An end-to-end edge side power line icing closed-loop monitoring method and system

Abstract

The application provides an end-to-end edge side power transmission line icing closed-loop monitoring method and system, and belongs to the technical field of computer vision. The application is used for solving the problems of excessive dependence on network bandwidth, high delay, poor quantitative calculation accuracy and inability to form a local early warning closed loop in the existing power transmission line icing closed-loop monitoring. The method comprises the following steps: S1, reference calibration and pixel equivalent calculation; S2, edge side visual acquisition and preprocessing; S3, pixel-level inference based on a lightweight instance segmentation model; S4, icing quantitative evaluation and thickness calculation; and S5, local closed-loop arbitration and multi-modal early warning. The system comprises an image acquisition module, an edge side intelligent calculation module, an icing quantitative evaluation module and a multi-modal closed-loop output module.

Description

Technical Field

[0001] This invention relates to a closed-loop monitoring method and system for end-to-end edge-side transmission line icing, belonging to the field of computer vision technology. Background Technology

[0002] High-voltage and ultra-high-voltage transmission lines are exposed to the natural environment year-round, making them highly susceptible to icing in winter or in high-altitude mountainous areas. Severe icing can cause transmission cables to sag, gallop, or even break and collapse, posing a devastating threat to power grid safety. Therefore, real-time monitoring of the icing status of transmission lines is crucial. Current technologies mainly involve the following three approaches and their corresponding drawbacks: 1) Indirect Measurement Method Based on Tension / Micrometeorological Sensors: This method calculates icing thickness by installing load cells or tension sensors on the insulator strings of the tower, combined with micrometeorological data such as wind speed and temperature. Disadvantages: This method is indirect and highly susceptible to wind deviation and line aging, resulting in extremely high errors. It also fails to visually reflect icing patterns (such as rime, hoarfrost, and eccentric icing), and the sensors are prone to damage, leading to high maintenance costs. 2) Visual Monitoring Method Based on Cloud Processing: Cameras are installed on the towers, and periodically captured images or real-time video streams are transmitted via 4G / 5G networks to a remote server or cloud platform for image recognition by a large cloud-based model. Disadvantages: Network signals are extremely unstable or nonexistent in high-altitude and remote areas, and transmitting high-definition images / videos incurs significant communication bandwidth costs. Once the network is down, the monitoring system immediately fails, preventing real-time alarms and causing significant security delays. 3) Traditional lightweight edge detection-based solutions: Simple object detection (such as traditional YOLOv3 / v4's Bounding Box detection) is used at the edge to select the icy area. Bounding Box detection is a bounding box detection method. Limitations: Object detection can only output rectangular boxes, which cannot accurately match the actual edge contours of the cable and the icing. Therefore, it can only make a qualitative judgment of "whether there is icing" and cannot perform quantitative calculations of icing thickness at the millimeter level, lacking engineering value for guiding de-icing operations.

[0003] Among existing patents, patent CN118967628A discloses a method for monitoring icing on power transmission lines based on AI image segmentation. This solution uses the YOLOv8-Seg model for image segmentation and calculates the icing thickness based on the pixel height ratio. All image analysis, calculation, and alarm logic are completed on a backend server, representing a centralized cloud processing model. This solution is highly dependent on network transmission and cannot function properly in high-altitude mountainous areas or remote areas without network access. Monitoring and alarm functions will completely fail when the network is interrupted. Furthermore, its thickness calculation relies solely on single-column pixel height statistics without considering the overall cable morphology for correction, resulting in significant calculation errors when dealing with complex conditions such as ice nodules, eccentric icing, and cable sag. Patent CN111951520A proposes a system and method for detecting the icing thickness of power transmission line conductors. This solution uses traditional image processing algorithms such as Canny edge detection, Hough line detection, and improved K-means clustering to extract the conductor contour and combines this with a monocular vision mapping model to calculate the icing thickness. Similarly, it deploys the core calculations and early warning functions on a remote platform. Traditional image processing algorithms have weak anti-interference capabilities and are easily affected by rain, snow, debris, and complex backgrounds, resulting in limited contour extraction accuracy. Furthermore, the entire system relies on wireless network data transmission, and loses local early warning capabilities in the event of a network outage, only able to perform basic thickness calculations and unable to form a closed-loop response at the end. Summary of the Invention

[0004] To address the problems of excessive reliance on network bandwidth, high latency, poor quantitative calculation accuracy, and inability to form a local early warning closed loop in existing power transmission line icing closed-loop monitoring, this invention proposes an end-to-end edge-side power transmission line icing closed-loop monitoring method and system.

