Fan blade damage rapid detection method based on inverse finite element reconstruction image recognition
By combining inverse finite element method reconstruction image recognition with convolutional neural network algorithm, wind turbine blade damage can be detected in real time, solving the problems of low efficiency and insufficient accuracy in existing technologies, and realizing efficient and rapid wind turbine blade damage detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2022-09-16
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies struggle to achieve real-time, rapid, and accurate damage detection on wind turbine blades, especially the identification of early, minute damage. Furthermore, traditional methods are inefficient under complex operating conditions and struggle to distinguish between vibrations caused by damage and those caused by environmental factors.
An image recognition method based on inverse finite element reconstruction is adopted. Local strain information is measured by strain sensors on the surface of wind turbine blades to reconstruct the global displacement/strain field of the three-dimensional structure. Damage detection is performed by combining convolutional neural network image recognition algorithm, providing an end-to-end damage detection method. The accuracy of the parameters is calibrated by inverse finite element strain field reconstruction.
It enables efficient and rapid detection of wind turbine blade damage, improving detection efficiency and accuracy. It is suitable for complex working conditions and does not require material properties or load information, making it universally applicable.
Smart Images
Figure CN115575104B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a rapid detection method for wind turbine blade damage, specifically a rapid detection method for wind turbine blade damage based on inverse finite element reconstruction image recognition. Background Technology
[0002] During continuous operation, wind turbine blades are susceptible to structural deformation and fatigue damage due to factors such as wind, heavy rain, lightning strikes, icing, insufficient blade material strength, fatigue loads, or human errors during manufacturing and installation. Early, minor damage or defects are difficult to detect, leading to reduced aerodynamic efficiency and shortened service life. More seriously, as wind turbine blades fatigue and degrade, cracks appear on the blade surface and propagate rapidly, significantly reducing the reliability and safety performance of the wind turbine structure, ultimately causing severe economic losses and safety accidents. Therefore, real-time and rapid monitoring of its health status is crucial. Among these, structural strain field reconstruction and damage identification are key issues in real-time structural health monitoring systems.
[0003] However, traditional damage identification methods based on vibration signal analysis are inefficient. Achieving accurate detection typically requires a large amount of measurement data and modal information before and after damage, leading to complex models, time-consuming analysis, and hindering real-time online detection. Furthermore, they lack sensitivity for detecting early, minor damage. This is particularly true for wind turbine blade structures, whose complex operating environment and conditions, such as wind speed, blade rotation speed, temperature, and environmental load, can significantly affect the blade's dynamic characteristics (natural frequencies, mode shapes, and damping). This makes it difficult to distinguish between vibrations caused by damage and those caused by environmental and operational conditions, requiring more complex algorithms, further increasing analysis time and failing to meet the requirements of real-time, rapid monitoring. Therefore, in situations where modal information before wind turbine blade damage is lacking or the material properties and initial damage / crack location are unknown, reconstructing the displacement / strain field of the wind turbine blade structure by monitoring limited strain information, and then accurately detecting internal material damage, predicting crack length and its dynamic propagation path until overall material failure, is a crucial problem to be solved. Summary of the Invention
[0004] To address the problems existing in the background technology, this invention provides a rapid damage detection method for wind turbine blades based on inverse finite element reconstruction image recognition. This method can reconstruct the global displacement / strain field of the three-dimensional structure in real time based on local strain information measured by a small number of strain sensors arranged on the surface of the wind turbine blade structure. Furthermore, it proposes for the first time to use the inverse finite element strain field reconstruction image for training the convolutional neural network image recognition algorithm model for damage detection, thus providing an end-to-end damage detection method for wind turbine blades with high detection efficiency. Simultaneously, it provides a method for calibrating the parameters of the reconstructed inverse finite element strain field of wind turbine blades in real time, ensuring the accuracy of the reconstructed strain field and further guaranteeing the accuracy of the final wind turbine blade damage detection. In summary, this wind turbine blade damage detection method has a simple model and rapid analysis, effectively improving the efficiency and accuracy of existing wind turbine blade damage detection methods. Moreover, the model does not require structural material properties and load information, therefore, this damage detection method has universal applicability under complex working conditions.
[0005] The technical solution adopted in this invention is:
[0006] The rapid detection method for wind turbine blade damage of the present invention includes the following steps:
[0007] Step 1: Arrange several strain sensors evenly at intervals on the surface of the wind turbine blades. Collect strain distribution data of different locations and degrees of damage on the wind turbine blades through each strain sensor, and divide the data into training strain distribution dataset and calibration strain distribution dataset.
