Non-contact blade vibration recognition system based on tower visual response

Abstract

This invention relates to the field of blade vibration monitoring technology, specifically to a non-contact blade vibration identification system based on the visual response of the tower. It includes a tower vibration compensation module configured to acquire frontal image data of the tower and determine a vibration compensation vector based on the frontal image data; and an anomaly identification module configured to acquire back and frontal image data of the blades and the vibration compensation vector, and identify abnormal features on the blade surface based on these data. This invention acquires tower vibration data through a stable acquisition source at the bottom of the tower to generate a precise vibration compensation vector, achieving efficient elimination of tower vibration interference and thus solving the problem of false vibration signal interference caused by tower vibration in traditional monitoring.

Description

Technical Field

[0001] This invention relates to the field of blade vibration monitoring technology, and more specifically, to a non-contact blade vibration identification system based on tower visual response. Background Technology

[0002] Traditional blade vibration monitoring relies heavily on contact sensors or single-parameter measurement techniques, which have limitations such as complex installation, poor environmental adaptability, and limited monitoring dimensions. Contact sensors are susceptible to mechanical wear and electromagnetic interference, resulting in insufficient long-term stability.

[0003] In related technologies, a non-contact visual displacement detection method predicts the integrity of the marker's edge and then matches the corresponding center coordinates with the pixel offset calculation method. This method avoids the problems of physical contact and accuracy defects caused by partial loss of the marker due to uneven lighting, occlusion, or other reasons.

[0004] However, when it comes to wind turbine towers, cameras mainly collect image data of the blades. If the vibrations generated by the tower are not taken into account, it will directly lead to a complete distortion of all the core parameters of the blade monitoring. Summary of the Invention

[0005] The purpose of this invention is to provide a non-contact identification system for blade vibration based on the visual response of the tower. By collecting image data of the tower, the vibration compensation vector of the tower is calculated, and the vibration compensation vector is used to eliminate the interference of tower vibration on the blade, so as to solve the problem of parameter distortion in blade monitoring caused by tower vibration.

[0006] The system includes: The first image data acquisition module is configured to acquire image data of the back of the blade. The second image data acquisition module is configured to acquire frontal image data of the tower body and frontal image data of the blades. The tower vibration compensation module is configured to acquire frontal image data of the tower and determine the vibration compensation vector of the tower based on the frontal image data. The anomaly identification module is configured to acquire image data of the back of the blade, image data of the front of the blade, and vibration compensation vector, and to identify abnormal features on the blade surface based on the image data of the back of the blade, image data of the front of the blade, and vibration compensation vector; wherein the vibration compensation vector is used to eliminate tower vibration interference.

[0007] Considering that wind turbines are installed outdoors, and there are too many uncontrollable factors in the outdoor geographical environment, the image data acquisition source can be optionally installed on the tower. This way, the installation environment of the acquisition source is located on the tower, avoiding interference from uncontrollable geographical factors and reducing the installation cost of the acquisition source.

[0008] The acquisition source refers to the original source of image data acquisition. It can be composed of a single image acquisition device or a combination of multiple image acquisition devices distributed at different locations, heights, or angles. The acquisition source is used to uniformly represent the overall collection of image acquisition devices that acquire image data, rather than referring to a single image acquisition device or a single physical installation point.

[0009] When the data acquisition source includes multiple image acquisition devices, these devices are deployed at the same or different heights within the tower and work collaboratively in an array to synchronously acquire image data from the same monitoring area. This array-based deployment and synchronous collaboration of multiple image acquisition devices compensates for the limitations of a single image acquisition device, such as limited viewing angle and insufficient anti-interference capabilities. It provides multi-dimensional, highly reliable image data support for tower vibration compensation and blade anomaly identification, ensuring the integrity and accuracy of the system's blade condition monitoring.

[0010] Optionally, the image acquisition equipment uses an industrial-grade camera. Industrial-grade cameras have a high protection rating and wide operating temperature range, and can withstand wind, sand, rain, snow, ultraviolet radiation, and extreme temperature changes. Therefore, industrial-grade cameras are the best choice for complex outdoor environments and tower vibration conditions. With their multi-dimensional performance advantages, they can provide stable, clear, and distortion-free image data for tower vibration compensation and blade anomaly identification.

[0011] Considering that the object being acquired by the acquisition source, namely the blade, is located at the top of the tower, the height of the acquisition source can be set according to actual needs. Optionally, the acquisition source for the image data of the back of the blade is located in the first height range of the tower and is positioned on the back of the blade.

[0012] The image data of the blade's back can be acquired from one or more sources, and the height of these sources is all within a first height range. In one possible approach, the one or more sources acquire image data of the blade's back from the same or different regions.

[0013] To reduce interference from the acquisition angle, image data is typically acquired continuously when the blade is almost at its lowest position. At this point, only the tilt angle exists, without interference from left or right lateral tilt angles. When acquiring image data from the lower side upwards at this time, the tilt angle at the lowest point of the blade is minimal. Optionally, the height of the first height interval should not be lower than the height of the lowest point during blade rotation. This avoids problems such as perspective distortion and edge deformation in blade imaging caused by excessive tilt angles, and also ensures that at least a portion of the blade is directly facing the acquisition source. However, considering that the first height interval is already within the blade's rotation path, the acquisition source is placed on the back of the blade to avoid interference from the acquisition source with the blade's rotation path.

[0014] Optionally, a data acquisition source can be installed at the height of the key monitoring area, so that the key monitoring area is directly facing the data acquisition source, thereby improving the accuracy of image data acquisition for the key monitoring area. The key monitoring area includes the critical structural parts of the blade, the core area that is prone to deformation and vibration, and the key parts that are prone to surface anomalies, such as the blade tip, blade root, and blade middle, as well as the windward side, leeward side, and edge contour of the blade, which are prone to surface defects or vibration deformation.

[0015] In reality, the image source for the back of the blade is positioned so that the camera is taken from the tower towards the blade. In other words, the camera's shooting direction originates from the tower. However, to acquire the front image data of the blade, the acquisition source must be positioned on the front of the blade. Therefore, the camera cannot be positioned from the tower; it must be taken directly towards the tower. Thus, when acquiring the front image data of the blade, it is possible to acquire the front image data of the tower. Therefore, the front image data of the tower and the front image data of the blade share the same acquisition source, which is located on the front of the blade. Furthermore, this same acquisition source for both the front image data of the tower and the front image data of the blade is positioned within the second height range of the tower.

