A method for identifying damage of a power distribution network conductor insulation sleeve

Abstract

The application discloses a kind of identification methods for the damage of distribution network wire insulation sleeve, it is related to the field of power system automation, the method comprises: synchronous acquisition wire laser point cloud and infrared thermal imaging data and carry out timing alignment;By the improved Gaussian filter denoising based on spatial distance and surface normal angle joint weighting and execute spatial fusion registration, construct the multi-dimensional feature point cloud model containing geometry and thermal characteristics;Based on the damage score evaluation screening suspected damaged area of geometric distortion and abnormal temperature rise fusion;Utilize deep fusion convolutional neural network to identify damage category, and calculate damaged area, depth and volume parameters by combining alpha shape surface reconstruction algorithm;Based on geospatial information, realize the accurate positioning of damaged point and generate identification report.The application realizes the fine quantitative analysis and automatic inspection of damaged area by deep coupling of geometry and thermal characteristics, improves the detection robustness and power operation intelligent level.

Description

Technical Field

[0001] This invention relates to the field of power system automation, and specifically to a method for identifying damage to the insulation sleeves of distribution network conductors. Background Technology

[0002] With the large-scale development of power systems, the application scope of overhead insulated conductors in distribution networks is constantly expanding. They play a key role in solving the problem of narrow line corridors and improving power supply reliability. Among them, online identification technology for damaged conductor insulation is crucial to ensuring the safe operation of lines. This technology monitors and analyzes the surface condition of conductors to promptly detect sheath damage caused by lightning strikes, tree line friction, atmospheric corrosion, and mechanical stress. This high-precision identification method can effectively prevent leakage and line breakage accidents, and achieve accurate location of potential safety hazards and scientific assessment of the degree of damage.

[0003] Traditional online identification technologies for damaged conductor insulation in power distribution networks rely heavily on visible light or infrared images to build image recognition models to determine the damaged area. In scenarios with severe smoke obscuring the view or heavy background noise interference, the segmentation accuracy for conductor boundaries and micro-cracks is extremely low. During damage assessment, sunlight can cause conductors to heat up, leading to the identification of damage as conductor damage. Furthermore, existing technologies can only roughly determine the presence or absence of damage through image recognition models, making it difficult to identify specific damage characteristics. This over-reliance on manual identification makes it difficult to achieve fully automated operation covering the entire process from data acquisition to damage identification and location. Moreover, it requires large-scale structural modifications to existing cables, significantly increasing implementation costs. In long-distance lines, it is difficult to achieve accurate location and qualitative analysis of the degree of damage.

[0004] Therefore, this invention provides a method for identifying damage to the insulation sleeve of power distribution network conductors. It can effectively preserve minute crack features while removing noise, thus significantly improving detection accuracy. It introduces a new evaluation method, reducing false identifications caused by external environmental factors. Through surface reconstruction technology, it achieves a leap from undamaged conductors to the specific degree of conductor damage, forming a fully automated process. Furthermore, it does not require any structural modifications to existing insulated conductors, has good compatibility, and reduces implementation costs. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying damage to the insulation sleeves of power distribution network conductors, which can effectively solve the problems of low segmentation accuracy, high probability of misjudgment, large limitations in damage judgment, insufficient automation of processing procedures, and high implementation costs in the aforementioned background technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying damage to the insulation sleeve of a power distribution network conductor includes the following specific steps:

[0008] Step 1: Collect laser point cloud data and infrared thermal imaging data of power distribution network conductors: The conductors in the target area are scanned synchronously by UAV-borne laser radar sensor and infrared thermal sensor to obtain the original three-dimensional point cloud dataset with geographic coordinate information and the corresponding multispectral thermal feature image. The point cloud dataset and thermal feature image are time-aligned based on the high-precision global satellite navigation system timestamp. The pulse signal output by the global positioning system is simultaneously connected to the trigger interface of the laser radar and infrared camera to ensure that the data streams of the two sensors are hard synchronized under a unified clock reference.

[0009] Step 2, Preprocessing multi-source data and performing spatial fusion and registration: An improved Gaussian filtering algorithm is used to denoise the original point cloud. The algorithm's weights are determined by spatial distance and the surface normal angle. Burrs, micro-cracks, and insulation damage edges on the conductors appear as high-frequency noise in the point cloud, indicating abrupt changes in spatial coordinates. Traditional filtering, if using strong smoothing, would smooth out both noise and micro-cracks; if using weak smoothing, both noise and micro-cracks would be preserved. The algorithm adds the normal angle to remove noise and preserve minute features. The characteristics of noise are large spatial distance and random normal angle. The crack edge is characterized by its short spatial distance and drastic change in normal vector. By using a filtering algorithm that combines spatial distance and normal angle, outlier interference points that are not wire targets in the environment were successfully removed. The infrared thermal features were mapped to the point cloud space using the calibrated extrinsic parameter matrix. By identifying the wire edge in the infrared image and the wire geometric contour in the point cloud, a spatial transformation matrix containing rotation matrix and translation vector was constructed. A multi-dimensional feature point cloud model containing geometric structure information, thermal distribution information and physical properties was constructed, and each three-dimensional point cloud coordinate was assigned a corresponding absolute temperature value and emissivity attribute.

[0010] Step 3: Extract the geometric and thermal anomaly features of the damaged insulation sleeve. Calculate the local curvature, normal deviation, and local temperature change rate of each point in the multi-dimensional feature point cloud model. Calculate the deviation between the measured and theoretical values ​​based on the preset standard morphological parameters of the insulation sleeve. In a purely visual scheme, abnormal curvature is reported as wire damage; in a purely infrared scheme, wire damage is reported if the detected temperature exceeds a threshold. For complex field environments, this single-factor judgment is prone to misjudgment. To improve the accuracy of wire damage judgment, a damage score composed of the curvature change rate and temperature gradient deviation is introduced. The assessment logic considers that when the external environment is hot, the surface temperature of the conductor insulation sleeve increases dramatically, but the geometry of the conductor remains unchanged. For micro-cracks, the temperature may not be high yet, but the geometry has changed. By screening out suspected damaged areas with significant geometric distortion or abnormal temperature rise, and dynamically adjusting the weighting coefficients in the damage score assessment logic based on ambient wind speed, light intensity, and operating load current, this fusion mechanism utilizes the physical property that geometric information is not affected by light. In strong light environments, it suppresses the weighting of thermal characteristics and effectively avoids false hot spot false alarms caused by specular reflection on the conductor surface.

