Offshore wind turbine blade damage quantitative monitoring method based on transfer learning

Abstract

The invention belongs to the field of wind power equipment intelligent structure health monitoring, and particularly discloses an offshore wind turbine blade damage quantitative monitoring method based on transfer learning, which comprises the following steps of guided wave data acquisition, coarse graining feature extraction, cross-domain feature alignment, damage quantitative model construction and damage size output. Target domain labeling information is not needed, coarse graining features are combined, the sample labeling and detection calculation cost is greatly reduced, and the method is adaptive to an existing guided wave collection system; the generalization ability is enhanced through cross-domain feature alignment and multi-scale model fusion, complex marine environment interference can be resisted, and cross-regional stable monitoring is achieved; damage can be accurately quantified, scientific operation and maintenance are supported, fault loss and excessive maintenance are reduced, the operation reliability of the wind turbine generator is improved, and remarkable engineering value and economic benefits are achieved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent structural health monitoring of wind power equipment, specifically involving a quantitative monitoring method for damage to offshore wind turbine blades based on transfer learning. Background Technology

[0002] As a core area of ​​global renewable energy development, offshore wind power, with its abundant resource reserves, high annual generating hours, and zero land space occupation, has become a core engine driving technological innovation and large-scale development in wind power. Wind turbine blades, as key components for wind energy capture and conversion, directly determine the overall power generation efficiency and operational safety of the turbine, serving as the core power source for offshore wind turbines. However, offshore wind turbine blades are exposed to the complex marine environment of high humidity, high salt spray, and strong wind loads for extended periods, making them highly susceptible to environmental corrosion, fatigue load impacts, and other factors, leading to various types of damage such as delamination, cracks, and fiber breakage. Such damage not only causes aerodynamic performance degradation and reduced power generation efficiency but can also trigger catastrophic accidents such as blade breakage, resulting in tens of millions of dollars in economic losses and even casualties. Therefore, achieving real-time monitoring, accurate quantification, and early warning of blade structural damage is a crucial requirement for ensuring reliable operation of offshore wind turbines, reducing maintenance costs, and improving the economic benefits throughout their entire lifecycle.

[0003] Traditional manual blade inspection methods rely on the subjective experience of inspectors, resulting in low efficiency, high cost, and difficulty in covering the interior and hidden areas of the blade, failing to meet the inspection needs of large-scale wind power projects. While conventional non-destructive testing (NDT) technologies can achieve some non-destructive testing, the equipment deployment cost is high, they are poorly adaptable to blades with complex curved surfaces, and they cannot achieve online real-time monitoring, making it difficult to cope with the damage evolution during the dynamic service process of blades. Currently, the mainstream solution for existing structural health monitoring technology is an online monitoring mode that integrates piezoelectric sensors on complex structures and uses Lamb waves as the damage information transmission medium. Although it has been applied to the full life cycle monitoring of complex structures, it still faces two major bottlenecks for offshore wind turbine blades. First, the complex damage forms of blade composite materials and the strong interference characteristics of the marine environment lead to low monitoring accuracy and insufficient robustness. Second, blade damage samples are extremely scarce. There are few damage cases in actual service, and the annotation cost is high. The number of samples available for model training is usually less than a hundred, which cannot meet the dependence of traditional supervised learning on big data, resulting in poor model generalization ability and easy overfitting or underfitting problems when monitoring damage across regions and operating conditions.

[0004] Therefore, it is essential to develop a new method for quantitative monitoring of damage to offshore wind turbine blades. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for quantitative monitoring of damage to offshore wind turbine blades based on transfer learning. In the case of extremely limited sample availability, the invention uses Lamb waves, a special guided wave, as the main transmission medium for structural damage information, thus solving the problem of low accuracy in damage quantification under the dual constraints of "small sample constraints and large-scale structures".

[0006] As an emerging machine learning paradigm, transfer learning's core mechanism lies in accurately characterizing and fully utilizing the potential and complex relationships between the target task and the source task, thereby efficiently extracting transfer knowledge that can be shared across domains, with the aim of significantly improving the overall performance of structural health monitoring tasks in the target domain.

[0007] To achieve the above objectives, the technical solution of the present invention is: a method for quantitative monitoring of damage to offshore wind turbine blades based on transfer learning, comprising the following steps:

[0008] S1: Guided wave data acquisition: Use guided wave data acquisition equipment to acquire Lamb wave signals at 3-7 different damage sizes with damage sizes of 15-35mm. Acquire 15-25 damage samples for each damage size to obtain the source domain dataset; then randomly acquire Lamb wave signals at other locations at the above damage sizes. Considering the sparsity of the damage distribution, during the random acquisition of damage samples, acquire 5-10 damage samples for each damage size to serve as the target domain dataset.

