Overhead line anomaly detection method and device based on unmanned aerial vehicle inspection and medium

Abstract

The invention discloses an overhead line anomaly detection method and device based on unmanned aerial vehicle inspection and a medium, and the method comprises the steps: firstly obtaining electromagnetic field information of an overhead line, planning an initial flight path of an unmanned aerial vehicle based on the electromagnetic field information, then controlling the unmanned aerial vehicle to fly along the initial flight path, and collecting dynamic electromagnetic field data of the overhead line; and then correcting the initial flight path of the unmanned aerial vehicle according to the dynamic electromagnetic field data to obtain a corrected flight path. And then, the unmanned aerial vehicle is controlled to fly along the corrected flight path, after a line image of the overhead line is collected, feature extraction is performed on the line image to obtain an overhead line area image, and then the overhead line area image is detected according to a preset overhead line anomaly detection network to obtain an anomaly detection result. Through electromagnetic field navigation, a flight path is optimized to guarantee precision, an overhead line area is accurately positioned, and tiny anomalies are identified, so that the problems of unmanned aerial vehicle inspection flight interference, difficulty in route planning and inaccurate anomaly detection are effectively solved.

Description

Technical Field

[0001] The embodiments of this application relate to, but are not limited to, the fields of drones and power system inspection, and particularly to methods, equipment and media for detecting anomalies in overhead lines based on drone inspections. Background Technology

[0002] In the current field of power system inspection, although drone inspection can reduce the burden of manual labor, it has many problems: the resolution of images and videos depends on the distance to the overhead lines. If the distance is too far, the resolution of key equipment will be low. If the distance is too close, the electric field of the overhead lines will seriously interfere with the electromagnetic compatibility and data transmission of the drone. Manually planning inspection routes requires operators to be familiar with the mountainous terrain and the direction of the overhead lines, which is difficult. Moreover, when drones fly long distances in complex areas such as mountains, GNSS signals are affected, the flight accuracy error increases, and the inspection effect is affected.

[0003] In related technologies, deep learning-based anomaly detection methods (such as image reconstruction) have shown good results in fields such as industrial quality inspection and medical imaging, but their performance is poor when directly applied to power line inspection. Ordinary feature alignment methods struggle to align target areas of overhead lines in different images, leading to large reconstruction errors. Defects are often minor geometric or texture disturbances, difficult to discern visually, and traditional reconstruction networks struggle to focus on anomalous areas, masking true defect signals. Particularly in small-sample anomaly detection tasks, the model has low sensitivity to differences in feature space between supporting and query samples, making it unable to align the wire structure and accurately reconstruct query image details. Furthermore, conventional reconstruction residuals are insufficient to amplify weak defect signals, affecting detection sensitivity and reliability. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0005] This application provides a method, equipment, and medium for detecting anomalies in overhead power lines based on UAV inspection, which can effectively solve the problems of flight interference, difficulty in flight path planning, and inaccurate anomaly detection in UAV inspections.

[0006] In a first aspect, embodiments of this application provide a method for detecting anomalies in overhead power lines based on unmanned aerial vehicle (UAV) inspection, comprising: acquiring electromagnetic field information of the overhead power line and planning an initial flight path of the UAV based on the electromagnetic field information; controlling the UAV to fly along the initial flight path and collecting dynamic electromagnetic field data of the overhead power line; correcting the initial flight path of the UAV according to the dynamic electromagnetic field data to obtain a corrected flight path; controlling the UAV to fly along the corrected flight path and collecting images of the overhead power line; extracting features from the images of the images of the images of the images of the images of the overhead power line to obtain an image of the overhead power line region; and detecting the image of the overhead power line region according to a preset overhead power line anomaly detection network to obtain anomaly detection results.

[0007] In one embodiment of this application, the step of correcting the initial flight path of the UAV based on the dynamic electromagnetic field data to obtain a corrected flight path includes: acquiring the flight state of the UAV at the previous moment while flying along the initial flight path; predicting the expected flight state of the UAV at the current moment along the initial flight path based on the previous moment's flight state; updating the previous moment's flight state based on the dynamic electromagnetic field data to obtain the current moment's flight state; obtaining the actual current flight state based on the expected flight state and the current moment's flight state; and correcting the initial flight path based on the actual current flight state to obtain the corrected flight path.

[0008] In one embodiment of this application, the step of extracting features from the line image to obtain an overhead line area image includes: extracting depth features from the line image to obtain a key feature map; performing semantic segmentation processing on the key feature map to obtain a region boundary enhancement map; and performing binary structure masking processing on the region boundary enhancement map to obtain an overhead line area image.

[0009] In one embodiment of this application, the step of detecting the overhead line area image according to a preset overhead line anomaly detection network to obtain anomaly detection results includes: inputting the overhead line area image into the overhead line anomaly detection network, performing coarse anomaly localization on the overhead line area image based on full-image structure registration, generating a preliminary anomaly heatmap; and performing anomaly detection based on the preliminary anomaly heatmap through local re-registration to obtain anomaly detection results.

[0010] In one embodiment of this application, the step of performing coarse anomaly localization on the overhead line area image based on full-image structural registration to generate a preliminary anomaly heatmap includes: performing full-image structural registration processing on the overhead line area image and a preset normal sample image to calculate the registration residual; and performing anomaly region marking processing on the overhead line area image based on the registration residual to generate a preliminary anomaly heatmap.

[0011] In one embodiment of this application, the anomaly detection based on the preliminary anomaly heatmap through local re-registration to obtain an anomaly detection result includes: performing residual region extraction processing based on the preliminary anomaly heatmap to obtain potential anomaly regions; performing mask generation processing on the potential anomaly regions to obtain saliency masks; performing cropping processing on the potential anomaly regions according to the saliency masks to obtain multiple local blocks; performing re-registration processing on the corresponding regions of each local block and normal sample images to obtain local re-registration results; performing anomaly score calculation processing on each local block based on the local re-registration results to obtain anomaly scores for each local block; and performing weighted fusion processing on the anomaly scores of all local blocks to obtain an anomaly detection result.

