CN108537780A - A kind of insulator breakdown detection method based on the full convolutional neural networks of second order - Google Patents

Abstract

A kind of insulator breakdown detection method based on the full convolutional neural networks of second order, is first marked to obtain a large amount of label images insulation subregion, is learnt to characteristics of image using single order FCN, and then is split to complex background insulation subregion；Then operation Optimized Segmentation is rebuild using mathematical morphology as a result, to obtain the accurate positionin of insulation subregion；Based on the segmentation result of region insulation, insulator breakdown region detection is carried out using second-order F CN networks, realizes the accurate positionin of fault zone, characteristics of image need not artificially be extracted, calculation amount is effectively reduced, and can effectively inhibit the interference of complex background, improves the accuracy rate of insulator breakdown identification.

Description

A Fault Detection Method for Insulators Based on Second-Order Fully Convolutional Neural Network

technical field

The invention belongs to the technical field of image segmentation and processing, in particular to an insulator fault detection method based on a second-order fully convolutional neural network.

Background technique

In order to ensure the safe and reliable operation of the entire transmission line, it is necessary to conduct timely and effective inspections of the transmission line insulators and find out the faults. With the development of smart grid technology, the application of UAV inspection technology in transmission line inspection has become more and more mature, and aerial insulator image recognition has become an important basis for judging the operation status of transmission lines. At present, there are relevant research results on the fault detection of aerial insulator images. The traditional methods are mainly divided into the following two categories, one is the threshold segmentation method based on contour, color, texture and shape; the other is the insulator based on supervised learning. segmentation. And some scholars have proposed a third type of insulator fault identification method based on deep learning.

In the first category, the threshold-based segmentation method is used to detect insulator faults by extracting the contour, color, and texture features of the insulator image. Li Bing-feng et al. proposed to binarize the original image, locate the position of the insulator, and extract the insulator outline in the binary image. However, this method can only detect the overall outline of the insulator, but cannot extract the detailed outline of the insulator, which is not conducive to later fault identification. Therefore, Lin Jucai et al. proposed a glass insulator defect diagnosis method based on color images, using the blue-green color features of glass insulators, transforming the color insulator image into the HSV visual metric space, and identifying insulator faults by searching for connected regions containing blue-green pixels. The method is sensitive to environmental changes, and is only applicable to glass insulators, and has poor versatility. Therefore, Yang Cuiru and others proposed to use the gray level co-occurrence matrix to extract the texture features of insulators to realize insulator recognition. Although this method avoids the influence of environmental factors on the characteristics of insulators and can realize the detection of insulator faults, it only uses a single texture feature to detect insulators. For fault identification, the identification rate of insulators with inconspicuous textures is low. In this regard, Jiang Yuntu et al. proposed an insulator fault identification based on multi-feature fusion, which can detect insulator faults by fusing various features such as contour, color, texture and shape, and improve the accuracy of fault identification. However, this method is sensitive to parameters.

For the second type of insulator segmentation method based on supervised learning, Shan Cheng et al. proposed the insulator defect detection combined with morphological features and BP neural network. Compared with unsupervised learning, it has a strong adaptive self-learning ability. Although this method learns the characteristics of insulators from training samples and realizes the classification and recognition of insulators, the convergence speed of this method is slow and it is easy to fall into local optimum. Therefore, Cheng Haiyan et al. used the invariant moments of the insulator images of the positive and negative samples to train the AdaBoost classifier. AdaBoost classifier is used to locate and identify insulators, and multiple weak classifiers are accumulated into a strong classifier to improve classification accuracy. Although supervised learning can effectively improve the accuracy of insulator fault detection, it still needs to manually extract the image features of insulators. When the image background is complex and changeable, it is difficult for feature descriptors to effectively extract image features, resulting in a low fault recognition rate.