[0008] Step 2: Input the training strain distribution dataset into the inverse finite element strain field reconstruction algorithm model for processing, and calibrate the training strain distribution dataset using the inverse finite element strain field reconstruction parameter calibration method according to the calibration strain distribution dataset. The inverse finite element strain field reconstruction algorithm model finally outputs the wind turbine blade strain field reconstruction image set.
[0009] Step 3: Perform image preprocessing on the wind turbine blade strain field reconstruction image set to obtain a preprocessed wind turbine blade strain field reconstruction image set.
[0010] Step 4: Input the pre-processed wind turbine blade strain field reconstruction image set into the convolutional neural network image recognition algorithm model for training, and obtain the trained convolutional neural network image recognition algorithm model.
[0011] Step 5: Collect strain distribution data of the wind turbine blades when they are undamaged and real-time strain distribution data through various strain sensors. Input the strain distribution data of the wind turbine blades when they are undamaged and real-time strain distribution data into the trained convolutional neural network image recognition algorithm model for processing. After processing, output the damage location and damage degree of the current wind turbine blade to achieve rapid detection of wind turbine blade damage.
[0012] In step one, strain distribution data of different positions and degrees of damage to the wind turbine blades are collected by various strain sensors and divided into training strain distribution datasets and calibration strain distribution datasets. Specifically, strain distribution data of the wind turbine blades at the same position and degree of damage are collected by each strain sensor each time. For the strain distribution data collected by each strain sensor each time, a portion of the strain distribution data collected by each strain sensor each time is divided into training strain distribution datasets, and the other portion of the strain distribution data collected by each strain sensor each time is divided into calibration strain distribution datasets.
[0013] The training strain distribution dataset includes strain distribution data collected by a portion of the strain sensors, while the calibration strain distribution dataset includes strain distribution data collected by another portion of the strain sensors. Each strain distribution dataset includes strain in the x-axis direction, strain in the y-axis direction, and shear strain in the xy-axis direction.
[0014] In step two, the training strain distribution dataset and the wind turbine blade structure type are input into the inverse finite element strain field reconstruction algorithm model for processing. The training strain distribution dataset is then calibrated using the inverse finite element strain field reconstruction parameter calibration method based on the calibration strain distribution dataset. Specifically, for each strain sensor in the training strain distribution dataset and each strain sensor in the calibration strain distribution dataset, the strain distribution data collected each time is input into the inverse finite element strain field reconstruction algorithm model for processing. After processing, the preprocessed wind turbine blade strain field reconstruction image of the strain distribution data reconstructed by the strain parameters is output, and the first weighting coefficient w of the current inverse finite element strain field reconstruction algorithm model is obtained. e Second weighting coefficient w k and the third weighting coefficient w g The preprocessed wind turbine blade strain field reconstruction image includes reconstructed strain parameters. The error between the reconstructed strain parameters of the preprocessed wind turbine blade strain field reconstruction image and the strain parameters of the strain distribution data acquired by the strain sensor in the calibration strain distribution dataset is calculated. When the error is greater than 2%, the first weighting coefficient w of the current inverse finite element strain field reconstruction algorithm model is adjusted using an adaptive weight adjustment algorithm. e Second weighting coefficient w k and the third weighting coefficient w gThe value is obtained by inputting the strain distribution data collected by each strain sensor in the training strain distribution dataset into the adjusted inverse finite element strain field reconstruction algorithm model and repeating step two to calibrate the training strain distribution dataset until the error obtained after multiple iterations is less than or equal to 2%. The calibration is then completed, and the preprocessed wind turbine blade strain field reconstruction image at this time is obtained as the final wind turbine blade strain field reconstruction image.
[0015] The reconstructed strain field images of each wind turbine blade constitute the wind turbine blade strain field reconstructed image set.
[0016] Each strain sensor in the training strain distribution dataset collects strain distribution data each time, and then performs strain field parameter reconstruction processing in the inverse finite element strain field reconstruction algorithm model. First, the global displacement field reconstruction is obtained. The global displacement field reconstruction is then transformed by the strain-displacement transformation matrix to obtain the global strain field reconstruction. Finally, the strain distribution data is output as a wind turbine blade strain field reconstruction image.
[0017] The core of the inverse finite element strain field reconstruction algorithm is based on minimizing the least squares error function between the constructed theoretical strain and the actual measured strain, using weighting coefficients w. e w k w g The weighting coefficients control the degree of consistency between theoretical strain and actual measured strain. When there are relatively few or no strain sensors, the value of the weighting coefficients is particularly important. However, in the current inverse finite element strain field reconstruction algorithm, the weighting coefficients are usually simply taken as a small constant value, which does not have the ability to be dynamically adjusted and has certain limitations.
