A wind power remote sensing monitoring system and a monitoring method

Abstract

The application belongs to the technical field of wind power monitoring, and discloses a wind power remote sensing monitoring system and a monitoring method. The system is autonomously flown in a wind farm by a robot gliding module and is stably attached to a fan blade. Data is collected by a radar and a depth camera in a multi-modal data fusion module, the data is input into a damage detection module after fusion processing, a three-dimensional model of the blade is generated, and damage data is obtained according to the three-dimensional model of the blade. The damage detection module constructs a three-dimensional model of the blade by using the fused data, and obtains damage data according to the three-dimensional model of the blade. The damage evolution prediction module is used to predict the service life of the fan blade and give early warning. The dynamic path planning and autonomous control module optimizes the inspection strategy according to the environmental temperature and humidity, wind field disturbance data, damage data and service life prediction data. Through the mutual cooperation of the above modules, the whole-chain wind power monitoring function from data acquisition, damage identification, service life prediction to intelligent decision-making is realized.

Description

Technical Field

[0001] This invention belongs to the field of wind power monitoring technology, and in particular relates to a wind power remote sensing monitoring system and monitoring method. Background Technology

[0002] Wind farms are typically located in remote areas, making it difficult and time-consuming to transport the necessary maintenance equipment and personnel. Onshore wind turbines are also more susceptible to extreme weather events such as heavy rain, lightning strikes, and hail. Traditional wind turbine inspections primarily rely on visual checks, including high-powered telescopes, high-altitude descent inspections, and inspections from maintenance platforms. Due to the high position of the turbine blades, minor faults such as cracks are difficult to detect. Traditional operations require shutdown for inspection, leading to operational interruptions and losses in power generation.

[0003] In recent years, robotics technology has been widely used in the monitoring of onshore wind power facilities. However, existing technologies have the following problems:

[0004] (1) Existing monitoring systems mostly use single vision or radar technology, lacking multimodal data fusion capabilities, making it difficult to accurately identify microcracks and structural damage warnings of wind turbine blades. Single vision technology is prone to image distortion under complex weather conditions such as strong light, rain and fog, while single radar technology, although it can penetrate surface dirt, lacks sufficient accuracy in analyzing the direction and depth of cracks.

[0005] (2) The robot has defects in maneuverability in high-altitude flow field environment, especially in positioning accuracy and wind resistance under the disturbance of wind turbine wake, which makes it difficult for existing technology to achieve dynamic inspection without stopping the machine.

[0006] (3) In terms of wind power operation and maintenance, the power generation loss caused by traditional shutdown detection has become a pain point in the industry. Existing detection equipment mostly relies on manual climbing or fixed sensors, which has low deployment efficiency and limited coverage, making it difficult to meet the high-frequency detection needs of large wind farms. In addition, existing equipment has poor adaptability to extreme weather conditions such as icing, resulting in a decrease in the reliability of monitoring data. These problems seriously restrict the level of intelligent operation and maintenance of wind power facilities.

[0007] Therefore, there is an urgent need to design an efficient and accurate wind power monitoring system and method. Summary of the Invention

[0008] The purpose of this invention is to solve the problems of insufficient identification accuracy, insufficient robot stability, and poor adaptability in existing monitoring systems, and to provide a wind power remote sensing monitoring system and monitoring method.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A wind power remote sensing monitoring system, comprising: The robot gliding module is used to enable robot movement and stable attachment. The module includes a robot, a glider, and a magnetic-gas composite adsorption unit. The glider is mounted on the robot for movement and climbing. A gripping device with a torque sensor is installed on the glider to measure the torque of the gripping device in real time. The magnetic-gas composite adsorption unit enables the robot to stably attach to the blade. This unit includes a neodymium iron boron permanent magnet array unit and a negative pressure adsorption unit. The neodymium iron boron permanent magnet array unit provides the initial adsorption force, while the negative pressure adsorption unit ensures stable attachment of the glider to the blade surface. The multimodal data fusion module is used to acquire radar data of internal blade cracks and visual image data of the blade surface, and fuse the radar data and visual image data to obtain fused data. The multimodal data fusion module includes radar, depth camera, meteorological detection unit and communication unit. The radar is used to acquire radar data of internal blade cracks and to measure the three-dimensional wake velocity field and wind field disturbance data of the wind turbine wake field in real time. The depth camera is used to acquire visual image data of the blade surface and to measure the clamping diameter and gripping angle of the clamping device in real time. The meteorological detection unit includes an anemometer and temperature and humidity sensor. The anemometer is used to detect the wind field velocity in real time. The temperature and humidity sensor is used to acquire ambient temperature and humidity data. The damage detection module uses fused data to construct a three-dimensional damage model of the blade and acquire damage data; The damage evolution prediction module is used to predict the lifespan of wind turbine blades, obtain lifespan prediction data, and provide early warnings. The dynamic path planning and autonomous control module optimizes the inspection strategy based on environmental temperature and humidity, wind field disturbance data, damage data, and life prediction data. Anti-interference and communication modules ensure the stability and reliability of data transmission.