[0005] The technical solution adopted by this invention to solve the above problems is: An end-to-end edge-side icing closed-loop monitoring method for transmission lines proposed in this invention, comprising: Step 1: Perform baseline calibration and calculate pixel equivalents using the image acquisition module; Step 2: According to the preset image acquisition strategy, real-time monitoring images of the transmission line are acquired through the deployed image acquisition module, and adaptive histogram equalization and normalization preprocessing are performed on the real-time monitoring images to obtain the input tensor that meets the input requirements of the instance segmentation model. Step 3: Input the input tensor into the lightweight instance segmentation model deployed in the edge-side intelligent computing module to extract pixel-level masks of the icy cable region in the real-time monitoring image, thus obtaining the pixel-level mask region of the icy cable. ; Step 4: Traverse the mask region Perform quantitative assessment and thickness calculation of icing, and output the average icing thickness of the current transmission line; Step 5: Compare the average icing thickness with the preset safety threshold, and execute the corresponding local response operation through the multimodal closed-loop output module based on the comparison result.

[0006] Furthermore, step 1 specifically includes: In the un-iced state, obtain the known true physical outer diameter. A baseline image of an uniced power transmission line; Extract the average pixel diameter of unice-covered power transmission lines in the reference image. ; The pixel equivalent is calculated based on the ratio of the known actual physical outer diameter to the average pixel diameter. .

[0007] Furthermore, step 2 specifically includes: Step 2.1: Based on the monitoring scenario of the transmission line, determine whether the image acquisition mode is timed snapshot mode or continuous recording frame extraction mode to obtain real-time monitoring images. In the timed capture mode, high-definition still images are captured at preset time intervals; in the continuous recording and frame extraction mode, the video stream is continuously decoded and key frames are dynamically extracted at preset frame intervals. Step 2.2: Process the real-time monitoring images Adaptive histogram equalization is performed, and the input tensor of the instance segmentation model requires resizing and normalizing the image after adaptive histogram equalization to obtain the input tensor. .

[0008] Furthermore, step 3 specifically includes: The deployed edge lightweight instance segmentation model is pre-optimized and compiled, and the tensor is... Input a pre-optimized and mutated lightweight edge instance segmentation model. After forward propagation, the lightweight edge instance segmentation model outputs an irregular set of pixel-level masks. ; Extract pixel-level mask set The mask area for icing cables is categorized as medium. .

[0009] Furthermore, step 4 specifically includes: Traverse the mask area of ​​the icing cable Calculate the total number of pixels within the region to obtain the pixel area of ​​the icy cable in the two-dimensional projection of the image. ; Extracting mask regions using skeleton extraction algorithms The central skeleton line of the icing cable, wherein the pixel length of the central skeleton line is ; Based on pixel area The pixel length of the center skeleton line is Calculate the average pixel outer diameter of the icing cable ; Average pixel outer diameter based on icing cable and pixel equivalent Calculate the actual physical total outer diameter of the icing cable ; Based on the actual physical total outer diameter of the icing cable and known true physical outer diameter Calculate the average icing thickness of the current transmission line; The formula for calculating the average icing thickness of current transmission lines is as follows: (1); In formula (1), This represents the average icing thickness of the current transmission line.

