A wind turbine blade edge trajectory abnormality fault diagnosis method and system

Abstract

This application relates to a fault diagnosis method for abnormal edge trajectory of wind turbine blades. It employs an improved wavelet packet multi-scale analysis of blade operating state images for filtering and noise reduction; it obtains blade trajectory edge features through invariant moment features to achieve translational, scaling, and rotational invariance recognition; and it utilizes a particle swarm optimization algorithm to obtain images with large edge trajectory deviations, calibrating abnormal states for fault diagnosis. This method solves the problems in existing technologies, such as the inability of edge detection operator thresholds to adapt, poor noise resistance, and difficulty in extracting abnormal blade edge trajectory features, leading to ineffective fault diagnosis. It improves fault diagnosis accuracy and, by filtering redundant data through the optimization algorithm, balances the accuracy and efficiency of intelligent recognition, providing reliable technical support for wind turbine blade fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of wind turbine abnormality diagnosis, and in particular to a method, system, computer equipment, and computer-readable storage medium for diagnosing abnormal blade edge trajectory faults in wind turbines. Background Technology

[0002] As a core component of energy conversion, the operating status of wind turbine blades directly affects the safety, stability, and power generation efficiency of the unit. Diagnosing abnormal conditions in blade edge trajectories is a crucial step in ensuring reliable unit operation. Currently, blade fault diagnosis largely relies on image edge detection technology, which extracts blade edge trajectory features to identify operational anomalies.

[0003] In existing technologies, commonly used edge detection operators include Roberts operator, Sobel operator, Prewitt operator, Laplace operator, and Log operator. Although they can extract image edges, they have significant drawbacks: First, the threshold cannot be adaptively adjusted according to the image, making them sensitive to noise and having poor noise resistance, which can easily lead to distortion of edge detection results due to noise interference. Second, the extracted edges are not fine enough, often exhibiting widening or discontinuity of boundaries, making it difficult to form complete trajectory lines and affecting the accurate identification of blade running status characteristics.

[0004] Although the Canny operator based on optimization algorithms has certain advantages in signal-to-noise ratio and detection accuracy, it is still difficult to effectively solve the problem of incomplete edge feature extraction caused by image noise and blurring in the complex operating environment of the blade. This makes it impossible to accurately capture the abnormal state features of the blade edge trajectory, which in turn makes it difficult to effectively diagnose abnormal faults, thus restricting the reliability and efficiency of wind turbine blade fault diagnosis.

[0005] Therefore, there is an urgent need for a technical solution that can efficiently process noise and accurately extract edge trajectory features to improve the accuracy of blade fault diagnosis. Summary of the Invention

[0006] This application provides a method, system, computer equipment, and computer-readable storage medium for diagnosing abnormal edge trajectory faults in wind turbine blades, in order to at least solve the problem of insufficient accuracy in extracting edge trajectory features in related technologies.

[0007] In a first aspect, embodiments of this application provide a method for diagnosing abnormal edge trajectories of wind turbine blades, the method comprising:

[0008] The blade image of the wind turbine is acquired, and the blade image is then subjected to multi-scale filtering using an improved wavelet packet to obtain the filtered image.

[0009] Edge trajectory extraction is performed on the filtered image to obtain the blade edge trajectory image, and edge invariant moment feature extraction is performed on the blade edge trajectory image to output multiple invariant moment feature sequences;

[0010] Initial abnormal samples are determined based on multiple invariant moment feature sequences. A particle swarm optimization model is constructed based on the invariant moment features of the initial abnormal samples. The particle swarm optimization model is used for iterative optimization to obtain key abnormal samples for blade fault diagnosis.

[0011] In some embodiments, multi-scale filtering of the leaf image using improved wavelet packets includes:

[0012] Images of the blade assembly are acquired from two orthogonal positions of the wind turbine to obtain blade sequence data at the two positions;

[0013] After setting the two leaf sequence data as the same variable, the leaf sequence data is subjected to multi-level wavelet packet decomposition under a three-level wavelet basis to obtain a one-dimensional wavelet packet coefficient matrix at each scale.

[0014] The wavelet packet decomposition uses a function that satisfies a preset dual-scale equation, which includes two sets of conjugate filter banks, and the coefficients of the conjugate filter banks are orthogonal.

[0015] The one-dimensional wavelet packet coefficient matrix obtained after filtering and reconstruction is defined as two new sequence data, which are used as the filtered image.

[0016] In some embodiments, edge trajectory extraction is performed based on the filtered image to obtain a blade edge trajectory image, including:

[0017] The pixel values ​​of the filtered image are converted into grayscale values, and edge points reflecting grayscale changes are extracted based on the grayscale values.

[0018] Redundant boundary points in the edge points are removed and boundary discontinuities are filled to obtain the first edge map sequence and the second edge map sequence in the orthogonal direction of the blade unit under the same state.

[0019] The first edge map sequence and the second edge map sequence are subtracted pixel by pixel. The pixels with non-zero values ​​in the subtraction result are set to 1. At the same time as setting the pixel to 1, the position of the pixel is defined as the trajectory point and the trajectory point value is 1, thus obtaining the blade edge trajectory image sequence.

[0020] In some embodiments, edge invariant moment feature extraction of the blade edge trajectory image includes:

[0021] The blade edge trajectory image is defined as a binary blade edge trajectory, where trajectory point value = 1 and non-trajectory point value = 0;

[0022] Based on the blade edge trajectory, the original moment set is output by calculating the spatial distribution statistics of the blade edge point set, wherein the original moment set includes: zero-order moment, first-order moment and higher-order moments;

[0023] Based on the original set of matrices, the centroid coordinates of the blade edge trajectory image are calculated, and the discretized central moments are calculated based on the centroid coordinates.

[0024] The central moments are scaled and normalized to obtain normalized central moments. Based on the standardized invariant moment feature set studied by Hu, the normalized central moments are processed to generate rotationally invariant moments in multiple dimensions. The rotationally invariant moments are defined as the invariant moment feature sequence.

[0025] In some embodiments, determining the initial anomalous sample based on multiple invariant moment feature sequences includes:

[0026] Select the first invariant moment feature sequence and the second invariant matrix feature sequence that are continuous in time or space to construct anomaly comparison samples;

[0027] Based on the aforementioned anomaly comparison samples, the weights of each invariant moment feature are calculated;

[0028] By combining the ETD distance formula of the weights, the comparative difference between the first invariant moment feature sequence and the second invariant matrix feature sequence, and the benchmark difference of the first invariant moment feature sequence are calculated.

