Few-sample wind power gear box fault diagnosis method based on data generation

Abstract

The invention relates to the field of wind power gear box fault diagnosis, in particular to a few-sample wind power gear box fault diagnosis method based on data generation, and the method comprises the steps: collecting vibration signal data of a gear box in the operation process through a vibration sensor disposed on the gear box of a wind turbine generator, and carrying out the preprocessing of the collected data; in combination with a historical fault report of the wind turbine generator gearbox, marking a fault label on the preprocessed data as training data; constructing a fault diagnosis model which comprises a generator, a feature extractor, a fault type classifier and a true and false data discriminator; a generative adversarial network composed of the generator, the feature extractor and the true and false data discriminator generates generated data similar to true data; training a fault diagnosis network composed of a feature extractor and a fault type classifier by using the generated data and the real data, and judging the fault type of the sample to be detected by using the fault diagnosis network; according to the invention, powerful technical guarantee can be provided for intelligent operation and maintenance in the wind power industry.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine gearbox fault diagnosis, and in particular to a small sample wind turbine gearbox fault diagnosis method based on data generation. Background Art

[0002] Wind turbines operate under alternating loads and other harsh conditions for a long time, which can easily lead to failures in their key components. Among them, gearbox failures not only cause the longest downtime, but also the highest maintenance costs, resulting in the most significant economic losses. Therefore, it is particularly important to implement real-time fault diagnosis and monitoring of wind turbine gearboxes. Timely identification of faults and adoption of corresponding maintenance strategies can not only effectively control the further deterioration of faults, but also significantly improve the overall operating efficiency and reliability of wind turbines.

[0003] In the past few years, wind turbine gearbox fault diagnosis methods based on deep learning have made significant progress. These methods rely on the powerful functions of neural networks to perform feature dimensionality reduction and pattern recognition, thereby effectively identifying faults in vibration signals. Compared with traditional fault diagnosis algorithms, deep learning methods have stronger high-dimensional and nonlinear data feature extraction capabilities, higher pattern recognition accuracy, and no need to rely on manual feature extraction. The application of these technologies, such as autoencoders, convolutional neural networks, and recurrent neural networks, has achieved good results in the field of wind turbine gearbox fault diagnosis with relatively sufficient labeled samples. However, in practical applications, obtaining high-quality labeled samples often faces high costs and time constraints, which makes labeled samples scarce, which in turn seriously affects the accuracy and reliability of fault identification. Therefore, it is of great theoretical significance and practical value to carry out research on wind turbine gearbox fault diagnosis under the condition of limited labeled samples. Summary of the invention

[0004] In order to solve the challenges faced by wind turbine gearbox fault diagnosis in the prior art, including the problem that labeled fault samples are difficult to obtain due to high costs, and the diagnosis model trained under limited labeled fault samples is prone to overfitting, thus failing to accurately identify wind turbine gearbox faults, the present invention proposes a small sample wind turbine gearbox fault diagnosis method based on data generation, which specifically includes the following steps:

[0005] The vibration sensor installed on the gearbox of the wind turbine collects the vibration signal data of the gearbox during operation and pre-processes the collected data;

[0006] Combined with the historical fault reports of wind turbine gearboxes, the preprocessed limited data is labeled with fault labels as training data;

[0007] Construct a fault diagnosis model, which includes a generator, a feature extractor, a fault type classifier, and a true and false data discriminator;

[0008] The generative adversarial network consisting of a generator, a feature extractor, and a true and false data discriminator generates generated data that is similar to the real data;

[0009] The fault diagnosis network consisting of feature extractor and fault type classifier is trained using generated data and real data;

[0010] Finally, the trained fault diagnosis network is used to determine the fault type of the sample to be detected; the present invention can provide a powerful technical guarantee for the intelligent operation and maintenance of the wind power industry.

[0011] Furthermore, the process of preprocessing the collected data includes: segmenting the collected data, that is, cutting the collected data into vibration signal segments of fixed length, and normalizing each vibration signal segment.

