CN110782431A

CN110782431A - High-voltage wire icing area detection method based on deep learning

Info

Publication number: CN110782431A
Application number: CN201910949919.8A
Authority: CN
Inventors: 张三元; 吴书楷; 祁忠琪; 涂凯
Original assignee: Zhejiang University ZJU
Current assignee: Zhejiang University ZJU
Priority date: 2019-10-08
Filing date: 2019-10-08
Publication date: 2020-02-11
Anticipated expiration: 2039-10-08
Also published as: CN110782431B

Abstract

The invention discloses a high-voltage wire icing area detection method based on deep learning. The method comprises the steps of preprocessing an original picture of the high-voltage wire shot by a camera, selecting a picture of a region to be detected, judging the class of the picture by a deep learning classification method, detecting key points of the boundary of the high-voltage wire icing region in the picture by the deep learning detection method if the picture shot by the camera is judged to be normal, and printing abnormal warning information on a screen if the picture shot by the camera is judged to be abnormal. The experimental result shows that the method can realize relatively accurate detection on the icing area of the high-voltage wire in the picture, is convenient for workers to subsequently evaluate the icing thickness, can effectively prevent the occurrence of the fault of the power facility, and has relatively high practical application value.

Description

High-voltage wire icing area detection method based on deep learning

Technical Field

The invention relates to a detection method for an icing area of a high-voltage wire, in particular to a detection method for the icing area of the high-voltage wire based on deep learning.

Background

The high-voltage wire as a power transmission facility plays an indispensable role in social production and life, and plays a vital role in the development of national economy. High-voltage wires widely exist in various natural environments, and have certain loopholes and difficulty in supervision. In addition, in cold weather in winter, the surface of the electric wire is often frozen, the transmission performance of the electric wire is affected, and even the electric wire is completely damaged. Therefore, the high-voltage wire is subjected to image monitoring and the icing area is automatically detected, so that the worker can conveniently evaluate the thickness of the icing area, and the stable operation of the electric power facility is ensured to a great extent. However, at present, no intelligent mode can monitor the external condition of the high-voltage wire, and workers often choose to check after power failure, so that a large amount of manpower, material resources and time are consumed.

The traditional area detection method depends on threshold segmentation, is limited by pixel values to a great extent, and is easily influenced by irresistible external factors such as natural environment and the like. In addition, the camera mounted on the high-voltage wire adopts an overlooking shooting angle, and the detection of the icing area by an object on the ground can cause great interference. Therefore, the conventional method has poor effect in practical application.

Disclosure of Invention

In order to solve the problems in the background art, the invention provides a high-voltage wire icing area detection method based on deep learning. In the method, the classification and the boundary key point detection of the target picture are calculated through a deep neural network, and the icing area in the selected range in the picture can be automatically and effectively positioned. The detection process can be divided into two steps, firstly, the normal class or the abnormal class of the preprocessed pictures is judged through a classification network, if the pictures are judged to be normal, the detection network is continuously used for detecting the icing area, and otherwise, warning information is printed on a screen in time.

The technical scheme adopted by the invention comprises the following steps:

step 1) acquiring a high-voltage wire picture as a sample picture by using a camera on a high-voltage wire pole, then preprocessing the sample picture, and constructing a classification data set and a wire boundary key point data set according to all the sample pictures;

step 2) establishing a deep classification network and training and testing;

step 3), establishing a deep detection network, and training and testing;

and 4) preprocessing the picture to be detected, inputting the preprocessed picture to the depth classification network in the step 2) for class judgment, outputting warning information if the classification network judges that the picture to be detected is abnormal, and inputting the picture to be detected to the depth detection network in the step 3) for obtaining the coordinate value of the key point of the boundary of the icing area of the electric wire if the depth classification network judges that the picture to be detected is normal, so that boundary position information is obtained, and a worker can conveniently evaluate the icing thickness.

The pretreatment in the step 1) and the step 4) comprises the following steps: intercepting a picture with one fifth height at the bottom of the picture, and zooming to a set size of 240 multiplied by 37;

the step 1) is specifically as follows:

dividing the preprocessed sample picture into a normal picture and an abnormal picture; the normal pictures are high-voltage wire pictures shot under the condition that the camera normally operates, wherein the high-voltage wire pictures comprise frozen high-voltage wire pictures and are given as normal picture labels 1; the abnormal pictures are high-voltage wire pictures shot under the condition that the camera is in an internal fault or the camera lens is frozen, and an abnormal picture label 0 is given; all classified sample pictures and labels thereof form a classified data set;

and marking coordinate values of four boundary key points of the upper left boundary, the upper right boundary, the lower right boundary and the lower left boundary of the electric wire icing area in the normal class pictures, wherein all the marked normal class pictures form an electric wire boundary key point data set.