[0029] Calculate the correlation coefficient of the abnormal comparison samples in the time series, and define the sequence similarity of the invariant moment feature sequence based on the correlation coefficient;

[0030] The difference parameter ratio is calculated based on the comparison difference and the benchmark difference. Based on the difference parameter ratio and the similarity, it is determined whether the abnormal comparison sample is an abnormal sample. If so, it is defined as an initial abnormal sample.

[0031] In some embodiments, calculating the weights of each invariant moment feature based on the anomaly comparison samples includes:

[0032] For each dimension of invariant moment features, calculate the absolute difference between the first invariant moment feature sequence and the second invariant matrix feature sequence;

[0033] Based on the absolute difference, the invariant moment features of each dimension are arranged in ascending order from smallest to largest to obtain the sorting result;

[0034] Based on the sorting results, the invariant moment features of the first k dimensions are assigned progressively decreasing weights, where k is a preset threshold.

[0035] In some embodiments, based on the initial anomaly samples, iterative optimization is performed using a particle swarm optimization algorithm to obtain key anomaly samples, including:

[0036] A particle swarm optimization model is constructed based on the invariant moment characteristics of the initial abnormal samples, wherein the invariant moment characteristic parameters of the initial abnormal samples are used as particles, and the number of particles, the maximum number of iterations, the learning factor, and the inertia weight are set.

[0037] A fitness function is constructed based on the difference parameter ratio and the correlation coefficient as key indicators, wherein the fitness function value is used to reflect the significance of abnormal features of the sample;

[0038] Iterative optimization in particle swarm optimization includes: updating the position and velocity of particles through individual extreme values ​​and global extreme values, adjusting the search direction according to the current fitness function value, and screening out target optimization samples with significant abnormal features and representativeness. The particle swarm optimization algorithm adopts a dynamic inertia weight strategy, and the inertia weight decays exponentially with the number of iterations.

[0039] When the iteration reaches the preset maximum number of iterations or the fitness function value tends to stabilize, the abnormal sample corresponding to the globally optimal particle is output as a key abnormal sample for constructing blade fault diagnosis.

[0040] Secondly, embodiments of this application provide a fault diagnosis system for abnormal edge trajectory of wind turbine blades, the system comprising: an acquisition module, a feature extraction module, and a detection module, wherein:

[0041] The acquisition module is used to acquire images of wind turbine blades and perform multi-scale filtering on the blade images using improved wavelet packets to obtain filtered images.

[0042] The feature extraction module is used to extract edge trajectories from the filtered image to obtain a blade edge trajectory image, and to extract edge invariant moment features from the blade edge trajectory image to output multiple invariant moment feature sequences.

[0043] The detection module is used to determine initial abnormal samples based on multiple invariant moment feature sequences, construct a particle swarm optimization model based on the invariant moment features of the initial abnormal samples, and perform iterative optimization processing through the particle swarm optimization model to obtain key abnormal samples for blade fault diagnosis.

[0044] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect above.

[0045] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0046] Compared to related technologies, the present application provides a method for diagnosing abnormal edge trajectories of wind turbine blades. This method employs improved wavelet packet multi-scale analysis of blade operating state images for filtering and noise reduction; it obtains blade trajectory edge features through invariant moment features to achieve translational, scaling, and rotational invariance recognition; and it utilizes a particle swarm optimization algorithm to obtain images with large edge trajectory deviations, calibrating abnormal states for fault diagnosis. This method solves the problems in existing technologies, such as the inability of edge detection operator thresholds to adapt, poor noise resistance, and difficulty in extracting abnormal blade edge trajectory features, leading to ineffective fault diagnosis. It improves fault diagnosis accuracy and, by filtering redundant data through optimization algorithms, balances the accuracy and efficiency of intelligent recognition, providing reliable technical support for wind turbine blade fault diagnosis. Attached Figure Description

[0047] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0048] Figure 1 This is a flowchart of a method for diagnosing abnormal edge trajectory faults of wind turbine blades according to an embodiment of this application;

[0049] Figure 2 This is a flowchart of wavelet packet multi-scale filtering of leaf images according to an embodiment of this application;

[0050] Figure 3 This is a schematic diagram of wavelet packet decomposition according to the embodiments of this application;

[0051] Figure 4 This is a structural block diagram of a wind turbine blade edge trajectory abnormality fault diagnosis system according to an embodiment of this application;

[0052] Figure 5 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0054] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0055] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0056] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0057] This application provides a method for diagnosing abnormal edge trajectory faults in wind turbine blades. Figure 1 This is a flowchart of a method for diagnosing abnormal edge trajectory faults of wind turbine blades according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:

[0058] S101, acquire the blade image of the wind turbine, and perform multi-scale filtering on the blade image using improved wavelet packets to obtain the filtered image;

[0059] Among them, the measured image signal of the blade is composed of a linear combination of the real signal and noise, and the wavelet transform of the measured signal is also composed of the wavelet transform of the signal and the wavelet transform of the noise. As the scale of the wavelet increases, the maximum points of the noise decrease significantly, and the amplitude of its wavelet transform decreases with the increase of scale. At large scales, the wavelet will mainly belong to the signal.

[0060] Therefore, in this embodiment, a coarse-to-fine strategy is used to track the maxima of wavelet transform at each scale, identify the parts belonging to the signal, remove the noise, and reconstruct the signal from these selected maxima points; the specific process of wavelet packet multi-scale filtering is as follows: Figure 2 As shown; based on Figure 2 Based on the principle, this embodiment constructs an improved wavelet model to refine the features of the blade trajectory graphic.

[0061] Figure 2 This is a flowchart of wavelet packet multi-scale filtering of leaf images according to an embodiment of this application, as follows: Figure 2 As shown, the process consists of four steps: ① Selecting the wavelet basis function (e.g., db4, balancing smoothness and computational efficiency) and the number of decomposition layers N (balancing frequency resolution and computational power); ② Wavelet packet decomposition, splitting the image signal into high- and low-frequency sub-bands at multiple scales (more refined than traditional wavelets, fully preserving edge details); ③ Thresholding quantization of the coefficients of each sub-band (suppressing noise-dominant small coefficients through soft / hard thresholding); ④ Layer-by-layer reconstruction to restore the denoised image. Its core function is to filter out noise, enhance the true contour of the blade edges, solve the "false edge / fracture" problem of traditional edge detection, provide high-fidelity data for subsequent invariant moment feature extraction, and directly improve the accuracy of fault diagnosis.