[0016] Considering that the acquisition source located in the second height range needs to collect frontal image data of the tower and use this data to determine the tower's vibration compensation vector to eliminate tower vibration interference, the accuracy of the frontal image data acquisition directly determines the accuracy of the vibration compensation vector. Optionally, the second height range is smaller than the first height range, thus avoiding its location and improving the diversity of image data acquisition. The second height range is the height range where the bottom of the tower is located. In other words, since the second height range is set at the bottom of the tower, the same acquisition source for both the tower's frontal image data and the blade's frontal image data will be installed at the bottom of the tower. The bottom of the tower is a fixed end rigidly connected to the foundation, and its structural stability is much higher than that of the upper part of the tower. Installing the acquisition source here reduces the possibility of tower vibration causing displacement or vibration itself. Under a stable acquisition source baseline, the acquired frontal image data of the tower can accurately deduce the tower's vibration characteristics at different times. Only the vibration compensation vector calculated based on these real vibration characteristics can accurately offset the interference of tower vibration on blade monitoring.

[0017] Optionally, only one acquisition source is set for the same acquisition source for the tower front image data and the blade front image data. Considering that the image data acquired by this acquisition source can almost cover the upper area of ​​the tower, there is no need to set up multiple acquisition sources to acquire data in different areas.

[0018] It should be noted that the target corresponds to the data acquisition source. The target is an artificial feature marker pre-fixed on the front surface of the tower (it can be a single high-recognition feature point, a feature group composed of multiple feature points, or a marker area with a regular geometric shape). Its core function is to serve as a reference marker for visual tracking, compensating for the shortcomings of the tower surface's simple natural texture and insufficient feature points for extraction, and providing a stable and clear identification object for the visual measurement of tower vibration.

[0019] In reality, both wind turbine blade vibration and tower vibration are dynamic, temporal processes. Therefore, calculating the tower's vibration displacement vector, extracting blade vibration characteristics, and identifying surface anomalies essentially involve analyzing the changes in the object's position and shape over time. Thus, a sequence of image frames arranged chronologically can completely record the visual changes of the tower / blades over a continuous period. Therefore, all image data needs to be preprocessed to obtain an image frame sequence. This preprocessing records the visual changes of the tower / blades over a continuous period and also preprocesses image data from different sources into a timestamp-synchronized frame sequence, ensuring accurate matching of the blade's back image, blade's front image, and tower's front image at the same time point.

[0020] Image data refers to all the raw visual data acquired by the acquisition source from the back of the blade, the front of the blade, and the front of the tower. An image frame is the basic static unit that constitutes dynamic visual data. Each frame corresponds to a precise point in time and contains the complete static visual features of the tower / blade at that moment (such as the pixel position of the tower target, the texture / defects on the blade surface, and the outline shape of the blade). An image frame sequence is a time-series visual data set formed by arranging multiple preprocessed image frames in the order of their acquisition time.

[0021] In one possible technical solution, the tower vibration compensation module includes: The image preprocessing unit is configured to receive the front image data of the tower body and preprocess the front image data of the tower body to obtain the front image frame sequence of the tower body. The target feature point localization and tracking unit is configured to receive the tower front image frame sequence and extract the two-dimensional pixel coordinates of multiple predefined feature points on the target in each tower front image frame, establish the correspondence of the same feature point in the time series, so as to obtain the motion trajectory of each feature point in the image plane. The tower surface vibration displacement calculation unit is configured to receive the motion trajectory of each feature point in the image plane, and back estimate the vibration parameters of the tower at the target position at each moment based on the motion trajectory. The vibration compensation vector generation unit is configured to receive the vibration parameters of the tower at the target position at each moment and input the vibration parameters into the tower vibration transmission model. The tower vibration transmission model calculates and outputs the high-level vibration compensation vector of the tower based on the acquisition source of the tower front image data.

[0022] This scheme generates the vibration compensation vector at the current moment through a vibration compensation vector generation unit. At the algorithm level, it can separate the tower-following component of the blade from the blade's autonomous vibration component, thereby solving the problem of pseudo vibration signals caused by tower vibration in the blade monitoring data. This ensures that the extracted blade vibration features only reflect the structural state of the blade itself, rather than the mixed motion of the tower and the blade.

[0023] In one possible technical solution, the anomaly identification module includes: A multi-source image preprocessing unit is configured to receive image data from the back of the blade and image data from the front of the blade, and to preprocess the image data from the back of the blade and image data from the front of the blade to obtain a time-stamped sequence of image frames from the back of the blade and a sequence of image frames from the front of the blade. The leaf feature extraction unit is configured to receive the leaf back image frame sequence and the leaf front image frame sequence, extract features from the leaf back image frame sequence and the leaf front image frame sequence, and integrate the features from different acquisition sources at the same time into the original leaf feature dataset. The decoupled calculation unit is configured to receive the original feature dataset of the blade and embed the vibration compensation vector into the feature points for decoupled calculation, so that the calculated net features and net state parameters are decoupled from the tower vibration displacement. An anomaly feature recognition unit is configured to receive net features and net state parameters, and to perform anomaly recognition based on the net features and net state parameters.

[0024] Traditional techniques often attempt to correct vibrating images through image registration or affine transformations. However, this not only consumes enormous computational resources and causes system latency, but also introduces unnecessary image blurring and artifacts due to pixel interpolation, destroying the details of minute defects. This solution, on the other hand, achieves a dynamic correction at the mathematical level by directly embedding the tower vibration compensation vector into the already extracted feature point coordinates. This design cleverly bypasses cumbersome pixel-level reconstruction, compensating only the lightweight coordinate data while preserving the integrity of the original image details, resulting in an exponential improvement in computational efficiency.

[0025] This decoupling method ultimately yields net features and net state parameters. For the subsequent anomaly identification unit, it no longer receives cluttered signals affected by tower vibration, but rather the blade's intrinsic response after eliminating common-mode interference from the tower. This decoupling design not only allows the system to sensitively detect surface cracks or subtle vibration frequency shifts on the blade, but also fundamentally avoids visual false alarms caused by tower vibration.

[0026] In one possible technical solution, based on the two previous solutions, the decoupling calculation unit verifies the net features and net state parameters obtained from the blade back image frame sequence using the net features and net state parameters obtained from the blade front image frame sequence during decoupling calculation. This solution considers that the acquisition source located in the second height range is installed at the bottom of the tower, which is almost unaffected by the flexible vibration of the tower. Therefore, the blade front image data acquired from this perspective has extremely high benchmark authenticity.

