A method and system for analyzing wind turbine blade images based on spatial trajectory

Abstract

The present invention discloses a method and system for splicing and analyzing wind turbine blade images based on spatial trajectories, comprising: first obtaining pictures taken by a drone with latitude, longitude and altitude marks, calculating the actual distance offset and pixel conversion ratio of the centers of adjacent pictures, obtaining the pixel offset to adjust the pictures, and splicing to obtain a panoramic picture of the blades. The trained abnormality recognition model is called to obtain the abnormal part results, and then the estimated degradation results of the surrounding parts are obtained, so as to realize accurate analysis and evaluation of the wind turbine blades.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine blade detection, and in particular to a wind turbine blade image stitching and analysis method and system based on spatial trajectory. Background Art

[0002] In the field of wind turbine blade inspection, traditional image stitching methods are often based on feature points, which are difficult to deal with single backgrounds and have poor real-time performance. With the development of drone aerial photography technology, it is possible to obtain multi-angle images of wind turbine blades, but how to accurately stitch these images and conduct effective analysis is an urgent problem to be solved. Existing technologies have deficiencies in processing the stitching and analysis of wind turbine blade images, which affects the accurate judgment of blade conditions and maintenance decisions. Summary of the invention

[0003] The object of the present invention is to provide a method and system for stitching and analyzing wind turbine blade images based on spatial trajectories.

[0004] In a first aspect, an embodiment of the present invention provides a method for analyzing wind turbine blade images based on spatial trajectories, comprising:

[0005] Acquire a plurality of drone-photographed pictures, wherein the plurality of drone-photographed pictures are obtained by drones for aerial photography of target wind turbine blades, and each of the drone-photographed pictures is marked with a shooting latitude and longitude and a shooting altitude;

[0006] According to the shooting longitude and latitude and the shooting altitude, the actual distance offset between the centers of two adjacent drone-shot pictures is calculated;

[0007] Calculating a pixel conversion ratio according to the actual center distance offset and the lens parameters of the drone;

[0008] Calculating a pixel offset according to the pixel conversion ratio, and adjusting the plurality of drone-photographed pictures according to the pixel offset to obtain a plurality of adjusted drone-photographed pictures;

[0009] Stitching the multiple adjusted drone-photographed images to obtain a panoramic image of the target wind turbine blades;

[0010] Calling a pre-trained abnormality recognition model to obtain an abnormal part recognition result corresponding to the leaf panoramic image;

[0011] According to the abnormal part identification result, the estimated degradation result of the corresponding surrounding part is obtained.

[0012] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is used to execute the method described in the first aspect.

[0013] Compared with the prior art, the beneficial effects provided by the present invention include: adopting a wind turbine blade image stitching analysis method and system based on spatial trajectory disclosed by the present invention, by obtaining pictures taken by drones with latitude, longitude and altitude marks, calculating the actual distance offset and pixel conversion ratio of the centers of adjacent pictures, obtaining the pixel offset to adjust the pictures, and stitching to obtain a panoramic picture of the blades. Calling the trained abnormality recognition model to obtain the abnormal part results, and then obtaining the estimated degradation results of the surrounding parts, to achieve accurate analysis and evaluation of the wind turbine blades. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0015] Figure 1 A schematic flow chart of the steps of a method for stitching and analyzing wind turbine blade images based on spatial trajectories provided in an embodiment of the present invention;

[0016] Figure 2 A schematic diagram of stitching pictures taken by a drone according to an embodiment of the present invention;

[0017] Figure 3 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0019] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0020] In order to solve the technical problems in the aforementioned background technology, Figure 1 A schematic flow chart of a method for stitching and analyzing wind blade images based on spatial trajectories provided in an embodiment of the present disclosure is provided. The method for stitching and analyzing wind blade images based on spatial trajectories is introduced in detail below.

[0021] Step S201, obtaining a plurality of pictures taken by a drone, wherein the plurality of pictures taken by a drone are obtained by aerial photography of a target wind turbine blade by a drone, and each picture taken by the drone is marked with a shooting latitude and longitude and a shooting altitude;

[0022] Step S202, calculating the actual center distance offset of each of two adjacent drone-photographed pictures according to the shooting longitude and latitude and the shooting altitude;

[0023] Step S203, calculating a pixel conversion ratio according to the center actual distance offset and the lens parameters of the drone;

[0024] Step S204, calculating a pixel offset according to the pixel conversion ratio, and adjusting the plurality of drone-photographed pictures according to the pixel offset to obtain a plurality of adjusted drone-photographed pictures;

[0025] Step S205, stitching the multiple adjusted drone-photographed images to obtain a panoramic image of the target wind turbine blade;

[0026] Step S206, calling a pre-trained abnormality recognition model to obtain an abnormal part recognition result corresponding to the leaf panoramic image;

[0027] Step S207: obtaining estimated degradation results of corresponding peripheral parts according to the abnormal part identification results.

[0028] In an embodiment of the present invention, for example, in a large wind farm, a blade of a wind turbine needs to be inspected. In order to comprehensively and accurately obtain the image information of the wind turbine blades, multiple drones are used for aerial photography. Each drone is equipped with a high-precision positioning system and image shooting equipment. The server first establishes a communication connection with the ground station that controls the drone, and the ground station transmits the image data taken by the drone to the server. These images are stored in the form of digital files, and each file contains some relevant information. After receiving these image data, the server stores them in a special database. For example, the server receives a certain number of pictures taken by drones, and these pictures are numbered in sequence. Each picture records the relevant information about the location of the drone when it was taken. The server starts to process the stored image data. Taking two of the pictures as an example, the server extracts the relative information related to their shooting locations from the metadata of the pictures, and then uses the relevant algorithms and formulas of the geographic information system (GIS) to calculate the relative distances between the two shooting points in the horizontal and vertical directions. The server calculates the relative distances between all the shooting points of the pictures taken by the drones in the same way. After the server has obtained the relative distance between the shooting points of two adjacent pictures, it needs to calculate the pixel conversion ratio in combination with the lens parameters of the drone. For example, the focal length of the lens used by the drone is 35mm, the sensor size is 24mm×36mm, and the resolution is 4864x3648 pixels. First, based on the principle of optical imaging, the server uses the relationship of similar triangles and parameters such as the focal length and sensor size of the lens to calculate the conversion relationship between the relative distance and the distance on the imaging plane. Assuming that there is a certain relative distance between the shooting points of the two pictures in the horizontal and vertical directions, then, in order to obtain the pixel conversion ratio, the server performs the following calculation: In the horizontal direction, the width of the imaging plane corresponds to 4864 pixels, while the actual width of the sensor is 24mm, that is, 0.024 meters. Through the proportional relationship calculation, the number of pixels corresponding to 1 meter in the horizontal direction is approximately 4864÷0.024≈202667 pixels (here is only a principle calculation example). However, due to various errors and influencing factors in actual shooting, after correction and comprehensive calculation, the pixel conversion ratio in the horizontal direction is finally determined. Similarly, in the vertical direction, the height of the imaging plane corresponds to 3648 pixels, and the actual height of the sensor is 36mm, or 0.036 meters. Through proportional calculation, the number of pixels corresponding to 1 meter in the vertical direction is approximately 3648÷0.036≈101333 pixels. After correction and comprehensive consideration, the pixel conversion ratio in the vertical direction is finally determined (also based on the relative distance and the corrected relationship). Based on the pixel conversion ratio calculated previously, the server continues to calculate the pixel offset.For two example images, the pixel conversion ratios in the horizontal and vertical directions are known. According to the relative distances in the horizontal and vertical directions obtained by converting the relative distances between the shooting points calculated previously, the pixel offsets in the horizontal and vertical directions are calculated. The server adjusts the images according to these pixel offsets. For example, for image 2, the server moves it 9000 pixels to the left in the horizontal direction and 300 pixels upward in the vertical direction. The server adjusts all the drone-photographed images in the same way to obtain multiple adjusted drone-photographed images. The server has completed the position adjustment of all drone-photographed images and then performs the stitching operation. The server first selects an image as the reference image, for example, selects the image numbered 1. Then, according to the order in which the images were taken, the adjusted other images are stitched with the reference image in sequence. During the stitching process, the server merges and matches the overlapping parts of adjacent images according to the pixel offsets. Through the image algorithm, the color, brightness, contrast, etc. are finely adjusted so that the transition at the stitching is natural and there is no obvious stitching mark. Please refer to it in conjunction. Figure 2After a series of stitching processes, the server finally obtains a complete panoramic image of the target wind turbine blade. This panoramic image shows the whole picture of the wind turbine blade, including the surface, edge, texture and other details of the blade. After obtaining the panoramic image of the wind turbine blade, the server calls the pre-trained abnormality recognition model to analyze it. For example, this abnormality recognition model is a convolutional neural network (CNN) model based on deep learning, which has been trained on a large amount of wind turbine blade image data with abnormal parts marked. The server inputs the panoramic image of the blade into the abnormality recognition model. The model first extracts features from the input image, and extracts low-level features (such as edges, textures, etc.) and high-level features (such as shapes, patterns, etc.) in the image through multi-layer convolution and pooling operations. Then, the model classifies and locates according to the extracted features, determines whether there are abnormal parts in the image, and determines the location and type of the abnormal parts. For example, the model recognizes that there is a crack at the tip of the blade in the panoramic image of the blade, and outputs the location and type information of this abnormal part to the server as the recognition result. After obtaining the abnormal part recognition result, the server begins to analyze the estimated degradation of the corresponding surrounding parts. For example, the abnormal part identified is a crack at the tip of the blade. The server first obtains attribute information related to the abnormal part, such as the length, width, depth, etc. of the crack. At the same time, the server also obtains factor information of the operating environment of the wind turbine blade, such as wind speed, wind direction, temperature, humidity and other operating environment attributes. The server encodes the abnormal part attributes and the operating environment attributes through the encoding unit and converts them into parameter forms that can be calculated and analyzed. For example, the length of the crack is encoded as a value of 5 cm, and the wind speed is encoded as a value of 10 m / s. Then, the server integrates the abnormal part parameters and the operating environment parameters through the integration unit to obtain integrated feature parameters. For example, the crack length of 5 cm and the wind speed of 10 m / s are integrated into a comprehensive feature vector. The server inputs the integrated feature parameters into multiple professional knowledge networks to perform initial feature extraction processing. These professional knowledge networks can include networks based on physical models, networks based on statistical analysis, etc. Each network generates corresponding initial feature data. Then, the server determines the first weighting coefficient corresponding to each professional knowledge network according to the first decision component corresponding to each peripheral degradation prediction indication. For example, for a network based on a physical model, the first weighting coefficient is 0.6; for a network based on statistical analysis, the first weighting coefficient is 0.4. According to these weighting coefficients, the starting feature data generated by each professional knowledge network is weighted and summed to obtain the first summary feature data corresponding to each peripheral degradation estimation indication. Next, the server performs context-adaptive conversion on each first summary feature data through multiple evaluation anchor points to obtain the primary feature set generated by each evaluation anchor point. For example, evaluation anchor point 1 converts the first summary feature data into a new feature vector 1, and evaluation anchor point 2 converts the first summary feature data into another new feature vector 2.The server again determines the second weighting coefficient corresponding to each professional knowledge network according to the second decision component corresponding to each peripheral degradation prediction indication, and weights and sums the initial feature data generated by each professional knowledge network according to these weighting coefficients to obtain the second summary feature data corresponding to each peripheral degradation prediction indication. Afterwards, the server uses the temporal association model to perform temporal association feature extraction processing on each second summary feature data. For example, the temporal association model includes two temporal association secondary models, each of which includes a parallel attention mechanism unit and a forward conduction neural network unit. The server inputs each second summary feature data into the parallel attention mechanism unit for weight allocation calculation to obtain the weight allocation data corresponding to each second summary feature data. Then, these weight allocation data are input into the forward conduction neural network unit for feature refinement to obtain the secondary feature set generated by the temporal association model. For example, there are two peripheral degradation prediction indications, a first peripheral degradation prediction indication and a second peripheral degradation prediction indication, and the first peripheral degradation prediction indication precedes the second peripheral degradation prediction indication in timing. The server first performs a time series correlation feature extraction process on the second summary feature data corresponding to the first peripheral degradation prediction indication according to the time series correlation model to obtain a secondary feature set corresponding to the first peripheral degradation prediction indication. Then, the server performs a time series correlation feature extraction process on the second summary feature data corresponding to the first peripheral degradation prediction indication and the second summary feature data corresponding to the second peripheral degradation prediction indication according to the time series correlation model to obtain a secondary feature set corresponding to the second peripheral degradation prediction indication. The server determines the weighting coefficient corresponding to the primary feature set and the weighting coefficient corresponding to the secondary feature set according to the third decision component corresponding to each peripheral degradation prediction indication. For example, for the first peripheral degradation prediction indication, the weighting coefficient of the primary feature set is 0.7, and the weighting coefficient of the secondary feature set is 0.3. Then, according to these weighting coefficients, the primary feature set and the secondary feature set are weighted and summed to obtain the third summary feature data corresponding to each peripheral degradation prediction indication. Finally, the server performs a mapping process on the third summary feature data corresponding to each peripheral degradation prediction indication through a feature mapping unit to obtain the estimated degradation result of the peripheral part corresponding to each peripheral degradation prediction indication. For example, it is estimated that the area within 10 cm around the crack at the tip of the blade will further deteriorate in the next month, with a moderate degree of deterioration. In summary, through a series of complex and precise calculations and analyses, the server has completed the acquisition and splicing of drone-photographed images, the identification of abnormal parts, and the acquisition of estimated deterioration results of surrounding parts, providing strong support for the maintenance and overhaul of wind turbine blades.

[0029] In the embodiment of the present invention, the calling of the pre-trained abnormality recognition model to obtain the abnormal part recognition result corresponding to the leaf panoramic picture can be implemented through the following example.

[0030] Obtain leaf features corresponding to the leaf panoramic image;

[0031] Through the target recognition unit in the abnormal recognition model, according to multiple retrieval feature sets and the leaf features, the identification retrieval features corresponding to each retrieval feature in the multiple retrieval feature sets are obtained; the multiple retrieval feature sets are obtained by grouping each pending retrieval feature corresponding to the target recognition unit; the pending retrieval feature is obtained according to the original retrieval feature, or is selected from each identification retrieval feature generated by the previous level recognition unit of the target recognition unit based on a preset feature standard, and the target recognition unit is a first level recognition unit in the abnormal recognition model;

[0032] According to each recognition retrieval feature generated by the final recognition unit in the abnormality recognition model, the abnormal part recognition result corresponding to the leaf panoramic picture is obtained.