[0062] Specifically, step S101 includes the following sub-steps:

[0063] S1, acquire images of the blade unit from two orthogonal positions of the wind turbine unit to obtain blade sequence data at the two positions;

[0064] In this embodiment, the acquired blade unit images are specifically image sequences of the blade operating status acquired from two different orthogonal positions of the wind turbine unit, respectively obtaining image sequence data x(n) and y(n) at time n, where n is time and x and y are image numbers; the purpose of acquiring images at orthogonal positions is to comprehensively capture the spatial features of the blade from different angles, providing raw data for subsequent multi-dimensional analysis, ensuring the integrity of the blade image information, and laying the foundation for subsequent filtering and edge extraction.

[0065] S2. After setting the two leaf sequence data as the same variable, perform multi-layer wavelet packet decomposition on the leaf sequence data under the three-layer wavelet basis to obtain the one-dimensional wavelet packet coefficient matrix at each scale. The function used for wavelet packet decomposition satisfies the preset dual-scale equation. The preset dual-scale equation includes two sets of conjugate filter groups, and the coefficients of the conjugate filter groups have an orthogonal relationship.

[0066] Specifically, firstly, the two image sequence data x(n) and y(n) obtained in step S1 are uniformly set as variable U(n) for unified processing; then, processing is performed based on the wavelet packet function that satisfies the preset dual-scale equation, which is:

[0067]

[0068] In the formula, For a conjugate filter bank, where That is, the two coefficients are also orthogonal, which can guarantee the stability and orthogonality of the decomposition.

[0069] Subsequently, U(n) is subjected to a three-level wavelet packet multi-scale decomposition, which decomposes the signal into subspaces of different scales, and finally extracts the one-dimensional wavelet packet coefficient matrix at each scale. It can be understood that the purpose of the three-level decomposition is to distinguish between signal and noise through multi-scale analysis (the amplitude of noise decreases at large scales), and to provide a coefficient basis for subsequent noise filtering.

[0070] It should also be noted that, Figure 3 This is a schematic diagram of wavelet packet decomposition based on an embodiment of this application, such as... Figure 3 As shown, the level 3 of the U(n) wavelet packet decomposition is marked as the signal after image filtering. Then the signal Perform a 3-level wavelet packet decomposition. Wavelet space. That is

[0071]

[0072] Where m = 0, 1, 2, ..., ; by subspace Subspaces are divided into different scales j. Synthetic Hilbert space That is This indicates that the entire space of square-integrable functions can be completely covered by these wavelet subspaces. The inverse process (i.e., wavelet packet decomposition) involves transforming the original signal from... Decomposing the signal into each Wj subspace provides a mathematical basis for subsequent noise removal and effective signal retention through threshold quantization, ensuring that the three-layer decomposition can effectively separate the signal and noise, laying the foundation for filtering and reconstruction.

[0073] S3 defines the one-dimensional wavelet packet coefficient matrix obtained after filtering and reconstruction as two new sequence data, which are used as the filtered image.

[0074] Specifically, wavelet packet multi-scale decomposition is performed on x(n) and y(n) under a three-level wavelet basis, and the one-dimensional wavelet packet coefficient matrix at each scale is extracted. Next, further "filtering" is performed: based on the difference in coefficient amplitude between noise and signal (noise coefficient amplitude is small), threshold quantization (such as setting a reasonable threshold and setting the noise coefficient below the threshold to zero) is used to remove redundant coefficients corresponding to noise, and only the effective coefficients reflecting the characteristics of the blade trajectory are retained.

[0075] Subsequently, through a "reconstruction" operation (i.e., the inverse wavelet packet decomposition process, based on subspace synthesis theory), the filtered... The coefficient matrix is ​​recombined to finally obtain the denoised leaf image signal, namely the filtered images x(n) and y(n).

[0076] Step S3 purifies the original image signal, removes noise interference, and makes x(n) and y(n) more clearly retain the true features of the blade trajectory, providing high-quality preprocessing data for subsequent "blade image edge trajectory extraction" (such as grayscale conversion and edge point extraction), and ensuring the accuracy of edge detection.

[0077] In step S101 above, by analyzing the wavelet transform characteristics of the real signal and noise in the measured image signal of the blade, a method is adopted to separate the signal and noise and reconstruct the signal by tracking the maximum value of the wavelet transform at each scale from coarse to fine. This constructs a wavelet model to purify the features of the blade trajectory graphic, providing a high-quality signal basis after denoising for subsequent multi-scale filtering of wavelet packets and edge trajectory extraction.

[0078] S102, extract the edge trajectory from the filtered image to obtain the blade edge trajectory image, and extract the edge invariant moment features from the blade edge trajectory image to output multiple invariant moment feature sequences;

[0079] The process of extracting the edge trajectory from the filtered image to obtain the blade edge trajectory image includes the following sub-steps:

[0080] S1 converts the pixel values ​​of the filtered image into grayscale values ​​and extracts edge points that reflect grayscale changes based on the grayscale values;

[0081] It is understandable that in step S1, based on the filtered leaf image signals x(n) and y(n), the RGB values ​​of the image pixels are converted to grayscale values ​​using the color image RGB to grayscale formula (gray = R * 0.299 + G * 0.587 + B * 0.114), thus simplifying the image from color to grayscale and reducing the data dimensionality to improve the efficiency of subsequent processing.

[0082] Based on this, edge points are extracted according to the characteristics of gray value changes. The location of the gray value change is the potential location of the leaf edge. This step locates the edge by capturing gray value changes, providing an initial set of edge points for subsequent edge refinement processing.

[0083] S2, remove redundant boundary points from the edge points and fill in the boundary discontinuities to obtain the first edge map sequence and the second edge map sequence in the orthogonal direction of the blade unit under the same state;

[0084] Considering that the initial edge points extracted in step S1 may contain redundant boundary points (non-real edges) and boundary discontinuities (edge ​​discontinuities) caused by residual noise or image blurring, this step optimizes the accuracy and continuity of the edges by removing redundant points (such as filtering out pseudo-edges with insignificant gray-level changes) and filling in discontinuities (such as connecting adjacent edge segments with consistent gray-level change trends).