[0010] Furthermore, step 5 specifically includes: The edge control unit reads the set icing thickness threshold. If the current average icing thickness of the transmission line Less than the ice thickness threshold If the current average icing thickness of the transmission line is within a safe range, then it is in a safe state. The corresponding timestamp is written to the local log of the edge control unit, and the green LED is kept on via GPIO; if the average icing thickness of the current transmission line... Greater than or equal to the set icing thickness threshold If it does, it is in a dangerous state, triggering a closed-loop early warning mechanism.

[0011] Furthermore, the closed-loop early warning mechanism includes visual output, hardware control, and link uploading; Visual output includes: rendering the original image on the local display device, overlaying a masked area of ​​semi-transparent icing cables. The profile is shown, along with the average icing thickness of the current transmission line. Numerical value; Hardware control includes: the edge computing board in the edge control unit outputs a high level through GPIO pins to trigger a relay and drive a bright red LED to flash; at the same time, it outputs a PWM waveform to drive a buzzer to emit a high-frequency alarm sound; Link upload includes: capturing the current alarm image with results and wirelessly transmitting the alarm image back to the backend as a basis for verification.

[0012] Furthermore, this invention also proposes an end-to-end edge-side icing closed-loop monitoring system for power transmission lines, comprising: The image acquisition module is deployed on one of the transmission towers or inspection equipment to acquire real-time monitoring images of the transmission line according to a preset image acquisition strategy. The edge-side intelligent computing module is connected to the image acquisition module and has a built-in hardware acceleration unit for preprocessing real-time monitoring images and extracting pixel-level masks of icy cable areas through a lightweight instance segmentation model. The icing quantitative assessment module, connected to the edge-side intelligent computing module, is used to calculate the pixel area of ​​the iced cable and the pixel length of the central skeleton line based on the pixel-level mask, and to calculate the average icing thickness of the current transmission line in combination with the pre-calibrated pixel equivalent. The multimodal closed-loop output module is connected to the icing quantitative assessment module. It is used to compare the average icing thickness with the preset icing thickness threshold and execute the corresponding local response operation based on the comparison result.

[0013] Furthermore, the edge-side intelligent computing module includes a model optimization and deployment unit, which is used to convert the instance segmentation model into an intermediate representation format and then compile it through a compiler-level toolchain to generate an inference engine file adapted to one of the NPUs or GPUs. The compiler-level toolchain includes at least the TensorRT toolchain and the RKNN toolchain. RKNN is the model format of Rockchip, NPU is a neural network processor, and GPU is a graphics processor.

[0014] Furthermore, the multimodal closed-loop output module includes: The local display unit is used to render and overlay the pixel-level mask outline and average icing thickness value of the iced cable; The audible and visual alarm unit is connected to a relay via GPIO pins and is used to drive LED warning lights and a buzzer to perform audible and visual alarms. The remote communication unit is used to send alarm data, including device identification and icing thickness values, to a remote command center via a low-power wide area network.

[0015] The beneficial effects of this invention are: 1. This invention completely solves the problems of high latency and loss of control when the network is disconnected. It does not rely on cloud servers. All complex AI instance segmentation inference and mathematical modeling are completed within 100 milliseconds at the edge (camera side), achieving true "what you see is what you get". It still has the self-protection capability of sound and light alarm in the case of network disconnection.

[0016] 2. This invention achieves an exponential improvement in algorithm accuracy, abandons the traditional rough estimation method of "rectangular box (Bbox)" for target detection, and innovatively adopts the pixel-level calculus idea of ​​"mask area divided by skeleton length" for instance segmentation, which perfectly avoids the geometric calculation errors caused by cable tilt, sag, and irregular ice nodules.

[0017] 3. The system established by this invention has strong software and hardware ecosystem compatibility. The system is not only compatible with the high-performance NVIDIA edge ecosystem, but also perfectly adaptable to the domestically developed and controllable Rockchip ecosystem, which is conducive to flexible implementation in power grid projects with different cost constraints and domestic substitution requirements.