[0027] When capturing the features of the blade's back side, the raw data from the high-position camera, which vibrates synchronously with the tower, contains complex nonlinear couplings. The net features of the blade's front side, acquired through a bottom steady-state acquisition source, can serve as a standard reference for real-time reverse verification of the decoupling results of the back side features. Specifically, the decoupling calculation unit compares the blade's net state parameters calculated from the front (steady-state view) and back (follow-up view) at the same moment. If both show high consistency within the compensated data envelopment, it proves that the current vibration compensation vector and decoupling algorithm are accurate and effective. If significant deviations occur, the system triggers an adaptive correction procedure, utilizing the wide-area field of view advantage at the bottom to recalibrate the spatial mapping relationship of the high-position camera.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows: Vibration data of the tower is acquired by a stable acquisition source at the bottom of the tower to generate a precise vibration compensation vector, thereby efficiently eliminating tower vibration interference and solving the problem of spurious vibration signal interference caused by tower vibration in traditional monitoring. Specifically, decoupled calculations effectively separate the autonomous vibration characteristics of the blades caused by structural, load, and material factors, accurately retaining the core information reflecting the abnormal state of the blades. Ultimately, this improves the accuracy of blade vibration monitoring, avoids the risk of false alarms, and ensures the accuracy of anomaly identification, providing reliable support for early warning of blade damage and health status assessment. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the non-contact identification system of the present invention; Figure 2 This is a schematic diagram of the tower vibration compensation module structure of the present invention; Figure 3 This is a schematic diagram of the anomaly identification module structure of the present invention.

[0030] Numbering on the map: 100. First image data acquisition module; 200. Second image data acquisition module; 300. Tower vibration compensation module; 310. Image preprocessing unit; 320. Target feature point localization and tracking unit; 330. Tower surface vibration displacement calculation unit; 340. Vibration compensation vector generation unit; 400. Anomaly identification module; 410. Multi-source image preprocessing unit; 420. Blade feature extraction unit; 430. Decoupling calculation unit; 440. Anomaly feature identification unit. Detailed Implementation

[0031] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0032] Figure 1 A non-contact blade vibration identification system based on tower visual response is shown, including: The first image data acquisition module 100 is configured to acquire image data of the back of the blade. The second image data acquisition module 200 is configured to acquire frontal image data of the tower body and frontal image data of the blades. The tower vibration compensation module 300 is configured to acquire frontal image data of the tower and determine the vibration compensation vector of the tower based on the frontal image data of the tower. The anomaly identification module 400 is configured to acquire image data of the back of the blade, image data of the front of the blade, and vibration compensation vector, and to identify abnormal features on the blade surface based on the image data of the back of the blade, image data of the front of the blade, and vibration compensation vector; wherein the vibration compensation vector is used to eliminate tower vibration interference.

[0033] In one possible embodiment, the image data of the back of the blade acquired by the first image data acquisition module 100 is collected by one or more acquisition sources installed on the tower. The acquisition source can be a single image acquisition device or a combination of multiple image acquisition devices distributed at different locations, heights, or angles.

[0034] During implementation, the number of data acquisition sources should be determined based on actual operating conditions such as the length of a single blade, the number of monitoring locations, and outdoor environmental conditions. For example, when the blade length of a wind turbine is relatively short, such as within 30 meters, the monitoring coverage of a single data acquisition source can meet the monitoring needs of the entire blade area, and only one data acquisition source is required. However, when the blade length is relatively long, such as greater than 30 meters, the areas to be monitored are widely distributed, and a single data acquisition source may not be able to cover the entire monitoring area. In this case, 2-3 data acquisition sources can be set up according to the actual situation.

[0035] Simultaneously, the impact of other actual operating conditions on the number of data acquisition sources must be comprehensively considered. For example, if the blade length is less than 30 meters, but the number of parts to be monitored is large and their distribution is scattered, relying solely on a single data acquisition source may not be sufficient to effectively cover all critical parts. In this case, the number of data acquisition sources should be appropriately increased according to monitoring needs to ensure no omissions in monitoring. Furthermore, the acquisition capability of the data acquisition source itself is also an important consideration. If a single data acquisition source is equipped with multiple image acquisition devices, and the placement of each device is reasonable with a wide shooting angle coverage, it can achieve simultaneous acquisition of a large area, significantly improving its overall acquisition capability. Based on this, for blades longer than 30 meters, even with many monitoring parts, it may be possible to meet the monitoring needs of the entire area with only one data acquisition source, avoiding resource waste due to blindly adding more data acquisition sources.

[0036] Preferably, when the acquisition source includes multiple image acquisition devices, these devices are deployed at the same or different heights on the tower and work collaboratively in an array to synchronously acquire image data of the same monitoring area. For example, for wind turbine blades longer than 30 meters, the acquisition source can evenly deploy 2-3 image acquisition devices circumferentially on a mounting frame at the same height on the tower. The shooting angles of each device are interconnected to form an array coverage, synchronously capturing images from more angles during the blade's rotation, thus compensating for the limitations of a single device's shooting angle. However, due to limitations in the installation space and hardware integration scale of the acquisition source, the number of image acquisition devices in a single acquisition source will not be too large, and its coverage and monitoring capabilities still have clear boundaries.

[0037] Preferably, the height difference between the image acquisition devices of a single acquisition source is typically controlled within a very small range of 0.5 meters to 2 meters. For example, on an installation platform at a fixed height on the tower, two image acquisition devices can be fixed on the upper and lower sides of the mounting frame, respectively, with a height difference of only 1 meter. This array-based layout achieves complementary horizontal viewing angles. If the height difference between the devices exceeds 3 meters, the excessive vertical span of the shooting angles will lead to overlapping blind spots or acquisition distortion in the images of the same blade monitoring area, thus reducing the effectiveness of monitoring. Therefore, the height distribution of image acquisition devices within a single acquisition source is always limited to integration onto the same mounting carrier and a reasonably small height difference, ensuring the consistency and coverage effectiveness of the acquisition source as a single monitoring source.

[0038] Preferably, in this embodiment, the image acquisition device uses an industrial-grade camera; for example, a Basler ace 2 series 20-megapixel global shutter industrial camera is selected, which has anti-vibration and wide-temperature operating characteristics, adapts to outdoor wind and sand and temperature changes, and can stably capture details of high-speed blade rotation.

[0039] The above industrial-grade cameras are merely examples. Those skilled in the art can select other industrial-grade cameras that meet the functional characteristics according to actual monitoring needs, environmental adaptability, and imaging performance requirements, and all of them fall within the protection scope of this invention.

[0040] In one possible embodiment, the tower front image data and blade front image data acquired by the second image data acquisition module 200 are acquired by one or more acquisition sources installed on the tower. The acquisition source can be a single image acquisition device, or it can be a combination of multiple image acquisition devices distributed at different locations, heights, or angles.

[0041] In practice, because the acquisition sources used to collect images of the front of the tower and the blades need to cover a wide area of ​​the monitoring region, they are typically deployed at the bottom of the tower, away from the blades. This installation location ensures that the field of view of the acquisition source covers the main area of ​​the tower front and the blades when they rotate to the front, and the installation conditions at the bottom are more stable. With this deployment method, the number of acquisition sources is usually not too large; the focus is on ensuring the acquisition accuracy of the image acquisition equipment. For example, when the wind turbine blades are longer than 30 meters and clear capture of blade surface defects is required, if only one acquisition source is set at the bottom of the tower, its industrial-grade camera must have high resolution, a wide field of view, and strong environmental adaptability to ensure that clear and complete image data can still be obtained under conditions such as high-speed blade rotation, changes in outdoor lighting, and wind and sand interference, providing reliable data support for subsequent anomaly identification.

