Blade trajectory calculation unit stress balance algorithm based on big data machine learning

Abstract

The invention relates to the technical field of wind power generation operation and maintenance, and discloses a blade trajectory calculation unit stress balance algorithm based on big data machine learning, and the algorithm constructs a dynamic visual reference system attached to a cabin, and solves the real-time pose parameters of a camera relative to a hub to eliminate vibration interference. Under the reference system, leaf tip space relative coordinates are extracted and decomposed into waving and shimmy deformation, and meanwhile, normalized torsion characteristic quantity is constructed; a standard phase grid is further established for phase domain resampling, and a multi-blade differential topological matrix for filtering environment common-mode interference is constructed; and finally, mass imbalance, pneumatic imbalance and structural rigidity damage discrimination indexes are calculated based on the matrix, and inversion diagnosis is carried out on the stress state of the blade. According to the method, through visual pose decoupling and multi-dimensional differential topology analysis, the influence of tower shaking and environmental wind speed fluctuation on measurement is effectively overcome, and accurate positioning and type identification of various unbalanced faults of the fan blade are realized.

Description

Technical Field

[0001] This invention relates to the field of wind power operation and maintenance technology, specifically to a force balance algorithm for wind turbine unit based on big data machine learning for blade trajectory estimation. Background Technology

[0002] As the capacity of wind turbine generators continues to increase, blade sizes are gradually increasing, and their flexibility is becoming increasingly prominent. During operation, blades are subjected to complex alternating loads, making them prone to aerodynamic imbalances, mass imbalances, and structural stiffness damage. This not only affects the generator's power generation efficiency but can also lead to catastrophic accidents such as tower collapse or blade breakage in severe cases. Therefore, real-time monitoring of blade trajectory and stress balance is of significant engineering importance.

[0003] Traditional blade monitoring methods primarily rely on contact sensors such as strain gauges and fiber Bragg gratings. These sensors need to be embedded or affixed to the blade surface, making them prone to detachment or failure in harsh outdoor environments over long periods. They are also susceptible to lightning strikes, resulting in high maintenance costs and difficulty in meeting monitoring needs throughout the entire blade lifecycle. In recent years, non-contact monitoring technology based on machine vision has attracted considerable attention due to its advantages such as convenient installation and rich information content. This technology typically uses cameras installed in the nacelle or at the base of the tower to acquire images of blade movement and then uses image processing algorithms to extract blade displacement information.

[0004] However, existing visual monitoring technologies still face many challenges in practical applications. When cameras are installed on top of the nacelle, the tower experiences irregular pitch and lateral movement under wind loads, causing the camera's coordinate system to sway. This camera motion is superimposed on the blade images. Without an effective dynamic reference frame construction and pose decoupling mechanism, the directly extracted image displacement will contain a large number of spurious displacement components caused by tower vibration, severely affecting measurement accuracy. Simultaneously, wind turbines typically operate under variable speed conditions, with blade rotational speeds fluctuating in real time with wind speed. This results in non-stationary time-series data with inconsistent periods, making direct lateral comparisons through time-domain or frequency-domain analysis difficult.

[0005] Furthermore, existing technologies primarily focus on single-dimensional waving direction displacement monitoring, lacking simultaneous extraction and correlation analysis of waving direction deformation and torsional deformation. In reality, mass imbalance, aerodynamic imbalance, and structural stiffness damage exhibit significant coupling and differences in their manifestations across multiple physical dimensions. Relying solely on single-dimensional displacement amplitude changes is insufficient for effectively distinguishing and accurately tracing the causes of these fault types. Existing monitoring schemes often struggle to eliminate common-mode interference from environmental wind speed variations, leading to false alarms or missed alarms under complex weather conditions, and failing to accurately assess the unit's true stress balance state. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a blade trajectory estimation algorithm for turbine force balance based on big data machine learning. It aims to solve the technical problems of measurement distortion caused by tower vibration and difficulty in distinguishing multi-dimensional blade fault types in existing non-contact monitoring technologies.

[0007] The first aspect of this invention provides a blade trajectory estimation algorithm for wind turbine force balance based on big data machine learning, comprising: acquiring a sequence of blade images during wind turbine operation, constructing a dynamic visual reference system attached to the rigid structure of the nacelle, and solving for the real-time pose parameters of the camera relative to the hub center. This process defines a hub reference coordinate system with its origin at the geometric center of the wind turbine hub, identifies the rigid geometric features of the hub region in the image and extracts feature points, establishes the correspondence between the pixel coordinates of the feature points in the image pixel coordinate system and their spatial coordinates in the hub reference coordinate system; based on a pinhole camera imaging model, a system of equations concerning the rotation matrix and translation vector is constructed using a perspective N-point localization algorithm, and the rotation matrix and translation vector characterizing the real-time rigid body motion of the hub reference coordinate system relative to the camera coordinate system are solved by minimizing the reprojection error.