[0012] In one embodiment of this application, obtaining the electromagnetic field information of an overhead line includes: obtaining the three-phase current of the overhead line; calculating the magnetic field information of the overhead line based on the three-phase current; obtaining the electric field information above the overhead line; and obtaining the electromagnetic field information of the overhead line based on the magnetic field information and the electric field information.

[0013] In one embodiment of this application, the method further includes: transmitting the anomaly detection result to an overhead line anomaly intelligent management platform, wherein the anomaly detection result includes an original overhead line image, an anomaly mask image, and anomaly category information; obtaining corresponding anomaly information based on the original overhead line image, the anomaly mask image, and the anomaly category information; synchronizing the anomaly information to a relevant application; and confirming the anomaly areas marked in the anomaly information based on the anomaly information received by the relevant application and in conjunction with the actual situation on site.

[0014] On the other hand, embodiments of this application provide an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the overhead line anomaly detection method as described above.

[0015] On the other hand, embodiments of this application provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the overhead line anomaly detection method as described above.

[0016] This application provides an overhead power line anomaly detection method, electronic device, and computer-readable storage medium based on UAV inspection. First, the electromagnetic field information of the overhead power line is acquired, and an initial flight path of the UAV is planned based on this information. Then, the UAV is controlled to fly along the initial flight path, collecting dynamic electromagnetic field data of the overhead power line. The initial flight path is then corrected based on the dynamic electromagnetic field data to obtain a corrected flight path. Subsequently, the UAV is controlled to fly along the corrected flight path, and after acquiring images of the overhead power line, feature extraction is performed on the images to obtain an image of the overhead power line region. Then, a preset overhead power line anomaly detection network is used to detect anomalies in the image of the overhead power line region to obtain anomaly detection results. This application uses electromagnetic field information as a navigation signal to plan the initial flight path of the UAV and flexibly corrects the initial flight path based on dynamic electromagnetic field data, dynamically optimizing the trajectory and ensuring flight accuracy. Feature extraction from the line image can accurately locate the overhead power line region, improving the accuracy of anomaly detection. Detection of the overhead power line region image using the overhead power line anomaly detection network can accurately identify micro-cracks and low-intensity anomalies, thereby improving the accuracy of anomaly detection. Attached Figure Description

[0017] Figure 1 This is a flowchart of an overhead line anomaly detection method provided in one embodiment of this application; Figure 2 This is provided in one embodiment of the present application. Figure 1 The detailed flowchart of step 130; Figure 3 This is a flowchart of a drone inspection self-navigation method provided in one embodiment of this application; Figure 4 This is provided in one embodiment of the present application. Figure 1 The detailed flowchart of step 150; Figure 5 This is a flowchart of feature extraction from a line image provided in one embodiment of this application; Figure 6 This is a comparison diagram of a conventional residual structure and an inverted residual structure provided in one embodiment of this application; Figure 7 This is a diagram of the overall framework of U-Net provided in one embodiment of this application; Figure 8 This is provided in one embodiment of the present application. Figure 1 The detailed flowchart of step 160; Figure 9 This is a flowchart illustrating the structure of a RegAD model provided in one embodiment of this application; Figure 10 This is provided in one embodiment of the present application. Figure 8 The detailed flowchart of step 820; Figure 11 This is a flowchart of an overhead line anomaly detection method provided in another embodiment of this application; Figure 12 This is an overall flowchart of an overhead line anomaly detection method provided in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be noted that although the flowchart shows a logical order, in some cases, the steps shown or described may be performed in a different order than that shown in the flowchart. The terms "first," "second," etc., used in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the structures, proportions, sizes, etc., depicted in the drawings are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effects and purposes achieved by this application, should still fall within the scope of the technical content disclosed in this application. Similarly, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are only for clarity of description and are not used to limit the scope of implementation of this application. Changes or adjustments in their relative relationships, without substantially altering the technical content, should also be considered within the scope of implementation of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] In the current field of power system inspection, while drone inspections can alleviate the burden of manual inspections, the resolution of images and videos is heavily dependent on the distance between the drone and overhead lines and their surroundings. If the drone is too far from the overhead lines, the image resolution of critical equipment can be low; if the drone is too close, the electric field of the overhead lines will severely interfere with the drone's electromagnetic compatibility and data transmission. Manually planning inspection routes requires drone operators to have a deep understanding of mountainous terrain and the direction of the overhead lines; route planning is a significant challenge for frontline drone operators. Furthermore, long-distance flight in complex areas such as mountains can affect GNSS signal strength, leading to increased flight accuracy errors, which also impacts the effectiveness of overhead line inspections.

[0022] In related technologies, deep learning-based anomaly detection methods, such as those based on image reconstruction or feature matching, have achieved good results in fields such as industrial quality inspection and medical imaging. However, directly applying these methods to power line inspection often yields poor results. On the one hand, ordinary feature alignment methods (such as optical flow estimation and feature distance matching) are difficult to align with target areas of overhead lines in different images, leading to inaccurate reconstruction errors. On the other hand, defects usually manifest as minute geometric or texture perturbations, which are extremely difficult to distinguish visually, making it difficult for traditional reconstruction networks to focus on abnormal areas, thus masking the true defect signals.