[0009] S2: Coarse-grained feature extraction: The source domain dataset and target domain dataset obtained in step S1 are subjected to coarse-grained feature extraction. First, the original waveform is segmented by a multi-scale sliding window with lengths of 10, 100, and 1000, which covers three orders of magnitude to capture the transient local fluctuations, intermediate transition dynamics, and long-term evolution trend of the signal. This can effectively represent the multi-scale temporal characteristics of the signal with low computational complexity and avoid feature redundancy between adjacent scales. Then, mean features are extracted in each of the above scales. Finally, a cross-scale coarse-grained source domain feature representation and target domain feature representation are constructed to provide concise and robust damage features for subsequent damage quantification.

[0010] S3: Cross-domain feature alignment: The coarse-grained source domain features and target domain features obtained in step S2 are aligned in subspace. Under this alignment framework, a second-order statistical alignment strategy is adopted. By calculating and comparing the covariance matrices of the source domain and the target domain, and using the covariance matrix as a metric, the second-order statistical differences between the source domain and the target domain are systematically reduced. At the same time, this alignment process does not depend on any label information of the target domain. Even when the target domain label is missing, the alignment strategy can still effectively reduce the distribution offset between domains, thereby narrowing the spatial distance between the source domain and the target domain in the feature space.

[0011] S4: Damage Quantization Model Construction: The source domain features obtained in step S3 after subspace alignment are normalized and simultaneously concatenated with the damage quantization labels corresponding to the feature samples. Then, the normalized features and their labels are input into a Transformer neural network based on a self-attention mechanism to train a damage quantization model with high generalization ability in an end-to-end manner. At the same time, the coarse-grained transferable source domain features obtained by sliding windows at different scales in step S2 are input into the Transformer neural network to obtain the damage quantization models at the corresponding scales.

[0012] S5: Damage Size Output: The subspace-aligned target domain features obtained in step S3 are normalized. Then, the normalized target domain data are sequentially and independently input into each Transformer neural network trained at different scales in step S4 to obtain the independent damage size prediction value output by the model at each scale. Then, a weighted average fusion operation is performed on the damage size estimation results given by all scale models to integrate multi-scale information. Finally, the weighted fusion value is output as the quantitative result of the target structure damage.

[0013] In a preferred embodiment of the present invention, step S2, which involves constructing coarse-grained source domain feature representations and target domain feature representations across scales, comprises the following steps:

[0014] (1) Signal sequence setting: The signal sequence is (L is the total length of the signal);

[0015] (2) Multi-scale sliding window setting: The signal X is segmented into non-overlapping multi-scale segments according to the window length n, resulting in M ​​sub-segments:

[0016] ;

[0017] (3) Multi-scale signal mean calculation: For each segment of signal Calculate the mean:

[0018]

[0019] (4) Construction of coarse-grained feature vectors: Construct a coarse-grained feature matrix from the mean features at all scales:

[0020]

[0021] In a preferred embodiment of the present invention, the subspace alignment mechanism in step S3 can significantly enhance the domain invariance between source domain features and target domain features, and simultaneously improve the transferability of the two types of features.

[0022] In a preferred embodiment of the present invention, the processing step S3 is as follows: the samples of the source domain and the target domain are adaptively aligned based on the CORAL transformation, as shown in the following formula:

[0023]

[0024] This formula yields the aligned feature map samples, where... yes The transpose of the matrix, It is the Frobenius norm of the matrix.

[0025] In a preferred embodiment of the present invention, the specific processing steps of steps S4 and S5 are as follows: the Transformer neural network used in the single-scale model training consists of an input layer, a position encoding layer, an addition layer, two self-attention layers (causal attention and global attention layers), an index layer, a fully connected layer, and a regression layer; the damage quantification results obtained by the single-scale Transformer neural network model are respectively... , where h is the number of models.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. This invention does not require target domain annotation information, and by combining coarse-grained features, it significantly reduces the cost of sample annotation and detection calculations, and is compatible with existing guided wave acquisition systems;

[0028] 2. This invention enhances generalization ability through cross-domain feature alignment and multi-scale model fusion, which can resist interference from complex marine environments and achieve stable cross-regional monitoring;

[0029] 3. This invention can accurately quantify damage to support scientific operation and maintenance, reduce failure losses and over-maintenance, improve the operational reliability of wind turbine units, and has significant engineering value and economic benefits. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the technical process of the present invention.