[0033] In an embodiment of the present invention, exemplarily, in a large wind farm monitoring system, a server receives a panoramic image of a wind turbine blade that has been stitched. The server first preprocesses the panoramic image of the blade, including operations such as denoising and contrast enhancement, so as to extract features more accurately. Then, the server uses a series of image processing algorithms and deep learning models to extract the features of the blade. For example, the server extracts features from the image through a convolutional neural network (CNN). CNN learns different levels of features from the image, such as edges, shapes, textures, etc. These features are combined into a high-dimensional feature vector, which represents the features of the panoramic image of the blade. The abnormal recognition model contains multiple target recognition units, each of which is responsible for processing specific features and retrieval tasks. For example, the current target recognition unit is a unit responsible for detecting crack features on the surface of the blade. First, the server obtains the pending retrieval features corresponding to this target recognition unit. These pending retrieval features can be obtained based on the original retrieval features, such as some pre-defined basic features about cracks, such as length, width, direction, etc.; or they can be selected from the recognition retrieval features generated by the previous level recognition unit based on preset feature standards. Then, the server divides these pending retrieval features into groups to form multiple retrieval feature sets. For example, the features related to the crack length are divided into one group, and the features related to the crack width are divided into another group. Next, the server inputs the features of the panoramic image of the blade and these retrieval feature sets into the target recognition unit. The target recognition unit generates corresponding identification retrieval features for each retrieval feature through complex calculations and analysis. For example, for the retrieval feature of the crack length, the target recognition unit may determine that the identification retrieval feature of the crack length is "longer" based on the input blade features; for the retrieval feature of the crack width, the generated identification retrieval feature may be "wider". As the abnormal recognition model is processed layer by layer, it finally reaches the final recognition unit. The final recognition unit integrates the identification retrieval features generated by the previous recognition units at all levels and makes the final judgment and analysis. For example, the final recognition unit receives a series of identification retrieval features such as "longer crack", "wider crack", and "crack direction is perpendicular to the blade axis". The server obtains the final abnormal part identification result through the processing logic of the final recognition unit. For example, the final recognition unit determines that there is a serious crack abnormality in the middle of the blade, with a length of about 50 cm and a width of about 5 mm, which poses a great threat to the performance and safety of the wind turbine blade. The server outputs and records the location (middle), type (crack), specific description (length, width, etc.) of the abnormal part as the final abnormal part identification result, so as to facilitate more in-depth analysis and processing later, such as evaluating the impact on surrounding parts and formulating maintenance plans.During the entire process, the server continuously performs data processing and model calculations, fully utilizing the powerful functions of the anomaly recognition model to accurately identify abnormal parts from the panoramic images of the blades, providing key information support for ensuring the normal operation of the wind farm and the maintenance of wind turbine blades.

[0034] In an embodiment of the present invention, when the target recognition unit is a recognition unit other than the first level in the abnormal recognition model, each pending retrieval feature corresponding to the target recognition unit is obtained by the following process, which can be implemented through the following example.

[0035] According to each recognition retrieval feature generated by the previous level recognition unit, obtaining an abnormal part recognition result corresponding to each recognition retrieval feature; the abnormal part recognition result includes a recognition confidence corresponding to the recognition part indicated by the recognition retrieval feature;

[0036] According to the recognition confidence in the abnormal part recognition result corresponding to each of the recognition retrieval features, the recognition retrieval feature that meets the preset feature standard is selected from each of the recognition retrieval features and determined as the pending retrieval feature corresponding to the target recognition unit.

[0037] In an embodiment of the present invention, illustratively, the server is processing a panoramic image of a wind turbine blade in a large wind farm to detect possible anomalies. The server has run the previous recognition units of the anomaly recognition model. For example, the previous recognition unit generates the following recognition retrieval features: Feature A represents the color change on the blade surface, Feature B represents the irregularity of the blade texture, and Feature C represents the wear on the blade edge. The server performs further analysis and judgment on each recognition retrieval feature. For Feature A, the server uses complex algorithms and models to obtain the recognition result of the abnormal part corresponding to the feature as "the local color of the blade surface darkens, there may be slight corrosion, and the recognition confidence is 0.7". For Feature B, the recognition result is "the blade texture is obviously interrupted and chaotic, there may be internal damage, and the recognition confidence is 0.85". For Feature C, the recognition result is "there are a few traces of wear on the blade edge, which may be normal wind erosion, and the recognition confidence is 0.4". The server pre-sets preset feature standards, for example, features with a recognition confidence greater than 0.7 are considered important and require further analysis. In the above example, the recognition confidence of feature A is 0.7, which just meets the preset standard; the recognition confidence of feature B is 0.85, which is significantly higher than the preset standard; the recognition confidence of feature C is 0.4, which does not meet the preset standard. Therefore, the server selects feature A and feature B and determines them as pending retrieval features corresponding to the current target recognition unit. However, feature C is excluded from the pending retrieval features due to its low recognition confidence. Next, the server further processes and analyzes these pending retrieval features to provide a basis for more accurate identification of abnormal parts of the blade. For example, in another scenario, the previous level recognition unit generates feature D to represent tiny protrusions on the surface of the blade, feature E to represent the peeling of the blade coating, and feature F to represent signs of looseness at the connection of the blade. For feature D, the abnormal part identification result obtained by the server is "there are sporadic tiny protrusions on the surface of the blade, which may be a material defect, and the recognition confidence is 0.65". For feature E, the recognition result is "a large area of ​​peeling of the blade coating may affect the protection performance, and the recognition confidence is 0.8". For feature F, the recognition result is "the blade connection is slightly loose, which requires close attention, and the recognition confidence is 0.75". According to the preset feature standard (recognition confidence is greater than 0.7), the server selects features E and F as pending retrieval features corresponding to the target recognition unit, and feature D is excluded because the recognition confidence does not meet the standard. In this way, the server can select important features for subsequent analysis based on the previous level of recognition results, thereby improving the accuracy and efficiency of abnormality recognition. In actual large-scale wind farm monitoring, the server needs to process image data of hundreds or thousands of wind turbine blades at the same time.For each panoramic image of the blade, such a detailed analysis and screening process must be carried out to ensure that any abnormal situation that may affect the normal operation and safety of the wind turbine is discovered in a timely and accurate manner. For example, another new panoramic image of the wind turbine blade is transmitted to the server. The previous level recognition unit generates feature G to represent scratches on the blade surface, feature H to represent slight deformation of the blade shape, and feature I to represent suspected voids inside the blade. The server's analysis result for feature G is "deep scratches on the blade surface may weaken the structural strength, and the recognition confidence is 0.82". The analysis result of feature H is "slight deformation of the blade shape may be caused by uneven force, and the recognition confidence is 0.72". The analysis result of feature I is "there are suspected small voids inside the blade, which need further confirmation, and the recognition confidence is 0.6". According to the preset feature standard, the server determines that feature G and feature H are pending retrieval features corresponding to the target recognition unit, and feature I is temporarily not listed as a pending retrieval feature due to its relatively low recognition confidence. As more and more panoramic images of wind turbine blades are processed, the server continues to accumulate experience and data, optimize the preset feature standards and recognition models, and improve the detection capability and accuracy of wind turbine blade anomalies. For example, after processing a large amount of data, the server found that for certain types of wind turbine blades, the originally set recognition confidence threshold needs to be adjusted appropriately. Alternatively, different recognition retrieval features are given different weights according to different wind turbine models, operating environments and other factors to more accurately screen out valuable pending retrieval features. In a complex wind farm environment, the server received a panoramic image of a wind turbine blade that has been in operation for many years. The previous level recognition unit gave feature J to represent rust on the blade surface, feature K to represent the bending of the blade tip, and feature L to represent crack signs at the root of the blade. The result of the server's analysis of feature J is "there are large areas of rust on the blade surface, which may accelerate corrosion, and the recognition confidence is 0.78". The result of feature K is "the blade tip is obviously bent, affecting aerodynamic performance, and the recognition confidence is 0.88". The result of feature L is "there are fine cracks at the root of the blade, which poses a safety hazard, and the recognition confidence is 0.75". According to the current preset feature standard, the server selects feature K and feature J as pending retrieval features for subsequent processing. Although feature L is also of certain importance, it will be considered again in the subsequent comprehensive evaluation because its recognition confidence is slightly lower than the standard, or used as auxiliary information to more comprehensively judge the overall condition of the blade. The server continuously performs such in-depth and detailed analysis and screening on each panoramic image of the blade, providing accurate and reliable blade abnormality information to the operation and maintenance personnel of the wind farm, ensuring the stable operation and efficient power generation of the wind farm.

[0038] In an embodiment of the present invention, the loaded content of the previous-level recognition unit includes a plurality of retrieval feature sets obtained by grouping the pending retrieval features corresponding to the previous-level recognition unit; the generated content of the previous-level recognition unit includes an identification retrieval feature set corresponding one-to-one to the plurality of retrieval feature sets; according to the recognition confidence in the abnormal part recognition result corresponding one-to-one to each of the identification retrieval features, an identification retrieval feature that meets the preset feature standard is selected from each of the identification retrieval features and determined as the pending retrieval feature corresponding to the target recognition unit, which can be implemented through the following examples.

[0039] For each of the identification retrieval feature sets, based on the recognition confidence in the abnormal part recognition result corresponding to each identification retrieval feature, the identification retrieval feature that meets the preset feature standard is selected from the identification retrieval feature set and determined as the pending retrieval feature corresponding to the target recognition unit.

[0040] In an embodiment of the present invention, illustratively, in a large wind farm monitoring system, the server is processing a series of image data of wind turbine blades. The server first obtains the relevant information of the previous level recognition unit. The loaded content of the previous level recognition unit includes multiple retrieval feature sets obtained by grouping the corresponding pending retrieval features. For example, three retrieval feature sets are divided: set A includes features related to blade surface damage, set B includes features related to blade structural deformation, and set C includes features related to blade material aging. The generated content of the previous level recognition unit includes the recognition retrieval feature sets corresponding to these multiple retrieval feature sets. For set A, the generated recognition retrieval features can be the length, depth, number, etc. of the scratches on the blade surface; for set B, it can be the angle, position, etc. of the blade bending; for set C, it can be the degree of material fading, hardness change, etc. Next, the server will select the recognition retrieval features that meet the preset feature standards from each recognition retrieval feature according to the recognition confidence in the abnormal part recognition result corresponding to each recognition retrieval feature, and determine it as the pending retrieval feature corresponding to the target recognition unit. Take set A as an example, the identification retrieval features included therein are: feature a1 indicates that the length of the scratch on the blade surface is 10 cm, and the corresponding identification confidence in the abnormal part identification result is 0.8; feature a2 indicates that the depth of the scratch on the blade surface is 2 mm, and the corresponding identification confidence is 0.6; feature a3 indicates that the number of scratches on the blade surface is 5, and the corresponding identification confidence is 0.7. The feature standard preset by the server is that the identification confidence is greater than 0.7. Then, in set A, features a1 and a3 meet the preset feature standards and are selected as pending retrieval features corresponding to the target identification unit, while feature a2 is not selected because it does not meet the standard due to its identification confidence of 0.6. Let's look at set B again, where the identification retrieval features are: feature b1 indicates that the blade bending angle is 10 degrees, and the corresponding identification confidence is 0.75; feature b2 indicates that the blade bending position is in the middle, and the corresponding identification confidence is 0.9; feature b3 indicates that the blade bending influence range is 30 cm, and the corresponding identification confidence is 0.65. According to the preset standard, features b1 and b2 in set B are selected as pending retrieval features, and feature b3 is not selected. For set C, for example, the identification retrieval features are: feature c1 indicates that the degree of material fading is 30%, and the corresponding identification confidence is 0.7; feature c2 indicates that the material hardness decreases by 20%, and the corresponding identification confidence is 0.85; feature c3 indicates that cracks appear due to aging of the material, and the corresponding identification confidence is 0.6. Then features c1 and c2 in set C are determined as pending retrieval features, and feature c3 is not determined. By analyzing and screening each retrieval feature set one by one, the server accurately determines the pending retrieval features that meet the preset standards, providing a valuable data basis for the subsequent processing of the target recognition unit. For example, a new set of wind turbine blade images was processed, and the retrieval feature sets divided by the previous level recognition unit were different.For example, set D contains features related to blade attachments, set E contains features related to blade temperature distribution, and set F contains features related to blade vibration frequency. The identification retrieval features in set D are: feature d1 indicates that the area of ​​blade attachments is 5 square centimeters, and the corresponding identification confidence is 0.78; feature d2 indicates that the weight of the attachments is 10 grams, and the corresponding identification confidence is 0.62; feature d3 indicates that the attachments are located at the tip of the blade, and the corresponding identification confidence is 0.8. According to the preset standard, features d1 and d3 in set D are determined as pending retrieval features, and feature d2 is not determined. The identification retrieval features in set E are: feature e1 indicates that the area of ​​the high temperature area of ​​the blade is 20 square centimeters, and the corresponding identification confidence is 0.85; feature e2 indicates that the maximum temperature is 80 degrees Celsius, and the corresponding identification confidence is 0.72; feature e3 indicates that the uneven temperature distribution is 30%, and the corresponding identification confidence is 0.68. In set E, features e1 and e2 are determined as pending retrieval features, and feature e3 is not determined. The identification and retrieval features in set F are: feature f1 indicates that the vibration frequency of the blade is 100 Hz, and the corresponding identification confidence is 0.7; feature f2 indicates that the vibration amplitude is 5 mm, and the corresponding identification confidence is 0.82; feature f3 indicates that the abnormal vibration area is at the root of the blade, and the corresponding identification confidence is 0.75. Features f2 and f3 in set F are determined as pending retrieval features, and feature f1 is not determined. In the process of continuously processing various wind turbine blade images, the server continues to perform such detailed analysis and screening. For each new data set, the preset feature standards and processes are strictly followed to accurately select valuable pending retrieval features from multiple retrieval feature sets, providing strong support for timely and accurate detection of abnormal conditions of wind turbine blades. For example, a more complex set of wind turbine blade image data is processed. Set G contains features related to blade coating wear, set H contains features related to the internal stress distribution of the blade, and set I contains features related to the cleanliness of the blade surface. The identification and retrieval features in set G are: feature g1 indicates that the coating wear depth is 1 mm, and the corresponding identification confidence is 0.76; feature g2 indicates that the coating wear area is 8 square centimeters, and the corresponding identification confidence is 0.88; feature g3 indicates that the coating wear position is at the leading edge of the blade, and the corresponding identification confidence is 0.72. According to the preset standard, features g2 and g1 in set G are determined as pending retrieval features, and feature g3 is not determined. The identification and retrieval features in set H are: feature h1 indicates that the maximum internal stress is 500 MPa, and the corresponding identification confidence is 0.8; feature h2 indicates that the stress concentration area is in the middle of the blade, and the corresponding identification confidence is 0.75; feature h3 indicates that the stress distribution unevenness coefficient is 0.3, and the corresponding identification confidence is 0.65. In set H, features h1 and h2 are determined as pending retrieval features, and feature h3 is not determined.The identification and retrieval features in set I are: feature i1 indicates that the dust coverage of the blade surface is 20%, and the corresponding identification confidence is 0.7; feature i2 indicates that the surface oil area is 3 square centimeters, and the corresponding identification confidence is 0.68; feature i3 indicates that the overall score of surface cleanliness is 7 points (out of 10 points), and the corresponding identification confidence is 0.82. Features i1 and i3 in set I are determined as pending retrieval features, and feature i2 is not determined. With the continuous increase in the amount of data processed, the server has accumulated rich experience and can more accurately select pending retrieval features based on the identification confidence and preset standards. At the same time, the server can also optimize and adjust the preset standards based on these large amounts of data to meet the needs of abnormal detection of wind turbine blades of different types and operating conditions. For example, a set of wind turbine blade images from an offshore wind farm is processed. Set J contains features related to the degree of blade corrosion, set K contains features related to traces of lightning strikes on blades, and set L contains features related to the impact of marine salt spray on blades. The identification and retrieval features in set J are: feature j1 indicates that the corrosion depth of the blade is 0.5 mm, and the corresponding identification confidence is 0.78; feature j2 indicates that the corrosion area is 5 square centimeters, and the corresponding identification confidence is 0.85; feature j3 indicates that the corrosion location is at the root of the blade, and the corresponding identification confidence is 0.7. According to the preset standard, features j2 and j1 in set J are determined as pending retrieval features, and feature j3 is not determined. The identification and retrieval features in set K are: feature k1 indicates that the length of the lightning strike mark is 3 centimeters, and the corresponding identification confidence is 0.8; feature k2 indicates the material change in the lightning damage area, and the corresponding identification confidence is 0.72; feature k3 indicates that the lightning strike occurs at the tip of the blade, and the corresponding identification confidence is 0.9. In set K, features k1 and k3 are determined as pending retrieval features, and feature k2 is not determined. The identification and retrieval features in set L are: feature l1 indicates that the thickness of salt spray deposition is 0.2 mm, and the corresponding identification confidence is 0.75; feature l2 indicates the degree of material performance degradation caused by salt spray, and the corresponding identification confidence is 0.65; feature l3 indicates the distribution of the salt spray affected area on the blade surface, and the corresponding identification confidence is 0.82. Features l1 and l3 in set L are determined as pending retrieval features, and feature l2 is not determined. By processing these wind turbine blade image data of different types and scenarios, the server continuously improves and optimizes its anomaly detection algorithm and process, improves the accuracy and reliability of wind turbine blade abnormality detection, and provides a strong guarantee for the safe and stable operation of wind farms.