[0011] Step 4: Identify damage type and calculate damage severity parameters: Input the suspected damaged area into a pre-trained deep fusion convolutional neural network model. Through the continuous three-dimensional convolutional layers and collaborative attention mechanism inside the model, the weight ratio of geometric features and thermal features is automatically assigned to identify the damage type of the insulation sleeve. A triangular mesh model of the damaged area is constructed using surface reconstruction technology based on the Alpha Shape Algorithm. The area, depth, and volume parameters of each damaged area are calculated. The specific physical model is mapped to the geometric parameters through the Alpha Shape Algorithm to restore the real damage situation. This is helpful in judging the remaining life of the insulation sleeve, predicting the time of breakdown risk, and assisting in building a fully automated process from data acquisition to damage report generation.

[0012] Step 5, output the identification results and achieve accurate positioning: Based on the geospatial information of the multi-dimensional feature point cloud model, determine the absolute geographic coordinates of each damaged point and its relative position on the conductor segment, generate a distribution network conductor insulation sleeve damage identification report containing damage coordinates, damage type, damage geometric parameters and on-site panoramic photos, and combine with the power geographic information system ledger to locate the tower section where the damaged point is located.

[0013] In step 1, the pulse frequency of the UAV-borne lidar sensor is set to a preset pulse frequency to ensure that the density of the scanning point cloud formed on the surface of the conductor is greater than or equal to a preset density threshold, and the single-point ranging accuracy is within a preset error range. The wavelength of the infrared thermal sensor is within a predetermined wavelength range, the temperature measurement range is adapted to the preset ambient temperature range in the power distribution network operating environment, and the thermal sensitivity is below a preset sensitivity threshold. During the data acquisition process, the UAV, as a mobile sensing platform, maintains its flight altitude within a preset vertical distance from the conductor and its flight speed within a preset inspection speed range. The data streams of the two sensors are hard-synchronized using a high-precision global satellite navigation system timestamp, and the synchronization error is less than a preset synchronization time threshold.

[0014] In step 2, the improved Gaussian filtering algorithm updates the coordinates of the current point using the weighted average of points within the neighborhood window. The weight function is determined by the distance from the point to the center and the angle between the surface normals, and the weight calculation formula is introduced as follows:

[0015]

[0016] In the weight calculation formula, To improve the weights in the Gaussian filtering algorithm and reflect the contribution of neighborhood points to the final coordinate result, For spatial distance, The angle between the normal and the direction of motion. and With preset parameters to control the smoothness, the algorithm can retain feature information of the insulation sleeve edge and micro-cracks while removing noise.

[0017] In step 2, spatial fusion registration adopts a combination of homography transformation based on feature points and perspective projection transformation. By identifying the edge of the conductor in the infrared image and the geometric contour of the conductor in the point cloud, a spatial transformation matrix is ​​constructed. The projection error after registration is controlled within a preset projection error threshold. Each three-dimensional point cloud coordinate is assigned a corresponding absolute temperature value and emissivity attribute, forming a four-dimensional point cloud structure with physical properties.

[0018] In step 3, the screening of suspected damaged areas adopts a multi-criteria fusion judgment logic, calculating the local curvature change rate and local temperature gradient deviation of each point in the point cloud data, and introducing a damage score evaluation formula:

[0019]

[0020] In the injury score assessment formula, The loss score reflects the likelihood that the point belongs to the area of ​​insulation damage. The higher the loss score, the more likely the point is to be the actual location of insulation damage, and vice versa. The normalized rate of change of curvature reflects the degree of abnormality in the geometric morphology of the conductor surface; The normalized local temperature gradient deviation reflects the degree of abnormality in the thermal distribution on the conductor surface. and For the preset weighting coefficients, when When the damage score exceeds the preset threshold, the point and its neighborhood are marked as suspected damaged areas; otherwise, the point and its neighborhood are marked as normal areas, effectively eliminating false hotspot interference caused by ambient light reflection.

[0021] In step 4, the deep fusion convolutional neural network model adopts a three-dimensional residual network structure. This model extracts spatial geometric features through a predetermined number of consecutive three-dimensional convolutional layers, and introduces a collaborative attention mechanism in the feature map fusion stage to automatically allocate the weight ratio of geometric features and thermal features. The model training set contains actual inspection damage samples of a predetermined size, covering the insulation sleeve damage state under different lighting, different humidity and different wire specifications.

[0022] In step 4, when calculating the parameters of the damaged area, a surface reconstruction technique based on the Alpha Shape Algorithm is used. By constructing a compact triangular mesh model, the volume of the damaged pits is calculated, and a volume quantization formula is introduced:

[0023]

[0024] In the volume quantization formula The volume of the constructed damaged area model, Let the area of ​​the grid element be... The area is calculated by summing the areas of all triangular facets located within the damaged boundary, representing the normal displacement of the center point of the infinitesimal element relative to the surface of the standard cylindrical conductor.

[0025] In step 5, the precise positioning system integrates a differential GPS receiver, achieving dynamic positioning accuracy that meets preset accuracy requirements. The identification report records the longitude, latitude, and altitude of the damaged point in detail. Simultaneously, the system automatically links to the power distribution network geographic information system ledger to determine which two towers the damaged point is located between, as well as the predetermined horizontal distance to the nearest tower. The system can also assess the risk level of the identified damage by calculating the ratio of the identified damage depth to the total thickness of the insulation sleeve. If the ratio is greater than the first preset ratio threshold and accompanied by a sudden increase in local electric field strength and temperature, a level one alarm is automatically triggered. If the ratio is within the preset ratio range, a level two warning is triggered. The system background automatically provides maintenance priority suggestions based on historical maintenance records.