[0008] The interference of camera motion on imaging is eliminated by utilizing the real-time pose parameters, and the spatial relative coordinates of the blade tip are extracted under the dynamic visual reference system. Specifically, the blade tip feature points in the image are identified and their pixel coordinates are obtained. The orthogonality of the rotation matrix and the translation vector are used to construct a reverse projection ray from the camera coordinate system to the hub reference coordinate system. A geometric constraint condition is introduced that the distance from the blade tip to the hub center is equal to the rotor rotation radius. The intersection point of the reverse projection ray and the sphere centered at the hub center is obtained. The component of this spatial relative coordinate along the rotor main axis is taken as the flapping direction deformation. The difference between the actual azimuth angle and the rigid azimuth angle of the blade tip in the rotor rotation plane is calculated, and this difference is converted into the tangential displacement in the rotation plane as the flapping direction deformation. Simultaneously, local sub-images containing the leaf tip are extracted and the closed edge contour of the leaf tip is obtained. The secondary axis direction perpendicular to the blade's spanwise extension direction is determined through principal component analysis. The maximum projection span of the leaf tip contour in this secondary axis direction is extracted as the pixel projection chord length. Combined with the camera focal length, the physical chord length of the blade, and the leaf tip depth value, perspective distortion correction is performed on the pixel projection chord length to construct a normalized torsional feature quantity.

[0009] The real-time azimuth angle of the wind turbine is acquired, and a standard phase grid is established. Phase domain resampling is performed on the flapping direction deformation, oscillation direction deformation, and normalized torsional characteristic. A non-uniform sampling point set is constructed using the real-time azimuth angle of the wind turbine as the independent variable. The data is mapped onto the standard phase grid through periodic extension processing and periodic cubic spline interpolation algorithms to generate standardized feature functions. The characteristic mean vector of the three blades at the same standard phase point is calculated, and the difference between the standardized feature function of each blade and this characteristic mean vector is calculated. These are combined to form a multi-blade differential topology matrix that filters out environmental common-mode interference.

[0010] The stress balance state of the blades is inverted and diagnosed based on the multi-blade differential topology matrix. The energy amplitude of the differential eigenvectors of the blades in the flapping direction is calculated. Multiple sets of energy amplitudes at different rotor speeds are statistically analyzed, and the Pearson correlation coefficient between these amplitudes and the square of the rotor speed is calculated. When this coefficient exceeds a preset mass correlation threshold and the torsional dimension energy is below a preset threshold, a mass imbalance fault is determined. The normalized cross-correlation coefficient between the differential eigenvectors in the flapping and torsional directions is calculated. When the absolute value of this coefficient exceeds a preset aerodynamic coupling threshold and the differential eigenvector amplitude is positively correlated with wind speed or pitch angle, an aerodynamic imbalance fault is determined. A phase plane trajectory is constructed with the differential eigenvectors in the flapping and torsional directions as the ordinate. The area of ​​the closed region enclosed by the trajectory is calculated using the Discrete Green's formula as the hysteresis loop area index. When the deviation ratio of the hysteresis loop area index of a blade relative to the mean exceeds a preset damage threshold, the blade is determined to have structural stiffness damage.

[0011] This invention provides a force balance algorithm for turbine unit calculation based on blade trajectory estimation using big data machine learning. It has the following beneficial effects:

[0012] 1. This invention constructs a dynamic visual reference system attached to the rigid structure of the nacelle and uses a perspective N-point positioning algorithm to calculate the real-time pose of the camera relative to the hub, thus achieving motion decoupling between the measurement coordinate system and the camera itself. This mechanism effectively eliminates camera sway errors caused by wind turbine tower vibration, allowing the system to obtain high-precision absolute spatial coordinates of the blade tip in dynamic operating environments without the need for expensive hardware image stabilization equipment or laser trackers, ensuring the reliability of monitoring data under complex load conditions.

[0013] 2. This invention constructs a multi-dimensional force inversion model based on physical mechanisms, which can separate the deformation features of three dimensions—flailing, oscillation, and torsion—from a single visual image, and establish fault criteria using indicators such as hysteresis loop area and aerodynamic coupling degree. This method overcomes the limitation of traditional vibration monitoring, which can only report the total vibration amplitude, and can accurately distinguish between three completely different fault types: mass imbalance, aerodynamic imbalance, and structural stiffness damage, significantly improving the resolution and specificity of unit fault diagnosis.