[0085] Meanwhile, for the orthogonally filtered images x(n) and y(n) of the blade under the same state, the corresponding first edge map sequence E1 (x, y) and second edge map sequence E2 (z, y) are generated respectively, capturing edge features from the orthogonal direction to ensure the integrity of edge information.

[0086] S3, perform pixel-by-pixel subtraction on the first edge map sequence and the second edge map sequence, set the non-zero value of the pixels in the subtraction result to 1, and define the position of the pixel as the trajectory point while setting the pixel to 1, and set the trajectory point value to 1, to obtain the blade edge trajectory image sequence.

[0087] Specifically, the first edge image sequence E1 and the second edge image sequence E2 obtained in step S2 are subtracted pixel by pixel. This difference operation filters out redundant edges (such as background or non-critical structure edges) that overlap between the two edge images. The numerical part of the subtraction result is set to 1, which represents the trajectory point. The core of this step is to suppress ineffective edges with weak differences, further highlighting effective edges with significant grayscale changes. Finally, a refined blade edge trajectory image sequence E3(x, y) is obtained, which is the blade edge trajectory image f(x, y) output by edge trajectory extraction. This provides a clear and accurate edge trajectory foundation for subsequent invariant moment feature extraction.

[0088] Through the above steps S1 to S3, the filtered image is converted to grayscale, edge points are extracted and optimized, and the edge map is thinned by subtraction, thus extracting a clear and accurate blade edge trajectory image, which provides a reliable edge trajectory basis for subsequent blade trajectory invariant moment feature extraction and abnormal fault diagnosis.

[0089] Furthermore, edge invariant moment feature extraction is performed on the blade edge trajectory image to output multiple invariant moment feature sequences, including the following sub-steps:

[0090] S1, define the blade edge trajectory image as a binary blade edge trajectory, where trajectory point value = 1 and non-trajectory point value = 0;

[0091] Specifically, the blade edge trajectory image f(x,y) is defined as a binary trajectory: the gray value of each coordinate point traversed by the blade edge trajectory is set to 1, and the gray value of coordinate points not traversed is set to 0. Binarization can significantly simplify the image data dimension, retain only the spatial location information of the edge trajectory, filter out irrelevant gray-level fluctuations, provide a clear and unified input basis for subsequent moment feature calculation, and ensure that feature extraction focuses on the geometric distribution of the trajectory itself.

[0092] S2, based on the blade edge trajectory, outputs the original moment set by calculating the spatial distribution statistics of the blade edge point set, where the original moment set includes the zeroth moment, the first moment, and higher-order moments;

[0093] Specifically, based on the binarized blade edge trajectory, the (p+q)th order primitive moment is calculated, expressed as follows: This includes the zeroth moment (p=0, q=0, reflecting the overall existence of the trajectory) and the first moment (p+q=1, such as...). These primitive moments (p+q≥2, reflecting the basic distribution related to the centroid of the trajectory) and higher-order moments (p+q≥2, reflecting the detailed features of the trajectory shape) serve as statistics of the spatial distribution of edge point sets, fully characterizing the position, range, and morphological features of the trajectory, and are the basic data for subsequent calculation of invariant moments.

[0094] S3, based on the original matrix set, calculates the centroid coordinates of the blade edge trajectory image, and calculates the discretized invariant line moments based on the centroid coordinates;

[0095] First, the centroid coordinates of the blade edge trajectory image are calculated based on the original moments:

[0096] ,in It is the zeroth moment. (The first moment), the centroid coordinates reflect the geometric center position of the trajectory.

[0097] Furthermore, the discretized central moments are calculated with the centroid as the origin:

[0098] It is understandable that the central moments eliminate the influence of trajectory translation on the features (no matter where the trajectory is translated, the moments with the centroid as the origin remain unchanged), laying the foundation for subsequent scaling and rotation invariance processing.

[0099] S3. The central moments are scaled and normalized to obtain normalized invariant linear moments. Based on the standardized invariant linear moment feature set studied by Hu, the normalized central moments are processed to generate multi-dimensional rotational invariant moments. The multiple rotational invariant moments are defined as an invariant moment feature sequence.

[0100] In this step, the invariant linear moments are scaled: by... Eliminate the influence of trajectory scale changes (regardless of whether the trajectory is enlarged or reduced, the normalized moment value remains stable).

[0101] Next, based on Hu's theory of normalized invariant moments, seven rotationally invariant moments are derived using the second and third order normalized central moments. μ 1 to μ7) Its expression ensures that the moment value remains unchanged during trajectory rotation. The seven rotationally invariant moments ultimately generated constitute an invariant moment feature sequence, achieving invariance identification of translational, scaling, and rotational changes in the blade edge trajectory. Each trajectory can be uniquely labeled with a feature sequence, providing a stable feature benchmark for subsequent anomaly detection. The seven Hu invariant moments are defined as follows:

[0102] (4.8)

[0103] Through the above steps S1 to S3, seven invariant moment features of the blade image edge are obtained, realizing that each blade trajectory pattern is marked by a unique edge invariant moment. Therefore, the recognition of edge invariant moments completes the recognition of the blade trajectory pattern.

[0104] S103. Initial abnormal samples are determined based on multiple invariant moment feature sequences. A particle swarm optimization model is constructed based on the invariant moment features of the initial abnormal samples. The particle swarm optimization model is used for iterative optimization to obtain key abnormal samples for blade fault diagnosis.

[0105] Determining the initial outlier sample based on multiple invariant moment feature sequences includes the following sub-steps:

[0106] S1. Select the first invariant moment feature sequence and the second invariant matrix feature sequence that are continuous in time or space to construct anomaly comparison samples, and calculate the weight of each invariant moment feature based on the anomaly comparison samples.

[0107] Among the invariant moment feature sequences output in step S102, two sets of invariant moment feature sequences that are continuous in time or space are first selected, namely A. i (i=1,…,7) and B i+1 (i=1,…,7) are used as anomaly comparison samples, where A i and B i+1 All contain invariant moment features in 7 dimensions (μ1 to μ7).

[0108] Furthermore, calculating the weights of each invariant moment feature includes:

[0109] S1.1, For each dimension of invariant moment features, calculate the absolute difference between the first invariant moment feature sequence and the second invariant matrix feature sequence;

[0110] S1.2, Based on absolute interpolation, sort the invariant moment features of each dimension in ascending order from smallest to largest to obtain the sorting result;

[0111] S1.3, based on the ranking results, assign successively decreasing weights to the invariant moment features of the first k dimensions, where k is a preset threshold.