[0018] 4. This invention significantly reduces construction and operation costs, reducing the bandwidth requirement of transmitting a 2Mbps real-time video stream to transmitting only a few tens of bytes of "thickness parameter data packet", thus saving operators' traffic costs. Attached Figure Description

[0019] Figure 1 A schematic diagram of a closed-loop monitoring method for icing on end-to-end edge-side power transmission lines; Figure 2 This is a structural block diagram of an end-to-end edge-side power transmission line icing closed-loop monitoring system; Figure 3 This diagram illustrates the adaptation and unified deployment architecture of a closed-loop monitoring system for power transmission line icing on different underlying edge processor chips. Detailed Implementation

[0020] Specific implementation method one: Combining Figure 1 This embodiment describes a closed-loop monitoring method for end-to-end edge-side transmission line icing, comprising the following steps: S1: Reference calibration and pixel equivalent calculation; Under clear, ice-free weather conditions, the image acquisition module was used to capture images containing a known true physical outer diameter. A baseline image of an unice-covered transmission line. The average pixel diameter of the unice-covered transmission line in the image is extracted. Calculate pixel equivalent , where pixel equivalent That is, the actual physical size represented by each pixel, in mm / pixel, calculated using the following formula: (1).

[0021] S2: Edge-side visual acquisition and preprocessing; Real-time monitoring images are acquired at the edge according to a preset strategy (timed photo capture or video frame extraction). The image undergoes edge preprocessing, including Adaptive Histogram Equalization (CLAHE) to eliminate the effects of strong light or low illumination. The image is then resized and normalized according to the input tensor requirements of the instance segmentation model to obtain the input tensor. .

[0022] S3: Pixel-level inference based on a lightweight instance segmentation model; tensor Input is fed into an edge lightweight instance segmentation model that has been pre-optimized and compiled using TensorRT or RKNN (such as a model pruned from an architecture like YOLOv8-Seg).

[0023] After forward propagation, the model outputs an irregular set of pixel-level masks. Extract the masked area categorized as "iced cable", and denote it as... Because of instance segmentation, the masked area precisely matches the irregular physical outline of the icy surface, rather than a rough rectangular bounding box.

[0024] S4: Quantitative assessment and thickness calculation of icing; Traversal Calculate the total number of pixels contained within the masked area to obtain the pixel area of ​​the icy cable in the two-dimensional projection of the image. .

[0025] Using skeleton extraction algorithms (such as the Zhang-Suen thinning algorithm) to... The process involves extracting the central skeleton line of the icy cable, calculating its pixel length, and denoting it as... .

[0026] Calculate the average pixel outer diameter of the icing cable : (2); Combined with pixel equivalent in S1 Calculate the actual total physical outer diameter of the icy cable at this time. : (3); Then, the current average icing thickness of the transmission line can be calculated. : (4).

[0027] S5: Local closed-loop arbitration and multimodal early warning; The edge control unit reads the set icing thickness threshold. ,Compare and : like The system determines that it is in a safe state and only writes the timestamp and thickness value to the local log of the edge device, and controls the green LED to stay on via GPIO.

[0028] like The system determines that it is in a dangerous state and triggers a closed-loop early warning mechanism: 1) Visual Output: Rendering the original image on a local micro-screen (such as an LCD / OLED screen). LCD stands for Liquid Crystal Display, and OLED stands for Organic Light Emitting Diode, with a semi-transparent overlay. Outline and red highlighted text display Numerical value; 2) Hardware control: The edge computing board outputs a high level through the GPIO pin to trigger the relay and drive the bright red LED to flash; at the same time, it outputs a PWM waveform to drive the buzzer to emit a high-frequency alarm sound; 3) Link Upload: Capture the alarm image with the current result and compress it significantly, then send it back to the backend via 4G / wireless bridge or other means as a basis for review.

[0029] Specific Implementation Method Two: Combining Figure 2 and Figure 3 This embodiment will be described as follows: Figure 2 As shown, the structure of the end-to-end edge-side power transmission line icing closed-loop monitoring system described in this embodiment includes: The system comprises an image acquisition module, an edge-side intelligent computing module, an icing quantitative assessment module, and a multimodal closed-loop output module. The operational flow of each module during system operation is as follows: Figure 3 As shown. The functions of each module are as follows: The image acquisition module is deployed on one of the transmission towers or inspection equipment to acquire real-time monitoring images of the transmission line according to a preset image acquisition strategy, and supports dual-mode acquisition.