[0031] Figure 2This invention relates to a wind turbine blade guided wave data acquisition system.

[0032] Figure 3 This invention uses microwave absorbing adhesive to simulate five types of damage.

[0033] Figure 4 This is a schematic diagram of the sensor network setup of the present invention;

[0034] Figure 5 This is a root mean square error map of the target domain data of the present invention with different degrees of damage.

[0035] Figure 6 This is a mean absolute error diagram of different degrees of damage to the target domain data of this invention;

[0036] Figure 7 This is a relative error distribution diagram of different degrees of damage to the target domain data of this invention. Detailed Implementation

[0037] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in more detail below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited to these embodiments.

[0038] Example 1

[0039] A quantitative monitoring method for damage to offshore wind turbine blades based on transfer learning, specifically including the following steps:

[0040] (1) Guided wave data acquisition: The data used in this invention comes from, for example Figure 2 The wind turbine blade guided wave data acquisition system shown uses Lamb waves (elastic waves propagating in thin plates / layered media) as the primary medium for transmitting structural damage information. In the experiment, because the changes in structural stiffness, wave energy absorption characteristics, and ultrasonic guided wave propagation parameters (such as phase shift and amplitude attenuation) caused by the absorbing adhesive when it is applied to the structural surface are similar to the actual damage effects, the absorbing adhesive is used to simulate structural damage. Figure 3 As shown. Five damage propagation modes were set, corresponding to damage diameters of 15mm, 20mm, 25mm, 30mm, and 35mm, respectively. The data acquisition system is equipped with 28 PZT piezoelectric sensors, which are bonded to a complex-structured fiberglass sandwich panel using Hysol EA 9396 adhesive. The sandwich material is mainly composed of synthetic polymer foam or high-performance wood. The sensor network is deployed as follows... Figure 4As shown in the figure. During the acquisition process, the PZT piezoelectric sensor acts as both an exciter and a receiver. The excitation signal is a five-cycle Hanning window modulated signal with a frequency of 12MHz and an amplitude of 120V, and the received signal has a length of 6000 sampling points. To significantly improve the signal-to-noise ratio and effectively reduce measurement noise interference, a repetitive sampling strategy is adopted: each target signal is sampled three times independently, and then the three sampled waveforms are arithmetically averaged in the time domain to obtain high-fidelity final data.

[0041] On the fiberglass sandwich panel of the wind turbine blade, source domain data and target domain data were collected in different regions to verify the cross-regional quantization effect of the method. Specifically, guided wave data acquisition equipment was used to collect Lamb wave signals under five different damage sizes in a specific region, with 20 samples collected for each damage size, totaling 100 samples, which constituted the source domain dataset. Subsequently, Lamb wave signals under five damage sizes were randomly collected in other regions; considering the sparsity of the actual damage distribution, only 7 samples were collected for each damage size, totaling 35 samples, which constituted the target domain dataset.

[0042] (2) Coarse-grained feature extraction: To fully explore the damage-sensitive features of Lamb wave signals at different time scales, this invention employs a multi-scale sliding window analysis strategy to extract coarse-grained features from the source domain dataset and the target domain dataset obtained in step S1. Specifically, three sliding windows with window lengths of 10, 100, and 1000 sampling points are set. This multi-scale segmented processing mechanism can simultaneously capture the transient response features, local mode transition features, and overall wave packet evolution features during the propagation of Lamb waves, thereby achieving a comprehensive characterization of damage information. Subsequently, the statistical mean of the signal amplitude is calculated within each window at each scale as a coarse-grained feature representation, and the features extracted from the three scales are cascaded and fused to finally construct source domain coarse-grained features and target domain coarse-grained features with cross-scale characterization capabilities. This multi-scale coarse-grained strategy reduces computational complexity by averaging features, while statistical averaging effectively suppresses random noise and improves the signal-to-noise ratio. Furthermore, cross-scale feature fusion preserves damage information at different temporal resolutions, enhancing the sensitivity of features to different types and degrees of damage. This provides a concise, robust, and efficient feature foundation for subsequent quantitative damage assessment based on transfer learning.

[0043] (3) Cross-domain feature alignment: To construct a feature subspace mapping relationship between the source and target domains, this invention employs a correlation alignment (CORAL) strategy based on second-order statistics to adaptively align the source and target domain features obtained in step S2. The core idea of ​​this strategy is to achieve inter-domain alignment by minimizing the covariance difference between the feature distributions of the source and target domains. In terms of algorithm implementation, the covariance matrices of the source and target domain features are first calculated as second-order statistical representations describing the feature distributions within their respective domains. Then, the Frobenius norm between the two covariance matrices is minimized by optimizing the objective function, thereby systematically reducing the distribution differences between domains. It is worth emphasizing that this method belongs to the category of unsupervised domain adaptation, and the alignment process does not depend on any label information of the target domain. This is of great significance for structural health monitoring scenarios with high labeling costs in practical engineering applications. Through the aforementioned subspace alignment mechanism, this module can effectively minimize the distribution offset between the source domain and the target domain in the feature space, enhance the cross-domain transferability of the model, and enable the damage quantification model trained in the source domain to effectively generalize to the target domain, providing a theoretical basis and technical support for cross-regional damage diagnosis knowledge transfer.

[0044] (4) Damage Quantization Model Construction: A damage quantization model is constructed using a strategy combining multi-scale feature fusion and deep learning. First, the source domain features that have undergone domain adaptive transformation in step S3 are normalized and preprocessed to eliminate the dimensional differences between different feature dimensions. Then, the normalized features are synchronously mapped and concatenated with the damage quantization labels that strictly correspond to each sample to construct feature-label pairs. Based on this, the preprocessed feature vectors and their corresponding labels are input into a Transformer neural network architecture based on a self-attention mechanism. The long-range dependencies between features are fully explored through an end-to-end training paradigm, thereby constructing a damage quantization model with strong generalization ability. To further improve the multi-scale perception capability of the model, a parallel processing strategy is adopted: the source domain transferable features extracted through the sliding window mechanism at different time scales are input into the Transformer network to train the corresponding damage quantization sub-models. This multi-scale modeling method can effectively capture the response features of structural damage in different time-frequency domains, significantly enhancing the model's ability to identify and quantify complex damage patterns. In terms of model architecture, an encoder paradigm with stacked multi-layer self-attention modules is adopted. Deep representation learning of sequence data is achieved through positional encoding mechanism and multi-head attention mechanism. The specific parameters are set as follows: maximum sequence length of positional encoding layer is 256, number of attention heads is 4, key-value projection dimension is 32, dimension of each head is 8, initial learning rate is 0.0001, and maximum number of training rounds is 200.

[0045] (5) Damage Scale Output: A multi-model fusion strategy is adopted to achieve accurate quantification of damage in the target domain structure. The specific implementation steps are as follows: First, the target domain features after adaptive domain transformation are standardized to ensure the consistency of feature distribution and numerical stability. Second, the standardized target domain feature vectors are input into a pre-trained multi-scale Transformer model set. These models are fully trained on the source domain data for feature representation at different time scales, and can capture the dynamic characteristics of damage evolution from multiple observation perspectives. Each scale-specific Transformer model will independently generate the corresponding damage level prediction value, forming a multi-dimensional damage assessment result set. To fully utilize the complementary advantages of multi-scale models, an adaptive weighted fusion mechanism is introduced. By optimizing the allocation of weight coefficients for each scale model, intelligent integration of heterogeneous prediction results is achieved. This fusion strategy not only considers the difference in prediction accuracy of each model under different damage modes, but also takes into account the synergistic effect of cross-scale features. Finally, the comprehensive evaluation value after weighted fusion processing will be output as the damage quantification index of the target structure, providing a reliable basis for the accurate diagnosis of structural health status. The root mean square error, mean absolute error, and relative error distribution of the target domain data with different damage levels obtained from the experiment are shown below. Figure 5-7 As shown in the quantitative analysis results, the proposed model exhibits high quantification accuracy across all damage sizes. Specifically, for 15mm and 30mm damage, the RMSE is only 1.411mm and 1.684mm, respectively, indicating excellent recognition ability for small and medium-sized damage. For 25mm damage, the RMSE is 2.015mm, maintaining high quantification accuracy as well. While the RMSEs for 20mm and 35mm damage are 3.916mm and 3.001mm, respectively, showing slight fluctuations, the overall error remains within an acceptable range. Furthermore, the MAE trend is consistent with the RMSE, further validating the model's stability in damage quantification. Furthermore, the error analysis results for all target domain training samples are shown in Table 1. The error analysis results show that the model's RMSE is 2.5780 mm and MAE is 2.1214 mm. Both values ​​are close and at a low level, indicating a small deviation between the predicted and actual damage sizes, and the quantization results have good stability. The MAPE is 9.06%, meaning the average relative prediction error is controlled within 10%, indicating that the model can maintain high relative quantization accuracy under different damage size conditions. The coefficient of determination R² reaches 0.8671, indicating a strong linear correlation between the predicted and actual values. In summary, the proposed model achieves high-precision quantification of wind turbine blade structural damage sizes using only a small number of target domain labeled samples. All error indicators are at a good level, verifying the effectiveness and reliability of this method in cross-regional damage quantification tasks, and demonstrating good engineering application prospects.