[0010] A wind power remote sensing monitoring method includes the following steps: S1. Robot Movement and Attachment: The robot's autonomous flight is controlled by a gliding module. The robot's flight attitude is adjusted using a blended wing-body configuration design and a PID control model. Once the robot reaches the vicinity of the wind turbine blade surface, the clamping force of the gripping device is adjusted using torque calculated by a depth camera and real-time torque measured by a torque sensor. The robot is then attached to the wind turbine blade surface via a magnetic-gas composite adsorption unit. S2. Multimodal Data Synchronous Acquisition and Fusion: After attachment, radar and a depth camera are used for synchronous sampling to obtain radar data of internal blade cracks and visual image data of the blade surface. The radar data and visual image data are aligned using a spatiotemporal registration model, and a dynamic weighted fusion algorithm is used to fuse the radar data and visual image data to obtain fused data. S3. Crack Depth Quantization and 3D Modeling: Based on the fused data, the scattering characteristics of wind turbine blade surface cracks to radar signals are analyzed using a radar scattering model. Crack length, crack width, crack depth, crack center location, and crack direction are extracted. The 3D morphology of the crack is inverted, and a 3D model of the blade is generated using the fused data. Damage data is then obtained based on the 3D model of the blade. S4. Damage evolution modeling and life prediction: The crack growth rate is quantified based on the improved crack propagation model; and the crack length and ambient temperature and humidity are time-series modeled using an LSTM network to predict the life of the wind turbine blades, obtain life prediction data, adapt to different models through a transfer learning framework, generate life reports and provide early warnings. S5. Dynamic Path Optimization: Based on environmental temperature and humidity, damage data, and life prediction data, the inspection path is optimized by improving the A* algorithm and Q learning strategy. At the same time, the damage data and predicted life data are input into the operation and maintenance management platform through the anti-interference and communication module to obtain maintenance priority suggestions.

[0011] Furthermore, step S1 is detailed as follows: S101. Control the robot's autonomous flight through the glider in the robot's gliding module; S102. During the initial or stable flight phase, the robot adjusts its flight attitude using a blended wing-body configuration design method, where the glide ratio maximization objective function is used in this method. for: In the formula, the lift coefficient and drag coefficient Through angle of attack respectively It is obtained by combining trigonometric functions and exponential functions; During the dynamic flight phase, the three-dimensional wake velocity field and wind disturbance data of the wind turbine wake are acquired by radar; using a PID control model, the pitch angle correction is calculated based on the three-dimensional wake velocity field to adjust the robot's flight attitude; the PID control model is specifically as follows: in, The three-dimensional wake velocity field, The integral of the wake velocity field norm; S103. After the robot reaches the surface of the wind turbine blade, it uses a depth camera to measure the clamping diameter of the glider in real time. and the angle of capture Real-time torque calculation ,in , Torque coefficient; Simultaneously, the torque of the clamping device is measured in real time using a torque sensor. According to torque and torque Obtaining torque difference Through torque difference The gripping force of the clamping device is adjusted based on the set threshold. The neodymium iron boron permanent magnet array unit, utilizing a magnetic-gas composite adsorption unit, provides initial adsorption force in conjunction with a negative pressure adsorption unit, ensuring the reliable adhesion of the glider to the blade surface; the minimum clamping force of the magnetic-gas composite adsorption unit must meet the following requirements: In the formula, The coefficient of friction of the adsorption surface, The effective working area of ​​the magnetic-gas composite adsorption unit, For pressure coefficient, For robot quality, This is the blade tilt angle.

[0012] Furthermore, step S2 specifically involves: S201, Multimodal Data Synchronous Acquisition: After the robot attaches to the blade, the radar and depth camera sample synchronously in time and space at a frequency of 100Hz; the radar emits millimeter waves to penetrate the dirt layer on the blade surface and obtain radar data of internal cracks in the blade; the depth camera obtains visual image data of blade surface details, scratches or minor deformations. S202, Spatiotemporal Registration and Error Correction: Three-dimensional coordinates are obtained through the checkerboard calibration method, and radar data and visual image data are aligned using a spatiotemporal registration model; S203, Dynamic Weighted Data Fusion: The radar data and visual image data processed in step S202 are fused using a dynamic weighted fusion algorithm to obtain fused data.

[0013] Furthermore, the chessboard calibration method and spatiotemporal registration model in step S202 are as follows: Chessboard grid marking method: A high-precision checkerboard calibration board is deployed in a static scene; the radar coordinate system and visual coordinate system are calibrated using radar and a depth camera, and the intrinsic parameters of the depth camera are also calibrated; the coordinates of the checkerboard corner points in the radar coordinate system and the visual coordinate system are recorded; the pixel coordinates of the checkerboard corner points are extracted through image processing, and then back-projected onto the three-dimensional coordinates in the visual coordinate system using the depth camera intrinsic parameters. ; Extracting the three-dimensional coordinates of checkerboard corner points from point cloud data using millimeter-wave radar. ; Constructing a spatiotemporal registration model: in, and Here is the Euler angle rotation matrix. The time synchronization compensation term is used to correct coordinate deviations caused by sampling time differences; This indicates the positional offset between the visual sensor and the radar; Yaw angle; It is the pitch angle.

[0014] Furthermore, the dynamic weight fusion algorithm in step S203 is as follows: The weighting coefficients are dynamically adjusted based on the signal-to-noise ratio. In the formula, For the dynamic weighting coefficients of radar data, For radar data Perform normalization processing; These are the dynamic weighting coefficients for visual image data. For visual image data Perform Laplacian sharpening; , ;in, The signal-to-noise ratio of radar data. The signal-to-noise ratio of visual image data; For radar signal power, Noise power; This represents the maximum brightness value of the image. This represents the standard deviation of noise.

[0015] Furthermore, in step S3, the radar scattering model for: in, For radar wavelength, The complex dielectric constant of fiberglass is... For the first Crack length, The location of the crack center. For penetration depth, This represents the total number of crack segments. Wave number; The imaginary unit; These are the spatial coordinates in the horizontal direction.