[0041] In an embodiment of the present invention, when the target recognition unit is a first-level recognition unit in the abnormality recognition model, each pending retrieval feature corresponding to the target recognition unit is obtained by the following process, which can be implemented through the following example.

[0042] Determine each of the original search features as each pending search feature corresponding to the target recognition unit;

[0043] Alternatively, the feature reliability corresponding to each of the original search features is obtained; based on the feature reliability corresponding to each of the original search features, the original search features that meet the feature screening criteria are selected from each of the original search features and determined as the pending search features corresponding to the target recognition unit.

[0044] In an embodiment of the present invention, for example, in a monitoring center of a large wind farm, the server is preparing to perform abnormality recognition processing on a newly received wind turbine blade image. When the target recognition unit is the first-level recognition unit in the abnormality recognition model, there are two ways to determine the pending retrieval features. The first way: the server directly determines each original retrieval feature as each pending retrieval feature corresponding to the target recognition unit. For example, the original retrieval features include the color, texture, shape, surface flatness, etc. of the blade. The server does not perform additional screening and judgment, and directly identifies all of these original retrieval features as pending retrieval features for subsequent more in-depth analysis and processing. For example, the color features in the original retrieval features include the overall hue and color uniformity of the blade; the texture features include the clarity and regularity of the texture; the shape features include the curvature and symmetry of the blade; the surface flatness features include whether there are convexities and concavities. In this way, the server does not consider any other attributes of these features, such as reliability or importance, and directly includes them all in the category of pending retrieval features, providing comprehensive basic data for subsequent model processing. The second way: the server first obtains the feature reliability corresponding to each original retrieval feature. For example, for the color feature, the server determines that its feature reliability is 0.8 (full score is 1) by analyzing historical data and prior knowledge, which means that the color feature has a high reliability in abnormality recognition. For the texture feature, the feature reliability is evaluated as 0.7; for the shape feature, the feature reliability is 0.9; and for the surface flatness feature, the feature reliability is 0.75. Then, the server selects the pending retrieval features according to the pre-set feature screening criteria. For example, the feature screening criteria is that the feature reliability is greater than 0.75. Based on this criterion, the server selects the original retrieval features that meet the criteria from the above original retrieval features. In this example, the color feature, shape feature, and surface flatness feature are determined as the pending retrieval features corresponding to the target recognition unit, while the texture feature is temporarily not determined as the pending retrieval feature because its feature reliability is 0.7 and does not meet the screening criteria. In another processing, the original retrieval features are different, including the material hardness of the blade, the coating thickness, the tightness of the connection, etc. The server obtains the feature reliability of 0.6 for the material hardness, 0.85 for the coating thickness, and 0.78 for the tightness of the connection. For example, the feature screening standard is still feature reliability greater than 0.75. Then, in this process, coating thickness and connection tightness are determined as pending retrieval features, while material hardness is not determined as a pending retrieval feature because its feature reliability is 0.6 and does not meet the standard.In this way, the server can select more valuable and reliable original retrieval features as pending retrieval features in a targeted manner according to the reliability of the features and the preset screening criteria, providing a more accurate and effective data basis for the subsequent abnormal recognition model processing, thereby improving the accuracy and efficiency of abnormal recognition of wind turbine blades. For example, there is another set of new original retrieval features, such as the vibration frequency, sound characteristics, and heating conditions of the blades. The server evaluates the feature reliability of the vibration frequency to be 0.72, the feature reliability of the sound characteristics to be 0.8, and the feature reliability of the heating condition to be 0.78. If the feature screening criteria remain unchanged (greater than 0.75), the sound characteristics and heating conditions are determined as pending retrieval features, and the vibration frequency is not determined. In actual large-scale wind farm monitoring, the server needs to continuously process various different wind turbine blade images and data. For each new processing task, the pending retrieval features must be accurately determined based on the specific original retrieval features and their corresponding feature reliability, as well as the preset screening criteria, to ensure the effective operation and accurate judgment of the abnormal recognition model. For example, when processing a set of image data from a new model of wind turbine blades, the original retrieval features include the surface roughness of the blades, the integrity of the internal fiber structure, the chemical stability of the protective coating, etc. The server evaluates the feature reliability of the surface roughness to be 0.82, the feature reliability of the internal fiber structure integrity to be 0.9, and the feature reliability of the chemical stability of the protective coating to be 0.7. According to the feature screening criteria (greater than 0.75), the surface roughness and the internal fiber structure integrity are determined as pending retrieval features, and the chemical stability of the protective coating is not determined. With the continuous development of technology and the accumulation of data, the server will adjust the evaluation method and screening criteria of feature reliability to adapt to different detection needs and improve the performance of abnormal identification. For example, after one optimization, for another set of original retrieval features, such as the electromagnetic characteristics, wind resistance index, lubrication condition, etc. of the blades, the server re-evaluates its feature reliability and determines the pending retrieval features according to the updated screening criteria (for example, increased to greater than 0.8). Through continuous optimization and improvement, the server can better cope with the complex and changeable wind turbine blade detection tasks, detect potential abnormalities in a timely manner, and ensure the safe and stable operation of wind farms.

[0045] In the embodiment of the present invention, the retrieval feature set is obtained by the following process, which can be implemented through the following examples.

[0046] According to a predetermined number of group divisions, each pending retrieval feature corresponding to the target recognition unit is evenly distributed to the multiple retrieval feature sets;

[0047] Obtaining the feature reliability of each pending retrieval feature corresponding to the target recognition unit; dividing each pending retrieval feature into the plurality of retrieval feature sets according to the feature reliability of each pending retrieval feature according to the predetermined number of group divisions;

[0048] The feature space position of each pending retrieval feature corresponding to the target recognition unit in the feature domain is obtained; and each pending retrieval feature is divided into the plurality of retrieval feature sets according to the feature space position of each pending retrieval feature in the feature domain according to a predetermined number of group divisions.

[0049] In an embodiment of the present invention, illustratively, in a large wind farm monitoring system, the server is processing the abnormal identification task of wind turbine blades. First, the server evenly distributes each pending retrieval feature corresponding to the target identification unit to multiple retrieval feature sets according to the predetermined number of group divisions. For example, the predetermined number of group divisions is 3, and there are 15 pending retrieval features, namely features A1 to A15. The server will evenly distribute these 15 features to 3 retrieval feature sets, that is, each set contains approximately 5 features. For example, retrieval feature set 1 contains features A1, A2, A3, A4, and A5; retrieval feature set 2 contains features A6, A7, A8, A9, and A10; retrieval feature set 3 contains features A11, A12, A13, A14, and A15. Next, the server obtains the feature reliability of each pending retrieval feature corresponding to the target identification unit. For example, the feature reliability of feature A1 is 0.8, the feature reliability of feature A2 is 0.7, the feature reliability of feature A3 is 0.9, and so on. Then, the server divides each pending retrieval feature into multiple retrieval feature sets according to the predetermined number of group divisions and the feature reliability corresponding to each pending retrieval feature. For example, the division rule set by the server is: features with feature reliability higher than 0.85 are divided into retrieval feature set 1, features with feature reliability between 0.7 and 0.85 are divided into retrieval feature set 2, and features with feature reliability lower than 0.7 are divided into retrieval feature set 3. Based on this rule, feature A1 (reliability 0.8) and feature A2 (reliability 0.7) will be divided into retrieval feature set 2, and feature A3 (reliability 0.9) will be divided into retrieval feature set 1. Then, the server obtains the feature space position of each pending retrieval feature corresponding to the target recognition unit in the feature domain. For example, the feature space positions of features A1 to A15 in the feature domain are (10, 20), (30, 40), (50, 60), (70, 80), (90, 100), (110, 120), (130, 140), (150, 160), (170, 180), (190, 200), (210, 220), (230, 240), (250, 260), (270, 280), and (290, 300). The server divides each pending retrieval feature into multiple retrieval feature sets according to the feature space position of each pending retrieval feature in the feature domain according to the predetermined number of group divisions. For example, if the division rule set by the server is: features with a horizontal coordinate of the feature space position less than 100 are divided into retrieval feature set 1, features with a horizontal coordinate between 100 and 200 are divided into retrieval feature set 2, and features with a horizontal coordinate greater than 200 are divided into retrieval feature set 3.Then, features A1, A2, A3, A4, and A5 will be divided into retrieval feature set 1, features A6, A7, A8, A9, and A10 will be divided into retrieval feature set 2, and features A11, A12, A13, A14, and A15 will be divided into retrieval feature set 3. In another processing, the server faces different pending retrieval features. For example, there are 20 pending retrieval features, namely features B1 to B20. The predetermined number of group divisions is still 3. The server first distributes these 20 features roughly evenly into 3 retrieval feature sets. For example, retrieval feature set 1 contains features B1 to B7, retrieval feature set 2 contains features B8 to B14, and retrieval feature set 3 contains features B15 to B20. Then the feature reliabilities of these features are obtained. For example, the feature reliability of feature B1 is 0.75, the feature reliability of feature B2 is 0.92, the feature reliability of feature B3 is 0.68, and so on. According to the newly set division rule based on feature reliability, for example, the feature reliability higher than 0.8 is divided into retrieval feature set 1, the feature reliability between 0.6 and 0.8 is divided into retrieval feature set 2, and the feature reliability lower than 0.6 is divided into retrieval feature set 3, and these features are re-divided. Then, their feature space positions in the feature domain are obtained. For example, the feature space position of feature B1 is (50,50), the feature space position of feature B2 is (150,150), and so on. According to the newly set division rule based on feature space position, for example, the feature space position with a horizontal coordinate less than 120 is divided into retrieval feature set 1, the feature reliability between 120 and 200 is divided into retrieval feature set 2, and the feature reliability greater than 200 is divided into retrieval feature set 3, and these features are divided again. In this way, the server can comprehensively consider factors such as the number of pending retrieval features, feature reliability, and feature space position, and reasonably divide them into different retrieval feature sets, providing a more targeted and effective data organization method for subsequent abnormality identification and analysis. In actual operation, the server will continuously adjust the division rules and the number of group divisions to adapt to different types of wind turbine blades, different detection requirements, and changing environmental conditions. For example, when processing a batch of images from a specific model of wind turbine blades, the server found that the previous division method was not accurate enough, so it reset the division thresholds of feature reliability and feature spatial position based on the characteristics of this batch of blades to improve the accuracy and efficiency of anomaly identification. For another example, when the environmental conditions of a wind farm change significantly, such as a significant increase in wind speed or a more severe climate, the server will increase the number of group divisions to analyze the characteristics of the blades in more detail and promptly detect new types of anomalies that may appear. In short, the server continuously optimizes the anomaly detection and analysis capabilities of wind turbine blades through a flexible and precise feature set division method to ensure the stable operation and safety of wind farms.

[0050] In an embodiment of the present invention, the target recognition unit in the abnormal recognition model obtains the identification retrieval feature corresponding to each retrieval feature in the multiple retrieval feature sets according to the multiple retrieval feature sets and the leaf features, which can be implemented through the following examples.

[0051] By means of the target recognition unit, based on a multi-layer perceptron, according to each of the retrieval feature sets, a candidate retrieval feature set corresponding to each of the retrieval feature sets is obtained; the candidate retrieval feature set includes candidate retrieval features corresponding to each of the retrieval features in the corresponding retrieval feature set;

[0052] Through the target recognition unit, based on the feedforward network, according to each of the candidate retrieval feature sets and the leaf features, the recognition retrieval feature set corresponding to each of the candidate retrieval feature sets is obtained; the recognition retrieval feature set includes a one-to-one corresponding recognition retrieval feature for each candidate retrieval feature in the corresponding candidate retrieval feature set.