[0014] 3. This invention employs a non-contact image acquisition scheme to replace traditional blade root strain gauges or accelerometers, avoiding the problems of contact sensors being susceptible to lightning strikes, detachment, and difficult maintenance on the blade surface. Simultaneously, by resampling in the phase domain and constructing a multi-blade differential topology matrix, the algorithm utilizes the consistency characteristics of the three blades to filter out environmental common-mode interference caused by wind speed fluctuations and turbulence, reducing the algorithm's dependence on meteorological data. This achieves highly robust online force balance monitoring at a lower hardware cost. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system architecture and overall process of the present invention;

[0016] Figure 2 This is a flowchart illustrating the construction and coordinate decoupling of the dynamic visual reference system of the present invention.

[0017] Figure 3 This is a flowchart of the leaf tip multidimensional feature extraction process of the present invention;

[0018] Figure 4 This is a flowchart of the phase domain resampling and differential feature matrix construction process of the present invention;

[0019] Figure 5 This is a flowchart of the force balance inversion diagnosis based on physical mechanisms of the present invention.

[0020] Among them, 101 is the image acquisition device; 102 is the data transmission interface; 103 is the data processing terminal; 201 is the dynamic benchmark construction module; 202 is the multi-dimensional feature extraction module; 203 is the phase normalization module; and 204 is the force inversion analysis module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 , Figure 1 This is a schematic diagram of the system architecture and process of a blade trajectory estimation and turbine force balance algorithm based on big data machine learning according to an embodiment of the present invention. The present invention provides a method and system for blade force inversion based on hub visual reference dynamic decoupling and phase difference topology analysis. The system includes an image acquisition device 101, a data transmission interface 102, and a data processing terminal 103.

[0023] The image acquisition device 101 is fixedly installed on the outer wall of the wind turbine tower or on the ground base. Its optical lens has a field of view covering the central area of ​​the wind turbine hub and the complete swept plane of the rotating blades. The image acquisition device 101 is configured to continuously acquire sequential image data during the operation of the wind turbine at a set sampling frequency, and send the data to the data processing terminal 103 through the data transmission interface 102.

[0024] The data processing terminal 103 is equipped with a processor and a memory. The memory stores computer program instructions. When the processor executes the instructions, it implements the following functional modules: dynamic benchmark construction module 201, multi-dimensional feature extraction module 202, phase normalization module 203, and force inversion analysis module 204.

[0025] The dynamic reference construction module 201 receives sequential image data from the image acquisition device 101. In each frame of the image, the dynamic reference construction module 201 identifies the rigid geometric features of the wind turbine hub, including the hub center point, the hub cover edge contour, or the blade root flange bolt ring. Based on the identified rigid geometric features and a preset hub geometric model, the dynamic reference construction module 201 uses a perspective transformation algorithm to calculate the pose transformation matrix of the image acquisition device 101 relative to the wind turbine hub at each moment. This pose transformation matrix includes a rotation matrix and a translation vector, used to characterize the rigid body motion components of the image acquisition device 101 relative to the wind turbine hub caused by the vibration of the tower.

[0026] The multi-dimensional feature extraction module 202 is connected to the dynamic reference construction module 201. Using the pose transformation matrix, the multi-dimensional feature extraction module 202 inversely maps the blade tip pixel coordinates identified in the image to a hub reference coordinate system with the hub center as the origin, obtaining the blade tip spatial relative coordinates after removing the tower vibration component. The multi-dimensional feature extraction module 202 decomposes this spatial relative coordinate into a flapping direction displacement component perpendicular to the wind turbine rotation plane and a swaying direction displacement component parallel to the wind turbine rotation plane. Simultaneously, the multi-dimensional feature extraction module 202 extracts the edge contour of the blade tip region in the image, calculates the projection chord length of the blade tip contour on the imaging plane, and performs geometric correction on the projection chord length using the pose transformation matrix, generating torsional feature quantities characterizing the blade's torsional deformation state.

[0027] The phase normalization module 203 receives the time-domain feature data output by the multi-dimensional feature extraction module 202. The phase normalization module 203 calculates the real-time azimuth angle of the wind turbine based on the blade root position or hub angle encoding information. Using the azimuth angle as the independent variable, the phase normalization module 203 performs spatial interpolation resampling on the flapping direction displacement component, the oscillation direction displacement component, and the torsional feature quantity, mapping the operating data of the three blades in different time dimensions to a unified azimuth phase domain, generating a standardized phase domain feature curve.