[0023] Especially in small-sample anomaly detection tasks, the model has low sensitivity to spatial differences in features between support samples and query samples, and cannot effectively align wire structures, making it impossible for normal support samples to accurately reconstruct detailed regions of the query image. At the same time, due to the small area and weak signal of the defect region, the conventional reconstruction residual cannot sufficiently amplify the anomalous signal, affecting the sensitivity and reliability of the final detection.

[0024] In view of this, embodiments of this application provide a method, electronic device, and computer-readable storage medium for detecting anomalies in overhead power lines based on UAV inspection. The method first acquires the electromagnetic field information of the overhead power line and plans the initial flight path of the UAV based on this information. Then, the UAV is controlled to fly along the initial flight path, collecting dynamic electromagnetic field data of the overhead power line. The initial flight path of the UAV is then corrected based on the dynamic electromagnetic field data to obtain a corrected flight path. Subsequently, the UAV is controlled to fly along the corrected flight path. After acquiring images of the overhead power line, feature extraction is performed on the images to obtain an image of the overhead power line region. Then, an anomaly detection network is used to detect the image of the overhead power line region to obtain an anomaly detection result. Embodiments of this application use electromagnetic field information as a navigation signal to plan the initial flight path of the UAV and flexibly correct the initial flight path based on dynamic electromagnetic field data, dynamically optimizing the trajectory and ensuring flight accuracy. Feature extraction from the line image can accurately locate the overhead power line region, improving the accuracy of anomaly detection. Detection of the image of the overhead power line region using the anomaly detection network can accurately identify micro-cracks and low-intensity anomalies, thereby improving the accuracy of anomaly detection.

[0025] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0026] Reference Figure 1 , Figure 1 This is a flowchart of the overhead line anomaly detection method provided in the embodiments of this application. The process may specifically include, but is not limited to, steps 110 to 160.

[0027] Step 110: Obtain electromagnetic field information of overhead lines and plan the initial flight path of the UAV based on the electromagnetic field information; Step 120: Control the UAV to fly along the initial flight path and collect dynamic electromagnetic field data of the overhead lines; Step 130: Correct the initial flight path of the UAV based on the dynamic electromagnetic field data to obtain the corrected flight path; Step 140: Control the drone to fly along the corrected flight path and collect images of the overhead power lines; Step 150: Extract features from the line image to obtain the overhead line area image; Step 160: Detect the overhead line area image according to the preset overhead line anomaly detection network to obtain the anomaly detection results.

[0028] Steps 110 to 160 will be described in detail below.

[0029] In a feasible embodiment, in step 110, the electromagnetic field information of the overhead line refers to the combined information of the alternating magnetic field generated by the current in the surrounding space and the electric field above the line when the overhead line is running in the power system.

[0030] Understandably, traditional navigation methods typically rely on GPS positioning. However, in complex environments such as wind farms, GNSS signals are often affected by terrain and electromagnetic interference, leading to increased positioning errors. Introducing electromagnetic field information as a navigation signal to guide drones can improve navigation reliability. In one feasible embodiment, when acquiring the electromagnetic field information of an overhead line, the three-phase current of the overhead line (i.e., the current data in the A, B, and C phase conductors) can be acquired first, and then the magnetic field information of the overhead line can be calculated based on the three-phase current. After acquiring the electric field information above the overhead line, the electromagnetic field information of the overhead line can be integrated based on this electric field information and the magnetic field information.

[0031] Step 110 will be described in detail below with specific embodiments.

[0032] In some implementations, the three-phase current of an overhead line can be calculated using Ampere's circuit theorem. The magnetic field generated in the surrounding space. Each phase of the current at the location of the drone... The generated magnetic fields are as follows:

[0033] in, These represent the distances between the drone and phases A, B, and C, respectively. The drone in phase current The magnetic flux density generated under the action.

[0034] Furthermore, the synthetic magnetic field at the UAV point under sinusoidal steady state It is the vector superposition result of the magnetic fields generated by the three-phase currents, expressed as:

[0035] Where μ is the magnetic permeability, and its value can be taken as 4π×10⁻⁷ Henry / meter. These are the current phasors for phases A, B, and C, respectively, in amperes. Their physical meaning is the vector superposition of the magnetic fields generated by the three-phase currents.

[0036] It is understandable that the electric field is generated by the electrical charge on the overhead power lines. During the inspection of overhead power lines, changes in the electric field are closely related to the drone's flight altitude and distance. By incorporating the electric field information of the overhead power lines, more precise distance control can be achieved, ensuring that the drone maintains a safe distance from the overhead power lines throughout the inspection process. This effectively improves the safety and reliability of the inspection, ensuring stability during flight and the quality of data acquisition.

[0037] In some implementations, when obtaining electric field information above overhead lines, the mirror method can be used to calculate the electric field, and the distance between the UAV and the overhead lines can be determined by combining electric field measurements. Among them, the change in electric field strength is closely related to the flight altitude of the UAV, and the formulas for calculating electric field strength E are shown in equations (3) and (4):

[0038]

[0039] in, The distance between the drone and the phase line. The distance between the drone and the mirror image of the phase line. The horizontal angle between the drone and the phase line. The horizontal angle is the line connecting the drone to the mirror image of the phase line. It is caused by the charge on the phase line. The electric field generated at the drone along the x-axis, It is caused by the charge on the phase line. The electric field generated at the drone along the y-axis.

[0040] Furthermore, the combined electric field E at the UAV under sinusoidal steady state can be expressed as equation (5):

[0041] in, and These are the real and imaginary parts of the alternating electric field, respectively. For phase factor, and These are unit phasors in the x-axis and y-axis directions, respectively.

[0042] In some implementations, electromagnetic field information provides a novel navigation signal for UAVs, enabling precise path guidance in complex environments. By calculating the distance between the UAV and overhead power lines and combining this with real-time electromagnetic field measurements, the shortcomings of traditional GNSS navigation methods under electromagnetic interference can be effectively avoided.