[0112] In this embodiment, when calculating the weights of the invariant moment features of each dimension, the absolute difference between Ai and Bi+1 is first calculated for each dimension (e.g., μ1 to μ7) to reflect the difference between the two sets of sequences in that dimension. Then, the absolute differences of each dimension are sorted from smallest to largest. According to the principle of "nearer is larger and farther is smaller", the first k dimensions with smaller differences are assigned progressively decreasing weights. Among them, the smaller the difference, the stronger the representativeness of the trajectory features, and the higher the weight. Finally, the weight set W={wi}(i=1,…,7) is obtained, which provides the weight basis for subsequent difference calculation.

[0113] S2, by combining the ETD distance formula of the weights, calculate the comparative difference between the first invariant moment feature sequence and the second invariant matrix feature sequence, and the benchmark difference of the first invariant moment feature sequence;

[0114] Here, the weight W={wi} obtained from S1 is used to calculate the difference between the two sets of sequences using the defined ETD distance formula:

[0115] Among them, "comparison difference" refers to the ETD distance between the first invariant moment feature sequence (Ai) and the second invariant moment feature sequence (Bi+1), i.e., D (A,B). In the formula, the differences of each dimension are weighted by weight wi to highlight the influence of important dimensions (dimensions with high weight) on the overall difference.

[0116] Specifically, the ETD distance from A to B is defined as follows:

[0117] :

[0118] The weighting function is selected based on the principle of near-large and far-small: ,in, The predicted value of point N+1 in sequence S is: ;

[0119] Furthermore, the "benchmark difference" refers to the distance between the first invariant moment feature sequence (Ai) and its own ETD, i.e., D(A,A), which serves as a benchmark value for measuring the difference (theoretically, the difference between itself and its counterpart is 0, but in practice it is used for the standardized ratio). In this embodiment, the high-dimensional invariant moment feature difference is transformed into a quantifiable value through ETD distance calculation, providing data support for subsequent judgment.

[0120] S3, calculate the correlation coefficient of the abnormal comparison samples in the time series, and define the sequence similarity of the invariant moment feature sequences based on the correlation coefficient;

[0121] Based on the differences in each dimension obtained from S2, the following correlation coefficient formula is used:

[0122]

[0123] Calculate the correlation coefficient between the two time series A and B. The correlation coefficient reflects the overall similarity between the two series: the closer the value is to 1, the more consistent the trend of the series changes (high similarity); the lower the value, the greater the difference in the trend (low similarity). This correlation coefficient provides a basis for the correlation of the series for anomaly judgment.

[0124] S4. Calculate the difference parameter ratio based on the comparison difference and the baseline difference, and determine whether the abnormal comparison sample is an abnormal sample based on the difference parameter ratio and similarity. If so, define it as the initial abnormal sample.

[0125] Understandably, this step first calculates the difference parameter ratio, i.e., the ratio of the contrast difference to the baseline difference, D(A,B) / D(A,A), and then combines it with the sequence similarity obtained in S3 for a comprehensive judgment.

[0126] If the ratio is ≥80% and the correlation coefficient shows high similarity, it indicates that the two sets of sequences have similar characteristics and are judged to be normal.

[0127] If the ratio is less than 80%, or the correlation coefficient shows low similarity, it indicates that the two sets of sequences have significant differences in characteristics and are judged as abnormal. The comparison sample is then defined as the initial abnormal sample.

[0128] Steps S1 to S3, through quantified difference ratios and sequence similarity, achieve preliminary screening of abnormal states of blade edge trajectories, laying the foundation for subsequent optimization algorithms to further accurately identify abnormal samples.

[0129] In one exemplary embodiment, Table 1 below is a table of edge trajectory values ​​for abnormal leaf samples:

[0130]

[0131] Table 2 below shows the numerical values ​​of the edge trajectories of unknown leaf samples:

[0132]

[0133] Table 1 lists the seven invariant moments (μ1-μ7) values ​​of four known normal leaves. Its core function is to serve as a "benchmark feature library". The μ values ​​of normal samples are generally small in range (e.g., μ1 is between 0.0108 and 0.0145, and μ2 is between 0.0597 and 0.1026), and the values ​​are stable without significant positive or negative fluctuations. This reflects the typical invariant moment characteristics of the edge trajectory of normal leaves and provides a "normal state reference" for comparison with unknown samples.

[0134] Table 2, the edge trajectory numerical table of unknown leaf samples, contains 7 invariant moment values ​​of 16 samples to be diagnosed. These samples are the targets for anomaly screening, and their μ values ​​are significantly different from those of normal samples.

[0135] For example, the μ1 of unknown sample 1 is 0.7099 (about 49 times the maximum value of μ1 of normal samples), and μ3 is -1.8390 (μ3 of normal samples is positive and ≤0.1184). The overall numerical range is larger and the positive and negative fluctuations are obvious. By calculating the ratio of D(A,B) / D(A,A) with normal samples, it is possible to preliminarily determine which samples deviate from normal characteristics.

[0136] It is understandable that the values ​​in Tables 1 and 2 are the input data for the subsequent particle swarm optimization of edge-invariant moments. By comparing the invariant moment values ​​of 16 unknown samples with those of 4 normal samples, samples that differ significantly from the normal samples (i.e., the ratio < 80%) are selected. This provides a basis for the subsequent particle swarm algorithm to further refine key abnormal samples (such as the 8 samples retained in the final document optimization), ultimately achieving redundant data removal and improving the accuracy and efficiency of blade fault diagnosis.

[0137] Furthermore, based on the initial anomaly samples, the key anomaly samples are obtained through iterative optimization using the particle swarm optimization algorithm, including the following detailed steps:

[0138] S1. Construct a particle swarm optimization model based on the invariant moment characteristics of the initial abnormal samples, where the invariant moment characteristic parameters of the initial abnormal samples are used as particles, and the number of particles, the maximum number of iterations, the learning factor, and the inertia weight are set.

[0139] Specifically, the seven invariant moment feature parameters (μ1 to μ7) of the initial abnormal sample are treated as "particles". Each particle is represented by a triple (xi, vi, pbest_i), where xi is the current position of the particle (the subscript of the corresponding invariant moment feature i=1,2,…,7), vi is the current velocity of the particle (the value of the corresponding invariant moment feature), and pbest_i is the optimal position found by the particle itself (i.e. the best combination of invariant moment features in the history of the particle).