[0018] The error between the reconstructed strain parameters of the preprocessed wind turbine blade strain field reconstruction image and the strain parameters of the strain distribution data collected by the strain sensor in the calibration strain distribution dataset is as follows:
[0019]
[0020] Where, ε iFEM ε represents the reconstructed strain parameter in the preprocessed wind turbine blade strain field reconstruction image. exp This represents the strain parameter in the strain distribution data acquired by the strain sensor in the current test within the calibration strain distribution dataset.
[0021] The reconstructed strain parameters in the preprocessed wind turbine blade strain field reconstruction image specifically include the reconstructed strain in the x-axis direction, the reconstructed strain in the y-axis direction, and the shear reconstructed strain in the xy-axis direction. The strain parameters of the strain distribution data collected by the strain sensor in the calibration strain distribution dataset include the strain in the x-axis direction, the strain in the y-axis direction, and the shear strain in the xy-axis direction. The x-axis direction error, y-axis direction error, and xy-axis direction error are calculated between the reconstructed strain parameters in the preprocessed wind turbine blade strain field reconstruction image and the strain parameters of the strain distribution data collected by the strain sensor in the calibration strain distribution dataset. The calibration is completed when the x-axis direction error, y-axis direction error, and xy-axis direction error are all less than or equal to 2%.
[0022] In step three, the calibrated wind turbine blade strain field reconstruction image set is preprocessed to obtain a preprocessed wind turbine blade strain field reconstruction image set. For each calibrated wind turbine blade strain field reconstruction image in the calibrated wind turbine blade strain field reconstruction image set, the calibrated wind turbine blade strain field reconstruction image is normalized and Gaussian noise is added to obtain a preprocessed wind turbine blade strain field reconstruction image. All the preprocessed wind turbine blade strain field reconstruction images constitute the preprocessed wind turbine blade strain field reconstruction image set.
[0023] First, the image is normalized to a range of 0 to 1 to accelerate the training and convergence speed of the convolutional neural network image recognition algorithm model in step four. Then, Gaussian noise is added to improve the model's ability to learn mapping rules from the input space, as well as its generalization ability and fault tolerance.
[0024] In step five, the strain distribution data of the undamaged wind turbine blade and the real-time strain distribution data are input into the trained convolutional neural network image recognition algorithm model for processing. After processing, the real-time damage location and damage degree of the current wind turbine blade are output. For each detection location of the wind turbine blade, the real-time damage degree D at the detection location is as follows:
[0025]
[0026] Where, ε vm，c ε represents the real-time equivalent strain distribution on the surface at the detection location of the wind turbine blade. vm，u This represents the equivalent strain distribution on the surface at the detection location when the wind turbine blade is undamaged.
[0027] When the real-time damage level D at the detection location of the wind turbine blade is 0, it indicates that there is no damage at the detection location. When the real-time damage level D at the detection location of the wind turbine blade is greater than 0, it indicates that damage has occurred at the detection location. The trained convolutional neural network image recognition algorithm model outputs the damage location coordinates and real-time damage level D at the detection location of the wind turbine blade.
[0028] The real-time equivalent strain distribution ε of the surface at the detection location vm，c Equivalent strain distribution ε of the surface at the detection location when the wind turbine blade is undamaged vm，u Specifically as follows:
[0029]
[0030]
[0031] Where, ε xx，c and ε xx，u ε represents the real-time strain along the x-axis at the detection location of the wind turbine blade and the strain along the x-axis at the detection location when the wind turbine blade is undamaged, respectively. yy，c and ε yy，u ε represents the real-time y-axis strain at the detection location of the wind turbine blade and the y-axis strain at the detection location when the wind turbine blade is undamaged, respectively. zz，c and ε zz，u γ represents the real-time strain along the z-axis at the detection location of the wind turbine blade and the strain along the z-axis at the detection location when the wind turbine blade is undamaged, respectively. xy，c and γ xy，u These represent the real-time shear strain along the x and y axes at the detection location of the wind turbine blade, and the shear strain along the x and y axes at the detection location when the wind turbine blade is undamaged, respectively.
[0032] The real-time strain ε in the z-axis direction at the detection position of the wind turbine blade. zz，c And the strain ε in the z-axis direction at the detection location when the wind turbine blades are undamaged. zz，u Specifically as follows:
[0033]
[0034]
[0035] Where v represents the Poisson's ratio of the wind turbine blade material, which is determined according to the wind turbine blade structure type.
[0036] In step two, the wind turbine blade structure types include plate shell, rod beam, and sandwich.