[0012] Furthermore, a generative adversarial network is formed by a generator, a feature extractor and a true or false data discriminator. In the network, the generator generates a data according to the input random noise z; the feature extractor extracts the features of the data, and the true or false data discriminator determines whether the data is generated data according to the extracted features. If the true or false data discriminator cannot determine whether the data is generated data, the generative adversarial network is fully trained and can be used to generate data to expand the data set.

[0013] Furthermore, when the adversarial generative network is trained, the loss function of the generator can be expressed as:

[0014] L ad_G =E z～P(z) [log(D(G(z)))]

[0015] Among them, L ad_G is the loss function of the generator, E[·] represents the mathematical expectation, G represents the generator, z represents the random noise input to the generator, P(z) represents the data distribution of the random noise, z~P(z) represents the data z that conforms to the data distribution P(z); D represents the true and false data discriminator;

[0016] The loss function of the discriminator consisting of a feature extractor and a true and false data discriminator can be expressed as:

[0017] L ad_F&D =-E x～P(x) [log(D(x))]-E z～P(z) [log(1-D(G(z)))]

[0018] Among them, L ad_F&Dis the loss function of the discriminator, x represents the real sample data, P(x) represents the data distribution of the real sample data, and x~P(x) represents the data x that conforms to the data distribution P(x);

[0019] During the training process, both the generator and the discriminator are trained in the direction of minimizing their loss functions. The value function of the adversarial generative network during adversarial training is expressed as:

[0020]

[0021] Among them, L ad is the value function of the adversarial generative network during adversarial training; the generator and the discriminator play a minimax game during training; when the Nash equilibrium is reached, the discriminator can no longer distinguish between the generated data and the real data, indicating that the generator can generate sufficiently real data.

[0022] Furthermore, a fault diagnosis network is constructed by a feature extractor and a fault type classifier, and the feature extractor of the fault diagnosis network and the feature extractor in the generative adversarial network share parameters. Features are extracted from the samples to be detected by the feature extractor, and the fault identification network determines the fault type of the samples to be detected according to the extracted features. The fault diagnosis network is trained based on data generated by the generative adversarial network and limited real data.

[0023] Furthermore, the loss function of the fault diagnosis network during training is expressed as:

[0024]

[0025] Among them, L ce represents the loss function of the fault diagnosis network; N G represents the number of training samples, Indicates the prediction label of the i-th training sample With the true label The cross entropy between .

[0026] Furthermore, the feature extractor is composed of multiple cascaded feature organization units, each of which is composed of cascaded bidirectional gate recurrent units and multi-scale convolution modules; the bidirectional gate recurrent unit ensures that the long-term dependencies in the time series data are fully captured, while the multi-scale convolution module can effectively mine the spatial features of data at different scales. The combined application of the bidirectional gate recurrent unit and the multi-scale convolution module enables the feature organization unit to jointly learn time series and spatial features, enhancing its ability to mine and restore the complexity of real data.

[0027] Furthermore, the bidirectional gate recurrent unit updates the feature vector of the input bidirectional gate recurrent unit according to the hidden features, and the updated feature vector serves as the output of the feature extraction unit; the multi-scale convolution module uses multiple convolution kernels of different sizes to extract multiple different features from the input data, and fuses the extracted features by feature splicing, and then uses the cascaded pooling layer, batch normalization layer, and ReLU activation function to further process the fused features as the output of the multi-scale convolution module.

[0028] Furthermore, the multi-scale convolution module uses at least convolution layers with convolution kernel sizes of 1, 3, 5, and 7 to extract features of different sizes from the input features respectively, and then concatenates the extracted features together and fuses them through an average pooling layer with a pooling kernel of 2. Finally, the batch normalization layer and the ReLU activation function are used to obtain the feature map output by the multi-scale convolution module.