[0140] Furthermore, key parameters are set, including the number of particles (determined by the initial number of outlier samples), the maximum number of iterations, learning factors C1 and C2 (positive numbers, used to adjust the weights of particles learning towards individual and global extrema), and the initial and final inertia weights Wmax and Wmin (used to balance the global exploration and local exploitation capabilities of particles). This step provides a basic model framework for particle swarm optimization, ensuring that the optimization process revolves around the invariant moment characteristics of outlier samples.

[0141] S2, a fitness function is constructed based on the difference parameter ratio and correlation coefficient as key indicators, whereby the fitness function value is used to reflect the significance of abnormal features of the sample;

[0142] Specifically, a fitness function is constructed using the "difference parameter ratio (D(A,B) / D(A,A))" and the "correlation coefficient (N(A,B))" as core indicators. The fitness function value directly reflects the significance of the abnormal characteristics of the sample: if the difference parameter ratio of a sample is smaller (the further it deviates from the normal sample) and the correlation coefficient is lower (the greater the difference in trend from the normal sample), then the fitness function value is smaller (or larger according to the setting), indicating that the abnormal characteristics of the sample are more significant.

[0143] The optimization objective of this embodiment is to "obtain a sample set Xi with the smallest difference from the data to be identified" (here, "smallest difference" means the highest degree of matching with abnormal features). Therefore, the fitness function design should prioritize the selection of samples with a difference parameter ratio of <80% and a low correlation coefficient to provide a quantitative evaluation standard for subsequent iterative optimization.

[0144] S3, iterative optimization in the particle swarm includes: updating the position and velocity of particles through individual extreme values ​​and global extreme values, adjusting the search direction according to the current fitness function value, and screening out target optimization samples with significant abnormal features and representativeness. Among them, the particle swarm algorithm adopts a dynamic inertia weight strategy, and the inertia weight decays exponentially with the number of iterations.

[0145] The core of this step is to filter representative anomalous samples by iteratively updating the particle state:

[0146] Specifically, during the iteration process, each particle updates its velocity and position according to the formula by tracking the "individual extreme value pbest_i" (its own historical best position) and the "global extreme value gbest" (the historical best position of the entire particle swarm): the velocity update combines the learning factor, random number and extreme value deviation, and the position update is the current position plus the new velocity, ensuring that the particle searches for better abnormal feature regions.

[0147] A dynamic inertia weight strategy is adopted: the inertia weight w decreases as the number of iterations increases according to the formula w=Wmax - (Wmax - Wmin)*t / tmax (t is the current iteration number and tmax is the maximum iteration number). In each iteration, the search direction is adjusted according to the fitness function value: if the fitness function value of the current particle is better than the historical extreme value, then pbest_i or gbest is updated, and the target samples with the most significant abnormal features and the most representative are gradually selected.

[0148] S4. When the iteration reaches the preset maximum number of iterations or the fitness function value tends to stabilize, the abnormal sample corresponding to the globally optimal particle is output as a key abnormal sample for constructing blade fault diagnosis.

[0149] It is understandable that when the number of iterations reaches the preset maximum, or when the fitness function value changes less than the threshold (tends to stabilize) in multiple consecutive iterations, the optimization stops: the abnormal sample represented by the particle corresponding to the global extreme value gbest is output, which is the "key abnormal sample". These samples have been filtered out of redundant data (e.g., 8 out of 16 unknown samples in the document are ultimately retained after optimization), and fully contain the characteristics of all abnormal states ("the optimization sample library fully contains the edge trajectory samples of the four leaf graphics of all abnormal states").

[0150] Those skilled in the art will recognize that the output of key abnormal samples provides high-quality training data for subsequent blade fault diagnosis models (such as BP neural networks), balancing the accuracy and efficiency of intelligent identification, and ultimately improving the reliability of wind turbine blade fault diagnosis.

[0151] In one exemplary embodiment, a specific optimization process includes the following steps:

[0152] Step 1: Each individual entity in the search space is called a "particle". Each particle has its own position and velocity, which can be represented by triples. It means that, among them Indicates the current position of the particle. This indicates the particle's current velocity. This indicates the best position that the particle itself has searched for;

[0153] Step 2: Iterate repeatedly between the particle's current position and velocity, using D(A,B) / D(A,A) as the final condition for determining anomalies. If the ratio is greater than or equal to 80%, the particle is considered similar. If the ratio is less than 80%, the particle is considered similar.

[0154] If a better solution is found, it is used as a basis to find the next solution, until the optimal solution that satisfies the conditions is found; in each iteration, the particle updates itself by tracking two "extremes": the first is the best solution found by the particle itself, called the individual extreme point (using...). (Indicates its location).

[0155] Step 3: The other extreme point is the best solution found so far for the entire population, called the global extreme point (using...). (Indicates its position); Assume the search space of the problem is a square in two-dimensional space, denoted by S, and S = a1, b1 × ... × aD, bD. After finding the two best solutions in each iteration, the particle... Update your speed and position according to the following formula:

[0156]

[0157] In the formula: Let be the optimal position traversed by particle i. It represents the best position that all particles in the swarm can reach; A positive constant is the learning factor. The values ​​are random numbers in the interval [0,1]; thus, the optimal particle swarm that satisfies the conditions is obtained.

[0158] Step 4: By comparing the edge moment data of the 16 acquired images with the 4 measured images, it was found that the values ​​of the corresponding positions of the invariant moments in similar blade edge trajectory images were extremely similar. Therefore, when the sample invariant moments were grouped into a set Ai (i=1,2,…,7) and the invariant moments of the image to be identified were grouped into a set Bi (i=1,2,…,7), the analysis of the datasets revealed that at the same position i, the invariant moment values ​​of Ai and Bi for similar blade edge trajectories exhibit approximately equal maxima and minima. The greater the difference between the blade edge trajectory images, the greater the difference between the maxima and minima of their invariant moments and their corresponding positions.