[0043] In one feasible embodiment, after obtaining the electromagnetic field information of the overhead line, an initial flight path can be planned based on the electromagnetic field information. In this process, the route and electromagnetic field distribution characteristics can be combined to initially plan a path that avoids strong interference and ensures the image acquisition effect.

[0044] In one feasible embodiment, in step 120, the dynamic electromagnetic field data of the overhead line refers to the electromagnetic field change data around the overhead line measured in real time by sensors during the real-time flight of the UAV. The sensors can be mounted on the UAV or be fixed sensors pre-deployed along the overhead line.

[0045] In one feasible embodiment, in step 130, the Kalman filter (KF) algorithm can be used to correct the UAV's flight trajectory based on real-time measured electromagnetic field data. It should be noted that the Kalman filter is a powerful recursive algorithm that can effectively correct errors in the flight path by dynamically estimating the UAV's position and attitude.

[0046] In one feasible embodiment, such as Figure 2 As shown, step 130 uses the Kalman filter algorithm to correct the initial flight path of the UAV based on the dynamic electromagnetic field data to obtain the corrected flight path. This process may include, but is not limited to, steps 210 to 250.

[0047] Step 210: Obtain the flight status of the UAV at the moment before it flies along the initial flight path; Step 220: Based on the flight status at the previous moment, predict the expected flight status of the UAV along the initial flight path at the current moment; Step 230: Update the flight status of the previous moment based on the dynamic electromagnetic field data to obtain the flight status of the current moment; Step 240: Based on the expected flight status and the current flight status, obtain the actual current flight status; Step 250: Correct the initial flight path based on the actual current flight status to obtain the corrected flight path.

[0048] In a feasible embodiment, in step 210, the flight state at the previous moment refers to the parameters reflecting the motion state of the UAV at the moment that has just passed (such as time t-1), such as its position, attitude angles (such as pitch angle, roll angle, and heading angle), and flight speed.

[0049] In a feasible embodiment, in step 220, the expected flight state refers to the position, attitude, and other states that the UAV should theoretically be in at the current moment (e.g., time t) based on the planning logic of the initial path (e.g., preset speed, heading) and the actual state at the previous moment (i.e., the ideal state assuming that the UAV flies completely along the initial path).

[0050] In one feasible embodiment, in step 230, the dynamic electromagnetic field data is the electromagnetic field change data around the overhead line measured in real time by the UAV (reflecting the actual environmental conditions). By correcting the flight state at the previous moment using this real-time data, the current flight state can be obtained that is more in line with the actual environment.

[0051] In a feasible embodiment, in step 240, the actual current flight state is the result of fusing the expected state (theoretical value under the ideal path) and the current flight state (actual environmental correction value), which can take into account both the planning logic of the initial path and the authenticity of the real-time environment, and is equivalent to the optimal state obtained by combining the two.

[0052] Step 130 will be described in detail below with specific embodiments.

[0053] In some implementations, the state of the UAV at time t can be represented by its position (x(t), y(t), z(t)) and attitude angles (Φ(t), θ(t), Ψ(t)), and the UAV path planning problem can be equivalent to a discrete system, whose Kalman filter state equation is shown in equation (6):

[0054] in, It is the state phasor of the UAV at time k, consisting of the UAV's position, flight state, and calculated electromagnetic field values. Let be the input variable matrix of the UAV state at time k, composed of measured electric and magnetic fields, and A and B be system parameters. This represents system state noise.

[0055] In some implementations, the current position of the UAV can be predicted using the state at the previous moment. During the cruise, the motion of the UAV conforms to Newton's second law, and the state equation and measurement equation are continuously updated, with the predicted state values ​​constantly corrected based on the actual values. The prediction equations for the UAV flight process are shown in equations (7) and (8):

[0056]

[0057] in, It is the state prediction phasor of the UAV at time k. Let be the prediction error covariance at time k, and Q be the state noise covariance.

[0058] In some implementations, the state of the UAV is updated by electromagnetic field measurements (i.e., dynamic electromagnetic field data), and the measurement equation is shown in equation (9):

[0059] in, Here, C represents the measured value at time k, and C is the measurement system parameter. For measuring noise.

[0060] In some implementations, comparing the measured value (i.e., the current flight state) with the predicted value (i.e., the expected flight state) can yield a more accurate state.

[0061] Calculate the Kalman gain (a key parameter in Kalman filtering used to balance prediction and measurement errors):

[0062] Updated state estimate:

[0063] Update the covariance matrix:

[0064] in, Let be the correction gain matrix at time k, and R be the measurement noise covariance.

[0065] In this embodiment, the Kalman filter algorithm can provide the UAV with powerful real-time path correction capabilities. By continuously updating the state estimation, the stability and accuracy of the UAV in electromagnetic interference environments can be greatly improved, ensuring that the UAV can accurately perform inspection tasks.

[0066] See Figure 3 , Figure 3This is a flowchart of a UAV inspection self-navigation method based on Kalman filtering provided in one embodiment of this application. This process corresponds to the implementation of step 130. The specific process is as follows: First, initialize the UAV's position and simultaneously set a safe distance from the overhead line (for example, set to 4 meters directly above the B-phase conductor of the overhead line; this distance avoids interference from strong electric fields while ensuring image acquisition resolution). Second, the UAV's onboard edge computing box (a device responsible for real-time data processing and control command generation) establishes a connection with the electromagnetic field probe (a sensor used to detect the intensity and distribution of the electromagnetic field around the overhead line). The probe collects dynamic electromagnetic field data of the overhead line in real time. Based on this, combined with the previously calculated magnetic and electric field estimates, the current position of the UAV is planned, and the predicted position is initially corrected using the real-time collected dynamic electromagnetic field data. To smooth out errors during operation, the UAV's flight attitude (including parameters such as pitch and roll angles) can be adjusted every 0.5 seconds. Next, by dynamically adjusting the Kalman gain, the state equation (Equation (6)) and prediction equation (Equations (7) and (8)) are updated to further correct the deviation between the predicted position value and the actual measured value. The updated pose information (including position coordinates and attitude angles) is extracted from this information and used as a basis for real-time navigation to guide the UAV flight. Finally, the airborne edge computing box generates specific control commands based on the above series of calculation results to drive the UAV to fly along the corrected path and complete the automatic inspection task of the overhead line.