[0053] In an embodiment of the present invention, exemplarily, in an advanced wind farm monitoring system, the server is focusing on an in-depth analysis of a newly acquired wind turbine blade image to detect possible anomalies. The server first works through the target recognition unit in the anomaly recognition model. For example, the current task of the target recognition unit is to analyze the surface damage characteristics of the blade. In the first step, the server obtains candidate retrieval feature sets corresponding to each retrieval feature set based on the multi-layer perceptron through the target recognition unit according to each retrieval feature set. For example, the current retrieval feature set includes a length feature set, a depth feature set, and a distribution feature set of scratches on the blade surface. The multi-layer perceptron receives these retrieval feature sets as input and performs complex calculations and pattern recognition. For the length feature set, the original retrieval features can be a series of specific length values, such as 5 cm, 8 cm, 10 cm, etc. After processing by the multi-layer perceptron, the corresponding candidate retrieval feature set is output, which can include candidate retrieval features such as "shorter scratches", "medium length scratches", and "longer scratches", and each candidate retrieval feature corresponds to the input original retrieval feature one by one. For the depth feature set, the original retrieval features are, for example, 1 mm, 2 mm, 3 mm, etc. After being processed by the multi-layer perceptron, the candidate retrieval feature set obtained can be "shallow scratches", "medium scratches", "deep scratches". If the original retrieval features of the distribution feature set are "concentrated at the front end of the blade", "evenly distributed on the blade surface", "mainly in the middle of the blade", etc., the processed candidate retrieval feature set can be "local concentrated distribution", "overall uniform distribution", "central dense distribution". In the second step, the server obtains the identification retrieval feature set corresponding to each candidate retrieval feature set based on the feedforward network and each candidate retrieval feature set and the leaf features through the target recognition unit. The feedforward network receives the candidate retrieval feature set generated above and the comprehensive features extracted from the leaf image, such as color, texture, shape and other overall leaf features. Taking the candidate retrieval feature of "shorter scratches" as an example, the feedforward network analyzes the overall leaf features, and the possible output identification retrieval feature is "minor surface damage". For "medium length scratches", the possible output is "moderate surface damage risk". For "longer scratches", it can be "suspicion of serious surface damage". Similarly, for the candidate retrieval features of the depth feature set "shallow scratches", "medium scratches", and "deep scratches", the corresponding identification retrieval feature sets may be "slight surface wear", "may affect blade performance", and "significantly reduce blade structural strength", respectively. The candidate retrieval features of the distribution feature set "local concentrated distribution", "overall uniform distribution", and "central intensive distribution" may correspond to the identification retrieval feature sets "risk of local uneven force", "overall condition is relatively stable", and "the middle part needs special attention". In another processing, for example, the task of the target recognition unit is to analyze the material aging characteristics of the blade. The retrieval feature set includes the material hardness feature set, the elasticity feature set, and the brittleness feature set.Through the multi-layer perceptron, for the material hardness feature set, the original retrieval features such as 50 Rockwell hardness, 60 Rockwell hardness, 70 Rockwell hardness, etc., can be converted into candidate retrieval feature sets such as "low hardness", "medium hardness", and "high hardness". Through further processing by the feedforward network and combined with the blade features, "low hardness" can correspond to the identification retrieval feature "obvious material aging", "medium hardness" can correspond to "certain signs of aging", and "high hardness" can correspond to "good material performance". Similar processing and transformation are also performed for the elasticity feature set and the embrittlement degree feature set to generate meaningful identification retrieval features. Through such a fine and gradually in-depth analysis process, the server can extract valuable information from complex blade features, accurately identify potential anomalies and problems, and provide strong support for the safe operation and maintenance of wind farms. In actual large-scale wind farm monitoring, the server needs to repeat this processing process continuously to deal with a large amount of wind turbine blade image data. Each processing can accumulate experience and data, and continuously optimize the performance and accuracy of the model. For example, when processing a batch of new wind turbine blade images, it was found that the surface damage characteristics of some blades were unique. The server will adjust the parameters of the multi-layer perceptron and feedforward network based on these new data, so that the subsequent analysis is more in line with the actual situation. Or, when it detects changes in environmental conditions, such as significant fluctuations in temperature and humidity, the server will update the analysis mode of the aging characteristics of the blade material accordingly to more accurately capture possible anomalies. In short, the server ensures the stable operation and timely maintenance of wind turbine blades in wind farms through efficient and accurate anomaly recognition model processing.

[0054] In the embodiment of the present invention, each layer of recognition unit in the abnormal recognition model is the target recognition unit;

[0055] Alternatively, the previous preset order of magnitude identification unit in the abnormality identification model is the target identification unit.

[0056] In an embodiment of the present invention, for example, in a large wind farm monitoring system, the server is using an abnormal recognition model to analyze the image of a wind turbine blade. For example, the abnormal recognition model has a total of five layers of recognition units. In the first case, each layer of recognition unit in the abnormal recognition model is a target recognition unit. The server starts processing from the first layer of recognition units. The first layer of recognition units is responsible for preliminary analysis of the basic appearance features of the blade, such as the color and overall shape of the blade. It converts these features into corresponding retrieval features. Then, the second layer of recognition units receives the output of the first layer, analyzes and processes more detailed features such as the texture and surface flatness of the blade, and also generates corresponding retrieval features. The third layer of recognition units may focus on the edge condition of the blade, such as whether there is wear and deformation. The fourth layer of recognition units goes further and analyzes possible structural problems inside the blade. Finally, the fifth layer of recognition units combines the retrieval features of all previous layers to give the final abnormal part recognition result. For example, the first layer of recognition units finds that the color of the blade is locally uneven and passes it to the second layer as a retrieval feature. Based on this and the analysis of the texture, the second layer of recognition units finds that the texture is abnormally interrupted in some areas and continues to pass it to the third layer. The third-layer recognition unit confirms that the blade edge is slightly worn and passes it to the fourth layer. The fourth-layer recognition unit, through a more in-depth analysis, suspects that there are tiny cracks inside the blade. The fifth-layer recognition unit combines this information and finally determines that the blade has a certain degree of abnormality and needs further inspection and maintenance. In the second case, the first preset order of recognition units in the abnormality recognition model are the target recognition units. For example, the preset number is three layers. The server first focuses on the first three layers of recognition units. The first-layer recognition unit still analyzes the basic appearance characteristics of the blade, such as finding that the overall shape of the blade is slightly distorted. The second-layer recognition unit conducts a detailed analysis of the surface quality of the blade and finds that there are many small scratches on the surface. The third-layer recognition unit focuses on analyzing the connection parts of the blade and finds that the connection parts show signs of looseness. The fourth and fifth-layer recognition units can integrate, verify, or perform some auxiliary processing on the results of the previous three layers. For example, the fourth-layer recognition unit integrates and preliminarily judges the retrieval features of the first three layers, and the fifth-layer recognition unit gives the final accurate abnormality recognition results and related suggestions based on more comprehensive considerations and more complex algorithms. In actual operation, the server will flexibly select the above two methods according to different needs and performance characteristics of the model to effectively identify abnormal conditions of wind turbine blades and provide reliable protection for the stable operation of wind farms.

[0057] In the embodiment of the present invention, obtaining the estimated degradation result of the corresponding surrounding part according to the abnormal part identification result can be implemented through the following example.

[0058] Obtain the abnormal part attribute corresponding to the abnormal part identification result and the operating environment attribute corresponding to the fan blade flying operating environment factor;

[0059] Performing initial feature extraction processing on the abnormal part attribute and the operating environment attribute according to multiple professional knowledge networks to obtain initial feature data generated by each of the professional knowledge networks;

[0060] According to the evaluation anchor points corresponding to the multiple peripheral degradation prediction indications, the initial feature data generated by each of the professional knowledge networks are subjected to context adaptation conversion to obtain the primary feature set generated by each of the evaluation anchor points; and according to the time series association model, the initial feature data generated by each of the professional knowledge networks are subjected to time series association feature extraction processing to obtain the secondary feature set generated by the time series association model; the peripheral degradation prediction indication is used to indicate the simulation demonstration of the abnormal part identification result on the wind turbine blade operating environment factor; there is a time series association relationship between the multiple peripheral degradation prediction indications;

[0061] A multi-directional prediction process is performed according to the primary feature set and the secondary feature set to obtain the predicted degradation results of the surrounding parts corresponding to each of the peripheral degradation prediction indications.

[0062] In an embodiment of the present invention, illustratively, in a large wind farm, the server receives the identification result of the abnormal part of the wind turbine blade. The server first obtains the identification result of the abnormal part. For example, the abnormal part is a large crack on the surface of the blade, and its abnormal part attributes include detailed information such as the length, width, depth, position and shape of the crack. At the same time, the server obtains the operating environment attributes corresponding to the operating environment factors of the wind turbine blade, such as the current wind speed, wind direction, temperature, humidity, and the operating time and power of the wind turbine. The server calls multiple professional knowledge networks, which may include networks based on material science, networks based on mechanical analysis, networks based on meteorological data, etc. After receiving the abnormal part attributes and operating environment attributes, the network based on material science extracts the initial feature data related to material fatigue and degradation, such as predicting the possibility of further cracking of the material based on the morphology and material properties of the crack. The network based on mechanical analysis extracts the initial feature data about the change in the force distribution of the blade based on the size, position and wind speed of the crack, and judges the impact of this force change on the area around the crack. The network based on meteorological data extracts the initial feature data that may accelerate or slow down the degradation of the blade based on environmental factors such as temperature and humidity. For example, there are multiple peripheral degradation prediction indicators, such as short-term degradation prediction indicators, medium-term degradation prediction indicators, and long-term degradation prediction indicators. Each prediction indicator has a corresponding evaluation anchor point. The evaluation anchor point of the short-term degradation prediction indicator performs a context-adaptive transformation on the starting feature data. For example, it may focus on the possibility of crack expansion under the current wind speed and temperature conditions in the next week. After the transformation, the primary feature set generated by the short-term evaluation anchor point is obtained. The evaluation anchor point of the medium-term degradation prediction indicator considers the impact of seasonal changes on blade degradation in the next month and generates the corresponding primary feature set. The evaluation anchor point of the long-term degradation prediction indicator integrates more long-term factors, such as the natural aging of materials, the overall operation plan of the wind turbine, etc., to obtain its primary feature set. At the same time, the time series association model performs time series association feature extraction processing on the starting feature data generated by each professional knowledge network. For example, the time series association model contains the degradation data of the same wind turbine blade in different seasons and different operating conditions in the past year. It associates the current starting feature data with the historical data and extracts the secondary feature set related to the time series. For example, it was found that the characteristics of the current cracks were related to the degradation pattern under similar operating conditions last winter, thus predicting the possible degradation rate in the coming winter. The server combines the primary feature set generated by each evaluation anchor point and the secondary feature set generated by the time series correlation model to perform multi-directional prediction processing. For short-term degradation prediction indications, the server combines the primary feature set of the short-term evaluation anchor point and the recent secondary feature set in the time series correlation model to predict that in the next week, the material within 1 cm around the crack may deteriorate slightly and the strength may decrease by 5%.For the medium-term degradation prediction indication, the primary feature set of the medium-term assessment anchor point and the secondary feature set containing medium-term historical data are combined to predict that within the next month, obvious degradation may occur within 3 cm around the crack, which requires close attention. For the long-term degradation prediction indication, the primary feature set of the long-term assessment anchor point and the secondary feature set covering the long-term trend are combined. It is estimated that within the next six months, the crack will expand to a certain extent, affecting the performance of the entire blade, and it is recommended to arrange a maintenance plan. In another scenario, the abnormal part is the corrosion area at the root of the blade. The server obtains the abnormal part attributes such as the size, depth, and composition of the corrosion material of the corrosion area, as well as the operating environment attributes such as the current air quality, rainfall, and vibration frequency of the fan. The material corrosion network in the professional knowledge network extracts the initial feature data based on this information, such as predicting the development speed of corrosion based on the composition of the corrosion material and the ambient humidity. The mechanical performance evaluation network extracts the initial feature data on the impact on the stability of the blade structure based on the location of the corrosion area and the vibration of the fan. The evaluation anchor point corresponding to the short-term indication in the peripheral degradation prediction indication takes into account the upcoming rainy season and the current wind turbine operating load, converts the initial feature data to obtain the short-term primary feature set, and predicts that in the next three days, the surface rust may increase within 0.5 cm around the corrosion area. The evaluation anchor point of the medium-term indication combines seasonal changes and equipment maintenance cycles to generate a medium-term primary feature set, and predicts that the corrosion area will expand by 1 square centimeter in the next two months. The evaluation anchor point of the long-term indication is based on the overall life planning of the wind turbine and industry experience, and obtains the long-term primary feature set. It is estimated that in the next year, if no measures are taken, corrosion will cause the blade root strength to decrease by 10%. The time series association model extracts the secondary feature set by analyzing the development trend of similar corrosion situations in different time periods in the past, and further supports the prediction results of each peripheral degradation prediction indication. The server finally integrates the primary feature set and the secondary feature set to provide specific and accurate estimated degradation results for the short-term, medium-term and long-term peripheral degradation prediction indications, providing a strong basis for the maintenance decision of the wind farm. In the actual operation of wind farms, such analysis and prediction are constantly carried out. For example, wear on the tip of the blade was detected again. The server obtains the abnormal part attributes such as the degree, location, and shape of wear, as well as the operating environment attributes such as the wind speed distribution, atmospheric particle concentration, and fan speed at the time. The friction and wear network in the professional knowledge network extracts the initial feature data based on these, such as evaluating the aggravating factors of wear based on the wear shape and particle concentration. The structural strength analysis network analyzes the initial features that affect the overall strength of the blade based on the wear location and fan speed. The evaluation anchor point of the short-term degradation prediction indication takes into account the upcoming change in wind direction and the fan operation plan in the short term, converts the initial feature data into a short-term primary feature set, and predicts that the surface roughness within 0.3 cm around the wear will increase in the next two days.The mid-term assessment anchor combines the weather forecast and equipment maintenance schedule for the next few weeks to generate a mid-term primary feature set, predicting that the wear area will deepen by 0.5 mm in the next month and a half. The long-term assessment anchor obtains the long-term primary feature set based on the long-term service goals of the wind turbine and the general wear development laws of the industry. It is estimated that in the next eight months, wear will affect the aerodynamic performance of the blades, resulting in a 2% reduction in power generation efficiency. The time series association model refers to historical data of similar wear conditions in the past, extracts secondary feature sets, and assists in improving the estimated results for each time period. Through comprehensive processing, the server gives accurate estimated degradation results for each peripheral degradation estimation indication, helping wind farms to scientifically arrange maintenance work and ensure the safe, stable operation and efficient power generation of wind turbines.