[0016] Furthermore, step S4 specifically involves: S401. Using the normalized crack length, temperature, and humidity, and an improved crack propagation model based on the modified Paris formula, analyze the temperature and humidity parameters based on the improved crack propagation model to quantify the crack growth rate. ; In the formula, The length of the crack; For ambient temperature and humidity; The rate constant is related to ambient temperature and humidity; Indices related to materials or the environment; The improved crack propagation model is as follows: In the formula, This is expressed as the effective stress intensity factor amplitude, quantifying the crack propagation driving force. The larger the crack, the faster the crack propagation rate. The faster; This represents the maximum cyclic stress. S402. A 64-32-16 layer LSTM network is used to perform time-series modeling of crack length, temperature, and humidity to predict the lifespan of wind turbine blades and obtain lifespan prediction data; the 64-32-16 layer LSTM network is as follows: In the formula, The temporal characteristics of crack length, temperature, and humidity; The correlation weights between crack evolution and crack length, temperature, and humidity; For the period of time , The current crack length and temperature; The bias is used to adjust the model output; S403 introduces a transfer learning framework to adapt to different models. The system combines environmental temperature and humidity, damage data, and life prediction data to generate a remaining life report and triggers a fault warning 72 hours in advance. The transfer learning framework is as follows: In the formula It is a task-related loss function; It is a domain-dependent loss function; It is the loss function of the regularization term.

[0017] Furthermore, the improved A* algorithm in step S5 is as follows: The cumulative cost function of the improved A* algorithm is defined as: In the formula, Represents a node The square of the magnitude of the three-dimensional wind speed vector at that location; This represents the cumulative cost from the parent node to the starting point. To move to node The increase in energy consumption, of which , air density; For nodes Wind speed at the location; This is the aerodynamic drag coefficient; The equivalent windward area of ​​the robot; From the parent node to the current node The distance traveled.

[0018] Furthermore, the improved algorithm in step S5 introduces the Q-learning reinforcement learning framework, with the following Q-value update rule: In the formula, For instant rewards, This is the current state. For action, For the new state, This is a new action.

[0019] The beneficial effects of this invention are: (1) In view of the problem that the detection capability of a single sensor in the prior art is insufficient under complex weather conditions such as strong light and rain and fog, the present invention effectively overcomes the limitations of a single mode by using the deep fusion technology of radar imaging and high-resolution vision (depth camera) combined with dynamic weight data fusion method, significantly improving the identification accuracy of microcracks and structural damage, and reducing the risk of misjudgment caused by environmental interference.

[0020] (2) In view of the robot’s lack of maneuverability under the disturbance of the wind turbine wake, the present invention adopts the collaborative control of the robot gliding module and the multi-degree-of-freedom clamping system, combined with the dynamic path planning algorithm and the wind-resistant magnetic-gas composite adsorption unit to ensure that the system can stably attach and autonomously transfer in extreme wind speed environment, thereby supporting continuous monitoring of the wind turbine without stopping, and greatly reducing the power generation loss caused by detection.

[0021] (3) In view of the problems of low efficiency and limited coverage of traditional shutdown detection, this invention realizes rapid and efficient inspection of multiple wind turbine units through the collaboration of autonomous transfer mechanism and intelligent path planning technology. At the same time, it predicts the life of components in advance through damage evolution prediction model, reduces the frequency of unplanned shutdowns, and significantly improves the operation and maintenance efficiency and economy of large wind farms.

[0022] (4) By replacing manual high-altitude operations with non-contact remote sensing monitoring technology, the risk of personnel falling has been completely eliminated; and by combining damage evolution models to provide early warning of potential faults, the accident rate has been effectively reduced, providing a reliable guarantee for the safe operation of wind farms. Attached Figure Description

[0023] Figure 1 This is a flowchart of the wind power remote sensing monitoring method provided in the embodiments of the present invention; Figure 2 This is a flowchart of the gliding deployment process provided in an embodiment of the present invention; Figure 3 This is a flowchart of the multimodal data fusion process provided in an embodiment of the present invention; Figure 4 This is a flowchart of intelligent damage prediction provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] The application principle of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0026] This application constructs an integrated air-space-ground intelligent monitoring system for wind power facilities through the synergy of four core technologies: biomimetic robotic gliding structure, multimodal data fusion, intelligent damage prediction, and dynamic path planning. The system adopts a closed-loop architecture of "perception-decision-execution," achieving stable attachment to high-altitude flow fields based on a biomimetic glider, realizing sub-millimeter-level damage detection through data fusion of radar and depth cameras, and predicting the remaining lifespan of components by combining damage evolution models, ultimately forming a full-chain solution covering "data acquisition-analysis-early warning-decision".

[0027] A wind power remote sensing monitoring system includes: a robot gliding module, a multimodal data fusion module, a damage detection module, a damage evolution prediction module, a dynamic path planning and autonomous control module, and an anti-interference and communication module. Details are as follows:

[0028] The robot gliding module enables robot movement and stable attachment. It comprises a robot, a glider, and a magnetic-gas composite adsorption unit. The robot is equipped with a glider for movement and climbing; the glider has a gripping device with a torque sensor for real-time torque measurement. The magnetic-gas composite adsorption unit ensures stable attachment of the robot to the blade. This unit includes a neodymium iron boron permanent magnet array and a negative pressure adsorption unit. The neodymium iron boron permanent magnet array provides initial adsorption force, while the negative pressure adsorption unit ensures stable attachment of the robot to the blade surface.

[0029] The multimodal data fusion module is used to acquire radar data of internal blade cracks and visual image data of the blade surface, and then fuse the radar data and visual image data to obtain fused data. The multimodal data fusion module includes a radar, a depth camera, a meteorological detection unit, and a communication unit. The radar is used to acquire radar data of internal blade cracks and to measure the three-dimensional wake velocity field and wind field disturbance data of the wind turbine wake in real time; the depth camera is used to acquire visual image data of the blade surface and to measure the clamping diameter and gripping angle of the clamping device in real time. The meteorological detection unit includes an anemometer and a temperature and humidity sensor; the anemometer is used to detect the wind field velocity in real time; the temperature and humidity sensor is used to acquire ambient temperature and humidity data.