[0067] In one feasible embodiment, after controlling the UAV to fly along the corrected flight path and acquiring images of the overhead power lines, feature extraction can be further performed on the images to obtain an image of the overhead power line area. For example... Figure 4 As shown, the execution process of step 150 may include, but is not limited to, steps 410 to 430.

[0068] Step 410: Extract deep features from the line image to obtain key feature maps; Step 420: Perform semantic segmentation on the key feature maps to obtain region boundary enhancement maps; Step 430: Perform binary structure masking on the region boundary enhancement map to obtain the overhead line region image.

[0069] In a feasible embodiment, in step 410, the line image refers to the original image of the overhead line (including the line, towers, background environment, etc.) collected by the UAV while flying along the corrected path; "deep feature extraction" refers to mining deep, discriminative features in the image (such as the edge texture, shape contour, spatial distribution pattern, etc. of the overhead line, rather than simple shallow features such as color and brightness) through a deep learning model (such as convolutional neural network); the key feature map is a visual representation of these deep features, which can highlight the essential differences between the overhead line and other background elements. It is understood that overhead lines usually occupy a small area of ​​the image against a complex natural background, and the key feature map obtained by deep feature extraction of the line image can accurately locate the overhead line in the image.

[0070] In a feasible embodiment, in step 420, semantic segmentation processing refers to using an algorithm to classify and label the pixels in the key feature map according to semantic categories (such as "overhead line area", "background area", "tower area", etc.) to clarify the spatial range of different targets; the region boundary enhancement map is the image obtained after segmentation, in which the boundary between the overhead line and other areas is enhanced (such as clearer edge lines and higher contrast), further distinguishing the target area from non-target areas.

[0071] In a feasible embodiment, in step 430, "binary structure mask processing" refers to converting the region boundary enhancement map into a binary image (mask) containing only black and white colors. The region belonging to the overhead line is marked with one color (such as white), and the background region that is not the overhead line is marked with another color (such as black), forming a mask template. The overhead line region image is the region containing only the overhead line that is accurately extracted from the original image through this mask, eliminating irrelevant background interference and focusing the core target for subsequent anomaly detection.

[0072] Step 150 will be described in detail below with specific embodiments.

[0073] In this embodiment, the lightweight segmentation network MobileUNet can be used to accurately locate overhead power lines in an image. The segmentation mask is then incorporated into the subsequent feature registration process, allowing modeling and anomaly detection only for the overhead power line region. This effectively suppresses background noise interference and improves the accuracy and robustness of anomaly detection. MobileUNet is a lightweight semantic segmentation network that integrates the MobileNet and U-Net architectures. It uses MobileNetV2 as the encoder backbone, and the decoder borrows the symmetric skip connection mechanism of U-Net to fuse semantic information from different levels of the encoder with upsampled features. Multiple deconvolution layers are then used to gradually restore spatial resolution, generating a segmentation mask of the same size as the original image. The output layer uses either a Sigmoid or Softmax activation function to achieve pixel-level accurate differentiation between the overhead power line region and the background.

[0074] See Figure 5 The specific process of using MobileUNet to extract features from overhead line images to obtain images of overhead line areas includes: First, the original images of overhead lines are collected by a drone; then, the original images are preprocessed, including filtering (to suppress image noise) and normalization (to unify the distribution range of image data), to make the images more suitable for subsequent network calculations; then, the preprocessed images are input into the MobileUNet network, and after network calculation, the probability field of each pixel belonging to the overhead line is output (i.e., the probability distribution of each pixel corresponding to the "overhead line" category, reflecting the possibility that the pixel is an overhead line); subsequently, the probability field is thresholded (e.g., setting the threshold to 0.5, if the "overhead line probability" of a pixel is ≥0.5, then the pixel is determined to belong to the overhead line, otherwise it is determined to be the background); finally, a binary mask image is generated (overhead line areas are marked as 1, and background areas are marked as 0), thereby accurately extracting the image of the area containing only overhead lines.

[0075] In some implementations, MobileUNet employs the lightweight and efficient deep feature extractor MobileNetV2. Through depthwise separable convolution and inverted residual blocks, this extractor maintains good feature extraction capabilities even under computationally limited conditions. Specifically: Traditional convolution operates simultaneously in both spatial and channel dimensions, resulting in high computational and parameter counts. MobileNetV2's Depthwise Separable Convolution divides standard convolution into two stages: Depthwise Convolution and Pointwise Convolution, significantly reducing computational complexity. In the Depthwise Convolution stage, each input channel uses a separate kernel for spatial convolution, without mixing information between channels. This reduces both computational cost and model parameters. Specifically, assuming the input feature map has a spatial dimension of DxD, the kernel size is KxK, and the number of input channels is M, the computational complexity of Depthwise Convolution is:

[0076] Pointwise convolution is used to mix information from different channels after depthwise convolution. It uses a 1x1 convolution kernel to perform a linear combination of channel dimensions at each spatial location. If the number of output channels is N, the computational complexity of pointwise convolution is:

[0077] Therefore, the total computational complexity of the depthwise separable convolution is:

[0078] Compared to the computational complexity of traditional convolution... Depthwise separable convolutions significantly reduce computational and parameter overhead. This efficient architecture is well-suited for deployment on computationally limited devices such as drones, and is particularly suitable for efficiency-critical vision tasks such as semantic segmentation.

