Power transmission tower automatic inspection method and system based on edge computing and target identification

Abstract

The application discloses a power transmission tower automatic inspection method and system based on edge computing and target identification, the method comprises the following steps: S1, system initialization; S2, route cruising and tower coarse positioning; S3, tower main body identification and accurate positioning, and accurate hovering is completed; S4, accessory identification and detail shooting, all accessories on the tower are identified through a target detection model, and detail shooting is completed according to the accessory size; S5, eight-direction flying full-dimension detection; S6, multi-tower continuous inspection, until all tower inspections are completed; S7, task completion, the edge computing box stores the inspection data locally and synchronizes to the cloud, and waits for subsequent defect detection analysis; the system is applied to the automatic inspection method, and comprises an image acquisition module, an edge computing module, a communication module and a power supply module; the inspection method realizes full-process unmanned inspection, improves real-time identification accuracy through edge computing and full-dimension coverage, reduces inspection cost and improves inspection efficiency.

Description

Technical Field

[0001] This invention relates to the field of power transmission line inspection technology, and in particular to an automatic inspection method and system for transmission towers based on edge computing and target recognition. Background Technology

[0002] Currently, transmission line tower inspections are mainly divided into two methods: manual inspection and traditional drone inspection. Manual inspection relies on inspectors climbing the tower or visually observing from the ground, which is inefficient, labor-intensive, and poses safety hazards due to high-altitude operations. It also has low accuracy in identifying defects in small components such as insulators and clamps, leading to missed or false inspections. A single tower inspection typically takes 10-15 minutes, resulting in high costs. Traditional drone inspections mostly adopt a "pilot control + post-analysis" model. The drone's flight trajectory depends on the pilot's experience, and the gimbal camera's focus and attitude need manual adjustment, making full automation impossible. Furthermore, video streams are mostly transmitted to the cloud for processing, resulting in high latency and dependence on network conditions. It also lacks accurate identification and automatic detail shooting capabilities for tower components, making it difficult to meet the needs of large-scale, refined inspections.

[0003] Among the existing publicly available technologies, such as the intelligent inspection method for overhead transmission lines by UAV formation with publication number CN116700357B, the focus is mainly on the adaptation of UAV formation to meteorological environment, without involving real-time target recognition at the edge and automatic control of gimbal focus; another example is the UAV inspection method and system for overhead transmission lines with publication number CN115268499B, which focuses on path planning and coordinate transformation, without realizing closed-loop control for tower accessory recognition and detailed shooting, and without realizing localized real-time processing.

[0004] In summary, considering existing technology and actual inspection needs, the current transmission line tower inspection methods have the following shortcomings:

[0005] (1) Low level of automation. Traditional drone inspections rely on pilot control. Pole positioning, gimbal adjustment and focus control all require manual intervention, making it impossible to achieve fully automated inspection without human intervention. Inspection efficiency is greatly affected by human factors.

[0006] (2) Poor real-time performance. Existing technologies mostly use cloud processing for video streams, which has a delay of more than 200ms. It cannot quickly respond to the target recognition results and drive the drone and gimbal to make immediate adjustments, which can easily lead to blurry shooting of accessory details and positioning deviation.

[0007] (3) Insufficient recognition accuracy. There is a lack of dedicated recognition models for transmission tower accessories (insulators, crossarms, clamps, etc.). The recognition accuracy of small targets (clamps, anti-vibration hammers) is less than 85%, which is difficult to meet the pre-defect detection requirements.

[0008] (4) The inspection coverage is not comprehensive. Most solutions do not achieve 360° all-dimensional fly-around inspection of the tower. They can only take pictures in a single direction, which easily leads to the problem of missing inspection of the accessories on the back and side of the tower.

[0009] (5) Poor hardware adaptability: existing edge computing solutions mostly use general-purpose chips, making it difficult to balance computing power and power consumption. They cannot adapt to the limited power supply resources of drones and have not been specifically optimized for power transmission inspection scenarios, making deployment difficult. Summary of the Invention

[0010] To overcome the above problems, the purpose of this invention is to provide an automatic inspection method and system for power transmission towers based on edge computing and target recognition. This inspection method realizes unmanned inspection throughout the entire process, improves the real-time performance and expected recognition accuracy of the inspection through edge computing and full-dimensional coverage, reduces inspection costs, and improves inspection efficiency.

[0011] The technical solution adopted in this invention is:

[0012] An automatic inspection method for power transmission towers based on edge computing and target recognition includes the following steps:

[0013] S1: System initialization, drone power-on, gimbal camera self-test, edge computing box starts and loads optimized target detection model.

[0014] S2: Route navigation and coarse tower positioning. The drone automatically flies to the first tower according to the preset route. After reaching the preset inspection starting point, it hovers and starts to collect video stream from the gimbal camera, which is then transmitted to the edge computing box in real time.