[0037] The beneficial effects of this invention are:
[0038] 1) This invention can reconstruct the global displacement / strain field of a three-dimensional structure in real time based on the local strain information measured by a small number of strain sensors arranged on the surface of the wind turbine blade structure. It also proposes for the first time to use the inverse finite element strain field reconstruction image for damage detection training of the convolutional neural network image recognition algorithm model, thereby providing an end-to-end damage detection method for wind turbine blades with high detection efficiency.
[0039] 2) This invention provides a method for calibrating parameters of the reconstructed strain field of a wind turbine blade in reverse finite element method to calibrate the reconstructed strain field parameters in real time, thereby ensuring the accuracy of the reconstructed strain field of the wind turbine blade in reverse finite element method, and further ensuring the accuracy of the final wind turbine blade damage detection.
[0040] 3) The model of this invention is simple and the analysis is fast, which can effectively improve the efficiency and accuracy of existing wind turbine blade damage detection. Moreover, the model does not require the use of structural material properties and load information. Therefore, this damage detection method has universal applicability under complex working conditions. Attached Figure Description
[0041] Figure 1 This is a flowchart of the rapid detection method for wind turbine blade damage of the present invention;
[0042] Figure 2 This is a diagram of the inverse finite element model architecture of the wind turbine blades of this invention;
[0043] Figure 3 This is a flowchart of the method for reconstructing parameters of the inverse finite element strain field of wind turbine blades according to the present invention;
[0044] Figure 4 (a) is a schematic diagram of the inverse finite element mesh generation and strain sensor arrangement on the wind turbine blades of the present invention;
[0045] Figure 4 (b) is a schematic diagram of different damage locations on the wind turbine blades of the present invention;
[0046] Figure 4 (c) is a schematic diagram of different degrees of damage on the wind turbine blades of the present invention. Detailed Implementation
[0047] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0048] like Figure 1 As shown, the rapid detection method for wind turbine blade damage of the present invention includes the following steps:
[0049] Step 1: Arrange several strain sensors evenly at intervals on the surface of the wind turbine blades. Collect strain distribution data of different locations and degrees of damage on the wind turbine blades through each strain sensor, and divide the data into training strain distribution dataset and calibration strain distribution dataset.
[0050] In step one, strain distribution data of different locations and degrees of damage to the wind turbine blades are collected by various strain sensors and divided into training strain distribution datasets and calibration strain distribution datasets. Specifically, strain distribution data of the wind turbine blades at the same location and degree of damage are collected by each strain sensor each time. For the strain distribution data collected by each strain sensor each time, a portion of the strain distribution data collected by each strain sensor each time is divided into training strain distribution datasets, and the other portion of the strain distribution data collected by each strain sensor each time is divided into calibration strain distribution datasets.
[0051] The training strain distribution dataset includes strain distribution data collected by a portion of the strain sensors, while the calibration strain distribution dataset includes strain distribution data collected by another portion of the strain sensors. Each strain distribution dataset includes strain in the x-axis direction, strain in the y-axis direction, and shear strain in the xy-axis direction.
[0052] Step 2: Input the training strain distribution dataset into the inverse finite element strain field reconstruction algorithm model for processing. Then, calibrate the training strain distribution dataset using the inverse finite element strain field reconstruction parameter calibration method based on the calibration strain distribution dataset. The inverse finite element strain field reconstruction algorithm model ultimately outputs a set of reconstructed strain field images of the wind turbine blades. Wind turbine blade structure types include plate shells, struts, and sandwich structures.
[0053] like Figure 2 and Figure 3 As shown, in step two, the training strain distribution dataset and the wind turbine blade structure type are input into the inverse finite element strain field reconstruction algorithm model for processing. The training strain distribution dataset is then calibrated using the inverse finite element strain field reconstruction parameter calibration method based on the calibration strain distribution dataset. Specifically, for each strain sensor in the training strain distribution dataset and each strain sensor in the calibration strain distribution dataset, the strain distribution data collected each time is input into the inverse finite element strain field reconstruction algorithm model for processing. After processing, the preprocessed wind turbine blade strain field reconstruction image of the strain distribution data reconstructed by the strain parameters is output, and the first weighting coefficient w of the current inverse finite element strain field reconstruction algorithm model is obtained. e Second weighting coefficient w k and the third weighting coefficient w g The preprocessed wind turbine blade strain field reconstruction image includes reconstructed strain parameters. The error between the reconstructed strain parameters of the preprocessed wind turbine blade strain field reconstruction image and the strain parameters of the strain distribution data acquired by the strain sensor in the calibration strain distribution dataset is calculated. When the error is greater than 2%, the first weighting coefficient w of the current inverse finite element strain field reconstruction algorithm model is adjusted using an adaptive weight adjustment algorithm.e Second weighting coefficient w k and the third weighting coefficient w g The value is obtained by inputting the strain distribution data collected by each strain sensor in the training strain distribution dataset into the adjusted inverse finite element strain field reconstruction algorithm model and repeating step two to calibrate the training strain distribution dataset until the error obtained after multiple iterations is less than or equal to 2%. The calibration is then completed, and the preprocessed wind turbine blade strain field reconstruction image at this time is obtained as the final wind turbine blade strain field reconstruction image.