[0026] This invention is applied to an all-weather automated inspection system. The drone autonomously avoids obstacles along a pre-set power line route, with the route overlap set to a preset threshold. The inspection flight speed is a preset inspection speed. The backend edge computing unit performs real-time data slicing. The total time from data acquisition to generating a preliminary identification report is less than a preset time threshold. Simultaneously, the all-weather automated inspection system also includes multi-temporal evolution analysis of the same conductor segment. By storing a multi-dimensional feature point cloud model of historical inspections, a health record for the conductor insulation is established. Using a long short-term memory neural network model based on meteorological parameters and operating current data, the evolution trend of discovered minor damage is predicted, and the predicted lead time to reach critical breakdown risk is greater than or equal to a preset prediction time threshold.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] This invention deeply integrates the geometric features of laser point clouds with the thermal anomaly features of infrared thermal imaging, effectively solving the problem of low accuracy in traditional visible light image recognition under complex lighting, smoke obstruction, and background interference. The accuracy of identifying micro-cracks and hidden damage is improved to a preset high accuracy level, and the false alarm rate is significantly reduced. In particular, by introducing the surface normal angle as a weight penalty term, the technical problem of micro-crack features disappearing due to edge smoothing in conventional filtering algorithms is solved, so that micro-cracks on the surface of the insulating sleeve can still be completely preserved after filtering.

[0029] This invention introduces a damage score assessment mechanism based on the fusion of geometric distortion and abnormal temperature rise. By utilizing the physical property that geometric information is unaffected by light, it automatically suppresses the weight of thermal features under strong light conditions and enhances the weight of thermal features under low wind speed and stable operating current conditions. This effectively eliminates false hot spot interference caused by specular reflection on the conductor surface, significantly reducing the false alarm rate of inspections. At the same time, it achieves dual physical confirmation of geometric morphology and thermal state, which greatly improves the detection rate of early hidden damage.

[0030] This invention utilizes surface reconstruction technology based on the Alpha shape algorithm to achieve quantitative analysis of the damaged area, depth, and volume of the insulation sleeve. The measurement error is within a very small preset error range, which provides scientific data support for subsequent insulation performance evaluation and maintenance decisions, and changes the limitation of traditional methods that can only qualitatively determine whether there is damage.

[0031] This invention automates the entire process from autonomous drone inspection and synchronous data collection to intelligent identification and positioning at the back end. The efficiency of line inspection is significantly improved compared with traditional manual visual inspection and ordinary image inspection. The real-time push of identification reports greatly shortens the response time from fault discovery to fault handling. At the same time, it does not require any structural modification to the existing insulated wires and is fully compatible with the huge number of various types of wires in the power distribution network, which avoids the high cost of large-scale cable replacement.

[0032] This invention achieves precise output of the absolute coordinates of the damaged point through deep integration of differential positioning technology and power geographic information system, ensuring that maintenance personnel can quickly and accurately reach the work location and eliminating the problem of difficulty in finding the damaged point in long-distance lines.

[0033] This invention, through multi-temporal evolution analysis and life prediction functions, realizes the transformation from post-fault handling to preventive maintenance, timely detects and addresses initial minor wear, effectively avoids wire breakage accidents caused by insulation sleeve damage, and greatly improves the safety factor of power grid operation and power supply reliability. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall technical solution architecture of a method for identifying damage to the insulation sleeve of a power distribution network conductor according to an embodiment of this application;

[0035] Figure 2 This is a schematic diagram of the core principle framework of multi-source feature deep fusion and damage degree quantitative analysis in a method for identifying damage to the insulation sleeve of a power distribution network conductor according to an embodiment of this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments:

[0037] Example 1

[0038] In the method for identifying damaged insulation sleeves of power distribution network conductors, the first step is to collect laser point cloud data and infrared thermal imaging data of the power distribution network conductors. This step involves simultaneously scanning the conductors within the target area using a UAV-borne laser radar sensor and an infrared thermal sensor. The aim is to obtain a raw 3D point cloud dataset with high-precision geographic coordinate information and corresponding multispectral thermal feature images. During data acquisition, the UAV, acting as a mobile sensing platform, is set at a vertical distance of 5 to 10 meters from the conductor, with a flight speed maintained between 3 and 5 m / s. The pulse frequency of the UAV-borne laser radar sensor is set to 200 kHz to ensure that the density of the scanned point cloud formed on the conductor surface is greater than 500 points per square meter, ensuring accurate single-point ranging. The accuracy is strictly controlled within 20mm. Meanwhile, the infrared thermal sensor operates in the wavelength range of 8μm to 14μm, with a thermal sensitivity NETD of less than 50mK and a temperature measurement range set from -20℃ to 150℃, effectively adapting to various ambient temperature ranges in the power distribution network operation environment. To achieve time consistency of multi-source data, the system performs hard synchronization of the data streams of the two sensors through a high-precision global satellite navigation system timestamp. Specifically, the PPS pulse signal output by the GNSS receiver is simultaneously connected to the trigger interfaces of the lidar and the infrared camera, ensuring that the acquisition time difference between each frame of infrared image and each group of laser point cloud is less than 1ms, thereby performing precise time alignment processing on the point cloud dataset and thermal feature images based on the synchronization clock.