For the third type of insulator fault identification method based on deep learning, Chen Qing et al. applied CNN to insulator fault identification. Compared with the above two traditional methods, CNN can automatically learn in layers, and the shallower convolution layer perceives The domain is small, which is used to learn local area features; the deep convolutional layer has a large perceptual domain, which is used to learn more abstract high-level semantic features. The deep convolutional network considers the local and overall information of the image. At the same time, the abstract features learned by the deeper convolutional layer are less sensitive to the size, position and direction of the object, which is helpful for insulator fault identification. Nevertheless, this method essentially classifies image pixels, uses the image blocks around the pixel as the input of CNN for training and prediction, and indirectly obtains segmentation results. Due to the redundancy in adjacent pixel blocks, the amount of image input data is too large and the processing speed is slow; at the same time, the size of the pixel neighborhood is difficult to determine, and the spatial position information in the image cannot be taken into account, so it cannot be well identified. It is difficult to accurately segment the contour. In this regard, Jonathan Long et al. proposed a segmentation method based on a fully convolutional neural network (FCN). Compared with the traditional CNN segmentation method, FCN is an end-to-end network that does not limit the size of the input image and can be used without full convolutional neural networks. In the case of connected layers, dense prediction can be performed, and segmentation maps of arbitrary sizes can be generated, which improves the processing speed. However, due to the expansion of the field of view of the pooling layer in the FCN network, the detailed information such as the target position is lost. For the insulator image with a complex background, it is difficult to accurately locate the insulator in the area when the fault detection is performed directly, causing part of the background to be falsely detected as the foreground, thus It is difficult to effectively improve the accuracy of fault identification of insulators.

All in all, traditional insulator fault detection methods rely on the underlying feature extraction and classifier selection of images, and it is difficult to achieve effective fault detection for insulator images with complex backgrounds. Classical deep learning (CNN, FCN) methods are less affected by complex background interference. Large, the improvement of the recognition rate is limited.

Contents of the invention

In order to overcome the shortcomings of the above-mentioned prior art, the object of the present invention is to provide a method for detecting insulator faults based on a second-order fully convolutional neural network, which does not require artificial extraction of image features, effectively reduces the amount of calculation, and can effectively suppress complex backgrounds. Interference, improve the accuracy of insulator fault identification.

In order to achieve the above object, the technical scheme that the present invention takes is:

A kind of insulator fault detection method based on second-order fully convolutional neural network, comprising the following steps:

Step 1: Input the aerial insulator image with complex background, and normalize the image to 400×600 size;

Step 2: Initialize the fully convolutional (FCN) network, where the convolution kernel size is 3×3, the learning rate is 10 ^-14 , and the number of iterations is 100,000;

Step 3: Input the training set, training set label, test set and test set label into the first-order FCN network for training and testing;

Step 4: Reconstruct and filter the preliminary segmented insulator image to obtain the regional insulator image;

Step 5: Multiply the reconstructed and filtered image with the original image to obtain the regional insulator image with the background removed;

Step 6: Use the regional insulator images and fault labels as the input of the second-order FCN network for training and testing;

Step 7: Output the insulator fault detection result.

The FCN network of described step 3 is trained and comprises the following process:

(a) FCN network training forward propagation process:

The forward propagation process of FCN is to calculate the actual output after the training samples are transmitted layer by layer. The specific calculation formula is as follows:

z ^l+1 ＝w ^l+1 α ^l +b ^l+1

α ^l+1 ＝f(z ^l+1 )

Among them, l is the number of network layers, z ^l+1 is the weighted input of the neurons of the l+1 layer, α is the input data of each layer of the corresponding image, w, b are the neurons of each layer of the fully convolutional neural network The weight and bias of , f is the linear correction function ReLU.

(b) FCN network training backpropagation process:

The backpropagation process of FCN is the backpropagation process of gradient or error, and the specific calculation formula is as follows:

Among them, J(w, b) is the objective function, m is the number of samples, h _{w, b} x ⁱ is the standard output, y ⁱ is the predicted value, and the optimal w and b are found by the stochastic gradient descent method to minimize the objective function.