[0079] In some implementations, such as Figure 6 As shown, this embodiment employs an inverted residual block structure, which includes the following steps: First, the input feature channels are expanded using 1×1 point convolutions to increase the feature dimension and enhance expressive power; then, 3×3 depthwise separable convolutions are used to extract spatial information from the expanded features, reducing computational complexity while preserving key feature information; finally, 1×1 point convolutions are used again to compress the channel dimension back to its original size for subsequent network processing. This structure effectively improves feature extraction efficiency and non-linear expressive power while ensuring a lightweight model.

[0080] In some implementations, the U-Net architecture is used to accurately reconstruct the target boundary. U-Net is a widely used structure in the field of medical image semantic segmentation, particularly adept at handling tasks where the target shape is continuous but the contour is complex. Its core advantages lie in skip connections and symmetrical structure, making it suitable for segmenting overhead lines and complex backgrounds. For example... Figure 7 As shown, the U-Net structure consists of two symmetrical parts: the encoder extracts deep abstract semantic information through convolution and downsampling operations; the decoder gradually restores spatial resolution through upsampling and convolution. For thin, linear, or even partially occluded overhead lines, each layer of the U-Net decoder concatenates or adds the corresponding encoder feature maps to form skip connections. This operation compensates for the loss of edge information during the decoder's upsampling process, making the model more accurate in restoring the boundaries of the target structure.

[0081] In some implementations, overhead power line regions can be extracted using a binarization strategy. During the inference phase, MobileUNet performs pixel-level semantic segmentation on the input image, outputting a probability map of the same size as the original image (i.e., the probability of an overhead power line existing at each pixel location, with a value range of [0,1]). This probability map is essentially the network's confidence prediction of each pixel belonging to the "overhead power line" category. To transform this continuous probability output into a discrete binary structure mask, a fixed threshold is used: typically, an empirical threshold (e.g., 0.5) is set for each pixel location. If its probability value If the signal is positive, it is considered foreground (overhead line), and the mask value is set to 1; otherwise, it is considered background, and the mask value is set to 0. This process can be represented as:

[0082] in This represents the final binary mask image. The threshold value is set, usually 0.5.

[0083] After binarization, the resulting mask image has clear structural boundaries and can be directly used to extract the target region from the original image. Specifically, the mask image is multiplied pixel-by-pixel by the original image, i.e.:

[0084] Through the above operations, pixels in the image with a mask value of 1 (i.e., the overhead power line area) are retained, while all background pixels are set to zero, thus obtaining a "structural region image" that contains only the overhead power line area. This image is then used as input for feature extraction and feature enhancement operations, ensuring that the model focuses only on the actual target area in subsequent steps, eliminating background interference.

[0085] In one feasible embodiment, after obtaining an image of the overhead line area, the image can be input into an overhead line anomaly detection network for detection, thereby obtaining the corresponding anomaly detection results. For example... Figure 8 As shown, the execution process of step 160 may include, but is not limited to, steps 810 to 820.

[0086] Step 810: Input the overhead line area image into the overhead line anomaly detection network, perform coarse anomaly localization on the overhead line area image based on full-image structure registration, and generate a preliminary anomaly heatmap; Step 820: Perform local re-registration anomaly detection based on the preliminary anomaly heatmap to obtain anomaly detection results.

[0087] In one feasible embodiment, the overhead line anomaly detection network can be a dual-registration anomaly detection framework designed for overhead line scenarios, based on the RegAD model. RegAD (Registration-based Few-Shot Anomaly Detection) is a few-shot anomaly detection model for industrial vision and natural image scenarios. Its core idea is to achieve accurate localization of anomaly regions by relying only on a small number of normal samples through a registration prediction mechanism. Figure 9 As shown, RegAD mainly consists of a shared feature encoder and a prediction module, employing a Siamese architecture. The input is a pair of images of the same category, including a normal image (support) and a query image. This architecture does not require image reconstruction or rely on dense annotations, making it suitable for anomaly detection tasks of structural targets (such as industrial surfaces and overhead power lines) under limited sample conditions. Specifically, the overhead power line anomaly detection network, while inheriting the advantages of RegAD's limited sample detection, adopts a "dual registration" mechanism. For example, on the one hand, spatial registration aligns the image of the overhead power line region to be detected with the normal sample image in spatial location (e.g., correcting the angle and scale differences of the line); on the other hand, feature registration achieves matching at a deeper feature level (e.g., aligning key features such as the texture and shape of the line). Through this dual registration, it can more accurately capture the unique anomaly patterns of overhead power lines (such as wear and foreign object attachment), efficiently distinguishing normal lines from abnormal areas even with a limited number of normal samples, ultimately achieving accurate detection and localization of overhead power line anomalies.

[0088] In a feasible embodiment, in step 810, after inputting the image of the overhead line area into the overhead line anomaly detection network, the network can first start from the overall structure of the entire image (such as the overall direction and shape layout of the overhead line), match and align the input image with the standard structure of a normal overhead line, thereby roughly finding the area that does not conform to the normal structure, completing the coarse (preliminary, approximate) anomaly localization, and generating a preliminary anomaly heat map (in the form of a heat map, the color depth reflects the probability of an anomaly in the area).

[0089] In a feasible embodiment, in step 820, the suspected abnormal local areas in the heat map are further subjected to more refined structural matching and alignment (i.e., "local re-registration"). Through this precise local detection, the specific location and type of the abnormality are finally determined, and the abnormality detection result is obtained (such as determining where there are broken strands, foreign objects attached, or other abnormalities).

