Wind turbine generator bearing fault diagnosis method and device

Abstract

The invention discloses a wind turbine generator bearing fault diagnosis method and device, and relates to the technical field of wind turbine generator bearing fault diagnosis. The method comprises the steps of collecting vibration signals of a wind turbine generator bearing in a normal state and in different fault types in a resampling mode; combining an intrinsic mode function obtained by decomposing the vibration signal through variational mode decomposition and a frequency domain signal obtained through fast Fourier transform to serve as input, and forming a data set together with a label of the vibration signal; a meta learning-HyperBand Bayesian optimizer is adopted to optimize the hyper-parameters of the CNN-Transformer hybrid model, and the hyperparameters of the CNN-Transformer hybrid model are optimized; training the optimized CNN-Transform hybrid model by using the data set to obtain a wind turbine generator bearing fault diagnosis model; a fault diagnosis result of the wind turbine generator bearing can be obtained by using the wind turbine generator bearing fault diagnosis model. According to the invention, the accuracy and efficiency of fault diagnosis can be improved.

Description

Technical Field

[0001] This application relates to the technical field of bearing fault diagnosis for wind turbines, and particularly to a method and device for bearing fault diagnosis of wind turbines. Background Art

[0002] As a clean energy source, wind power has been widely used in recent years. As the core equipment of the wind power generation system, the operating performance of wind turbines directly affects the efficiency and stability of the entire power generation process. One of the main components of wind turbines - bearings, bear multiple pressures such as rotor imbalance, external shocks, and mechanical wear during long-term operation. Due to the complex working environment and large load changes, bearings are prone to various faults, such as wear, cracks, spalling, etc. These faults not only affect the stability and efficiency of the wind turbines, but may also lead to serious economic losses and safety hazards. Therefore, timely detection and diagnosis of bearing faults are of great significance for ensuring the normal operation of wind turbines and reducing maintenance costs.

[0003] Traditional fault diagnosis methods usually rely on the analysis and feature extraction of vibration signals, but these methods often rely on manual experience and have certain limitations in dealing with complex signals and noises. With the development of computing technology and artificial intelligence, fault diagnosis methods based on deep learning have gradually emerged. These methods can automatically learn effective features from a large amount of data and have strong robustness, which can effectively improve the accuracy of fault diagnosis. However, deep learning methods still face challenges such as complex signal noises, data imbalance, and long model training time in practical applications. Therefore, how to improve the performance of existing fault diagnosis methods remains an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this application is to provide a method and device for bearing fault diagnosis of wind turbines, which can improve the accuracy and efficiency of fault diagnosis.

[0005] To achieve the above purpose, this application provides the following solutions:

[0006] In a first aspect, the present application provides a method for diagnosing faults in a wind turbine bearing, including: collecting vibration signals of the wind turbine bearing in a normal state and different fault types by means of resampling, and labeling the normal state or fault type as the label of the vibration signal; using variational mode decomposition to decompose each vibration signal into multiple intrinsic mode functions; using fast Fourier transform to convert each vibration signal from the time domain to the frequency domain to obtain a frequency domain signal; combining the multiple intrinsic mode functions and frequency domain signals corresponding to each vibration signal as inputs, and together with the label of each vibration signal, constituting a data set; according to the data set, using meta-learning - HyperBand Bayesian optimizer to optimize the hyperparameters of the CNN-Transformer hybrid model; training the optimized CNN-Transformer hybrid model using the data set to obtain a wind turbine bearing fault diagnosis model; collecting real-time vibration signals during the operation of the wind turbine bearing, and through variational mode decomposition and fast Fourier transform, obtaining multiple intrinsic mode functions and frequency domain signals corresponding to the real-time vibration signals; inputting the multiple intrinsic mode functions and frequency domain signals corresponding to the real-time vibration signals into the wind turbine bearing fault diagnosis model to obtain a fault diagnosis result of the wind turbine bearing; the fault diagnosis result is a normal state or a fault type.