[0029] The present invention effectively solves the problem of scarcity of labeled samples and provides richer and more realistic data support for fault diagnosis of wind turbines, thereby reducing the maintenance cost of wind turbines, improving the reliability and safety of equipment, and providing strong technical support for intelligent operation and maintenance of the wind power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of a method for diagnosing wind turbine gearbox faults based on a small number of samples generated based on data in the present invention;

[0031] Figure 2 It is a structural diagram of a bidirectional door circulation unit of the present invention;

[0032] Figure 3 This is a structural diagram of the multi-scale convolution module of the present invention;

[0033] Figure 4 This is a schematic diagram of the fault diagnosis model structure and training process of the present invention;

[0034] Figure 5 The application process of the fault diagnosis model of the present invention;

[0035] Figure 6 It is the source of the data used in the verification test of the present invention;

[0036] Figure 7 A schematic diagram showing the comparison between the data generated by the fault diagnosis model of the present invention and the original data;

[0037] Figure 8 Schematic diagram comparing the diagnostic accuracy of the present invention and other fault diagnosis and identification methods. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] The present invention proposes a small sample wind turbine gearbox fault diagnosis method based on data generation, such as Figure 1 , specifically including the following steps:

[0040] The vibration sensor installed on the gearbox of the wind turbine collects the vibration signal data of the gearbox during operation and pre-processes the collected data;

[0041] Combined with the historical fault reports of wind turbine gearboxes, the preprocessed limited data is labeled with fault labels as training data;

[0042] Construct a fault diagnosis model, which includes a generator, a feature extractor, a fault type classifier, and a true and false data discriminator;

[0043] The generative adversarial network consisting of a generator, a feature extractor, and a true and false data discriminator generates generated data that is similar to the real data;

[0044] The fault diagnosis network consisting of feature extractor and fault type classifier is trained using generated data and real data;

[0045] Finally, the trained fault diagnosis network is used to determine the fault type of the sample to be detected; the present invention can provide a powerful technical guarantee for the intelligent operation and maintenance of the wind power industry.

[0046] This embodiment provides a specific implementation of the present invention, that is, a wind turbine gearbox fault diagnosis method based on data generation under the condition of scarce fault samples, comprising the following steps:

[0047] S1: Data collection and data preprocessing: The vibration signal data of the gearbox during operation is collected by installing a vibration sensor on the wind turbine gearbox; the continuous vibration signal is segmented according to the preset length to ensure that each segmented signal segment contains sufficient vibration information for subsequent analysis; each segmented vibration signal segment is normalized to normalize its amplitude range to eliminate the numerical differences between samples and facilitate comparison and analysis of different samples.

[0048] Preferably, the method for collecting the vibration signal of the gearbox of the wind turbine generator set includes signal collection through a vibration condition monitoring system (Condition Monitoring System, CMS). For example, during the daily operation of the wind turbine generator set, the CMS samples the vibration signal of the gearbox at a sampling frequency of 25600 Hz and a sampling duration of 5.12 seconds, thereby obtaining 131072 sampling data points. In order to obtain more samples, for each type of fault type, a total of 655360 data points of vibration data collected within five days before the gearbox fails are processed. The continuous vibration signal is segmented according to the preset sample length of 1024 data points, and each segmented vibration signal segment is normalized to normalize its amplitude range to eliminate the numerical differences between samples, which is convenient for comparison and analysis of different samples. Finally, each type of fault type generates 640 valid samples of this type of fault.

[0049] S2: Data labeling and data division: Combined with the historical fault reports of wind turbine gearboxes, specific fault category labels are added to some vibration samples with known faults, and the corresponding fault types are marked; the fault samples with fault category labels are divided into training sets and test sets.

[0050] Preferably, the method of labeling some fault samples with fault category labels based on the historical fault report of the wind turbine gearbox includes: assigning corresponding fault category labels to some fault samples based on the inspection report of the wind turbine gearbox and the guidance of experts. These fault category labels specifically include the fault types corresponding to the fault samples, such as gearbox tooth surface wear, inner ring fault, rolling element fault, and through cracks. In this way, the accuracy of the labeling can be ensured, providing a reliable reference for subsequent fault diagnosis.

[0051] S3: Offline fault diagnosis model training: Build a fault diagnosis model, train the fault diagnosis model through the training set, verify the fault diagnosis model with the help of the test set and save the optimal parameters.