[0054] The reconstructed strain field images of each wind turbine blade constitute the wind turbine blade strain field reconstructed image set.
[0055] Each strain sensor in the training strain distribution dataset collects strain distribution data each time, and then performs strain field parameter reconstruction processing in the inverse finite element strain field reconstruction algorithm model. First, the global displacement field reconstruction is obtained. The global displacement field reconstruction is then transformed by the strain-displacement transformation matrix to obtain the global strain field reconstruction. Finally, the strain distribution data is output as a wind turbine blade strain field reconstruction image.
[0056] The core of the inverse finite element strain field reconstruction algorithm is based on minimizing the least squares error function between the constructed theoretical strain and the actual measured strain, using weighting coefficients w. e w k w g The weighting coefficients control the degree of consistency between theoretical strain and actual measured strain. When there are relatively few or no strain sensors, the value of the weighting coefficients is particularly important. However, in the current inverse finite element strain field reconstruction algorithm, the weighting coefficients are usually simply taken as a small constant value, which does not have the ability to be dynamically adjusted and has certain limitations.
[0057] The error between the reconstructed strain parameters of the preprocessed wind turbine blade strain field reconstruction image and the strain parameters of the strain distribution data acquired by the strain sensor in the current acquisition is calculated as follows:
[0058]
[0059] Where, ε iFEM ε represents the reconstructed strain parameter in the preprocessed wind turbine blade strain field reconstruction image. exp This represents the strain parameter in the strain distribution data acquired by the strain sensor in the current iteration of the calibration strain distribution dataset;
[0060] The reconstructed strain parameters in the preprocessed wind turbine blade strain field reconstruction image specifically include the reconstructed strain in the x-axis direction, the reconstructed strain in the y-axis direction, and the shear reconstructed strain in the xy-axis direction. The strain parameters of the strain distribution data collected by the strain sensor in the calibration strain distribution dataset include the strain in the x-axis direction, the strain in the y-axis direction, and the shear strain in the xy-axis direction. The x-axis direction error, y-axis direction error, and xy-axis direction error are calculated between the reconstructed strain parameters in the preprocessed wind turbine blade strain field reconstruction image and the strain parameters of the strain distribution data collected by the strain sensor in the calibration strain distribution dataset. The calibration is completed when the x-axis direction error, y-axis direction error, and xy-axis direction error are all less than or equal to 2%.
[0061] Step 3: Perform image preprocessing on the wind turbine blade strain field reconstruction image set to obtain a preprocessed wind turbine blade strain field reconstruction image set.
[0062] In step three, the calibrated wind turbine blade strain field reconstruction image set is preprocessed to obtain a preprocessed wind turbine blade strain field reconstruction image set. For each calibrated wind turbine blade strain field reconstruction image in the calibrated wind turbine blade strain field reconstruction image set, the calibrated wind turbine blade strain field reconstruction image is normalized and Gaussian noise is added to obtain a preprocessed wind turbine blade strain field reconstruction image. All the preprocessed wind turbine blade strain field reconstruction images constitute the preprocessed wind turbine blade strain field reconstruction image set.
[0063] First, the image is normalized to a range of 0 to 1 to accelerate the training and convergence speed of the convolutional neural network image recognition algorithm model in step four. Then, Gaussian noise is added to improve the model's ability to learn mapping rules from the input space, as well as its generalization ability and fault tolerance.
[0064] Step 4: Input the pre-processed wind turbine blade strain field reconstruction image set into the convolutional neural network image recognition algorithm model for training, and obtain the trained convolutional neural network image recognition algorithm model.
[0065] Step 5: Collect strain distribution data of the wind turbine blades when they are undamaged and real-time strain distribution data through various strain sensors. Input the strain distribution data of the wind turbine blades when they are undamaged and real-time strain distribution data into the trained convolutional neural network image recognition algorithm model for processing. After processing, output the damage location and damage degree of the current wind turbine blade to achieve rapid detection of wind turbine blade damage.
