Overhead line fault visual diagnosis positioning method based on unmanned aerial vehicle technology

Abstract

The invention relates to the technical field of intelligent inspection of power systems, and particularly discloses an overhead line fault visual diagnosis positioning method based on an unmanned aerial vehicle technology, which comprises the following steps: carrying out dynamic image sequence acquisition on an overhead power line, and carrying out de-noising processing to obtain a second image group; separating a foreground electric power component area according to the second image group, and marking a potential key electric power component area; using the convolutional neural network model to classify and identify the electric power parts, determining the types of the classified electric power parts, and associating the same electric power parts in the continuous frames of images; calculating the adjacent frame deformation of the power component according to the adjacent frame deformation of the power component; and performing classification by adopting a support vector machine model to obtain a power fault type. According to the overhead line fault visual diagnosis and positioning method based on the unmanned aerial vehicle technology, the problems that accurate identification of key parts of a line and careful evaluation of fault features are difficult to achieve in a dynamic and complex environment in an existing mode are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for power systems, and more specifically, to a method for visual diagnosis and location of faults in overhead lines based on unmanned aerial vehicle (UAV) technology. Background Technology

[0002] In the vast field of power system operation and maintenance, the inspection and fault diagnosis of overhead lines occupy a crucial position. These lines, like the "blood vessels" of the power grid, bear the heavy responsibility of power transmission, and their safety and stability are directly related to the power supply security of thousands of households and the normal operation of the social economy. However, the diversity of line components and environmental interference make the identification process exceptionally difficult, and this difficulty further exacerbates the complexity of judging the severity of faults. Existing methods are insufficient to accurately identify key line components and conduct detailed assessments of fault characteristics in dynamic and complex environments. For example, during inspections, issues such as broken conductor strands may be misjudged due to blurred images or background interference, and it may even be impossible to accurately assess their impact on the overall safety of the line, thus affecting the safety and reliability of power grid operation. Summary of the Invention

[0003] In order to overcome the shortcomings of the existing technology, the present invention provides a visual diagnosis and localization method for overhead line faults based on UAV technology, which aims to solve the problems in the above-mentioned existing technology.

[0004] The technical solution adopted by this invention to solve its technical problem is: a visual diagnosis and location method for overhead line faults based on UAV technology, comprising the following steps: S1: Acquire dynamic image sequences of overhead power lines and generate the first image group; S2: Denoise the first image group to obtain the second image group; S3: Separate the foreground power component region from the second image group to obtain the third image group, and mark the potential key power component region in the third image group to obtain the power component contour feature set; S4: Use a convolutional neural network model to classify and identify the power components from the power component contour feature set, determine the type of the classified power components, and associate the same power components in consecutive frame images to obtain the power component tracking sequence; S5: Calculate the deformation amount of adjacent frames of each power component in the tracking sequence, generate a set of power component deformation parameters based on the deformation amount of adjacent frames of the power component, and classify the power fault type using a support vector machine model according to the power component type and the set of deformation parameters.

[0005] Preferably, in step S1, a dynamic image sequence containing multiple frames of images is obtained by using a drone to collect dynamic image sequences of overhead power lines. Position coordinate data are recorded synchronously when acquiring the dynamic image sequence; Extract consecutive multi-frame images from the dynamic image sequence as a first image group; match the corresponding overhead line station number according to the interval where the position coordinate data in the first image group is located, and obtain the first image group containing the overhead line station number.

[0006] Specifically, in step S2, a Gaussian filter is used to convolve the images in the first image group to obtain a denoised second image group.

[0007] It is worth noting that in step S2, for the denoised second image group, the pixel grayscale value of each image is calculated to determine the pixel grayscale distribution characteristics of the second image group. Based on the pixel grayscale distribution characteristics, the pixel values ​​of the denoised second image group are adjusted using a histogram equalization function to generate a second image group with enhanced contrast.