[0090] In a feasible embodiment, the execution process of step 810 may include, but is not limited to: first, performing full-image structural registration processing on the overhead line area image and the preset normal sample image, and calculating the registration residual; then, performing abnormal region marking processing on the overhead line area image based on the registration residual, and generating a preliminary abnormal heat map.

[0091] In one feasible embodiment, such as Figure 10 As shown, the execution process of step 820 may include, but is not limited to, steps 1010 to 1060.

[0092] Step 1010: Based on the preliminary anomaly heatmap, perform residual region extraction processing to obtain potential anomaly regions; Step 1020: Perform mask generation processing on potential abnormal regions to obtain saliency masks; Step 1030: Based on the saliency mask, the potential abnormal regions are cropped to obtain multiple local blocks; Step 1040: Perform re-registration processing on the corresponding regions of each local block and the normal sample image to obtain the local re-registration result; Step 1050: Based on the local re-registration results, perform anomaly score calculation for each local block to obtain the anomaly score for each local block; Step 1060: Perform weighted fusion processing on the anomaly scores of all local blocks to obtain the anomaly detection results.

[0093] Step 160 will be described in detail below with specific embodiments.

[0094] Combination Figure 9In the first stage, the overhead line anomaly detection network based on the RegAD model first extracts features from a pair of input images of the same category (such as the image to be monitored and support images), and then constructs feature representations layer by layer through three convolutional residual modules. The image to be monitored is the image of the overhead line region to be detected, and the support images refer to a small number of normal sample images (usually known to be without anomalies) belonging to the same "overhead line" category as the image to be monitored. These support images provide the model with a reference benchmark for the structure and features of "normal overhead lines," helping the model to locate anomalies through "registration (image alignment) + feature comparison." To enhance structural alignment capabilities, a spatial transformation network is integrated after each residual module to perform affine transformations on the feature maps, enabling the support images to be progressively registered to the spatial structure of the query image at multiple scales.

[0095] After registration, the resulting deep feature pairs , The input will be fed into a Siamese shared encoder for structural alignment discrimination. This encoder consists of a parameter-shared backbone network E, which extracts structural features from two input branches. Simultaneously, to construct a direction-aware feature prediction relationship, a prediction head P is added to one branch to generate a prediction vector; the other branch employs a stopped gradient propagation operation to avoid information leakage, forming an asymmetric prediction mechanism. The negative cosine similarity loss is defined in the first direction as:

[0096] in and These represent the predicted feature map of the registered image to be detected and the feature map of the supporting image, respectively. This represents L2 norm normalization. The loss function reflects the consistency of the registered structure in one direction. To enhance training stability and avoid degeneracy, RegAD further borrows from the SimSiam framework and constructs a symmetric loss function, namely:

[0097] By comparing the consistency of aligned features in the encoding space, the model generates a structural matching response map at each spatial location. Regions that still exhibit significant feature inconsistencies after registration are identified as anomalous candidate regions. Based on this response map, a preliminary anomaly heatmap can be generated, completing coarse localization at the whole-map level.

[0098] Next, refined anomaly detection based on local re-registration of the anomaly region is performed, as follows: First, regions with large residuals are extracted from the registration results of the first stage as potential anomaly regions. These regions typically correspond to low-intensity anomalies such as microcracks and surface wear. To further focus on these regions, a saliency mask is generated:

[0099] in, Indicates position The residual value, For adaptive thresholding, it is typically set to the residual plot mean plus a certain number of standard deviations, such as... , and , where are the mean and standard deviation of the residual plot, respectively, and k is the adjustment coefficient.

[0100] Next, based on the saliency mask, the anomalous region is cropped into several smaller patches to improve local detail representation. These patches are upsampled or enlarged to enhance their spatial resolution, allowing for a more complete expression of local details. For each high-resolution local patch, a spatial transformation network is used to re-register it with the corresponding region in the supporting image. Due to the smaller registration range and more localized deformation, the spatial transformation network can capture more subtle structural deviations. This process is equivalent to performing a high-precision alignment again in regions where anomalies are known to exist, to confirm whether they are genuine anomalies and to identify finer attributes such as their boundaries, shape, and intensity.

[0101] Ultimately, by employing this fusion strategy, which takes into account both global and local anomaly information, high-quality anomaly detection results with sharp boundaries and clear structures can be obtained.

[0102] In one feasible embodiment, after detecting anomalies in the overhead line area image using an overhead line anomaly detection network and obtaining anomaly detection results, depth anomaly judgment processing can be performed based on these results, such as... Figure 11 As shown, the process may include, but is not limited to, steps 1110 to 1140.

[0103] Step 1110: Transmit the anomaly detection results to the overhead line anomaly intelligent management platform. The results include the original overhead line image, anomaly mask image, and anomaly category information. Step 1120: Obtain the corresponding anomaly information based on the original overhead line image, anomaly mask image, and anomaly category information; Step 1130: Synchronize the exception information to the relevant applications; Step 1140: Based on the abnormal information received by the relevant application, confirm the abnormal areas marked in the abnormal information in combination with the actual situation on site.

[0104] Specifically, after the anomaly detection results are transmitted to the overhead line anomaly intelligent management platform, the platform can extract key information such as the corresponding original overhead line image, anomaly mask (marking the anomaly location), and anomaly category information (such as broken strands, foreign objects, etc.) from the anomaly detection results. Then, based on the image, mask, and category information, it organizes and generates clear anomaly information, and then synchronizes the organized anomaly information to the relevant application used by the staff. After seeing the anomaly information through the application, the staff can further determine whether there are any overhead line anomalies in the corresponding area based on the actual situation on site.