[0066] In step five, the strain distribution data of the undamaged wind turbine blade and the real-time strain distribution data are input into the trained convolutional neural network image recognition algorithm model for processing. After processing, the real-time damage location and damage degree of the current wind turbine blade are output. For each detection location of the wind turbine blade, the real-time damage degree D at the detection location is as follows:
[0067]
[0068] Where, ε vm，c ε represents the real-time equivalent strain distribution on the surface at the detection location of the wind turbine blade. vm，u This represents the equivalent strain distribution on the surface at the detection location when the wind turbine blade is undamaged.
[0069] When the real-time damage level D at the detection location of the wind turbine blade is 0, it indicates that there is no damage at the detection location. When the real-time damage level D at the detection location of the wind turbine blade is greater than 0, it indicates that damage has occurred at the detection location. The trained convolutional neural network image recognition algorithm model outputs the damage location coordinates and real-time damage level D at the detection location of the wind turbine blade.
[0070] Real-time equivalent strain distribution ε of the surface at the detection location vm，c Equivalent strain distribution ε of the surface at the detection location when the wind turbine blade is undamaged vm，u Specifically as follows:
[0071]
[0072]
[0073] Where, ε xx，c and ε xx，u ε represents the real-time strain along the x-axis at the detection location of the wind turbine blade and the strain along the x-axis at the detection location when the wind turbine blade is undamaged, respectively. yy，c and ε yy，u ε represents the real-time y-axis strain at the detection location of the wind turbine blade and the y-axis strain at the detection location when the wind turbine blade is undamaged, respectively. zz，c and ε zz，u γ represents the real-time strain along the z-axis at the detection location of the wind turbine blade and the strain along the z-axis at the detection location when the wind turbine blade is undamaged, respectively. xy，c and γ xy，u These represent the real-time shear strain along the x and y axes at the detection location of the wind turbine blade, and the shear strain along the x and y axes at the detection location when the wind turbine blade is undamaged, respectively.
[0074] Real-time strain ε in the z-axis direction at the detection location of the wind turbine blade zz，cAnd the strain ε in the z-axis direction at the detection location when the wind turbine blades are undamaged. zz，u Specifically as follows:
[0075]
[0076]
[0077] Where v represents the Poisson's ratio of the wind turbine blade material, which is determined according to the wind turbine blade structure type.
[0078] Specific embodiments of the present invention are as follows:
[0079] like Figure 4 As shown in (a), this invention takes a wind turbine blade with a length of L, a bottom width of B, and a top width of W as an example, and divides it into 14 inverse finite element meshes. The solid circles represent strain sensors (fiber optic strain sensors or strain gauges), arranged at the centroid of each inverse finite element mesh. Figure 4 As shown in (b), the solid cubes represent the damaged areas. In the first experimental configuration of the embodiment, the damage degree of the damaged areas corresponding to the six different damage locations is all set to D1; as Figure 4 As shown in (c), the solid cubes represent the damaged areas, and the size of the solid cubes represents the degree of damage. As the second experimental configuration, the degree of damage of the damaged areas at the six different damage locations is set sequentially from the bottom edge to the top edge along the length direction, namely D6, D5, D4, D3, D2, D1, and the degree of damage decreases sequentially, i.e., D6>D5>D4>D3>D2>D1.
[0080] Based on the two sets of experimental configurations mentioned above, the experimental configuration for simulating wind turbine blade damage is shown in Table 1. The six damage locations on the wind turbine blade are denoted sequentially by their distance from the bottom edge as a, a+b, a+b+c, a+b+c+d, a+b+c+d+e, and a+b+c+d+e+f. Simultaneously, damage areas with different degrees of damage are sequentially set at the corresponding damage locations. In the first experimental configuration, the damage degree at all six damage locations is set to D1. In the second experimental configuration, the damage degrees at the six damage locations are set to D6, D5, D4, D3, D2, and D1, respectively, with the damage degree decreasing sequentially, i.e., D6>D5>D4>D3>D2>D1.
[0081] Based on the two sets of experimental configurations mentioned above, the experimental configuration for simulating wind turbine blade damage is shown in Table 1. The six damage locations on the wind turbine blade are denoted sequentially by their distance from the bottom edge as a, a+b, a+b+c, a+b+c+d, a+b+c+d+e, and a+b+c+d+e+f. Simultaneously, damage areas with different degrees of damage are sequentially set at the corresponding damage locations. In the first experimental configuration, the damage degree at all six damage locations is set to D1. In the second experimental configuration, the damage degrees at the six damage locations are set to D6, D5, D4, D3, D2, and D1, respectively, with the damage degree decreasing sequentially, i.e., D6>D5>D4>D3>D2>D1.
