Composite insulator defect diagnosis method based on U-net segmentation

Abstract

The present invention relates to a composite insulator defect diagnosis method based on U-net segmentation, comprising the following steps: step S1 constructs a training sample set; step S2: based on the U-net segmentation algorithm, establishes a composite insulator infrared image segmentation model, and performs training based on the training sample set; step S3: inputs the image to be tested into the trained composite insulator infrared image segmentation model to obtain a binary image, and finds the corresponding composite insulator center axis according to the binary image; step S4: based on the composite insulator center axis, combined with the image temperature data matrix, obtains the composite insulator center axis temperature data; step S5: based on the heating characteristics of different composite insulator defect types, the composite insulator defect type is judged by the composite insulator center axis temperature curve. The present invention can quickly and effectively extract the overall heating characteristics of the composite insulator, and accurately diagnose and classify the defects.

Description

Technical Field

[0001] The present invention relates to the field of high-voltage composite insulator defect detection, and in particular to a composite insulator defect diagnosis method based on U-net segmentation. Background Art

[0002] Composite insulators play the role of mechanical fixation and electrical insulation in UHV transmission systems due to their light weight, high mechanical strength, and superior performance in anti-pollution flashover. However, in actual operation, composite insulators will gradually deteriorate due to the combined effects of electrical, mechanical, and external environments, resulting in varying degrees of heating. Infrared thermal imaging, as a portable, efficient, and highly accurate detection method, plays an important role in defect detection of composite insulators. However, the current means of defect detection is to import infrared images one by one into the analysis software of the thermal imager and manually draw lines to obtain the overall temperature distribution characteristics of the insulator. The detection efficiency is slow and is greatly affected by human subjective factors, making it difficult to meet the efficiency and accuracy requirements of infrared inspection of composite insulators. Summary of the invention

[0003] In view of this, the object of the present invention is to provide a composite insulator defect diagnosis method based on U-net segmentation, which can quickly and effectively extract the overall heating characteristics of the composite insulator and perform accurate diagnosis and classification of defects.

[0004] To achieve the above object, the present invention adopts the following technical solution:

[0005] A composite insulator defect diagnosis method based on U-net segmentation includes the following steps:

[0006] Step S1: Collect infrared images of composite insulators and perform preprocessing, parse the temperature data matrix corresponding to each infrared image, and construct a training sample set;

[0007] Step S2: Based on the U-net segmentation algorithm, a composite insulator infrared image segmentation model is established, and training is performed based on a training sample set;

[0008] Step S3: input the image to be tested into the trained composite insulator infrared image segmentation model to obtain a binary image, and find the corresponding central axis of the composite insulator based on the binary image;

[0009] Step S4: Based on the central axis of the composite insulator and in combination with the image temperature data matrix, obtain the temperature data of the central axis of the composite insulator;

[0010] Step S5: Based on the heating characteristics of different composite insulator defect types, the defect type of the composite insulator is determined by the temperature curve of the central axis of the composite insulator.

[0011] Furthermore, the temperature data matrix analysis in step S1 is as follows:

[0012] Step (1): Use the Thermal SDK development tool to extract the temperature information of the infrared image;

[0013] Step (2): Decode the extracted byte-type temperature matrix information and perform bit operations to obtain a Raw matrix;

[0014] Step (3): Create a blank matrix of size N×M and fill the Raw matrix row by row to obtain a temperature data matrix of the same size corresponding to the entire image.

[0015] Furthermore, the image preprocessing uses Labelme to perform polygonal outlining on the composite insulator contour in the infrared image, and names it Insulator, saves and generates a Json label file, which is the infrared image dataset, and stores a binary label map as a label for network model training.

[0016] Furthermore, the composite insulator infrared image segmentation model includes an encoder and a decoder;

[0017] In the encoder, there are 5 blocks in total, the number of channels in each block is the same, all convolutional layers use the same 3×3 convolution kernel, and the step size is 1; batch normalization is performed before the activation function, and the Relu function is used at the end of each convolutional layer for activation; the kernel size of the maximum pooling layer is 2×2, and the step size is 2;

[0018] In the decoder, symmetrical with the encoder, 3×3 convolution and upsampling operations are used to perform channel-by-channel feature fusion with the feature layer of the same dimension output in each block in the encoder. The resolution is gradually increased to the original input size. The last layer uses a 1*1 convolution kernel to make the number of channels consistent with the number of image categories, thereby achieving accurate segmentation of the target segmentation area of ​​the input image.