[0105] See Figure 12 , Figure 12 This is an overall flowchart of an overhead power line anomaly detection method provided in one embodiment of this application. The overall process is as follows: After acquiring the electromagnetic field information of the overhead power line, the initial flight path of the UAV is first planned based on this information (e.g., setting a safe distance parameter from the line). Then, the initial flight path is corrected by using a Kalman filter algorithm combined with the dynamic electromagnetic field data collected by the UAV in real time, and finally, a corrected flight path adapted to the real-time environment is formulated. While controlling the UAV to fly along the corrected flight path, the original image data of the overhead power line is collected simultaneously. After preprocessing the original image (e.g., denoising and normalization), it is input into the lightweight segmentation network MobileUNet for deep feature extraction. The overhead power line region is segmented through an encoder-decoder structure, and the region image containing only the overhead power line is obtained after binary structure masking. Then, the overhead power line region image is input into the overhead power line anomaly detection network designed based on RegAD for detection, and the anomaly detection result is obtained. When the anomaly detection result shows an anomaly, the platform generates clear anomaly information (such as an anomaly segmentation map) based on the result and synchronizes it to the relevant applications of the staff. After the staff makes a final confirmation of the anomaly area in combination with the actual situation on site, they take targeted measures (such as arranging maintenance, review and investigation).

[0106] This application also discloses an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable by the processor. When the computer program is executed by the processor, it implements the overhead line anomaly detection method described above.

[0107] This application also discloses a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the overhead line anomaly detection method described above.

[0108] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting anomalies in overhead power lines based on unmanned aerial vehicle (UAV) inspection, characterized in that, include: The electromagnetic field information of the overhead power line is obtained, and the initial flight path of the UAV is planned based on the electromagnetic field information; Control the UAV to fly along the initial flight path and collect dynamic electromagnetic field data of the overhead line; The initial flight path of the UAV is corrected based on the dynamic electromagnetic field data to obtain the corrected flight path; Control the drone to fly along the corrected flight path and collect images of the overhead power line; Feature extraction is performed on the line image to obtain an image of the overhead line area; The image of the overhead line area is detected by a preset overhead line anomaly detection network to obtain anomaly detection results.

2. The overhead line anomaly detection method according to claim 1, characterized in that, The step of correcting the initial flight path of the UAV based on the dynamic electromagnetic field data to obtain the corrected flight path includes: Obtain the flight status of the UAV at the moment before it flies along the initial flight path; Based on the flight state at the previous moment, the expected flight state of the UAV at the current moment along the initial flight path is predicted; The flight state at the previous moment is updated based on the dynamic electromagnetic field data to obtain the flight state at the current moment; Based on the expected flight state and the current flight state, the actual current flight state is obtained; The initial flight path is corrected based on the actual current flight state to obtain the corrected flight path.

3. The overhead line anomaly detection method according to claim 1, characterized in that, The step of extracting features from the line image to obtain an overhead line area image includes: Depth feature extraction is performed on the line image to obtain a key feature map; The key feature map is semantically segmented to obtain a region boundary enhancement map; The region boundary enhancement map is processed by binary structure masking to obtain the overhead line region image.

4. The overhead line anomaly detection method according to claim 1, characterized in that, The step of detecting anomalies in the overhead line area image according to a preset overhead line anomaly detection network to obtain anomaly detection results includes: The image of the overhead line area is input into the overhead line anomaly detection network. Based on the full-image structure registration, coarse anomaly localization is performed on the image of the overhead line area to generate a preliminary anomaly heatmap. Anomaly detection is performed based on the preliminary anomaly heatmap, with local re-registration, to obtain anomaly detection results.

5. The overhead line anomaly detection method according to claim 4, characterized in that, The process of performing coarse anomaly localization on the overhead line area image based on full-image structure registration to generate a preliminary anomaly heatmap includes: The image of the overhead line area is subjected to full-image structural registration with a preset normal sample image, and the registration residual is calculated. Based on the registration residual, the image of the overhead line area is processed to mark abnormal regions, generating a preliminary abnormal heatmap.

6. The overhead line anomaly detection method according to claim 4, characterized in that, The anomaly detection based on the preliminary anomaly heatmap, through local re-registration, yields anomaly detection results, including: Based on the preliminary anomaly heatmap, residual region extraction processing is performed to obtain potential anomaly regions; The potential abnormal regions are subjected to mask generation processing to obtain a saliency mask; Based on the saliency mask, the potential abnormal region is cropped to obtain multiple local blocks; The corresponding regions of each local block and the normal sample image are re-registered to obtain the local re-registration result; Based on the local reregistration results, anomaly score calculation is performed on each local block to obtain the anomaly score for each local block. The anomaly scores of all the local blocks are weighted and fused to obtain the anomaly detection results.

7. The overhead line anomaly detection method according to claim 1, characterized in that, The acquisition of electromagnetic field information of overhead lines includes: Obtain the three-phase current of the overhead line; The magnetic field information of the overhead line is calculated based on the three-phase current. Obtain the electric field information above the overhead line; The electromagnetic field information of the overhead line is obtained based on the magnetic field information and the electric field information.

8. The overhead line anomaly detection method according to claim 1, characterized in that, The method further includes: The anomaly detection results are transmitted to the overhead line anomaly intelligent management platform. The anomaly detection results include the original overhead line image, anomaly mask image, and anomaly category information. Based on the original overhead line image, the anomaly mask image, and the anomaly category information, the corresponding anomaly information is obtained; Synchronize the abnormal information to the relevant applications; Based on the abnormal information received by the relevant application, the abnormal areas marked in the abnormal information are confirmed in combination with the actual situation on site.

9. An electronic device, wherein, The electronic device includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the overhead line anomaly detection method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the overhead line anomaly detection method as described in any one of claims 1 to 8.

Images