three color channel data are summed to obtain the global feature pattern. This process can be expressed as:

[0086]

[0087] Among them, W and b represent the local feature coefficients learned by SAE, and x′ represents the value of each sub-region to be convolved in the training image sample. Represents the ZCA whitening coefficient in the whitening preprocessing stage.

[0088] Step S3 also includes:

[0089] Average pooling is used to transform the (l-n+1)×(l-n+1)-dimensional global feature pattern into p×p-dimensional.

[0090] The pooling layer in CNN aggregates features at different positions in the previous layer, for example, taking the average value (or maximum value) of a certain area in the feature pattern of the previous layer as the statistical representation of the area. This operation can reduce dimensional overfitting and make the aggregated features invariant to scale and rotation in space. Figure 5 As shown, the (l-n+1)×(l-n+1)-dimensional global feature pattern will become p×p-dimensional after the pooling operation, and obtain p×p×k-dimensional features. The most commonly used pooling method is to take the mean or maximum value of the pooled area as the statistics of the aggregation result. Therefore, under the condition of whitening preprocessing, the convolutional autoencoder is more suitable for the average pooling method. The embodiment of the present invention adopts this principle and selects the average pooling method.

[0091] S4. Input the global features of the image into a preset softmax regression model to identify and classify broken bone defects of the high-voltage transmission line to obtain broken bone defect identification results.

[0092] Before step S4, the method further includes the following steps:

[0093] The preset softmax regression model is trained using the unlabeled dataset.

[0094] Specifically, Figure 6As shown, the training process of the preset softmax regression model is as follows:

[0095] First, initialize the classifier model, set the weight w and bias b to random numbers, and randomly select 100 samples from the data set each time as a group and input them into the classifier model. Suppose the model estimates that the image category with broken strands is M, and the image category without broken strands is N. First, use the initialized random weights and biases to calculate the weighted input, and the activation function used in the hidden layer is the Relu function:

[0096] Relu(Z)=max(0,Z) (14);

[0097] Due to the linear non-saturation characteristics of the Relu activation function, the convergence speed and operation speed of the Relu activation function are improved, and the slope of Relu is a constant, which helps solve the convergence problem of deep networks. The output layer is classified by the Softmax function, so the activation value of the jth neuron on the Lth layer is for:

[0098]

[0099] Furthermore, the sum of the activation value output vectors of each neuron in the Lth layer is 1:

[0100]

[0101] According to formula (16), the final output layer is obtained. After Softmax classification, two types of vectors are output, representing the two categories of the data set. The output vector is compared with the target given true vector through the cross entropy cost function, which is defined as:

[0102]

[0103] Among them, y j represents the jth value in the actual label; m j Represents the value of the jth element in the Softmax output vector.

[0104] The weight derivative training update is:

[0105]

[0106] Bias derivative The training update is as follows:

[0107]

[0108] Finally, the loss function LOSS is calculated by taking the average value of the cross entropy cost function. If the loss function does not meet the accuracy requirements, the sample is returned to continue training; if the loss function meets the accuracy requirements, the current weight and bias values ​​are saved, and the classification parameter training is completed.

[0109] The present invention also provides a comparative example of the SAE-based high-voltage transmission line broken strand defect identification and detection method and the prior art, which is as follows:

[0110] After the original image set data is processed by Sobel operator image gradient, a new data set of 900 images is obtained, each image size is 150×150 pixels. The experiment divides the data set into two parts: training set and test set, with a ratio of 8:1, which is 800 and 100 images respectively. Figure 7 shown.

[0111] In order to compare the effect of the embodiment of the present invention, the non-transfer learning SAE classification algorithm is selected as the comparison algorithm. 900 images of high-voltage transmission line with broken strand defects are directly randomly sampled into 10,000 8×8 image blocks for local feature learning. The embodiment of the present invention first collects the same number of image blocks as the non-transfer learning for local feature learning based on the transfer learning SAE algorithm. The comparison process is as follows: Figure 8 shown.

[0112] Among the relevant parameters, the regularization constant of the whitening preprocessing stage is set to 0.1, the number of SAE hidden layer units s = 400 (i.e., the corresponding 400 self-learning features), the weight decay coefficient λ = 3e-3, the sparse penalty weight β = 5, and the target sparsity value = 0.035; when extracting global features, the pooling area size of the convolutional network is set to 19×19; when classifying defective images, the number of iterations of the softmax regression model is set to 10, 30, 50, 70, and 100, respectively, to test the stability of the algorithm in this paper. Finally, 100 images with and without broken strand defects are selected as the test set for classification experiments, and 5 cross-validations are performed under different iteration conditions. The classification performance of the transfer learning and non-transfer learning algorithms is evaluated with the accuracy index. Fig. 9 This is a schematic diagram of the local features learned by the comparison algorithm from the broken strand defect image. Fig.10 This is a schematic diagram of local features learned from a broken strand defect image by the method proposed in an embodiment of the present invention.

[0113] It can be seen that Fig.10 In the example, the edges of local features learned from the database are clearer; Fig. 9The learned local feature weights are relatively fuzzy, and the edge performance of each local feature vector is relatively weak. Since the edge performance of the local feature vector reflects the similarity between local features, the greater the similarity, the better the effect of local feature extraction using SAE. The experimental results of the Accuracy index of the method proposed in the embodiment of the present invention and the comparison algorithm under different iteration numbers are shown in Table 1.