[0052] The fault diagnosis model proposed in the present invention can effectively improve the fault diagnosis accuracy and robustness of the wind turbine gearbox. Generative adversarial networks solve the problem of scarcity of labeled samples by generating high-quality synthetic fault samples, thereby enhancing the learning ability of the model and reducing the risk of overfitting. The fault diagnosis network uses synthetic samples for effective training. This method not only improves the ability to extract fault features, but also enhances the adaptability of the model under complex working conditions, making fault diagnosis more accurate. This innovative solution provides more reliable data support for the operation and maintenance of wind turbines, helps to reduce maintenance costs, and improve the reliability and safety of equipment.

[0053] Preferably, the fault diagnosis model combines the generative adversarial network with the fault diagnosis network, and realizes cross-task knowledge transfer and sharing by sharing feature extractors. The model structure includes a generator, a feature extractor, a fault type classifier, and a true and false data discriminator, wherein the generator generates high-quality fault data using a bidirectional gate recurrent unit and a multi-scale convolution module to expand the training set. The feature extractor is shared, which not only serves the fault diagnosis network, but also provides a more effective feature representation for the generative adversarial network. By sharing the feature extractor, the fault diagnosis network and the generative adversarial network can benefit from each other. The discriminator of the generative adversarial network improves the sensitivity to true and false data by learning the classification information in the fault diagnosis task, thereby enhancing the ability of the generator to generate real data. At the same time, the fault diagnosis network improves the diagnostic accuracy and the generalization ability of the model with the help of the high-quality data generated by the generative adversarial network. This design not only improves the resource utilization efficiency, but also optimizes the feature representation, making fault diagnosis more efficient in complex industrial environments. The fault diagnosis model not only enhances the model's ability to recognize complex fault modes, but also promotes effective learning under the condition of scarce samples, making the fault diagnosis of wind turbine gearboxes more accurate and reliable.

[0054] This embodiment provides an implementation process of building and training an offline fault diagnosis model, which specifically includes:

[0055] S31: The fault diagnosis model of this embodiment includes four parts: a generator, a feature extractor, a fault type classifier, and a true and false data discriminator. The generator is mainly used to map low-dimensional random noise to a high-dimensional data space from a random noise vector as input to generate a sample that is as realistic as possible; the discriminator composed of the feature extractor and the true and false data discriminator is mainly used to distinguish whether a given sample is a real data set or generated data output by the generator; the generative adversarial network composed of the generator and the discriminator is mainly used to expand a limited data set. The fault diagnosis network composed of the feature extractor and the fault type classifier is mainly used to distinguish which type of fault a given sample belongs to. It should be pointed out that the fault diagnosis network and the generative adversarial network share a feature extractor, and the fault diagnosis network and the generative adversarial network are combined to achieve knowledge sharing and effective knowledge transfer. The generative adversarial network enhances the diversity and authenticity of the training data of the diagnosis network by generating high-quality fault data, while the classification task of the fault diagnosis network helps the discriminator improve the ability to identify true and false data. Sharing the feature extractor enables the two networks to benefit from each other, improves the quality of feature representation, and thus improves the accuracy of fault diagnosis and the reliability of generated data.

[0056] In this embodiment, the generator and feature extractor are mainly composed of a bidirectional gate recurrent unit and a multi-scale convolution module, and the fault type classifier and true and false data discriminator are composed of a fully connected network, such as Figure 2,The bidirectional gate recurrent unit considers the context information before and after each element in the vibration time series data, which can more comprehensively understand the dynamic changes in the sequence, thus improving the quality of feature representation; Figure 3 The multi-scale convolution module applies convolution kernels at different levels, so that richer and more multi-level information can be mined in the process of high-dimensional feature extraction, improving its performance in diagnosis and identification tasks; in the process of data generation, it can more accurately reproduce the complexity of real data, ensuring that the generated fault samples reach a high level in terms of diversity and authenticity. The four parts of the generator, feature extractor, fault type classifier, and true and false data discriminator are defined as G, F, C, and D respectively. The discriminator can be expressed as F&D, the generative adversarial network can be expressed as G&F&D, and the fault diagnosis network can be expressed as F&C.