[0007] In a second aspect, the present application provides a device for diagnosing faults in a wind turbine bearing, including: a data collection module for collecting vibration signals of the wind turbine bearing in a normal state and different fault types by means of resampling, and labeling the normal state or fault type as the label of the vibration signal; a decomposition module for using variational mode decomposition to decompose each vibration signal into multiple intrinsic mode functions; a conversion module for using fast Fourier transform to convert each vibration signal from the time domain to the frequency domain to obtain a frequency domain signal; a merging module for combining the multiple intrinsic mode functions and frequency domain signals corresponding to each vibration signal as inputs, and together with the label of each vibration signal, constituting a data set; an optimization module for optimizing the hyperparameters of the CNN-Transformer hybrid model according to the data set using meta-learning - HyperBand Bayesian optimizer; a training module for training the optimized CNN-Transformer hybrid model using the data set to obtain a wind turbine bearing fault diagnosis model; a real-time acquisition module for collecting real-time vibration signals during the operation of the wind turbine bearing, and through variational mode decomposition and fast Fourier transform, obtaining multiple intrinsic mode functions and frequency domain signals corresponding to the real-time vibration signals; a diagnosis module for inputting the multiple intrinsic mode functions and frequency domain signals corresponding to the real-time vibration signals into the wind turbine bearing fault diagnosis model to obtain a fault diagnosis result of the wind turbine bearing; the fault diagnosis result is a normal state or a fault type.

[0008] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0009] The present application provides a method and device for fault diagnosis of a wind turbine bearing. The combination of variational mode decomposition and fast Fourier transform is used for non-stationary signal decomposition and frequency-domain feature extraction, effectively solving the problem of dealing with non-stationary signals and noise interference in traditional methods; the CNN-Transformer hybrid model combines a convolutional neural network and a Transformer network, enhancing the local feature extraction ability and long-term dependence modeling, and improving the accuracy of fault diagnosis; at the same time, the meta-learning-HyperBand Bayesian optimizer accelerates the hyperparameter optimization through meta-learning, reduces the time and resources required for manual adjustment, and significantly improves the efficiency of the diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0011] Figure 1 It is a schematic flowchart of a method for fault diagnosis of a wind turbine bearing provided in an embodiment of the present application;

[0012] Figure 2 It is a schematic diagram of a partial time-domain signal spectrum decomposed by variational mode decomposition provided in another embodiment of the present application;

[0013] Figure 3 It is a schematic diagram of a partial frequency-domain signal spectrum generated by fast Fourier transform provided in another embodiment of the present application;

[0014] Figure 4 It is a schematic diagram of the structure of the Transformer network provided in another embodiment of the present application;

[0015] Figure 5 It is a detailed flowchart of a method for fault diagnosis of a wind turbine bearing provided in an embodiment of the present application;

[0016] Figure 6 It is a schematic diagram of the confusion matrix of the fault diagnosis output provided by the present application;

[0017] Figure 7 It is a schematic diagram of the ablation experiment of variational mode decomposition-fast Fourier transform provided by the present application;

[0018] Figure 8Schematic diagram of the accuracy comparison of different models provided in this application with and without using an optimizer. Detailed implementation manners

[0019] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0020] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0021] In an exemplary embodiment, as Figure 1 shown, a method for diagnosing bearing faults of a wind turbine is provided, including the following steps 101 to 108. Among them:

[0022] Step 101: Collect vibration signals of the bearings of the wind turbine in the normal state and different fault types by using resampling, and label the normal state or fault type as the label of the vibration signal.

[0023] Step 102: Use variational mode decomposition to decompose each vibration signal into multiple intrinsic mode functions.

[0024] Step 103: Use the fast Fourier transform to convert each vibration signal from the time domain to the frequency domain to obtain a frequency domain signal.

[0025] Step 104: Combine the multiple intrinsic mode functions and frequency domain signals corresponding to each vibration signal as inputs, and together with the label of each vibration signal, form a data set.