[0039] After data acquisition, step 2 is performed to preprocess the multi-source data and perform spatial fusion and registration. Due to the complexity of the field environment, the original point cloud often contains a large number of noisy points. Therefore, an improved Gaussian filtering algorithm is used to denoise the original point cloud, removing outlier interference points caused by trees, towers, and birds in the environment. This improved Gaussian filtering algorithm updates the coordinates of the current point using the weighted average of points within the neighborhood window. Its core lies in the construction of the weight function, which is determined by the distance from the point to the center and the angle between the surface normals. It aims to retain the feature information of the insulation sleeve edge and micro-cracks. The weight calculation formula is introduced as follows:

[0040]

[0041] In the weight calculation formula, This represents the weights in the improved Gaussian filter algorithm. This represents the Euclidean spatial distance between the point to be processed and the center point of its neighborhood. This represents the angle between the normal vector of the point to be processed and the average normal vector of its neighborhood. and The preset parameters representing the smoothness are typically set within the ranges of 0.01 to 0.05 and 10 to 20 degrees, respectively, based on the initial density of the point cloud and the surface roughness of the conductor. After denoising, the infrared thermal features are mapped to the point cloud space using the calibrated extrinsic parameter matrix. Spatial fusion registration employs a combination of homography transformation and perspective projection transformation based on feature points. By identifying the pixel coordinates of the conductor edge in the infrared image and the three-dimensional coordinates of the conductor's geometric contour in the point cloud, a spatial transformation matrix containing the rotation matrix R and the translation vector t is constructed. The projection error after registration is limited to within a one-pixel threshold. Thus, each three-dimensional point cloud coordinate is assigned a corresponding absolute temperature value and emissivity attribute, constructing a multi-dimensional feature point cloud model containing geometric structure information and thermal distribution information, forming a four-dimensional point cloud structure with physical properties.

[0042] Step 3 then extracts the geometric and thermal anomaly features of the damaged insulation sheath. By analyzing the microscopic properties of each point in the multi-dimensional feature point cloud model, suspected damaged areas are identified. First, the local curvature of each point is calculated. The degree of surface curvature is determined by analyzing the eigenvalues ​​of the covariance matrix of that point and its k-neighborhood points. Then, the normal deviation is calculated, i.e., the angle of deviation of the normal vector of each point relative to the normal of the standard cylindrical axis of the conductor. Finally, combined with infrared thermal data, the local temperature change rate is calculated, reflecting the thermal gradient distribution on the conductor surface in the spatial dimension. The system calculates the deviation between the measured and theoretical values ​​based on preset standard morphological parameters of the insulation sheath. To achieve multi-criteria fusion judgment, a damage score evaluation formula is introduced:

[0043]

[0044] In the injury score assessment formula, For the loss score, The normalized rate of change of curvature. This represents the normalized local temperature gradient deviation. and The preset weighting coefficients are dynamically adjusted based on ambient wind speed, light intensity, and operating load current. For example, geometric feature weights are added in strong light environments. ,when When the damage score exceeds the preset threshold, the system marks the point and its 50mm radius neighborhood as a suspected damaged area. This effectively eliminates false hotspot interference caused by specular reflection of ambient light, ensuring that the selected area has significant geometric distortion or abnormal temperature rise.

[0045] After identifying the suspected damaged area, step 4 is executed to identify the damage type and calculate the damage severity parameters. The system inputs a subset of features from the suspected damaged area into a pre-trained deep fusion convolutional neural network model. This model employs a three-dimensional residual network structure, extracting complex spatial geometric features through 32 to 64 consecutive three-dimensional convolutional layers. A collaborative attention mechanism is introduced during the feature map fusion stage, automatically allocating the weight ratios of geometric and thermal features based on the signal-to-noise ratio of the input data. The deep fusion convolutional neural network model can identify four main types of damage: lightning burns manifested as localized high-temperature carbonization and craters, and long strips of... The system identifies the types of damage, including dendritic friction with rudimentary wear surfaces and periodic hot spots, atmospheric corrosion manifested as diffuse surface flaking and slight overall temperature rise, and mechanical stress damage manifested as cracks or localized depressions. Simultaneously, it utilizes surface reconstruction technology based on the Alpha Shape Algorithm to perform detailed modeling of the damaged area. By constructing a compact triangular mesh model, the volume of the damaged pits is calculated. The area is calculated by accumulating the areas of all triangular facets within the damage boundary. The depth parameter is taken as the maximum normal displacement of each point within the region relative to the standard cylindrical surface of the conductor. In this way, a quantitative description of the degree of damage is achieved.

[0046] Finally, step 5 is executed to output the identification results and achieve precise positioning. Based on the geospatial information inherent in the multi-dimensional feature point cloud model, the system determines the absolute geographic coordinates of each damaged point, including longitude, latitude, and altitude, as well as its relative position on the entire conductor. The precise positioning system integrates a differential global positioning system receiver, and with the RTK correction data provided by the ground reference station, the dynamic positioning accuracy reaches within 20mm. After the identification is completed, the system automatically generates a report on the damage identification of the distribution network conductor insulation sleeve. The report includes the coordinates of the damaged point, the type of damage, the geometric parameters of the damage, and a panoramic photo of the site. This report is pushed to the operation and maintenance management platform in real time through a 5G or private network wireless communication link. At the same time, the system automatically links to the distribution network geographic information system ledger to determine which two specific towers the damaged point is located between, and the predetermined horizontal distance from the nearest tower.

[0047] The invention also includes a risk level assessment of the identified damage level. The system calculates the ratio of the identified damage depth to the total thickness of the insulating sleeve. If the ratio is greater than 0.6 and accompanied by a temperature anomaly caused by a surge in local electric field strength (i.e., the temperature rise exceeds 15K above the ambient temperature), a first-level alarm is automatically triggered. If the ratio is between 0.3 and 0.6, a second-level warning is triggered. The system background automatically provides maintenance priority suggestions based on historical maintenance records and fault evolution patterns to assist operators in formulating troubleshooting plans.

[0048] This invention is applied to an all-weather automated inspection system. The drone autonomously avoids obstacles by following a pre-set CAD flight path for the power line, with the path overlap set to over 80% to ensure full coverage of the point cloud data. The backend edge computing unit performs real-time data slicing processing. The total time from data acquisition to generating a preliminary identification report is controlled within 180 seconds. Furthermore, the system includes a multi-temporal evolution analysis module. By storing multi-dimensional feature point cloud models of the same conductor on different dates, a health record for the conductor insulation is established. Using a long short-term memory neural network model, based on meteorological parameters including humidity, salt spray concentration, and UV index, as well as operating current data, the system predicts the evolution trend of discovered minor damage. When the predicted lead time to critical breakdown risk is greater than 30 days, the system automatically issues a preventative maintenance reminder.

[0049] In a specific application scenario, an inspection was conducted on a 10kV distribution network section. This section is 1.5km long and located in a typical mountainous forest area. A drone equipped with a RIEGL VUX-1 lidar and a FLIR A65 infrared sensor flew parallel to the conductor at a speed of 4m / s, 5m above it. During the inspection, the lidar acquired the three-dimensional coordinates of the conductor surface at a line scan frequency of 100Hz, achieving a point cloud density of 850 points per meter. The infrared sensor captured real-time radiation temperature measurement images of the conductor. In step 2, the improved Gaussian filtering algorithm removed the point cloud obstructed by tree branches 2m above the flight path. In step 3, the system identified a point at a distance of 752m. The suspected area with a value of 0.88 was identified as a lightning strike burn after deep fusion convolutional neural network analysis in step 4. The burn area was measured to be 45.2 mm² and the maximum depth was 1.8 mm using a surface reconstruction algorithm. The damage ratio was 0.51 relative to the insulation thickness of 3.5 mm for this type of conductor. Since the local temperature rise was measured to be 8.5 K, the system classified it as a level 2 warning and generated an identification report showing its geographical location as XX degrees North latitude and XX degrees East longitude, with a horizontal distance of 15.6 m from the nearest tower No. 12.

[0050] Example 2

[0051] Based on Example 1, this example provides an implementation method for identifying damage to conductor insulation sleeves in power distribution networks based on edge computing and cloud collaboration, aiming to further improve response speed and data processing efficiency in large-scale power distribution network inspections.

[0052] When performing step 1 to collect laser point cloud data and infrared thermal imaging data of power distribution network conductors, the UAV-borne control system integrates a high-performance edge computing module, such as an NVIDIA Jetson series processor. This processor is directly connected to the Ethernet interface of the LiDAR and the GMSL2 interface of the infrared camera. During the flight of the UAV, the edge computing module receives the raw bit stream generated by the sensor in real time and performs preliminary data structuring processing. The raw packet data generated by the LiDAR is parsed into point cloud frames with intensity information and multiple echo markers. Each point cloud frame contains 32,000 discrete points. The 16-bit Raw format image output by the infrared thermal sensor is converted into a temperature matrix with emissivity correction.

[0053] In step 2, during the preprocessing of multi-source data and the execution of spatial fusion registration, the edge computing module executes an improved Gaussian filtering algorithm through a hardware acceleration unit. To improve real-time performance, the algorithm adopts a voxel grid-based index structure, dividing the space into small cubes of 10mm×10mm×10mm. Only voxels containing wire feature points participate in the weight calculation formula, thereby significantly reducing computational overhead. During spatial fusion registration, the system pre-stores the UAV load installation deviation matrix obtained through precise laboratory calibration, including rotational deviations composed of roll angle, pitch angle, and yaw angle, as well as translational displacement deviations in the X, Y, and Z axis directions. During real-time processing, the system combines the instantaneous attitude data output by the IMU to dynamically compensate for the registration offset caused by UAV vibration, ensuring that the projection error between the infrared image and the point cloud remains within the preset error threshold throughout the entire flight.

[0054] In step 3, when extracting the geometric morphological and thermal anomaly features of the damaged insulation sleeve, the edge computing module uses a parallel processing architecture to simultaneously calculate the curvature and temperature gradient of tens of thousands of points. To further improve the screening accuracy, the system introduces a dynamic adaptive threshold mechanism. Based on the weather sensor feedback for the day, when the ambient wind speed is greater than 5 m / s, the system automatically lowers the thermal weight coefficient in the damage score assessment formula. To reduce the risk of missed detections due to insignificant thermal characteristics caused by enhanced convective heat transfer, the system will increase the weight of geometric morphology features. It relies heavily on the rate of change of curvature to identify geometric distortions caused by external forces.

[0055] When performing step 4 to identify the damage category and calculate the damage severity parameters, the edge computing module runs a lightweight deep fusion convolutional neural network model that has been quantized and compressed. This model uses pruning and knowledge distillation techniques to compress the number of model parameters to 15% of the original size while maintaining the recognition accuracy. The model takes less than 50ms to infer the suspected damaged area. The damage parameters identified, consisting of area, depth, and volume, are encapsulated in a lightweight JSON data format. For severe damage such as deep ablation where the insulation layer is about to break through, the edge computing module triggers a high-priority transmission task to upload the original high-definition point cloud slice and the original infrared image of the area to the cloud server for secondary verification through 5G network slicing technology.

[0056] When executing step 5 to output the identification results and achieve precise positioning, the cloud server receives inspection data from multiple drone nodes. The cloud system has a more powerful power geographic information system processing capability, which can realize automatic matching of power grid ledgers throughout the province. The system projects the absolute geographic coordinates of each damaged point to the local power spatial coordinate system and automatically calculates the distance between the damaged point and the conductor suspension point with an accuracy of 10mm. The cloud platform will also combine the historical maintenance database to query the lightning strike frequency, tree line friction history and corrosion environment level of the line section in the past 3 years, thereby providing auxiliary verification for the damage category output in step 4.

[0057] Furthermore, the multi-temporal evolution law analysis in this embodiment runs on a cloud big data platform. The system analyzes the growth rate of the damaged area by comparing historical inspection data from the same period. If the depth of a mechanical stress failure increases by more than 20% between two inspections at a 3-month interval, the system will raise its evolution risk level from "low" to "high" even if the current depth ratio has not yet reached the first-level alarm threshold. This dynamic trend-based assessment mechanism can detect the accelerated deterioration of the insulation sleeve earlier.

[0058] In a specific application scenario, for a high-risk power distribution network section crossing a highway, the inspection system adopted an edge computing mode. After the drone detected wear on the insulation sleeve of a conductor located directly above the highway, the edge computing module completed the preliminary identification within 5ms, determining it to be mechanical stress damage with a damaged area of ​​120mm². Since this location is a particularly important area, the edge computing module immediately called on the 5G uplink bandwidth to transmit the three-dimensional mesh model of the damaged point to the cloud. The cloud expert system, combined with the frequent crosswind weather records of the area, determined that the wear was caused by wind deflection and predicted that it might lead to insulation failure within the next two months. The system then generated a maintenance task order which was directly sent to the inspector's mobile terminal. The task order included the specific path to navigate to the highway shoulder below the fault point.

[0059] Example 3

[0060] Based on the above embodiments, this embodiment describes in detail a special identification scheme for high humidity salt spray corrosion environment. This scheme further refines the performance and processing flow of multi-dimensional feature point cloud model in complex environment.

[0061] In step 1, for the salt fog environment in coastal areas, the wavelength range of the infrared thermal sensor was optimized to 10.5μm to 12.5μm to reduce the absorption effect of water vapor in the atmosphere on radiation transmission. The lidar sensor on the UAV was equipped with the function of transmitting multiple echoes at once. By analyzing the energy distribution of the first and last echoes, the surface of the insulating sleeve covered by salt scale can be identified more effectively. During the data acquisition process, the UAV's global satellite navigation system timestamp was adjusted to correct for ionospheric delay, ensuring that the synchronization error under strong magnetic field interference is still less than the preset synchronization time threshold.

[0062] In step 2, the preprocessing stage, the improved Gaussian filtering algorithm adds processing for diffuse reflection noise generated by salt spray particles, and the parameters in the weight calculation formula are adjusted accordingly. The algorithm is set as a function related to environmental visibility. When visibility is low, the algorithm increases the size of the neighborhood window to suppress particle noise through a wider range of spatial smoothing. During the spatial fusion registration process, the system specifically introduces an emissivity correction model. Since salt scale can significantly change the emissivity of the insulating sleeve surface, the system inverts the surface roughness through laser echo intensity and then corrects the absolute temperature value corresponding to each point cloud coordinate in real time, thereby eliminating spurious temperature differences caused by changes in surface emissivity.

[0063] In step 3, for damage extraction targeting atmospheric corrosion, the system adds a texture feature dimension. Besides curvature and normal deviation, the system also calculates the uniformity of point cloud density distribution within local micro-elements. Atmospheric corrosion typically manifests as uneven surface pitting; this microscopic feature exhibits high-frequency fluctuations in the normal deviation distribution map, which introduces… In the evaluation logic, a sensitive item for corrosion characteristics has been added. In the screening of suspected damaged areas, the system uses a multi-scale sliding window algorithm to capture not only millimeter-level pitting corrosion, but also centimeter-level flaking areas.

[0064] In step 4, when identifying damage categories, the deep fusion convolutional neural network model incorporated a large number of damage samples from salt spray environments into the training set. For the identified "atmospheric corrosion" category, the system not only calculated the conventional area and depth but also introduced the calculation of the surface roughness index. By analyzing the dispersion of the 3D point cloud within the damaged area, the severity of corrosion was assessed. When calculating the parameters of the damaged area, the parameters of the Alpha Shape Algorithm were used. The irregularity of the corroded surface is dynamically adjusted to construct a non-convex envelope surface that better matches the actual corrosion morphology.

[0065] In step 5, the output of the identification results integrates environmental factors such as salt spray deposition rate. The identification report will include an "environmental corrosion correlation" index, which reflects the correlation between the current damage and the geographical environment. While locating the damage point, the system will automatically query the distance of the location from the coastline and the prevailing wind direction in the area. The precise positioning system uses differential positioning technology, which can ensure that the relative position error of the damage point on the conductor segment is within a very small range, even under the condition of obvious multipath effect at the seaside.

[0066] For risk level assessment, in salt spray environment, the system lowers the temperature threshold for triggering the first level alarm. Since salt scale has a certain conductivity in humid environment, even if the insulation sleeve is not completely damaged, local small arc discharge will generate unique high-frequency hot spots. After the system recognizes this feature, it will immediately combine the electric field strength evolution model to determine whether surface flashover will occur, and push the warning to the operation and maintenance management platform in real time.

[0067] In a specific application scenario, an identification was performed on a 10kV overhead line along a coast. This line is subjected to long-term sea winds, and the surface of the insulation sleeve is covered with obvious salt deposits. The multi-dimensional feature point cloud model collected by the drone shows that between towers 3 and 4, the conductor surface exhibits a large area of ​​minor geometric distortions; the extraction in step 3... The threshold was exceeded in multiple locations; the depth residual network in step 4 identified these areas as having severe atmospheric corrosion, with an average corrosion depth of 0.8 mm and a maximum local depth of 1.5 mm. Since the line is the main power supply line for the local industrial area, the system automatically upgraded the alarm level to level two based on the risk assessment model, even though the depth ratio was only 0.43, considering the risk of surface flashover caused by salt spray, and gave a maintenance suggestion of "live cleaning or spraying RTV coating within one week".

[0068] Example 4

[0069] This embodiment provides a scheme for identifying damage to the insulation sleeves of power distribution network conductors in forest fire prevention areas with high vegetation coverage. In this scheme, the segmentation of point cloud data and the identification logic for specific damage to the insulation sleeves are optimized.

[0070] In step 1, the UAV-borne lidar sensor is set to a high-density scanning mode, and the pulse frequency is increased to 400kHz. This is because in forest areas, the conductor is often blocked by leaves and branches. Increasing the point cloud density can increase the probability that the laser pulse will penetrate the gap and hit the surface of the conductor. The infrared thermal sensor simultaneously records the thermal contrast between the forest background and the conductor. Since the thermal emission of the forest background is relatively uniform, it is beneficial to reverse-locate the position of the conductor in complex geometric environments by using thermal characteristics.

[0071] In the preprocessing stage of step 2, the improved Gaussian filtering algorithm adds a point cloud classification sub-step based on eigenvalue decomposition. By analyzing the local covariance matrix, the point cloud is classified into linear features such as conductors, planar features such as tower surfaces, and scattered features such as tree canopies. This step prioritizes the retention of linear feature point clouds and removes outlier scattered feature points, thereby effectively solving the problem of outlier points caused by tree interference. In the spatial fusion registration process, the system uses the linear geometric features of conductors as strong constraints and performs fine registration through the iterative nearest point algorithm from point to line. The registered four-dimensional point cloud structure can accurately reflect the real state of the conductor insulation sleeve in the area where tree-line friction occurs frequently.

[0072] In step 3, during feature extraction, the system introduces dynamic geometric envelope analysis for the specific damage type of "tree line friction." Since tree line friction causes a flattened wear surface on the insulating sleeve, this feature manifests as a specific minimum value in curvature calculation. Simultaneously, the heat generated by friction appears as a long strip-shaped abnormal temperature rise zone extending along the conductor axis in the thermal dimension of the point cloud. The evaluation formula now includes a weighting for the spatial correlation of this elongated heat distribution.

[0073] In step 4, the deep fusion convolutional neural network model adds a temporal context awareness layer. For suspected damaged areas, the system not only analyzes the current static features, but also combines the dynamic analysis of the video stream captured by the UAV's forward-facing camera. If the damaged area is identified as being accompanied by overlapping tree branches, the model will determine it with a high probability as mechanical stress damage and potential discharge burns caused by "tree line friction". When calculating volume parameters, for the plane formed by wear, the local least squares fitting method is used to determine the normal depth of the wear surface. The calculation accuracy is further guaranteed by increasing the sampling point density.

[0074] In step 5, the generated identification report adds "surrounding vegetation risk assessment". Based on the multi-dimensional feature point cloud model, the system automatically calculates the clearance distance between the tree branch and the damaged point of the wire. If the distance is less than the preset safety threshold, the system will push the damage identification report and coordinate with the forestry department or line maintenance unit to suggest that offline tree pruning be carried out at the same time. The precise positioning system uses a differential global positioning system receiver and combines it with the forest fire prevention geographic information system to determine the precise rescue path of the damaged point in the forest area.

[0075] In the risk level assessment stage, the system adds a fire hazard weight to the identification results of forest areas. If it is found that the insulation sleeve has a local high temperature of more than 80°C caused by tree line friction, the system will immediately trigger the highest level forest fire prevention alarm and push the location coordinates to the fire prevention command center in real time through the operation and maintenance management platform.

[0076] In a specific application scenario, during a line inspection in a mountainous forest area, a drone successfully acquired the geometric outline of the conductor using multi-echo laser technology under the dense shade of trees. After removing tens of thousands of interference points from leaves in step 2, step 3 identified a suspected wear area with a length of 30cm in the center of a certain span. In step 4, the neural network model, combined with the flat curvature characteristics and thermal anomalies of the point cloud at this location, determined that long-term tree line friction had caused the insulation layer to thin. The measured wear volume was 1250mm³, and the minimum remaining insulation thickness was only 0.5mm. Due to the large amount of dry vegetation around the location, the report generated by the system directly triggered the emergency repair process, and within 1 hour, it guided personnel with infrared thermal imagers to the designated location to eliminate the hidden danger.

[0077] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying damage to the insulation sleeve of a power distribution network conductor, characterized in that, The method includes the following specific steps: Step 1: Collect laser point cloud data and infrared thermal imaging data of power distribution network conductors: Obtain raw point cloud data and corresponding multispectral thermal feature images through UAV, and perform time-series alignment processing on point cloud data and thermal feature images based on a synchronous clock. Step 2, preprocess multi-source data and perform spatial fusion registration: The collected point cloud data is subjected to an improved Gaussian filtering algorithm based on the joint weighting of spatial distance and surface normal angle to denoise the original point cloud and retain the edge features of the insulating sleeve. The infrared thermal features are mapped to the point cloud space using the calibrated external parameter matrix. Through spatial fusion registration, a multi-dimensional feature point cloud model containing three-dimensional geometric structure information and thermal distribution information is constructed. Step 3: Extract the geometric morphological features and thermal anomaly features of the damaged insulation sleeve: calculate the curvature, normal deviation and local temperature change rate of each point in the multi-dimensional feature point cloud model, and introduce a damage score evaluation formula based on the fusion of geometric distortion and abnormal temperature rise to screen suspected damaged areas. Step 4, identify the damage type and calculate the damage degree parameters: Input the suspected damaged area into the pre-trained deep fusion convolutional neural network model. Through the three-dimensional residual network structure and collaborative attention mechanism, identify the damage type of the insulating sleeve. Then, use the adaptive surface reconstruction technology based on the alpha shape algorithm to construct the triangular mesh of the damaged area and calculate the area, depth and volume parameters of the damaged area. Step 5: Output the identification results and achieve accurate positioning: Based on the geospatial information of the multi-dimensional feature point cloud model, generate a report on the identification of damage to the insulation sleeves of the power distribution network conductors and push it to the operation and maintenance management platform in real time.

2. The method for identifying damage to the insulation sleeve of power distribution network conductors according to claim 1, characterized in that, In step 2, the improved Gaussian filtering algorithm updates the coordinates of the current point using the weighted average of points within the neighborhood window. The weight function includes the spatial distance from the point to the center and the surface normal angle. The function is constructed by multiplying the negative exponential function of the Euclidean spatial distance between the point to be processed and the neighborhood center point, and the negative exponential function of the angle between the normal vector of the point to be processed and the average normal vector of the neighborhood. The spatial distance term ensures that the closer the points are in the neighborhood, the greater their weight. The normal angle term significantly reduces the weight of the neighborhood point when the angle between the normal vector of the point to be processed and the average normal vector of the neighborhood increases. During the execution of the algorithm, the algorithm identifies and retains the feature information of the edge of the insulating sleeve and the micro-cracks, and filters out the point cloud noise belonging to vegetation, buildings, towers and biological targets.

3. The method for identifying damage to the insulation sleeve of power distribution network conductors according to claim 1, characterized in that, In step 2, spatial fusion registration adopts a combination of homography transformation based on feature points and perspective projection transformation. Specifically, the spatial transformation matrix is ​​calculated by identifying the pixel coordinates of the conductor edge in the infrared image and the three-dimensional coordinates of the conductor's geometric contour in the point cloud. During the registration process, the system uses the pre-stored UAV payload installation deviation matrix, including rotational and translational deviations, and combines it with the instantaneous attitude data output by the inertial measurement unit to dynamically compensate for the registration offset caused by UAV vibration and flight attitude changes. This ensures that the projection error between the infrared image and the point cloud after spatial mapping is within a preset pixel threshold, thereby achieving accurate alignment of thermal properties in the three-dimensional geometric topology.

4. The method for identifying damage to the insulation sleeve of power distribution network conductors according to claim 1, characterized in that, In step 3, the screening process for suspected damaged areas adopts a multi-criteria fusion judgment logic. The degree of surface curvature is determined by analyzing the eigenvalues ​​of the covariance matrix of the point and its neighboring points, and the deviation angle of the normal vector of each point relative to the normal of the standard cylindrical axis of the conductor is calculated. The damage score assessment logic is achieved by weighted summation of the normalized rate of curvature change and the normalized local temperature gradient deviation. When the damage score is greater than the preset damage score threshold, the point and all points in its preset radius neighborhood are jointly marked as suspected damaged areas.

5. The method for identifying damage to the insulation sleeve of power distribution network conductors according to claim 1, characterized in that, In step 4, the deep fusion convolutional neural network model adopts a three-dimensional residual network structure to extract spatial geometric features from the suspected damaged area through a predetermined number of consecutive three-dimensional convolutional layers. Then, through a collaborative attention mechanism, the weights of geometric and thermal features are dynamically allocated according to the signal-to-noise ratio of the input data. The identified damage types include lightning burns, tree line friction, atmospheric corrosion, and mechanical stress damage. During the training phase, the model introduces quantization compression, pruning, and knowledge distillation to reduce the number of model parameters while maintaining recognition accuracy. The backend edge computing unit performs real-time slice inference on the collected data, so that the time from data collection to the generation of preliminary recognition results is within the preset time threshold.

6. The method for identifying damage to the insulation sleeve of power distribution network conductors according to claim 1, characterized in that, In step 4, surface reconstruction technology based on the Alpha shape algorithm is used to calculate the parameters of the damaged area. The shape parameters are dynamically adjusted according to the geometric characteristics of the damage type to construct a differentiated non-convex envelope surface that fits the lightning strike crater, tree line wear surface, or atmospheric corrosion pit. The area of ​​the damaged area is obtained by accumulating the areas of all triangular facets located within the damage boundary, and the depth parameter is obtained by the maximum normal displacement of each point in the area relative to the standard cylindrical surface of the conductor. For atmospheric corrosion damage, the system further calculates the density distribution uniformity of the point cloud in the local micro-element and the surface roughness index. By analyzing the dispersion of the three-dimensional point cloud in the damaged area, the severity of corrosion is evaluated, and the normal depth of the wear surface is determined by the local least squares fitting method.

7. The method for identifying damage to the insulation sleeve of power distribution network conductors according to claim 1, characterized in that, In step 5, the precise positioning system integrates a differential GPS receiver and works with real-time dynamic differential correction data provided by a ground reference station to ensure that the dynamic positioning accuracy is within the preset centimeter-level accuracy requirement. The system automatically generates an identification report that includes the longitude, latitude, and altitude of the damaged point and automatically links it to the power grid geographic information system to determine which two towers the damaged point is located between and the horizontal distance to the nearest tower. For forest-covered areas, the identification report adds a risk assessment of the surrounding vegetation and calculates the clearance distance between the tree branches and the damaged point on the conductor. When the clearance distance is less than a preset safety threshold, the system links with the line maintenance unit to trigger a tree pruning task.

8. The method for identifying damage to the insulation sleeve of power distribution network conductors according to claim 1, characterized in that, The method also includes a risk level assessment of the identified damage extent: the system calculates the ratio of the identified damage depth to the total thickness of the insulating sleeve. If the ratio is greater than a first preset ratio threshold and is accompanied by a local temperature rise exceeding the preset ambient temperature threshold, a level one alarm is automatically triggered; if the ratio is within the preset ratio range, a level two warning is triggered. In a salt spray corrosion environment, the system uses laser echo intensity to invert the surface roughness, corrects the emissivity corresponding to each point cloud coordinate in real time, and lowers the temperature threshold for triggering a level one alarm to cope with high-frequency hot spots caused by the tiny arc discharge of salt scale in a humid environment. The system background also provides maintenance priority suggestions based on historical maintenance records.

9. The method for identifying damage to the insulation sleeve of a power distribution network conductor according to claim 1, characterized in that, The method also includes multi-temporal evolution analysis of the same conductor segment: a health record of the conductor insulation sleeve is established by storing a multi-dimensional feature point cloud model of historical inspections; a long short-term memory neural network model is used to predict the evolution trend of the discovered damage based on meteorological parameters and operating current data, and to analyze the growth rate of the damaged area; when the predicted time advance to reach the critical breakdown risk is above the preset prediction time threshold, the system automatically issues a preventive maintenance prompt; the multi-temporal evolution analysis runs on a cloud big data platform, and by comparing historical inspection data from the same period, when the rate of increase in damage depth exceeds the preset growth threshold, the system automatically raises the evolution risk level of the damaged point.

Images