[0030] The damage detection module uses fused data to construct a three-dimensional damage model of the blade and acquire damage data.

[0031] The damage evolution prediction module is used to predict the lifespan of wind turbine blades, obtain lifespan prediction data, and provide early warnings.

[0032] The dynamic path planning and autonomous control module optimizes inspection strategies based on environmental temperature and humidity, wind field disturbance data, damage data, and lifespan prediction data.

[0033] The anti-interference and communication module ensures the stability and reliability of data transmission, copes with signal interference under complex weather conditions, provides real-time communication support for all modules, and ensures the coordinated operation of all parts of the system.

[0034] like Figures 1-4 As shown, a wind power remote sensing monitoring method specifically includes the following steps: S1. Robot movement and attachment: The system controls the robot's autonomous flight through a robotic gliding module. The robot carries a glider and adjusts its flight attitude based on a blended wing-body design and a PID control model to ensure stable gliding in highly turbulent environments. The glider is equipped with a torque sensor and a magnetic-gas composite adsorption unit. The torque sensor provides real-time feedback on the torque generated by the gripping device to adjust the gripping force, while the magnetic-gas composite adsorption unit ensures reliable attachment of the robot to the wind turbine blade surface.

[0035] The specific technical solution is as follows: S101. The operator starts the wind power remote sensing monitoring system, which performs a self-check and initializes various parameters. The robot's gliding module controls the deployment of the glider, and the glider's surface is fully extended, forming a stable blended wing-body configuration, enabling the robot to fly autonomously.

[0036] S102. During the initial or stable flight phases, the robot adjusts its flight attitude using a blended wing-body configuration design to ensure stable gliding in highly turbulent environments. The blended wing-body configuration design employed by the glider effectively matches the nonlinear characteristics of the high-altitude flow field, enhancing stability and wind resistance during gliding.

[0037] The core of the blended wing-body configuration design method is to establish an objective function that maximizes the glide ratio, by optimizing the lift coefficient. and drag coefficient Modeling is performed to match the nonlinear characteristics of the high-altitude flow field, and an objective function for maximizing the glide ratio is established. : In the formula, The angle of attack was measured using a depth camera. The nonlinear characteristic specifically refers to the lift coefficient in the high-altitude flow field. and drag coefficient The complex and non-proportional relationship exhibited with changes in angle of attack. A genetic algorithm (GA) was used to globally optimize the objective function of maximizing the glide ratio to obtain the optimal individual. The fitness value (glide ratio) was calculated for each individual. Fitness value The larger the size, the better the individual. Individual refers to the configuration of a glider with specific design parameters (such as angle of attack, lift coefficient, drag coefficient, etc.).

[0038] The global optimization of the genetic algorithm (GA) uses a tournament selection method: 200 initial individuals (angle of attack) are randomly generated. The offspring are evenly distributed within the range of 0° to 30°. Each time, a number of individuals (e.g., 5) are randomly selected from the population, and the individual with the highest fitness is chosen as the parent. This process is repeated until a sufficient number of parents are selected (ensuring a stable population size). Arithmetic crossover is performed on the selected parents with a crossover probability of 0.85, and Gaussian mutation is performed on the offspring with a mutation probability of 0.02. The newly generated offspring are merged with the parents, and the 200 individuals with the highest fitness are retained for the next generation. This process is repeated until 500 generations have been reached, at which point the optimal individual is output. Then the corresponding glide ratio can be calculated. Glide ratio It is 42% better than the traditional NACA airfoil.

[0039] The radar samples at a frequency of 100Hz to monitor the three-dimensional wake velocity field and wind disturbance data of the wind turbine wake in real time. The three-dimensional wake velocity field is then used to optimize the glider's flight path, ensuring its stability under wake disturbances. During dynamic flight, especially when encountering wake disturbances or strong wind changes, a PID control model is used to calculate the pitch angle correction based on the three-dimensional wake velocity field, adjusting the robot's flight attitude to ensure stable flight even under wake disturbances. Under level 12 wind conditions, this control model can suppress the attitude angle error to within ±0.8°.

[0040] Among them, the PID control model is: in, The three-dimensional wake velocity field, Three-dimensional wake velocity field Time derivative, The integral of the wake velocity field norm, The second time derivative of the wake field is sampled by radar at a frequency of 100Hz, and the control model outputs the pitch angle correction. .

[0041] S103. After the robot carrying the glider reaches the vicinity of the wind turbine blade surface, it uses a depth camera to measure the clamping diameter of the glider's clamping device in real time. and the angle of capture According to the clamping diameter and the angle of capture Real-time torque calculation The torque formula is as follows:

[0042] in, Torque, This is the torque coefficient.

[0043] The torque of the glider's clamping device is measured in real time using a torque sensor. According to torque and torque Obtaining torque difference The torque difference is compared with the set threshold. If it exceeds the threshold, the gripping force of the clamping device is adjusted to ensure that the clamping force is neither too high (to prevent blade damage) nor too low (to prevent detachment).

[0044] Simultaneously, the glider's magnetic-gas composite adsorption unit aligns with the blade surface, ensuring stable adhesion. The magnetic-gas composite adsorption unit on the glider activates, with the neodymium iron boron permanent magnet array providing initial adsorption force. Simultaneously, the negative pressure adsorption unit begins operation, generating a vacuum of -80 kPa, working in conjunction with the neodymium iron boron permanent magnet array to ensure reliable adhesion of the glider to the blade surface. The negative pressure adsorption unit is a mature technology and will not be elaborated upon here.