[0028] The force inversion analysis module 204 is connected to the phase normalization module 203. The force inversion analysis module 204 calculates the characteristic differences of the three blades at the same phase angle and constructs a differential topology matrix. The force inversion analysis module 204 determines the type of force imbalance based on the correlation between each component in the differential topology matrix and the unit's operating parameters. These operating parameters include rotor speed and wind speed.

[0029] Specifically, when the force inversion analysis module 204 detects that the amplitude of the flare direction displacement component deviation of a blade is linearly correlated with the square of the rotor speed, and its torsional characteristic deviation is lower than a preset threshold, it determines that the blade has a mass imbalance fault. When the force inversion analysis module 204 detects that the flare direction displacement component deviation of a blade is positively correlated with the wind speed, and its torsional characteristic deviation fluctuates in phase with the flare direction displacement component deviation, it determines that the blade has an aerodynamic imbalance fault. When the force inversion analysis module 204 detects that the area of ​​the hysteresis loop formed by the flare direction displacement component deviation and the torsional characteristic deviation of a blade exceeds a preset threshold, it determines that the blade has structural stiffness damage.

[0030] Through the collaborative work of the aforementioned dynamic benchmark construction module 201, multi-dimensional feature extraction module 202, phase normalization module 203, and force inversion analysis module 204, the system achieves decoupled analysis and fault type identification of the stress state of wind turbine blades in an environment where no contact sensors are required and tower vibration interference exists.

[0031] See attached document Figure 2 , Figure 2 This is a schematic diagram of the dynamic visual reference system construction and coordinate decoupling process according to an embodiment of the present invention.

[0032] The dynamic benchmark construction module 201 performs the coordinate system construction and decoupling process, which includes steps S201 to S204.

[0033] In step S201, a reference coordinate system is established. The camera coordinate system is defined. Its origin The optical center is located at the image acquisition device 101, with the Z-axis pointing towards the fan along the optical axis. Define the image pixel coordinate system. Its origin Located in the top left corner of the image, with coordinate variables as follows: Define the hub reference coordinate system. Its origin Located at the geometric center of the wind turbine hub, this is the connection node between the rotating components of the wind turbine and the stationary components of the nacelle. In the hub reference coordinate system... In the middle, the definition The shaft points outward from the nacelle along the axis of the wind turbine's main shaft. The axis points upwards in the vertical plane, and The axis is determined according to the right-hand rule. Hub reference coordinate system. It is attached to the rigid structure of the nacelle and moves with the overall movement of the tower and the nacelle.

[0034] In step S202, rigid geometric features are identified and feature points are extracted. The dynamic benchmark construction module 201 performs grayscale conversion and histogram equalization on the single-frame image acquired by the image acquisition device 101 to enhance the image edge contrast. Edge information is extracted using the Canny operator, and the circular or elliptical structural contour of the wind turbine hub area is located using the Hough Circle Transform or least squares ellipse fitting algorithm. The structural contour corresponds to the hub cover edge, fairing contour, or bolt ring at the blade root flange connection in physical space. The system pre-constructs the geometric model of the hub and selects... indivual( Feature points are used as control points, such as equidistant points on the flange ring or specific bolt positions. These control points are located in the hub reference coordinate system. The spatial coordinates in the equation are predetermined and fixed, denoted as . Feature pixels extracted from the image are established using template matching or azimuth-based corner sorting algorithms. With space control points A one-to-one correspondence between them.

[0035] In step S203, the real-time pose of the camera relative to the wheel hub is calculated. Based on the pinhole camera imaging model, the spatial coordinates... with pixel coordinates The following projection relationship must be satisfied:

[0036] ;

[0037] in, Scale factor; The camera intrinsic parameter matrix includes focal length, principal point coordinates, and tilt factor. This matrix is ​​obtained in advance using the Zhang Zhengyou calibration method or other camera calibration techniques. For a moment The rotation matrix characterizes the hub reference coordinate system. Relative to camera coordinate system The rotational posture; For a moment The translation vector characterizes the hub reference coordinate system. The origin is in the camera coordinate system The position within the range. The Efficient Perspective-n-Point (EPnP) algorithm is used to construct a framework for... and The system of linear equations is solved, and nonlinear optimization is performed using the Gauss-Newton method or the Levenberg-Marquardt algorithm to minimize the reprojection error. The solution obtained and The relative rigid body motion between the image acquisition device 101 and the wind turbine hub was quantified, including the pitch and sway components of the tower.