[0015] S3: Pole and tower body recognition and precise positioning. The edge computing box uses a target detection model to identify the main body of the pole and tower in the video stream in real time, outputs pixel coordinates, calculates its deviation from the center of the screen, controls the drone to move and adjust, and places the entire pole and tower in the exact center of the video screen to achieve precise hovering.

[0016] S4: Accessory recognition and detail shooting. The target detection model identifies all accessories on the tower and completes detailed shooting based on the size of the accessories.

[0017] S5: Eight-directional circling full-dimensional detection. With the tower as the center, it flies around at a constant speed within the detection radius, and repeats S3-S4 every 45° to complete the eight-directional detection.

[0018] S6: Continuous inspection of multiple towers. After the inspection of a single tower is completed, the system automatically generates an inspection completion mark, the edge computing box issues a flight path continuation command, the drone automatically returns to the preset flight path, flies to the next tower, and repeats S2-S5 until all towers are inspected.

[0019] S7: Mission complete. After all tower inspections are completed, the drone automatically returns to the takeoff point. The edge computing box stores the inspection data locally and synchronizes it to the cloud, awaiting subsequent defect detection and analysis.

[0020] As a further description of the present invention, the specific process of S4 is as follows:

[0021] S41: The edge computing box identifies all the accessories on the tower through the object detection model and outputs the category, pixel coordinates and confidence score of each accessory.

[0022] S42: Convert the pixel coordinates of the accessory into the tilt / rotation angle of the gimbal, issue a command to drive the gimbal to rotate, and move the target accessory to the center of the screen.

[0023] S43: Automatically adjusts the focal length according to the size of the accessories, with small accessories adjusted to the maximum optical zoom and large accessories adjusted to a moderate focal length.

[0024] S44: Automatically captures high-resolution detail images and stores them locally after focus lock.

[0025] As a further description of the present invention, the automatic adjustment of the focal length according to the accessory size in S43 is accomplished by an adaptive control algorithm of gimbal and focal length. The adaptive control algorithm establishes a coordinate-angle transformation model based on the proportional relationship between the accessory pixel coordinates and the image size to achieve precise gimbal alignment; based on the accessory category and size, preset focal length adjustment rules are set to achieve automatic focal length adaptation without manual intervention.

[0026] As a further description of the present invention, the construction steps of the coordinate-angle transformation model are as follows:

[0027] Step 1: Based on the images captured by the drone's camera, output target detection values ​​using the target detection model. .

[0028] Step 2: Pixel deviation normalization, calculate the normalized offset of the target relative to the center of the image.

[0029] X-direction (yaw) normalized offset .

[0030] Y-direction (pitch) normalized offset .

[0031] Step 3: Determine whether the offset in the X and Y directions meets the center alignment requirement. If it does, the target accessory has been moved to the center of the screen, completing the precise alignment of the gimbal. If it does not meet the requirement, proceed to step 4.

[0032] Step 4: Coordinate-angle mapping. The required rotation angle is calculated by normalizing the offset and the maximum angle of the gimbal.

[0033] Gimbal yaw control angle .

[0034] Gimbal pitch control angle .

[0035] Step 5: The value is sent to the gimbal for rotation and then returns to the first step to recapture the image and update the position.

[0036] As a further description of the present invention, the focal length focusing rule is as follows:

[0037] Prioritize matching the reference focal length according to accessory category; within the same category, make fine adjustments according to the physical size of the accessory; after fine adjustments, make closed-loop corrections according to the proportion of the image.

[0038] As a further description of the present invention, the accessory categories include large accessories, medium accessories, and small accessories. The physical dimensions of large accessories are greater than or equal to 25mm, the physical dimensions of medium accessories are greater than or equal to 10mm and less than 25mm, and the physical dimensions of small accessories are less than 10mm.

[0039] As a further description of the present invention, the detection radius in S5 is in the range of 8m-12m.

[0040] As a further description of the present invention, the edge computing box adopts RK3588 or RK3576.

[0041] As a further description of the present invention, the method collects distance data between the drone and the tower in real time, sets a safe distance threshold and a low battery threshold, and automatically issues hovering or backing commands when the distance is less than the threshold to prevent collision; when the remaining battery is less than the low battery threshold, it triggers automatic return to home to ensure the safety of the drone.

[0042] An automatic inspection system for power transmission towers based on edge computing and target recognition is applied to an automatic inspection method for power transmission towers based on edge computing and target recognition. The system includes:

[0043] The image acquisition module uses a gimbal camera mounted on a drone. It employs a variable-focus optical pod, supports RTSP video stream output, and has pitch and rotation adjustment functions. The optical zoom is ≥20x, the resolution is ≥4K, and it supports the gimbal control protocol. It receives attitude and focal length control commands from the edge module to complete the real-time acquisition of image data of power transmission towers and their accessories.

[0044] The edge computing module, using an edge computing box with RK3588 or RK3576 chips, carries a dedicated operating system image and completes video stream decoding, target detection model inference, and control command issuance. It is installed in the drone and imports KML coordinates to generate a cruise route.