[0045] To meet the adhesion requirements under extreme wind speeds, the minimum clamping force of the magnetic-gas composite adsorption unit must satisfy the following: Substitute parameters The coefficient of friction of the adsorption surface, Effective working area of ​​the magnetic-gas composite adsorption unit For pressure coefficient, For robot quality, For the blade tilt angle, when Calculated when the wind speed is Wind speed was measured using an anemometer. The actual system employed a combined approach of neodymium iron boron permanent magnet array (surface magnetic flux density 1.2T) and negative pressure adsorption (vacuum degree -80kPa).

[0046] S2. Synchronous acquisition and fusion of multimodal data: Once the adhesion is stable, the system initiates multimodal data acquisition. The radar penetrates the dirt layer on the blade surface to capture the three-dimensional structural information of internal cracks, acquiring radar data; the depth camera focuses on surface details, identifying scratches or minute deformations, obtaining visual image data. To address interference from complex weather conditions (such as rain, fog, and strong light) on the radar or depth camera, the system aligns radar data and visual image data using spatiotemporal registration technology to eliminate spatiotemporal discrepancies. Furthermore, a dynamic weighted fusion algorithm is introduced to adjust the weights of radar data and visual image data in real time based on the signal-to-noise ratio.

[0047] The specific technical solution is as follows: S201, Multimodal Data Synchronous Acquisition After the robot stably attaches to the wind turbine blades, it synchronizes with radar and a depth camera. The radar transmits millimeter waves (wavelength...) The radar penetrates the dirt layer on the blade surface to capture the three-dimensional structural information of internal cracks, acquiring radar data; the depth camera focuses on the details of the blade surface, capturing visible defects such as cracks, scratches, and deformations on the wind turbine blade surface, obtaining visual image data. The radar and depth camera sample synchronously at a frequency of 100Hz to ensure consistency of spatiotemporal data. Simultaneously, ambient temperature and humidity are obtained through temperature and humidity sensors in the meteorological monitoring unit.

[0048] S202, Spatiotemporal Registration and Error Correction Three-dimensional coordinates are obtained using the checkerboard calibration method. To eliminate the spatiotemporal discrepancy between radar data and visual image data, a spatiotemporal registration model is established to ensure accurate alignment between the radar data and visual image data. The specific steps are as follows: A high-precision checkerboard calibration board is deployed in a static scene, ensuring it is simultaneously detected by radar and vision sensors. The radar and vision coordinate systems are calibrated using radar and a depth camera, and the depth camera's intrinsic parameters are also calibrated. The coordinates of the checkerboard corner points in both the radar and vision coordinate systems are recorded. The pixel coordinates of the checkerboard corner points are extracted through image processing and back-projected onto the vision coordinate system using the depth camera's intrinsic parameters, resulting in their 3D coordinates. The three-dimensional coordinates of the checkerboard corner points were extracted from the point cloud data of millimeter-wave radar. .

[0049] Spatiotemporal registration model: in, and Here is the Euler angle rotation matrix. The time synchronization compensation term is used to correct the coordinate deviation caused by the sampling time difference. This coefficient is determined through experimental calibration. It is assumed that the time difference is linearly related to time, and the compensation amount increases linearly with the cumulative time. For about the vertical axis (yaw angle) A rotation matrix is ​​used to align the horizontal direction; For the horizontal axis (pitch angle) A rotation matrix for aligning the vertical direction; This represents the positional offset between the visual sensor and the radar. The spatiotemporal registration model is modeled using an Euler angle rotation matrix and a time synchronization compensation term, calculating the residual between the registered depth camera coordinates and the original radar coordinates; the time error is calculated based on the acquisition time of radar data and visual image data; and the spatial error is calculated based on the radar coordinate system and the visual coordinate system. If the spatial (coordinate) error (RMS) exceeds the threshold (±1.2mm) or the time error (the time between the visual sensor and the radar acquiring data) (≤5ms) does not meet the standard, the parameters are readjusted and iteratively optimized until convergence. The time error is mainly used to measure the time difference between the visual sensor and the radar acquiring data.

[0050] S203, Dynamic Weighted Data Fusion To address the interference of complex weather conditions on radar or depth cameras, the radar data corrected in step S202 is fused with visual image data using a dynamic weighted fusion algorithm to obtain fused data.

[0051] The dynamic weight fusion algorithm is as follows: The weighting coefficients are dynamically adjusted based on the signal-to-noise ratio. In the formula, For the dynamic weighting coefficients of radar data, For radar data Normalization is performed to eliminate dimensional differences. These are the dynamic weighting coefficients for visual image data. For visual image data Laplacian sharpening is applied to enhance image edge features. , .in, The signal-to-noise ratio of radar data. The signal-to-noise ratio of visual image data; For radar signal power, Noise power; The maximum brightness value of the visual image. The value represents the noise standard deviation. Experiments show that the algorithm improves the crack recognition rate from 67% to 93% under rain and fog conditions, and reduces the false alarm rate from 8.7% to 2.1%.

[0052] S3, Crack Depth Quantization and 3D Modeling Based on fused data, the three-dimensional morphology of the crack is reconstructed using a radar scattering model. The radar scattering model analyzes the scattering characteristics of radar signals from cracks on the wind turbine blade surface, extracting key information such as crack length, width, depth, center location, and trajectory to reconstruct the three-dimensional morphology of the crack. Finally, the system generates a three-dimensional model of the blade using the fused data, visually displaying the overall blade structure and crack details. Damage location, type (e.g., fatigue crack or impact damage), and morphological features are labeled on the three-dimensional model, obtaining damage data containing damage location, type, and morphological characteristics, providing visualization support for subsequent life prediction. The radar scattering model is detailed below:

[0053] To address the challenge of quantifying crack depth, a radar scattering model was established. : in, For radar wavelength, The complex dielectric constant of fiberglass is... For the first Crack length, The location of the crack center. For penetration depth, This represents the total number of crack segments. Wave number; The imaginary unit; These are the spatial coordinates in the horizontal direction; The phase delay term represents the wave's penetration depth. Phase change during round-trip propagation.