The concrete implementation steps of described step 4 are as follows:

(a) initialization; B ₁ is a circular structural element with a size of (2i+1)×(2i+1), i=1, k=1; wherein the radius of the i circular structural element, k is the number of reconstructions;

(b) Define the mask image as f _mask , mark the image f _marker ; the reconstructed circular structural element The input image is f,

f _mask =f

f _marker = fΘB _i

h ₁ =f _marker

Among them, f _marker is the marked image reconstructed by the initial corrosion and recorded as h ₁ ;

(c) Corrosion reconstruction operation, the specific calculation formula is as follows:

Among them, h _k is the result after the kth corrosion reconstruction, as the marked image of the k+1th reconstruction, and h _k+1 is the result after the k+1th corrosion reconstruction;

(d) Judgment, if h _k+1 =h _k , then get the corrosion reconstruction result f ^ε =h _k ; otherwise, k=k+1, return to step (c);

(e) transform the mask image f _mask and the marker image f _marker ; i=1, k=1;

f′ _mask ＝(f ^ε ) ^c

f′ _marker ＝(f ^ε ) ^c ΘB _i

h′ ₁ = f′ _marker

Among them, (f ^ε ) ^c is the complement operation of f ^ε , and f′ _marker is the marked image reconstructed by the first expansion, denoted as h′ ₁ ;

(f) Dilation reconstruction operation, the specific calculation formula is as follows:

Among them, h′ _k is the result of the k-th expansion and reconstruction, as the marked image of the k+1-th reconstruction, and h′ _k+1 is the result of the k+1-th expansion and reconstruction;

(g) Judgment, if h′ _k+1 =h′ _k , get the open reconstruction result f _rec ; otherwise, k=k+ 1, return to step (f).

The beneficial effects of the present invention compared with prior art are:

In view of the problem that traditional insulator fault detection methods need to manually extract features and select classifiers, the second-order FCN fault detection model adopted in the present invention can automatically extract effective features of insulator images in layers, without manual extraction of image features, reducing the amount of calculation At the same time, the accuracy of fault identification is improved; for the classic CNN and FCN segmentation methods are easily disturbed by complex backgrounds, and some backgrounds are misdetected as defects in the insulator region, the present invention performs mathematical morphology on the insulator image segmented by the first-order FCN The reconstruction can effectively filter out the interference of the complex background and obtain the accurate positioning of the insulator area, thereby improving the accuracy of fault identification. The present invention can be applied to power network transmission line fault inspection, and lays a theoretical foundation for realizing the intelligence of the power network.

Description of drawings

Fig. 1(a) is an insulator test image 1 in the experiment of the present invention.

Fig. 1(b) is the result of fault segmentation on insulator image 1 using the contrastive method CNN.

Figure 1(c) is the superimposition of the fault segmentation result of the insulator image 1 under the CNN method and the original image.

Fig. 1(d) is the fault segmentation result of insulator image 1 using the contrastive method FCN.

Figure 1(e) is the superimposition of the fault segmentation results of the insulator image 1 under the FCN method and the original image.

Fig. 1(f) is the fault segmentation result of the insulator image 1 using the method of the present invention.

Fig. 1(g) is the fault segmentation result of the insulator image 1 under the method of the present invention and the original image superimposed.

Fig. 2(a) is an insulator test image 2 in the experiment of the present invention.

Fig. 2(b) is the result of fault segmentation on insulator image 2 using the contrastive method CNN.

Figure 2(c) is the superimposition of the fault segmentation results of the insulator image 2 under the CNN method and the original image.

Fig. 2(d) is the fault segmentation result of insulator image 2 using the contrastive method FCN.

Figure 2(e) is the superimposition of the fault segmentation results of the insulator image 2 under the FCN method and the original image.

Fig. 2(f) is the fault segmentation result of the insulator image 2 using the method of the present invention.

Fig. 2 (g) is the fault segmentation result of the insulator image 2 under the method of the present invention and the original image superimposed.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and embodiments.

Embodiment one:

In order to test the effectiveness and superiority of the present invention for color image segmentation, the experimental environment of the present invention is IW4206-2Q deep learning workstation, Ubuntu 16.0464-bit operating system, 62.8GB memory, NVIDIA GeForce GTX1080*2 graphics card, CPU E5-1602V4, and finally The construction of the second-order fcn network model is realized under the caffe deep learning framework.

Firstly, the first-order FCN network is used to segment the insulator area, and the segmented insulator image is mathematically reconstructed and filtered to obtain the regional insulator; on the basis of the regional insulator image, the second-order FCN is used for fault identification. The specific implementation steps are as follows:

Step 1: Input the aerial insulator image with complex background, and normalize the image to 400×600 size.