[0045] The communication module combines a drone data transmission radio with a gigabit network port on the edge box, with a data transmission distance of ≥5km and a network port transmission rate of ≥1000Mbps, enabling bidirectional real-time transmission of video streams and control commands with a transmission latency of ≤50ms.

[0046] The power supply module adopts a customized step-down power supply module for drones. The input voltage is the drone battery voltage, and the output voltage is 12V / 24V. It provides stable power supply for edge computing boxes and gimbal cameras, with a flight time of ≥2 hours.

[0047] The beneficial effects of this invention are:

[0048] This invention presents an automated inspection method for power transmission towers based on edge computing and target recognition. This method achieves fully unmanned operation through tower positioning, centering, component identification, detailed imaging, fly-around detection, and continuous inspection of multiple towers, eliminating the need for drone operator intervention and overcoming the shortcomings of existing technologies that rely on manual operation. This significantly reduces the average time spent on manual inspection of a single tower, improves the overall inspection efficiency of towers, significantly enhances the degree of automation, and reduces the risks associated with manual tower climbing.

[0049] This invention relates to an automatic inspection method for power transmission towers based on edge computing and target recognition. Through 360° fly-around and eight-directional fixed-point detection, it solves the problems of incomplete inspection coverage and easy omissions in existing technologies. The inspection coverage of towers and accessories reaches 100%, ensuring the reliability of inspection results.

[0050] This invention relates to an automatic inspection method for power transmission towers based on edge computing and target recognition. It uses RK3588 or RK3576 low-power edge chips to balance computing power and power consumption, adapts to the limited power supply resources of drones, has low deployment difficulty, good hardware compatibility, and high stability. Attached Figure Description

[0051] Figure 1 This is a flowchart of the automatic inspection method for power transmission towers based on edge computing and target recognition proposed in this invention.

[0052] Figure 2 This is a schematic diagram of an eight-directional detection method for automatic inspection of power transmission towers based on edge computing and target recognition proposed in this invention.

[0053] Figure 3 This is a hardware connection topology diagram of the automatic inspection system for power transmission towers based on edge computing and target recognition proposed in this invention.

[0054] Figure 4 This is a software layered architecture diagram of the automatic inspection system for power transmission towers based on edge computing and target recognition proposed in this invention. Detailed Implementation

[0055] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0058] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0059] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0060] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0061] like Figures 1-4 As shown, it illustrates a specific embodiment of the present invention:

[0062] Example 1:

[0063] An automatic inspection method for power transmission towers based on edge computing and target recognition includes the following steps:

[0064] S1: System initialization, drone power-on, gimbal self-test, edge computing box starts and loads optimized target detection model, completes video stream and control link connection, and ensures that each module works normally.

[0065] S2: Route navigation and coarse tower positioning. The drone automatically flies to the first tower according to the preset route (generated by importing the KML coordinates of the power transmission line). After reaching the preset inspection starting point (10-15m away from the tower), it hovers and starts to collect video stream from the gimbal camera, which is then transmitted to the edge computing box in real time.

[0066] S3: Pole tower main body recognition and precise positioning. The edge computing box identifies the main body of the pole tower in the video stream in real time through the target detection model, outputs pixel coordinates, calculates its deviation from the center of the screen, and sends fine-tuning instructions to the drone flight control center to control the drone to move and adjust, placing the entire pole tower in the exact center of the video screen to complete precise hovering.

[0067] S4: Accessory recognition and detail shooting: Identify all accessories on the tower through the target detection model and complete the detail shooting according to the size of the accessories;

[0068] Specifically, the process of S4 is as follows:

[0069] S41: The edge computing box identifies all accessories on the tower, including insulators, crossarms, clamps, and vibration dampers, using a target detection model, and outputs the category, pixel coordinates, and confidence score for each accessory (confidence threshold ≥ 0.85).

[0070] S42: Convert the pixel coordinates of the accessory into the pitch / rotation angle of the gimbal, and send a command to drive the gimbal to rotate, so that the target accessory moves to the center of the screen;

[0071] S43: Automatically adjusts the focal length according to the size of the accessories, with small accessories adjusted to the maximum optical zoom and large accessories adjusted to a moderate focal length.

[0072] S44: Automatically captures high-resolution detail images and stores them locally after focus lock.

[0073] S5: Eight-directional circling full-dimensional detection. With the tower as the center, it circles at a constant speed within the detection radius. It hovers at 45° and repeats S3-S4 to complete the eight-directional detection.

[0074] In this embodiment, the specific process of S5 is as follows:

[0075] After the detailed shooting of the accessories in one direction is completed, the edge computing box issues a flight command to control the drone to fly around the tower at a constant speed with a radius of 8-12m, hovering in eight core directions: front, right, back, and left. S3-S4 are repeated in each direction to identify the tower and accessories from multiple angles and shoot detailed images to ensure that nothing is missed.