[0019] Furthermore, the model training is as follows: during the training process, the optimizer selects Adam Optimize, the initial learning rate is 0.0001, the learning rate is adjusted in equal intervals of step_size=1, the adjustment multiple gamma=0.92, Batchsize=4, the number of iterations (epoch) is K times, the loss function selects the binary cross entropy loss, the U-net algorithm is iterated once, and the predicted binary image is generated and compared with the label file to obtain the loss value. As the number of iterations increases, the network weights and bias parameters of each layer make the loss value continuously updated in the direction of decreasing; when the iteration conditions are met, the training ends and the optimal weight is saved.

[0020] Furthermore, the step S3 is specifically as follows:

[0021] According to the predicted binary image, use the connected region function of Opencv to find the center point of the insulator area;

[0022] Similarly, cut off a small area at both ends of the insulator, use the same method to find the two end points of the insulator, calculate the slope, and substitute the center point to find the center axis equation.

[0023] Furthermore, the step S4 is specifically as follows: based on the linear equation of the central axis, the value range is the two end points of the two composite insulators, the horizontal and vertical coordinate values ​​are rounded, the coordinate temperature value corresponding to the temperature data matrix is ​​read, the length of the central axis is the horizontal coordinate, and the temperature value of each pixel point on the central axis is the vertical coordinate, and the temperature data of the central axis of the composite insulator is obtained, that is, the overall temperature distribution characteristics of the composite insulator.

[0024] Furthermore, the defect types include core rod brittleness, surface contamination, and high-voltage measuring sheath aging and moisture.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The present invention automatically extracts the overall heating characteristics of the composite insulator, overcoming the disadvantages of low efficiency of manual line drawing detection and being greatly affected by human subjective factors.

[0027] 2. The present invention can effectively diagnose whether the composite insulator has defects such as core rod rot, surface contamination, and high-voltage test sheath aging and moisture, and classify the defects;

[0028] 3. The present invention improves the inspection efficiency of on-site operation and maintenance personnel, formulates subsequent operation and maintenance strategies for the heated composite insulators, and improves the continuous power supply capacity, stability, and safe operation level of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a method flow chart of an embodiment of the present invention.

[0030] Figure 2 This is the principle framework of the U-net algorithm model used in the embodiment of the present invention.

[0031] Figure 3 It is a segmentation prediction effect diagram in an embodiment of the present invention.

[0032] Figure 4 This is the process of obtaining the central axis of the composite insulator in the embodiment of the present invention.

[0033] Figure 5 It is the infrared spectrum and central axis temperature curve of the typical defect heating type in the embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0035] Please refer to Figure 1 The present invention provides a composite insulator defect diagnosis method based on U-net segmentation, comprising the following steps:

[0036] Step S1: using a drone equipped with an infrared thermal imager to inspect composite insulators on 110 kV and above transmission lines to obtain a large number of infrared images, parsing the temperature data matrix corresponding to each infrared image, and establishing a composite insulator infrared image sample library;

[0037] Preferably, in this embodiment, the temperature matrix parsing steps are as follows:

[0038] Step (1): Use the Thermal SDK development tool to extract the temperature information of the infrared image.

[0039] Step (2): Decode the extracted byte-type temperature matrix information and perform bit operations to obtain a Raw matrix.

[0040] Step (3): Create a blank matrix of 640×512 size and fill the Raw matrix row by row to obtain a temperature data matrix of the same size corresponding to the entire image.

[0041] Preferably, in this embodiment, the image preprocessing uses Labelme to perform polygonal outline on the composite insulator contour in the infrared image, and names it "Insulator", saves and generates a Json label file, which is the infrared image dataset, and the file stores the binary label map as the label for network model training.