[0042] In this embodiment, the industrial-grade camera preferably adopts the Basler acA2500-14gmNIR series 25-megapixel global shutter industrial camera. This camera is equipped with a high-resolution lens and a high-sensitivity CMOS sensor, which can clearly reproduce the fine texture and defect features of the blade surface under long-distance shooting conditions, meeting the high-precision imaging requirements of long-distance monitoring.

[0043] The above industrial-grade cameras are merely examples. Those skilled in the art can select other industrial-grade cameras that meet the functional characteristics according to actual monitoring needs, environmental adaptability, and imaging performance requirements, and all of them fall within the protection scope of this invention.

[0044] In one possible embodiment, the blade is 40 meters long and the tower is 100 meters high. It is assumed that the center of rotation of the blade is at the top of the tower, i.e., at 100 meters. In this embodiment, since the blade length exceeds 30 meters, two acquisition sources are chosen to ensure the quality of image data acquisition on the back of the blade. Furthermore, considering that image acquisition devices typically acquire image data when the blade is almost at its lowest point, one acquisition source is positioned at 60 meters on the tower, and the other at 95 meters, close to the root of the blade, to accurately capture the blade tip and root.

[0045] In this embodiment, the height range of the acquisition source equal to or higher than 60 meters is defined as the first height range, and the acquisition source of the image data on the back of the blade is required to be set within the first height range of the tower body, and the acquisition source is set on the back of the blade to avoid interference with the rotation of the blade.

[0046] In one possible embodiment, since the front of the tower and the front of the blades appear in the same image, the acquisition of the image data of the front of the tower and the image data of the front of the blades share a common acquisition source. This acquisition source is set on the front of the blades. Preferably, a single acquisition source is set up to acquire the image data of the front of the tower and the image data of the front of the blades, and the height of the acquisition source is between 5 meters and 10 meters. This height range of 5 meters to 10 meters is defined as the second height range, which can avoid some infrastructure at the bottom of the tower, increase the difficulty of human damage, and provide a certain degree of protection for the acquisition source. At the same time, it can also utilize the stability of the bottom of the tower to ensure the authenticity of the acquired image data.

[0047] In one possible embodiment, Figure 2 The tower vibration compensation module 300 is shown, including: The image preprocessing unit 310 is configured to receive the front image data of the tower body and preprocess the front image data of the tower body to obtain the front image frame sequence of the tower body. The target feature point localization and tracking unit 320 is configured to receive the tower front image frame sequence, and extract the two-dimensional pixel coordinates of multiple predefined feature points on the target in each tower front image frame, establish the correspondence of the same feature point in the time series, so as to obtain the motion trajectory of each feature point in the image plane. The tower surface vibration displacement calculation unit 330 is configured to receive the motion trajectory of each feature point in the image plane and estimate the vibration parameters of the tower at the target position at each moment based on the motion trajectory. The vibration compensation vector generation unit 340 is configured to receive the vibration parameters of the tower at the target position at each moment and input the vibration parameters into the tower vibration transmission model. The tower vibration transmission model calculates and outputs the high-level vibration compensation vector of the tower observed by the acquisition source based on the front image data of the tower.

[0048] Figure 3 An anomaly detection module 400 is shown, including: The multi-source image preprocessing unit 410 is configured to receive blade back image data and blade front image data, and preprocess the blade back image data and blade front image data to obtain a time-stamped blade back image frame sequence and blade front image frame sequence. The leaf feature extraction unit 420 is configured to receive the leaf back image frame sequence and the leaf front image frame sequence, extract features from the leaf back image frame sequence and the leaf front image frame sequence, and integrate the features from different acquisition sources at the same time into the original leaf feature dataset. Decoupling calculation unit 430 is configured to receive the original feature dataset of the blade and embed the vibration compensation vector into the feature points for decoupling calculation, so as to decouple the calculated net features and net state parameters from the tower vibration displacement. Anomaly feature recognition unit 440 is configured to receive net features and net state parameters, and to perform anomaly recognition based on the net features and net state parameters.

[0049] For ease of understanding, the acquisition source set at 60 meters on the tower is defined as the first acquisition source, the acquisition source set at 95 meters is defined as the second acquisition source, and the acquisition source set at 10 meters is defined as the third acquisition source. The first, second, and third acquisition sources work simultaneously. The first and second acquisition sources acquire image data of the back of the blades, while the third acquisition source acquires image data of the front of the tower and the front of the blades.

[0050] First, the image preprocessing unit 310 receives the front image data of the tower body and preprocesses the front image data of the tower body, including the following steps: The target location coordinates in the original front image of the tower are identified by using target detection algorithms based on YOLO and SSD. Based on these coordinates, the region of interest (ROI) containing the target is delineated, and redundant areas in the tower background that are not related to vibration calculation are removed to reduce the amount of subsequent data processing, while focusing on the core feature areas of the target.

[0051] To address the issue of target feature blurring caused by changes in outdoor lighting (strong light, weak light, shadow), an adaptive histogram equalization (CLAHE) algorithm is used to enhance the contrast of the ROI region, improving the recognition of key features such as target edges and textures. At the same time, the grayscale levels of the image are preserved to avoid target feature distortion due to over-enhancement, ensuring that the distinction between the target and the background in the image meets the accuracy requirements for vibration displacement calculation.

[0052] The enhanced image is denoised by using a bilateral filtering algorithm. This can filter out high-frequency noise introduced by wind and sand, equipment vibration, etc., while preserving low-frequency features such as the edge and contour of the target to the greatest extent. This avoids the impact of noise interference on the dynamic displacement measurement of the target and ensures the accuracy of vibration parameter calculation.

[0053] Based on the stability of target features, continuously acquired image frames are filtered to remove invalid frames with blurred, occluded, or motion-blurred target features. For the retained valid frames, subpixel-level frame alignment is performed with the target center as the reference to eliminate inter-frame offsets caused by slight shaking of the acquisition equipment or small displacement of the tower, ensuring that the position reference of the same target is consistent in each image frame, and providing time-consistent image data for accurate calculation of subsequent vibration displacement.

[0054] The valid image frames processed as described above are arranged in chronological order of acquisition time to generate a sequence of image frames of the front of the tower.