[0076] S6: Continuous inspection of multiple towers. After the inspection of a single tower is completed, the system automatically generates an inspection completion mark, the edge computing box issues a flight path continuation command, the drone automatically returns to the preset flight path, flies to the next tower, repeats S2-S5, until all towers are inspected;

[0077] S7: Mission complete. After all tower inspections are completed, the drone automatically returns to the takeoff point. The edge computing box stores the inspection data (detailed images, videos, logs) locally and synchronizes it to the cloud, awaiting subsequent defect detection and analysis.

[0078] In this embodiment, the inspection method achieves fully unmanned operation through tower positioning, centering, accessory identification, detailed imaging, fly-around detection, and continuous inspection of multiple towers, eliminating the need for pilot intervention and overcoming the shortcomings of existing technologies that rely on manual operation. This significantly reduces the average time spent on manual inspection of a single tower, improves the overall inspection efficiency of towers, significantly enhances the degree of automation, and reduces the risk of manual tower climbing. Through 360° fly-around and eight-directional fixed-point detection, the method solves the problems of incomplete inspection coverage and easy omissions in existing technologies, achieving 100% inspection coverage of towers and accessories, ensuring the reliability of inspection results. Utilizing the RK3588 or RK3576 low-power edge chip balances computing power and power consumption, adapts to the limited power resources of drones, has low deployment difficulty, good hardware compatibility, and high stability.

[0079] Example 2:

[0080] Specifically, in S43, the automatic adjustment of the focal length based on the accessory size is accomplished using an adaptive control algorithm of the gimbal and the focal length. The adaptive control algorithm establishes a coordinate-angle transformation model based on the ratio between the accessory pixel coordinates and the image size to achieve precise gimbal alignment. Based on the accessory type and size, a preset focal length adjustment rule is established to achieve automatic focal length adaptation without manual intervention.

[0081] In this embodiment, the core of the coordinate-angle transformation model is to convert the pixel coordinates of the accessory in the image into the gimbal pitch / yaw control angle, thereby achieving precise gimbal alignment. The entire process is based on imaging geometry and scale normalization calculations.

[0082] In this embodiment, the basic definitions include the following:

[0083] Category 1: Image baseline parameters.

[0084] Screen resolution: W (width, pixels), H (height, pixels).

[0085] Center pixel coordinates of the image: , The center pixel coordinates of the image are the desired alignment point of the gimbal.

[0086] The second category: detecting output parameters.

[0087] Accessory target pixel coordinates: .

[0088] The third category: gimbal control dimension.

[0089] Yaw axis: Align left and right (corresponding to the X-axis of the screen);

[0090] Pitch: Aligns vertically (corresponding to the Y-axis of the screen).

[0091] Category 4: Proportioning coefficients and restrictions

[0092] Maximum yaw angle of gimbal: (°), Maximum pitch angle: (°).

[0093] Conversion gain: K (value range is 0~1, used for smoothing / overshoot prevention, default is 0.8~0.95).

[0094] The steps for constructing the coordinate-angle transformation model are as follows:

[0095] Step 1: Based on the images captured by the drone's camera, output target detection values ​​using the target detection model. ;

[0096] Step 2: Pixel deviation normalization, calculate the normalized offset of the target relative to the center of the image: (the offset value ranges from...) This eliminates the impact of different resolutions and screen sizes.

[0097] X-direction (yaw) normalized offset ;

[0098] Y-direction (pitch) normalized offset ;

[0099] The physical meaning of this step is: This indicates that the target is on the far right of the screen; This indicates that the center is aligned.

[0100] Step 3: Determine whether the offset in the X and Y directions meets the center alignment requirement. If it does, the target accessory has been moved to the center of the screen, completing the precise alignment of the gimbal. If it does not meet the requirement, proceed to step 4.

[0101] Step 4: Coordinate-angle mapping, calculate the required rotation angle by normalizing the offset and the maximum angle of the gimbal;

[0102] Gimbal yaw control angle ;

[0103] Gimbal pitch control angle ;

[0104] In this embodiment, the mapping relationship between coordinates and angles is transformed into a pixel offset ratio, which is equal to the gimbal angle ratio. The calculation of this relationship is based on linear mapping, has an extremely fast calculation speed, and can be adapted to real-time control scenarios.

[0105] Step 5: The value is sent to the gimbal for rotation and then returns to the first step to recapture the image and update the position.

[0106] In this embodiment, the actual coordinate-angle transformation model is implemented with constraints of a central dead zone and angle limiting to ensure stability and prevent jitter.

[0107] ①Center dead zone: When the offset is <2%~3%, no action is taken (to avoid micro-jitter), that is, when hour, .

[0108] ② Angle limit: the final calculated angle limit The output value shall not exceed the physical limits of the gimbal.

[0109] Specifically, the focus adjustment rule is as follows:

[0110] Prioritize matching the reference focal length according to accessory category; within the same category, make fine adjustments according to the physical size of the accessory; after fine adjustments, make closed-loop corrections according to the proportion of the image.