Step 2: Initialize the fully convolutional (FCN) network, where the convolution kernel size is 3×3, the learning rate is 10 ^-14 , and the number of iterations is 100,000 times.

Step 3: Input the training set, training set label, test set and test set label into the first-order FCN network for training and testing. The specific process is as follows:

(a) FCN network training forward propagation process:

The forward propagation process of FCN is to calculate the actual output after the training samples are transmitted layer by layer. The specific calculation formula is as follows:

z ^l+1 ＝w ^l+1 α ^l +b ^l+1

α ^l+1 ＝f(z ^l+1 )

Among them, l is the number of network layers, z ^l+1 is the weighted input of the neurons of the l+1 layer, α is the input data of each layer of the corresponding image, w, b are the neurons of each layer of the fully convolutional neural network The weight and bias of , f is the linear correction function ReLU.

(b) FCN network training backpropagation process:

The backpropagation process of FCN is the backpropagation process of gradient or error, and the specific calculation formula is as follows:

Among them, J(w, b) is the objective function, m is the number of samples, h _{w, b} x ⁱ is the standard output, y ⁱ is the predicted value, and the optimal w and b are found by the stochastic gradient descent method to minimize the objective function.

Step 4: Reconstruct and filter the preliminary segmented insulator image to obtain the regional insulator image. The specific steps are as follows:

(a) initialization; B ₁ is a circular structural element with a size of (2i+1)×(2i+1), i=1, k=1; wherein the radius of the i circular structural element, k is the number of reconstructions;

(b) Define the mask image as f _mask , and mark the image as f _marker ; it is the reconstructed circular structural element The input image is f,

f _mask =f

f _marker = fΘB _i

h ₁ =f _marker

Among them, f _marker is the marked image reconstructed by the initial corrosion and recorded as h ₁ ;

(c) Corrosion reconstruction operation, the specific calculation formula is as follows:

Among them, h _k is the result after the kth corrosion reconstruction, as the marked image of the k+1th reconstruction, and h _k+1 is the result after the k+1th corrosion reconstruction;

(d) Judgment, if h _k+1 =h _k , then get the corrosion reconstruction result f ^ε =h _k ; otherwise, k=k+1, return to step (c);

(e) transform the mask image f _mask and the marker image f _marker ; i=1, k=1;

f′ _mask ＝(f ^ε ) ^c

f′ _marker ＝(f ^ε ) ^c ΘB _i

h′ ₁ = f′ _marker

Among them, (f ^ε ) ^c is the complement operation of f ^ε , and f′ _marker is the marked image reconstructed by the first expansion, denoted as h′ ₁ ;

(f) Dilation reconstruction operation, the specific calculation formula is as follows:

Among them, h′ _k is the result of the k-th expansion and reconstruction, as the marked image of the k+1-th reconstruction, and h′ _k+1 is the result of the k+1-th expansion and reconstruction;

(g) Judgment, if h′ _k+1 =h′ _k , get the open reconstruction result frec; otherwise, k=k+ 1, return to step (f).

Step 5: Multiply the reconstructed and filtered image with the original image to obtain the regional insulator image with the background removed.

Step 6: Use the regional insulator images and fault labels as the input of the second-order FCN network for training and testing.

Step 7: Output the insulator fault detection result.

Referring to Fig. 1(a), the insulator test image 1 in the experiment of the present invention. The test image 1 is segmented by using three comparison methods: CNN, FCN and the method of the present invention. In the simulation experiment, in order to unify the parameters, both the CNN segmentation method and the FCN segmentation method use a 3×3 convolution kernel, the learning rate is 10 ^-14 , and the number of iterations is 100,000 times. The experimental results are shown in Figure 2(b) and 2(c). Because of the CNN segmentation method, the image blocks around the predicted pixels are used as the input of CNN for training and prediction, and the segmentation results are obtained indirectly; due to the redundancy in adjacent pixel blocks , the size of the pixel neighborhood is difficult to determine, the spatial position information cannot be considered, and the result of fault segmentation is rough, especially when the distance between two faults is relatively close, it is easy to be misdetected as one fault.

See Figure 2(d) and Figure 2(e). In contrast, the FCN segmentation method combines high-dimensional features and low-level feature information to improve segmentation accuracy, but due to the expansion of the pooling layer's sensory field of view, the target position, etc. Loss of detailed information, part of the background is falsely detected as a faulty part.