[0055] Preferably, 2 to 4 targets are set around the same height of the tower, ensuring that at least one target remains unobstructed regardless of the angle the blades rotate to, thus preventing the entire row of targets from being invisible due to blade occlusion. Black and white checkerboard targets or cross-shaped targets are selected (matching the high-contrast features of the pre-processed image to maximize feature point recognition). 4-8 feature points are pre-defined on the target surface, prioritizing points with unique geometric recognition, such as the intersection of target edges, the center of the cross, and the corners of the checkerboard, avoiding points with blurred surface textures or those susceptible to sudden changes in lighting. The physical relative positions of each feature point on the target (e.g., the actual straight-line distance and angle between adjacent feature points) are pre-measured and recorded. Simultaneously, this set of feature points is bound to the target ROI region defined in the pre-processing stage, limiting the feature point extraction range to the target area within the ROI, thus preventing the extraction of irrelevant feature points from the tower background from the outset.

[0056] Next, the target feature point localization and tracking unit 320 uses the Shi-Tomasi corner detection algorithm to identify candidate feature points on the target ROI region for each frame of the tower's frontal image. Based on the predefined relative positional relationship (actual distance, angle) of the feature points, the initially extracted candidate feature points are geometrically verified. For example, the pixel distance / angle between candidate feature points is calculated and compared with the predefined physical relative relationship (after conversion to pixel scale). Invalid candidate points with a deviation of more than 5% are eliminated, and only valid feature points that correspond one-to-one with the predefined feature points are retained. Simultaneously, the preprocessed ROI boundaries are used to filter out background pseudo-feature points that exceed the target area. Then, for the selected effective feature points, the Zernike matrix subpixel localization algorithm is used to calibrate the initial pixel coordinates of the feature points, improving the coordinate accuracy to the 0.1 pixel level. The subpixel-level frame alignment in the preprocessing stage has eliminated small inter-frame offsets, which can further ensure the calibration accuracy. Finally, the two-dimensional pixel coordinates (x, y) of each effective feature point in the current image frame are output, where the origin of the coordinates is the upper left corner of the image frame, x is the horizontal pixel axis, and y is the vertical pixel axis.

[0057] The first frame of the image frame sequence is selected as the initial tracking frame. A unique identifier (ID, such as ID1, ID2…IDn) is assigned to all valid feature points extracted in this frame. The two-dimensional pixel coordinates and geometric feature descriptors (such as corner response values ​​and neighborhood grayscale distribution) of each feature point are recorded to establish an initial feature point file. For each subsequent frame, the extracted valid feature points are matched with the feature points of the previous frame. The KNN (k-nearest neighbor) matching algorithm is preferred. The geometric descriptors of the feature points are compared, and an Euclidean distance threshold is set (to accommodate the low noise characteristics of the pre-processed image; the threshold is set below 20). Only candidate matching pairs with a matching distance less than the threshold are retained. Simultaneously, the geometric consistency of the matching pairs is verified by combining the predefined relative positional relationships of the feature points (e.g., the pixel distance between ID1 and ID2 must be consistent with the predefined value), eliminating false matches. If some feature points are lost due to temporary leaf occlusion or sudden changes in illumination, the remaining targets and redundant feature points (4-8) on the targets can be used to ensure that even if a single feature point is lost, it can still be completed through the relative positions of adjacent feature points, ensuring that the temporal correspondence of the same feature point is not interrupted.

[0058] For each feature point with a unique ID, its two-dimensional pixel coordinate sequence across all frames is calculated according to the acquisition time sequence of the image frames. , … By concatenating these coordinates sequentially, a sequence of discrete coordinates of the feature point in the image plane is formed: By using polynomial fitting, the discrete coordinate sequence is transformed into a continuous curve, thus obtaining the complete motion trajectory of the feature point.

[0059] The tower surface vibration displacement calculation unit 330 includes: Based on the pixel equivalent k, the motion trajectory of the feature points on the image plane is converted into the actual instantaneous vibration displacement of the tower at the target installation position: Select the pixel coordinates of feature points during the steady-state vibration phase As the reference pixel position when the tower is stationary; Instantaneous vibration displacement and The calculation formula is as follows: ; ; In the formula, The instantaneous horizontal vibration displacement of the tower body at the target position at time t (perpendicular to the tower body axis), in mm; The instantaneous vertical vibration displacement of the tower body at the target position at time t (parallel to the tower body axis), unit: mm; in, In the formula, The actual physical straight-line distance between adjacent feature points pre-recorded on the target; It represents the linear distance between pixels of the same set of feature points in an image frame.

[0060] Based on the instantaneous vibration displacement sequence, the core vibration parameters of the tower body are extracted, including: Vibration amplitude: ; In the formula, The amplitude of the horizontal vibration (mm); The vertical vibration amplitude is (mm); max is the maximum value of the sequence, and min is the minimum value of the sequence.

[0061] Vibration frequency is the frequency of vibration. and Perform a Fast Fourier Transform to obtain the principal vibration frequencies of the tower body, and then take the frequency value corresponding to the peak value in the frequency domain. and (Hz).

[0062] Vibration phase: ; In the formula, The vibration phase (rad) at time t in the horizontal direction. The vibration phase (rad) at time t in the vertical direction.

[0063] Instantaneous vibration velocity: ; In the formula, Let t be the instantaneous horizontal vibration velocity (mm / s). The instantaneous vibration velocity in the vertical direction at time t (mm / s); The frame acquisition interval.

[0064] For each feature point i, calculate its instantaneous vibration displacement using the original method. , ;amplitude , ;frequency , Phase , ;speed , By combining the weights of the relative positions of feature points (the closer the actual distance between adjacent feature points, the higher the weight), the mean vibration parameters of the entire target are calculated. , , , , , , , , and .

[0065] The vibration compensation vector generation unit 340 inputs the average vibration parameters into the tower vibration transmission model. The core calculation formulas in the tower vibration transmission model include: ; ; ; ; ; ; ; ; In the formula, Let be the actual horizontal vibration displacement of the tower at time t. Let t be the actual vertical vibration displacement of the tower body at time t; The X-axis is the deformation transfer coefficient of the tower structure; The deformation transfer coefficient of the tower structure along the Y-axis. and Related to the material stiffness and dynamic inertia of the tower body, during the system initialization phase, the three-dimensional model of the tower body design drawings is input and static simulation calculations are performed using finite element analysis (FEA) software; or empirical values ​​are assigned by fitting the actual displacement ratio between the bottom and top through a static push-pull experiment with a known tensile force applied on site. The X-axis tower body harmonic resonance gain coefficient; The Y-axis tower harmonic resonance gain coefficient. and To characterize the amplification effect of simple harmonic vibration caused by wind load or impeller rotation frequency close to the natural frequency of the tower, the modal analysis report of the wind turbine generator is used to obtain the natural frequencies of the tower at each order. After the initial installation of the system, continuous vibration data under a typical working condition is recorded, and the peak ratio under a specific frequency band is extracted by fast Fourier transform (FFT) for calibration. The time delay from the vibration to the acquisition source; The X-axis velocity feedforward time constant (unit: s); Y-axis velocity feedforward time constant (unit: s). and As the core of the velocity compensation term, this constant is introduced to convert instantaneous velocity into displacement compensation. The theoretical delay reference value is pre-calculated based on the tower height divided by the structural sound velocity, and during actual system operation, the timestamp difference between the sensors at the bottom and top ends is dynamically fitted and updated using a Kalman filter algorithm. The height difference between the target and the acquisition source; The height of the tower foundation is a pre-set constant, fixed at 10 meters; This represents the static deformation offset of the tower body along the X-axis. This represents the static deformation offset of the tower body along the Y-axis. and The system characterizes the permanent or extremely low-frequency static tilt displacement of the tower body under the action of gravity, long-term prevailing wind pressure, or foundation settlement. Under the condition of calm wind with the unit shut down and the ambient wind speed below 3m / s, the system continuously collects and averages the reference coordinate deviation of the bottom camera for multiple frames, and fixes it in the model as a static system constant. The installation calibration error of the vision system characterizes the mechanical alignment error between the optical center axis of the high-position camera and the theoretical center axis of the tower during the physical installation process. After the camera equipment is installed on-site, the camera extrinsic parameters are calibrated using a target, and the fixed deviation value in the installation translation matrix is ​​calculated and directly assigned.