[0008] Optionally, in step S3, the Otsu method is used to determine the foreground and background thresholds, and OpenCV's threshold function is used to classify the pixels of the second image group based on the foreground and background thresholds to separate the foreground power component region and obtain the third image group. The edges of power components are extracted from the third image group using OpenCV's contour lookup function to obtain multiple potential key power component regions; For each of the potential critical power component regions, the boundary pixels are traversed, continuous segments are counted, the maximum number of pixels in a continuous segment and the total number of boundary pixels are obtained, and the ratio of the maximum number of pixels in a continuous segment to the total number of boundary pixels is calculated as the boundary integrity value. If the boundary integrity value is lower than the preset integrity threshold, then the potential critical power component region is closed using OpenCV's morphological operation function with a 3x3 structuring element to repair the broken boundary and obtain a complete boundary power component region; if the boundary integrity is higher than or equal to the preset integrity threshold, then the potential critical power component region is taken as a complete boundary power component region. Generate a set of power component contour features based on all the complete boundary power component regions.

[0009] Specifically, in step S4, the complete boundary power component region is extracted from the power component contour feature set as the first feature descriptor; The first feature descriptor is used to extract features through a pre-trained deep learning network to obtain a feature vector, and the feature vector is processed through a classification layer to determine the type of power component. For each type of power component, the center coordinates of the corresponding complete boundary power component region are extracted from consecutive frame images. An image processing algorithm is used to calculate the similarity based on the center coordinates to associate power components of the same type. Based on the associated power components, a tracking path is generated, resulting in a power component tracking sequence.

[0010] Preferably, in step S5, the center position coordinates of adjacent frames are obtained from the power component tracking sequence, the Euclidean distance between the center position coordinates is calculated, and the displacement parameters of adjacent frames of the power component are obtained. Based on the displacement parameters of adjacent frames of the power component, the boundary coordinates of the complete boundary power component region of the adjacent frame are obtained. The boundary coordinates of the complete boundary power component region of the next frame are subtracted from the displacement parameters of the adjacent frames of the power component to obtain the adjustment boundary coordinates of the next frame. The corresponding points in the boundary coordinates of the complete boundary power component region of the previous frame and the adjustment boundary coordinates of the next frame are calculated by Euclidean distance and then the average value is obtained to obtain the deformation amount of the power component between adjacent frames. Based on the deformation amount of adjacent frames of the power component, the deformation amount of adjacent frames of the power component and the code corresponding to the power component type are combined to form a deformation vector (XI,Y), where XI is the deformation amount of adjacent frames of the power component between the I-th frame and the previous frame, I is greater than 1, and Y is the code corresponding to the power component type. Based on all deformation vectors generated by all adjacent frame pairs in the power component tracking sequence, a set of power component deformation parameters is formed. For the set of deformation parameters of power components, the SVC class in the sklearn library is used to process the deformation parameter set, where the deformation vector is used as input and the classification label is used as output. The power fault type is determined by the output classification label.

[0011] Specifically, after step S5, step S6 is also included: obtaining the deformation amount of all adjacent frames from the power component tracking sequence, calculating the average value of the deformation amount, and obtaining the power fault severity score; Power maintenance priorities are determined based on the preset intervals in which the severity scores fall, where each preset interval corresponds to a power maintenance priority. Obtain the matching overhead line station number from the first image group corresponding to the power component tracking sequence; Power line inspection reports are generated based on power fault type, power maintenance priority, and overhead line station number.

[0012] The beneficial effects of this invention are as follows: In the above-mentioned aerial line fault visual diagnosis and localization method based on UAV technology, a complete solution is constructed by integrating image processing, deep learning, and data analysis technologies to address the difficulties in dynamic monitoring and complex fault identification in aerial power line inspections. This solution first ensures data quality through dynamic image sequence acquisition and denoising. Then, it separates the foreground power component region and extracts contour features, using a convolutional neural network model to achieve accurate classification and tracking of the power components. Based on this, it calculates the displacement and deformation of adjacent frames of the components and combines this with a support vector machine model to classify fault types, thereby achieving the purpose of aerial line fault localization and diagnosis. By deeply integrating dynamic image analysis with fault diagnosis, inspection efficiency and fault assessment accuracy are improved, providing strong technical support for the safe operation of power lines. Attached Figure Description

[0013] Figure 1 This is a flowchart of a visual diagnosis and localization method for overhead line faults based on UAV technology.

[0014] Figure 2 This is a flowchart of the steps in step S3.

[0015] Figure 3 This is a flowchart of the steps in step S4.