[0030] The dual modes include timed snapshot mode and continuous recording frame extraction mode. Timed snapshot mode: triggers high-definition still image acquisition according to the specified frequency set by the system (such as once every 10 minutes). Continuous recording frame extraction mode: continuously decodes the video stream and dynamically extracts key frames according to the set frame interval. It is suitable for fast-moving inspection scenarios (such as drones and inspection robots).

[0031] The edge-side intelligent computing module connects to the image acquisition module and incorporates a high-efficiency AI acceleration unit (NPU / GPU) optimized for edge device instruction sets, equipped with a lightweight visual instance segmentation model. This module receives data from the image acquisition module and performs pixel-level classification and mask extraction of "uniced high-voltage transmission lines," "iced cables," and "background" in the image.

[0032] The icing quantitative assessment module, connected to the edge-side intelligent computing module, is the core algorithm execution unit of the system. It receives mask data from the intelligent computing module, extracts the morphological features (such as projected area and central axis length) of iced and uniced cables through pixel-level geometric calculation algorithms, and calculates the actual average icing thickness by combining it with a spatial scale conversion model. The multimodal closed-loop output module connects to the icing quantitative assessment module and compares the thickness value output by the icing quantitative assessment module with a preset safety threshold. If the threshold is exceeded, the local peripheral hardware is triggered to perform an end-to-end direct response, including but not limited to: driving the local screen to highlight the icing parameters and alarm box, triggering the high-brightness LED warning light to flash, driving a high-power buzzer to perform an acoustic alarm, and simultaneously sending an alarm command with a very small amount of data to the command center through a low-power wide area network (such as NB-IoT / short message). NB-IoT stands for Narrow Band Internet of Things.

[0033] The multimodal closed-loop output module includes: The local display unit is used to render and overlay the pixel-level mask outline and average icing thickness value of the iced cable; The audible and visual alarm unit is connected to a relay via GPIO pins and is used to drive LED warning lights and a buzzer to perform audible and visual alarms. The remote communication unit is used to send alarm data, including device identification and icing thickness values, to a remote command center via a low-power wide area network.

[0034] Example 1 To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description is provided in conjunction with specific embodiments. The closed-loop monitoring system for power transmission line icing established by this invention, in its physical deployment, includes an industrial-grade all-weather waterproof camera connected to an edge computing host (Edge AI Box) via a MIPI / CSI interface or RTSP network streaming protocol. MIPI stands for Mobile Industry Processor Interface, a development interface of the MIPI Alliance; CSI stands for Camera Serial Interface, also a development interface of the MIPI Alliance; and RTSP stands for Real Time Streaming Protocol. The edge computing host motherboard integrates an NPU or GPU, pre-installs a Linux operating system such as Ubuntu / Debian, and mounts a background daemon process written in C++ or Python.

[0035] The daemon process consists of three threads: the acquisition thread, the inference calculation thread, and the hardware control thread.

[0036] 1) Acquisition thread, responsible for sending instructions to the camera to acquire images; 2) The inference computation thread calls the YOLOv8-Seg instance segmentation model optimized with INT8 quantization or FP16 to extract the mask, and substitutes it into the aforementioned physical formula. Perform pure mathematical operations; 3) The hardware control thread controls the external GPIO pin levels by reading the motherboard's Sysfs file system. External relays are connected to these pins, which in turn connect to a 12V or 24V strobe light, a high-decibel buzzer, and an LCD screen for local debugging. This constitutes a fully self-consistent closed-loop monitoring station.

[0037] Example 2 Fixed monitoring of ultra-high voltage transmission towers in extremely cold conditions based on NVIDIA's edge ecosystem: Scenario: A 500kV ultra-high voltage transmission corridor, located in an extremely cold mountainous area with year-round wind and snow, which places extremely high demands on the computational performance of the algorithm, and requires precise capture of even the smallest ice accretions.