[0057] Assume f in is the input feature, then the operation of the multi-scale convolution module can be expressed as:

[0058] MC(F in )=cat(ψ1(f in ),...,ψ nk (f in ))

[0059] Among them, n k Represents the number of different convolution kernel sizes on the multi-scale convolution layer, n k = 1 means that feature extraction is performed using a convolution layer with a convolution kernel size of 1. In this embodiment, n k ={1,3,5,7};ψ i (·) represents a function that performs convolution operations using a convolution kernel of size i; cat(·) represents a concatenation function that concatenates multiple tensors together.

[0060] The operation process of the multi-layer convolution structure can be expressed as:

[0061] MCB(f in )=ReLU(BN(Pooling(MC(F in )))

[0062] Among them, Pooling(·), BN(·) and ReLU(·) represent the pooling operation, batch normalization and ReLU activation function, respectively.

[0063] Preferably, the generator is trained to generate realistic data from random noise z. The loss function of the generator can be expressed as:

[0064] L ad_G =E z～P(z) [log(D(G(x)))]

[0065] Among them, E represents mathematical expectation; P represents data distribution, the generator represents the mapping from the latent space to the real data space, and the discriminator represents the probability that a given input originates from real data rather than being synthesized by the generator. The goal of the generator is to minimize the loss function L ad_G , thereby generating samples that are very similar to real data. The discriminator is trained to distinguish between real data samples x and samples generated by the generator based on the input it receives.

[0066] The loss function of the discriminator can be expressed as:

[0067] L ad_F&D =-E x～P(x) [log(D(x))]-E z～P(z) [log(1-D(G(z)))]

[0068] The goal of the discriminator is to minimize L ad_F&D , which is directly opposite to the goal of the generator. Therefore, the generator and the discriminator play a minimax game during training. Ideally, the discriminator can no longer distinguish between generated data and real data, indicating that the generator has successfully learned to generate real samples. The value function of this adversarial process can be expressed as:

[0069]

[0070] By combining data augmentation with generative adversarial networks, this study effectively addresses the challenge of data scarcity by generating a sufficient amount of synthetic data to train fault diagnosis models. This approach not only increases the amount of training data, but also ensures that the synthetic samples maintain high fidelity, thereby improving the robustness and generalization ability of the diagnosis model.

[0071] S312: Structure Similarity Index Measure (SSIM) is an indicator used to evaluate the similarity between two signals. The index provides a similarity measurement method by comprehensively considering multiple aspects such as brightness, contrast, and structure. The value of SSIM ranges from 0 to 1, where 1 means exactly the same and 0 means completely different. In specific applications, SSIM can effectively capture the structural information in the signal, making it play an important role in fault sample generation and quality assessment. By using SSIM, the similarity between the generated fault samples and the real samples can be measured more accurately, thereby optimizing the generation effect of the model and improving the reliability and accuracy of fault diagnosis.

[0072] S32: The process of training the generative adversarial network based on the training set aims to optimize the model parameters so that the structural similarity index (SSIM) between the generated fault samples and the training set samples reaches or exceeds 0.75. Through the adversarial training mechanism, the model's performance in data generation ability is enhanced, thereby improving the quality and authenticity of the generated samples to better reflect the actual fault characteristics.

[0073] S33: Based on the trained generative adversarial network, corresponding fault samples are generated for each fault type to build a comprehensive generated fault dataset. This dataset not only covers the characteristics of different fault types, but also provides rich sample resources for subsequent model training and evaluation.

[0074] Preferably, N G The generated dataset of samples can be generated by a trained generative adversarial network, where and Represents fault samples and their labels. Finally, it can be further trained with real data sets.

[0075] S34: Based on the training set and the generated fault data set, the fault diagnosis network is trained to enhance its ability to identify different fault types. After the training process is completed, the fault diagnosis network is tested and verified using the test set to evaluate the performance of the model on unseen data. In this process, the model parameters with the best test accuracy will be saved to ensure that the highest recognition accuracy can be achieved in actual applications.

[0076] like Figure 4 , the fault diagnosis network is trained based on the training set and the generated fault data set. The real and generated fault samples are first input into the feature extractor to map the fault samples from the data space to the high-dimensional feature space. Then, the high-dimensional is input into the fault type classifier to obtain the diagnosis result. The whole diagnosis process can be expressed as:

[0077]

[0078] Then the cross entropy loss L of the fault diagnosis network is ce The structure is:

[0079]

[0080] Among them, J represents the cross entropy function.

[0081] In particular, if Figure 4In this embodiment, the generator is composed of four cascaded first feature organization units, each of which is composed of a cascaded bidirectional gate recurrent unit and a multi-scale convolution module; the feature extractor is composed of four cascaded second feature organization units, each of which is composed of a cascaded multi-scale convolution module and a bidirectional gate recurrent unit.

[0082] After the training process is completed, the fault diagnosis network is tested and verified using the test set to evaluate the performance of the model on unseen data. During this process, the model parameters with the best test accuracy will be saved to ensure the highest recognition accuracy in practical applications.

[0083] S4: Online fault diagnosis, such as Figure 5 , obtain the target vibration signal of the wind turbine gearbox, and perform normalization processing on it, and adjust its amplitude to the same range as the training sample; input the normalized target vibration signal into the previously trained fault diagnosis model, and the model obtains the fault diagnosis result of the wind turbine gearbox by analyzing and calculating the target signal.

[0084] In order to verify the effectiveness and advancement of a small sample wind turbine gearbox fault diagnosis method based on data generation under the condition of scarce fault samples disclosed in the present invention, a comparative test is conducted on the vibration signal data of the wind turbine gearbox. Figure 6 The wind turbine gearbox shown collects vibration signal data; as shown in the following table, the vibration signal data has a total of different wind turbine gearbox health states, each category contains 640 samples (each sample contains 1024 data points).

[0085] Table 1 Description of wind turbine gearbox vibration signal data

[0086]

[0087] To simulate the situation of limited labeled data of known fault types, only a subset of the total samples was used in the comparison test. Specifically, 20 fault samples were randomly selected from the complete sample set of each fault type as labeled training data to simulate the few-sample scenario in real operation and maintenance, and used to train the fault diagnosis model to provide the necessary information for supervised learning. The remaining samples of each fault type were designated as test samples to evaluate the performance of the model during training and to ensure that the model can generalize well to unseen data.

[0088] Based on the above sample sets, the proposed method for fault diagnosis of wind turbine gearboxes based on data generation under the condition of scarce fault samples is compared with seven other methods, including PN described in the literature "Prototypical Networks for Few-shot Learning (Snell et al., 2017)", ST-DAGANs-CapNet described in the literature "Adeep capsule neural network with data augmentation generative adversarial networks for single and simultaneous fault diagnosis of wind turbine gearbox (Liang et al., 2023)", FTGAN described in the literature "FTGAN: A Novel GAN-Based Data Augmentation Method Coupled Time–Frequency Domain for Imbalanced Bearing Fault Diagnosis (Wang et al., 2023)", CWGAN-GP-VAE described in the literature "Chiller Fault Diagnosis Based on VAE-Enabled Generative Adversarial Networks (Yan et al., 2022)", and Data Augmentation and Intelligent Fault ILoFGAN is described in "Diagnosis of Planetary GearboxUsing ILoFGAN Under Extremely Limited Samples (Chen et al., 2022)".

[0089] First, a comparative analysis of the quality of samples produced by different generative models is conducted. This analysis uses statistical methods to evaluate the diversity and authenticity of generated samples and explores the differences in the ability of different methods to synthesize data. All generative models are trained with a limited number of training fault samples and generate fault samples for the gearbox dataset, such as Figure 7As shown. The results show that the samples generated by the fault diagnosis model have a high similarity with the original samples, verifying its ability to generate high-quality fault samples. In addition, to further verify the superiority of the proposed fault diagnosis model, SSIM was applied to evaluate the quality of the generated samples, and the evaluation results are shown in Table 2. The analysis of the SSIM results further emphasizes the effectiveness of the fault diagnosis model in generating high-quality samples. The SSIM values ​​of the samples generated by the fault diagnosis model are always close to those of the original samples, indicating that their structural similarity is high. In contrast, the SSIM values ​​of ST DAGAN's CapNet and FTGAN show significant deviations, highlighting their limitations in capturing basic features and maintaining the fidelity of the original data. This difference further shows that the advanced architecture of the fault diagnosis model, especially its multi-scale convolutional module, plays a vital role in preserving the complex patterns in the fault samples.

[0090] Table 2 SSIM of samples generated by different generation techniques

[0091]

[0092]

[0093] The diagnostic performance of various methods in the context of wind turbine fault diagnosis was then explored in depth. A comprehensive evaluation was conducted to assess the effectiveness of each method in identifying and classifying faults based on the synthetic data generated in the previous analysis. To further improve the fairness and persuasiveness of the experiment, the compared methods were subjected to five independent experiments based on the same experimental parameters, datasets, and feature extractors. For detailed experimental results, see Figure 8 This repeated experiment design is designed to ensure the stability and reliability of the results, reduce the impact of accidental factors on the experimental results, and thus provide stronger support and evidence for the effectiveness of the method.

[0094] According to the comparison results in Table 3, the performance of each method can be analyzed in detail. The data in the table includes the mean and variance of the quality of the generated samples of each method. The mean reflects the overall quality of the generated samples, while the variance indicates the stability and consistency of the generated results.

[0095] Table 3 Comparison of test accuracy of different methods

[0096]

[0097] The fault diagnosis model has the highest average test accuracy of 97.19%, showing the best performance of this method in terms of the quality of generated samples. This means that the fault samples generated by the fault diagnosis model are closest to the original samples, have high similarity, and can better preserve the key features and complex patterns in the original data. FTGAN and CWGAN-GP-VAE followed closely with 93.60% and 92.32%, respectively, and their performance was also relatively good, but still not as good as the fault diagnosis model. This shows that these two methods have certain advantages in the quality of generated samples, but have not reached the level of the fault diagnosis model, and may have some limitations in capturing complex patterns or detailed features. The average test accuracy of ILoFGAN and ST-DAGANs-CapNet is 88.14% and 86.13%, respectively, which is relatively low, indicating that these two methods perform poorly in the quality of generated samples, and the generated samples may be insufficient in clarity and detail fidelity. PN has the lowest average test accuracy of only 75.20%, indicating that the PN method is difficult to effectively capture the complex features and details of the samples under the condition of limited fault samples.

[0098] In summary, the fault diagnosis model performs best in terms of generation quality, but the variance is large, which may indicate that the quality of samples generated by this method fluctuates greatly in some cases. However, overall, it is still the best choice and can generate high-quality fault samples that are closest to real samples, especially for application scenarios that require high quality of generated samples.

[0099] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A small sample wind turbine gearbox fault diagnosis method based on data generation, characterized in that: The specific steps include: The vibration sensor installed on the gearbox of the wind turbine collects the vibration signal data of the gearbox during operation and pre-processes the collected data; Combined with the historical fault reports of wind turbine gearboxes, the preprocessed limited data is labeled with fault labels as training data; Construct a fault diagnosis model, which includes a generator, a feature extractor, a fault type classifier, and a true and false data discriminator; The generative adversarial network consisting of a generator, a feature extractor, and a true and false data discriminator generates generated data that is similar to the real data; The fault diagnosis network consisting of feature extractor and fault type classifier is trained using generated data and real data; Finally, the trained fault diagnosis network is used to determine the fault type of the sample to be detected; the present invention can provide a powerful technical guarantee for the intelligent operation and maintenance of the wind power industry.

2. The method for diagnosing wind turbine gearbox faults based on a small number of samples generated by data according to claim 1, characterized in that: The process of preprocessing the collected data includes: segmenting the collected data, that is, cutting the collected data into vibration signal segments of fixed length, and normalizing each vibration signal segment.