[0026] Step 105: According to the data set, use the meta-learning - HyperBand Bayesian optimizer to optimize the hyperparameters of the CNN-Transformer hybrid model.

[0027] Step 106: Use the data set to train the optimized CNN-Transformer hybrid model to obtain a bearing fault diagnosis model for the wind turbine.

[0028] Step 107: Collect real-time vibration signals during the operation of the bearings of the wind turbine, and through variational mode decomposition and fast Fourier transform, obtain multiple intrinsic mode functions and frequency domain signals corresponding to the real-time vibration signals.

[0029] Step 108: Input the multiple intrinsic mode functions and frequency domain signals corresponding to the real-time vibration signal into the wind turbine bearing fault diagnosis model to obtain the fault diagnosis result of the wind turbine bearing; the fault diagnosis result is a normal state or a fault type.

[0030] Implement the above Steps 101 to 108. By combining the time-frequency domain signal characteristics of variational mode decomposition (VMD) and fast Fourier transform (FFT), and the deep learning model of convolutional neural networks (CNN)-Transformer, the problems of non-stationary signal processing and noise interference in wind turbine bearing fault diagnosis are effectively solved. VMD can accurately decompose complex vibration signals and extract more accurate fault features, while FFT helps to further analyze and identify fault signals in the frequency domain. Through the combined application of CNN-Transformer, more complex fault patterns can be automatically extracted and learned, improving the accuracy and robustness of diagnosis. In addition, the resampling technique balances the class distribution in the dataset, further enhancing the generalization ability of the model. Combining meta learning (Meta Learning, abbreviated as Meta)-HyperBand Bayesian optimization (BO) accelerates hyperparameter tuning, reduces manual intervention, significantly improves the model training efficiency and accuracy, and thus greatly enhances the practical application effect of wind turbine bearing fault diagnosis. Among them, HyperBand Bayesian optimization can be abbreviated as BOHB, and meta learning-HyperBand Bayesian optimization is abbreviated as Meta-BOHB.

[0031] In another exemplary embodiment of the present application, in order to expand the dataset to obtain more reliable experimental results, resampling is used to divide the data to obtain multiple sample numbers:

[0032]

[0033] where N is the number of samples, L is the total length of the vibration signal, T is the length of a single sample signal, and S is the step size.

[0034] The bearing fault diagnosis public dataset of Case Western Reserve University is processed using resampling to increase the available number of samples and make the experimental results more persuasive.

[0035] In another exemplary embodiment of the present application, the data set is labeled and the data is annotated according to different fault types. The data set of Case Western Reserve University usually contains multiple bearing fault categories, such as inner race fault, outer race fault, rolling element fault, etc. By carefully dividing the labels of each data point, the model can learn more accurate fault features.

[0036] After resampling is completed, the data labels for the drive end are made, including 10 types of data, among which there are 9 types of fault data, namely: ba_7, ba_14, ba_21, ir_7, ir_14, ir_21, or_7, or_14, or_21 and 1 type of normal state data nc; where ba is the rolling element fault, ir is the inner race fault, or is the outer race fault, and the numerical suffix is the target fault size in inches. Then, the process of labeling the normal state or fault type as the label of the vibration signal in step 101 above can be as follows:

[0037] If the fault type is: inner race fault with a fault size of 0.007 inches, then the label of the vibration signal is marked as 1; if the fault type is: inner race fault with a fault size of 0.014 inches, then the label of the vibration signal is marked as 2; if the fault type is: inner race fault with a fault size of 0.021 inches, then the label of the vibration signal is marked as 3; if the fault type is: outer race fault with a fault size of 0.007 inches, then the label of the vibration signal is marked as 4; if the fault type is: outer race fault with a fault size of 0.014 inches, then the label of the vibration signal is marked as 5; if the fault type is: outer race fault with a fault size of 0.021 inches, then the label of the vibration signal is marked as 6; if the fault type is: ball fault with a fault size of 0.007 inches, then the label of the vibration signal is marked as 7; if the fault type is: ball fault with a fault size of 0.014 inches, then the label of the vibration signal is marked as 8; if the fault type is: ball fault with a fault size of 0.021 inches, then the label of the vibration signal is marked as 9; if it is in the normal state, then the label of the vibration signal is marked as 10.