[0063] In the embodiment of the present invention, the initial feature extraction process is performed on the abnormal part attributes and the operating environment attributes according to multiple professional knowledge networks to obtain the initial feature data generated by each professional knowledge network, which can be implemented through the following examples.

[0064] According to the encoding unit, encoding operations are performed on the abnormal part attribute and the operating environment attribute respectively to obtain abnormal part parameters corresponding to the abnormal part attribute and operating environment parameters corresponding to the operating environment attribute;

[0065] Performing an integration operation on the abnormal part parameter and the operating environment parameter according to the integration unit to obtain an integrated characteristic parameter;

[0066] The integrated feature parameters are input into a plurality of the professional knowledge networks to perform initial feature extraction processing, and initial feature data generated by each of the professional knowledge networks are obtained.

[0067] In an embodiment of the present invention, exemplarily, in a monitoring system of a large wind farm, the server receives detailed information about abnormal parts of wind turbine blades and environmental data of wind turbine operation. First, the server obtains the attributes of the abnormal part, such as the abnormal part is a deep crack on the surface of the blade, and its attributes include the length of the crack is 15 cm, the width is 2 cm, the depth is 1 cm, the position is in the middle of the blade, and the direction of the crack is longitudinal. At the same time, the server also obtains the operating environment attributes, such as the current wind speed is 10 meters per second, the wind direction is northeast wind, the temperature is 25 degrees Celsius, the humidity is 60%, and the wind turbine has been running for 5000 hours. Next, the server uses the encoding unit to perform encoding operations on these abnormal part attributes and operating environment attributes. During the encoding process, the server converts the values ​​of the length, width, depth, etc. of the crack into a specific encoding form to obtain the abnormal part parameters. For example, the crack length of 15 cm is encoded as "001011", the width of 2 cm is encoded as "00010", and the depth of 1 cm is encoded as "00001". Similarly, for the operating environment attributes, the wind speed of 10 meters per second is encoded as "1010", the wind direction of northeast wind is encoded as "011", the temperature of 25 degrees Celsius is encoded as "01011", the humidity of 60% is encoded as "110", and the wind turbine operating time of 5000 hours is encoded as "100101010". Then, the server uses the integration unit to perform integration operations on the abnormal part parameters and the operating environment parameters. During integration, the server combines and fuses these encoded parameters according to certain rules and algorithms. For example, the codes of the length, width, and depth of the crack can be connected in sequence, and then spliced ​​together with the codes of wind speed, wind direction, temperature, humidity, and wind turbine operating time to form an integrated feature parameter, such as "0010110001000001010100110101011110100101010". After that, the server inputs this integrated feature parameter into multiple professional knowledge networks to perform initial feature extraction processing. For example, there are three professional knowledge networks, namely, a network based on material mechanics, a network based on meteorology, and a network based on wind turbine engineering. After receiving the integrated characteristic parameters, the network based on material mechanics extracts the initial characteristic data related to the blade material performance and crack expansion through its internal calculation and analysis mechanism. For example, based on the size of the crack and the characteristics of the material, the stress distribution of the material in the current environment is calculated, the speed and direction of further crack expansion are predicted, and the initial characteristic data "material stress is concentrated, the crack expands quickly, and it is expected to expand by 2 cm within a week" is generated. The network based on meteorology analyzes the impact of current meteorological conditions on the blades based on the integrated characteristic parameters. The initial characteristic data it may extract is "high humidity may accelerate the corrosion of materials around the crack, and the corrosion depth is expected to increase by 0.5 cm within a month."The network based on wind turbine engineering extracts the initial feature data related to the blade performance and the overall operation stability of the wind turbine from the perspective of wind turbine operation, such as "the long operation time of the wind turbine, combined with the crack situation, may affect the power generation efficiency, and the efficiency is expected to decrease by 3% within two months". Through such a detailed processing process, the server can extract valuable initial feature data from the complex abnormal part attributes and operating environment attributes with the help of the professional knowledge network, providing a basis for subsequent more accurate peripheral degradation estimation. In another processing, the abnormal part is the wear area at the root of the blade. The abnormal part attributes include wear depth of 0.5 cm, wear area of ​​3 square centimeters, location at the root close to the connection part, etc. The operating environment attributes include the current average wind speed of 8 meters per second, variable wind direction, temperature of 30 degrees Celsius, humidity of 50%, and the wind turbine has been running for 8,000 hours. The server also performs encoding operations first, converting the values ​​of wear depth, area, location, etc. and environmental attributes into codes. Then perform integration operations to form integrated feature parameters. After inputting the professional knowledge network, the initial feature data extracted by the network based on material mechanics can be "the wear area is close to the connection part, the structural strength is affected, and the strength is expected to drop by 10% within half a month". The initial feature data of the network based on meteorology can be "the high temperature accelerates the oxidation of the wear area, and the thickness of the oxide layer is expected to increase by 0.2 cm within 20 days". The initial feature data given by the network based on wind turbine engineering can be "the wind turbine has been running for a long time, and the root wear affects the overall balance. It is expected that maintenance and adjustment will be required within three months". The server continuously processes various abnormal conditions and operating environment data, and provides accurate and reliable initial feature data for the maintenance and management of wind farms through precise coding, integration and analysis of the professional knowledge network, ensuring the safe operation and power generation efficiency of wind turbine blades.

[0068] In an embodiment of the present invention, the context-adaptive conversion of the initial feature data generated by each of the professional knowledge networks according to the evaluation anchor points corresponding to the multiple peripheral degradation prediction indications is performed to obtain the primary feature set generated by each of the evaluation anchor points, which can be implemented through the following examples.

[0069] Determining a first weighting coefficient corresponding to each of the professional knowledge networks according to the first decision component corresponding to each of the surrounding degradation estimation indications, and performing weighted summation on the initial feature data generated by each of the professional knowledge networks according to the first weighting coefficient corresponding to each of the professional knowledge networks to obtain first summary feature data corresponding to each of the surrounding degradation estimation indications;

[0070] A context-adaptive transformation is performed on each of the first summary feature data according to the plurality of evaluation anchor points to obtain a primary feature set generated by each of the evaluation anchor points.

[0071] In an embodiment of the present invention, exemplarily, in a complex wind farm monitoring system, the server is busy processing the degradation prediction work of the wind turbine blades. For example, there are three peripheral degradation prediction indications, namely short-term (within 1 month), medium-term (within 3 months) and long-term (within 6 months) degradation prediction indications. The server first determines the first weighting coefficient corresponding to each professional knowledge network according to the first decision component corresponding to each peripheral degradation prediction indication. For the short-term degradation prediction indication, the first decision component may consider that the professional knowledge network based on material mechanics has a greater weight on the degradation in the short term, and sets its first weighting coefficient to 0.4; the influence of the professional knowledge network based on meteorology is second, with a weighting coefficient of 0.3; the influence of the professional knowledge network based on wind turbine engineering is relatively small, with a weighting coefficient of 0.3. For the medium-term degradation prediction indication, the judgment of the first decision component may be different, and it is considered that the weight of the professional knowledge network based on meteorology is increased to 0.4, the weight of the professional knowledge network based on material mechanics is adjusted to 0.35, and the weight of the professional knowledge network based on wind turbine engineering is 0.25. For long-term degradation prediction indications, the first decision component can pay more attention to factors related to wind turbine engineering, and set the weighting coefficient of the professional knowledge network based on wind turbine engineering to 0.4, and the weighting coefficients of the professional knowledge networks based on material mechanics and meteorology to 0.3 respectively. After determining the weighting coefficients, the server begins to weight and sum the initial feature data generated by each professional knowledge network. For example, the initial feature data generated by the professional knowledge network based on material mechanics is "the crack expansion speed is fast in the short term, and it is expected to expand by 1 cm in 10 days", the initial feature data generated by the professional knowledge network based on meteorology is "the current humidity is high, which may accelerate corrosion, and the corrosion depth is expected to be 0.3 cm in 20 days", and the initial feature data generated by the professional knowledge network based on wind turbine engineering is "the wind turbine is in general operation and has little impact on the blades". For short-term degradation prediction indications, the server calculates according to the weighting coefficients: (0.4*quantified value of the fast crack expansion speed in the short term) + (0.3*quantified value of the current high humidity accelerating corrosion) + (0.3*quantified value of the general wind turbine operation status), and obtains the first summary feature data corresponding to the short term. For medium-term and long-term degradation prediction indications, the server also performs similar calculations according to their respective weighting coefficients to obtain the first summary feature data corresponding to the medium-term and long-term, respectively. Next, the server performs context-adaptive conversion on each first summary feature data based on multiple evaluation anchor points. For example, the evaluation anchor points are set according to different environmental conditions and wind turbine operation modes. For short-term evaluation anchor points, it will focus on the upcoming severe weather conditions. If heavy rain and strong winds are predicted in the short term, it will adjust the short-term first summary feature data to further increase the possible degree of degradation and generate the primary feature set of short-term evaluation anchor points. For medium-term evaluation anchor points, attention will be paid to the impact of seasonal changes on blades.For example, if the rainy season is about to come, it will adjust the first summary feature data in the medium term accordingly to generate the primary feature set of the medium-term evaluation anchor. For the long-term evaluation anchor, the overall life cycle and maintenance plan of the wind turbine will be considered. According to these factors, the long-term first summary feature data is converted to obtain the primary feature set of the long-term evaluation anchor. In another process, for example, the peripheral degradation prediction indication is for the wear of the blade surface. For the short-term degradation prediction indication, the weighting coefficient determined by the first decision component can be 0.5 based on the material surface treatment expertise network, 0.3 based on the operational data analysis expertise network, and 0.2 based on the environmental monitoring expertise network. The server weights and sums the corresponding starting feature data according to these weighting coefficients to obtain the short-term first summary feature data. Then, the short-term evaluation anchor performs a context-adaptive conversion on the first summary feature data according to the recent increase in wind turbine load, and generates a short-term primary feature set, indicating that wear may accelerate in the short term. For the medium-term and long-term degradation prediction indications, similar weighted summation and context-adaptive conversion processes are also carried out to obtain their respective primary feature sets, which provide strong support for accurately evaluating the degradation of blades in different time periods. Through such sophisticated and comprehensive processing, the server continuously provides accurate and reliable blade degradation prediction information to the operation and maintenance personnel of the wind farm, ensuring the safe and stable operation of the wind turbine and efficient power generation.

[0072] In the embodiment of the present invention, the time series correlation feature extraction process is performed on the initial feature data generated by each professional knowledge network according to the time series correlation model to obtain the secondary feature set generated by the time series correlation model, which can be implemented through the following examples.

[0073] Determining a second weighting coefficient corresponding to each of the professional knowledge networks according to the second decision component corresponding to each of the peripheral degradation estimation indications, and performing weighted summation on the initial feature data generated by each of the professional knowledge networks according to the second weighting coefficient corresponding to each of the professional knowledge networks to obtain second summary feature data corresponding to each of the peripheral degradation estimation indications;

[0074] A time series correlation feature extraction process is performed on each of the second summary feature data according to the time series correlation model to obtain a secondary feature set generated by the time series correlation model.

[0075] In an embodiment of the present invention, exemplarily, in a monitoring system of a large wind farm, the server is processing a complex task of predicting the degradation of wind turbine blades. For example, there are three peripheral degradation prediction indicators, namely, short-term (1 month), medium-term (3 months) and long-term (6 months) degradation predictions. First, the server determines the second weighting coefficient corresponding to each professional knowledge network according to the second decision component corresponding to each peripheral degradation prediction indicator. For the second decision component of the short-term degradation prediction indicator, it can be considered that in the short term, the professional knowledge network based on material mechanics is more important for the prediction of degradation, so the second weighting coefficient is assigned to it as 0.4; the influence of the professional knowledge network based on meteorology is relatively small, and the weighting coefficient is set to 0.3; the weighting coefficient of the professional knowledge network based on wind turbine engineering is 0.3. For the second decision component of the medium-term degradation prediction indicator, more attention will be paid to the influence of meteorology, so the second weighting coefficient of the professional knowledge network based on meteorology is set to 0.4, the one based on material mechanics is set to 0.3, and the one based on wind turbine engineering is set to 0.3. For the second decision component of the long-term degradation prediction indication, more attention can be paid to the long-term impact of wind turbine engineering. The second weighting coefficient of the professional knowledge network based on wind turbine engineering is set to 0.4, the one based on material mechanics is 0.3, and the one based on meteorology is 0.3. After determining the weighting coefficient, the server begins to weight the initial feature data generated by each professional knowledge network to obtain the second summary feature data corresponding to each peripheral degradation prediction indication. For example, the initial feature data generated by the professional knowledge network based on material mechanics is "slight increase in material fatigue in the short term", the one based on meteorology is "recent climate stability, small impact on blades", and the one based on wind turbine engineering is "the short-term operation load of the wind turbine is normal". For the short-term degradation prediction indication, the server calculates according to the above second weighting coefficient: (0.4*quantitative value of slight increase in material fatigue in the short term) + (0.3*quantitative value of small impact of recent climate stability on blades) + (0.3*quantitative value of normal short-term operation load of the wind turbine), and obtains the second summary feature data corresponding to the short term. For the medium-term and long-term degradation estimation indications, the server also performs similar calculations according to the respective second weighting coefficients to obtain the second summary feature data corresponding to the medium-term and long-term, respectively. Next, the server performs time series correlation feature extraction processing on each second summary feature data according to the time series correlation model. For example, the time series correlation model stores the blade degradation data of the same wind turbine in different seasons and under different operating conditions in the past year. For the short-term second summary feature data, the server inputs it into the time series correlation model. The model compares the current short-term features with the features of the same time period in the past (such as last month) to find similarities and differences. If the blade degradation rate was faster under similar conditions last month, and the current short-term second summary feature data shows that the situation is better, then the model will extract the secondary feature that "the short-term degradation rate is expected to slow down."For the medium-term second summary feature data, the model will consider the trend of the past few months. If the blade degradation has gradually increased under similar meteorological conditions in the past few months, and the current medium-term second summary feature data has not improved significantly, the model may extract the secondary feature of "increased medium-term degradation risk". For the long-term second summary feature data, the model will integrate the long-term data of the past six months or even a year. If the blade performance has gradually declined under similar engineering conditions in the past long-term operation, and the current second summary feature data has not improved significantly, the model may extract the secondary feature of "long-term focus on blade performance degradation". Through such processing, the server can generate a valuable secondary feature set for each peripheral degradation prediction indication from the perspective of time series association, providing support for more accurate and comprehensive blade degradation prediction. In another scenario, for example, the starting feature data generated by the professional knowledge network is different. The one based on material mechanics is "there are tiny cracks inside the material with a tendency to expand", the one based on meteorology is "the rainy season is about to begin, and the humidity is increasing", and the one based on fan engineering is "the fan is running at high load for a long time". For the short-term degradation prediction indication, the server calculates according to the newly determined second weighting coefficient to obtain the short-term second summary feature data. When processing the short-term second summary feature data, the time series correlation model found that the climate conditions during the same period this year were more conducive to crack propagation than the same period last year, thereby extracting the secondary feature of "higher risk of short-term crack propagation". For the medium-term and long-term degradation prediction indicators, the server also performs corresponding calculations and processing. Based on the historical data and the current second summary feature data, the model extracts the secondary features of "medium-term humidity impact may increase degradation" and "long-term high-load operation aggravates performance degradation". By continuously and accurately executing these steps, the server provides the operation and maintenance personnel of the wind farm with more forward-looking and reliable blade degradation prediction information, which helps to take effective maintenance measures in a timely manner to ensure the normal operation and power generation efficiency of the wind turbine.