3. The method for diagnosing wind turbine gearbox faults based on a small number of samples generated by data according to claim 1, characterized in that: A generative adversarial network is formed by a generator, a feature extractor, and a true or false data discriminator. In the network, the generator obtains a data based on the input random noise z; the feature extractor is used to extract the features of the data, and the true or false data discriminator determines whether the data is generated data based on the extracted features. If the true or false data discriminator cannot determine whether the data is generated data, the generative adversarial network training is complete.

4. The method for diagnosing wind turbine gearbox faults based on a small number of samples generated by data according to claim 3, characterized in that: When the adversarial generative network is trained, the loss function of the generator is expressed as: L ad_G =E z～P(z) [log(D(G(z)))] Among them, L ad_G is the loss function of the generator, E[·] represents the mathematical expectation, G represents the generator, z represents the random noise input to the generator, P(z) represents the data distribution of the random noise, z~P(z) represents the data z that conforms to the data distribution P(z); D represents the true and false data discriminator; The loss function of the discriminator consisting of a feature extractor and a true and false data discriminator can be expressed as: L ad_F&D =-E x～P(x) [log(D(x))]-E z～P(z) [log(1-D(G(z)))] Among them, L ad_F&D is the loss function of the discriminator, x represents the real sample data, P(x) represents the data distribution of the real sample data, and x~P(x) represents the data x that conforms to the data distribution P(x); During the training process, both the generator and the discriminator are trained in the direction of minimizing their loss functions. The value function of the adversarial generative network during adversarial training is expressed as: Among them, L ad It is the value function of the adversarial generation network during adversarial training.

5. The method for diagnosing wind turbine gearbox faults based on a small number of samples generated by data according to claim 3, characterized in that: The fault diagnosis network is composed of a feature extractor and a fault type classifier, and the feature extractor of the fault diagnosis network and the feature extractor in the adversarial generation network share parameters. The feature extractor extracts features from the sample to be detected, and the fault identification network determines the fault type of the sample to be detected based on the extracted features.

6. A method for diagnosing wind turbine gearbox faults based on a small number of samples generated by data according to claim 5, characterized in that: The loss function of the fault diagnosis network during training is expressed as: Among them, L ce represents the loss function of the fault diagnosis network; N G represents the number of training samples, Indicates the prediction label of the i-th training sample With the true label The cross entropy between .

7. A method for diagnosing wind turbine gearbox faults based on a small number of samples generated based on data according to claim 3 or 5, characterized in that: The generator and feature extractor are composed of multiple cascaded feature organization units, each of which is composed of a cascaded bidirectional gate recurrent unit and a multi-scale convolution module. The multi-scale convolution module in the feature extractor includes downsampling the input feature data to achieve dimensionality reduction processing of the data, and the multi-scale convolution module in the generator includes upsampling the input feature data to achieve dimensionality increase processing of the data.

8. The method for diagnosing wind turbine gearbox faults based on a small number of samples generated by data according to claim 7, characterized in that: The bidirectional gate recurrent unit updates the feature vector of the input bidirectional gate recurrent unit according to the hidden features, and the updated feature vector serves as the output of the bidirectional gate recurrent unit; the multi-scale convolution module uses multiple convolution kernels of different sizes to extract multiple different features from the input data, and fuses the extracted features by feature splicing, and then uses the cascaded pooling layer, batch normalization layer, and ReLU activation function to further process the fused features as the output of the multi-scale convolution module.

9. The method for diagnosing wind turbine gearbox faults based on a small number of samples generated by data according to claim 8, characterized in that: The multi-scale convolution module uses at least convolution layers with convolution kernel sizes of 1, 3, 5, and 7 to extract features of different sizes from the input features, respectively. The extracted features are then concatenated together and fused through an average pooling layer with a pooling kernel of 2. Finally, the batch normalization layer and the ReLU activation function are used to obtain the feature map output by the multi-scale convolution module.

Images