[0042] Step S2: Based on the U-net segmentation algorithm, a composite insulator infrared image segmentation model is established, and a training sample set is imported for training;

[0043] Preferably, in this embodiment, the input image size is fixed to 512×512. In the encoder, there are 5 blocks, each with the same number of channels. All convolutional layers use the same 3×3 convolution kernel with a step size of 1. Batch Normalization is performed before the activation function to speed up the convergence process and improve the model's capacity. The Relu function is used at the end of each convolutional layer to avoid gradient disappearance. The kernel size of the max-pooling layer is 2×2 with a step size of 2, which reduces the size of the feature map and avoids feature redundancy. As the convolutional layers in the encoder continue to move downward, the size of the feature map becomes smaller and smaller, the number of channels increases, and the feature content output by the convolution kernel becomes more and more abstract.

[0044] In the decoder, which is basically symmetrical with the encoder, 3×3 convolution and upsampling operations are used to perform channel-by-channel feature fusion with the feature layer of the same dimension output in each block of the encoder. The resolution is gradually increased to the original input size. The last layer uses a 1*1 convolution kernel to make the number of channels consistent with the number of image categories, and finally achieve accurate segmentation of the target segmentation area of ​​the input image. Shallow features are used for segmentation, and deep features are used for positioning. This idea fully integrates shallow and deep features, so that the image can achieve a better segmentation effect.

[0045] During the training process, the optimizer selected Adam Optimize with a faster convergence speed, the initial learning rate was 0.0001, the learning rate was adjusted in equal intervals of step_size=1, the adjustment multiple gamma=0.92, Batch size=4, the number of iterations (epoch) was 200 times, and the loss function selected the binary cross entropy loss. The U-net algorithm was iterated once, and the predicted binary image was generated and compared with the label file to obtain the loss value. As the number of iterations increases, the network weights and bias parameters of each layer are continuously updated in the direction of decreasing the loss value. When the iteration conditions are met, the training ends and the optimal weight is saved.

[0046] Step S3: input the image to be tested into the trained defect detection model to obtain a binary image, and find the corresponding central axis of the composite insulator based on the binary image;

[0047] Preferably, the generated binary image is predicted, and the connected component function (connectedComponentsWithStats) of Opencv is used to find the center point of the insulator area. Similarly, a small area is intercepted at both ends of the insulator, and the two end points of the insulator are found by the same method, the slope is calculated, and the center point is brought into the equation of the center axis.

[0048] Step S4: Based on the central axis of the composite insulator and in combination with the image temperature data matrix, obtain the temperature data of the central axis of the composite insulator.

[0049] Preferably, in this embodiment, based on the equation of the central axis line, the value range is the two end points of the two composite insulators, the horizontal and vertical coordinate values ​​are rounded, the coordinate temperature values ​​corresponding to the temperature data matrix are read, the central axis length is the horizontal coordinate, and the temperature value of each pixel point on the central axis is the vertical coordinate, and the temperature data of the central axis of the composite insulator is obtained, that is, the overall temperature distribution characteristics of the composite insulator.

[0050] Step 5: Based on the heating characteristics of different composite insulator defect types, the defect type of the composite insulator is determined by the temperature curve of the central axis of the composite insulator.

[0051] Preferably, in this embodiment, the heating defects are classified into three categories: core rod rot, surface dirt, and high-voltage measuring sheath aging and moisture. The heating characteristics are:

[0052] Normal insulator: The central axis temperature curve fluctuates within the range of ±1K. Due to the irregular "U"-shaped distribution of the electric field of the composite insulator, the temperature curve may rise slightly at both ends.

[0053] Core rod brittleness: The heating point of this type of defect is mostly close to the high-voltage side of the composite insulator, spanning multiple sheds, and the maximum temperature rise value ranges from 5K to 47.6K.

[0054] Surface contamination: The heating positions of this type of defective composite insulators are scattered, and the range spans a small number of shed sheets, mostly between two adjacent groups of sheds, and the maximum temperature rise range is 2.6K~11.6K.

[0055] The high-voltage side sheath is aging and damp: The heating position of this type of defect is located in the high-voltage end sheath, ranging from the end hardware to 1~2 shed skirts at the high-voltage end. Most of the maximum temperature rise values ​​are below 5K.

[0056] With reference to the heating characteristics of the above three types of defects, the specific defect type of the composite insulator can be given according to the heating characteristics of the center axis temperature.