[0016] Figure 4 This is a flowchart of the steps in step S5. Detailed Implementation

[0017] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0018] Combination Figures 1 to 4 The method for visual diagnosis and localization of overhead line faults based on UAV technology, as shown, includes the following steps: S1: Acquire dynamic image sequences of overhead power lines and generate the first image group; S2: Denoise the first image group to obtain the second image group; S3: Separate the foreground power component region from the second image group to obtain the third image group, and mark the potential key power component region in the third image group to obtain the power component contour feature set; S4: Use a convolutional neural network model to classify and identify the power components from the power component contour feature set, determine the type of the classified power components, and associate the same power components in consecutive frame images to obtain the power component tracking sequence; S5: Calculate the deformation amount of adjacent frames of each power component in the tracking sequence, generate a set of power component deformation parameters based on the deformation amount of adjacent frames of the power component, and classify the power fault type using a support vector machine model according to the power component type and the set of deformation parameters.

[0019] In the aforementioned aerial power line fault visual diagnosis and localization method based on UAV technology, a complete solution is constructed to address the challenges of dynamic monitoring and complex fault identification in aerial power line inspections by integrating image processing, deep learning, and data analysis techniques. This solution first ensures data quality through dynamic image sequence acquisition and denoising. Then, it separates the foreground power component region and extracts contour features, using a convolutional neural network model to achieve accurate classification and tracking of the power components. Based on this, it calculates the displacement and deformation of adjacent frames of the components and combines this with a support vector machine model to classify fault types, thereby achieving the goal of aerial power line fault localization and diagnosis. By deeply integrating dynamic image analysis with fault diagnosis, inspection efficiency and fault assessment accuracy are improved, providing strong technical support for the safe operation of power lines.

[0020] It is worth noting that in step S1, a dynamic image sequence containing multiple frames of images is obtained by using a drone to collect dynamic image sequences of overhead power lines. Position coordinate data are recorded synchronously when acquiring the dynamic image sequence; Extract consecutive multi-frame images from the dynamic image sequence as a first image group; match the corresponding overhead line station number according to the interval where the position coordinate data in the first image group is located, and obtain the first image group containing the overhead line station number.

[0021] When collecting dynamic image sequences of overhead power lines using drones, multi-rotor drones equipped with high-resolution cameras can be used to inspect the high-voltage transmission lines along their flight path, capturing a video stream of 30 frames per second to obtain a dynamic sequence containing hundreds of images. Specifically, location coordinate data is recorded simultaneously with the acquisition of the dynamic image sequence. The drone's built-in GPS module can be used to associate each frame with its corresponding latitude and longitude coordinates. For example, during the inspection of a 500kV overhead line from city A to city B, the drone records coordinate data every 0.5 seconds during its flight, forming a coordinate dataset corresponding to the image sequence. This synchronization mechanism is based on timestamp matching, ensuring that each frame has corresponding spatial location information, thus providing geographical reference for subsequent analysis. In practical applications, this brings technical benefits such as facilitating fault location and reducing on-site troubleshooting time.

[0022] Extracting consecutive frames from the dynamic image sequence as the first image group involves preliminary screening in image processing. For example, assuming the dynamic image sequence is 10 minutes long and contains 18,000 frames, a frame extraction function in video editing software such as the OpenCV library can be used to select a consecutive 500 frames as the first image group. This extraction principle is based on the principle of continuity, ensuring that the images within the group reflect dynamic changes. For instance, consecutive frames between the tower piles of two adjacent overhead power lines can be extracted as a first image group, facilitating subsequent pile number location and matching.

[0023] Specifically, when matching the corresponding overhead line station number based on the location coordinate data interval in the first image group, the coordinate interval can be mapped to a pre-stored line database. For example, the latitude and longitude range (e.g., 120.1 to 120.2 east longitude) can be mapped to the station number T001 of the starting tower of this line, thus associating the first image group with station number T001. In actual operation, if the coordinates of the first image group fall within the coordinate interval corresponding to a certain station number, the system will automatically associate the group with the specific station number. This helps with subsequent defect identification and report generation, improving maintenance efficiency.

[0024] Preferably, in step S2, a Gaussian filter is used to convolve the images in the first image group to obtain a denoised second image group.