[0082] Based on this, strain distribution data at different damage locations and damage levels can be collected by strain sensors arranged on the blade surface, and divided into training strain distribution datasets and calibration strain distribution datasets.
[0083] Table 1. Configuration of Wind Turbine Blade Damage Simulation Experiment
[0084]
[0085] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A rapid detection method for wind turbine blade damage based on inverse finite element reconstruction image recognition, characterized in that: Includes the following steps: Step 1: Arrange several strain sensors evenly at intervals on the surface of the wind turbine blades. Collect strain distribution data of different positions and degrees of damage on the wind turbine blades through each strain sensor, and divide them into training strain distribution dataset and calibration strain distribution dataset. Step 2: Input the training strain distribution dataset into the inverse finite element strain field reconstruction algorithm model for processing, and calibrate the training strain distribution dataset using the inverse finite element strain field reconstruction parameter calibration method according to the calibration strain distribution dataset. The inverse finite element strain field reconstruction algorithm model finally outputs the wind turbine blade strain field reconstruction image set. Step 3: Perform image preprocessing on the wind turbine blade strain field reconstruction image set to obtain a preprocessed wind turbine blade strain field reconstruction image set; Step 4: Input the pre-processed wind turbine blade strain field reconstruction image set into the convolutional neural network image recognition algorithm model for training, and obtain the trained convolutional neural network image recognition algorithm model; Step 5: Collect strain distribution data of the wind turbine blades when they are undamaged and real-time strain distribution data through various strain sensors. Input the strain distribution data of the wind turbine blades when they are undamaged and real-time strain distribution data into the trained convolutional neural network image recognition algorithm model for processing. After processing, output the damage location and damage degree of the current wind turbine blade to achieve rapid detection of wind turbine blade damage.
2. The method for rapid detection of wind turbine blade damage based on inverse finite element reconstruction image recognition according to claim 1, characterized in that: In step one, strain distribution data of different positions and degrees of damage to the wind turbine blades are collected by various strain sensors and divided into training strain distribution datasets and calibration strain distribution datasets. Specifically, strain distribution data of the wind turbine blades at the same position and degree of damage are collected by each strain sensor each time. For the strain distribution data collected by each strain sensor each time, a part of the strain distribution data collected by each strain sensor each time is divided into training strain distribution datasets and the other part of the strain distribution data collected by each strain sensor each time is divided into calibration strain distribution datasets. The training strain distribution dataset includes strain distribution data collected by a portion of the strain sensors, while the calibration strain distribution dataset includes strain distribution data collected by another portion of the strain sensors. Each strain distribution dataset includes strain in the x-axis direction, strain in the y-axis direction, and shear strain in the xy-axis direction.
3. The method for rapid detection of wind turbine blade damage based on inverse finite element reconstruction image recognition according to claim 2, characterized in that: In step two, the training strain distribution dataset and the wind turbine blade structure type are input into the inverse finite element strain field reconstruction algorithm model for processing. The training strain distribution dataset is then calibrated using the inverse finite element strain field reconstruction parameter calibration method based on the calibration strain distribution dataset. Specifically, for each strain sensor in the training strain distribution dataset and each strain sensor in the calibration strain distribution dataset, the strain distribution data collected each time is input into the inverse finite element strain field reconstruction algorithm model for processing. After processing, the preprocessed wind turbine blade strain field reconstruction image of the strain distribution data reconstructed by the strain parameters is output, and the first weighting coefficient w of the current inverse finite element strain field reconstruction algorithm model is obtained. e Second weighting coefficient w k and the third weighting coefficient w g The preprocessed wind turbine blade strain field reconstruction image includes reconstructed strain parameters. The error between the reconstructed strain parameters of the preprocessed wind turbine blade strain field reconstruction image and the strain parameters of the strain distribution data acquired by the strain sensor in the calibration strain distribution dataset is calculated. When the error is greater than 2%, the first weighting coefficient w of the current inverse finite element strain field reconstruction algorithm model is adjusted using an adaptive weight adjustment algorithm. e Second weighting coefficient w k and the third weighting coefficient w g The value is obtained by inputting the strain distribution data collected by each strain sensor in the training strain distribution dataset into the adjusted inverse finite element strain field reconstruction algorithm model and repeating step two to calibrate the training strain distribution dataset until the error obtained after multiple iterations is less than or equal to 2% to complete the calibration. The preprocessed wind turbine blade strain field reconstruction image at this time is then obtained as the final wind turbine blade strain field reconstruction image. The reconstructed strain field images of each wind turbine blade constitute the wind turbine blade strain field reconstructed image set.