[0038] In step S204, the spatial relative coordinates of the leaf tip are reverse mapped and calculated. Leaf tip feature points are identified in the image, and their pixel coordinates are obtained. Using the pose parameters obtained in step S203 and Construct from the camera coordinate system to the hub reference coordinate system The inverse transformation. First, calculate the normalized direction vector in the camera coordinate system. :

[0039] ;

[0040] Utilizing the orthogonality property of rotation matrices ( ),Will Transform to hub reference coordinate system Downward direction vector :

[0041] ;

[0042] Calculate the camera's optical center in the hub reference coordinate system Position coordinates in :

[0043] ;

[0044] Blade tip in hub reference coordinate system Spatial position Located in Starting point, direction On the ray. Based on the geometric parameters of the wind turbine, a length constraint is introduced: the Euclidean distance from the blade tip to the hub center is equal to the known rotor radius. :

[0045] ;

[0046] Seeking answers regarding The quadratic equation in one variable is solved by taking its positive real solutions, and then the relative spatial coordinates of the leaf tip are calculated. :

[0047] ;

[0048] Should This refers to the position coordinates of the blade tip in the hub reference frame after eliminating the effects of tower vibration and camera shake, which characterizes the physical deformation state of the blade relative to the hub.

[0049] See attached document Figure 3 The process of multidimensional feature extraction module 202 performing multidimensional feature extraction of leaf tip includes steps S301 to S303.

[0050] In step S301, the relative spatial coordinates of the blade tip are orthogonally decomposed to obtain the deformation components of the flapping direction and the oscillation direction. The multidimensional feature extraction module 202 receives the relative spatial coordinates of the blade tip from the automatic reference construction module 201. Based on the hub reference coordinate system The definition, in which The shaft runs along the direction of the wind turbine's main shaft. The plane is the plane of rotation of the wind turbine. Multidimensional feature extraction module 202 extracts... Components deform as the direction of the swing This component characterizes the bending displacement of the blade along the main axis under aerodynamic thrust:

[0051] ;

[0052] Simultaneously, the multidimensional feature extraction module 202 calculates the actual azimuth angle of the blade tip in the plane of rotation. The calculation formula is: The multi-dimensional feature extraction module 202 obtains the rigid azimuth angle of the blade at a given moment when it theoretically has not undergone elastic deformation by periodically recognizing the features of the blade root image or receiving the azimuth angle signal from the SCADA system of the receiver unit. The multi-dimensional feature extraction module 202 calculates the deformation in the oscillation direction. This component characterizes the tangential displacement of the blade in the plane of rotation under the influence of gravity and centrifugal force:

[0053] ;

[0054] in, Let be the known radius of rotation of the wind turbine. Through the above decomposition, the three-dimensional coordinate information is decoupled into two independent physical quantities corresponding to the direction of the force.

[0055] In step S302, the projected geometric features of the leaf tip image region are extracted. The multidimensional feature extraction module 202 uses the pose parameters and camera imaging model calculated in step S203 to extract the relative spatial coordinates of the leaf tip. The image is reprojected back to the pixel coordinate system to determine the center position of the blade tip in the image. Using this center position as a reference, a region of interest (ROI) of a preset size is defined, and a local sub-image containing the blade tip airfoil is cropped from the original image. Binarization segmentation and Canny edge detection are performed on the local sub-image to obtain the closed edge contour of the blade tip. The multidimensional feature extraction module 202 performs principal component analysis (PCA) on this closed edge contour to determine the principal axis direction (corresponding to the blade spanwise extension direction) and the secondary axis direction (corresponding to the blade chord direction). The maximum projection span of the blade tip contour in the secondary axis direction is calculated and denoted as the pixel projection chord length. The pixel projection chord length It represents the projected width of the physical chord length of the leaf tip onto the camera's imaging plane.

[0056] In step S303, a normalized torsional feature quantity is constructed to characterize the blade's torsional deformation. This is due to the pixel projection chord length. The value depends not only on the twist angle of the blade, but also on the scaling effect of the distance (depth) between the blade tip and the camera, and the interference caused by the depth change needs to be eliminated. The multidimensional feature extraction module 202 first calculates the blade tip in the camera coordinate system. The depth value below Using the rotation matrix obtained in step S203 Translation vector The calculation is as follows:

[0057] ;

[0058] Pick The third component of the vector is used as the depth value. Subsequently, the multi-dimensional feature extraction module 202 combines the camera focal length... and the known physical chord length of the leaf tip Calculate the normalized torsional characteristic. :

[0059] ;

[0060] The normalized torsional characteristic This is a dimensionless physical quantity. Its numerical change reflects the change in the cosine of the central angle between the blade tip chord and the camera's optical axis. When the blade undergoes torsional deformation, this angle changes, leading to a change in the width of the projected cross-section, thereby causing... The numerical fluctuations. Therefore, The time-varying component directly maps to the aeroelastic torsion angle of the blade. This enables non-contact measurement of the blade's torsional modes. Finally, the multi-dimensional feature extraction module 202 outputs data including... A multidimensional state feature vector sequence.