[0025] In this embodiment, the smoothing properties of the Gaussian function are utilized to suppress noise interference in the image. Specifically, a Gaussian filter is a linear filter based on a Gaussian distribution. Its kernel is typically a matrix with the standard deviation as a parameter. For example, when the standard deviation is set to 1.5, the kernel size is 5x5. By convolving this kernel with the image pixels, i.e., weighting each pixel and its neighborhood, random noise is blurred while edge details are preserved. In the business scenario of power line inspection, such as image sequences collected by drones, high-frequency noise, such as granular interference, is often introduced due to flight jitter or atmospheric interference. In this case, applying a Gaussian filter for convolution can yield clearer images. Arranging these images by timestamps yields a second image group, supporting the accuracy of subsequent defect identification. For example, when processing an image group containing tower insulators, the convolution process first traverses the pixel matrix of each frame, assigning the highest weight to the center pixel, and decreasing the weight of neighboring pixels according to a Gaussian curve. This reduces the noise variance of the processed image from the original 25 to about 5, ensuring that the conductor texture is not excessively blurred.

[0026] Optionally, in step S2, for the denoised second image group, the pixel grayscale value of each image is calculated to determine the pixel grayscale distribution characteristics of the second image group. Based on the pixel grayscale distribution characteristics, the pixel values ​​of the denoised second image group are adjusted using a histogram equalization function to generate a second image group with enhanced contrast.

[0027] Specifically, pixel grayscale value refers to the brightness level of each pixel in an image, typically ranging from 0 to 255. By traversing all frames of the second image group after denoising, the frequency of occurrence of each grayscale level is calculated to form a histogram distribution. For example, in an image of a power line, the conductor area may be concentrated in high grayscale values, while the background sky tends to be low grayscale. This distribution feature can reveal the problem of insufficient contrast.

[0028] In one embodiment, for a second image group of 500 frames, an image processing library is used to calculate the average grayscale value. For example, if the overall average is 128, but the distribution is concentrated in the 100-150 range, it indicates the image is too dark, which helps identify feature loss due to uneven lighting. Based on the pixel grayscale distribution characteristics, the pixel values ​​are adjusted using a histogram equalization function to generate a second image group with enhanced contrast. This is a non-linear transformation method whose principle is to remap grayscale values ​​to make the distribution more uniform, thereby stretching the contrast. Specifically, histogram equalization first calculates the cumulative distribution function, and then maps the original grayscale value to a new value. For example, the original grayscale value g is mapped to the new value g'=(L-1)*CDF*g, where L is the grayscale level and CDF is the cumulative probability, thus expanding the dynamic range. In power line operations, for example, for image groups inspected at night, after equalization, the contrast between the conductor and the background increases from the original 20:1 to 50:1, facilitating the detection of minute cracks.

[0029] Specifically, in step S3, the Otsu method is used to determine the foreground and background thresholds, and OpenCV's threshold function is used to classify the pixels of the second image group based on the foreground and background thresholds to separate the foreground power component region and obtain the third image group. The edges of power components are extracted from the third image group using OpenCV's contour lookup function to obtain multiple potential key power component regions; For each of the potential critical power component regions, the boundary pixels are traversed, continuous segments are counted, the maximum number of pixels in a continuous segment and the total number of boundary pixels are obtained, and the ratio of the maximum number of pixels in a continuous segment to the total number of boundary pixels is calculated as the boundary integrity value. If the boundary integrity value is lower than the preset integrity threshold, then the potential critical power component region is closed using OpenCV's morphological operation function with a 3x3 structuring element to repair the broken boundary and obtain a complete boundary power component region; if the boundary integrity is higher than or equal to the preset integrity threshold, then the potential critical power component region is taken as a complete boundary power component region. Generate a set of power component contour features based on all the complete boundary power component regions.

[0030] The Otsu method is an automatic thresholding algorithm that works by finding the optimal threshold to classify image pixels into foreground and background by maximizing the inter-class variance. Specifically, this method assumes a bimodal grayscale distribution of image pixels. By iterating through all possible thresholds and calculating the mean and weights of the foreground and background, it finds the threshold that maximizes the inter-class variance. For example, in power line inspection images, Otsu can effectively distinguish bright conductor components from the dark background sky. This method is suitable for images captured by drones with large variations in lighting conditions, ensuring the robustness of the threshold.