Referring to Figure 2(f), in contrast, the method of the present invention uses mathematical morphology reconstruction operations to filter out complex backgrounds in the first-order FCN segmentation results, thereby obtaining accurate positioning of insulator regions; based on the segmentation results of regional insulators, using two The first-order FCN network is used to detect the insulator fault area and realize the accurate positioning of the fault area. Compared with the classic CNN and FCN segmentation methods, it can effectively suppress the interference of complex backgrounds and improve the accuracy of insulator fault identification.

Claims (3)

1. a kind of insulator fault detection method based on second-order full convolutional neural network, is characterized in that, comprises the steps: Step 1: Input the aerial insulator image with complex background, and normalize the image to 400×600 size; Step 2: Initialize the fully convolutional (FCN) network, where the convolution kernel size is 3×3, the learning rate is 10 ^-14 , and the number of iterations is 100,000; Step 3: Input the training set, training set label, test set and test set label into the first-order FCN network for training and testing; Step 4: Reconstruct and filter the preliminary segmented insulator image to obtain the regional insulator image; Step 5: Multiply the reconstructed and filtered image with the original image to obtain the regional insulator image with the background removed; Step 6: Use the regional insulator images and fault labels as the input of the second-order FCN network for training and testing; Step 7: Output the insulator fault detection result. 2. a kind of insulator fault detection method based on second-order full convolutional neural network according to claim 1, is characterized in that, the FCN network of described step 3 is trained and comprises following process: (a) FCN network training forward propagation process: The forward propagation process of FCN is to calculate the actual output after the training samples are transmitted layer by layer. The specific calculation formula is as follows: z ^l+1 ＝w ^l+1 α ^l +b ^l+1 α ^l+1 ＝f(z ^l+1 ) Among them, l is the number of network layers, z ^l+1 is the weighted input of the neurons of the l+1 layer, α is the input data of each layer of the corresponding image, w, b are the neurons of each layer of the fully convolutional neural network The weight and bias of , f is the linear correction function ReLU. (b) FCN network training backpropagation process: The backpropagation process of FCN is the backpropagation process of gradient or error. The specific calculation formula is as follows: Among them, J(w, b) is the objective function, m is the number of samples, h _{w, b} x ⁱ is the standard output, y ⁱ is the predicted value, and the optimal w and b are found by the stochastic gradient descent method to minimize the objective function. 3. a kind of insulator fault detection method based on second-order full convolutional neural network according to claim 1, is characterized in that, the specific implementation steps of described step 4 are as follows: (a) initialization; B ₁ is a circular structural element whose size is (2i+1)×(2i+1), i=1, k=1; wherein the radius of the i circular structural element, k is the number of reconstructions; (b) Define the mask image as f _mask , mark the image f _marker ; the reconstructed circular structural element The input image is f, f _mask = f f _marker = fΘB _i h ₁ =f _marker Among them, f _marker is the marked image reconstructed by the initial corrosion and recorded as h ₁ ; (c) Corrosion reconstruction operation, the specific calculation formula is as follows: Among them, h _k is the result after the kth corrosion reconstruction, as the marked image of the k+1th reconstruction, and h _k+1 is the result after the k+1th corrosion reconstruction; (d) Judgment, if h _k+1 =h _k , then get the corrosion reconstruction result f ^ε =h _k ; otherwise, k=k+1, return to step (c); (e) transform the mask image f _mask and the marker image f _marker ; i=1, k=1; f′ _mask ＝(f ^ε ) ^c f′ _marker ＝(f ^ε ) ^c ΘB _i h′ ₁ = f′ _marker Among them, (f ^ε ) ^c is the complement operation of f ^ε , and f′ _marker is the marked image reconstructed by the first expansion, denoted as h′ ₁ ; (f) Dilation reconstruction operation, the specific calculation formula is as follows: Among them, h′ _k is the result of the k-th expansion and reconstruction, as the marked image of the k+1-th reconstruction, and h′ _k+1 is the result of the k+1-th expansion and reconstruction; (g) Judgment, if h′ _k+1 =h′ _k , get the open reconstruction result f _rec ; otherwise, k=k+1, return to step (f).