[0076] In an embodiment of the present invention, multiple peripheral degradation prediction indications include a first peripheral degradation prediction indication and a second peripheral degradation prediction indication, and the first peripheral degradation prediction indication is located before the second peripheral degradation prediction indication in a timing set having a timing correlation relationship; the timing correlation feature extraction processing is performed on each of the second summary feature data according to the timing correlation model to obtain a secondary feature set generated by the timing correlation model, which can be implemented through the following example.

[0077] performing a time series correlation feature extraction process on the second summary feature data corresponding to the first surrounding degradation prediction indication according to the time series correlation model to obtain a secondary feature set corresponding to the first surrounding degradation prediction indication, and performing a time series correlation feature extraction process on the second summary feature data corresponding to the first surrounding degradation prediction indication and the second summary feature data corresponding to the second surrounding degradation prediction indication according to the time series correlation model to obtain a secondary feature set corresponding to the second surrounding degradation prediction indication;

[0078] The secondary feature set corresponding to the first surrounding degradation prediction indication and the secondary feature set corresponding to the second surrounding degradation prediction indication are determined as the secondary feature sets generated by the time series association model.

[0079] In an embodiment of the present invention, exemplarily, in an operation and maintenance system of a large wind farm, the server is working to accurately estimate the degradation of wind turbine blades. For example, there are two peripheral degradation estimation indications, namely the first peripheral degradation estimation indication (short-term, within 1 month) and the second peripheral degradation estimation indication (medium-term, within 3 months). Moreover, the first peripheral degradation estimation indication is located before the second peripheral degradation estimation indication in the time series set. First, the server obtains the second summary feature data corresponding to the first peripheral degradation estimation indication. For example, these data reflect that some tiny cracks have appeared on the blade surface in the current short term, as well as changes in environmental factors such as wind speed and temperature. The server inputs these data into the time series association model. The model will search for similar short-term situations in past historical data, and compare and analyze them. For example, it was found that under similar environmental conditions and the initial state of the blade last month, the cracks expanded to a certain extent within a week. Based on such time series association feature extraction processing, the server obtains the secondary feature set corresponding to the first peripheral degradation estimation indication, which may include "the crack expansion speed is fast in the short term, and close attention is required". Next, the server obtains the second summary feature data corresponding to the second peripheral degradation prediction indication. In addition to including the blade condition in the current medium term, these data also integrate the information during the first peripheral degradation prediction indication. Then, the server inputs these data into the time series association model again. The model not only considers the current medium-term situation, but also combines the results of the previous first peripheral degradation prediction indication. For example, historical data shows that when similar crack expansion occurs in the short term, if the wind speed continues to be high and the temperature fluctuates greatly in the medium term, the degradation of the blade will be significantly aggravated. Through such comprehensive analysis and time series association feature extraction processing, the server obtains the secondary feature set corresponding to the second peripheral degradation prediction indication, which can be "the risk of blade degradation in the medium term has increased significantly, and it is recommended to maintain in advance." Finally, the server combines the secondary feature set corresponding to the first peripheral degradation prediction indication and the secondary feature set corresponding to the second peripheral degradation prediction indication to determine the secondary feature set generated by the time series association model. For example, in another processing, the second summary feature data corresponding to the first peripheral degradation prediction indication shows that there is slight wear on the blade surface and a high humidity environment in the short term. The server uses the time series association model to extract the secondary feature set corresponding to the first estimated peripheral degradation indication as "wear may be slightly aggravated by high humidity in the short term." For the second summary feature data corresponding to the second estimated peripheral degradation indication, it not only includes more signs of wear and possible climate change in the current medium term, but also combines short-term conditions. Through comprehensive analysis, the model extracts the secondary feature set corresponding to the second estimated peripheral degradation indication as "wear may accelerate and affect blade performance in the medium term due to short-term wear base and potential climate impacts."The server combines these two secondary feature sets to form a complete secondary feature set, providing more comprehensive and accurate blade degradation prediction information for operation and maintenance personnel, helping them to formulate reasonable maintenance strategies and plans. In actual operation and maintenance scenarios, this kind of time series correlation analysis is constantly carried out. For example, the first peripheral degradation prediction indication found that the blade was hit by a strong wind in the short term, resulting in local deformation. Based on historical data, the time series correlation model extracts the secondary feature set corresponding to the first peripheral degradation prediction indication as "short-term strong wind impact may cause local performance degradation". The second summary feature data corresponding to the second peripheral degradation prediction indication combines the short-term impact situation and the possible continuous load and environmental changes in the medium term. Based on this, the model extracts the secondary feature set corresponding to the second peripheral degradation prediction indication as "due to short-term impact and subsequent load in the medium term, local deformation may worsen and affect the overall structural stability". The server determines these two secondary feature sets as the final secondary feature sets, providing strong support for the maintenance decision of the wind farm, ensuring the safe and reliable operation of wind turbine blades in different time periods, improving power generation efficiency and reducing failure risks.

[0080] In an embodiment of the present invention, the temporal association model includes at least one temporal association secondary model, each of which includes a parallel attention mechanism unit and a forward conduction neural network unit; the temporal association feature extraction process is performed on each of the second summary feature data according to the temporal association model to obtain a secondary feature set generated by the temporal association model, which can be implemented by the following example. Each of the second summary feature data is input into the parallel attention mechanism unit for weight allocation calculation to obtain weight allocation data corresponding to each of the second summary feature data;

[0081] The weight allocation data corresponding to each of the second summary feature data is input into the forward conduction neural network unit for feature refinement to obtain a secondary feature set generated by the time series association model.

[0082] In an embodiment of the present invention, illustratively, in a busy wind farm monitoring center, the server is concentrating on the critical task of predicting the degradation of wind turbine blades. For example, the server receives multiple second summary feature data, which come from different professional knowledge networks and reflect the conditions of wind turbine blades in different aspects. First, the server inputs these second summary feature data into the parallel attention mechanism unit in the time series association model. The parallel attention mechanism unit begins to perform weight distribution calculations on each input second summary feature data. For example, there are three second summary feature data: feature data A mainly describes the degree and trend of wear on the blade surface; feature data B focuses on the impact of recent climate change on the blade; feature data C focuses on the correlation between the vibration of the wind turbine during operation and the blade structure. The parallel attention mechanism unit analyzes the importance and relevance of these data. If the current wind farm is located in an area that is about to enter a windy season, and historical data show that the wear of the blade surface is more critical to degradation in this case, then it will assign a higher weight to feature data A, such as 0.4; because climate change is relatively stable, the weight assigned to feature data B can be 0.3; and the vibration of the wind turbine is relatively minor in the current analysis, and the weight assigned to feature data C can be 0.3. After calculation, the server obtains the weight distribution data corresponding to each second summary feature data. Next, the server inputs these weight distribution data into the forward conduction neural network unit for feature refinement. After receiving the weight distribution data, the forward conduction neural network unit begins to conduct in-depth analysis and processing of these data. It will use its complex neural network structure and learning ability to extract and optimize the features of the input data. For example, it will find that the change trend of the degree of wear in feature data A shows a certain periodicity, which is similar to the situation in similar seasons in history; although the climate change in feature data B is relatively stable, some subtle fluctuations may have a potential association with blade degradation; although the vibration in feature data C has a lower weight, some specific frequency vibrations may indicate potential problems in the internal structure of the blade. By refining and mining these features, the forward conduction neural network unit eventually generates a secondary feature set of the time series association model. For example, this secondary feature set includes key information such as the prediction of the degradation rate of the blades in the future, the indication of key areas where degradation may occur, and the comparative analysis with similar situations in history. In another processing, the second summary feature data received by the server was different. Feature data D describes the fatigue condition of the blade material; feature data E reflects the potential impact of recent electric field maintenance operations on the blades; feature data F involves the possibility of corrosion on the blades due to the concentration of pollutants in the surrounding environment.The parallel attention mechanism unit assigns a weight of 0.4 to feature data D based on the current actual situation, for example, material fatigue is a gradual accumulation process and is relatively critical; since the effect of recent maintenance operations needs further observation, a weight of 0.3 is assigned to feature data E; considering that the influence of pollutant concentration is relatively indirect, a weight of 0.3 is assigned to feature data F. The forward conduction neural network unit processes these weighted distribution data and may find that the material fatigue condition is highly similar to a certain period in the past, predicting that future degradation may accelerate; although the maintenance operation has no obvious impact at present, there are potential risks; the slow increase in pollutant concentration may cause severe corrosion in the long run. The resulting secondary feature set provides more targeted and forward-looking decision support for operation and maintenance personnel, helping them to take effective prevention and maintenance measures in a timely manner to ensure the safe and stable operation and efficient power generation of wind turbine blades.

[0083] In the embodiment of the present invention, the multi-directional estimation processing is performed according to the primary feature set and the secondary feature set to obtain the estimated degradation result of the surrounding part corresponding to each of the peripheral degradation estimation indications, which can be implemented through the following examples.

[0084] Determining a weighting coefficient corresponding to the primary feature set and a weighting coefficient corresponding to the secondary feature set according to the third decision component corresponding to each of the surrounding degradation prediction indications, and summing the primary feature set and the secondary feature set according to the weighting coefficient corresponding to the primary feature set and the weighting coefficient corresponding to the secondary feature set to obtain third summary feature data corresponding to each of the surrounding degradation prediction indications;

[0085] The feature mapping unit performs mapping processing on the third summary feature data corresponding to each of the peripheral degradation prediction indications to obtain the estimated degradation result of the peripheral part corresponding to each of the peripheral degradation prediction indications.

[0086] In an embodiment of the present invention, exemplarily, in a large-scale wind farm, the server is busy processing the peripheral degradation estimation work of the wind turbine blades. For example, there are three peripheral degradation estimation indications, namely short-term (1 month), medium-term (3 months) and long-term (6 months). The server has obtained the primary feature set and secondary feature set corresponding to each peripheral degradation estimation indication. First, the server determines the weighting coefficients corresponding to the primary feature set and the secondary feature set according to the third decision component corresponding to each peripheral degradation estimation indication. For the third decision component of the short-term peripheral degradation estimation indication, it can be considered that in the short term, the primary feature set has a more direct and significant impact on the degradation estimation, so the weighting coefficient of the primary feature set is determined to be 0.6, and the weighting coefficient of the secondary feature set is determined to be 0.4. For the third decision component of the medium-term peripheral degradation estimation indication, it can be considered that the importance of the primary feature set and the secondary feature set is relatively balanced, and the weighting coefficients of the two are determined to be 0.5 respectively. For the third decision component of the long-term peripheral degradation prediction indication, since the long-term prediction is more dependent on historical trends and comprehensive factors, the weight coefficient of the secondary feature set may be increased to 0.6, and the weight coefficient of the primary feature set may be reduced to 0.4. After determining the weight coefficient, the server begins to weight the primary feature set and the secondary feature set. For example, for the short-term peripheral degradation prediction indication, for example, the data contained in the primary feature set indicates that the blade surface has slight wear in the short term, and the quantitative value is 80; the data of the secondary feature set shows that the recent wind speed change will slightly accelerate the wear, and the quantitative value is 60. The server calculates according to the weight coefficient: (0.6*80)+(0.4*60)=72, and obtains the third summary feature data corresponding to the short-term peripheral degradation prediction indication. For the medium-term and long-term peripheral degradation prediction indications, the server also performs similar calculations according to the weight coefficients determined by each, and obtains the third summary feature data corresponding to the medium-term and long-term peripheral degradation prediction indications respectively. Next, the server performs mapping processing on the third summary feature data corresponding to each peripheral degradation prediction indication according to the feature mapping unit. The feature mapping unit stores a series of mapping rules and functions for converting the third summary feature data into specific estimated degradation results. For example, for the third summary feature data 72 corresponding to the short-term peripheral degradation prediction indication, the feature mapping unit may map it to "the degree of blade surface wear is expected to increase by 5% in the short term, and close observation is required" according to the preset rules. For the third summary feature data corresponding to the medium-term peripheral degradation prediction indication, for example, 85, the feature mapping unit may map it to "local micro cracks may appear on the blades in the medium term, and a detailed inspection is recommended after 1 month." For the third summary feature data corresponding to the long-term peripheral degradation prediction indication, such as 90, the feature mapping unit may map it to "in the long run, the structural stability of the blades may be affected, and a maintenance plan needs to be formulated within 3 months." In another scenario, for example, the data of the primary feature set and the secondary feature set are different.For the short-term peripheral degradation prediction indication, the primary feature set shows that the blade surface is slightly hit, with a quantitative value of 70; the secondary feature set indicates that the climate is dry in the short term, which may affect the material performance of the blade, with a quantitative value of 50. The server calculates the weighted sum: (0.6*70)+(0.4*50)=62, and obtains the third summary feature data. The feature mapping unit maps it as "the material performance of the impacted part of the blade may be slightly reduced in the short term, pay attention to monitoring". For the medium-term peripheral degradation prediction indication, the primary feature set points out that there is potential stress concentration inside the blade, with a quantitative value of 75; the secondary feature set indicates that there may be strong wind weather in the medium term, with a quantitative value of 65. The weighted sum: (0.5*75)+(0.5*65)=70, after being processed by the feature mapping unit, the third summary feature data may be obtained as "the stress concentration area of ​​the blade may develop into cracks in the medium term, and maintenance materials need to be prepared in advance". For the long-term peripheral degradation prediction indication, the primary feature set indicates that the overall aging signs of the blade are obvious, with a quantitative value of 80; the secondary feature set shows that the long-term operating environment is harsh, with a quantitative value of 70. Calculate the weight: (0.4*80)+(0.6*70)=74, and obtain "After long-term operation, the blades need to be fully overhauled or replaced, and the feasibility should be evaluated within 6 months" through the feature mapping unit. Through such fine and comprehensive processing, the server can provide accurate, detailed and targeted estimated degradation results for the surrounding degradation estimation indications in different time periods, helping the operation and maintenance personnel of the wind farm to make scientific and reasonable decisions, ensuring the safe and stable operation of the wind turbine blades and efficient power generation.