[0031] In this embodiment, cv2.threshold is used as the OpenCV threshold function for pixel classification. This function accepts the image and foreground / background thresholds and outputs a binarized image. Specifically, in a power line inspection scenario, the threshold function is applied to each frame of the second image group. Pixels above the foreground / background threshold are set to 255 (foreground), and those below are set to 0 (background), thereby separating the foreground power component area and generating the third image group. For example, when processing a tower image, the threshold function quickly traverses the pixel matrix, highlighting the conductor and insulator areas while filtering out background noise.

[0032] In this embodiment, cv2.findContours is used as the OpenCV contour finding function to extract the edges of power components from the third image group. This function detects connected regions in the binary image and returns a set of contour points. Specifically, cv2.findContours uses a boundary tracking algorithm to scan the image and identify external and internal contours. For example, in the separated image, the function can extract multiple closed curves to represent potential power components such as wires or insulators, thereby obtaining multiple potential key power component regions.

[0033] Specifically, the connectivity of adjacent pixels is checked along the contour point sequence of a potential critical power component region to identify breakpoints. The ratio of the number of pixels in the largest continuous segment to the total number of boundary pixels is calculated as the boundary integrity value. For example, if the total boundary of a potential critical power component region is 1000 pixels and the largest continuous segment is 800 pixels, the ratio is 0.8. If the preset threshold is 0.9, it is considered incomplete. If the boundary integrity value is lower than the preset threshold, this embodiment uses cv2.morphologyEx as the OpenCV morphological operation function to perform a closing operation with a 3x3 structuring element. This operation first dilates and then erodes, filling small holes and connecting breaks. Specifically, in a potential critical power component region, the closing operation can repair boundary gaps caused by noise. For example, when processing images of insulators affected by bird nest interference, the 3x3 kernel slides over the breakpoint to generate a continuous boundary, thus obtaining a complete boundary power component region. This processing improves the accuracy of defect detection in business operations and provides a reliable foundation for power system maintenance. If the boundary integrity value is higher than or equal to the threshold, the potential region is directly regarded as a complete boundary power component region without additional repair.

[0034] It is worth noting that in step S4, the complete boundary power component region is extracted from the power component contour feature set as the first feature descriptor; The first feature descriptor is used to extract features through a pre-trained deep learning network to obtain a feature vector. The feature vector is then processed by a classification layer to determine the type of power component. The types of power components include conductors, lightning protection wires, insulators, and towers, etc. For each type of power component, the center coordinates of the corresponding complete boundary power component region are extracted from consecutive frame images. An image processing algorithm is used to calculate the similarity based on the center coordinates to associate power components of the same type. Based on the associated power components, a tracking path is generated, resulting in a power component tracking sequence.

[0035] In this embodiment, ResNet-50 is used as a pre-trained deep learning network to extract features from the first feature descriptor. ResNet-50 is a residual network whose core function is to use residual blocks to solve the gradient vanishing problem in deep networks. By using skip connections, the input is directly added to the output, thereby extracting multi-level features. Specifically, the first feature descriptor is first input into the network's convolutional layer, and low-level features such as edges and textures are extracted layer by layer, up to high-level semantic features. Finally, a fixed-dimensional feature vector is output. For example, for the first feature descriptor corresponding to a complete boundary power component region, ResNet-50 will generate a 2048-dimensional vector through its 50-layer structure to capture the unique pattern of the power component. Then, the vector is processed by additional classification layers such as fully connected layers and a softmax function to calculate the probability distribution to determine the type. For example, the output probability of a conductor is 0.8, and that of an insulator is 0.15, thus classifying it as a conductor. This improves the accuracy of component identification in business operations and provides a basis for maintenance decisions.