[0111] In this embodiment, the design principle of the focal length adjustment rule is: the larger / closer the accessory, the smaller the focal length (wide-angle); the smaller / farther the accessory, the larger the focal length (telephoto); to achieve "the optimal size of the accessory, occupying 60%~80% of the frame".

[0112] Table 1 below shows the specific requirements for prioritizing the matching of reference focal lengths according to accessory categories.

[0113] Table 1. Prioritizing Matching Reference Focal Length by Accessory Category

[0114]

[0115] After matching the base focal length according to the accessory category, secondary optimization is performed within the same category, with fine adjustments made according to the physical dimensions of the accessories. Specific requirements are shown in Table 2 below:

[0116] Table 2 Secondary Optimization Based on Physical Dimensions

[0117]

[0118] Finally, a closed-loop correction is performed based on the screen ratio to ensure optimal composition. The goal is to ensure that accessories occupy 60% to 80% of the effective area of ​​the screen. When the screen ratio is >85% (too large), the focal length is set to -1 stop (wide-angle); when the screen ratio is <50% (too small), the focal length is set to +1 stop (telephoto); when the screen ratio is between 60% and 80%, the current focal length is maintained.

[0119] In this embodiment, to ensure the reliability of the shooting results after automatic focal length adaptation, the lower limit of the focal length is set to be no less than 24mm to avoid distortion; the upper limit of the focal length is set to be no more than 135mm to avoid shaking / out of focus; during the focusing process, the switching interval is ≥300ms to prevent frequent zooming from causing deviations in the shooting results.

[0120] The following is an example under this automatic focal length adaptation rule:

[0121] When a large accessory, such as a 35cm battery, is detected, the focus is automatically adjusted to 35mm.

[0122] When a medium-sized accessory, such as a 15cm remote control, is detected, the focus is automatically adjusted to 50mm.

[0123] When a small accessory, such as a 3cm screw, is detected, the system automatically adapts and selects a focal length of 135mm.

[0124] When the accessories occupy 90% of the frame, reduce the focal length by one stop.

[0125] In this embodiment, a safety protection mechanism is set up to collect real-time distance data between the drone and the tower, and a safe distance threshold (≥5m) is set. When the distance is less than the threshold, a hovering or backward command is automatically issued to prevent collision. A low battery threshold (remaining battery ≤20%) is preset to trigger automatic return to home, ensuring the safety of the drone.

[0126] Example 3:

[0127] An automatic inspection system for power transmission towers based on edge computing and target recognition is applied to an automatic inspection method for power transmission towers based on edge computing and target recognition. The system includes:

[0128] The image acquisition module uses a gimbal camera mounted on a drone. It employs a variable-focus optical pod, supports RTSP video stream output, and has pitch and rotation adjustment functions. The optical zoom is ≥20x, the resolution is ≥4K, and it supports the gimbal control protocol. It receives attitude and focal length control commands from the edge module to complete the real-time acquisition of image data of power transmission towers and their accessories.

[0129] The edge computing module, using an edge computing box with RK3588 or RK3576 chips, carries a dedicated operating system image and completes video stream decoding, target detection model inference, and control command issuance. It is installed in the drone and imports KML coordinates to generate a cruise route.

[0130] The communication module combines a drone data transmission radio with a gigabit network port on the edge box, with a data transmission distance of ≥5km and a network port transmission rate of ≥1000Mbps, enabling bidirectional real-time transmission of video streams and control commands with a transmission latency of ≤50ms.

[0131] The power supply module adopts a customized step-down power supply module for drones. The input voltage is the drone battery voltage, and the output voltage is 12V / 24V. It provides stable power supply for edge computing boxes and gimbal cameras, with a flight time of ≥2 hours.

[0132] In this embodiment, the system achieves a closed-loop operation of "autonomous flight of UAV + real-time edge processing + target recognition + automatic control + continuous inspection of multiple towers" through two parts: hardware system and software system.

[0133] The hardware system consists of a drone platform, an edge computing box, a gimbal camera, a communication module, and a power supply module. The connection relationships and parameter limitations of each module are as follows:

[0134] (1) Unmanned Aerial Vehicle Platform: An industry-grade multi-rotor UAV is selected, which has autonomous flight, hovering and circling functions, supports flight control SDK / API, can receive flight control commands issued by the edge box, preset cruise speed of 5-8m / s, and hovering accuracy of ±0.5m.

[0135] (2) Edge computing box: RK3588 or RK3576 chip is selected. The RK3588 chip is an 8nm process and has 6TOPS AI computing power; the RK3576 chip has 4TOPS AI computing power. Both support NPU hardware acceleration, are equipped with a dedicated operating system image, and are used for video stream decoding, target detection model inference, and control command issuance. The size is adapted to the space of the drone and the power consumption is ≤15W.