[0114] Table 1 Accuracy index experimental results of the method proposed in the embodiment of the present invention and the comparison algorithm

[0115]

[0116] It can be seen from Table 1 that: under the same parameters, as the number of iterations increases, the overall performance of the non-transfer learning SAE classification algorithm (i.e., the comparison algorithm) and the SAE-based high-voltage transmission line broken strand defect recognition and detection method proposed in the embodiment of the present invention tends to be stable. However, from the Accuracy index of the classification result, it can be seen that the method of the embodiment of the present invention is better than the comparison algorithm. This is because the number of samples is only 900. After local feature extraction is performed using SAE, the similarity between each local feature vector is very small, and overfitting is prone to occur. Therefore, the global feature vector of the target sample image obtained by convolving the local feature vector with small similarity onto each target sample image cannot well represent all the features of the broken strand defect, resulting in poor recognition effect. The comparison example of the embodiment of the present invention illustrates that under small samples, the use of the SAE-based high-voltage transmission line broken strand defect recognition and detection method can achieve better recognition and classification effects than the non-transfer learning method.

[0117] The beneficial effect achieved by the present invention is that a method for detecting broken strand defects in high-voltage transmission lines based on transfer learning convolutional sparse autoencoder is proposed. The method obtains high-level local features by unsupervised feature learning of the obtained transmission line images, and selects high-level local visual features by using a correlation analysis algorithm to reduce the correlation between features; then, the local high-level feature vector learned by SAE is convolved with the target object to be identified in the target field by using a CNN model to obtain its global features; finally, the global feature response is sent to a classifier to complete the classification and identification of the target object. Compared with the related art, the recognition accuracy of the present invention is improved when the number of aerial images is limited.

[0118] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).

[0119] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0120] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0121] The embodiments of the present invention are described above in conjunction with the accompanying drawings. What is disclosed is only the preferred embodiment of the present invention. However, the present invention is not limited to the above-mentioned specific implementation manner. The above-mentioned specific implementation manner is only illustrative rather than restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms and equivalent changes without departing from the scope of protection of the purpose of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A method for identifying and detecting broken strand defects in high-voltage transmission lines based on SAE, characterized in that: The method comprises the following steps: S1, photographing an unlabeled data set of high-voltage transmission lines through drone inspection, and randomly collecting multiple image sub-blocks from the unlabeled data set, performing image gradient processing on the image sub-blocks using a preset gradient operator to enhance the edge characteristics of the image sub-blocks, and obtaining enhanced image sub-blocks; S2, inputting the enhanced image sub-block into a preset SAE neural network to obtain local features of the image sub-block, wherein the preset SAE neural network is trained using a back propagation method; S3, performing a convolution operation on the image of the unlabeled dataset according to the local features of the image sub-block to obtain a global feature of the image; S4. Input the global features of the image into a preset softmax regression model to identify and classify broken bone defects of the high-voltage transmission line to obtain broken bone defect identification results.

2. The SAE-based high-voltage transmission line broken strand defect identification and detection method according to claim 1 is characterized in that: The preset gradient operator is a Sobel gradient operator.

3. The SAE-based high-voltage transmission line broken strand defect identification and detection method according to claim 1 is characterized in that: The step of using the back propagation method to train the preset SAE neural network in step S2 specifically includes: Add constraints to the response of the hidden layer unit of the automatic encoder of the preset SAE neural network to whiten the local features of the image sub-block, and the response vector a of the hidden layer unit of the i-th dimension of the preset SAE neural network (i) satisfy: and (i) =σ(Wx (i) +b1)=σ(W SAE IN white x (i) +b1); Among them, W SAE is the input weight connecting the hidden layer unit and each enhanced image sub-block after whitening, b1 represents the input bias, σ() is the activation function, W=W SAE W white x is the overall weight coefficient after whitening, representing the mapping relationship between the hidden layer and the original data; Restored sample of the output of the preset SAE neural network after data reconstruction satisfy: in, is the i-th restored sample, is the output weight, b2 is the output bias; The back propagation method is used to find the minimum cost function for the preset SAE neural network, and the minimum cost function satisfies: Where N is the number of unlabeled training samples, λ is the weight decay coefficient, β is the sparse penalty weight; s is the number of hidden layer units, ρ is the target sparse value, is the average response of all training samples on the jth hidden layer unit; The average response calculated by the input weight in the enhanced image sub-block is the local feature a of the image sub-block corresponding to the enhanced image sub-block. T , which satisfies: a T =σ(W s x T +b 1S ); Among them, W S represents the overall weight including whitening processing, b 1S represents the input bias, x T represents the enhanced image sub-block under the target domain.

4. The SAE-based high-voltage transmission line broken strand defect identification and detection method according to claim 3 is characterized in that: Step S3 is specifically as follows: The K local features of the image sub-blocks obtained from the enhanced image sub-blocks of n×n dimensions are convolved pixel by pixel with a step size of 1 on the image of the unlabeled data set of size l×l to obtain K (l-n+1)×(l-n+1) dimensional global feature patterns, where K is the number of hidden layer units of the preset SAE neural network.

5. The SAE-based high-voltage transmission line broken strand defect identification and detection method according to claim 4 is characterized in that: In step S3, the local features of each image sub-block are split according to three color channels, and are respectively convolved pixel by pixel with a step size of 1 with the image of the unlabeled data set. The feature patterns obtained from the three color channel data are summed to obtain the global feature pattern.

6. The SAE-based high-voltage transmission line broken strand defect identification and detection method according to claim 4 is characterized in that: Step S3 also includes: Average pooling is used to transform the (l-n+1)×(l-n+1)-dimensional global feature pattern into p×p-dimensional.

7. The SAE-based high-voltage transmission line broken strand defect identification and detection method according to claim 4 is characterized in that: Before step S4, the method further includes the following steps: The preset softmax regression model is trained using the unlabeled dataset.

Images