Fan blade robot maintenance method and system based on cloud edge collaboration

Abstract

The invention belongs to the field of intelligent robots, and discloses a fan blade robot maintenance method and system based on cloud edge collaboration, and the method comprises the steps: enabling an intelligent robot to collect a damaged image of a fan blade in real time, and carrying out the preliminary feature extraction; the edge node classifies damage types by using a multi-task learning support vector machine, robot positioning is performed and a preliminary maintenance strategy is generated by fusing multi-sensor data and an SLAM algorithm, and the robot executes a maintenance task according to an instruction of the edge node and feeds back the maintenance task to the cloud; the cloud retrains the classification model through the convolutional neural network, optimizes a loss function and model parameters, and pushes a parameter optimization result to the edge node; the edge nodes replace original model parameters, and a fan blade damage classification and maintenance strategy is optimized; according to the method, the problems of resource configuration and maintenance efficiency of fan blade maintenance can be effectively solved through a cloud side cooperative working method of processing a high-real-time task at an edge node and uploading a complex task to the cloud side.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent robots, and specifically relates to a wind turbine blade robot maintenance method and system based on cloud-edge collaboration. Background Art

[0002] With the annual increase in the capacity of wind turbine generator sets and the continuous expansion of installed capacity, the reliability maintenance of wind turbines and the efficient maintenance of wind turbine blades have gradually become the focus of industry development. Traditional technical inspections such as visual inspection, infrared thermal imaging inspection, and vibration inspection are time-consuming, and maintenance after inspection is complicated. Too long downtime will lead to loss of production capacity and reduced economic benefits, which does not meet people's actual needs. Therefore, it is of practical engineering significance to develop an efficient wind turbine blade maintenance robot maintenance system to reduce maintenance costs. Summary of the invention

[0003] In order to solve the above problems existing in the prior art, the present invention proposes a wind turbine blade robot maintenance method and system based on cloud-edge collaboration.

[0004] On the one hand, the present invention provides a wind turbine blade robot maintenance method based on cloud-edge collaboration, the method comprising:

[0005] S1: The robot collects images of wind turbine blade damage through high-resolution cameras and sensors and performs feature extraction;

[0006] S2: The edge computing node pre-processes the image data and uses a multi-task learning support vector machine (MTL-SVM) to classify the feature parameters and identify the damage type. It combines the SLAM algorithm with multi-sensor fusion to obtain the robot's position information and matches it with the high-precision map to determine the damage location, generate a preliminary maintenance strategy, and send it to the robot.

[0007] S3: The cloud server receives the damage type and preliminary maintenance strategy, retrains the multi-task learning support vector machine (MTL-SVM) through the convolutional neural network (CNN), optimizes the loss function and model parameters, and pushes the updated parameters to the edge node. The edge node performs new damage classification and maintenance strategy generation based on the optimized multi-task learning support vector machine (MTL-SVM) to obtain maintenance instructions;

[0008] S4: The robot performs maintenance operations according to the maintenance instructions of the edge node and feeds back the data results to the cloud.

[0009] On the other hand, the present invention provides a wind turbine blade robot maintenance system based on cloud-edge collaboration, the system comprising: an image acquisition module, an edge processing module, an execution control module, a position information acquisition module and a cloud processing module;

[0010] The image acquisition module is used to collect images of damaged locations of wind turbine blades through a high-resolution camera and sensor, and perform preliminary noise reduction and feature extraction processing on the images;

[0011] The edge processing module is used to further extract features and enhance the image of the damaged position of the fan blade transmitted by the image acquisition module, and classify the extracted features using a multi-task learning support vector machine (MTL-SVM) to determine the damage type of the fan blade;

[0012] The execution control module is used to control the robot to perform corresponding maintenance operations, including cleaning, grinding and spraying, according to the maintenance decision instructions generated by the edge processing module;

[0013] The position information acquisition module is used to match the real-time acquired position information with the high-precision map based on the data of the position sensor in combination with the SLAM algorithm of multi-sensor fusion, so as to achieve accurate positioning and path planning of the robot;

[0014] The cloud processing module is used to store the location information, damage classification results and robot execution feedback data uploaded by the edge processing module, and optimize the deep learning model based on these data, and push the optimized model parameters to the edge processing module so that it can apply the updated model parameters for more accurate classification and maintenance decisions.