[0036] In one embodiment, for each type of power component, the center coordinates of the corresponding complete boundary power component region are extracted from consecutive frame images. These center coordinates are obtained by calculating the average x-coordinate and average y-coordinate of all pixels within the region. For example, for a complete boundary power component region of a conductor type in a sequence of video frames, with 1000 pixels, an average x-coordinate of 500 and an average y-coordinate of 300, the center coordinates are (500, 300). Specifically, the ORB algorithm from the OpenCV library is used as the image processing algorithm to calculate similarity based on these center coordinates. ORB is a feature point detection and description method. Its principle is to detect key points in an image and generate binary descriptors, comparing similarity using Hamming distance. For example, in consecutive frames, ORB feature points are extracted near the center coordinates of two complete boundary power component regions of conductor types, and the number of descriptor matches is calculated. If the number exceeds a threshold such as 50, they are considered similar, thus associating power components of the same type. This helps in tracking situations where the shooting angle of components changes due to drone operation. Based on the associated power components, a tracking path is generated. Specifically, the associated center position coordinate sequence is connected into a path curve. For example, in a multi-frame image, the associated center position coordinates form a point chain from frame 1 to frame 10, which is fitted into a tracking path as the power component tracking sequence.

[0037] Preferably, in step S5, the center position coordinates of adjacent frames are obtained from the power component tracking sequence, the Euclidean distance between the center position coordinates is calculated, and the displacement parameters of adjacent frames of the power component are obtained. Based on the displacement parameters of adjacent frames of the power component, the boundary coordinates of the complete boundary power component region of the adjacent frame are obtained. The boundary coordinates of the complete boundary power component region of the next frame are subtracted from the displacement parameters of the adjacent frames of the power component to obtain the adjustment boundary coordinates of the next frame. The corresponding points in the boundary coordinates of the complete boundary power component region of the previous frame and the adjustment boundary coordinates of the next frame are calculated by Euclidean distance and then the average value is obtained to obtain the deformation amount of the power component between adjacent frames. Based on the deformation amount of adjacent frames of the power component, the deformation amount of adjacent frames of the power component and the code corresponding to the power component type are combined to form a deformation vector (XI,Y), where XI is the deformation amount of adjacent frames of the power component between the I-th frame and the previous frame, I is greater than 1, and Y is the code corresponding to the power component type. Based on all deformation vectors generated by all adjacent frame pairs in the power component tracking sequence, a set of power component deformation parameters is formed. For the deformation parameter set of power components, the SVC class in the sklearn library is used to process the deformation parameter set, where the deformation vector is used as input and the classification label is used as output. The power fault type is determined by the output classification label.

[0038] In this embodiment, the image sequence is acquired by the drone while it is in motion, therefore, the drone will undergo displacement during this process. The center position coordinates of adjacent frames are obtained from the power component tracking sequence. For example, for a set of insulator-type power component tracking sequences, the center coordinates of the first and second frames are first extracted. Assuming the coordinates of the first frame are (300, 200) and the second frame are (310, 205), the Euclidean distance between them is calculated. That is, the distance value of 11.2 is obtained by using the formula sqrt((310-300)^2+(205-200)^2). This represents the displacement that occurs when the drone acquires the image, reflecting the change in the position of the power component in the two adjacent frames.

[0039] Based on the displacement parameters of adjacent frames of the power component, the boundary coordinates of the complete boundary region of the power component in adjacent frames are obtained. For example, for an insulator component, the boundary coordinate set of the previous frame may include multiple points such as (290, 190), (310, 210), etc., and the corresponding coordinates of the next frame are (300, 195), (320, 215), etc. Then, the displacement parameters of the adjacent frames of the power component are subtracted from each boundary coordinate of the next frame. For example, if the displacement parameters of the adjacent frames of the power component are 11.2, the adjusted boundary coordinates become (288.8, 183.8), (308.8, 203.8), etc. The Euclidean distance is calculated to be 28.3, which reflects the boundary of the next frame after removing the influence of the positional change of the power component in the adjacent frames caused by the movement of the drone. After calculating the Euclidean distances of all corresponding points in all adjacent frames, the average of all obtained Euclidean distances is used to obtain the deformation of the power component between adjacent frames for this pair of adjacent frames, which represents the degree of change in the component shape.

[0040] In one embodiment, based on the deformation amount of adjacent frames of the power component, a deformation vector (XI, Y) is formed by combining the deformation amount of the power component with the encoding corresponding to the power component type. For example, if the deformation amount of the power component between the second frame and the first frame is 5.3 and the encoding corresponding to the insulator type is 1, then the vector is (5.3, 1). If the deformation amount of the power component between the third frame and the second frame is 4.9 and the encoding corresponding to the insulator type is 1, then the vector is (4.9, 1). The deformation vectors generated for all adjacent frame pairs in the sequence form a set of power component deformation parameters. This facilitates batch processing of multi-component data in the inspection system and ensures comprehensive capture of deformation trends.