[0038] Hardware configuration: Employs a high-magnification optical zoom gimbal camera, with the edge motherboard using the NVIDIA Jetson OrinNano development board.

[0039] Implementation details: 1) As it is an ultra-high voltage tower, power supply relies on large-capacity solar panels and batteries on the tower. The NVIDIA board power consumption is set to 15W mode.

[0040] 2) The camera outputs a 1080P video stream via RTSP. It employs a "continuous recording followed by frame extraction" mode, where the Jetson board internally calls the GStreamer hardware decoder to extract one frame per second and load it into memory.

[0041] 3) On the model side, after converting the trained instance segmentation model into ONNX format, it is compiled into an engine file using the NVIDIA TensorRT toolchain for FP16 half-precision accelerated inference on the GPU.

[0042] 4) Due to the high image resolution and abundant computing power, an edge smoothing algorithm is introduced during the pixel-level calculation of icing to eliminate the influence of snowflake noise on the mask edge.

[0043] 5) Indicator lights and external loudspeaker. When the calculated ice thickness reaches 15mm, Jetson triggers the loudspeaker via a pin to play a pre-recorded voice message: "Attention! This line is severely iced!" Simultaneously, the searchlight LEDs flash downwards to warn passing forest rangers and inspection personnel. The screen is located inside the chassis, allowing maintenance workers to view real-time data while climbing the tower for maintenance.

[0044] Example 3 Low-power monitoring of remote power distribution networks based on the domestic Rockchip ecosystem: Scenario: 10kV / 35kV medium and low voltage rural power grids and distribution network poles, which are extremely numerous and highly sensitive to the cost and power consumption of system equipment, and there is a need for domestic alternatives and independent control.

[0045] Hardware configuration: It adopts a fixed-focus, low-power camera, and the edge motherboard uses the domestic Rockchip RK3588 (or RK3568, etc.) industrial core board.

[0046] Implementation details: 1) Due to limited system power, a small-sized solar panel and lithium battery pack are used. At this time, the acquisition strategy is switched to "specified frequency shooting", that is, the edge motherboard is in deep sleep state by default, and the internal RTC timer wakes up the motherboard every 30 minutes.

[0047] 2) Upon wake-up, the camera captures an image the instant it powers on. The RK3588 utilizes its powerful internal NPU (with a computing power of 6 TOPS) for inference.

[0048] 3) To adapt to domestic NPUs, the instance segmentation model needs to be PTQ (post-training quantization) on the PC using RKNN-Toolkit, converting it into an INT8 precision .rknn model format. This results in extremely low power consumption per inference, requiring only tens of milliseconds.

[0049] 4) After calculating the ice thickness, the RK3588 connects to a low-power electronic ink screen (E-inkScreen) via serial port to refresh the current cable thickness value and a simple outline diagram on the ink screen (the image is still retained after power failure, making it convenient for the line inspector to view with binoculars).

[0050] 5) Closed-loop alarm: If the threshold is exceeded, the connected low-power buzzer is woken up to sound intermittently. Then, an alarm message of several tens of bytes (including device ID, ice thickness of 20mm, and battery level) is immediately sent through the NB-IoT module. The system then powers off again and enters sleep mode, greatly extending battery life.

[0051] Example 4 Dynamic viewpoint closed-loop monitoring based on drones / mobile inspection robots: Scenario: Quadruped robot dogs, rail-mounted inspection robots, or multi-rotor drones are dynamically inspecting power transmission lines. The viewpoint is constantly changing, the background is complex, and there are no fixed reference points.

[0052] Hardware combination: lightweight drone payload pod / robot gimbal, built-in high frame rate industrial camera, equipped with a highly integrated lightweight computing stick or edge AI coprocessor.

[0053] Implementation details: 1) This is a more complex dynamic implementation. Video streams are acquired using the robot's "continuous recording" function.