[0159] Based on the above principles, this paper proposes an improved global optimization algorithm for particle swarm optimization. First, the invariant moments are treated as particles in a particle swarm, with the particle's position corresponding to the invariant moment values ​​among the seven invariant moments (i.e., i = 1, 2, ..., 7), and the particle's velocity being the invariant moment value (i.e., Ai or Bi). Second, the velocities at corresponding positions (i.e., i = 1, 2, ..., 7) in the two particle swarms (i.e., Ai and Bi) are compared, and the particles with similar velocities at the seven positions are selected as the optimization particles. The optimized particle replaces the original particle Ai and is denoted as Xi. Then, the objective function value of the new particle is calculated. If f(Xi) ≤ f(Ai), Xi is set to Ai, and the inertia weight is updated using equation (4.10). Finally, it is determined whether the termination condition has been met. If so, the calculation is stopped, and the optimal solution is output. Otherwise, the iteration count is updated, and the calculation continues. This iteration is repeated until the global optimum is reached. Thus, the algorithm obtains a sample set Xi with the smallest difference from the data to be identified. Based on this, the accuracy and efficiency of the BP neural network recognition algorithm below are greatly improved.

[0160] The weighting factors for global optimization are:

[0161]

[0162] in: These are the initial and final values ​​of the inertia weight. t is the maximum number of iterations of the algorithm.

[0163] The optimized data is shown in Table 3 below: Table of leaf sample edge trajectory values ​​after particle swarm optimization:

[0164]

[0165] As can be seen from the above optimization results, 8 samples were obtained in the actual test. The removal of excessive redundant data allows the intelligent recognition algorithm to balance accuracy and efficiency. The key samples obtained after optimization are sample 1, sample 2, sample 6, sample 8, sample 10, sample 11, sample 15, and sample 16. Comparing the optimized graphic samples with the known sample library, it is found that the optimized sample library completely contains the edge trajectory samples of the four leaf graphics in all abnormal states.

[0166] Through steps S101 to S103, blade images are processed using wavelet packet multi-scale filtering, invariant moment features (including Hu invariant moments) are extracted to capture trajectory characteristics, and key abnormal samples are screened using particle swarm optimization, thus achieving blade fault diagnosis. This effectively solves the problems of traditional edge detection operators' inability to adapt thresholds and poor noise resistance. Specifically, wavelet packet multi-scale analysis accurately separates signals and noise, making blade image features clearer and improving edge extraction accuracy, laying a high-quality data foundation for subsequent diagnosis. Furthermore, the translation, scaling, and rotation invariance of invariant moment features ensures stable identification of blade trajectory features under different states, avoiding feature distortion caused by changes in blade motion posture. In addition, the particle swarm optimization algorithm eliminates redundant data, and the key abnormal samples selected are highly representative, significantly improving the efficiency of fault diagnosis while ensuring the integrity of abnormal features.

[0167] This application balances diagnostic accuracy with intelligent identification efficiency, providing reliable technical support for the timely detection and accurate determination of wind turbine blade faults, which helps reduce operation and maintenance costs and improve the safety and stability of unit operation.

[0168] Secondly, embodiments of this application also provide a fault diagnosis system for abnormal edge trajectory of wind turbine blades. Figure 4 This is a structural block diagram of a wind turbine blade edge trajectory abnormality fault diagnosis system according to an embodiment of this application, as shown below. Figure 4 As shown, the system includes: an acquisition module 40, a feature extraction module 41, and a detection module 42, wherein:

[0169] The acquisition module 40 is used to acquire images of wind turbine blades and perform multi-scale filtering on the blade images using improved wavelet packets to obtain filtered images.

[0170] The feature extraction module 41 is used to extract the edge trajectory based on the filtered image to obtain the blade edge trajectory image, and to extract the edge invariant moment features from the blade edge trajectory image to output multiple invariant moment feature sequences.

[0171] The detection module 42 is used to determine the initial abnormal sample based on multiple invariant moment feature sequences, construct a particle swarm optimization model based on the invariant moment features of the initial abnormal sample, and perform iterative optimization processing through the particle swarm optimization model to obtain key abnormal samples for blade fault diagnosis.

[0172] This system accurately separates signals and noise through wavelet packet multi-scale analysis, resulting in clearer blade image features and improved edge extraction accuracy, laying a high-quality data foundation for subsequent diagnosis. Furthermore, leveraging the translational, scaling, and rotational invariance of invariant moment features, it ensures stable identification of blade trajectory features under different states, avoiding feature distortion caused by changes in blade motion attitude. In addition, the particle swarm optimization algorithm eliminates redundant data, and the selected key abnormal samples are highly representative, significantly improving the efficiency of fault diagnosis while ensuring the integrity of abnormal features. This application's solution balances diagnostic accuracy and intelligent recognition efficiency, providing reliable technical support for the timely detection and accurate determination of wind turbine blade faults, helping to reduce operation and maintenance costs and improve the safety and stability of unit operation.

[0173] In one embodiment, Figure 5 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application, such as... Figure 5 As shown, an electronic device is provided, which can be a server, and its internal structure diagram can be as follows. Figure 5 As shown, the electronic device includes a processor, a network interface, internal memory, and non-volatile memory connected via an internal bus. The non-volatile memory stores the operating system, computer programs, and a database. The processor provides computing and control capabilities, the network interface communicates with external terminals via a network, the internal memory provides the environment for the operating system, the computer program is executed by the processor to implement a fault diagnosis method for abnormal edge trajectory of wind turbine blades, and the database stores data.

[0174] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0175] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0176] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0177] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for diagnosing abnormal edge trajectory faults in wind turbine blades, characterized in that, The method includes: A sequence of wind turbine blade images is obtained, and the blade image sequence is then subjected to multi-scale filtering using an improved wavelet packet to obtain a filtered image sequence. Edge trajectories are extracted from the filtered image sequence to obtain a blade edge trajectory image sequence. Edge invariant moment features are extracted from the blade edge trajectory image sequence to output multiple invariant moment feature sequences. Initial abnormal samples are determined based on multiple invariant moment feature sequences. A particle swarm optimization model is constructed based on the invariant moment features of the initial abnormal samples. The particle swarm optimization model is used for iterative optimization to obtain key abnormal samples for blade fault diagnosis. The step of determining the initial abnormal sample based on multiple invariant moment feature sequences includes: Select time-continuous first and second invariant moment feature sequences to construct anomaly comparison samples; Based on the aforementioned anomaly comparison samples, the weights of each invariant moment feature are calculated; By combining the ETD distance formula of the weights, the comparative difference between the first invariant moment feature sequence and the second invariant moment feature sequence, and the benchmark difference of the first invariant moment feature sequence are calculated. Calculate the correlation coefficient of the abnormal comparison samples in the time series, and define the sequence similarity of the invariant moment feature sequence based on the correlation coefficient; The difference parameter ratio is calculated based on the comparison difference and the benchmark difference. Based on the difference parameter ratio and the similarity, it is determined whether the abnormal comparison sample is an abnormal sample. If so, it is defined as an initial abnormal sample. The process of constructing a particle swarm optimization model based on the invariant moment features of the initial abnormal samples includes: using the invariant moment feature parameters of the initial abnormal samples as particles, and constructing a fitness function with the difference parameter ratio and the correlation coefficient as key indicators. The fitness function is used to reflect the significance of the abnormal features of the samples.