[0066] It should be noted that in the formula to It simply represents the individual terms in the formula and has no other meaning.

[0067] The final output vibration compensation vector is still a two-dimensional column vector: ; Let be the actual horizontal vibration displacement of the tower at time t. Let t be the actual vertical vibration displacement of the tower body at time t; the superscript T indicates that the expression in row vector form is converted into a standard two-dimensional column vector.

[0068] The multi-source image preprocessing unit 410 receives the back image data and front image data of the blade, and preprocesses the back image data and front image data of the blade to obtain a time-stamped sequence of back image frames and front image frames of the blade.

[0069] First, the original acquisition timestamps of the front and back images of the blades, as well as the acquisition timestamp of the front image of the tower, are extracted. Using the timestamp of the front image of the tower as a reference, the front and back images of the blades are time-stamp aligned, and asynchronous image frames with time deviations exceeding the threshold are removed. This ensures that the final generated sequence of front and back image frames of the blades is completely synchronized with the time series related to tower vibration compensation, providing a timing basis for subsequent decoupling calculations.

[0070] Preferably, it also includes some basic preprocessing operations, such as image denoising, adaptive contrast enhancement, and image format unification and normalization.

[0071] The decoupling calculation unit 430 embeds the vibration compensation vector into the feature points for decoupling calculation. Any feature point parameter (such as the original coordinates) in the original feature dataset of the blade essentially contains two independent components: ;in, It is directly related to the vibration compensation vector.

[0072] Based on the calibration rules of target feature points, a mapping relationship between the vibration compensation vector and the blade feature interference is established. Taking feature point coordinates as an example, the tower vibration compensation vector is calculated by subtracting the pixel scale calibration coefficient k (pre-calibrated, based on the ratio of the actual distance between adjacent target feature points to the pixel distance) from the actual distance of the target feature points. Interference values ​​converted to leaf feature dimensions: ; Numerical calculations are performed directly on the original blade feature data, eliminating tower vibration interference, without touching the original blade images throughout the process: Core calculation formula (taking the coordinates of blade feature points as an example): ; In the formula, Net characteristic components; These are the original feature components; The actual distance to the target feature point is denoted by the pixel scale calibration coefficient.

[0073] This process involves no image pixel modification, image displacement compensation, or image enhancement; it only performs calculations on the extracted feature data.

[0074] Based on the net characteristics, the state parameters of the blade are recalculated to obtain the net state parameters.

[0075] Specifically, feature decoupling is further subdivided into the following three independent mathematical models: Decoupling model for spatial geometric features (e.g., coordinates, distance, contour): This section follows the same logic. Since coordinates have explicit spatial vector properties, displacement compensation is performed directly using vector subtraction.

[0076] Decoupling model for surface attribute features (e.g., texture gradient, grayscale, area): Texture gradients (such as crack features) are themselves the result of pixel value differences and are unaffected by overall image translation. However, tower vibrations can cause the absolute coordinates of this texture feature to shift within the image. Therefore, for attribute-type features (let's say...), Instead of changing its calculated feature value, it decouples and redirects the spatial index coordinates it is bound to, that is: ; In the above formula, To decouple the net surface attribute eigenvalues ​​after redirection, the attribute eigenvalues ​​themselves do not undergo addition or subtraction. Instead, their positions in the digital space are reverse-translated according to the vibration compensation vector, so that they return to the true physical coordinates of the blade. These are the original feature values ​​of surface attributes extracted from the original image (e.g., original texture gradient abrupt changes, grayscale distribution values, crack area, etc. in a certain region), where the original image is in Abnormal texture feature values ​​extracted at coordinates In the decoupled net feature dataset, its spatial location should be remapped to This ensures the accurate positioning of surface anomalies such as cracks in the blade's physical coordinate system.

[0077] Decoupling model for time-series frequency domain features (e.g., vibration frequency, rotational speed fluctuations): Frequency characteristics are derived by analyzing time-series signals. For frequency and rotational speed, frame-by-frame decoupling must be performed in the time-domain spatial sequence before frequency domain transformation; direct subtraction in the frequency domain is strictly prohibited.

[0078] Suppose the original feature point is at The time series set of trajectories within a time period is .

[0079] The first step is to calculate the net time series trajectory set: ; The second step is to obtain the net vibration frequency using the Discrete Fourier Transform (DFT). : ; In the formula, The sampling time period selected for frequency domain analysis (or the total duration / total number of sampling frame sequences). For the original feature points in The collection of trajectory time-series signals within a time period (the original motion trajectory sequence contaminated by tower vibration); This is the set of net time-series trajectory signals obtained after frame-by-frame removal of tower displacement interference; The net autonomous vibration frequency of the blade (i.e., the structural response frequency of the blade itself) is extracted after the discrete Fourier transform (DFT) and the common-mode interference of the tower is removed. For frequency variables during frequency domain transformation; The imaginary unit is used to construct the complex field of the Fourier transform; The core complex exponential basis functions of the Discrete Fourier Transform are used to transform time-domain signals. Map and resolve to the frequency domain space.

[0080] Among them, Written at Below means "in this equation, "These are the variables we are changing and searching for." This is used to give the formula an instruction: please iterate through all possible frequencies. .

[0081] By using this mathematical model that first decouples in the time domain and then transforms in the frequency domain, the interference of tower harmonics on the analysis of the autonomous vibration frequency of the blades can be completely eliminated.