[0046] Table 1 Error analysis results for all target domain training samples

[0047] MSE (Mean Squared Error) RMSE (Root Mean Square Error) MAE (Mean Absolute Error) MAPE (Mean Absolute Percentage Error) R² (coefficient of determination) 6.6459 2.5780 2.1214 9.06% 0.8671

[0048] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for quantitative monitoring of damage to offshore wind turbine blades based on transfer learning, characterized in that, Includes the following steps: S1: Guided wave data acquisition: Use guided wave data acquisition equipment to acquire Lamb wave signals at 3-7 different damage sizes with damage size of 15-35mm. Acquire 15-25 damage samples for each damage size to obtain the source domain dataset; then randomly acquire Lamb wave signals at other locations at the above damage sizes. Acquire 5-10 damage samples for each damage size to obtain the target domain dataset. S2: Coarse-grained feature extraction: The source domain dataset and target domain dataset obtained in step S1 are subjected to coarse-grained feature extraction respectively. First, the original waveform is segmented by a multi-scale sliding window with lengths of 10, 100 and 1000 respectively. Then, mean features are extracted in each of the above scales, and finally, a cross-scale coarse-grained source domain feature representation and target domain feature representation are constructed. S3: Cross-domain feature alignment: The coarse-grained source domain features and target domain features obtained in step S2 are aligned in subspace. Under this alignment framework, a second-order statistical alignment strategy is adopted. By calculating and comparing the covariance matrices of the source domain and the target domain respectively, the second-order statistical differences between the source domain and the target domain are systematically reduced using the covariance matrix as a metric. S4: Damage Quantization Model Construction: The source domain features obtained in step S3 after subspace alignment are normalized and simultaneously concatenated with the damage quantization labels corresponding to the feature samples. Then, the normalized features and their labels are input into a Transformer neural network based on a self-attention mechanism to train a damage quantization model with high generalization ability in an end-to-end manner. At the same time, the coarse-grained transferable source domain features obtained by sliding windows at different scales in step S2 are input into the Transformer neural network to obtain the damage quantization models at the corresponding scales. S5: Damage Size Output: The subspace-aligned target domain features obtained in step S3 are normalized. Then, the normalized target domain data are sequentially and independently input into each Transformer neural network trained at different scales in step S4 to obtain the independent damage size prediction value output by the model at each scale. Then, a weighted average fusion operation is performed on the damage size estimation results given by all scale models to integrate multi-scale information. Finally, the weighted fusion value is output as the quantitative result of the target structure damage.

2. The method for quantitative monitoring of offshore wind turbine blade damage based on transfer learning as described in claim 1, characterized in that, The steps in step S2 to construct coarse-grained source domain feature representations and target domain feature representations across scales are as follows: (1) Signal sequence setting: The signal sequence is (L is the total length of the signal); (2) Multi-scale sliding window setting: The signal X is segmented into non-overlapping multi-scale segments according to the window length n, resulting in M ​​sub-segments: ； (3) Multi-scale signal mean calculation: For each segment of signal Calculate the mean: ； (4) Construction of coarse-grained feature vectors: Construct a coarse-grained feature matrix from the mean features at all scales: 。 3. The method for quantitative monitoring of offshore wind turbine blade damage based on transfer learning as described in claim 1, characterized in that, The subspace alignment mechanism in step S3 can significantly enhance the domain invariance between coarse-grained source domain features and target domain features, and simultaneously improve the transferability of the two types of features.

4. The method for quantitative monitoring of offshore wind turbine blade damage based on transfer learning as described in claim 1, characterized in that, The processing steps in step S3 are as follows: the samples in the source domain and the target domain are adaptively aligned based on the CORAL transformation method, as shown in the following formula: ； This formula yields the aligned feature map samples, where yes The transpose of the matrix, It is the Frobenius norm of the matrix.

5. The method for quantitative monitoring of offshore wind turbine blade damage based on transfer learning as described in claim 1, characterized in that, The specific processing steps of steps S4 and S5 are as follows: The Transformer neural network used in the single-scale model training consists of an input layer, a position encoding layer, an addition layer, two self-attention layers, an index layer, a fully connected layer, and a regression layer; the damage quantification results obtained by the single-scale Transformer neural network model are as follows: , where h is the number of models.

Images