The step 2) is specifically as follows:

2.1) constructing a depth classification network comprising an input layer, three convolution modules, a one-dimensional data conversion layer, two full-connection layers, a Dropout layer and an output layer; the input layer is connected to the one-dimensional data conversion layer through the first convolution module, the second convolution module and the third convolution module in sequence, the one-dimensional data conversion layer is used for converting input data into one-dimensional data, and output of the one-dimensional data conversion layer is processed through the first full connection layer, the Dropout layer and the second full connection layer in sequence to generate second classification output;

each convolution module comprises a convolution layer and a maximum pooling layer which are sequentially connected;

the first convolution module and the second convolution module both use 32 convolution kernels with the size of 3 x 3, the convolution layer of the third convolution module uses 64 convolution kernels with the size of 3 x 3, the step size of all the convolution layers is 1, the filling mode is VALID, and the result of each convolution is activated through a ReLU function. In addition, the maximum pooling layer uses a pooling window with the size of 2 × 2 and the step size of 1, the node retention rate of a Dropout layer is set to be 0.5, the number of nodes of the first full connection layer is 64, the ReLU function is used for activation, the number of nodes of the second full connection layer is 2, and the Softmax function is used for activation;

2.2) sending the classification data set constructed in the step 1) into a deep classification network, and training the deep classification network by adopting an Adam optimization algorithm until the error of the network reaches the minimum value; the loss function uses a cross entropy loss function, the initial value of the learning rate is set to be 0.001, the attenuation step length is set to be 300 steps, and the attenuation rate is set to be 0.9.

The step 3) is specifically as follows:

3.1) constructing a depth detection network comprising an input layer, two convolution layers, four convolution modules, two full-connection layers and an output layer;

the input layer is input into a first convolution module after passing through a first convolution layer, the output of the first convolution module is input into a fourth convolution module after sequentially passing through a second convolution module and a third convolution module, and the fourth convolution module is output from an output layer after sequentially passing through a first full-connection layer and a second full-connection layer;

the first convolution module and the fourth convolution module comprise three convolution layers, a connecting layer and a maximum pooling layer, wherein the second convolution layer and the third convolution layer are two parallel convolution layers, the output of the first convolution layer is respectively input into the two parallel convolution layers, the output results of the two parallel convolution layers are stacked along the channel direction after passing through the connecting layer to generate a new characteristic diagram, and finally the new characteristic diagram is output through the maximum pooling layer; the second convolution module removes the maximum pooling layer on the basis of the first convolution module, and the third convolution module removes the maximum pooling layer on the basis of the fourth convolution module;

the first convolution layer in the depth detection network uses 64 convolution kernels with the size of 3 multiplied by 3 and the step length of 2, the second convolution layer uses 1000 convolution kernels with the size of 1 multiplied by 1, the number of nodes of the first full-connection layer is 128, the activation function is ReLU, and the number of nodes of the second full-connection layer is 8;

the first convolution layer of the first convolution module uses 16 convolution kernels with the size of 1 × 1, the second convolution layer in the first convolution module uses 64 convolution kernels with the size of 1 × 1, the filling mode is VALID, the third convolution layer uses 64 convolution kernels with the size of 3 × 3, and the filling mode is SAME; the first convolution layer of the fourth convolution module uses 32 convolution kernels with the size of 1 multiplied by 1, and the number of the convolution kernels of the fourth convolution module is doubled relative to the number of the convolution kernels of the first convolution module; the maximum pooling layers of the first convolution module and the fourth convolution module use pooling windows with a size of 3 × 3 and a step size of 2;

all the convolution operations mentioned above default to 1 without the step size being described, the filling mode is VALID, and the result generated by each convolution layer is activated by the ReLU function.

3.2) sending the wire boundary key point data set constructed in the step 1) into a depth detection network, and training the depth detection network by adopting an Adam optimization algorithm until the error of the network reaches the minimum value; wherein the loss function uses a least square error loss function, the initial value of the learning rate is set to 0.001, the attenuation step is set to 500 steps, and the attenuation rate is set to 0.9.

The invention has the beneficial effects that:

1) the image classification and icing area detection functions are realized through a neural network algorithm, and compared with the traditional method, the method well learns the detail characteristics of the image and greatly improves the detection precision.

2) The method only needs 30-40ms for detecting one picture, has certain timeliness, can prevent the occurrence of power failure to a great extent, improves the safety coefficient, and reduces the consumption of manpower, material resources and time.