[0041] Specifically, SVC, or Support Vector Classifier, works by finding the hyperplane with the largest margin to separate data points of different categories. During training, deformation vectors are input as features, and classification labels such as "normal conductor," "normal insulator," "cracked insulator," "broken conductor," or "tilted tower" are output. For example, a pre-trained model might be trained using the power component deformation parameter set {(3.1,1), (6.5,1), ...} corresponding to the label "broken insulator," and the power component deformation parameter set {(1.1,1), (0.5,1), ...} corresponding to the label "normal insulator," etc. The labels used during training are expert opinions derived from historical power component deformation parameter sets. Then, new power component deformation parameter sets are predicted, and classification labels are output. These labels are used to determine the power fault type, such as "cracked insulator." This improves the automation level of fault diagnosis, reduces manual intervention, and increases response speed.

[0042] Optionally, after step S5, step S6 is also included: obtaining the deformation amount of all adjacent frames from the power component tracking sequence, calculating the average value of the deformation amount, and obtaining the power fault severity score; Power maintenance priorities are determined based on the preset intervals in which the severity scores fall, where each preset interval corresponds to a power maintenance priority. Obtain the matching overhead line station number from the first image group corresponding to the power component tracking sequence; Power line inspection reports are generated based on power fault type, power maintenance priority, and overhead line station number.

[0043] Specifically, calculating the average deformation of these adjacent frames involves summing all values ​​and dividing by the number. For example, if the deformation values ​​of adjacent frames are 3.2, 4.5, 2.8, 5.1, 3.9, 4.2, 2.7, 4.8, and 3.6, the average is approximately 3.9. This average is used as the power fault severity score. In power inspection operations, it quantifies the overall deformation trend. A higher score may indicate severe component damage, thus providing a data basis for subsequent decision-making. When determining the power maintenance priority based on the preset interval of the severity score, multiple intervals can be set, such as 0 to 2 for low priority, 2 to 5 for medium priority, and 5 and above for high priority. In actual operations, this division is based on historical data statistics to ensure reasonable resource allocation and prevent small problems from escalating into major faults.

[0044] In one embodiment, obtaining the matching overhead line station number from the first image group corresponding to the power component tracking sequence involves an information backtracking process. During the formation of the power component tracking sequence, a first image group is generated, and this first image group has a matching overhead line station number. This overhead line station number is then extracted and matched with the power component tracking sequence. Finally, the three key pieces of information—power fault type, power maintenance priority, and overhead line station number—are entered into a preset table to generate a power line inspection report.

[0045] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for visual diagnosis and localization of faults in overhead power lines based on unmanned aerial vehicle (UAV) technology, characterized in that, Includes the following steps: S1: Acquire dynamic image sequences of overhead power lines and generate the first image group; S2: Denoise the first image group to obtain the second image group; S3: Separate the foreground power component region from the second image group to obtain the third image group, and mark the potential key power component region in the third image group to obtain the power component contour feature set; S4: Use a convolutional neural network model to classify and identify power components from the power component contour feature set, determine the type of power component after classification, and associate the same power components in consecutive frame images to obtain a power component tracking sequence; S5: Calculate the deformation amount of adjacent frames of each power component in the tracking sequence, generate a set of power component deformation parameters based on the deformation amount of adjacent frames of the power component, and classify the power fault type using a support vector machine model based on the power component type and the set of deformation parameters.

2. The method for visual diagnosis and localization of overhead line faults based on UAV technology according to claim 1, characterized in that: In step S1, a dynamic image sequence containing multiple frames of images is obtained by using a drone to collect dynamic image sequences of overhead power lines. Position coordinate data are recorded synchronously when acquiring the dynamic image sequence; Extract consecutive multi-frame images from the dynamic image sequence as a first image group; match the corresponding overhead line station number according to the interval where the position coordinate data in the first image group is located, and obtain the first image group containing the overhead line station number.

3. The method for visual diagnosis and localization of overhead line faults based on UAV technology according to claim 1, characterized in that: In step S2, a Gaussian filter is used to convolve the images in the first image group to obtain a denoised second image group.