[0054] S4. Damage evolution modeling and lifetime prediction: Based on damage localization, the system combines physical models and deep learning algorithms to predict damage propagation trends. An improved crack propagation model (such as the Paris formula) is used to analyze operating parameters such as temperature and humidity to quantify crack growth rates. Simultaneously, an LSTM network performs time-series modeling of crack length, temperature, and humidity to predict the lifespan of wind turbine blades, obtaining lifespan prediction data. Furthermore, by adapting to different turbine models through a transfer learning framework, the system can generate lifespan reports across platforms and trigger early warnings 72 hours in advance, helping wind farms develop preventative maintenance plans and avoid power generation losses due to unplanned downtime.

[0055] The specific technical solution is as follows: The damage intelligent prediction module achieves accurate prediction of wind turbine blade life by fusing an improved crack propagation model with an LSTM network.

[0056] Multidimensional characteristic data such as S401, crack length, temperature, and humidity were normalized and then used to establish an improved crack propagation model based on the improved Paris formula. Based on this improved crack propagation model, operating parameters such as temperature and humidity were analyzed to quantify the crack growth rate. ; In the formula, The length of the crack; For ambient temperature and humidity; The rate constant is related to ambient temperature and humidity; These are indices related to materials and the environment.

[0057] S402 uses a 64-32-16 layer LSTM network to perform time-series modeling of crack length, temperature and humidity to predict the life of wind turbine blades. After training 20,000 sets of data with the Adam optimizer, the root mean square error (RMSE) of the prediction is ≤7.9%.

[0058] S403. To further enhance generalization capabilities, a transfer learning framework is introduced to adapt to different turbine models, compressing cross-model test error fluctuations to ±5.3%. Finally, by combining environmental temperature and humidity, wind farm disturbance data, damage data, and life prediction data, a remaining life report is generated, and a fault warning is triggered 72 hours in advance, supporting preventive operation and maintenance decisions for wind farms.

[0059] To overcome the limitations of traditional threshold alarms, this solution constructs an improved crack propagation model based on an improved Paris formula: in This is expressed as the effective stress intensity factor amplitude, quantifying the crack propagation driving force. The larger the crack, the faster the crack propagation rate. The faster; The maximum cyclic stress represents the ultimate load the blade can withstand. The model prediction error is ≤8.5%.

[0060] The 64-32-16 layer LSTM network is as follows: In the formula, The temporal characteristics of crack length, temperature, and humidity; The correlation weights between crack evolution and crack length, temperature, and humidity; For the period of time , The current crack length and temperature; The bias is used to adjust the model output. After training with 20,000 sets of data using crack length, temperature, and humidity, the lifetime prediction RMSE is ≤7.9%.

[0061] The specific steps of introducing a transfer learning framework are as follows: In the formula, It is a task-related loss function used to measure the difference between the model's performance in predicting the evolution of wind turbine blade damage and the expected performance. It reflects the degree of fit of the model to the current task. The smaller the value, the better the model fits the task data.

[0062] It is a domain-dependent loss function used to measure the difference between the source domain (existing wind field data, such as wind field data used for transfer learning) and the target domain (the wind field data to be predicted). It helps the model distinguish features from different domains. The smaller the value, the better the feature differences between the source and target domains are handled, which is beneficial to the effect of transfer learning.

[0063] This is the regularization term in the loss function, used to prevent overfitting. By adding a regularization term to the loss function, the complexity of the model is limited, which to some extent enables the model to have better generalization ability in both the source and target domains. The smaller the value, the stronger the model's generalization ability may be. Cross-model testing error fluctuation range is compressed to ±5.3%.

[0064] S5. Dynamic Path Optimization and Closed-Loop Operation and Maintenance Decisions: Based on environmental temperature and humidity, wind farm disturbance data, damage data, and lifespan prediction data, the system dynamically adjusts its inspection strategy. By improving the A* algorithm and Q-learning strategy, it comprehensively considers wind farm disturbance, energy consumption, and inspection priorities to optimize subsequent inspection paths and shorten the time required to cover multiple units. Simultaneously, damage data and lifespan prediction data are transmitted to the operation and maintenance management platform, automatically generating maintenance priority suggestions or shutdown plans, forming a closed-loop process of "detection-analysis-early warning-decision-making." For example, for high-risk damaged units, the system prioritizes re-inspection or maintenance to minimize power generation loss.

[0065] The specific technical solution is as follows: The traditional A* algorithm's cost function only considers path length or time cost, while the improved A* algorithm incorporates wind field disturbance data as a key factor in its calculation. The improved A* algorithm prioritizes areas with lower wind speeds or stable flow directions, reducing the risk of the robot being affected by wake disturbances and selecting the optimal path. The cumulative cost function of the improved A* algorithm is defined as follows:

[0066] In the formula, Represents a node The square of the magnitude of the three-dimensional wind speed vector at a given location is used to quantify the impact of the wind field on stability by weighting it with a coefficient of 0.05. This represents the cumulative cost from the parent node to the starting point. To move to node The increase in energy consumption is related to wind speed and travel distance, ensuring a balance between energy consumption and wind resistance in the path. , air density; For nodes Wind speed at the location; This is the aerodynamic drag coefficient; The equivalent windward area of ​​the robot; From the parent node to the current node The distance traveled.