[0061] See attached document Figure 4 The phase normalization module 203 performs data resampling and differential construction, including steps S401 to S403.

[0062] In step S401, the real-time azimuth angle of the wind turbine is calculated and a standard phase grid is established. The phase normalization module 203 obtains the real-time azimuth angle sequence of the wind turbine rotation. This azimuth angle The range of values ​​is The zero-phase baseline is set directly above the wind turbine (at the 12 o'clock position) or directly in front of the tower. Due to the fluctuation of natural wind speed, the wind turbine rotation speed is not constant, resulting in non-uniform spatial phase distribution of the time-interval sampling data acquired by the image acquisition device 101.

[0063] To this end, the phase normalization module 203 sets a standard phase grid. ,in , For the sampling resolution per lap (e.g.) .

[0064] In step S402, spatial interpolation resampling is performed to generate the phase domain characteristic curves of each blade. For the first... One leaf Phase normalization module 203 uses azimuth angle The independent variable is the multidimensional feature vector output in step S303. A non-uniform sampling point set is constructed with the data as the dependent variable. To ensure the continuity and smoothness of the interpolation curve at phases 0 and 2π, the phase normalization module 203 periodically extends the data point set, copying the first and last data segments by one period and splicing them to the ends of the original data. Using the Periodic Cubic Spline Interpolation algorithm, the above data is mapped to a standard phase grid. The standardized characteristic function that varies with the phase angle is obtained as follows:

[0065] ;

[0066] in, , and They represent the first Each blade rotates to the phase angle The flapping deformation, oscillation deformation, and normalized torsional characteristic at different positions are measured. This step eliminates the influence of speed fluctuations on the time-domain signal waveform and achieves alignment of different blades on a spatial phase reference.

[0067] In step S403, a multi-blade differential topology matrix is ​​constructed to eliminate common-mode interference. Since environmental factors such as wind shear, tower shadow effect, and gusts can have similar aerodynamic effects (common-mode interference) on all blades passing through the same spatial location, differential features need to be calculated to separate the independent state characteristics reflecting the structure and aerodynamic properties of a single blade. Phase normalization module 203 calculates the phase angle of the three blades at the same phase angle. The eigenmean vector under :

[0068] ;

[0069] Subsequently, the calculation of the first The difference feature vector of each leaf relative to the mean :

[0070] ;

[0071] in, , and These represent the relative deviations of the blade in the flapping, oscillating, and torsional dimensions, respectively. This is achieved by calculating all phase points. The phase normalization module 203 generates a multi-blade differential topology matrix from the differential vectors on the blade. This matrix filters out the environmental common-mode components and retains the individual differences in the blades, serving as input data for subsequent stress inversion analysis.

[0072] See attached document Figure 5 The process of force inversion analysis module 204 performing force state identification includes steps S501 to S504.

[0073] In step S501, the mass imbalance discrimination index is calculated. The force inversion analysis module 204 obtains the differential characteristic data of each blade from the phase normalization module 203. The mass imbalance fault manifests as a difference in centrifugal force, which generates a periodic oscillation component in the rotor's rotation plane. The force inversion analysis module 204 calculates the... The differential energy amplitude of each blade in the oscillation direction :

[0074] ;

[0075] The force inversion analysis module 204 constructs multiple sets of data within a set sampling period. The dataset, in which This corresponds to the wind turbine rotation speed. Calculation. square of rotational speed The Pearson correlation coefficient between them is denoted as the mass imbalance index. The force inversion analysis module 204 will... Quality correlation threshold Compare them.

[0076] The threshold The statistical distribution of data based on the unit's historical health status is set, for example, by taking the average correlation coefficient of historical normal operating data. If Furthermore, the torsional differential characteristic integral of the blade If the noise level is below the preset noise threshold, then the first... One blade exhibits a mass imbalance fault. This logic leverages the strong linear correlation between mass centrifugal force and the square of rotational speed, as well as the physical characteristic that mass load has a weak effect on the blade's torsional mode.