[0082] The anomaly identification unit 440 performs anomaly identification based on net features and net state parameters, specifically including: Identification of blade structural integrity anomalies (cracks, edge defects, localized deformation) A baseline feature library of blades of the same type under normal conditions is pre-established, including the contour feature values, relative distances / angles of feature points, and baseline values ​​of surface texture gradient of normal blades (the calibration of this baseline library has been completed by physical scale calibration based on the actual distance of the target and the pixel scale calibration coefficient). The decoupled net feature data is compared with the benchmark feature library dimension by dimension: Extract the gradient change value of the net texture features on the leaf surface. If the abrupt change value of the texture gradient in a certain area exceeds twice the baseline threshold, and the relative distance of the net feature points in that area deviates from the baseline value by more than 3%, and the abrupt change is detected in 3 consecutive frames, it is determined to be an abnormal leaf crack. By comparing the overlap between the net contour features of the leaf and the baseline contour, the pixel-scale contour loss length is converted into the physical scale through the actual distance calibration coefficient of the target feature points. If the local loss length is >5mm and has not been recovered for 3 consecutive frames, it is judged as an edge defect abnormality. The local deformation of the blade is calculated based on the net feature point coordinates. If the deformation is greater than 0.5 mm / ㎡ and persists for 10 consecutive timestamp frames, it is judged as a local deformation anomaly.

[0083] Identification of abnormal blade motion (excessive swing angle, abnormal torsion angle, uneven rotation speed) A motion parameter threshold library is established based on blade design parameters, including the net swing angle threshold (e.g., ±15°), net torsion angle threshold (e.g., ±8°), and net speed fluctuation threshold (e.g., ±3r / min) for normal operation. Real-time parsing of the decoupled net state parameters and comparison with the threshold library: If the net blade swing angle exceeds the design threshold for 10 consecutive frames, and the swing angle change rate is >2° / s, it is judged as an abnormal swing angle. Extract the relative angle of net feature points of different cross sections of the blade. If the angle deviates from the reference torsion angle by more than 8° and lasts for more than 5 frames, it is determined to be a torsion anomaly. The blade net rotational speed is calculated based on the inter-frame displacement of net feature points. If the difference in rotational speed between adjacent frames is greater than 5 r / min, and the rotational speed fluctuation is greater than 3 r / min for 20 consecutive frames, it is judged as an abnormality of uneven rotational speed.

[0084] Identification of abnormalities at blade connection points (loose blade roots, loose bolts) The reference coordinates of fixed feature points in the leaf root region are calibrated. Three feature points with unique identification at the leaf root are selected, and their net feature coordinates and relative distances under normal conditions are recorded. Real-time monitoring of the stability of net feature points in the leaf root region: Calculate the inter-frame fluctuation value of the relative distance between fixed feature points at the leaf root. After converting it to the physical scale through the target calibration coefficient, if the fluctuation value is >0.3mm and there is no convergence for 30 consecutive frames, it is judged as abnormal leaf root loosening. If a pre-calibrated leaf root bolt position net feature point disappears and is not recovered for 5 consecutive frames, and there are no structural defects in the area, it is determined to be an abnormal bolt detachment.

[0085] In one possible embodiment, when the decoupling calculation unit 430 performs decoupling calculation, it verifies the net features and net state parameters obtained from the front image frame sequence of the blade using the net features and net state parameters obtained from the back image frame sequence of the blade.

[0086] For example, from the front-side net feature dataset, select 3-5 fixed, uniquely identifiable net feature points (e.g., 2 feature points at the leaf root and 1 feature point in the middle of the leaf). Based on the pixel-physical scale coefficients pre-calibrated by the target, convert the relative distance and angle of these net feature points to physical scale (e.g., the net relative distance between two feature points at the leaf root is 50.0 mm, and the included angle is 90.0°) as a verification benchmark. From the back-side net feature dataset, select 3-5 net feature points that completely correspond to the front-side features. Similarly, convert them to relative distance and angle at the physical scale using the target calibration coefficients, and compare them with the front-side benchmark. If the relative distance deviation of the corresponding feature points on the back-side is ≤ ±0.3 mm and the relative angle deviation is ≤ ±0.5°, then the set of net feature points on the back-side is considered valid. If the deviation exceeds the threshold, it is considered abnormal data, and interpolation calibration is performed using the front-side benchmark data combined with valid data from adjacent frames on the back-side. If the threshold is exceeded for 3 consecutive frames, a re-decoupling calculation is triggered (verifying the validity of the vibration compensation vector and recalibrating the decoupling coefficients based on the relative position benchmark of the target feature points).

[0087] Assuming the net relative distance between the feature point in the middle of the leaf and the feature point at the leaf root on the front side is 1200.0 mm, and the net relative distance between the corresponding two points on the back side is 1200.2 mm, with a deviation of 0.2 mm (≤0.3 mm), the net feature point on the back side is determined to be valid. If the deviation on the back side is 0.6 mm, exceeding the threshold, then the 1200.0 mm on the front side is used as the reference, and combined with the 1199.9 mm and 1200.1 mm data of the adjacent frames on the back side, the interpolation calibration is performed to 1200.0 mm to ensure that the net feature data on the back side is consistent with that on the front side.

[0088] For example, using the blade's frontal net state parameters as a benchmark, the net swing angle, net torsion angle, and net rotational speed of each frame are extracted in real time (e.g., a net swing angle of 8.5°, a net torsion angle of 3.2°, and a net rotational speed of 18 r / min in a certain frontal frame). The net swing angle, net torsion angle, and net rotational speed of the same timestamp frame on the back side are extracted and compared with the frontal benchmark parameters. If the net swing angle deviation is ≤ ±0.3°, the net torsion angle deviation is ≤ ±0.2°, and the net rotational speed deviation is ≤ ±0.5 r / min, the back-side net state parameters are considered valid. If the deviation exceeds the threshold, such as a back-side net swing angle of 9.0° (deviation of 0.5°), the back-side net state parameters of that frame are considered abnormal. The back-side data is corrected using the frontal benchmark parameters, and the deviation value is recorded. If the deviation exceeds the threshold for 5 consecutive frames, the timestamp synchronization of the vibration compensation vector is verified (based on the time-series tracking data of target feature points, ensuring that the frontal and back data are consistent with the compensation vector timestamp).

[0089] If the net torsion angle of a certain frame on the front is 4.0° and the net torsion angle of the same frame on the back is 4.1°, with a deviation of 0.1° (≤0.2°), it is considered valid; if the net torsion angle on the back is 4.3°, with a deviation of 0.3°, then the back data is corrected to 4.0° and the frame is marked as "needs verification", and further verification is performed by combining data from multiple frames.