[0067] The improved A* algorithm incorporates a Q-learning reinforcement learning framework to adjust the path planning strategy in real time to adapt to the dynamic wind field environment. The Q-value update rule is as follows: In the formula, For instant rewards, r Related to path safety and energy efficiency (e.g., higher rewards for avoiding high-wind-speed areas) r The specific reward mechanism is as follows: Based on wind field data and obstacle information obtained using a depth camera, the safety of candidate paths is evaluated. If the path passes through a stable area with low wind speeds and no obstacles, a higher positive reward is given; if the path passes through an area with high wind speeds and potential collision risks, a negative reward or a lower positive reward is given. Additional positive rewards are given if the path enables the robot to reach high-priority damaged areas or cover more undetected areas more quickly. A certain negative reward is given for each period of time after which the task is not completed or no effective progress is made, prompting the robot to make decisions quickly and proceed along the effective path.

[0068] Using wind field data as state input, in the current state The robot then attempts to perform different actions. Receive immediate rewards based on the reward function. And observe the new state caused by the action. Based on the Q-learning update formula, it incorporates instant rewards. and based on the new state Maximum expected future reward for state-action pairs The Q value is updated to dynamically adjust the priority rules for selecting the next node to be explored (i.e., the next target point on the path), thereby improving adaptability in complex environments.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A wind power remote sensing monitoring system, characterized in that, include: The robot gliding module is used to enable robot movement and stable attachment. The module includes a robot, a glider, and a magnetic-gas composite adsorption unit. The glider is mounted on the robot for movement and climbing. A gripping device with a torque sensor is installed on the glider to measure the torque of the gripping device in real time. The magnetic-gas composite adsorption unit enables the robot to stably attach to the blade. This unit includes a neodymium iron boron permanent magnet array unit and a negative pressure adsorption unit. The neodymium iron boron permanent magnet array unit provides the initial adsorption force, while the negative pressure adsorption unit ensures stable attachment of the glider to the blade surface. The multimodal data fusion module is used to collect radar data of internal cracks in the blade and visual image data of the blade surface, and fuse the radar data and visual image data to obtain fused data; the multimodal data fusion module includes radar, depth camera, meteorological detection unit and communication unit; The radar is used to collect radar data on internal cracks in the blades and to measure the three-dimensional wake velocity field and wind field disturbance data of the wind turbine wake field in real time; the depth camera is used to acquire visual image data of the blade surface and to measure the clamping diameter and gripping angle of the clamping device in real time; the meteorological detection unit includes an anemometer and temperature and humidity sensors, the anemometer is used to detect the wind field velocity in real time; the temperature and humidity sensors are used to collect ambient temperature and humidity data. The damage detection module uses fused data to construct a three-dimensional damage model of the blade and acquire damage data; The damage evolution prediction module is used to predict the lifespan of wind turbine blades, obtain lifespan prediction data, and provide early warnings. The dynamic path planning and autonomous control module optimizes the inspection strategy based on environmental temperature and humidity, wind field disturbance data, damage data, and life prediction data. Anti-interference and communication modules ensure the stability and reliability of data transmission.

2. A wind power remote sensing monitoring method, characterized in that, Includes the following steps: S1. Robot Movement and Attachment: The robot's autonomous flight is controlled by a gliding module. The robot's flight attitude is adjusted using a blended wing-body configuration design and a PID control model. Once the robot reaches the vicinity of the wind turbine blade surface, the clamping force of the gripping device is adjusted using torque calculated by a depth camera and real-time torque measured by a torque sensor. The robot is then attached to the wind turbine blade surface via a magnetic-gas composite adsorption unit. S2. Multimodal Data Synchronous Acquisition and Fusion: After attachment, radar and a depth camera are used for synchronous sampling to obtain radar data of internal blade cracks and visual image data of the blade surface. The radar data and visual image data are aligned using a spatiotemporal registration model, and a dynamic weighted fusion algorithm is used to fuse the radar data and visual image data to obtain fused data. S3. Crack Depth Quantization and 3D Modeling: Based on the fused data, the scattering characteristics of wind turbine blade surface cracks to radar signals are analyzed using a radar scattering model. Crack length, crack width, crack depth, crack center location, and crack direction are extracted. The 3D morphology of the crack is inverted, and the 3D model of the blade is reconstructed using the fused data. Damage data is then obtained based on the 3D model of the blade. S4. Damage evolution modeling and life prediction: The crack growth rate is quantified based on the improved crack propagation model; and the crack length and ambient temperature and humidity are time-series modeled using an LSTM network to predict the life of the wind turbine blades, obtain life prediction data, adapt to different models through a transfer learning framework, generate life reports and provide early warnings. S5. Dynamic Path Optimization: Based on environmental temperature and humidity, damage data, and life prediction data, the inspection path is optimized by improving the A* algorithm and Q learning strategy. At the same time, the damage data and predicted life data are input into the operation and maintenance management platform through the anti-interference and communication module to obtain maintenance priority suggestions.

3. The wind power remote sensing monitoring method according to claim 2, characterized in that, Step S1 is as follows: S101. Control the robot's autonomous flight through the glider in the robot's gliding module; S102. During the initial or stable flight phase, the robot adjusts its flight attitude using a blended wing-body configuration design method, where the glide ratio maximization objective function is used in this method. for: In the formula, the lift coefficient and drag coefficient Through angle of attack respectively It is obtained by combining trigonometric functions and exponential functions; During the dynamic flight phase, the three-dimensional wake velocity field and wind field disturbance data of the wind turbine wake field are acquired by radar; using the PID control model, the pitch angle correction is calculated based on the three-dimensional wake velocity field to adjust the robot's flight attitude. The PID control model is specifically as follows: in, The three-dimensional wake velocity field, The integral of the wake velocity field norm; S103. After the robot reaches the surface of the wind turbine blade, it uses a depth camera to measure the clamping diameter of the glider in real time. and the angle of capture Real-time torque calculation ,in , Torque coefficient; Simultaneously, the torque of the clamping device is measured in real time using a torque sensor. According to torque and torque Obtaining torque difference Through torque difference The gripping force of the clamping device is adjusted based on the set threshold. The neodymium iron boron permanent magnet array unit, utilizing a magnetic-gas composite adsorption unit, provides initial adsorption force in conjunction with a negative pressure adsorption unit, ensuring the reliable adhesion of the glider to the blade surface; the minimum clamping force of the magnetic-gas composite adsorption unit must meet the following requirements: In the formula, The coefficient of friction of the adsorption surface, The effective working area of ​​the magnetic-gas composite adsorption unit, For pressure coefficient, For robot quality, This is the blade tilt angle.