2. The method according to claim 1, characterized in that, Multi-scale filtering of the leaf image sequence using improved wavelet packets includes: Image sequences of the blades are acquired from two orthogonal positions of the wind turbine, resulting in image sequences of the blades at the two positions. After setting the two leaf image sequences to the same variable, the leaf image sequences are subjected to multi-level wavelet packet decomposition under a three-level wavelet basis to obtain one-dimensional wavelet packet coefficient matrices at each scale. The wavelet packet decomposition uses a function that satisfies a preset dual-scale equation, which includes two sets of conjugate filter banks, and the coefficients of the conjugate filter banks are orthogonal. The one-dimensional wavelet packet coefficient matrix obtained after filtering and reconstruction is defined as two new sequence data, which serve as the filtered image sequence.

3. The method according to claim 1, characterized in that, Edge trajectory extraction is performed based on the filtered image sequence to obtain the following leaf edge trajectory image sequence: The pixel values ​​of the filtered image sequence are converted into grayscale values, and edge points reflecting grayscale changes are extracted based on the grayscale values. Redundant boundary points in the edge points are removed and boundary discontinuities are filled to obtain the first edge map sequence and the second edge map sequence in the orthogonal direction of the blade unit under the same state. The first edge map sequence and the second edge map sequence are subtracted pixel by pixel. The pixels with non-zero values ​​in the subtraction result are set to 1. At the same time as setting the pixel to 1, the position of the pixel is defined as the trajectory point and the trajectory point value is 1, thus obtaining the blade edge trajectory image sequence.

4. The method according to claim 1, characterized in that, The extraction of edge invariant moment features from the blade edge trajectory image sequence includes: Each leaf edge trajectory image is defined as a binary leaf edge trajectory, where trajectory point value = 1 and non-trajectory point value = 0; Based on the blade edge trajectory, the original moment set is output by calculating the spatial distribution statistics of the blade edge point set, wherein the original moment set includes: zero-order moment, first-order moment and higher-order moments; Based on the original set of moments, the centroid coordinates of the blade edge trajectory image are calculated, and the discretized central moments are calculated based on the centroid coordinates. The central moments are scaled and normalized to obtain normalized central moments. Based on the standardized invariant moment feature set studied by Hu, the normalized central moments are processed to generate rotational invariant moments in multiple dimensions. The rotational invariant moments are defined as the invariant moment feature sequence.

5. The method according to claim 4, characterized in that, Based on the aforementioned anomaly comparison samples, the weights of each invariant moment feature are calculated as follows: For each dimension of invariant moment features, calculate the absolute difference between the first invariant moment feature sequence and the second invariant moment feature sequence; Based on the absolute difference, the invariant moment features of each dimension are arranged in ascending order from smallest to largest to obtain the sorting result; Based on the sorting results, the invariant moment features of the first k dimensions are assigned progressively decreasing weights, where k is a preset threshold.

6. The method according to claim 4, characterized in that, A particle swarm optimization model is constructed based on the invariant moment characteristics of the initial abnormal samples. Iterative optimization is then performed using this model to obtain key abnormal samples for blade fault diagnosis, including: Set the number of particles, maximum number of iterations, learning factor, and inertia weight; The particle position and velocity are updated by individual extreme values ​​and global extreme values, and the search direction is adjusted according to the current fitness function value to screen out target optimization samples with significant abnormal features and representativeness. The particle swarm optimization model adopts a dynamic inertia weight strategy, and the inertia weight decays exponentially with the number of iterations. When the iteration reaches the preset maximum number of iterations or the fitness function value tends to stabilize, the abnormal sample corresponding to the globally optimal particle is output as the key abnormal sample for constructing blade fault diagnosis.

7. A fault diagnosis system for abnormal edge trajectory of wind turbine blades, characterized in that, The system includes: an acquisition module, a feature extraction module, and a detection module, wherein: The acquisition module is used to acquire a sequence of wind turbine blade images and to perform multi-scale filtering on the blade image sequence using an improved wavelet packet to obtain a filtered image sequence. The feature extraction module is used to extract edge trajectories based on the filtered image sequence to obtain a blade edge trajectory image sequence, and to extract edge invariant moment features from the blade edge trajectory image sequence to output multiple invariant moment feature sequences. The detection module is used to determine initial abnormal samples based on multiple invariant moment feature sequences, construct a particle swarm optimization model based on the invariant moment features of the initial abnormal samples, and perform iterative optimization processing through the particle swarm optimization model to obtain key abnormal samples for blade fault diagnosis. The step of determining the initial abnormal sample based on multiple invariant moment feature sequences includes: Select time-continuous first and second invariant moment feature sequences to construct anomaly comparison samples; Based on the aforementioned anomaly comparison samples, the weights of each invariant moment feature are calculated; By combining the ETD distance formula of the weights, the comparative difference between the first invariant moment feature sequence and the second invariant moment feature sequence, and the benchmark difference of the first invariant moment feature sequence are calculated. Calculate the correlation coefficient of the abnormal comparison samples in the time series, and define the sequence similarity of the invariant moment feature sequence based on the correlation coefficient; The difference parameter ratio is calculated based on the comparison difference and the benchmark difference. Based on the difference parameter ratio and the similarity, it is determined whether the abnormal comparison sample is an abnormal sample. If so, it is defined as an initial abnormal sample. The process of constructing a particle swarm optimization model based on the invariant moment features of the initial abnormal samples includes: using the invariant moment feature parameters of the initial abnormal samples as particles, and constructing a fitness function with the difference parameter ratio and the correlation coefficient as key indicators. The fitness function is used to reflect the significance of the abnormal features of the samples.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.

Images