[0015] Beneficial effects of the present invention:

[0016] The present invention uses edge computing nodes to preprocess image data and classify multi-task learning support vector machines (MTL-SVM), combined with a multi-sensor fusion SLAM algorithm, so that the edge node can accurately obtain the robot position, realize the rapid identification of wind turbine blade damage types and the generation of maintenance strategies, and significantly improve real-time performance and response speed. At the same time, the cloud server is responsible for complex data analysis and model optimization, and pushes updated model parameters to the edge node, further improving the classification accuracy and maintenance effect. The collaborative mechanism between the cloud and edge nodes realizes efficient resource allocation and intelligent optimization, so that the system achieves a good balance between computing efficiency and operational accuracy. The robot automatically performs operations such as cleaning, polishing, and spraying, which significantly improves the degree of automation of wind turbine blade maintenance and reduces the need for manual intervention. The closed-loop feedback mechanism of cloud-edge collaboration enables the system to adaptively adjust, continuously optimize maintenance effects, enhance the reliability and stability of the system, and thus greatly improve the operational safety and maintenance quality of the wind turbine. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a wind turbine blade robot maintenance method based on cloud-edge collaboration in an embodiment of the present invention;

[0018] Figure 2 This is a cloud-edge architecture diagram of a wind turbine blade robot maintenance system based on cloud-edge collaboration in an embodiment of the present invention;

[0019] Figure 3 is a flow chart of wind turbine blade damage classification based on MTL-SVM in an embodiment of the present invention;

[0020] Figure 4 This is a flow chart of wind turbine blade robot positioning based on cloud-edge collaboration in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] A wind turbine blade maintenance method based on edge computing, such as Figure 1 As shown, the method includes: the robot collects damaged images of wind turbine blades through high-resolution cameras and sensors, and extracts features; the extracted information is sent to the edge computing node; the edge node preprocesses the data and classifies the damage type using a multi-task learning support vector machine (MTL-SVM); the multi-sensor data is integrated with the SLAM algorithm to generate preliminary maintenance instructions; after the robot executes the preliminary instructions, the damage classification results and execution data are fed back to the cloud; the cloud retrains the classification model through a convolutional neural network (CNN), optimizes the loss function, and optimizes the model parameters; the edge node receives the updated model parameters and replaces the original (MTL-SVM) model parameters for subsequent wind turbine blade damage classification and maintenance strategies.

[0023] In this embodiment: First, high-resolution cameras and sensors are used to collect image data of wind turbine blades. The robot takes pictures of wind turbine blades through patrol inspection or fixed position to collect damaged images, including cracks, paint peeling, wear and other damage conditions of wind turbine blades. The collected wind turbine blade images will be used to extract key feature information of damage, such as crack width, damaged texture, etc., through feature extraction algorithms, and this information will be sent to edge computing nodes.

[0024] Subsequently, in the edge computing node, the received wind turbine blade image is preprocessed. First, Gaussian filtering is applied to the image data to reduce noise and improve image quality. Secondly, the image pixel values ​​are normalized to make the data range consistent. The Gray Level Coevolution Matrix (GLCM) is used to extract image feature parameters to extract key feature information of damage. The image feature parameters include: pixel value, gray value, texture feature, and correlation function, contrast function and entropy function feature extraction function are selected. The feature extraction function of selecting correlation function, contrast function and entropy function is characterized in that multi-region connections are established through correlation function, and the linear correlation of grayscale between pixels at specified positions is calculated. The greater the correlation, the greater the correlation in the pixels. The expression of the correlation function is:

[0025]

[0026] The correlation function is used to compare the intensity of adjacent pixels in the image and distinguish the differences in pixel light and color in different parts of the image. The contrast function is:

[0027]

[0028] The entropy function is used to detect pixel uncertainty in the image. The larger the entropy value, the richer the image texture information and the more dispersed the pixel distribution. The smaller the entropy value, the less image texture information and the more concentrated the pixel distribution. The entropy function is:

[0029]

[0030] Among them, a is the mean, v is the standard deviation, and h(m,n) is the number of times pixels with gray levels m and n appear at the specified position.