[0087] In the embodiments of the present invention, the following implementation modes are also provided.

[0088] Acquire multiple initial professional knowledge networks, multiple initial evaluation anchor points corresponding to the peripheral degradation prediction indications, an initial time series association model, training instances, and instance category labels corresponding to the training instances;

[0089] Combining a plurality of the initial professional knowledge networks, a plurality of initial evaluation anchor points corresponding to the peripheral degradation prediction indications, and the initial time series association model into an original multi-directional prediction model;

[0090] According to the training instances and the instance category labels corresponding to the training instances, the original multi-directional prediction model is learned and optimized to obtain a target multi-directional prediction model; the target multi-directional prediction model includes a plurality of the professional knowledge networks, a plurality of evaluation anchor points corresponding to the peripheral degradation prediction indications, and the timing association model.

[0091] In an embodiment of the present invention, exemplarily, first, the server obtains multiple initial professional knowledge networks. These initial professional knowledge networks may include a network based on material science for analyzing the performance and degradation trend of blade materials; a network based on mechanical principles for evaluating the stress conditions and possible structural damage of blades during operation; a network based on meteorology, considering the impact of environmental factors on blades, etc. At the same time, the server obtains multiple initial evaluation anchor points corresponding to peripheral degradation prediction indicators. For example, for the initial evaluation anchor points of short-term (within 1 month) peripheral degradation prediction indicators, the focus can be on recent weather changes and minor damage to the blade surface; for the initial evaluation anchor points of the medium-term (3-6 months), more attention can be paid to the impact of seasonal changes on blades and the development of potential internal defects; for the initial evaluation anchor points of the long term (more than 6 months), the overall aging of blades and the cumulative effects of large-scale environmental changes may be considered. The server also obtains an initial time series association model, which can contain the operating data, degradation conditions, and related environmental and operating parameters of wind turbine blades in different time periods in the past. In addition, the server obtains a wealth of training instances and instance category annotations corresponding to these training instances. The training instance can be a large number of detailed status data of wind turbine blades under different operating times and different environmental conditions, and the instance category annotation clearly indicates the actual degradation degree and type of the blade in each training instance, such as mild wear, moderate corrosion, severe cracks, etc. Then, the server combines multiple initial professional knowledge networks, initial evaluation anchor points corresponding to multiple peripheral degradation prediction indicators, and initial time series association models into an original multi-directional prediction model. Next, the server learns and optimizes the original multi-directional prediction model based on the training instance and the corresponding instance category annotation. The server inputs a training instance into the original multi-directional prediction model. For example, this training instance is the data of a wind turbine blade that has been running for 3 months, including a detailed inspection report of the blade (such as some minor scratches on the surface and slight stress concentration inside), environmental data such as the average wind speed, temperature, and humidity for these 3 months, and information such as the position and operating load of the blade in the entire wind farm. The instance category annotation shows that the blade is in a "mildly degraded" state. The initial professional knowledge networks in the original multi-directional prediction model begin to analyze and extract features from the input blade data. The material science network may point out that minor scratches may cause material fatigue; the mechanics network may calculate the potential risk of stress concentration areas; and the meteorology network may evaluate the promotion or inhibition of blade degradation by environmental factors. The initial assessment anchor points corresponding to each peripheral degradation estimate indication make preliminary degradation estimates based on the extracted features. The short-term assessment anchor point can be considered that in the next month, the minor scratches on the blade surface will deepen slightly; the medium-term assessment anchor point can predict that in the next 3-6 months, the stress concentration area will develop into more obvious damage; the long-term assessment anchor point can be based on the current overall condition of the blade and speculate that after more than 6 months, attention should be paid to the overall performance degradation of the blade.The initial time series association model compares and correlates the current input training instance with historical data of similar situations in the past, trying to find out the potential degradation rules and trends. Then, the original multi-directional prediction model outputs the degradation prediction result of this training instance. The server compares the degradation prediction result output by the model with the instance category label (i.e., "mild degradation"). If there is a deviation between the model's prediction result and the actual label, the server will adjust the model's parameters according to this deviation, such as adjusting the weights in each professional knowledge network, the judgment criteria for evaluation anchor points, and the association rules in the time series association model, so that the model's prediction result is closer to the actual label. The server repeats the above process and repeatedly trains and optimizes the original multi-directional prediction model using a large number of training instances. For example, another training instance is a fan blade that has been in operation for 6 months. There are a few corrosion spots on the surface, the internal structure is basically normal, and the environmental data shows that it has experienced large temperature fluctuations and high humidity during this period. The instance category is labeled as "moderate degradation". The model analyzes, estimates, compares and adjusts parameters again to continuously improve and improve its performance. After repeated optimization of a large number of training instances, the server finally obtains the target multi-directional prediction model. This target multi-directional prediction model includes multiple optimized professional knowledge networks that can more accurately analyze the various characteristics of blades; multiple evaluation anchor points corresponding to peripheral degradation prediction indicators, whose judgment criteria are more reasonable and accurate; and an optimized time series correlation model that can more effectively mine the historical laws and trends of blade degradation. Through such a learning optimization process, the server can provide wind farms with more accurate and reliable wind turbine blade peripheral degradation prediction services, helping operation and maintenance personnel to take effective maintenance measures in a timely manner to ensure the stable operation and efficient power generation of wind farms.

[0092] In an embodiment of the present invention, the training instance includes sample abnormal part attributes corresponding to the sample object and sample operating environment attributes corresponding to the sample recommended resource; the original multi-directional prediction model is learned and optimized based on the training instance and the instance category label corresponding to the training instance to obtain the target multi-directional prediction model, which can be implemented through the following example.

[0093] Performing initial feature extraction processing on the abnormal part attributes of the sample and the operating environment attributes of the sample according to the multiple initial professional knowledge networks to obtain initial feature data generated by each of the initial professional knowledge networks;

[0094] Performing context-adaptive conversion on the initial feature data generated by each of the initial professional knowledge networks according to the multiple initial evaluation anchor points to obtain a sample primary feature set generated by each of the initial evaluation anchor points, and performing temporal association feature extraction processing on the initial feature data generated by each of the initial professional knowledge networks according to the initial temporal association model to obtain a sample secondary feature set generated by the initial temporal association model;

[0095] Performing prediction processing according to the sample primary feature set corresponding to each of the surrounding degradation prediction indications to obtain a first predicted degradation result corresponding to each of the surrounding degradation prediction indications, and performing prediction processing according to the sample secondary feature set corresponding to each of the surrounding degradation prediction indications to obtain a second predicted degradation result corresponding to each of the surrounding degradation prediction indications;

[0096] Performing multi-directional estimation processing according to the sample primary feature set and the sample secondary feature set to obtain a third estimated degradation result corresponding to each of the peripheral degradation estimation indications;

[0097] The original multi-directional prediction model is learned and optimized according to the first prediction degradation result, the second prediction degradation result, the third prediction degradation result and the instance category label to obtain the target multi-directional prediction model.

[0098] In an embodiment of the present invention, illustratively, the server first obtains a series of training instances, each of which includes the sample abnormal part attributes corresponding to the sample object and the sample operating environment attributes corresponding to the sample recommended resource. For example, a training instance is about a specific wind turbine blade. The sample abnormal part attributes indicate that there are some fine cracks on the surface of the blade, and the specific parameters such as the length, width and depth of the cracks are recorded in detail. The sample operating environment attributes include the average wind speed at the location of the wind turbine, the frequency of wind direction changes, the fluctuation range of ambient temperature and humidity, and the operating time and load of the wind turbine. The server calls multiple initial professional knowledge networks to perform initial feature extraction processing on these sample abnormal part attributes and sample operating environment attributes. For example, the initial professional knowledge network based on materials science will extract the initial features related to material fatigue and fracture from the characteristics of the crack, which may include the estimation of the residual strength of the material in the current crack state. The initial professional knowledge network based on mechanics will extract the initial features such as the force distribution of the blade and the possible crack extension trend based on information such as wind speed, load and crack location. The initial professional knowledge network based on meteorology will focus on the influence of ambient temperature and humidity on the performance of blade materials and extract the corresponding initial feature data. After the processing of each initial professional knowledge network is completed, the server obtains the initial feature data generated by each of them. Then, multiple initial evaluation anchors begin to perform situational adaptation conversion on these initial feature data. For the initial evaluation anchor of the short-term peripheral degradation estimation indication, it will focus on the possible impact of minor changes in environmental factors and abnormal parts in the near future. For example, if the wind speed increases significantly in the short term and the blade crack is on the windward side, it will convert the initial feature data into a sample primary feature set indicating a high risk of crack extension in the short term. For the initial evaluation anchor of the medium-term peripheral degradation estimation indication, seasonal changes, wind turbine maintenance plans and other factors will be comprehensively considered for conversion to generate a sample primary feature set reflecting the medium-term degradation trend. At the same time, the initial time series association model performs time series association feature extraction processing on the initial feature data generated by each initial professional knowledge network to obtain a sample secondary feature set. For example, the initial time series association model stores the degradation data and environmental change data of similar wind turbine blades in different time periods in the past. By comparing the current starting feature data with historical data, the model may find that the crack characteristics of the current blade are similar to the degradation that occurred in a past season, thereby extracting a sample secondary feature set, suggesting that similar degradation development may occur in the same time period. The server then performs prediction processing based on the sample primary feature set corresponding to each peripheral degradation estimate indication. For short-term peripheral degradation estimate indications, the server predicts that the crack will expand slightly in the next week or two based on the sample primary feature set generated by the short-term evaluation anchor point, but will not have a significant impact on the blade performance, and obtains the first estimated degradation result.For the mid-term peripheral degradation prediction indication, according to the sample primary feature set generated by the mid-term evaluation anchor point, it is predicted that the crack will become significantly longer and wider in the next one or two months, and close attention is needed to obtain the second estimated degradation result. Similarly, the server performs prediction processing based on the sample secondary feature set corresponding to each peripheral degradation prediction indication, and obtains the second estimated degradation result corresponding to each peripheral degradation prediction indication. Next, the server performs multi-directional prediction processing based on the sample primary feature set and the sample secondary feature set, and obtains the third estimated degradation result corresponding to each peripheral degradation prediction indication. In this process, the server will comprehensively consider the short-term and medium-term prediction results, as well as the long-term trends in historical data, and make a more comprehensive and comprehensive prediction of the degradation conditions in different time periods in the future. Finally, the server learns and optimizes the original multi-directional prediction model based on the first estimated degradation result, the second estimated degradation result, the third estimated degradation result and the instance category annotation. For example, the instance category annotation shows that in the actual operation of this wind turbine blade, the crack extension degree is large after one month, which has had a certain impact on the performance of the blade. However, the first estimated degradation result predicted by the model is too optimistic and underestimates the degree of degradation. Based on this deviation, the server adjusts the weight parameters in the initial professional knowledge network to make the evaluation of crack extension more accurate; at the same time, it optimizes the conversion rules of the initial evaluation anchor point so that it can more keenly capture the early signals of degradation; it also improves the association algorithm of the initial time series association model to improve the efficiency of using historical data. By processing and optimizing a large number of such training examples, the server gradually obtains a more accurate and reliable target multi-directional prediction model, which can provide more accurate services for the degradation prediction of wind turbine blades, help operation and maintenance personnel to formulate effective maintenance strategies in advance, and ensure the stable operation of wind farms. For another example, in another training example, the sample abnormal part attribute shows that the blade root has a certain degree of wear, and the sample operating environment attribute indicates that the wind turbine is often in high wind speed and severe weather conditions. The server processes according to the above process. The initial professional knowledge network based on materials science and mechanics extracts the initial features related to root wear and force. The initial evaluation anchor point generates short-term, medium-term and long-term sample primary feature sets based on these starting features. The initial time series association model extracts the sample secondary feature set. The server obtains the first estimated degradation result, the second estimated degradation result, and the third estimated degradation result corresponding to each peripheral degradation estimation indication. If the actual situation is that the wear of the root of the blade develops more seriously in a short period of time, and the estimated result of the model is not accurate enough, the server will adjust the parameters of the model accordingly, such as increasing the weight of the impact of environmental factors on wear, optimizing the judgment threshold of the severity of the evaluation anchor point, etc. As various training examples are continuously processed, the server continues to optimize the original multi-directional estimation model. For some training examples, it may be corrosion spots on the surface of the blade; for others, it may be abnormal vibration of the entire blade.The server can gradually improve the accuracy and reliability of the prediction through model learning and optimization according to the specific situation. In the actual operation and maintenance scenario, the server may receive new training instance data every day and optimize and update the model in a timely manner. For example, a new training instance shows that a wind turbine blade has new damage after experiencing strong wind weather. The server further improves the model by processing and analyzing this instance, so that it can better cope with the impact of sudden harsh environment on the blade. Through continuous learning and optimization, the target multi-directional prediction model in the server can more and more accurately predict the degradation trend of wind turbine blades in various complex situations, providing strong support for the efficient and safe operation of wind farms.

[0099] In an embodiment of the present invention, the original multi-directional prediction model is learned and optimized according to the first prediction degradation result, the second prediction degradation result, the third prediction degradation result and the instance category label to obtain the target multi-directional prediction model, which can be implemented through the following examples.

[0100] performing error calculation according to the first estimated degradation result and the instance category label to obtain a first error parameter, performing error calculation according to the second estimated degradation result and the instance category label to obtain a second error parameter, and performing error calculation according to the third estimated degradation result and the instance category label to obtain a third error parameter;

[0101] Determining a model error parameter according to the first error parameter, the second error parameter and the third error parameter;

[0102] The original multi-directional prediction model is learned and optimized according to the model error parameters until the model error parameters meet the error threshold, thereby obtaining a target multi-directional prediction model.