3) The method can quickly and effectively realize automatic detection of the icing area of the high-voltage electric wire, is convenient for workers to evaluate the icing thickness subsequently, can effectively prevent the occurrence of power facility faults, and has higher practical application value.

Drawings

FIG. 1 is a flow chart of the present invention.

FIG. 2 is a block diagram of the deep learning classification network of the present invention.

Fig. 3 is a block diagram of the deep learning detection network of the present invention.

FIG. 4 is a diagram of normal class pictures according to the present invention.

FIG. 5 is a schematic diagram of an exception class picture in accordance with the present invention.

FIG. 6 shows four different test images (a), (b), (c) and (d) according to the present invention.

FIG. 7 is a final detection result diagram of the present invention, wherein (a), (b), (c) and (d) are the detection result diagrams corresponding to the four different to-be-detected pictures in FIG. 6.

Detailed Description

The invention is further described with reference to the following figures and specific embodiments.

As shown in fig. 1, the implementation of the present invention is as follows:

the method comprises the following steps: sample pictures are collected through a high-definition camera installed on a high-voltage telegraph pole, then preprocessing operation is carried out on the pictures, a classification data set and a key point data set are established, and the specific process is as follows:

1.1) taking a picture with one fifth height at the bottom of the original picture, zooming the picture to a fixed size of 240 multiplied by 37, and dividing the picture into a normal category and an abnormal category. And labeling the pictures of each category, wherein the normal category label is 1, and the abnormal category label is 0. All pictures and their labels constitute training data for the classification network. Wherein, the normal class data and the abnormal class data are all according to 4: the scale of 1 is divided into a training set and a validation set.

1.2) because the shape of the high-voltage wire after icing is regular, only four boundary key points of the icing area of the wire are predicted during detection, so that pictures in a normal class are taken and marked by using marking software, and the marking process is that coordinates of the four key points at the upper left, the upper right, the lower right and the lower left of the boundary area of the wire are marked. All pictures and coordinate information constitute a set of key point data for detecting the network. The data in the key point data set are divided into a training set and a test set according to the ratio of 4: 1.

Step two: building a classification network, and sending the classification data set constructed in the step one into the classification network for training, wherein the specific process is as follows:

2.1) design the classification network as shown in FIG. 2. The classification network comprises three convolution modules, a one-dimensional data conversion layer, a first full connection layer, a Dropout layer and a second full connection layer. The data are firstly input into a first convolution module, processed and then sequentially sent into a second convolution module and a third convolution module, then converted into one-dimensional data through a one-dimensional data conversion layer, and the output of the one-dimensional data conversion layer is processed through a first full-connection layer, a Dropout layer and a second full-connection layer to generate second classification output;

the three convolution modules are basically the same in structure and comprise a convolution layer and a maximum pooling layer. The convolution layers of the first two convolution modules use 32 convolution kernels with the size of 3 × 3, the convolution layer of the third convolution module uses 64 convolution kernels with the size of 3 × 3, the default step length is 1, the filling modes are all VALID, and the result of each convolution is activated through a ReLU function. In addition, the maximum pooling layer uses a pooling window with the size of 2 × 2, the node retention rate of a Dropout layer is set to be 0.5, the number of nodes of the first full link layer is 64, the node retention rate is activated by using a ReLU function, the number of nodes of the second full link layer is 2, and the node retention rate is activated by using a Softmax function;

2.2) sending the classification data set constructed in the step one into a deep learning classification network, and training a neural network by adopting an Adam optimization algorithm until the error of the network reaches the minimum value. The loss function uses a cross entropy loss function, the initial value of the learning rate is set to be 0.001, the attenuation step length is set to be 300 steps, and the attenuation rate is set to be 0.9.

Step three: building a detection network, and sending the key point data set built in the step one into the detection network for training, wherein the specific process is as follows:

3.1) design the detection network as shown in FIG. 3. The depth detection network comprises a first convolution layer, four convolution modules, a second convolution layer and two full-connection layers. The data is firstly sent into a first convolution layer, then sent into four continuous convolution modules, then connected to two continuous full-connection layers after passing through a second convolution layer, and finally output is generated by the full-connection layers;

the first convolutional layer uses 64 convolutional kernels with the size of 3 x 3 and the step size of 2, and the first convolutional module comprises a convolutional layer using 16 convolutional kernels with the size of 1 x 1, and two parallel convolutional layers, a connecting layer and a maximum pooling layer. One of the parallel convolutional layers uses 64 convolutional kernels of 1 × 1 size and is filled with VALID, and the other uses 64 convolutional kernels of 3 × 3 size and is filled with SAME. The output results of the two convolutional layers are stacked along the channel direction to generate a new feature map, and the feature map is processed by using a pooling window with the size of 3 × 3 and the step size of 2. The second convolution module has the same structure as the first convolution module, but does not include the last maximum pooling layer; the number of convolution kernels of the third convolution module is doubled relative to that of convolution kernels of the second convolution module; the fourth convolution module adds a maximum pooling layer of 3 x 3 with a step size of 2 to the third convolution module. The second convolutional layer uses 1000 convolutional kernels of 1 × 1 size, the number of nodes of the first fully-connected layer is 128, the activation function is ReLU, and the number of nodes of the second fully-connected layer is 8. All the convolution operations are defaulted to 1 without explaining the step length, the filling mode is VALID, and the result generated by each convolution layer is activated through a ReLU function;