[0054] 2) Because the camera moves relative to the cable (e.g., during drone flight), the reference calibration in step S1 cannot be fixed. This embodiment introduces a binocular depth camera (RGB-D) or combines it with LiDAR ranging. The edge computing module obtains the depth distance between the camera and the cable in real time. .

[0055] 3) Dynamically corrected pixel equivalent formula: based on camera intrinsic parameters (focal length) ) and real-time depth Dynamically calculate the pixel equivalent of the current frame. This enables precise pixel-to-physical-size conversion during movement.

[0056] 4) After the robot's edge computing module runs the instance segmentation and calculates the thickness using the "area / length" formula, it no longer relies on external LEDs or buzzers for large-scale physical alarms.

[0057] 5) The closed-loop output method changes to: On the robot operator's local remote control screen (or image transmission screen), using edge device OSD technology, each frame of the original video stream is overlaid with a red and green semi-transparent mask layer and a floating thickness value label. The operator can intuitively see which cable segment is marked red and alarmed through the remote control screen (which belongs to the extended display output link of the edge system), thereby realizing an end-to-end closed-loop monitoring system of "flying, calculating, watching, and alarming simultaneously" at the drone / robot operation site.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A closed-loop monitoring method for icing on end-to-end edge-side transmission lines, characterized in that, include: Step 1: Perform baseline calibration and calculate pixel equivalents using the image acquisition module; Step 2: According to the preset image acquisition strategy, real-time monitoring images of the transmission line are acquired through the deployed image acquisition module, and adaptive histogram equalization and normalization preprocessing are performed on the real-time monitoring images to obtain the input tensor that meets the input requirements of the instance segmentation model. Step 3: Input the input tensor into the lightweight instance segmentation model deployed in the edge-side intelligent computing module to extract pixel-level masks of the icy cable region in the real-time monitoring image, thus obtaining the pixel-level mask region of the icy cable. ; Step 4: Traverse the mask region Perform quantitative assessment and thickness calculation of icing, and output the average icing thickness of the current transmission line; Step 5: Compare the average icing thickness with the preset safety threshold, and execute the corresponding local response operation through the multimodal closed-loop output module based on the comparison result.

2. The end-to-end edge-side icing closed-loop monitoring method for transmission lines according to claim 1, characterized in that, Step 1 specifically includes: In the un-iced state, obtain the known true physical outer diameter. A baseline image of an un-iced power transmission line; Extract the average pixel diameter of the un-iced power transmission line in the reference image. ; The pixel equivalent is calculated based on the ratio of the known actual physical outer diameter to the average diameter of the pixel. .

3. The end-to-end edge-side icing closed-loop monitoring method for transmission lines according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Based on the monitoring scenario of the transmission line, determine whether the image acquisition mode is timed snapshot mode or continuous recording frame extraction mode to obtain real-time monitoring images. In the timed capture mode, high-definition still images are captured at preset time intervals; in the continuous recording frame extraction mode, the video stream is continuously decoded and key frames are dynamically extracted at preset frame intervals. Step 2.2: Process the real-time monitoring images Adaptive histogram equalization is performed, and the input tensor of the instance segmentation model requires resizing and normalizing the image after adaptive histogram equalization to obtain the input tensor. .

4. The end-to-end edge-side icing closed-loop monitoring method for transmission lines according to claim 1, characterized in that, Step 3 specifically includes: The deployed edge lightweight instance segmentation model is pre-optimized and compiled, and the tensor is... Input a pre-optimized and mutated lightweight edge instance segmentation model. After forward propagation, the lightweight edge instance segmentation model outputs an irregular set of pixel-level masks. ; Extract pixel-level mask set The mask area for icing cables is categorized as medium. .