4. The method for visual diagnosis and localization of overhead line faults based on UAV technology according to claim 3, characterized in that: In step S2, for the denoised second image group, the pixel grayscale value of each image is calculated to determine the pixel grayscale distribution characteristics of the second image group. Based on the pixel grayscale distribution characteristics, the pixel values ​​of the denoised second image group are adjusted using a histogram equalization function to generate a second image group with enhanced contrast.

5. The method for visual diagnosis and localization of overhead line faults based on UAV technology according to claim 1, characterized in that: In step S3, the Otsu method is used to determine the foreground and background thresholds. Based on the foreground and background thresholds, the OpenCV threshold function is used to classify the pixels of the second image group, separate the foreground power component region, and obtain the third image group. The edges of power components are extracted from the third image group using OpenCV's contour lookup function to obtain multiple potential key power component regions; For each of the potential critical power component regions, the boundary pixels are traversed, continuous segments are counted, the maximum number of pixels in a continuous segment and the total number of boundary pixels are obtained, and the ratio of the maximum number of pixels in a continuous segment to the total number of boundary pixels is calculated as the boundary integrity value. If the boundary integrity value is lower than the preset integrity threshold, then the potential critical power component region is closed using OpenCV's morphological operation function with a 3x3 structuring element to repair the broken boundary and obtain a complete boundary power component region; if the boundary integrity is higher than or equal to the preset integrity threshold, then the potential critical power component region is taken as a complete boundary power component region. Generate a set of power component contour features based on all the complete boundary power component regions.

6. The method for visual diagnosis and localization of overhead line faults based on UAV technology according to claim 1, characterized in that: In step S4, the complete boundary power component region is extracted from the power component contour feature set as the first feature descriptor; The first feature descriptor is used to extract features through a pre-trained deep learning network to obtain a feature vector, and the feature vector is processed through a classification layer to determine the type of power component. For each type of power component, the center coordinates of the corresponding complete boundary power component region are extracted from consecutive frame images. An image processing algorithm is used to calculate the similarity based on the center coordinates to associate power components of the same type. Based on the associated power components, a tracking path is generated, resulting in a power component tracking sequence.

7. The method for visual diagnosis and localization of overhead line faults based on UAV technology according to claim 1, characterized in that: In step S5, the center position coordinates of adjacent frames are obtained from the power component tracking sequence, the Euclidean distance between the center position coordinates is calculated, and the displacement parameters of adjacent frames of the power component are obtained. Based on the displacement parameters of adjacent frames of the power component, the boundary coordinates of the complete boundary power component region of the adjacent frame are obtained. The boundary coordinates of the complete boundary power component region of the next frame are subtracted from the displacement parameters of the adjacent frames of the power component to obtain the adjustment boundary coordinates of the next frame. The corresponding points in the boundary coordinates of the complete boundary power component region of the previous frame and the adjustment boundary coordinates of the next frame are calculated by Euclidean distance and then the average value is obtained to obtain the deformation amount of the power component between adjacent frames. Based on the deformation amount of adjacent frames of the power component, the deformation amount of adjacent frames of the power component and the code corresponding to the power component type are combined to form a deformation vector (XI,Y), where XI is the deformation amount of adjacent frames of the power component between the I-th frame and the previous frame, I is greater than 1, and Y is the code corresponding to the power component type. Based on all deformation vectors generated by all adjacent frame pairs in the power component tracking sequence, a set of power component deformation parameters is formed. For the set of deformation parameters of power components, the SVC class in the sklearn library is used to process the deformation parameter set, where the deformation vector is used as input and the classification label is used as output. The power fault type is determined by the output classification label.

8. The method for visual diagnosis and localization of overhead line faults based on UAV technology according to claim 2, characterized in that: After step S5, step S6 is also included: obtaining the deformation amount of all adjacent frames from the power component tracking sequence, calculating the average value of the deformation amount, and obtaining the power fault severity score; Power maintenance priorities are determined based on the preset intervals in which the severity scores fall, where each preset interval corresponds to a power maintenance priority. Obtain the matching overhead line station number from the first image group corresponding to the power component tracking sequence; Power line inspection reports are generated based on power fault type, power maintenance priority, and overhead line station number.

Images