[0136] (3) Gimbal camera: It adopts a variable zoom optical pod, supports RTSP video stream output, has pitch and rotation adjustment functions, optical zoom magnification ≥20 times, resolution ≥4K, supports gimbal control protocol, and can receive attitude and focal length control commands sent by edge box.

[0137] (4) Communication module: The UAV data transmission radio is combined with the edge box gigabit network port. The data transmission distance is ≥5km and the network port transmission rate is ≥1000Mbps. It realizes bidirectional real-time transmission of video stream and control commands with a transmission delay of ≤50ms.

[0138] (5) Power supply module: The UAV adopts a customized step-down power supply module with an input voltage of the UAV battery voltage (22.2V / 3S) and an output voltage of 12V / 24V. It provides stable power supply for edge computing boxes and gimbal cameras, with a flight time of ≥2 hours, meeting the needs of continuous inspection of multiple towers.

[0139] In this embodiment, the hardware connection topology is as follows: the gimbal camera is connected to the edge box's network port via RTSP network stream; the edge box is connected to the drone flight controller via serial port / network port, and to the optoelectronic pod via gimbal control protocol; the drone battery supplies power to the edge box and gimbal camera respectively via step-down module.

[0140] In this embodiment, the system software adopts a layered and modular design, divided into a perception layer, an edge computing layer, a control layer, an application layer, a communication layer, and a storage layer. Each layer works collaboratively. The core technology stack includes a dedicated operating system, an edge computing inference framework, a video processing framework, a target detection model, and a UAV flight control interface. The specific module functions are as follows:

[0141] (1) Perception layer: responsible for collecting video streams from the gimbal camera (frame rate ≥ 30FPS) and UAV positioning / attitude data (GPS, gyroscope data), and transmitting the collected video streams and positioning data to the edge computing layer in real time.

[0142] (2) Edge computing layer: Deployed on RK3588 / RK3576 edge boxes, core functions include: real-time video stream decoding (using a dedicated video processing framework), object detection model inference (using an edge computing inference framework for optimization and enabling NPU acceleration), object coordinate calculation and transformation, inference speed ≥30FPS, processing latency ≤200ms.

[0143] (3) Control layer: Receives the recognition results of the edge computing layer and issues control commands, including UAV flight control (hovering, fine adjustment, and flyaround), gimbal attitude control (pitch and rotation), and automatic focus adjustment. Control accuracy: tower image centering error ≤5%, gimbal alignment error ≤1°.

[0144] (4) Application layer: responsible for task planning (importing KML coordinates to generate cruise route), defect detection data preprocessing, task status monitoring, and supporting single tower inspection completion marking and multi-tower continuous inspection triggering.

[0145] (5) Communication layer: responsible for data transmission between modules, including bidirectional transmission of video streams, control commands, positioning data and inspection logs, using encrypted transmission protocols to ensure data security.

[0146] (6) Storage layer: responsible for local high-definition image / video storage (using SD card / solid-state drive, storage capacity ≥128GB) and cloud synchronization. The image naming rule is "tower number_location_part category_time.jpg" to facilitate subsequent defect analysis.

[0147] Example 4:

[0148] In this embodiment, based on the RK3588 edge box, this embodiment selects 10 towers of a 220kV transmission line (6 straight towers and 4 tension towers) and conducts actual measurements under sunny and light wind (level 2) conditions. The implementation process is as follows:

[0149] (1) System initialization: The UAV is powered on, the gimbal performs a self-test, the edge box starts up, the optimized target detection model (dedicated inference format) is loaded, the video stream and control link are connected, and the system enters standby mode after the self-test is passed.

[0150] (2) Route cruise and coarse positioning: The UAV flies to the first base tower according to the preset route, hovers at 12m, turns on the gimbal camera, collects video stream (30FPS) in real time and transmits it to the edge box.

[0151] (3) Centering the tower: The edge box identifies the main body of the tower through the target detection model and calculates that the deviation between the main body of the tower and the center of the image is 3% (less than the threshold of 5%). No fine-tuning is required to complete the centering and hovering.

[0152] (4) Accessory identification and detail shooting: The model identified 2 sets of insulator strings, 1 crossarm, 4 wire clamps and 2 anti-vibration hammers on the tower, with a confidence level of ≥0.9. The edge box converted the coordinates of each accessory into the angle of the gimbal, drove the gimbal to align with each accessory in turn, adjusted the focus and took 4K detail pictures. A total of 8 detail pictures were taken and stored locally.

[0153] (5) 360° Flyaround Detection: The edge box issues a flyaround command, and the drone flies around the tower with a radius of 10m and a speed of 3m / s, hovering in front, right, back and left respectively. The above identification and shooting process is repeated in each position, and a total of 32 detailed pictures are taken.

[0154] (6) Multi-tower continuous inspection: After the first tower is inspected (takes 2 minutes and 40 seconds), the system generates a completion mark, the UAV automatically returns to the flight path, flies to the next tower, and repeats the above process until all 10 towers are inspected.