[0077] In step 5502, the aerodynamic imbalance criterion is calculated. Aerodynamic imbalance faults are usually caused by pitch angle errors or airfoil aerodynamic performance degradation, resulting in changes in the aerodynamic lift and aerodynamic torque on the blades. The force inversion analysis module 204 utilizes the coupling characteristics of flapping deformation and torsional deformation under aerodynamic loads to calculate the normalized cross-correlation coefficient between the flapping difference vector and the torsional difference vector, denoted as the aerodynamic coupling index. ;

[0078] ;

[0079] The force inversion analysis module 204 monitors the correlation between this indicator and the unit's operating conditions. If... Greater than the preset aerodynamic coupling threshold Furthermore, if the amplitude of the differential characteristic is positively correlated with the wind speed or pitch angle, then the first... One blade exhibits an aerodynamic imbalance fault. This step distinguishes between aerodynamic load anomalies and simple mass load anomalies by detecting the synchronicity (in-phase or out-of-phase) of the responses in the bending and torsional degrees of freedom.

[0080] In step S503, the structural stiffness damage discrimination index is calculated. The force inversion analysis module 204 constructs a system based on... For the horizontal axis, The phase plane trajectory is the vertical axis. When the blade structural stiffness decreases, a phase difference will occur between the bending response and the torsional response, causing the phase plane trajectory to form an open hysteresis loop. The force inversion analysis module 204 uses the discrete form of Green's formula to calculate the area of ​​the closed region enclosed by this trajectory, denoted as the hysteresis loop area index. :

[0081] ;

[0082] Among them, setting To form a closed loop. The force inversion analysis module 204 calculates the mean value of the hysteresis loop area index of the three blades. And calculate the first The relative deviation ratio of each leaf .like Exceeding the preset damage threshold (For example, 0.2, i.e., deviation exceeding 20%), then determine the first The blades have structural stiffness damage.

[0083] In step S504, the comprehensive diagnostic results are output. The force inversion analysis module 204 summarizes the above indicators. The system generates a blade health status report. When the system detects that any indicator exceeds the limit, it generates an alarm signal containing the fault type (quality / aerodynamic / structural), fault severity, and faulty blade number, and sends it to the remote monitoring center via the data transmission interface.

Claims

1. A blade trajectory estimation and turbine unit force balance algorithm based on big data machine learning, characterized in that, Includes the following steps: S1. Acquire the blade sequence images during the operation of the wind turbine, construct a dynamic visual reference system attached to the rigid structure of the nacelle, and calculate the real-time pose parameters of the camera relative to the hub center. S2. Using the real-time pose parameters to eliminate the interference of camera motion on imaging, under the dynamic visual reference system, extract the spatial relative coordinates of the leaf tip and decompose them into the deformation amount in the waving direction and the deformation amount in the swinging direction; at the same time, extract the projection geometric features of the leaf tip image area and construct the normalized torsional feature amount. S3. Obtain the real-time azimuth angle of the wind turbine, establish a standard phase grid, resample the deformation in the flapping direction, the deformation in the oscillation direction, and the normalized torsional characteristic in the phase domain, and construct a multi-blade differential topology matrix to eliminate environmental common-mode interference. S4. Based on the multi-blade differential topology matrix, calculate the mass imbalance discrimination index, aerodynamic imbalance discrimination index, and structural stiffness damage discrimination index to perform inversion diagnosis of the force balance state of the blade.

2. The blade trajectory estimation and turbine force balance algorithm based on big data machine learning according to claim 1, characterized in that, The specific process of constructing the dynamic visual reference system and calculating the real-time pose parameters in step S1 includes: Define a hub reference coordinate system with its origin at the geometric center of the wind turbine hub; identify the rigid geometric features of the hub region in the image and extract feature points, and establish the correspondence between the pixel coordinates of the feature points in the image pixel coordinate system and their spatial coordinates in the hub reference coordinate system; Based on the pinhole camera imaging model, a system of equations about the rotation matrix and translation vector is constructed using the perspective N-point positioning algorithm, and the rotation matrix and translation vector are solved by minimizing the reprojection error; the rotation matrix and translation vector represent the real-time rigid body motion of the hub reference coordinate system relative to the camera coordinate system.

3. The blade trajectory estimation and unit force balance algorithm based on big data machine learning according to claim 2, characterized in that, The extraction of the spatial relative coordinates of the leaf tip in step S2 specifically includes: Identify leaf tip feature points in the image and obtain their pixel coordinates; utilize the orthogonality of the rotation matrix and the translation vector to construct an inverse projection ray from the camera coordinate system to the hub reference coordinate system; By introducing the geometric constraint that the distance from the blade tip to the center of the hub is equal to the rotation radius of the wind turbine, the intersection point of the reverse projection ray and the sphere centered at the hub center is solved to obtain the spatial relative coordinates of the blade tip in the hub reference coordinate system.