5. The end-to-end edge-side icing closed-loop monitoring method for transmission lines according to claim 2, characterized in that, Step 4 specifically includes: Traverse the mask area of ​​the icing cable Calculate the total number of pixels within the region to obtain the pixel area of ​​the icy cable in the two-dimensional projection of the image. ; Extracting mask regions using skeleton extraction algorithms The central skeleton line of the icing cable, wherein the pixel length of the central skeleton line is ; Based on pixel area The pixel length of the center skeleton line is Calculate the average pixel outer diameter of the icing cable ; Average pixel outer diameter based on icing cable and pixel equivalent Calculate the actual physical total outer diameter of the icing cable ; Based on the actual physical total outer diameter of the icing cable and known true physical outer diameter Calculate the average icing thickness of the current transmission line; The formula for calculating the average icing thickness of current transmission lines is as follows: （1）； In formula (1), This represents the average icing thickness of the current transmission line.

6. The end-to-end edge-side icing closed-loop monitoring method for transmission lines according to claim 1, characterized in that, Step 5 specifically includes: The edge control unit reads the set icing thickness threshold. If the current average icing thickness of the transmission line Less than the ice thickness threshold If the current average icing thickness of the transmission line is within a safe range, then it is in a safe state. The corresponding timestamp is written to the local log of the edge control unit, and the green LED is kept on via GPIO. GPIO stands for General Purpose Input Output, and LED is a light-emitting diode. The average icing thickness of the current transmission line... Greater than or equal to the set icing thickness threshold If it does, it is in a dangerous state, triggering a closed-loop early warning mechanism.

7. The end-to-end edge-side icing closed-loop monitoring method for transmission lines according to claim 6, characterized in that, The closed-loop early warning mechanism includes visual output, hardware control, and link uploading; Visual output includes: rendering the original image on the local display device, overlaying a masked area of ​​semi-transparent icing cables. The profile is shown, along with the average icing thickness of the current transmission line. Numerical value; Hardware control includes: the edge computing board in the edge control unit outputs a high level through GPIO pins to trigger a relay and drive a bright red LED to flash; at the same time, it outputs a PWM waveform to drive a buzzer to emit a high-frequency alarm sound, where PWM stands for pulse width modulation. Link upload includes: capturing the current alarm image with results and wirelessly transmitting the alarm image back to the backend as a basis for verification.

8. An end-to-end edge-side transmission line icing closed-loop monitoring system, applied to the end-to-end edge-side transmission line icing closed-loop monitoring method according to any one of claims 1-7, characterized in that, include: The image acquisition module is deployed on one of the transmission towers or inspection equipment to acquire real-time monitoring images of the transmission line according to a preset image acquisition strategy. The edge-side intelligent computing module is connected to the image acquisition module and has a built-in hardware acceleration unit for preprocessing real-time monitoring images and extracting pixel-level masks of the icy cable area through a lightweight instance segmentation model. The icing quantitative assessment module is connected to the edge-side intelligent computing module. It is used to calculate the pixel area of ​​the icing cable and the pixel length of the central skeleton line based on the pixel-level mask, and to calculate the average icing thickness of the current transmission line in combination with the pre-calibrated pixel equivalent. The multimodal closed-loop output module is connected to the icing quantitative assessment module and is used to compare the average icing thickness with a preset icing thickness threshold and perform corresponding local response operations based on the comparison results.

9. The end-to-end edge-side icing closed-loop monitoring system for power transmission lines according to claim 8, characterized in that, The edge-side intelligent computing module includes a model optimization and deployment unit, which is used to convert the instance segmentation model into an intermediate representation format and then compile it into an inference engine file adapted to one of NPU and GPU through a compiler-level toolchain. The compiler-level toolchain includes at least the TensorRT toolchain and the RKNN toolchain. RKNN is the model format of Rockchip, NPU is a neural network processor, and GPU is a graphics processor.

10. The end-to-end edge-side icing closed-loop monitoring system for power transmission lines according to claim 8, characterized in that, The multimodal closed-loop output module includes: The local display unit is used to render and overlay the pixel-level mask outline and average icing thickness value of the iced cable; The audible and visual alarm unit is connected to a relay via GPIO pins and is used to drive LED warning lights and a buzzer to perform audible and visual alarms. The remote communication unit is used to send alarm data, including device identification and icing thickness values, to a remote command center via a low-power wide area network.

Images