[0103] In an embodiment of the present invention, exemplarily, the server has obtained the first estimated degradation result, the second estimated degradation result and the third estimated degradation result for a training instance of a wind turbine blade, and also obtained the accurate instance category labeling of the training instance. First, the server performs error calculation based on the first estimated degradation result and the instance category labeling. For example, the first estimated degradation result shows that the degradation degree of the blade is slight in the short term, while the instance category labeling shows that the actual degradation degree is moderate. The server obtains the first error parameter through a specific error calculation method, such as mean square error calculation. This error parameter reflects the degree of deviation between the short-term estimated result and the actual situation. Then, the server performs a similar error calculation based on the second estimated degradation result and the instance category labeling to obtain the second error parameter. For example, the second estimated degradation result predicts that the blade will have a certain degree of structural damage in the medium term, while the actual situation is that the degree of damage is more serious, and the corresponding error value is obtained by calculation. Then, the server performs error calculation based on the third estimated degradation result and the instance category labeling to obtain the third error parameter. For example, after the third estimated degradation result comprehensively considers various factors, it is estimated that the long-term blade performance will slowly decline, but the actual situation is that the decline is faster, so the error of this part is calculated. After that, the server determines the model error parameter based on the first error parameter, the second error parameter and the third error parameter obtained. This can be done by weighted averaging or other comprehensive calculations of these three error parameters to fully reflect the overall error of the model in different time periods. For example, the server assigns a weight of 0.3 to the short-term error parameter, a weight of 0.4 to the medium-term error parameter, and a weight of 0.3 to the long-term error parameter. After calculation, a comprehensive model error parameter is obtained. Then, the server learns and optimizes the original multi-directional prediction model based on this model error parameter. The server adjusts various parameters in the model, such as the weight coefficient in the professional knowledge network, the judgment threshold of the evaluation anchor point, and the correlation parameter in the time series correlation model. During an optimization process, the server finds that the short-term estimation error is large, which may be due to insufficient consideration of certain environmental factors. Therefore, the server increases the weight of environmental factors in the short-term evaluation and recalculates and estimates. Next, the server performs the above-mentioned error calculation and model optimization steps again, and iterates continuously. Until the model error parameters meet the preset error threshold, it indicates that the model's prediction results are sufficiently accurate and reliable. At this point, the server obtains the optimized target multi-directional prediction model. For example, in another training instance, the first estimated degradation result underestimates the development rate of blade surface wear in the short term, the second estimated degradation result underestimates the impact of internal stress concentration on the blade in the medium term, and the third estimated degradation result fails to fully predict the fatigue accumulation of blade materials in long-term operation. The server calculates the corresponding error parameters respectively, and after determining the comprehensive model error parameters, it makes targeted adjustments to the original multi-directional prediction model.For example, the algorithm for material wear and stress calculation in the professional knowledge network is optimized, and the ability of the time series correlation model to capture long-term trends is improved. After multiple iterations of optimization, the model error parameter is gradually reduced, and finally the error threshold is met, and a target multi-directional prediction model that can accurately predict the degradation of wind turbine blades is obtained. In actual operation, the server continuously processes a large number of training instances and continuously optimizes the model to adapt to the characteristics of different wind turbine blades and the complex and changing operating environment. For example, for a wind turbine blade located in a seaside wind farm, due to the influence of the high salt fog environment, its degradation pattern may be different from that of the blade in an inland wind farm. The server processes the training instances from the blade and optimizes the model, so that the target multi-directional prediction model can accurately capture the degradation characteristics in this special environment. Through continuous learning and optimization, the server can provide more accurate blade degradation prediction services for wind farms, helping operation and maintenance personnel to prepare for maintenance in advance, reduce the risk of failures, and improve power generation efficiency.

[0104] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the above-mentioned wind turbine blade image stitching analysis method based on spatial trajectory. Figure 3 As shown, Figure 3 The block diagram of the computer device 100 provided in the embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112 and a communication unit 113. To achieve data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are directly or indirectly electrically connected to each other. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0105] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. A method for analyzing wind turbine blade images based on spatial trajectories, characterized in that: include: Acquire a plurality of drone-photographed pictures, wherein the plurality of drone-photographed pictures are obtained by drones for aerial photography of target wind turbine blades, and each of the drone-photographed pictures is marked with a shooting latitude and longitude and a shooting altitude; According to the shooting longitude and latitude and the shooting altitude, the actual distance offset between the centers of two adjacent drone-shot pictures is calculated; Calculating a pixel conversion ratio according to the actual center distance offset and the lens parameters of the drone; Calculating a pixel offset according to the pixel conversion ratio, and adjusting the plurality of drone-photographed pictures according to the pixel offset to obtain a plurality of adjusted drone-photographed pictures; Stitching the multiple adjusted drone-photographed images to obtain a panoramic image of the target wind turbine blades; Calling a pre-trained abnormality recognition model to obtain an abnormal part recognition result corresponding to the leaf panoramic image; According to the abnormal part identification result, an estimated degradation result of the corresponding surrounding part is obtained; The calling of the pre-trained abnormality recognition model to obtain the abnormal part recognition result corresponding to the leaf panoramic picture includes: Obtain leaf features corresponding to the leaf panoramic image; Through the target recognition unit in the abnormality recognition model, based on the multi-layer perceptron, according to each retrieval feature set, a candidate retrieval feature set corresponding to each retrieval feature set is obtained; the candidate retrieval feature set includes candidate retrieval features corresponding to each retrieval feature in the corresponding retrieval feature set; each layer recognition unit in the abnormality recognition model is the target recognition unit, or the previous preset order of magnitude recognition unit in the abnormality recognition model is the target recognition unit; Through the target recognition unit, based on the feedforward network, according to each of the candidate retrieval feature sets and the leaf features, the recognition retrieval feature set corresponding to each of the candidate retrieval feature sets is obtained; the recognition retrieval feature set includes the recognition retrieval feature corresponding to each candidate retrieval feature in the corresponding candidate retrieval feature set; multiple retrieval feature sets are obtained by grouping each pending retrieval feature corresponding to the target recognition unit; the pending retrieval feature is obtained according to the original retrieval feature, or is selected from each recognition retrieval feature generated by the previous level recognition unit of the target recognition unit based on a preset feature standard, and the target recognition unit is a first-level recognition unit in the abnormal recognition model; According to each recognition retrieval feature generated by the final recognition unit in the abnormality recognition model, the abnormal part recognition result corresponding to the leaf panoramic picture is obtained.

2. The method according to claim 1, characterized in that When the target recognition unit is a recognition unit other than the first level in the abnormal recognition model, each pending retrieval feature corresponding to the target recognition unit is obtained by the following process, including: According to each identification retrieval feature generated by the previous level identification unit, an abnormal part identification result corresponding to each identification retrieval feature is obtained; the abnormal part identification result includes the identification confidence corresponding to the identification part indicated by the identification retrieval feature; the loaded content of the previous level identification unit includes a plurality of retrieval feature sets obtained by grouping the pending retrieval features corresponding to the previous level identification unit; the generated content of the previous level identification unit includes the identification retrieval feature set corresponding to the plurality of retrieval feature sets; For each of the identification retrieval feature sets, based on the recognition confidence in the abnormal part recognition result corresponding to each identification retrieval feature, the identification retrieval feature that meets the preset feature standard is selected from the identification retrieval feature set and determined as the pending retrieval feature corresponding to the target recognition unit.

3. The method according to claim 1, characterized in that When the target recognition unit is a first-level recognition unit in the abnormality recognition model, each pending retrieval feature corresponding to the target recognition unit is obtained by the following process, including: Determine each of the original search features as each pending search feature corresponding to the target recognition unit; Alternatively, the feature reliability corresponding to each of the original search features is obtained; based on the one-to-one feature reliability of each of the original search features, an original search feature that meets the feature screening criteria is selected from each of the original search features and determined as the pending search feature corresponding to the target recognition unit.

4. The method according to claim 1, characterized in that: The retrieval feature set is obtained by the following process: According to a predetermined number of group divisions, each pending retrieval feature corresponding to the target recognition unit is evenly distributed to the multiple retrieval feature sets; Obtaining the feature reliability of each pending retrieval feature corresponding to the target recognition unit; dividing each pending retrieval feature into the plurality of retrieval feature sets according to the feature reliability of each pending retrieval feature according to the predetermined number of group divisions; The feature space position of each pending retrieval feature corresponding to the target recognition unit in the feature domain is obtained; and each pending retrieval feature is divided into the plurality of retrieval feature sets according to the feature space position of each pending retrieval feature in the feature domain according to a predetermined number of group divisions.

5. The method according to claim 1, characterized in that The step of obtaining the estimated degradation result of the corresponding surrounding parts according to the abnormal part identification result includes: Obtain the abnormal part attribute corresponding to the abnormal part identification result and the operating environment attribute corresponding to the fan blade flying operating environment factor; According to the encoding unit, encoding operations are performed on the abnormal part attribute and the operating environment attribute respectively to obtain abnormal part parameters corresponding to the abnormal part attribute and operating environment parameters corresponding to the operating environment attribute; Performing an integration operation on the abnormal part parameter and the operating environment parameter according to the integration unit to obtain an integrated characteristic parameter; Inputting the integrated feature parameters into a plurality of professional knowledge networks to perform initial feature extraction processing to obtain initial feature data generated by each of the professional knowledge networks; Determining a first weighting coefficient corresponding to each of the professional knowledge networks according to the first decision component corresponding to each of the surrounding degradation estimation indications, and performing weighted summation on the initial feature data generated by each of the professional knowledge networks according to the first weighting coefficient corresponding to each of the professional knowledge networks to obtain first summary feature data corresponding to each of the surrounding degradation estimation indications; Performing context-adaptive conversion on each of the first summary feature data according to a plurality of evaluation anchor points to obtain a primary feature set generated by each of the evaluation anchor points; Determining a second weighting coefficient corresponding to each of the professional knowledge networks according to the second decision component corresponding to each of the peripheral degradation estimation indications, and performing weighted summation on the initial feature data generated by each of the professional knowledge networks according to the second weighting coefficient corresponding to each of the professional knowledge networks to obtain second summary feature data corresponding to each of the peripheral degradation estimation indications; Performing time series correlation feature extraction processing on each of the second summary feature data according to the time series correlation model to obtain a secondary feature set generated by the time series correlation model; the peripheral degradation prediction indication is used to indicate the simulation demonstration of the abnormal part identification result on the fan blade operating environment factor; there is a time series correlation relationship between multiple peripheral degradation prediction indications; Determining a weighting coefficient corresponding to the primary feature set and a weighting coefficient corresponding to the secondary feature set according to the third decision component corresponding to each of the surrounding degradation prediction indications, and summing the primary feature set and the secondary feature set according to the weighting coefficient corresponding to the primary feature set and the weighting coefficient corresponding to the secondary feature set to obtain third summary feature data corresponding to each of the surrounding degradation prediction indications; The feature mapping unit performs mapping processing on the third summary feature data corresponding to each of the peripheral degradation prediction indications to obtain the estimated degradation result of the peripheral part corresponding to each of the peripheral degradation prediction indications.

6. The method according to claim 5, characterized in that The plurality of peripheral degradation prediction indications include a first peripheral degradation prediction indication and a second peripheral degradation prediction indication, and the first peripheral degradation prediction indication is located before the second peripheral degradation prediction indication in a time series set having a time series correlation relationship; the performing time series correlation feature extraction processing on each of the second summary feature data according to the time series correlation model to obtain a secondary feature set generated by the time series correlation model includes: performing a time series correlation feature extraction process on the second summary feature data corresponding to the first surrounding degradation prediction indication according to the time series correlation model to obtain a secondary feature set corresponding to the first surrounding degradation prediction indication, and performing a time series correlation feature extraction process on the second summary feature data corresponding to the first surrounding degradation prediction indication and the second summary feature data corresponding to the second surrounding degradation prediction indication according to the time series correlation model to obtain a secondary feature set corresponding to the second surrounding degradation prediction indication; The secondary feature set corresponding to the first surrounding degradation prediction indication and the secondary feature set corresponding to the second surrounding degradation prediction indication are determined as the secondary feature sets generated by the time series association model.

7. The method according to claim 5, characterized in that The temporal association model includes at least one temporal association secondary model, each of which includes a parallel attention mechanism unit and a forward conduction neural network unit; performing temporal association feature extraction processing on each of the second summary feature data according to the temporal association model to obtain a secondary feature set generated by the temporal association model includes: Input each of the second summary feature data into the parallel attention mechanism unit to perform weight distribution calculation, and obtain weight distribution data corresponding to each of the second summary feature data; The weight allocation data corresponding to each of the second summary feature data is input into the forward conduction neural network unit for feature refinement to obtain a secondary feature set generated by the time series association model.

8. The method according to claim 5, characterized in that The step of obtaining an estimated degradation result of a corresponding peripheral part according to the abnormal part identification result further includes: Acquire multiple initial professional knowledge networks, multiple initial assessment anchor points corresponding to the peripheral degradation prediction indications, an initial time series association model, training instances, and instance category labels corresponding to the training instances; the training instances include sample abnormal part attributes corresponding to the sample objects and sample operating environment attributes corresponding to the sample recommended resources; Combining a plurality of the initial professional knowledge networks, a plurality of initial evaluation anchor points corresponding to the peripheral degradation prediction indications, and the initial time series association model into an original multi-directional prediction model; Performing initial feature extraction processing on the abnormal part attributes of the sample and the operating environment attributes of the sample according to the multiple initial professional knowledge networks to obtain initial feature data generated by each of the initial professional knowledge networks; Performing context-adaptive conversion on the initial feature data generated by each of the initial professional knowledge networks according to the multiple initial evaluation anchor points to obtain a sample primary feature set generated by each of the initial evaluation anchor points, and performing temporal association feature extraction processing on the initial feature data generated by each of the initial professional knowledge networks according to the initial temporal association model to obtain a sample secondary feature set generated by the initial temporal association model; Performing prediction processing according to the sample primary feature set corresponding to each of the surrounding degradation prediction indications to obtain a first predicted degradation result corresponding to each of the surrounding degradation prediction indications, and performing prediction processing according to the sample secondary feature set corresponding to each of the surrounding degradation prediction indications to obtain a second predicted degradation result corresponding to each of the surrounding degradation prediction indications; Performing multi-directional estimation processing according to the sample primary feature set and the sample secondary feature set to obtain a third estimated degradation result corresponding to each of the peripheral degradation estimation indications; performing error calculation according to the first estimated degradation result and the instance category label to obtain a first error parameter, performing error calculation according to the second estimated degradation result and the instance category label to obtain a second error parameter, and performing error calculation according to the third estimated degradation result and the instance category label to obtain a third error parameter; Determining a model error parameter according to the first error parameter, the second error parameter and the third error parameter; The original multi-directional prediction model is learned and optimized according to the model error parameters until the model error parameters meet the error threshold, thereby obtaining a target multi-directional prediction model; the target multi-directional prediction model includes a plurality of the professional knowledge networks, a plurality of evaluation anchor points corresponding to the peripheral degradation prediction indications, and the timing association model.

9. A server system, characterized in that: The method comprises a server, wherein the server is used to execute the method described in any one of claims 1 to 8.

Images