4. The blade trajectory estimation and unit force balance algorithm based on big data machine learning according to claim 1, characterized in that, The specific steps in step S2, which decompose the deformation in the swinging direction and the deformation in the oscillation direction, include: The component of the spatial relative coordinates of the blade tip along the main axis of the wind turbine is taken as the deformation in the flapping direction; Calculate the actual azimuth angle of the blade tip in the rotor rotation plane, obtain the rigid azimuth angle when the blade has not undergone elastic deformation, calculate the difference between the actual azimuth angle and the rigid azimuth angle, and convert this difference into the tangential displacement in the rotation plane as the deformation amount in the oscillation direction.

5. The blade trajectory estimation and unit force balance algorithm based on big data machine learning according to claim 2, characterized in that, The construction of the normalized torsional feature quantity in step S2 specifically includes: Extract a local sub-image containing the leaf tip and obtain the closed edge contour of the leaf tip. Perform principal component analysis on the closed edge contour to determine the secondary axis direction perpendicular to the blade span extension direction. Extract the maximum projection span of the leaf tip contour in the secondary axis direction as the pixel projection chord length. The depth value of the leaf tip in the camera coordinate system is calculated using the rotation matrix and translation vector. By combining the camera focal length and the physical chord length of the blade, the depth value is used to correct the perspective distortion of the pixel projection chord length, thereby obtaining a normalized torsional characteristic quantity that characterizes the change in the angle between the blade tip chord and the observation plane.

6. The blade trajectory estimation and unit force balance algorithm based on big data machine learning according to claim 1, characterized in that, The phase domain resampling in step S3 specifically includes: A standard phase grid containing several discrete phase points is set; a non-uniform sampling point set is constructed with the real-time azimuth angle of the wind turbine as the independent variable and the deformation in the flapping direction, the deformation in the oscillation direction, and the normalized torsional characteristic as the dependent variables. The non-uniform sampling point set is periodically extended, and the data is mapped onto the standard phase grid using a periodic cubic spline interpolation algorithm to generate a standardized characteristic function of each blade as the phase angle changes.

7. The blade trajectory estimation and unit force balance algorithm based on big data machine learning according to claim 1, characterized in that, The construction of the multi-leaf difference topology matrix in step S3 specifically includes: Calculate the characteristic mean vector of the three blades at the same standard phase point. The characteristic mean vector includes the mean components of three dimensions: flapping, oscillation, and torsion. The difference between the standardized feature function of each blade and the feature mean vector is calculated to obtain the difference feature vector of each blade in the complete rotation cycle. These vectors are then combined to form a multi-blade difference topology matrix that has filtered out common mode components.

8. The blade trajectory estimation and unit force balance algorithm based on big data machine learning according to claim 7, characterized in that, The calculation of the mass imbalance discrimination index in step S4 specifically includes: Calculate the energy amplitude of the differential eigenvector of the blade in the oscillation direction; The energy amplitudes at multiple different wind turbine speeds were statistically analyzed, and the Pearson correlation coefficient between the energy amplitudes and the square of the wind turbine speed was calculated as a mass imbalance discrimination index. If the mass imbalance discrimination index exceeds the preset mass correlation threshold, and the differential characteristic energy of the torsional dimension of the blade is lower than the preset threshold, then a mass imbalance fault is determined to exist.

9. The blade trajectory estimation and unit force balance algorithm based on big data machine learning according to claim 7, characterized in that, The calculation of the aerodynamic imbalance discrimination index in step S4 specifically includes: The normalized cross-correlation coefficient between the differential eigenvectors in the blade flapping direction and the differential eigenvectors in the torsional direction is calculated and used as an aerodynamic coupling index. If the absolute value of the aerodynamic coupling index exceeds the preset aerodynamic coupling threshold, and the amplitude of the differential feature is positively correlated with the wind speed or pitch angle, then an aerodynamic imbalance fault is determined to exist.

10. The blade trajectory estimation and unit force balance algorithm based on big data machine learning according to claim 7, characterized in that, The calculation of the structural stiffness damage discrimination index in step S4 specifically includes: Construct a phase plane trajectory with the differential eigenvector in the swing direction as the horizontal axis and the differential eigenvector in the torsional direction as the vertical axis; The area of ​​the closed region enclosed by the phase plane trajectory is calculated using the Discrete Green's Theorem and used as an index of the hysteresis loop area. If the deviation ratio of the hysteresis loop area index of a certain blade relative to the mean of all blades exceeds a preset damage threshold, then the blade is determined to have structural stiffness damage.

Images