[0090] In summary, using a wide-area camera installed at a low position on the tower (a relatively low and stable location on the tower) as the absolute reference, the extrinsic parameter transformation matrix of the high-position camera is dynamically updated. Specifically, a dynamic tolerance vector is defined. .set up At any given time, the net spatial coordinates calculated from the frontal image of the blade at the same physical monitoring point are: The net spatial coordinates calculated from the image on the back of the blade are: The deviation function is defined as follows: ; In the formula, This is a dynamic tolerance vector used to determine whether the spatial coordinates of feature points on the front and back sides are consistent, setting an upper limit for system tolerance. These represent the threshold components of the dynamic tolerance vector in the horizontal (X-axis) and vertical (Y-axis) directions, respectively. This is the net spatial coordinate vector calculated from the frontal image of the same physical monitoring point (such as a specific spot on the leaf) at time t. This is the net spatial coordinate vector calculated from the image on the back of the blade at the same physical monitoring point at time t. Let be the coordinate deviation function at time t, representing the magnitude of the spatial position difference of the same feature point after mapping from the front and back views; The initial homography matrix (static extrinsic parameter matrix) for front and back spatial mapping is calibrated during initialization and is used to project the back coordinates onto the front coordinate system when there is no complex deformation.

[0091] Decision logic: If N consecutive frames (e.g., N=5) satisfy... If the camera is affected by complex nonlinear torsion or structural micro-deformation, the original simple linear displacement compensation has failed, and the adaptive correction program is automatically triggered.

[0092] After the correction is triggered, extract the set of m high-confidence benchmark feature points captured by the camera (wide field of view) located in the low-position stable region of the tower. ,in, Represents a high-confidence benchmark feature point set The first high-confidence baseline feature point in the data. Represents a high-confidence benchmark feature point set The second high-confidence benchmark feature point, and so on, Represents a high-confidence benchmark feature point set The m-th high-confidence benchmark feature point in the dataset.

[0093] By constructing a local dynamic mapping matrix To replace the original static extrinsic matrix The least squares method is used for back-projection calibration, with the objective function being to minimize the reprojection error between the frontal reference point and the back reprojection point. ; In the above formula, m is the number of high-confidence baseline feature points extracted when adaptive correction is triggered; The index number of the reference feature point; The images were captured by a stable wide-area camera installed at a low position on the tower. Reference coordinate vectors of each reference feature point in the frontal image system; Captured by a high-position servo camera, and The corresponding number The original coordinate vectors of each reference feature point in the back-side image system; The local dynamic mapping matrix at time t, calculated in real time through reverse calibration, is used to replace the failed one. This eliminates nonlinear coupling errors.

[0094] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent scope of this application.

Claims

1. A non-contact identification system of blade vibration based on tower visual response, characterized in that, include: The first image data acquisition module (100) is configured to acquire image data of the back of the blade; The second image data acquisition module (200) is configured to acquire frontal image data of the tower body and frontal image data of the blades; The tower vibration compensation module (300) is configured to acquire frontal image data of the tower and determine the vibration compensation vector of the tower based on the frontal image data of the tower. as well as Anomaly identification module (400) is configured to acquire blade back image data, blade front image data and vibration compensation vector, and identify abnormal features on the blade surface based on the blade back image data, blade front image data and vibration compensation vector; The vibration compensation vector is used to eliminate tower vibration interference.

2. The tower visual response based non-contact identification system of blade vibration according to claim 1, characterized in that, The data acquisition source for the image of the back of the blade is located in the first height range of the tower body and is located on the back of the blade.

3. The non-contact blade vibration identification system based on tower visual response according to claim 2, characterized in that, The tower front image data and the blade front image data are acquired from the same source, which is located in the second height range of the tower and on the front of the blade.

4. The non-contact blade vibration identification system based on tower visual response according to claim 2 or 3, characterized in that, The acquisition source includes at least one image acquisition device; When the acquisition source includes multiple image acquisition devices, the multiple image acquisition devices are deployed in the same height range or different height ranges of the tower and work together in an array to synchronously acquire image data of the same monitoring area.

5. The non-contact blade vibration identification system based on tower visual response according to claim 3, characterized in that, The second altitude range is smaller than the first altitude range; The second height range is the height range where the bottom of the tower is located.

6. The non-contact blade vibration identification system based on tower visual response according to claim 1, characterized in that, The tower vibration compensation module (300) includes: The image preprocessing unit (310) is configured to receive the front image data of the tower body and preprocess the front image data of the tower body to obtain the front image frame sequence of the tower body. The target feature point localization and tracking unit (320) is configured to receive the tower front image frame sequence and extract the two-dimensional pixel coordinates of multiple predefined feature points on the target in each tower front image frame, establish the correspondence of the same feature point in the time series, so as to obtain the motion trajectory of each feature point in the image plane. The tower surface vibration displacement calculation unit (330) is configured to receive the motion trajectory of each feature point in the image plane and estimate the vibration parameters of the tower at the target position at each moment based on the motion trajectory. as well as The vibration compensation vector generation unit (340) is configured to receive the vibration parameters of the tower at the target position at each moment and input the vibration parameters into the tower vibration transmission model. The tower vibration transmission model calculates and outputs the high-level vibration compensation vector of the tower observed by the acquisition source based on the front image data of the tower.

7. The non-contact blade vibration identification system based on tower visual response according to claim 6, characterized in that, The formula for calculating the vibration compensation vector is: ； In the formula, Let be the vibration compensation vector at time t. Let be the actual horizontal vibration displacement of the tower at time t. Let t be the actual vertical vibration displacement of the tower body at time t; the superscript T indicates that the expression in row vector form is converted into a standard two-dimensional column vector.

8. The non-contact blade vibration identification system based on tower visual response according to claim 7, characterized in that, The anomaly detection module (400) includes: The multi-source image preprocessing unit (410) is configured to receive blade back image data and blade front image data, and preprocess the blade back image data and blade front image data to obtain a time-stamped blade back image frame sequence and blade front image frame sequence. The leaf feature extraction unit (420) is configured to receive the leaf back image frame sequence and the leaf front image frame sequence, extract features from the leaf back image frame sequence and the leaf front image frame sequence, and integrate the features from different acquisition sources at the same time into the original leaf feature dataset; The decoupling calculation unit (430) is configured to receive the original feature dataset of the blade and embed the vibration compensation vector into the feature points for decoupling calculation, so that the calculated net features and net state parameters are decoupled from the tower vibration displacement. as well as An anomaly feature identification unit (440) is configured to receive net features and net state parameters, and to perform anomaly identification based on the net features and net state parameters.

9. The non-contact blade vibration identification system based on tower visual response according to claim 8, characterized in that, The calculation formulas for decoupling include: ； In the formula, Net characteristic components; These are the original feature components; The actual distance to the target feature point is denoted by the pixel scale calibration coefficient.

10. The non-contact blade vibration identification system based on tower visual response according to claim 8, characterized in that, When the decoupling calculation unit (430) performs decoupling calculation, it uses the net features and net state parameters obtained from the front image frame sequence of the blade to verify the net features and net state parameters obtained from the back image frame sequence of the blade.

Images