3.2) sending the key point data set constructed in the step 1) into a detection network, and training a neural network by adopting an Adam optimization algorithm until the error of the network reaches the minimum value. Wherein the loss function uses a least square error loss function, the initial value of the learning rate is set to 0.001, the attenuation step is set to 500 steps, and the attenuation rate is set to 0.9.

Step four: and (3) acquiring a picture to be detected of the high-voltage wire, selecting the picture of the area to be detected, preprocessing the picture, namely, taking a part with the height of one fifth of the bottom, zooming the part to the fixed size of 240 multiplied by 37, and sending the part to the classification network in the step two. If the classified network judges that the picture shot by the camera is abnormal, warning information is printed on a screen, otherwise, the picture is continuously sent to a detection network in the third step for calculation, and coordinate values of four key points of the icing area of the electric wire, namely the upper left corner, the upper right corner, the lower right corner and the lower left corner, are obtained (a coordinate system is established by taking the upper left corner of the picture as an origin), so that boundary position information is obtained, and workers can conveniently evaluate the icing thickness according to related values.

The specific embodiment is as follows:

the picture shown in fig. 4 is used as the normal sample picture, and the picture shown in fig. 5 is used as the abnormal sample picture. And FIG. 6 shows a picture to be detected shot by the camera, the picture to be detected is sent to the depth classification network for category judgment after being preprocessed, if the picture to be detected is judged to be normal, the detection network is continuously used for calculation to obtain key point coordinates of the wire icing area, and if the picture to be detected is abnormal, warning information is directly printed on a screen. The detection results of fig. 6 are shown in fig. 7, the upper two graphs (a) (b) are printed with abnormal warning information because of the influence of icing when the camera is used for shooting, and the lower two graphs (c) (d) realize more accurate detection, wherein the area below the straight line is a picture sent into the network for detection, four endpoints below the straight line, namely, the upper left endpoint, the upper right endpoint, the lower right endpoint and the lower left endpoint, are key points obtained by detection, and the area surrounded by the four endpoints is an icing area of the electric wire.

The method can realize a relatively accurate detection function of the high-voltage electric wire icing area, can facilitate workers to evaluate the icing thickness in time, has important significance for preventing power facility faults, improves the safety coefficient, saves the consumption of manpower, material resources and time to a great extent, and has certain practical application value.

The foregoing detailed description is intended to illustrate and not limit the invention, which is intended to be within the spirit and scope of the appended claims, and any changes and modifications that fall within the true spirit and scope of the invention are intended to be covered by the following claims.

Claims

1. A high-voltage wire icing area detection method based on deep learning is characterized by comprising the following steps:

1) acquiring a high-voltage wire picture as a sample picture by using a camera on a high-voltage wire pole, preprocessing the sample picture, and constructing a classification data set and a wire boundary key point data set according to all the sample pictures;

2) establishing a deep classification network, and training and testing;

3) establishing a deep detection network, and training and testing;

4) preprocessing the picture to be detected, inputting the preprocessed picture to the depth classification network in the step 2) for class judgment, outputting warning information if the classification network judges that the picture to be detected is abnormal, and inputting the picture to be detected to the depth detection network in the step 3) for obtaining the coordinate value of the boundary key point of the electric wire icing area if the depth classification network judges that the picture to be detected is normal.

2. The method for detecting the icing area of the high-voltage electric wire based on the deep learning as claimed in claim 1, wherein the preprocessing in the step 1) and the step 4) is as follows: and (5) intercepting the picture with one fifth of height at the bottom of the picture, and zooming to a set size.

3. The method for detecting the icing area of the high-voltage electric wire based on the deep learning as claimed in claim 1, wherein the step 1) is specifically as follows:

4. The method for detecting the icing area of the high-voltage electric wire based on the deep learning as claimed in claim 1, wherein the step 2) is specifically as follows:

5. The method for detecting the icing area of the high-voltage electric wire based on the deep learning as claimed in claim 1, wherein the step 3) is specifically as follows: