Airborne visual recognition adaptive inductive power harvesting method for ultra-high voltage transmission lines

Abstract

This invention discloses an airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines. It employs dynamic threshold segmentation, multi-scale feature extraction, and a lightweight convolutional network to identify transmission components. Relative entropy image matching is used to calculate the alignment deviation between the conductor pose and the magnetic ring, and to label the line's electrical attributes. A PID visual servo closed-loop control mechanism is used to precisely couple the magnetic ring and the high-voltage conductor, establishing an image-linked closed-loop verification visual feedback mechanism. Based on the image matching results, the number of coil winding turns is dynamically switched to optimize power extraction for different conductor operating conditions. This invention sets two levels of safety thresholds for attitude fine-tuning and emergency demagnetization protection, combined with multi-condition shutdown logic. It deeply integrates machine vision and adaptive inductive power extraction, enabling fully autonomous online power replenishment for UAV inspections. This solves the problems of short flight time, poor power extraction adaptability, and insufficient operational safety of traditional UAVs, effectively improving the long-distance inspection radius and automation level of ultra-high voltage lines.

Description

Technical Field

[0001] This invention belongs to the field of ultra-high voltage transmission line inspection and UAV autonomous power replenishment technology, specifically involving an airborne visual recognition adaptive induction power extraction method for ultra-high voltage transmission lines. Background Technology

[0002] Ultra-high voltage transmission lines cover a wide area and have complex terrain, making manual inspection difficult and risky. Currently, the industry generally uses drones for autonomous patrols to detect line defects and inspect corridors. Existing inspection drones rely on onboard lithium batteries for power, which have limited capacity and short flight times per patrol, requiring frequent returns to replace batteries, significantly reducing the overall inspection efficiency. If ground charging stations are used for fixed-point power replenishment, the construction cost of supporting power supply facilities in the field transmission corridors is high and the deployment is difficult, making it unsuitable for long-distance continuous inspection needs.

[0003] Existing drone inductive power generation technology has the following drawbacks: Lacking a pre-screening mechanism for image recognition, drones cannot autonomously distinguish targets such as wires, spacers, hardware, and vegetation on mountains. They are prone to performing power-taking operations on non-conductive components, resulting in power-taking failure and the risk of collision. The system does not adapt to different voltage levels and conductor cross-section specifications. The number of turns in the magnetic ring coil is fixed. The high voltage line has excessively high induced voltage, which can easily damage the onboard circuit. The low voltage line has insufficient induced power, resulting in low charging efficiency. The power supply process lacks real-time visual closed-loop safety verification. After the drone is affected by airflow and its position drifts, the distance between the magnetic ring and the wire becomes out of control, which can easily lead to insulation breakdown and equipment short circuit damage. There is a lack of graded early warning and emergency demagnetization protection logic. Traditional image recognition algorithms are computationally intensive and cannot be adapted to the low-computing-power hardware embedded on drones. They also have low target recognition accuracy in bright light, fog, and dim environments, and invalid images consume a large amount of computing resources.

[0004] Therefore, there is an urgent need to propose a drone inductive power supply solution that integrates multi-scale image recognition, lightweight convolutional classification network, adaptive coil winding adjustment, and image-linked closed-loop security verification to solve the problem of continuous power supply for long-distance autonomous inspection of ultra-high voltage lines by drones. Summary of the Invention

[0005] To address the problems of short flight time of drones, lack of adaptive adjustment for inductive power extraction, absence of visual safety closed-loop monitoring, and poor target recognition accuracy in complex environments in existing technologies, this invention provides an airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines. This method achieves adaptive power extraction through airborne visual hierarchical recognition of transmission components, automatic classification and labeling of electrical parameters using lightweight neural networks, and dynamic switching of magnetic ring coil windings. Furthermore, it constructs a real-time closed-loop verification mechanism for the entire power extraction process, thereby improving the drone's long-distance autonomous inspection endurance and operational safety in the field.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows: An airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines, executed collaboratively by a UAV equipped with a visible light depth-of-field camera, an electromagnetic induction power extraction module, and an airborne image processing unit, includes the following steps: S1. Cruise Image Acquisition and Image Identification: Control the UAV to autonomously cruise along the ultra-high voltage transmission corridor. The visible light depth-of-field camera continuously acquires real-scene images along the line. The airborne image processing unit uses an adaptive dynamic brightness threshold segmentation method to complete image identification, distinguishing between mountains, vegetation, clouds, strong light reflection interference backgrounds and ultra-high voltage conductors, spacers, tension fittings, and other target areas. Invalid interference images without transmission components are removed, and a standardized set of valid target images is output. If the standardized set of valid target images for the current cruise segment is empty, the UAV continues to acquire images until a valid set is obtained before proceeding to S2. S2. Multi-scale layered image extraction: Layered image extraction is performed frame by frame on the effective target image set. The edge detection algorithm is used to extract the shallow gray-scale contour of the conductor, the scale-invariant feature algorithm is used to extract the local texture features of the hardware, and the bidirectional coding algorithm is used to extract the deep semantic features of the conductor split structure and erection height. The three layers of features are spliced ​​together to generate a multi-dimensional feature vector library for power transmission components. S3, Image Classification of Transmission Components: A lightweight improved convolutional classification network is built, which completes image classification with a multi-dimensional feature vector library as input. It automatically divides the images into four categories: single conductors, four-split conductors, spacers, and tension fittings. Simultaneously, electrical attribute labels such as line voltage level, conductor cross-section, and erection height are labeled for each subset. If the classified subset is a spacer or tension fitting, it is determined that the target does not meet the conditions for power collection. The UAV continues to move along the preset route and jumps to S1 to re-collect subsequent corridor real-scene images. Only the single conductor and four-split conductor image subsets enter S4. S4. Standard Prototype Image Matching: Retrieve the ultra-high voltage power transmission standard prototype image library stored locally on the airborne machine, and use the relative entropy map matching algorithm to match the classified image subset with the corresponding voltage level standard prototype image. Calculate the image matching similarity. If there is a slight deviation in the matching of a single frame, perform multi-frame cross-checking. If the matching similarity of multiple consecutive frames does not meet the standard, it is determined that the current conductor is not suitable for power extraction. The UAV continues to move along the preset route, returns to S1 to re-collect subsequent corridor real-scene images, filters conductor images with matching similarity higher than the preset threshold, locks the target conductor for induction power extraction, and outputs the target conductor pixel coordinates and relative distance. S5. Precise control of UAV power-taking attitude: Based on the pixel coordinates and relative distance of the target conductor output by image matching, combined with the data of the UAV positioning module and inertial measurement module, the three-axis deviation of the body is calculated. The pitch, roll and yaw attitude of the UAV is adjusted by proportional-integral-derivative closed-loop control. The opening and closing magnetic ring of the airborne electromagnetic induction power-taking module is driven to maintain a safe insulation distance with the target conductor. After alignment, it is positioned to the sag section of the conductor and readjusted and aligned. Then the magnetic ring is driven to clamp the conductor and lock it. S6. Adaptive inductive power extraction with wires: After the magnetic ring is clamped and locked, the preset power extraction parameter matrix is ​​retrieved according to the voltage and cross-section labels marked in the image classification. The electronic switching switch dynamically switches the number of turns of the coil winding inside the magnetic ring. The magnetic ring closes and cuts the ultra-high voltage alternating magnetic field to generate induced alternating current, which is then rectified, filtered and regulated to charge the onboard energy storage battery of the UAV. S7. Image-linked power supply closed-loop verification: During the power supply process, the depth-sensing camera continuously acquires real-time images of the target conductor, extracts local ROI areas of the conductor, and cyclically executes the entire process of image identification, image extraction, image classification, and image matching. The image matching similarity is updated in real time, and two levels of safety judgment thresholds are set. If the matching similarity is lower than the first-level warning threshold, S5 is called to fine-tune the aircraft attitude. If the similarity is lower than the second-level safety threshold, the power supply circuit is disconnected, the electromagnetic ring is separated, and S4 to S6 are re-executed for cyclic verification. The power supply operation ends when any of the following conditions are met: the energy storage battery is fully charged, a manual return command is given, an obstacle ahead of the line is identified, or the onboard battery is low. Low battery priority return protection is also set up.

[0007] By adopting the above technical solution, relying on the integrated collaboration of airborne vision and inductive power extraction, the endurance and operational safety of UAV inspections of ultra-high voltage lines are significantly improved. S1 dynamic threshold filtering filters out complex environmental interference, eliminating invalid images and reducing computational power consumption; S2 three-layer feature extraction combined with a lightweight convolutional network accurately identifies conductors and automatically labels electrical parameters, avoiding the risk of accidental power extraction from fittings. Target conductors are locked through relative entropy matching, and PID servo precisely controls the safe distance between the magnetic ring and the conductor. The number of coil turns is dynamically switched according to the conductor's operating conditions, adaptively adjusting the inductive power to solve the problem of charging efficiency imbalance between high and low voltage conductors. Lightweight cyclic verification of ROI throughout the power extraction process, two-level threshold-based early warning, automatic demagnetization protection during drift, and automatic termination of operations under multiple conditions construct a visual closed-loop safety mechanism, effectively overcoming the shortcomings of traditional UAVs such as short endurance, poor adaptability, lack of safety protection, and low recognition accuracy, expanding the range of long-distance autonomous inspection operations.

[0008] Furthermore, the image identification in S1 specifically includes: S11. Image lighting uniform correction processing: Brightness equalization correction is performed on all real-world images captured by the depth-sensing camera. After correction, the entire image is equally divided into several grid sub-regions. The texture gradient value is calculated for each sub-region. The power transmission components are distinguished from the background by combining the gradient shape length and width constraints. S12. Region Threshold Judgment and Invalid Image Removal: A dynamic discrimination threshold is generated based on the average gradient of all grid sub-regions in the entire image. Continuous connected regions with gradient values ​​greater than the dynamic discrimination threshold are judged as valid regions of transmission components. Other background regions are directly cropped and removed. Only images containing valid transmission component regions are retained as valid target images. Blank interference images without conductors or fittings are completely removed, and a standardized image set is output after image discrimination.

[0009] The above technical solution first eliminates recognition interference caused by uneven lighting conditions such as strong light, fog, and dim lighting through S11 brightness equalization correction. Texture gradients are calculated by grid partitioning, and morphological constraints are used to initially segment transmission components from complex backgrounds, reducing the probability of misidentification of mountain vegetation. Then, S12 generates dynamic thresholds based on the global gradient mean, which can adaptively adapt to different corridor environments, accurately select effective areas of component connectivity, crop redundant backgrounds, and remove invalid images without transmission line components. This reduces the computational load of subsequent image layer extraction and classification matching, lowers the computing power pressure on airborne embedded hardware, and provides a standardized image dataset for high-quality input for subsequent feature extraction and conductor recognition, thereby improving the stability and overall processing efficiency of target recognition in complex field scenarios from the source.

[0010] Furthermore, the method for generating the dynamic discrimination threshold in S12 specifically includes: S121. Statistically calculate the texture gradient values ​​corresponding to all grid sub-regions, extract the maximum, minimum and average values ​​of all texture gradient values ​​respectively, and remove abnormal gradient noise data based on the maximum and minimum values. The abnormal gradient noise data are gradient values ​​that exceed the preset floating range of the average value. The floating range is matched with the corridor voltage level and the environmental correction coefficient. S122. The remaining effective texture gradient values ​​after removing abnormal gradient noise data are calculated by weighted average, and the compensation and correction are performed in combination with the preset environmental correction coefficient of the ultra-high voltage transmission corridor to generate a real-time dynamic identification threshold that is adapted to the current shooting light and environmental scene. This dynamic identification threshold is simultaneously involved in the linkage correction of the S7 secondary safety threshold. When the light is dim, the matching judgment benchmark is simultaneously lowered. S123. Compare the real-time texture gradient values ​​of each grid sub-region with the dynamic discrimination threshold one by one, and select the continuous grid connected regions whose gradient values ​​are greater than the dynamic discrimination threshold. Combined with the region length, width and shape constraints, these regions are the effective target regions of the power transmission components.

[0011] The above technical solution first eliminates gradient noise exceeding the mean fluctuation range, removing abrupt interference caused by reflections and shadows. The fluctuation range is adapted to different ultra-high voltage corridor operating conditions by combining line voltage and environmental parameters. Then, a weighted average is applied, and an environmental correction coefficient is introduced to generate a dynamic threshold, which can adaptively adjust in real time according to lighting and terrain. In dim environments, the matching judgment standard is simultaneously lowered to avoid missed detections. Simultaneously, the threshold is linked to the S7 safety verification benchmark, achieving full-process parameter coordination. Combined with grid gradient comparison and morphological constraints, the component areas are accurately divided. This not only improves the accuracy of component recognition in complex scenarios such as clouds, fog, and strong light, but also unifies pre- and post-processing logic, reduces airborne image processing errors, and ensures that image recognition and power supply safety verification are simultaneously adapted to the field environment.

[0012] Furthermore, the image classification of power transmission components in S3 specifically includes: S31. Lightweight Improved Convolutional Classification Network Construction: Based on the basic convolutional neural network, a lightweight improvement is made by removing redundant convolutional layers and fully connected layers, and embedding an attention mechanism module and a feature fusion module. This reduces the network's computational load while enhancing the feature focusing ability of multi-dimensional feature vectors, making it suitable for the embedded computing environment of UAV airborne image processing units. S32, Feature Input and Intelligent Classification: The multi-dimensional feature vector library of power transmission components generated in S2 is batch input into a lightweight improved convolutional classification network. The network learns the differences in contour, texture and semantic features of different power transmission components. It iteratively calculates the input feature vectors through preset classification weight parameters to accurately distinguish four types of targets: single conductor, four-split conductor, spacer bar and tension fitting, and generates four independent image subsets accordingly. S33. Automatic labeling of electrical attributes: A database of correlation mapping between image features and electrical attributes of transmission components is pre-established. The electrical attributes include two data sources: one is the pre-stored ledger parameters of the transmission corridor, and the other is the erection height and conductor split structure obtained by deep semantic extraction in S2. Combining the image recognition results and the preset parameters of the transmission corridor, the corresponding line voltage level, conductor cross-sectional specifications, and conductor erection height are automatically matched and labeled for each image subset after classification. This forms a classified image dataset with complete attribute labels, providing unified data support for subsequent adaptive power extraction parameter adjustment and matching threshold correction at all levels.

[0013] By adopting the above technical solution, the convolutional network is lightweighted, redundant layers are removed, and attention and feature fusion modules are added, significantly reducing computational overhead and adapting to low-computing-power hardware on UAVs, while simultaneously enhancing the line feature extraction effect. Based on multi-dimensional feature vectors, four-class classification of components is completed, accurately filtering out invalid images of spacers and tension fittings, retaining only conductor images for the power extraction process. Combining ledger and image semantic dual data sources, voltage, cross-section, and overhead line height electrical labels are automatically labeled, constructing a standardized attribute-based dataset. This provides a reliable basis for subsequent adaptive switching of coil turns and dynamic correction of image matching thresholds, preventing non-target erroneous power extraction from the source, achieving automatic matching of power extraction parameters with on-site conductor conditions, and improving recognition accuracy and power extraction adaptability.

[0014] Furthermore, the S4 standard prototype image matching step includes: S41. Accurate matching of prototype image retrieval: Based on the classification image dataset after automatic electrical attribute labeling completed in step S33 above, according to the electrical attributes of the transmission component category, voltage level, and conductor specifications corresponding to the image dataset, retrieve the standard prototype image that matches the currently identified component and transmission operating condition parameters from the airborne local storage of the ultra-high voltage transmission standard prototype image library; the standard prototype image library is organized and archived in multiple levels according to the transmission component category and the adapted transmission parameters, and matches the corresponding classification image subsets with attribute labels to complete the bidirectional matching retrieval of component type and electrical parameters.

[0015] S42. Pre-processing of matching images: For the retrieved standard prototype image and the classified target image to be identified, uniformly perform image impurity filtering, image quality correction and image normalization processing to unify the image presentation state of the image to be matched and the standard prototype image, and construct the image matching operation basis.

[0016] S43. Image feature matching based on relative entropy algorithm: Relying on the relative entropy map matching algorithm, the core feature distribution information of the image texture, contour and structure of the image to be identified and the standard prototype image after pre-normalization are extracted respectively. The overlap and difference of the feature distribution of the two sets of images are compared, and the global image matching operation is performed to determine the matching status between the image to be identified and the standard prototype image.

[0017] S44. Matching result hierarchical screening and judgment: Based on the system's preset two-level matching judgment benchmark, the image matching fit status is hierarchically identified and judged. Slight feature deviations in a single frame enter the multi-frame review process. Images with multiple consecutive frames that do not match the features are defined as invalid images. The system reverts to S1 to re-acquire cruise real-scene images. Wire images that meet the features of the standard prototype image are defined as valid target images. The corresponding power transmission wire is locked as a candidate target for inductive power extraction. S45. Multi-frame image verification and target locking: Continuously acquire multiple sets of real-time images of the same power transmission line, and sequentially execute the process of prototype image retrieval, image normalization, feature matching, and result filtering. Cross-verify and comprehensively evaluate the matching results of multiple sets of images. After verification, lock the inductive power supply target line and output the position and spatial relative state information of the target line in the image for precise attitude control of the UAV.

[0018] The above technical solution first retrieves matching standard images layer by layer according to component type and electrical parameters to ensure that the matching benchmark matches the on-site working conditions. It also uniformly corrects the image quality of the two types of images, eliminating the interference of image quality differences on matching accuracy. A relative entropy algorithm is then used to quantify the differences in feature distribution, accurately measuring the conductor matching fit. Finally, a hierarchical judgment combined with multi-frame cross-verification avoids misjudgments caused by lighting and jitter in single-frame images, eliminating failed matching images and re-acquiring images. Multi-frame comprehensive verification locks the target conductor, outputting pixel coordinates and relative distances, providing precise visual input for UAV PID attitude control, improving magnetic ring alignment accuracy, reducing the risk of power extraction offset, and ensuring stable and safe operation of inductive power extraction.

[0019] Furthermore, the image feature matching based on the relative entropy algorithm in step S43 includes: S431. Image feature information extraction: Read the image to be identified and the standard prototype image after pre-normalization respectively, and extract all feature distribution information corresponding to the texture, contour and structure inside the two images; S432. Relative Entropy Matching Operation: The relative entropy map matching algorithm is used to compare two sets of feature distribution information to obtain the overlap and difference information between the two sets of image feature distributions. S433. Matching Status Determination: Based on the overlap and difference information obtained from the relative entropy calculation, determine the matching status between the image to be identified and the standard prototype image, and output the matching status to the subsequent matching result classification and screening determination step.

[0020] The above technical solution first extracts the texture, contour, and structural feature distributions of the image to be identified and the standard prototype image, preserving the key visual information of the guide wire and avoiding matching distortion caused by the loss of local features. The relative entropy algorithm is used to quantify the overlap and difference between the two sets of feature distributions. Compared with traditional feature matching, this method can more objectively quantify the overall similarity of the images and weaken local errors caused by lighting and shooting angle. Based on the calculation results, the degree of image fit is intuitively determined, and standardized matching status data is output to supply the hierarchical screening process. This provides a quantitative basis for multi-frame verification and target guide wire locking, improves the stability of guide wire recognition, reduces alignment errors caused by image deviations, and provides reliable quantitative support for UAV attitude adjustment and power supply safety verification.

[0021] Furthermore, the S6 conductor adaptation adaptive inductive power extraction specifically includes: S61 Power Parameter Retrieval: Read the line voltage and conductor cross-section electrical attribute labels marked in step S33, call the power parameter matrix pre-stored in the device, and the winding tap position in the power parameter matrix is ​​linked and adapted with the S1 dynamic identification threshold and the S7 two-level safety matching threshold. S62, Coil winding switching control: Controls the electronic switching switch to switch the number of turns connected to the coil winding inside the magnetic ring; S63. Alternating magnetic field induction power generation: A closed magnetic ring cuts the alternating magnetic field generated by the ultra-high voltage line to generate induced alternating current. S64, Power Conditioning and Charging Output: The induced AC power is sequentially sent to the rectifier unit, filter unit, and voltage regulator unit to complete the power processing. The processed power is then sent to the UAV's onboard energy storage battery to complete the charging.

[0022] By employing the above technical solution, relying on the voltage and cross-section tags obtained from prior identification to retrieve a linked power extraction parameter matrix, and coordinating the winding position with image recognition and safety verification thresholds, operational condition linkage control is achieved. The number of turns in the magnetic ring coil is dynamically switched via an electronic switch to adapt to the magnetic field strength of different conductor specifications, addressing the pain points of excessively high high-voltage induced voltage and insufficient low-voltage power extraction. The closed magnetic ring cuts the alternating magnetic field to generate alternating current, which is then rectified, filtered, and regulated through multiple stages of power conditioning to output a stable and compatible voltage for the onboard battery. The entire process achieves automatic adaptation of power extraction parameters to the conductor, improving charging efficiency, preventing circuit overload damage to equipment, and, in conjunction with the visual recognition system, forming an adaptive autonomous power replenishment system, extending the drone's inspection endurance and reducing power supply constraints during field operations.

[0023] Furthermore, the specific subdivision steps of the S62 coil winding switching control include: S621, Operating Parameter Matching and Comparison: Retrieve the power parameter matrix obtained from S61, and combine it with the voltage level and electrical properties of the current transmission line to match the working position of the coil winding of the corresponding appropriate specification. S622, Winding On / Off Switching Execution: Through the airborne electronic switching switch, according to the matching winding working position, the corresponding coil winding inside the magnetic ring is selectively connected or disconnected to complete the switching configuration of the number of coil turns connected. S623, Winding Status Locking: After completing the coil winding switching, lock the conduction state of the electronic switching switch, fix the current magnetic ring coil winding connection specification, and adapt to the current ultra-high voltage conductor induction power extraction conditions.

[0024] The above technical solution first matches the appropriate winding position based on the conductor voltage and cross-sectional parameters, ensuring precise matching of the coil configuration to the on-site magnetic field strength. An onboard electronic switching switch quickly increases or decreases the number of coil turns. Reducing the number of turns on high-voltage conductors prevents overvoltage breakdown, while increasing the number of turns on low-voltage conductors enhances induced power generation, resolving power imbalances between different lines. After switching, the switch is locked in the on state to stabilize the winding connection specifications and prevent incorrect winding position switching due to flight vibrations. The entire control logic is automatically executed based on the electrical tags output by image recognition, requiring no manual intervention. This achieves adaptive adjustment of inductive power intake, stabilizes charging output power, protects onboard electrical equipment, and improves the reliability of the UAV's autonomous power replenishment.

[0025] Furthermore, the S7 image-linked power-on closed-loop verification specifically includes: S71. Real-time continuous image acquisition: During the power supply operation of the UAV, real-time images of the target conductor are continuously acquired through a visible light depth-of-field camera. S72, Lightweight Loop Iterative Recognition of Local ROI: The local region of interest of the target conductor is extracted to replace the whole image in the calculation. The local image is cyclically substituted to perform a complete processing flow of image discrimination, multi-scale hierarchical image extraction, power transmission component image classification and standard prototype image matching. S73, Real-time matching status update: Continuously update the image matching and fit status of the target conductor, and synchronously compare the system's preset first-level warning and second-level safety judgment criteria. The two judgment criteria are adaptively adjusted according to the voltage and cross-sectional parameters marked in S3. S74. Emergency handling of abnormal working conditions: When it is detected that the image matching status does not meet the safety judgment benchmark, it is determined that the UAV has drifted, the airborne power supply circuit is disconnected, and the electromagnetic induction magnetic ring structure is separated. S75. Realignment and Power Collection Operation: Repeat the steps of standard prototype image matching, precise control of UAV power collection attitude, and adaptive induction power collection of wire in sequence to complete the repositioning and alignment of the target wire for power collection. S76. Multi-condition power supply termination judgment: The closed-loop verification and power supply process are executed cyclically, and three termination conditions are monitored simultaneously: full battery charge, manual return command, and obstacle in front of the line. If any one of these conditions is triggered, power supply will be terminated, the magnetic ring will disengage and the line will return. In addition, a low battery protection is set. When the battery charge is lower than the safe return threshold, the line will be forcibly terminated and the line will return first.

[0026] Using the above technical solution, continuous image acquisition of the conductor is performed throughout the power supply process. Local areas of interest (ROIs) are captured for recognition and computation, significantly reducing data volume and adapting to low-computing-power airborne hardware, enabling lightweight real-time detection. Matching thresholds are adaptively adjusted based on conductor electrical parameters, with tiered warning and safety benchmarks, allowing real-time detection of UAV position drift. Upon deviation, the power supply circuit is automatically cut off and the magnetic ring is separated to avoid insulation breakdown and short-circuit risks, followed by automatic repositioning and power supply recovery. Simultaneously, three termination conditions—full charge, return to base, and obstacle—are monitored in real-time, with a low-charge priority return-to-base protection feature that stops operation immediately upon fulfilling these conditions. A visual linkage closed-loop safety mechanism is constructed, enabling full autonomous monitoring and automatic anomaly protection, significantly improving the safety and autonomy of field power supply operations.

[0027] Furthermore, the identification method in S74 where the image matching state does not meet the security judgment criteria specifically includes: S741. Real-time matching result acquisition: Retrieve the target traverse image matching and fitting status data updated in real time by S73, and obtain the matching correspondence between the current frame image and the standard prototype image; S742, Two-level benchmark comparison and discrimination: The real-time image matching status is compared with the system's preset safety judgment benchmark item by item to verify the matching status of the structural features and positional features of the current target conductor; S743, Abnormal State Identification: When the real-time image matching status fails to meet the preset safety judgment benchmark, the image feature correspondence shifts, or the target wire identification features mismatch, the current image matching status is determined to not meet the safety judgment benchmark, and an abnormal identification signal is generated.

[0028] By employing the above technical solution, real-time image matching data of the conductor is retrieved to accurately obtain the conductor's position and structural matching degree. This data is then compared and verified item by item with two levels of safety thresholds, simultaneously checking the conductor's structural and positional characteristics. In the event of feature deviation or identification mismatch, aircraft drift can be quickly determined and an abnormal signal output, providing a reliable triggering basis for emergency demagnetization protection. Anomalies are identified based on quantitative matching results, avoiding false alarms or missed detections caused by interference from single frames, thus improving the sensitivity of position deviation identification. As the core of the power supply safety closed loop, it accurately identifies hazardous operating conditions, ensuring that the magnetic ring and conductor always maintain a safe distance, eliminating the risk of insulation breakdown and equipment short circuits from a visual monitoring perspective, and improving the safety redundancy of UAV-based power supply operations.

[0029] Compared with the prior art, the present invention has the following beneficial effects: 1. Significantly improved target recognition capabilities in complex environments, adaptable to low-computing-power airborne hardware: This invention employs illumination correction and dynamic thresholding of mesh texture gradients to filter images, combined with multi-scale hierarchical feature extraction and lightweight convolutional neural networks, to eliminate invalid images such as clouds, strong light, and vegetation, reducing the computational burden on airborne equipment. Relying on relative entropy image matching to quantify the similarity of conductor features, combined with multi-frame cross-verification, it solves the problems of low accuracy in field recognition, high computational cost, and easy misjudgment of hardware for power extraction in traditional solutions, thus avoiding the risks of collisions and power extraction failures from the source.

[0030] 2. Achieve adaptive adjustment of inductive power supply under all working conditions, balancing charging efficiency and equipment safety: Based on image recognition, the conductor voltage and cross-sectional electrical parameters are automatically labeled, and the number of turns of the magnetic ring coil is dynamically switched by linkage parameter matrix. The optimal winding position is matched for high voltage, low voltage, and different split conductors, overcoming the defects of high voltage overvoltage damage to the circuit and insufficient low voltage power generation caused by the fixed number of coil turns in traditional systems. After rectification, filtering, and voltage stabilization, the output of stable charging power is achieved, and it can autonomously adapt to various ultra-high voltage transmission conditions.

[0031] 3. Construct a visual linkage closed-loop safety protection system with full autonomous control: During the power supply phase, the system continuously captures and cyclically identifies the ROI of the conductor, sets two-level linkage safety thresholds to monitor the drone's position drift in real time, automatically fine-tunes the attitude if the matching degree is lower than the warning threshold, and immediately disconnects the power supply circuit and separates the magnetic ring if it is lower than the safety threshold, eliminating the risk of insulation breakdown and short circuit. It also integrates multi-condition shutdown logic for full charge, obstacles, low battery, and manual return, and sets up low battery priority return protection to achieve automatic handling of anomalies and autonomous start and stop of operations, making up for the shortcomings of traditional systems that lack real-time visual monitoring and tiered emergency protection.

[0032] 4. High degree of autonomous inspection, completely solving the endurance bottleneck: Through PID visual servo precise control of the drone's attitude, the magnetic ring and the conductor are safely and precisely coupled. The deep integration of machine vision and adaptive induction power supply enables the drone to autonomously replenish power online during long-distance cruise in ultra-high voltage corridors. There is no need for frequent return to change batteries or to deploy ground charging piles in the field. This greatly expands the inspection operation radius, reduces line maintenance costs, and improves the level of automated inspection of ultra-high voltage transmission lines. Attached Figure Description

[0033] The present invention can be further illustrated by the non-limiting embodiments given in the accompanying drawings; Figure 1 This is a flowchart illustrating the overall process of the airborne visual recognition adaptive inductive power harvesting method for ultra-high voltage transmission lines according to the present invention. Figure 2 This is a schematic diagram of the UAV onboard collaborative hardware and attitude control structure of the present invention; Figure 3 This is a flowchart of the cruise image acquisition and image identification subdivision processing of the present invention S1; Figure 4 This is a flowchart of the lightweight classification network and electrical attribute annotation for the S3 power transmission components of the present invention; Figure 5 This is a flowchart of the nested subdivision process for S4 standard prototype image matching and S43 relative entropy feature matching of the present invention. Figure 6 This is a flowchart of the S6 conductor adaptation adaptive inductive power harvesting hierarchical control process of the present invention; Figure 7 This is a complete flowchart of the S7 image-linked power-supply closed-loop verification process of the present invention. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0036] like Figure 1-7 As shown: This embodiment uses a quadcopter inspection drone equipped with an integrated hardware module. The complete set of hardware consists of three main parts: a visible light three-axis image-stabilized depth-of-field camera (A), an openable electromagnetic induction power-harvesting module (B), and an airborne embedded image processing unit. Auxiliary components include an airborne positioning module, an inertial measurement unit, an electronic switching switch, a rectifier, filter, voltage regulator, and power conditioning unit, and an airborne high-capacity lithium battery. All hardware relies on the airborne bus for data communication and command linkage. The entire recognition algorithm, standard prototype image library, and power-harvesting parameter matrix are all pre-stored in the local solid-state storage space of the airborne image processing unit, providing the underlying hardware support for the entire autonomous induction power-harvesting operation.

[0037] like Figure 2 As shown in the diagram, the entire airborne collaborative hardware and attitude control structure is clearly illustrated: a visible light depth-sensing camera is responsible for acquiring real-world images, while the positioning device and inertial measurement unit simultaneously acquire the aircraft's position and attitude data. These two types of data, along with the target conductor pixel coordinates and relative distance output by image recognition, are sent to a three-axis deviation calculation unit. The unit calculates pitch, roll, and yaw deviation signals, which are then transmitted to a proportional-integral-derivative (PID) closed-loop controller. The controller outputs commands to adjust the UAV's overall attitude and simultaneously controls the opening and closing magnetic ring to maintain a safe insulation distance between the magnetic ring and single or four-split ultra-high voltage conductors. The entire operation process is carried out sequentially: image acquisition, feature extraction, component classification, image matching, attitude adjustment, inductive power supply, and safety verification. The entire process relies on visual recognition to perceive the environment, electromagnetic induction for energy replenishment, and image closed-loop verification to ensure equipment safety. It can complete all operational stages—autonomous navigation of the ultra-high voltage transmission corridor, conductor identification, magnetic ring alignment, online power replenishment, and full-process safety monitoring—without real-time human intervention.

[0038] like Figure 1 , Figure 3 As shown: After receiving the UAV's command for a cruise mission along the ultra-high voltage transmission corridor from the ground station, the airborne flight control system initiates an autonomous cruise program, flying at a constant speed along the transmission corridor according to the preset route. The visible light depth-sensing camera activates continuous frame acquisition mode, continuously capturing real-world images along the corridor at a frequency of 25 frames per second. The captured images completely cover the conductors, spacers, tension fittings, mountains, forests, clouds, fog, and areas with strong light reflection. All original real-world images are transmitted in real time to the airborne image processing unit, which then initiates the entire image recognition processing workflow, employing methods adapted to varying outdoor lighting conditions, relying on lighting preprocessing, and using grid gradients. An adaptive dynamic brightness threshold segmentation algorithm with a four-level linkage logic of statistics, environmental linkage dynamic threshold generation, and grid region discrimination and filtering distinguishes power transmission component targets from various interference backgrounds. This algorithm abandons the traditional fixed global threshold mode and can automatically generate exclusive segmentation brightness thresholds based on the real-time brightness distribution of a single frame and the on-site fog and reflection conditions. It accurately divides the high-gradient target area of ​​metal power transmission components into low-gradient interference backgrounds of mountains, forests, clouds and fog, and simultaneously removes invalid blank images that do not contain line components. Finally, it outputs a standardized set of valid target images, providing clean image materials without redundancy and interference for subsequent multi-level image feature extraction processes.

[0039] After all the original real-scene images are transmitted, the onboard image processing unit will first carry out unified illumination correction processing to complete the pre-processing of all real-scene images. For each original real-scene image transmitted back by the depth-sensing camera, the device will first perform a global brightness equalization correction operation to adapt to three typical outdoor lighting scenarios: strong midday light reflection, dim morning fog, and shadow obstruction in the mountains and forests. The segmented grayscale stretching algorithm will be used to smooth out the difference in brightness between the images.

[0040] The complete execution logic of the global brightness equalization correction operation is as follows: First, the color real-world image output by the camera is uniformly converted into a single-channel grayscale image. The brightness and darkness interference brought by the color channel are removed, and only the pixel grayscale brightness and darkness information is retained as the basis for correction. Then, the grayscale distribution of all pixels in the entire grayscale image is statistically analyzed. Based on the differences in pixel grayscale brightness and darkness, three independent image areas are automatically divided: strong light reflection, light fog and darkness, and mountain shadow. Brightness adjustment is carried out separately for the three types of areas. The overall brightness of the image is reduced in the strong light reflection area to weaken the grayscale change caused by overexposed white pixels. The overall brightness of the image is increased in the light fog and darkness area to highlight the subtle grayscale changes of components in the dark. The light and dark boundary in the mountain shadow area is smoothed to eliminate large-area dark spots. After the brightness adjustment of each area is completed, grayscale fusion processing is performed on all pixels in the image to eliminate the light and dark boundary discontinuity between different correction areas, so that the grayscale transition of the entire image is natural and uniform, thus completing the global brightness equalization correction.

[0041] The accompanying segmented grayscale stretching algorithm independently performs differentiated grayscale adjustments for the three types of lighting zones to smooth out the differences in brightness within each zone. The overall operation process is as follows: For the strong light and reflective zone, first distinguish the grayscale boundaries between overexposed bright pixels and normal pixels of the power transmission components within the zone, compress the grayscale variation span of the bright pixels themselves, weaken the drastic grayscale fluctuations caused by strong light reflection, and at the same time completely preserve the edge grayscale contours of the conductors and fittings themselves, so as not to lose the component recognition features due to darkening the image. For the hazy and dim zone, distinguish between low-brightness dark pixels and faint components within the zone. The grayscale boundaries of pixels are widened to broaden the grayscale variation range corresponding to the pixels of the components, amplifying the subtle grayscale differences of the power transmission components in a light fog environment. This makes the grayscale changes of the originally blurred component outlines clearer, without over-brightening the image and causing excessive noise in the background of mountains and vegetation. For the shadow occlusion zones in the mountains and forests, the boundary pixels between the shadow and illuminated areas are identified, smoothing the grayscale variation of pixels on both sides of the boundary, reducing the grayscale difference between the shadow and bright areas, while preserving the inherent texture grayscale differences of the power transmission components within the shadows, preventing the components from being completely covered by shadows and causing target loss. After all three types of grayscale stretching are completed, a uniform grayscale smoothing transition is applied to all pixels in the image to eliminate the light and dark discontinuities caused by stretching different zones, resulting in a corrected image with uniform light and dark distribution and clear grayscale features of the components.

[0042] After the correction operation is complete, the algorithm automatically divides the entire corrected image into equal grid sub-regions, independently calculating the texture gradient value for each grid. The gradient value represents the intensity of grayscale changes within the grid, and is accompanied by gradient shape length and width constraints to assist in the judgment. If the length and width ratio of a continuous gradient region within the grid falls within the standard range of the power transmission component, it is marked as a suspected target area. If the length and width ratio deviates from the standard range of the component, it is directly identified as an irrelevant background grid such as vegetation or mountains. Relying on this preliminary division, a rough segmentation of power transmission components and complex backgrounds can be initially completed, significantly reducing the computational workload of the subsequent threshold discrimination stage.

[0043] After all mesh texture gradient values ​​are calculated, the system generates a unique dynamic discrimination threshold based on all mesh gradient data to accurately filter valid images containing power transmission components. The airborne unit first aggregates the texture gradient values ​​calculated from all mesh sub-regions and performs gradient data preprocessing: extracting the maximum, minimum, and overall arithmetic mean of all gradient values. Using the mean as a benchmark, a gradient floating interval with upper and lower boundary constraints is defined. This floating interval uses the global gradient arithmetic mean as its central benchmark. The system pre-stores a fixed base offset amplitude, generates the upper boundary of the interval by superimposing the base offset on the central benchmark, and generates the lower boundary of the interval by subtracting the base offset from the central benchmark. This completes the basic boundary division of the floating interval. Gradient values ​​distributed within the interval are considered normal and valid gradient data, while gradient values ​​exceeding the upper and lower boundaries are directly identified as abnormal noise data. The interval boundaries are not constant values; the interval values ​​are adaptively matched and adjusted based on the component size benchmark corresponding to the current corridor voltage level and real-time collected illumination and cloud / fog environment correction coefficients. The adaptive matching adjustment is executed in two progressive levels: the first level corrects the base offset amplitude according to the current corridor voltage level, reads the preset voltage parameters of the line to match the corresponding component standard texture gradient reference, and synchronously scales the base offset amplitude. The base offset amplitude is amplified in the case of high voltage large cross-section conductors and narrowed in the case of low voltage small cross-section conductors. The second level combines real-time lighting and cloud / fog reflection conditions to further scale the offset amplitude after the first level adjustment. Under strong light and reflection conditions, the vertical offset amplitude is reduced simultaneously, while under light fog and dim conditions, the vertical offset amplitude is amplified simultaneously. Finally, the formal boundary of the floating range is determined. During the matching adjustment, the standard texture gradient reference range of the conductors and fittings is locked based on the voltage level of the transmission corridor. Then, the vertical offset amplitude of the floating range is scaled in combination with real-time weather and light reflection intensity. Under strong light and reflection conditions, the range is narrowed to filter abrupt noise, while under light fog and dim conditions, the range is widened to retain weak component gradient information. All gradient values ​​that exceed the floating range are judged as abnormal gradient noise caused by reflection and shadow and are directly removed.

[0044] After removing noise and retaining all valid gradient data, the system performs a weighted average calculation on the remaining gradient values. The entire weighted average calculation follows a fixed step-by-step calculation logic: The first step is to define two types of weighting factors: distance weighting factor and visibility weighting factor. The distance weighting factor is assigned according to the distance of the grid relative to the camera. Grids with clear images at close range are assigned a high distance weight, while grids with blurred images at a distance are assigned a low distance weight. The visibility weighting factor is assigned according to the real-time weather conditions. The basic visibility weight is used on normal sunny days, while the visibility weight of the image edge grids is increased in dense fog, rain, and low visibility conditions. The second step is to multiply the effective gradient value of a single grid with the distance weight factor and visibility weight factor corresponding to that grid in sequence to obtain the weighted gradient value of the single grid. The third step is to sum the weighted gradients of all grid cells to obtain the total weighted gradient. The fourth step is to sum the distance weight factor and visibility weight factor of all valid grids to obtain the total weight summation value; The fifth step is to divide the sum of weighted gradients by the sum of total weights to finally output the weighted average gradient.

[0045] The weighted average calculation assigns differentiated weights based on the distance of the grid from the depth-sensing camera and the real-time weather conditions. Grids that are closer and have clearer images of components are assigned higher calculation weights. In dense fog, rain, and low visibility weather, the weights of edge grids are increased accordingly. After weighted summation, the result is divided by the sum of the weights of all effective grids to obtain the weighted average. The calculation result is then superimposed with corridor-specific environment correction coefficients adapted to the ultra-high voltage transmission scenario and related to voltage levels and light intensity to complete the illumination compensation correction. The corridor-specific environment correction coefficients are a pre-calibrated set of segmented parameters, including two independent parameters: voltage level sub-coefficients and illumination level sub-coefficients. The voltage level sub-coefficients are fixed values ​​based on the line voltage level, while the illumination level sub-coefficients are divided into multiple fixed values ​​based on the global grayscale mean of the image.

[0046] Illumination compensation correction follows a fixed process: first, the current line voltage level is read and the corresponding voltage level sub-coefficient is retrieved; then, the average grayscale value of the entire image pixels is statistically analyzed to match the illumination level sub-coefficients of the corresponding brightness level; finally, the gradient weighted average is multiplied by the two types of sub-coefficients in turn to complete the numerical compensation correction. Illumination compensation correction determines the current brightness level based on the overall grayscale average value of the image. In dim scenes, the gradient benchmark is amplified and the weight of component textures is increased by correcting the coefficients; in strong light and reflective scenes, the gradient extrema are compressed and the background interference gradient is suppressed by correcting the coefficients. Finally, a unique dynamic discrimination threshold is generated that is adapted to the current shooting lighting and environmental scene.

[0047] The dedicated dynamic discrimination threshold is a single gradient judgment value obtained by compensating the gradient weighted average with an environmental correction coefficient. This value is generated independently for each image. Each image corresponds to a dedicated judgment value under different lighting conditions, shooting distances, and corridor conditions. There is no unified fixed judgment standard for the entire scene. The dedicated dynamic discrimination threshold is a gradient judgment benchmark generated independently for each frame image and matched one-to-one with the current scene lighting, shooting distance, and corridor conditions. Unlike the universal fixed threshold for the entire scene, this dynamic discrimination threshold has a full-process linkage function. It will be synchronously transmitted to the subsequent power supply closed-loop verification module to realize the linkage correction of two-level safety judgment benchmarks. In rainy or dim low-light environments, the image matching judgment standard is automatically lowered to avoid the problem of missed detection of wire targets.

[0048] Once the dedicated dynamic discrimination threshold is generated, the device compares the real-time texture gradient value of each grid sub-region with the dynamic discrimination threshold one by one, filtering out continuous grid connected regions with gradient values ​​higher than the threshold. A second verification is then performed based on the region's length and width shape constraints. Connected regions whose length and width proportions match the shape characteristics of conductors and fittings are officially determined as valid target areas for transmission components. The image processing unit automatically crops all background areas within the image that do not contain valid targets, completely eliminating blank interference images that do not contain conductors, spacers, or tension fittings. A standardized set of valid target images is batch-output and automatically cached in the onboard image-specific cache partition. After the image materials are processed, the data is completely sent to the multi-scale layered image extraction process.

[0049] After the airborne image cache partition stores the standardized and valid target image set, the airborne image processing unit reads all the image data and performs a three-layer progressive multi-scale hierarchical image extraction operation on each valid image, extracting shallow contour features, mid-layer texture features, and deep semantic features respectively. Finally, the three layers of features are spliced ​​and integrated to generate a multi-dimensional feature vector library dedicated to power transmission components, providing complete feature input materials for lightweight convolutional classification networks.

[0050] The entire layered extraction process is carried out simultaneously in three layers. The first layer is shallow grayscale contour extraction, which uses an edge detection algorithm to traverse the entire effective image, capture the grayscale abrupt change points at the edges of the ultra-high voltage conductor, and fit the discrete edge points into continuous and complete conductor contour lines. It records basic contour information such as the direction, width, and number of conductor lines, forming the first layer of shallow contour feature data. The second layer is mid-layer texture local feature extraction, which uses a scale-invariant feature algorithm adapted to the near and far imaging scenarios of power transmission fittings to process the internal spacer bars and tension fitting metal structure areas of the image separately. The complete execution flow of this algorithm is as follows: First, a multi-scale Gaussian difference pyramid is constructed for the metal component area to simulate the target scaling changes caused by the drone's near and far shooting. The extreme key points are located by traversing each layer of the pyramid image. After removing invalid key points with low contrast at the edges, a principal direction is assigned around the effective key points to achieve rotation invariance. Then, a fixed-dimensional feature descriptor is generated based on the grayscale gradient of the key point's neighborhood. The entire process is unaffected by changes in shooting distance, shooting angle, and lighting conditions, accurately extracting local texture features such as the surface texture of the fittings, bolt points, and metal splicing gaps, generating a mid-layer texture feature dataset. The third layer is deep semantic feature extraction, which uses a bidirectional encoding algorithm to analyze the global structural information of the image. This algorithm consists of two main execution steps: forward spatial encoding and reverse semantic decoding. Forward encoding compresses the three layers of pre-features of global pixels, contours and textures from bottom to top to construct a global spatial correlation feature vector. Reverse encoding traces back the position, arrangement and distance relationship of components from top to bottom. The feature mapping and fusion are completed in both directions simultaneously, so as to automatically identify global semantic information such as the split arrangement of the conductor, the vertical erection height of the conductor and the span of the conductor, and output deep structural semantic features.

[0051] After all three layers of image features are extracted, the airborne unit initiates a feature vector concatenation and fusion operation. This process concatenates and integrates the shallow contour grayscale features, the mid-layer hardware texture local features, and the deep conductor structure semantic features in a fixed-dimensional order, transforming them into standardized multi-dimensional floating-point feature vectors. Each image corresponds to a set of independent multi-dimensional feature vectors. The feature vectors corresponding to all valid images are collected and stored, constructing a dedicated multi-dimensional feature vector library for power transmission components. Once the vector library data is organized, it is completely transmitted to the power transmission component image classification module as the input data source for network computation.

[0052] like Figure 4 As shown: After the complete multi-dimensional feature vector library of power transmission components is delivered to the classification module, the airborne image processing unit will retrieve all vector data and start the entire process of lightweight convolutional network classification. The process is divided into three sub-processes: network construction, intelligent feature classification, and automatic labeling of electrical attributes. This completes the classification of power transmission components into four categories and automatically binds electrical parameter labels, enabling automatic diversion of images of fittings with no power extraction value and directional downward flow of images of conductors that can be used for power extraction.

[0053] The first step in classification work is to build a lightweight improved convolutional neural network. This involves embedding and lightweighting a standard basic convolutional neural network to adapt to the low-computing-power hardware environment of UAVs. The first step involves removing deep, redundant convolutional and fully connected layers to reduce overall floating-point computation. The second step involves embedding a channel attention mechanism module in the middle layers of the network to automatically amplify the weights of key features such as wires and hardware, while weakening irrelevant background features such as vegetation and sky. The third step involves adding a multi-scale feature fusion module to unify and fuse the features extracted from the three layers in the previous stage, enhancing the ability to focus on and recognize multi-dimensional features. After the modification, the lightweight convolutional network occupies only 35% of the memory of the original network, perfectly adapting to the limited computing power environment of the embedded image processing unit on UAVs. After the modification, the network enters a ready-to-operate state.

[0054] After the lightweight network is built and debugged, the system will batch input all the multi-dimensional feature vector libraries generated in the previous stage into the lightweight convolutional classification network to start the feature input and intelligent classification process. Internally, the network performs iterative calculations on each set of feature vectors based on preset classification weight parameters. The complete execution steps of this iterative calculation are as follows: First, input a single set of multi-dimensional contour, texture, and semantic fusion feature vectors. Then, call the offline pre-trained and solidified initial weight matrix of the network to complete the linear mapping of features and output preliminary classification prediction values. The initial weight matrix of the network is a multi-level set of weight parameters derived from offline training and convergence based on massive ultra-high voltage transmission component image samples. It includes a complete set of weight sub-matrices and matching bias vectors corresponding to the input convolutional layer, attention module, multi-scale feature fusion module, and terminal classification output layer. It is pre-burned and stored in the local storage space of the onboard image processing unit, eliminating the need for random initialization on-site by the drone. It can directly establish a linear numerical mapping relationship between the input fused feature vectors and the predicted scores of each layer of neurons in the network and the four types of transmission components, quickly outputting a single set of features. The first step involves the initial classification prediction of the corresponding components. The second step calculates the loss error between the predicted classification result and the standard component label. This error value is then backpropagated layer by layer to fine-tune the classification weight parameters of each layer of the network. Simultaneously, the weight matrix and bias vector values ​​at each level are corrected to reduce the deviation between the predicted value and the actual component label. The third step iterates through all sample feature vectors, repeating the mapping, error calculation, and weight correction processes of the first two steps. Each complete iteration is counted as one round of iteration. Multiple rounds of iterative iteration continuously converge the loss error and optimize the weight parameters until the error value falls below a preset convergence threshold, at which point the iteration terminates. This allows the network to autonomously learn to distinguish the feature differences of four types of targets: a single conductor with a continuous, long, strip-shaped outline without metal attachments; a four-split conductor with the semantic features of four parallel, equidistant conductors; a spacer bar with multiple sets of conductors clamping metal texture features; and tension fittings with complex metal structures anchored at the line ends. After the iterative calculation, the network automatically divides all images into four independent image subsets: single conductor image subset, four-split conductor image subset, spacer bar image subset, and tension fitting image subset. The system has built-in automatic diversion logic. If the current classification output subset is spacer bars and tension fittings, which are two types of targets that do not have the conditions for power extraction, it will directly jump back to the cruise image acquisition process. The drone will continue to cruise forward to capture new real-world images of the corridor. Only the image subsets of single wires and four-split wires, which are two types of targets that can carry out inductive power extraction, will be sent down to the standard prototype image matching process.

[0055] Once the subset of conductor images usable for power extraction is segmented, the system simultaneously performs automatic electrical attribute labeling and subdivision, binding standardized electrical attribute tags to each conductor image and establishing a mapping database linking image features with electrical parameters. The electrical tag data source is collected synchronously from two sources: the first source is static parameters from the transmission corridor ledger pre-stored onboard, including the line design voltage level and standard conductor cross-sectional specifications. The second source is dynamic semantic data extracted from the previous stage's deep bidirectional coding, including the real-time conductor erection height and conductor splitting structure type. Both data sources are synchronously input into the mapping database for matching, automatically labeling each frame of single and four-split conductor image subsets with three core electrical attribute tags: line voltage level, conductor cross-sectional specifications, and conductor erection height. After all labeling is completed, a categorized conductor image dataset with complete electrical attribute labels is generated. This dataset is transmitted to the standard prototype image matching process for accurate standard image retrieval and cached in the parameter retrieval partition, providing unified electrical parameter data support for subsequent adaptive power extraction winding switching and dynamic correction of two-level safety judgment benchmarks.

[0056] like Figure 5 As shown: After the data set of conductor images with complete electrical annotations is transmitted, the airborne unit will read all the image data, retrieve the standard prototype image library stored locally based on the electrical parameters bound to the images, and complete the entire set of operations of image matching, hierarchical screening, and multi-frame verification by relying on the relative entropy algorithm. Finally, the target conductor is locked and the spatial coordinate data is output, providing a visual quantitative basis for the closed-loop control of UAV attitude. The entire process is divided into five sub-steps executed in sequence.

[0057] When the matching process officially begins, the system first performs a precise matching prototype image retrieval procedure. The onboard local solid-state storage partition contains a multi-level archived standard prototype image library for ultra-high voltage power transmission. The image library is categorized and archived according to three dimensions: transmission component type, voltage level, and conductor cross-sectional specifications. The system reads the electrical attribute tags carried by the conductor image to be matched, simultaneously identifies the component type, and accurately retrieves the standard prototype image that perfectly matches the on-site operating parameters according to the three-level matching logic of component type, voltage level, and conductor cross-section. This achieves bidirectional parameter matching between the real-world image to be identified and the standard reference image, avoiding the similarity calculation distortion problem caused by cross-category matching of conductor images of different specifications.

[0058] After the standard prototype image for the adapted working condition is retrieved, the device will simultaneously perform unified image quality preprocessing on both the real-scene guide image to be identified and the retrieved standard prototype image to establish a standardized matching calculation basis. The unified processing flow consists of three steps: First, image impurity filtering, removing random interference pixels such as birds and floating debris from within the image. Second, global image quality correction, unifying the contrast, brightness, and resolution of the two images. Third, image scale regularization, correcting perspective distortion in the real-scene image to unify the image presentation scale between the real-scene image and the standard prototype image, eliminating matching errors caused by differences in shooting distance and shooting angle. After the two images are regularized, the feature matching calculation stage will begin.

[0059] Once the image to be identified and the standard prototype image are both normalized, the system will initiate an image feature matching subdivision process based on the relative entropy algorithm. This process calculates the degree of image matching fit by quantifying the differences in feature distribution. The first step involves extracting image feature information. The normalized real-world image to be identified and the standard prototype image are read separately, and three core feature distribution information—wire texture, edge contour, and splitting arrangement structure—are extracted from both images, generating two independent feature distribution probability matrices. The second step involves relative entropy matching calculation. The relative entropy algorithm is used to perform a full-domain comparison of the two feature distribution matrices, quantifying the amount of overlap and difference information between the two sets of feature distributions. Lower values ​​indicate a higher degree of image feature overlap and a better matching fit. The third step involves determining the matching state. Based on the overlap and difference data output by the relative entropy calculation, a standardized matching fit score is output. This score is directly sent to the matching result grading and screening process as the core judgment criterion.

[0060] After the image matching score is calculated, the system will perform hierarchical screening and judgment of the matching results according to the preset two-level image matching judgment benchmarks, thus completing the hierarchical identification. If the matching score of a single frame image is only slightly lower than the first-level benchmark, it is judged to be a slight deviation caused by single-frame shooting shake or instantaneous lighting interference, and it will automatically enter the multi-frame cross-verification process. If the matching scores of five or more consecutive frames are all lower than the first-level benchmark, it is judged that the current image of the conductor is missing features and the recognition is invalid. The images are marked as invalid images, and the entire system is returned to the cruise image acquisition process to re-capture the actual corridor scene image. If the image matching score is higher than the first-level judgment benchmark, the conductor image of that frame is designated as a valid target image, and the corresponding ultra-high voltage conductor in the image is locked as a candidate target for inductive power extraction. Then, the multi-frame verification and locking process is entered.

[0061] After the single-frame valid conductor image screening is completed, the depth-sensing camera will continuously capture multiple sets of real-time images of the same target conductor to carry out the multi-frame image verification and target locking process. Each set of images completely repeats the entire calculation process of prototype image retrieval, image regularization, feature matching, and result screening. Cross-verification and comprehensive weighted analysis are carried out on multiple sets of matching scores. The cross-verification involves first comparing and verifying the matching scores of adjacent frames and interval sampling frames one by one, and eliminating abnormal deviations caused by single-frame jitter, instantaneous reflection, and local cloud and fog interference. Simultaneously, the conductor contour, texture, and semantics between multiple frames are verified. The system assesses the spatial continuity of the target's position. If the target coordinates in a frame deviate from the coordinates of the preceding and following frames by more than a preset spatial threshold, the frame is deemed an interference failure and discarded. A comprehensive weighted assessment assigns differentiated weights to the valid multi-frame matching scores remaining after cross-verification. Images with closer shooting distances, uniform lighting, and intact, unobstructed conductors are assigned higher weights, while blurred or partially occluded distant frames are assigned lower weights. The scores of each frame are multiplied by their corresponding weights, summed, and then divided by the sum of the weights of all valid frames to obtain the comprehensive judgment score. This comprehensive calculation eliminates judgment errors caused by random interference in single frames. After all multi-frame comprehensive verifications meet the standards, the system permanently locks the conductor as the final inductive power-collecting target conductor and calculates and outputs two core spatial data points: the pixel coordinates of the target conductor within the camera's view and the three-dimensional relative distance between the drone's body and the conductor. These two sets of spatial data are transmitted in real-time to the drone's precise power-collecting attitude control module.

[0062] After the target traverse pixel coordinates and 3D relative distance data are transmitted to the three-axis deviation calculation unit, the device simultaneously receives two other types of real-time data: global coordinate data from the airborne positioning device and raw attitude data of the fuselage pitch, roll, and yaw recorded by the inertial measurement device. The device calculates the fuselage's three-axis deviation in real time using a 3D spatial coordinate transformation model. This model uses a three-level coordinate transformation framework, consisting of camera imaging pinhole projection transformation, fuselage coordinate system, and world global coordinate system, as its core computational framework.

[0063] The first step is to use the pixel coordinates of the target traverse image combined with the gimbal's intrinsic parameters to transform the image pixel coordinate system into the camera coordinate system, and then calculate the three-dimensional local coordinates of the traverse relative to the depth-sensing camera. The second step involves constructing a coordinate rotation matrix by combining the pitch, roll, and yaw attitude angles output by the inertial measurement unit, and transforming the camera's local coordinates to the UAV's body coordinate system. The third step involves integrating the global geodetic coordinates output by the airborne positioning equipment to complete translation matrix compensation, mapping the position of the conductor in the airframe coordinate system to the global world coordinate system of the power transmission corridor, and simultaneously comparing the ideal power supply standard spatial position of the conductor to complete the difference solution, ultimately calculating the three-axis deviation.

[0064] The three-axis deviations represent the pitch angle deviation, the left and right roll offset of the fuselage, and the horizontal yaw difference of the nose of the UAV, respectively. The three-axis deviation values ​​are synchronously transmitted to the onboard proportional-integral-derivative closed-loop controller.

[0065] After receiving the three-axis deviation values, the proportional-integral-derivative (PID) closed-loop controller outputs a continuous and smooth control signal according to the PID control algorithm. This algorithm performs parallel and independent calculations on the pitch, roll, and yaw deviations in three separate paths. The complete calculation steps for a single path are as follows: The first step is to collect the instantaneous deviation of the current single axis in real time, multiply the corresponding proportional coefficient of the matched axis to obtain the proportional adjustment component, and rely on the proportional component to quickly offset the real-time spatial offset. The second step involves continuously accumulating all instantaneous deviations within the cycle to obtain the cumulative steady-state deviation, which is then multiplied by the integral coefficient to obtain the integral adjustment component, thus eliminating the static residual alignment error caused by long-term cruise. The third step is to compare the current deviation with the historical deviation of the previous cycle to calculate the rate of deviation change, and combine the differential coefficient to generate the differential adjustment component to predict the trend of offset change and buffer servo shock and body vibration. The fourth step involves linearly superimposing the proportional, integral, and derivative components, along with the static compensation reference, to generate the original single-axis control signal. The fifth step involves upper and lower limit processing of the original signal to avoid the risk of large attitude changes, and finally outputting a single-channel control signal that is smooth and continuous without any changes.

[0066] The three-axis signals are synchronously sent to the three arm motors of the UAV and the electromagnetic induction power-taking module servo mechanism on the gimbal, simultaneously adjusting the overall flight attitude of the UAV and the spatial position of the opening and closing magnetic ring: adjusting the pitch angle to match the vertical height of the conductor, correcting the roll angle to align with the horizontal centerline of the conductor, and changing the yaw angle so that the opening of the magnetic ring faces directly below the conductor. After completing the initial alignment, the UAV needs to further determine the optimal power-taking point. Since the ultra-high voltage conductor is distributed as a catenary between the towers, the conductor tension is the lowest and the spatial position is the lowest at its lowest sag point, which is convenient for the UAV to approach and minimizes the mechanical stress on the conductor when the magnetic ring clamps it. At the same time, the distance between the conductor and the ground is the largest at this point, which is conducive to maintaining a safe clearance. Therefore, the airborne image processing unit uses the deep semantic features extracted in stage S2 (including the conductor erection height and sag shape) combined with the preset line span parameters to calculate the spatial coordinates of the lowest sag point of the current target conductor and guide the UAV to fly to this position. After reaching the sag point, the above attitude control process is repeated to ensure that the opening of the magnetic ring faces directly below the sag section of the conductor, ensuring that the subsequent clamping operation is stable and reliable.

[0067] After the attitude control process is completed, the airborne power control unit first issues a magnetic ring clamping command. The opening and closing magnetic ring slowly closes under servo drive. Its inner insulating pad first contacts the outer insulation layer of the conductor, and then a preset clamping force is applied through an elastic clamping mechanism to securely wrap the conductor around the center of the magnetic ring. During this process, the inner wall of the magnetic ring and the metal conductor of the conductor maintain the designed insulation distance. The clamping force is only used to fix the relative position and prevent displacement caused by flight disturbances, without damaging the conductor insulation layer or affecting the electromagnetic induction air gap. After clamping in place, the magnetic ring position is locked, and the subsequent adaptive induction power harvesting process begins.

[0068] like Figure 6 As shown: After clamping and locking, the airborne power supply control unit officially starts the entire adaptive inductive power supply process. Based on the electrical parameters of the wires obtained from the image classification and labeling process, the number of turns of the magnetic ring coil is dynamically switched to complete magnetic field induction power generation and multi-level power conditioning to stably charge the airborne energy storage lithium battery. The entire process is divided into four subdivided steps that are executed sequentially.

[0069] The first step in initiating the power extraction process is to retrieve power extraction parameters. The onboard unit reads the line voltage level and conductor cross-section specifications stored in the electrical attribute labeling stage, and retrieves the multi-dimensional power extraction parameter matrix pre-stored locally. This parameter matrix has full-process linkage characteristics. Each coil winding parameter within the matrix is ​​bound to the dynamic threshold of the previous image recognition and the two-level safety matching benchmark of the subsequent closed-loop verification. Conductors with different electrical parameters will automatically match the dedicated winding position, realizing the coordinated linkage of image recognition, safety verification, and power extraction parameters.

[0070] Once the parameter matrix corresponding to the conductor's operating condition is retrieved, the equipment will perform a three-layer subdivision of coil winding switching control, completing the adaptive configuration of the number of coil turns connected inside the magnetic ring. The first layer involves matching and comparing operating condition parameters, comparing the current conductor voltage and cross-sectional parameters with each operating condition level within the parameter matrix. The voltage level is divided into high and low voltage levels, using 220kV as the common boundary for ultra-high voltage transmission lines. Voltage levels ≥220kV are classified as high-voltage conditions, corresponding to large-section transmission conductors; voltage levels <220kV are classified as low-voltage conditions, corresponding to small-section transmission conductors. The system automatically matches the coil winding operating level to the current magnetic field strength. The second layer executes winding on / off switching. The onboard electronic switching switch receives the level command and selectively connects or disconnects multiple independent coil windings inside the magnetic ring. High-voltage, large-section conductors automatically reduce the number of coil turns connected to prevent excessive induced voltage from damaging the onboard circuit; low-voltage, small-section conductors automatically increase the number of coil turns to enhance induced power generation. The third layer involves locking the winding status. After the winding switching is completed, the electronic switching switch has a built-in latch to lock the current conduction state, which isolates the problem of incorrect gear switching caused by airflow vibration and slight fuselage deviation. The magnetic ring coil connection specifications are adapted to the current power transmission conditions.

[0071] Once the coil windings inside the magnetic ring are locked in position, the flight control system issues a command to fully close the magnetic ring and lock the clamping force. The system then initiates the alternating magnetic field induction power generation process. The servo mechanism drives the two halves of the opening and closing magnetic ring to fully close and lock the clamping force. The inner insulating pad and the wire insulation layer are tightly bonded, ensuring the wire remains firmly in the center of the magnetic ring. The multi-turn coil inside the magnetic ring continuously cuts the alternating electromagnetic field generated by the current flowing through the high-voltage wire, generating alternating induced alternating current based on the principle of electromagnetic induction. This induced alternating current is transmitted to the downstream power conditioning unit through the conductive lead at the bottom of the magnetic ring.

[0072] After the alternating induced AC power is generated and delivered to the conditioning unit, the equipment performs power conditioning and charging output processes. The induced AC power is sequentially connected in series to a three-stage power processing unit: the first-stage rectifier unit converts the alternating AC power into unidirectional pulsating DC power; the second-stage filter unit filters out high-frequency noise and voltage fluctuation spikes within the DC voltage; and the third-stage voltage regulator unit stabilizes the fluctuating DC power into a standard constant voltage suitable for charging the drone's lithium battery. The stable power after three stages of conditioning is continuously output to the onboard energy storage battery charging interface, uninterruptedly replenishing the lithium battery. Simultaneously with the power extraction operation, an image-linked power extraction closed-loop verification process is initiated, monitoring the magnetic ring alignment safety status in real time throughout the process.

[0073] like Figure 7 As shown: After the drone officially begins online power collection, the visible light depth-of-field camera will continuously capture real-time images of the target conductor. The system will activate a lightweight visual closed-loop safety verification mechanism, relying on local key areas for cyclical identification to reduce equipment computing power consumption. Based on two-level safety judgment benchmarks, it will handle aircraft drift and equipment abnormality issues in a graded manner, and simultaneously monitor four types of shutdown termination conditions in real time. The entire closed-loop process includes six subdivided steps executed cyclically.

[0074] During the continuous operation of the verification process, the depth-sensing camera maintains a high-speed acquisition mode of 25 frames per second, continuously capturing local images of the coupling position between the magnetic ring and the wire. The captured raw real-time images are transmitted in real time to the onboard image processing unit, serving as the continuous data source for the closed-loop recognition work.

[0075] After real-time images of the power transmission line are continuously fed into the image processing unit, in order to reduce the computational burden on the onboard equipment, the algorithm automatically extracts only key local areas within the image that contain only the power transmission line and magnetic ring, discarding irrelevant areas such as the sky and distant mountains. It then performs a lightweight, iterative recognition process for these key local areas. The cropped images of these key local areas are then iteratively replicated to complete the entire processing flow, including image discrimination, multi-scale layered image extraction, power transmission component image classification, and standard prototype image matching. Throughout the process, only small, localized images are processed, reducing computational power consumption by more than 70%. Each iteration outputs a real-time updated image matching score.

[0076] After the local image matching operation is completed and a new matching score is generated, the system continuously refreshes the current conductor matching fit score, performs a real-time matching status update process, and simultaneously retrieves the conductor voltage and cross-sectional electrical parameters from the image classification annotation. It also adaptively adjusts the two sets of judgment criteria: the primary warning benchmark and the secondary safety benchmark. The judgment criteria for high-voltage, large-section conductors are appropriately raised to improve alignment accuracy requirements, while the judgment criteria for low-voltage conductors are moderately relaxed to adapt to operating scenarios with weak magnetic fields and slight image jitter. These two sets of dynamic judgment criteria will serve as the quantitative basis for subsequent anomaly detection.

[0077] Once the two-level dynamic judgment benchmarks are updated, the system automatically retrieves the real-time updated image matching and fit quantification scores, reads the matching correspondence between the current frame's conductor structure and position and the standard prototype image, and initiates the safety benchmark identification subdivision process. Relying on three layers of discrimination logic, it accurately identifies fuselage drift anomalies. The device compares the real-time matching score sequentially with the first-level warning benchmark and the second-level safety benchmark, simultaneously verifying two core features: the overall conductor structure matching degree and the magnetic ring's position matching degree relative to the conductor position. After the comparison, the anomaly status is determined. If the real-time matching score is lower than the first-level warning benchmark but higher than the second-level safety benchmark, it is determined to be a minor fuselage drift. If the score is lower than the second-level safety benchmark, it is determined that the conductor image features have significantly shifted and the target identification is completely mismatched, generating a high-priority anomaly identification signal and triggering the emergency demagnetization protection logic.

[0078] Once the system receives a high-priority anomaly identification signal, it will immediately initiate emergency response procedures for abnormal operating conditions, simultaneously issuing two sets of hardware control commands: the first disconnects the main switch of the onboard power supply circuit, cutting off the power transmission path between the magnetic ring coil and the lithium battery; the second drives the magnetic ring servo mechanism to quickly open and close the magnetic ring, completely separating the magnetic ring from the high-voltage conductor, thus eliminating the safety hazards of insulation breakdown, short circuit, and equipment burnout. After all emergency protection actions are completed, the UAV will automatically enter the re-alignment power supply operation process, sequentially executing the standard prototype image matching, precise UAV attitude control, and adaptive inductive power supply procedures, completing conductor repositioning, secondary alignment of the magnetic ring, and resuming online power replenishment. If the matching score is in the middle range between the first-level warning benchmark and the second-level safety benchmark, only a slight drift is judged, and there is no need to disconnect the power supply circuit. The UAV's pitch, roll, and yaw are directly adjusted to correct the magnetic ring position, and then the system returns to the real-time image acquisition stage to continue the cyclical verification, without interrupting the charging process.

[0079] While the entire closed-loop verification process runs continuously, the system simultaneously performs a multi-condition power extraction termination determination procedure, monitoring four types of shutdown trigger conditions in real time. If any one of these conditions is identified, the entire power extraction operation is permanently terminated: the first condition is that the onboard energy storage battery voltage reaches the full-charge standard threshold; the second condition is that the ground maintenance station issues a remote manual return control command; the third condition is that the image recognition process detects the presence of poles, trees, or large floating obstacles in front of the conductor; and the fourth condition is that the remaining charge of the onboard battery drops to the low-charge warning threshold. After the power extraction termination command is issued, the magnetic ring fully opens and releases its clamping force, separating from the conductor. The power extraction circuit remains disconnected, and the drone returns to the ground take-off and landing point according to the preset route. The entire process of image recognition and sensing power extraction for ultra-high-voltage transmission lines by drone is then completed.

[0080] The entire implementation process begins with autonomous drone cruise image acquisition. Dynamic threshold segmentation filters out interference from complex field environments. Multi-scale hierarchical feature extraction, combined with a lightweight neural network, accurately classifies power transmission components, automatically distinguishing between single and quadruple conductors suitable for power extraction and spacers and tension fittings without power extraction value. Relative entropy image matching quantifies the spatial position of the conductors, and proportional-integral-derivative visual servo closed-loop control precisely controls the drone's attitude, locating the lowest point of the conductor sag and aligning the magnetic ring clamp. Based on automatically labeled conductor electrical parameters from image recognition, the number of turns in the magnetic ring coil is dynamically switched to adapt to different line magnetic field strengths, achieving stable inductive charging. The entire power extraction process incorporates lightweight visual closed-loop verification for key local areas, sets up a two-level linkage safety judgment benchmark for graded handling of fuselage drift faults, and includes automatic demagnetization and release clamping emergency protection and multi-condition autonomous shutdown logic. This deeply integrates machine vision target recognition, electromagnetic induction adaptive power extraction, and full-process safety monitoring. The entire method comprehensively addresses four major industry pain points of traditional inspection drones: insufficient battery life, poor power supply adaptability, high safety risks in field operations, and low recognition accuracy in complex environments. It eliminates the need for manual remote real-time control throughout the entire process, enabling long-distance autonomous inspection of ultra-high voltage transmission corridors with continuous online power replenishment, and comprehensively improving the level of automated line inspection operations.

[0081] The above provides a detailed description of the airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines provided by this invention. The specific embodiments described are only intended to aid in understanding the method and its core concepts. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the scope of protection of the claims of this invention.

Claims

1. An airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines, characterized by the coordinated execution of a visible light depth-of-field camera, an electromagnetic induction power extraction module, and an airborne image processing unit mounted on a UAV, wherein... Includes the following steps: S1. Cruise Image Acquisition and Image Identification: Control the UAV to autonomously cruise along the ultra-high voltage transmission corridor. The visible light depth-of-field camera continuously acquires real-scene images along the line. The airborne image processing unit uses an adaptive dynamic brightness threshold segmentation method to complete image identification, distinguishing between mountains, vegetation, clouds, strong light reflection interference backgrounds and ultra-high voltage conductors, spacers, tension fittings, and other target areas. Invalid interference images without transmission components are removed, and a standardized set of valid target images is output. If the standardized set of valid target images for the current cruise segment is empty, the UAV continues to acquire images until a valid set is obtained before proceeding to S2. S2. Multi-scale layered image extraction: Layered image extraction is performed frame by frame on the effective target image set. The edge detection algorithm is used to extract the shallow gray-scale contour of the conductor, the scale-invariant feature algorithm is used to extract the local texture features of the hardware, and the bidirectional coding algorithm is used to extract the deep semantic features of the conductor split structure and erection height. The three layers of features are spliced ​​together to generate a multi-dimensional feature vector library for power transmission components. S3, Image Classification of Transmission Components: A lightweight improved convolutional classification network is built, which completes image classification with a multi-dimensional feature vector library as input. It automatically divides the images into four categories: single conductors, four-split conductors, spacers, and tension fittings. Simultaneously, electrical attribute labels such as line voltage level, conductor cross-section, and erection height are labeled for each subset. If the classified subset is a spacer or tension fitting, it is determined that the target does not meet the conditions for power collection. The UAV continues to move along the preset route and jumps to S1 to re-collect subsequent corridor real-scene images. Only the single conductor and four-split conductor image subsets enter S4. S4. Standard Prototype Image Matching: Retrieve the ultra-high voltage power transmission standard prototype image library stored locally on the airborne machine, and use the relative entropy map matching algorithm to match the classified image subset with the corresponding voltage level standard prototype image. Calculate the image matching similarity. If there is a slight deviation in the matching of a single frame, perform multi-frame cross-checking. If the matching similarity of multiple consecutive frames does not meet the standard, it is determined that the current conductor is not suitable for power extraction. The UAV continues to move along the preset route, returns to S1 to re-collect subsequent corridor real-scene images, filters conductor images with matching similarity higher than the preset threshold, locks the target conductor for induction power extraction, and outputs the target conductor pixel coordinates and relative distance. S5. Precise control of UAV power-taking attitude: Based on the pixel coordinates and relative distance of the target conductor output by image matching, combined with the data of the UAV positioning module and inertial measurement module, the three-axis deviation of the body is calculated. The pitch, roll and yaw attitude of the UAV is adjusted by proportional-integral-derivative closed-loop control. The opening and closing magnetic ring of the airborne electromagnetic induction power-taking module is driven to maintain a safe insulation distance with the target conductor. After alignment, it is positioned to the sag section of the conductor and readjusted and aligned. Then the magnetic ring is driven to clamp the conductor and lock it. S6. Adaptive inductive power extraction with wires: After the magnetic ring is clamped and locked, the preset power extraction parameter matrix is ​​retrieved according to the voltage and cross-section labels marked in the image classification. The electronic switching switch dynamically switches the number of turns of the coil winding inside the magnetic ring. The magnetic ring closes and cuts the ultra-high voltage alternating magnetic field to generate induced alternating current, which is then rectified, filtered and regulated to charge the onboard energy storage battery of the UAV. S7. Image-linked power supply closed-loop verification: During the power supply process, the deep-eye camera continuously acquires real-time images of the target conductor, extracts local ROI areas of the conductor, and cyclically executes the entire process of image identification, image extraction, image classification, and image matching. The image matching similarity is updated in real time, and two levels of safety judgment thresholds are set. If the matching similarity is lower than the first-level warning threshold, S5 is called to fine-tune the aircraft attitude. If the similarity is lower than the second-level safety threshold, the power supply circuit is disconnected, the electromagnetic ring is separated, and S4 to S6 are re-executed for cyclic verification. The power supply operation ends when any of the following conditions are met: the energy storage battery is fully charged, a manual return command is given, an obstacle ahead of the line is identified, or the onboard battery is low. Low battery priority return protection is also set up.

2. The airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines according to claim 1, characterized in that, The image identification in S1 specifically includes: S11. Image lighting uniform correction processing: Brightness equalization correction is performed on all real-world images captured by the depth-sensing camera. After correction, the entire image is equally divided into several grid sub-regions. The texture gradient value is calculated for each sub-region. The power transmission components are distinguished from the background by combining the gradient shape length and width constraints. S12. Region Threshold Judgment and Invalid Image Removal: A dynamic discrimination threshold is generated based on the average gradient of all grid sub-regions in the entire image. Continuous connected regions with gradient values ​​greater than the dynamic discrimination threshold are judged as valid regions of transmission components. Other background regions are directly cropped and removed. Only images containing valid transmission component regions are retained as valid target images. Blank interference images without conductors or fittings are completely removed, and a standardized image set is output after image discrimination.

3. The airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines according to claim 2, characterized in that, The method for generating the dynamic discrimination threshold in S12 specifically includes: S121. Statistically calculate the texture gradient values ​​corresponding to all grid sub-regions, extract the maximum, minimum and average values ​​of all texture gradient values ​​respectively, and remove abnormal gradient noise data based on the maximum and minimum values. The abnormal gradient noise data are gradient values ​​that exceed the preset floating range of the average value. The floating range is matched with the corridor voltage level and the environmental correction coefficient. S122. The remaining effective texture gradient values ​​after removing abnormal gradient noise data are calculated by weighted average, and the compensation and correction are performed in combination with the preset environmental correction coefficient of the ultra-high voltage transmission corridor to generate a real-time dynamic identification threshold that is adapted to the current shooting light and environmental scene. This dynamic identification threshold is simultaneously involved in the linkage correction of the S7 secondary safety threshold. When the light is dim, the matching judgment benchmark is simultaneously lowered. S123. Compare the real-time texture gradient values ​​of each grid sub-region with the dynamic discrimination threshold one by one, and select the continuous grid connected regions whose gradient values ​​are greater than the dynamic discrimination threshold. Combined with the region length, width and shape constraints, these regions are the effective target regions of the power transmission components.

4. The airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines according to claim 1, characterized in that, The image classification of power transmission components in S3 specifically includes: S31. Lightweight Improved Convolutional Classification Network Construction: Based on the basic convolutional neural network, a lightweight improvement is made by removing redundant convolutional layers and fully connected layers, and embedding an attention mechanism module and a feature fusion module. This reduces the network's computational load while enhancing the feature focusing ability of multi-dimensional feature vectors, making it suitable for the embedded computing environment of UAV airborne image processing units. S32, Feature Input and Intelligent Classification: The multi-dimensional feature vector library of power transmission components generated in S2 is batch input into a lightweight improved convolutional classification network. The network learns the differences in contour, texture and semantic features of different power transmission components. It iteratively calculates the input feature vectors through preset classification weight parameters to accurately distinguish four types of targets: single conductor, four-split conductor, spacer bar and tension fitting, and generates four independent image subsets accordingly. S33. Automatic labeling of electrical attributes: A database of correlation mapping between image features and electrical attributes of transmission components is pre-established. The electrical attributes include two data sources: one is the pre-stored ledger parameters of the transmission corridor, and the other is the erection height and conductor split structure obtained by deep semantic extraction in S2. Combining the image recognition results and the preset parameters of the transmission corridor, the corresponding line voltage level, conductor cross-sectional specifications, and conductor erection height are automatically matched and labeled for each image subset after classification. This forms a classified image dataset with complete attribute labels, providing unified data support for subsequent adaptive power extraction parameter adjustment and matching threshold correction at all levels.

5. The airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines according to claim 1, characterized in that, The S4 standard prototype image matching step includes: S41. Accurate matching of prototype image retrieval: Based on the classification image dataset after automatic electrical attribute labeling completed in step S33 above, according to the electrical attributes of the transmission component category, voltage level, and conductor specifications corresponding to the image dataset, retrieve the standard prototype image that matches the currently identified component and transmission condition parameters from the airborne local storage of the ultra-high voltage transmission standard prototype image library; the standard prototype image library is organized and archived in multiple levels according to the transmission component category and the adapted transmission parameters, and matches the corresponding classification image subsets with attribute labels to complete the bidirectional matching retrieval of component type and electrical parameters; S42. Pre-processing of matching images: For the retrieved standard prototype image and the classified target image to be identified, uniformly perform image impurity filtering, image quality correction and image normalization processing to unify the image presentation state of the image to be matched and the standard prototype image, and construct the image matching operation basis. S43. Image feature matching based on relative entropy algorithm: Relying on the relative entropy image matching algorithm, extract the core feature distribution information of image texture, contour and structure of the image to be identified and the standard prototype image after pre-normalization, compare the feature distribution overlap and difference of the two sets of images, perform full-domain image matching operation, and determine the matching fit status between the image to be identified and the standard prototype image. S44. Matching result hierarchical screening and judgment: Based on the system's preset two-level matching judgment benchmark, the image matching fit status is hierarchically identified and judged. Slight feature deviations in a single frame enter the multi-frame review process. Images with multiple consecutive frames that do not match the features are defined as invalid images. The system reverts to S1 to re-acquire cruise real-scene images. Wire images that meet the features of the standard prototype image are defined as valid target images. The corresponding power transmission wire is locked as a candidate target for inductive power extraction. S45. Multi-frame image verification and target locking: Continuously acquire multiple sets of real-time images of the same power transmission line, and sequentially execute the process of prototype image retrieval, image normalization, feature matching, and result filtering. Cross-verify and comprehensively evaluate the matching results of multiple sets of images. After verification, lock the inductive power supply target line and output the position and spatial relative state information of the target line in the image for precise attitude control of the UAV.

6. The airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines according to claim 5, characterized in that, The image feature matching based on the relative entropy algorithm in S43 includes: S431. Image feature information extraction: Read the image to be identified and the standard prototype image after pre-normalization respectively, and extract all feature distribution information corresponding to the texture, contour and structure inside the two images; S432. Relative Entropy Matching Operation: The relative entropy map matching algorithm is used to compare two sets of feature distribution information to obtain the overlap and difference information between the two sets of image feature distributions. S433. Matching Status Determination: Based on the overlap and difference information obtained from the relative entropy calculation, determine the matching status between the image to be identified and the standard prototype image, and output the matching status to the subsequent matching result classification and screening determination step.

7. The airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines according to claim 1, characterized in that, The S6 conductor adaptive inductive power supply specifically includes: S61 Power Parameter Retrieval: Read the line voltage and conductor cross-section electrical attribute labels marked in step S33, call the power parameter matrix pre-stored in the device, and the winding tap position in the power parameter matrix is ​​linked and adapted with the S1 dynamic identification threshold and the S7 two-level safety matching threshold. S62, Coil winding switching control: Controls the electronic switching switch to switch the number of turns connected to the coil winding inside the magnetic ring; S63. Alternating magnetic field induction power generation: A closed magnetic ring cuts the alternating magnetic field generated by the ultra-high voltage line to generate induced alternating current. S64, Power Conditioning and Charging Output: The induced AC power is sequentially sent to the rectifier unit, filter unit, and voltage regulator unit to complete the power processing. The processed power is then sent to the UAV's onboard energy storage battery to complete the charging.

8. The airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines according to claim 7, characterized in that, The specific steps for the S62 coil winding switching control include: S621, Operating Parameter Matching and Comparison: Retrieve the power parameter matrix obtained from S61, and combine it with the voltage level and electrical properties of the current transmission line to match the working position of the coil winding of the corresponding appropriate specification. S622, Winding On / Off Switching Execution: Through the airborne electronic switching switch, according to the matching winding working position, the corresponding coil winding inside the magnetic ring is selectively connected or disconnected to complete the switching configuration of the number of coil turns connected. S623, Winding Status Locking: After completing the coil winding switching, lock the conduction state of the electronic switching switch, fix the current magnetic ring coil winding connection specification, and adapt to the current ultra-high voltage conductor induction power extraction conditions.

9. The airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines according to claim 1, characterized in that, The S7 image-linked power-feeding closed-loop verification specifically includes: S71. Real-time continuous image acquisition: During the power supply operation of the UAV, real-time images of the target conductor are continuously acquired through a visible light depth-of-field camera. S72, Lightweight Loop Iterative Recognition of Local ROI: The local region of interest of the target conductor is extracted to replace the whole image in the calculation. The local image is cyclically substituted to perform a complete processing flow of image discrimination, multi-scale hierarchical image extraction, power transmission component image classification and standard prototype image matching. S73, Real-time matching status update: Continuously update the image matching and fit status of the target conductor, and synchronously compare the system's preset first-level warning and second-level safety judgment criteria. The two judgment criteria are adaptively adjusted according to the voltage and cross-sectional parameters marked in S3. S74. Emergency handling of abnormal working conditions: When it is detected that the image matching status does not meet the safety judgment benchmark, it is determined that the UAV has drifted, the airborne power supply circuit is disconnected, and the electromagnetic induction magnetic ring structure is separated. S75. Realignment and Power Collection Operation: Repeat the steps of standard prototype image matching, precise control of UAV power collection attitude, and adaptive induction power collection of wire in sequence to complete the repositioning and alignment of the target wire for power collection. S76. Multi-condition power supply termination judgment: The closed-loop verification and power supply process are executed cyclically, and three termination conditions are monitored simultaneously: full battery charge, manual return command, and obstacle in front of the line. If any one of these conditions is triggered, power supply will be terminated, the magnetic ring will disengage and the line will return. In addition, a low battery protection is set. When the battery charge is lower than the safe return threshold, the line will be forcibly terminated and the line will return first.

10. The airborne visual recognition adaptive inductive power extraction method for ultra-high voltage transmission lines according to claim 9, characterized in that, The specific methods for identifying image matching states that do not meet security criteria in S74 include: S741. Real-time matching result acquisition: Retrieve the target traverse image matching and fitting status data updated in real time by S73, and obtain the matching correspondence between the current frame image and the standard prototype image; S742, Two-level benchmark comparison and discrimination: The real-time image matching status is compared with the system's preset safety judgment benchmark item by item to verify the matching status of the structural features and positional features of the current target conductor; S743, Abnormal State Identification: When the real-time image matching status fails to meet the preset safety judgment benchmark, the image feature correspondence shifts, or the target wire identification features mismatch, the current image matching status is determined to not meet the safety judgment benchmark, and an abnormal identification signal is generated.

Images