[0155] (7) Mission complete: All tower inspections are completed, the UAV automatically returns to the takeoff point, and the edge box stores 400 detailed pictures, inspection videos and flight logs locally and synchronizes them to the cloud.

[0156] The specific test data is as follows:

[0157] (1) Identification accuracy: Insulator identification accuracy rate 96.8%, crossarm identification accuracy rate 97.2%, wire clamp identification accuracy rate 95.6%, vibration damper identification accuracy rate 95.4%, and average accessory identification accuracy rate 96.5%.

[0158] (2) Real-time performance: Video stream processing latency is 175-190ms, with an average of 182ms. The object detection model inference speed is 33-37FPS, with an average of 35FPS.

[0159] (3) Inspection efficiency: The inspection time for a single tower is 2 minutes 32 seconds to 2 minutes 48 seconds, with an average of 2 minutes 40 seconds. The total inspection time for 10 towers is 26 minutes 40 seconds, which saves about 98 minutes compared to manual inspection and about 41 minutes compared to traditional drone inspection.

[0160] (4) Inspection coverage: The eight-directional fly-around inspection coverage rate is 100%, with no missed or false inspections. A total of 400 4K detail images were taken, and the detail clarity meets the defect detection requirements (cracks larger than 0.5mm can be clearly identified).

[0161] (5) Stability: The flight is stable throughout the entire process with no risk of collision. The system operates continuously without crashes or jams. The hovering accuracy is stable within ±0.3m, and the gimbal alignment error is ≤0.8°, which meets the inspection requirements.

[0162] Example 5:

[0163] In this embodiment, based on the RK3576 edge box, 10 towers of a 110kV medium-low voltage transmission line were selected for actual testing under cloudy and light wind (level 3) conditions. The implementation process is the same as in embodiment four, except that a lightweight target detection model and RK3576 chip are used. The inspection time for a single tower is 2 minutes and 55 seconds, the accessory identification accuracy is 95.2%, and the video stream processing latency is 195ms, which meets the requirements of lightweight inspection and is suitable for medium-low voltage transmission line inspection scenarios.

[0164] The specific test data is as follows:

[0165] (1) Identification accuracy: Insulator identification accuracy rate 95.4%, crossarm identification accuracy rate 95.8%, wire clamp identification accuracy rate 94.8%, vibration damper identification accuracy rate 94.6%, and average accessory identification accuracy rate 95.2%.

[0166] (2) Real-time performance: Video stream processing latency is 188-198ms, with an average of 193ms. The inference speed of the lightweight object detection model is 28-32FPS, with an average of 30FPS.

[0167] (3) Inspection efficiency: The inspection time for a single tower is 2 minutes 45 seconds to 3 minutes 05 seconds, with an average of 2 minutes 55 seconds. The total inspection time for 10 towers is 29 minutes 10 seconds, which saves about 96 minutes compared to manual inspection and about 39 minutes compared to traditional drone inspection.

[0168] (4) Inspection coverage: The eight-directional fly-around detection coverage rate is 100%, with no missed or false detections, and the clarity of the detailed diagrams meets the requirements for defect detection of medium and low voltage transmission lines.

[0169] (5) Cost and stability: The hardware cost is reduced by 20% compared with the fourth embodiment. The system runs continuously without crashes or lags. The hovering accuracy is stable within ±0.5m, and the gimbal alignment error is ≤0.9°. It achieves lightweight deployment, is compatible with small and medium-sized UAVs, and meets the needs of large-scale inspection.

[0170] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

[0171] Many other changes and modifications can be made without departing from the concept and scope of this invention. It should be understood that this invention is not limited to the specific embodiments, and the scope of this invention is defined by the appended claims.

Claims

1. An automatic inspection method for power transmission towers based on edge computing and target recognition, characterized in that, Includes the following steps: S1: System initialization, drone power-on, gimbal camera self-test, edge computing box starts and loads the optimized target detection model; S2: Route cruise and tower coarse positioning. The drone automatically flies to the first tower according to the preset route. After reaching the preset inspection start point, it hovers and starts to collect video stream from the gimbal camera and transmits it to the edge computing box in real time. S3: Pole tower main body recognition and precise positioning. The edge computing box identifies the main body of the pole tower in the video stream in real time through the target detection model, outputs pixel coordinates, calculates its deviation from the center of the screen, controls the drone to move and adjust, and places the entire pole tower in the exact center of the video screen to complete precise hovering. S4: Accessory identification and detail shooting: Identify all accessories on the tower through the target detection model and complete the detail shooting according to the size of the accessories; S5: Eight-directional circling full-dimensional detection. With the tower as the center, it circles at a constant speed within the detection radius. It hovers at 45° and repeats S3-S4 to complete the eight-directional detection. S6: Continuous inspection of multiple towers. After the inspection of a single tower is completed, the system automatically generates an inspection completion mark, the edge computing box issues a flight path continuation command, the drone automatically returns to the preset flight path, flies to the next tower, repeats S2-S5, until all towers are inspected; S7: Mission complete. After all tower inspections are completed, the drone automatically returns to the takeoff point. The edge computing box stores the inspection data locally and synchronizes it to the cloud, awaiting subsequent defect detection and analysis.