[0031] Due to the huge size of the wind turbine blades, the camera field of view is limited, and the range of each photo is limited, so the pictures collected by the camera need to be stitched. The images are stitched by pixel cropping, and the adjacent images are guaranteed to have a 5% overlap area during stitching.

[0032] According to the analysis of the field of view size, a total of 4 field of view characteristic parameters are selected as shown in Table 1.

[0033] Table 1. Field of view characteristic parameters of smart cameras

[0034]

[0035] For the extracted feature parameters, the multi-task learning support vector machine (MTL-SVM) architecture is used in the edge computing node to handle multiple damage classification tasks at the same time. Each sub-model focuses on different types of damage (such as paint peeling, cracks, wear, etc.) and is processed in parallel in the edge node. Specifically, in the MTL-SVM architecture, in order to handle multiple wind turbine blade damage classification tasks at the same time, a joint loss function is designed, which combines the losses of the shared model and the task-specific model:

[0036]

[0037] Among them, L is the total loss function; L shared (W shared ) is the loss of the shared model part, which represents the loss function of sharing features between different tasks; is the loss function of the task-specific model, which is used to optimize the specific classification model for each task; i is the task weight coefficient; N is the number of tasks; W shared Represents shared model parameters; represents the parameters of a specific classification model; it depends not only on the parameters W of the shared model share , but also with task-specific parameters Correlation; i is the task weight coefficient, which is used to control the importance of each task in the overall model training. The system can adjust these weights according to the classification difficulty or real-time changes of different tasks. N is the number of tasks, which includes paint loss detection, crack detection, wear detection and other tasks in this application.

[0038] For each task sub-model, the classification function is used to transform the input feature s m Mapped to a specific damage category. The classification result calculation of each sub-model in MTL-SVM is:

[0039] T i (s)=(i,j)s m +b i

[0040] Among them, T i (s) represents the classification result of the i-th task (such as paint peeling detection, crack detection, wear detection); (i, j) is the weighted vector in the support vector machine model, which represents the weight of the input feature in the classification model, reflecting the importance of the feature in different tasks; s m is the input feature parameter, which represents the features extracted from the leaf damage image (such as gray value, texture features, etc.); b i is a bias term used to adjust the threshold of the classification result. The classification function T of each task i(s) can be optimized independently, while at the same time, through the synergy of shared models, the classification tasks of different damage types can be completed in parallel within the same architecture.

[0041] In multi-task learning, optimizing the shared parameter W shared The goal is to maximize the synergy between tasks. The optimization process can be expressed as:

[0042]

[0043] in, is the gradient of the shared parameters, which is used to guide the adjustment direction of the shared model between various tasks; is the gradient influence of the ith task on the shared parameters, that is, the contribution of the shared parameters to each task. By weighted summation of the gradients of each task, the shared parameters can be balanced between different tasks.

[0044] In order to adapt to the dynamic changes of blade damage, the system assigns the weight λ to each task. i To make adaptive adjustments:

[0045]

[0046] Among them, λ i (t) is the weight of the i-th task at time t; η is the learning rate, which controls the step size of weight adjustment; is the derivative of the i-th task loss function with respect to the weight, which is used to guide the direction of weight adjustment. Through the above steps, the MTL-SVM architecture improves the classification accuracy and computational efficiency through shared parameters and collaborative optimization between tasks, making the classification results more accurate. Efficiently implementing multi-task classification in edge computing nodes enables the model to better learn damage characteristics and adapt to different working conditions.

[0047] At the same time, in this embodiment, the robot combines multi-sensor fusion technology and uses a multi-sensor fusion SLAM algorithm to obtain position information in real time and match it with a high-precision map to determine the precise position of the robot. Specifically, it includes: installing a position sensor on the wind turbine blade robot, using the sensor to collect the position information of the robot during the sampling process, using a three-dimensional cloud map to match the real-time perception data to provide positioning information, and obtaining accurate information about the damaged position of the wind turbine blade photographed by the robot. The normal distribution transformation based on probability matching divides the target point into a multi-grid model, calculates the covariance and mean of the points in the grid, solves the mostly normal distribution of each grid and calculates the joint probability distribution model, that is, the probability that the current point falls in the target point grid during the rotation matching process between the current point and the target point. When the probability reaches the maximum, it is the best match.

[0048] The target area is divided into a multi-layer grid model, each grid layer has a different resolution, low-resolution matching is performed first, and the matching range is gradually narrowed. In the high-resolution grid, the covariance and mean of each grid point are carefully analyzed to ensure the balance between calculation accuracy and speed; for each grid, the covariance matrix and mean vector of the internal points are calculated, and the multidimensional normal distribution in the grid is solved based on these parameters to calculate the joint probability between the current point and the target point:

[0049]

[0050] Among them, P(x|μ,∑) represents the joint probability between the current point and the target point, x is the position of the current point, μ is the mean of the target point, ∑ is the covariance matrix, y represents the dimension of the data, and T represents the transpose operation of the matrix;

[0051] During the rotation matching process, the probability of the current point falling in the target point grid is calculated, and a joint probability distribution model is constructed based on this: the joint probability matrix between the candidate point and the target point is calculated, with each row of the matrix corresponding to a target point and each column corresponding to a candidate point. Then, the candidate point with the highest joint probability is selected from the matrix as the preliminary matching point, and a threshold (0.8) is set to exclude low-reliability matches. Next, the matching results are optimized by verifying the geometric relationship (distance and angle) between the points, and points that do not meet the consistency will be replaced with the second highest probability point. Finally, the verified and optimized matching pairs of the target point and the candidate point are output to ensure the accuracy and geometric consistency of the matching.

[0052] Through processing by edge computing nodes, preliminary maintenance strategies for damage types are generated, such as specific operation instructions such as cleaning, polishing, and spraying.

[0053] After receiving the preliminary maintenance instructions, the robot performs the corresponding maintenance operations according to the specific damage type. After the robot completes the execution, the maintenance results (including the classification results of the damage type, the operation execution status, the task success rate, etc.) will be sent back to the cloud server. The cloud further analyzes these data through a convolutional neural network (CNN), extracts the type, degree, repair results and specific conditions of the wind turbine blade damage, and uses them as new training samples to retrain the model. The damage type is used as a label, and the robot execution status data is used as an input feature. The cross entropy loss function L1 is used to minimize the classification error:

[0054]

[0055] Among them, y i is the true label, The probability predicted by the model; the model parameters are optimized through back propagation and gradient descent algorithm (Adam).

[0056] The cloud analyzes the classification accuracy and the success rate of the robot's task execution to adjust the model's weight parameters (i, j) and bias b. i Perform optimization and generate updated parameters (i, j) ′ and b i ′ After receiving the optimization parameters transmitted by the cloud, the edge node updates and replaces the model parameters during the idle time period:

[0057] New damage classification and maintenance strategy generation are performed based on the optimized model to ensure higher classification accuracy and decision-making efficiency in subsequent wind turbine blade maintenance tasks.

[0058] Generate new damage classification and maintenance strategies based on the optimized model:

[0059] T i ′(s)=(i,j)′·s m +b′ i

[0060] In this embodiment, a local map is used instead of a global map to participate in the matching, which reduces memory usage and improves computational efficiency. The local map is dynamically updated at a certain frequency as the robot moves, with the initial position of the robot as the starting point, to ensure high-precision positioning of the robot on the wind turbine blade.

[0061] In this embodiment, the edge collects the robot position information and the wind blade damage information, and transmits the maintenance information to the robot to perform wind blade maintenance.

[0062] In this embodiment, the MTL-SVM output result of 0 is defined as no operation, and the robot continues to inspect. 1 is defined as a repainting operation, and the robot performs repainting maintenance on the fan blades. The output result of 2 is defined as complex processing, and the robot performs other complex processing such as blade fracture repair.

[0063] In this embodiment, the cloud-edge collaboration feature is that the end-edge cloud architecture is divided into a device layer, an edge layer, and a cloud service layer.

[0064] The equipment layer includes: smart cameras, sensors, paint touch-up equipment, etc.

[0065] The edge layer includes: pocket microcomputers, edge gateways and other edge servers.

[0066] The cloud service layer includes: cloud processor, cloud storage, etc.

[0067] The pocket microcomputer has networking capabilities and can perform complex calculations to make local decisions quickly. It can also perform image classification processing and send instructions to improve the robot's execution capabilities.

[0068] The pocket microcomputer processes one frame of image every 3 seconds and transmits data to the edge layer every 5 seconds, ensuring that the edge computing system extracts and analyzes image information 12 times per minute.

[0069] The cloud-edge-end architecture of the wind turbine blade robot maintenance system established by the present invention is as follows: Figure 2 As shown in the figure, the classification process of wind turbine blade damage type based on MTL-SVM is as follows Figure 3 As shown in Figure 2, the robot positioning process is as follows: Figure 4 shown.

[0070] A wind turbine blade robot maintenance system based on cloud-edge collaboration, the system comprising: an image acquisition module, an edge processing module, an execution control module, a position information acquisition module and a cloud processing module;

[0071] The image acquisition module is used to acquire an image of the damaged position of the fan blade;

[0072] The edge processing module is used to process and classify the damaged position image of the wind turbine blade;

[0073] The execution control module is used to control the robot according to the edge layer decision instructions to complete the operation instructions;

[0074] The position information acquisition module is used to match the real-time acquired position information with the data in the high-precision map through multi-sensor fusion according to the signal sent by the position sensor, combined with the multi-sensor SLAM algorithm, to achieve accurate positioning of the robot;

[0075] The cloud processing module is used to store the collected location information and images, and to achieve secure transmission of data between the robot device, edge node and cloud server.

[0076] The system implementation of the present invention is the same as the method implementation.

[0077] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A wind turbine blade robot maintenance method based on cloud-edge collaboration, characterized in that: include: S1: The robot collects images of wind turbine blade damage through high-resolution cameras and sensors and performs feature extraction; S2: The edge computing node pre-processes the image data and uses a multi-task learning support vector machine (MTL-SVM) to classify the feature parameters and identify the damage type. It combines the SLAM algorithm with multi-sensor fusion to obtain the robot's position information and matches it with the high-precision map to determine the damage location, generate a preliminary maintenance strategy, and send it to the robot. S3: The cloud server receives the damage type and preliminary maintenance strategy, retrains the multi-task learning support vector machine (MTL-SVM) through the convolutional neural network (CNN), optimizes the loss function and model parameters, and pushes the updated parameters to the edge node. The edge node performs new damage classification and maintenance strategy generation based on the optimized multi-task learning support vector machine (MTL-SVM) to obtain maintenance instructions; S4: The robot performs maintenance operations according to the maintenance instructions of the edge node and feeds back the data results to the cloud.

2. According to the cloud-edge collaboration-based wind turbine blade robot maintenance method of claim 1, it is characterized in that: The edge computing node pre-processes the image data, including: Gaussian filtering is used on the image data to reduce noise and improve image quality. Secondly, the image pixel values ​​are normalized to make the data range consistent.

3. According to the cloud-edge collaboration-based wind turbine blade robot maintenance method of claim 1, it is characterized in that: The robot uses high-resolution cameras and sensors to collect images of wind turbine blade damage and perform feature extraction, including: The robot senses and identifies the wind blades through a camera, collects and obtains damaged image information of the wind blades; constructs a grayscale co-occurrence matrix based on the image grayscale value to capture the texture features in the damaged image of the wind blades; based on the constructed grayscale co-occurrence matrix, a feature extraction function is used to extract features of the damaged image information of the wind blades to obtain multiple feature parameters; the feature parameters include but are not limited to pixel values, grayscale values ​​and texture features; the feature extraction function includes but is not limited to correlation function, contrast function and entropy function.

4. The wind turbine blade robot maintenance method based on cloud-edge collaboration according to claim 3 is characterized in that: The correlation function expression is: The contrast function is: The entropy function is: Among them, Correlation is the correlation function, a is the mean, v is the standard deviation, h(m,n) is the number of times pixels with grayscale levels m and n appear at a specified position; Con is the contrast function, which measures the difference in grayscale levels in an image. The larger the value, the higher the contrast of the image; ENT is the entropy function, which indicates the complexity of image information. The larger the value, the higher the randomness of the image and the more information it contains.

5. The wind turbine blade robot maintenance method based on cloud-edge collaboration according to claim 1 is characterized in that: Multi-task learning support vector machine (MTL-SVM) is used to classify feature parameters and identify damage types, including: In the MTL-SVM architecture, in order to handle multiple wind turbine blade damage classification tasks simultaneously, a joint loss function is designed, which combines the losses of the shared model and the task-specific model: Among them, L is the total loss function; L shared (W shared ) is the loss of the shared model part, which represents the loss function of sharing features between different tasks; is the loss function of the task-specific model, which is used to optimize the specific classification model for each task; i is the task weight coefficient; N is the number of tasks; W shared Represents shared model parameters; represents a specific classification model parameter; For each task-specific classification model, its classification function is used to transform the input feature s m Mapped to specific damage categories; this formula describes the calculation of the classification results of each sub-model in MTL-SVM: T i (s)=(i,j)s m +b i Among them, T i (s) represents the classification result of the i-th task; (i,j) is the weight vector in the support vector machine model; s m is the input feature; b i is the bias term; In multi-task learning, optimizing the shared parameter W shared The goal is to maximize the synergy between tasks, and its optimization process can be expressed as: in, is the gradient of the shared parameters; is the gradient influence of the i-th task on the shared parameters; In order to adapt to the dynamic changes of blade damage, the system assigns the weight λ to each task. i To make adaptive adjustments: Among them, λ i (t+1) is the weight of the i-th task at time t+1; λ i (t) is the weight of the i-th task at time t; η is the learning rate, which controls the step size of weight adjustment; is the derivative of the loss function of the i-th task with respect to the weight, which is used to guide the direction of weight adjustment.

6. The wind turbine blade robot maintenance method based on cloud-edge collaboration according to claim 1 is characterized in that: Combined with multi-sensor fusion SLAM algorithm to obtain robot position information, including: Sensors are used to collect the robot's position information during the sampling process in real time. The SLAM algorithm integrates the data collected by multiple sensors on the robot in real time to obtain environmental perception information, and uses a probabilistic matching model to match the three-dimensional cloud map with the robot's position information to obtain accurate position information.

7. The wind turbine blade robot maintenance method based on cloud-edge collaboration according to claim 6 is characterized in that: A probabilistic matching model is used to match the 3D cloud map with the robot’s location information to obtain accurate location information, including: The target area is divided into a multi-layer grid model, each grid layer has a different resolution, low-resolution matching is performed first, and the matching range is gradually narrowed; in the high-resolution grid, the covariance and mean of each grid point are carefully analyzed to ensure the balance between calculation accuracy and speed; for each grid, the covariance matrix and mean vector of the internal points are calculated, and the multidimensional normal distribution in the grid is solved based on these parameters to calculate the joint probability between the current point and the target point: Among them, P(x|μ,∑) represents the joint probability between the current point and the target point, x is the position of the current point, μ is the mean of the target point, ∑ is the covariance matrix, y represents the dimension of the data, and T represents the transpose operation of the matrix; The joint probability between the current point and the target point in each grid is calculated to obtain the joint probability matrix between the candidate points and the target points. Each row of the matrix corresponds to a target point, and each column corresponds to a candidate point. The candidate point with the highest joint probability is selected from the matrix as the preliminary matching point, and a threshold is set to exclude low-reliability matches. Then, the matching results are optimized by verifying the geometric relationship (distance and angle) between the points. Points that do not meet the consistency will be replaced by the second highest probability point. Finally, the verified and optimized matching pairs of the target point and the candidate point are output to ensure the accuracy and geometric consistency of the matching.

8. The wind turbine blade robot maintenance method based on cloud-edge collaboration according to claim 1 is characterized in that: The wind turbine blade maintenance instructions include: when the support vector machine output is 0, the system does not generate a command and the robot continues to inspect; when the support vector machine output is 1, the system generates a repainting instruction; when the support vector machine output is 2, the system generates a repair instruction.

9. The wind turbine blade robot maintenance method based on cloud-edge collaboration according to claim 1 is characterized in that: The cloud server receives the damage type and preliminary maintenance strategy, retrains the multi-task learning support vector machine (MTL-SVM) through the convolutional neural network (CNN), optimizes the loss function and model parameters, and pushes the updated parameters to the edge node. The edge node performs new damage classification and maintenance strategy generation based on the optimized multi-task learning support vector machine (MTL-SVM) to obtain maintenance instructions, including: The cloud server uses CNN convolutional neural network to receive data from edge computing nodes and generates location information, operation success rate, and maintenance effect when performing maintenance tasks based on maintenance strategies; Using the damage types and preliminary maintenance strategies uploaded by edge nodes as training samples, the model is retrained, with damage types as labels and robot execution data as input features. The cross entropy loss function L is used to minimize the classification error, and the model parameters are optimized through back propagation and gradient descent algorithms (Adam); The cloud analyzes the classification accuracy and the success rate of the robot's task execution to adjust the model's weight parameters (i, j) and bias b. i Optimize and generate updated parameters (i,j)′ and b′ i ; After receiving the optimization parameters transmitted by the cloud, the edge nodes synchronously update and replace the model parameters: New damage classification and maintenance strategy generation are performed based on the optimized model to ensure higher classification accuracy and decision-making efficiency in subsequent wind turbine blade maintenance tasks.

10. A wind turbine blade robot maintenance system based on cloud-edge collaboration, the system is used to execute a wind turbine blade robot maintenance method based on cloud-edge collaboration as described in any one of claims 1 to 9, characterized in that: include: Image acquisition module, edge processing module, execution control module, location information acquisition module and cloud processing module; The image acquisition module is used to collect images of damaged locations of wind turbine blades through a high-resolution camera and sensor, and perform preliminary noise reduction and feature extraction processing on the images; The edge processing module is used to further extract features and enhance the image of the damaged position of the fan blade transmitted by the image acquisition module, and classify the extracted features using a multi-task learning support vector machine (MTL-SVM) to determine the damage type of the fan blade; The execution control module is used to control the robot to perform corresponding maintenance operations according to the maintenance decision instructions generated by the edge processing module; The position information acquisition module is used to match the real-time acquired position information with the high-precision map based on the data of the position sensor in combination with the SLAM algorithm of multi-sensor fusion, so as to achieve accurate positioning and path planning of the robot; The cloud processing module is used to store the location information, damage classification results and robot execution feedback data uploaded by the edge processing module, and optimize the deep learning model based on these data, and push the optimized model parameters to the edge processing module so that it can apply the updated model parameters for more accurate classification and maintenance decisions.

Images