4. The method for rapid detection of wind turbine blade damage based on inverse finite element reconstruction image recognition according to claim 3, characterized in that: The error between the reconstructed strain parameters of the preprocessed wind turbine blade strain field reconstruction image and the strain parameters of the strain distribution data collected by the strain sensor in the calibration strain distribution dataset is as follows: Where, ε iFEM ε represents the reconstructed strain parameter in the preprocessed wind turbine blade strain field reconstruction image. exp This represents the strain parameter in the strain distribution data acquired by the strain sensor in the current iteration of the calibration strain distribution dataset; The reconstructed strain parameters in the preprocessed wind turbine blade strain field reconstruction image specifically include the reconstructed strain in the x-axis direction, the reconstructed strain in the y-axis direction, and the shear reconstructed strain in the xy-axis direction. The strain parameters of the strain distribution data collected by the strain sensor in the calibration strain distribution dataset include the strain in the x-axis direction, the strain in the y-axis direction, and the shear strain in the xy-axis direction. The x-axis direction error, y-axis direction error, and xy-axis direction error are calculated between the reconstructed strain parameters in the preprocessed wind turbine blade strain field reconstruction image and the strain parameters of the strain distribution data collected by the strain sensor in the calibration strain distribution dataset. The calibration is completed when the x-axis direction error, y-axis direction error, and xy-axis direction error are all less than or equal to 2%.
5. The method for rapid detection of wind turbine blade damage based on inverse finite element reconstruction image recognition according to claim 1, characterized in that: In step three, the calibrated wind turbine blade strain field reconstruction image set is preprocessed to obtain a preprocessed wind turbine blade strain field reconstruction image set. For each calibrated wind turbine blade strain field reconstruction image in the calibrated wind turbine blade strain field reconstruction image set, the calibrated wind turbine blade strain field reconstruction image is normalized and Gaussian noise is added to obtain a preprocessed wind turbine blade strain field reconstruction image. All the preprocessed wind turbine blade strain field reconstruction images constitute the preprocessed wind turbine blade strain field reconstruction image set.
6. The method for rapid detection of wind turbine blade damage based on inverse finite element reconstruction image recognition according to claim 4, characterized in that: In step five, the strain distribution data of the undamaged wind turbine blade and the real-time strain distribution data are input into the trained convolutional neural network image recognition algorithm model for processing. After processing, the real-time damage location and damage degree of the current wind turbine blade are output. For each detection location of the wind turbine blade, the real-time damage degree D at the detection location is as follows: Where, ε vm，c ε represents the real-time equivalent strain distribution on the surface at the detection location of the wind turbine blade. vm，u This represents the equivalent strain distribution on the surface at the detection location when the wind turbine blade is undamaged; When the real-time damage level D at the detection location of the wind turbine blade is 0, it indicates that there is no damage at the detection location. When the real-time damage level D at the detection location of the wind turbine blade is greater than 0, it indicates that damage has occurred at the detection location. The trained convolutional neural network image recognition algorithm model outputs the damage location coordinates and real-time damage level D at the detection location of the wind turbine blade.
7. The method for rapid detection of wind turbine blade damage based on inverse finite element reconstruction image recognition according to claim 6, characterized in that: The real-time equivalent strain distribution ε of the surface at the detection location vm，c Equivalent strain distribution ε of the surface at the detection location when the wind turbine blade is undamaged vm，u Specifically as follows: Where, ε xx，c and ε xx，u ε represents the real-time strain along the x-axis at the detection location of the wind turbine blade and the strain along the x-axis at the detection location when the wind turbine blade is undamaged, respectively. yy，c and ε yy，u ε represents the real-time y-axis strain at the detection location of the wind turbine blade and the y-axis strain at the detection location when the wind turbine blade is undamaged, respectively. zz，c and ε zz，u γ represents the real-time strain along the z-axis at the detection location of the wind turbine blade and the strain along the z-axis at the detection location when the wind turbine blade is undamaged, respectively. xy，c and γ xy，u These represent the real-time shear strain along the x and y axes at the detection location of the wind turbine blade, and the shear strain along the x and y axes at the detection location when the wind turbine blade is undamaged, respectively.
8. The method for rapid detection of wind turbine blade damage based on inverse finite element reconstruction image recognition according to claim 7, characterized in that: The real-time strain ε in the z-axis direction at the detection position of the wind turbine blade. zz，c And the strain ε in the z-axis direction at the detection location when the wind turbine blades are undamaged. zz，u Specifically as follows: Where v represents the Poisson's ratio of the material of the wind turbine blade.
9. The method for rapid detection of wind turbine blade damage based on inverse finite element reconstruction image recognition according to claim 1, characterized in that: In step two, the wind turbine blade structure types include plate shell, rod beam, and sandwich.