2. The automatic inspection method for transmission towers based on edge computing and target recognition according to claim 1, characterized in that, The specific process of S4 is as follows: S41: The edge computing box identifies all the accessories on the tower through the object detection model and outputs the category, pixel coordinates and confidence score of each accessory; S42: Convert the pixel coordinates of the accessory into the pitch / rotation angle of the gimbal, and send a command to drive the gimbal to rotate, so that the target accessory moves to the center of the screen; S43: Automatically adjusts the focal length according to the size of the accessories, with small accessories adjusted to the maximum optical zoom and large accessories adjusted to a moderate focal length. S44: Automatically captures high-resolution detail images and stores them locally after focus lock.

3. The automatic inspection method for transmission towers based on edge computing and target recognition according to claim 2, characterized in that, In S43, the automatic adjustment of the focal length based on the accessory size is accomplished by an adaptive control algorithm of the gimbal and the focal length. The adaptive control algorithm establishes a coordinate-angle transformation model based on the ratio between the accessory pixel coordinates and the image size to achieve precise gimbal alignment. Based on the accessory type and size, a preset focal length adjustment rule is established to achieve automatic focal length adaptation without manual intervention.

4. The automatic inspection method for transmission towers based on edge computing and target recognition according to claim 3, characterized in that, The steps for constructing the coordinate-angle transformation model are as follows: Step 1: Based on the images captured by the drone's camera, output target detection values ​​using the target detection model. ; Step 2: Pixel deviation normalization, calculate the normalized offset of the target relative to the center of the image: X-direction (yaw) normalized offset ; Y-direction (pitch) normalized offset ; Step 3: Determine whether the offset in the X and Y directions meets the center alignment requirement. If it does, the target accessory has been moved to the center of the screen, completing the precise alignment of the gimbal. If it does not meet the requirement, proceed to step 4. Step 4: Coordinate-angle mapping, calculate the required rotation angle by normalizing the offset and the maximum angle of the gimbal; Gimbal yaw control angle ; Gimbal pitch control angle ; Step 5: The value is sent to the gimbal for rotation and then returns to the first step to recapture the image and update the position.

5. The automatic inspection method for transmission towers based on edge computing and target recognition according to claim 3, characterized in that, The focal length focusing rule is as follows: Prioritize matching the reference focal length according to accessory category; within the same category, make fine adjustments according to the physical size of the accessory; after fine adjustments, make closed-loop corrections according to the proportion of the image.

6. The automatic inspection method for transmission towers based on edge computing and target recognition according to claim 5, characterized in that, The accessory categories include large accessories, medium accessories, and small accessories. The physical dimensions of large accessories are greater than or equal to 25mm, the physical dimensions of medium accessories are greater than or equal to 10mm and less than 25mm, and the physical dimensions of small accessories are less than 10mm.

7. The automatic inspection method for transmission towers based on edge computing and target recognition according to claim 1, characterized in that, The detection radius in S5 ranges from 8m to 12m.

8. The automatic inspection method for transmission towers based on edge computing and target recognition according to claim 1, characterized in that, The edge computing box uses either RK3588 or RK3576.

9. The automatic inspection method for transmission towers based on edge computing and target recognition according to claim 1, characterized in that, The method collects distance data between the drone and the tower in real time, sets a safe distance threshold and a low battery threshold, and automatically issues hover or back command when the distance is less than the threshold to prevent collision; when the remaining battery is less than the low battery threshold, it triggers automatic return to home to ensure the safety of the drone.

10. An automatic inspection system for transmission towers based on edge computing and target recognition, applied to the automatic inspection method for transmission towers based on edge computing and target recognition as described in any one of claims 1 to 9, characterized in that, The system includes: The image acquisition module uses a gimbal camera mounted on a drone. It adopts a variable-focus optical pod, supports RTSP video stream output, has pitch and rotation adjustment functions, optical zoom of ≥20x, resolution of ≥4K, supports gimbal control protocol, receives attitude and focal length control commands issued by edge module, and completes real-time acquisition of image data of power transmission towers and their accessories. The edge computing module, using an edge computing box with RK3588 or RK3576 chips, is equipped with a dedicated operating system image to complete video stream decoding, target detection model inference, and control command issuance. It is installed in the drone and imports KML coordinates to generate a cruise route. The communication module combines a drone data transmission radio with a gigabit network port on the edge box, with a data transmission distance of ≥5km and a network port transmission rate of ≥1000Mbps, enabling bidirectional real-time transmission of video streams and control commands with a transmission latency of ≤50ms. The power supply module adopts a customized step-down power supply module for drones. The input voltage is the drone battery voltage, and the output voltage is 12V / 24V. It provides stable power supply for edge computing boxes and gimbal cameras, with a flight time of ≥2 hours.

Images