4. The wind power remote sensing monitoring method according to claim 2, characterized in that, Step S2 is as follows: S201, Multimodal Data Synchronous Acquisition: After the robot attaches to the blade, the radar and depth camera sample synchronously in time and space at a frequency of 100Hz; the radar emits millimeter waves to penetrate the dirt layer on the blade surface and obtain radar data of internal cracks in the blade; the depth camera obtains visual image data of blade surface details, scratches or minor deformations. S202, Spatiotemporal Registration and Error Correction: Three-dimensional coordinates are obtained through the checkerboard calibration method, and radar data and visual image data are aligned using a spatiotemporal registration model; S203, Dynamic Weighted Data Fusion: The radar data and visual image data processed in step S202 are fused using a dynamic weighted fusion algorithm to obtain fused data.

5. The wind power remote sensing monitoring method according to claim 4, characterized in that, The chessboard calibration method and spatiotemporal registration model in step S202 are as follows: Chessboard grid marking method: A high-precision checkerboard calibration board is deployed in a static scene; the radar coordinate system and visual coordinate system are calibrated using radar and a depth camera, and the intrinsic parameters of the depth camera are also calibrated; the coordinates of the checkerboard corner points in the radar coordinate system and the visual coordinate system are recorded; the pixel coordinates of the checkerboard corner points are extracted through image processing, and then back-projected onto the three-dimensional coordinates in the visual coordinate system using the depth camera intrinsic parameters. ; Extracting the three-dimensional coordinates of checkerboard corner points from point cloud data using millimeter-wave radar. ; Constructing a spatiotemporal registration model: in, and Here is the Euler angle rotation matrix. The time synchronization compensation term is used to correct coordinate deviations caused by sampling time differences; This indicates the positional offset between the visual sensor and the radar; Yaw angle; It is the pitch angle.

6. The wind power remote sensing monitoring method according to claim 4, characterized in that, The dynamic weight fusion algorithm in step S203 is as follows: The weighting coefficients are dynamically adjusted based on the signal-to-noise ratio. In the formula, For the dynamic weighting coefficients of radar data, For radar data Perform normalization processing; These are the dynamic weighting coefficients for visual image data. For visual image data Perform Laplacian sharpening; , ;in, The signal-to-noise ratio of radar data. The signal-to-noise ratio of visual image data; For radar signal power, Noise power; This represents the maximum brightness value of the image. This represents the standard deviation of noise.

7. The wind power remote sensing monitoring method according to claim 2, characterized in that, Radar scattering model in step S3 for: in, For radar wavelength, The complex dielectric constant of fiberglass is... For the first Crack length, The location of the crack center. For penetration depth, This represents the total number of crack segments. Wave number; The imaginary unit; These are the spatial coordinates in the horizontal direction.

8. The wind power remote sensing monitoring method according to claim 2, characterized in that, Step S4 is as follows: S401. Using the normalized crack length, temperature, and humidity, and establishing an improved Paris formula crack propagation model, analyze the temperature and humidity parameters based on the improved crack propagation model to quantify the crack growth rate. ; In the formula, The length of the crack; For ambient temperature and humidity; The rate constant is related to ambient temperature and humidity; Indices related to materials or the environment; The improved crack propagation model is as follows: In the formula, This is expressed as the effective stress intensity factor amplitude, quantifying the crack propagation driving force. The larger the crack, the faster the crack propagation rate. The faster; This represents the maximum cyclic stress. S402. A 64-32-16 layer LSTM network is used to perform time-series modeling of crack length, temperature, and humidity to predict the lifespan of wind turbine blades and obtain lifespan prediction data; the 64-32-16 layer LSTM network is as follows: In the formula, The temporal characteristics of crack length, temperature, and humidity; The correlation weights between crack evolution and crack length, temperature, and humidity; For the period of time , The current crack length and temperature; The bias is used to adjust the model output; S403 introduces a transfer learning framework to adapt to different models. The system combines environmental temperature and humidity, damage data, and life prediction data to generate a remaining life report and triggers a fault warning 72 hours in advance. The transfer learning framework is as follows: In the formula It is a task-related loss function; It is a domain-dependent loss function; It is the loss function of the regularization term.

9. The wind power remote sensing monitoring method according to claim 2, characterized in that, The improved A* algorithm for step S5 is as follows: The cumulative cost function of the improved A* algorithm is defined as: In the formula, Represents a node The square of the magnitude of the three-dimensional wind speed vector at that location; This represents the cumulative cost from the parent node to the starting point. To move to node The increase in energy consumption, of which , air density; For nodes Wind speed at the location; This is the aerodynamic drag coefficient; The equivalent windward area of ​​the robot; From the parent node to the current node The distance traveled.

10. The wind power remote sensing monitoring method according to claim 2, characterized in that, The improved algorithm in step S5 introduces the Q-learning reinforcement learning framework, with the Q-value update rule as follows: In the formula, For instant rewards, This is the current state. For action, For the new state, This is a new action.

Images