[0057] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made according to the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A composite insulator defect diagnosis method based on U-net segmentation, It is characterized in that The following steps are involved: Step S1: Collect infrared images of composite insulators and perform preprocessing, parse the temperature data matrix corresponding to each infrared image, and construct a training sample set; Step S2: Based on the U-net segmentation algorithm, a composite insulator infrared image segmentation model is established, and training is performed based on a training sample set; Step S3: input the image to be tested into the trained composite insulator infrared image segmentation model to obtain a binary image, and find the corresponding central axis of the composite insulator based on the binary image; Step S4: Based on the central axis of the composite insulator and in combination with the image temperature data matrix, obtain the temperature data of the central axis of the composite insulator; Step S5: based on the heating characteristics of different composite insulator defect types, the defect type of the composite insulator is determined by the temperature curve of the central axis of the composite insulator; The composite insulator infrared image segmentation model includes an encoder and a decoder; In the encoder, there are 5 blocks in total, the number of channels in each block is the same, all convolutional layers use the same 3×3 convolution kernel, and the step size is 1; batch normalization is performed before the activation function, and the Relu function is used at the end of each convolutional layer for activation; the kernel size of the maximum pooling layer is 2×2, and the step size is 2; In the decoder, symmetric to the encoder, 3×3 convolution and upsampling operations are used to perform channel-by-channel feature fusion with the feature layer of the same dimension output in each block of the encoder. The resolution is gradually increased to the original input size. The last layer uses a 1*1 convolution kernel to make the number of channels consistent with the number of image categories, so as to achieve accurate segmentation of the target segmentation area of ​​the input image. The training of the composite insulator infrared image segmentation model is as follows: During the training process, the optimizer selected Adam Optimize, the initial learning rate was 0.0001, the learning rate was adjusted in equal intervals with step_size=1, the adjustment multiple gamma=0.92, Batch size=4, the number of iterations (epoch) was K, and the loss function selected binary cross entropy loss. The U-net algorithm was iterated once to generate a predicted binary image and compare it with the label file to obtain the loss value. As the number of iterations increases, the network weights and bias parameters of each layer make the loss value continuously updated in the direction of decreasing; when the iteration conditions are met, the training ends and the optimal weights are saved.

2. The composite insulator defect diagnosis method based on U-net segmentation according to claim 1, It is characterized in that The temperature data matrix analysis in step S1 is as follows: Step (1): Use the Thermal SDK development tool to extract the temperature information of the infrared image; Step (2): Decode the extracted byte-type temperature matrix information and perform bit operations to obtain a Raw matrix; Step (3): Create a blank matrix of size N×M, and fill the Raw matrix row by row to obtain a temperature data matrix of the same size corresponding to the entire image.

3. The composite insulator defect diagnosis method based on U-net segmentation according to claim 1, It is characterized in that The image preprocessing uses Labelme to perform polygonal outlining on the composite insulator contour in the infrared image, and names it Insulator, saves and generates a Json label file, which is the infrared image dataset. The file stores a binary label map as a label for network model training.

4. The composite insulator defect diagnosis method based on U-net segmentation according to claim 1, It is characterized in that The step S3 is specifically as follows: According to the predicted binary image, use the connected region function of Opencv to find the center point of the insulator area; Similarly, cut off a small area at both ends of the insulator, use the same method to find the two end points of the insulator, calculate the slope, and substitute the center point to find the center axis equation.

5. The composite insulator defect diagnosis method based on U-net segmentation according to claim 4, It is characterized in that The step S4 is specifically as follows: according to the equation of the central axis straight line, the value range is the two end points of the two composite insulators, the horizontal and vertical coordinate values ​​are rounded, the coordinate temperature value corresponding to the temperature data matrix is ​​read, the central axis length is the horizontal coordinate, and the temperature value of each pixel point on the central axis is the vertical coordinate, and the temperature data of the central axis of the composite insulator is obtained, that is, the overall temperature distribution characteristics of the composite insulator.

6. The composite insulator defect diagnosis method based on U-net segmentation according to claim 1, It is characterized in that The defect types include core rod rot, surface dirt accumulation, and high-voltage measuring sheath aging and moisture.

Images