[0038] The made fault diagnosis data set is shown in Table 1.

[0039] Table 1 Fault Diagnosis Data Set

[0040]

[0041] In another exemplary embodiment of the present application, to address the issues of low accuracy and complexity in the fault diagnosis of fan bearings, variational mode decomposition and fast Fourier transform are used to process the fault vibration signals. Variational mode decomposition is used to decompose non-stationary signals and extract vibration characteristics in different frequency bands, while the fast Fourier transform helps to transform the signals into the frequency domain, thereby revealing potential fault characteristics. This processing step lays the foundation for subsequent feature extraction and classification, ensuring the accurate extraction of fault signals.

[0042] In one example, the goal of variational mode decomposition is to decompose a signal into several intrinsic mode functions (IMFs), that is:

[0043]

[0044] where, u k (t) represents the k-th mode function, and r(t) is the residual term. This formula represents the various mode functions and the remaining term obtained after decomposing the original signal by the variational mode decomposition method.

[0045] To obtain accurate mode functions, the optimization goal of variational mode decomposition is to minimize the following loss function:

[0046]

[0047] where, u k (t) represents the k-th intrinsic mode function, is the frequency domain representation of u k (t), α represents the balance factor, K represents the number of intrinsic mode functions, t represents time, and || || 2 represents the L2 norm. α is used to balance the processing weights in the time domain and the frequency domain. This formula is used to solve for each intrinsic mode function and ensure that the decomposition of the signal does not introduce excessive noise or distortion.

[0048] Partial time domain signal spectra decomposed by variational mode decomposition are as Figure 2 shown.

[0049] In another example, the fast Fourier transform is used to convert the original vibration signal from the time domain to the frequency domain to help extract frequency domain features. The formula for the fast Fourier transform is as follows:

[0050]

[0051] where: X(f) is the signal representation in the frequency domain, x(t n ) is the discrete signal in the time domain, N is the number of signal sampling points, and f is the signal frequency.

[0052] After processing, the obtained time-frequency domain signals are stacked and merged into a dataset containing intrinsic mode functions and frequency domain signals.

[0053] Partial frequency domain signal spectra generated by the fast Fourier transform are as Figure 3 shown.

[0054] In another exemplary embodiment of the present application, the CNN-Transformer hybrid model includes: a convolutional neural network and a Transformer network. The convolutional neural network is used to input multiple intrinsic mode functions and frequency domain signals corresponding to vibration signals and extract local features. The Transformer network is used to output the fault diagnosis result of the wind turbine bearing according to the local features extracted by the convolutional neural network. The Transformer network structure is as Figure 4 shown.

[0055] The convolutional neural network extracts local features from the input signal through convolutional operations. The forward propagation formula of the convolutional layer is:

[0056] y = f(W·x + b);

[0057] where x is the input data, W is the weight matrix of the convolutional kernel or filter, b is the bias term, f is the activation function used to introduce non-linearity, and y is the output after the convolutional operation. The goal of the CNN is to extract local features in the signal to help identify fault features.

[0058] In fault diagnosis, the self-attention mechanism of the Transformer network can capture long-range dependencies between input signals. Its calculation formula is as follows:

[0059]

[0060] where Attention() represents the obtained attention value, Q is the query vector, K is the key vector, V is the value vector, and d k is the dimension of the key. By calculating the dot product between the query and the key and normalizing it through softmax, the Transformer can dynamically adjust the weights of each signal part to capture important information in the signal.

[0061] In another exemplary embodiment of the present application, in order to further improve the accuracy and efficiency of fault diagnosis, the Meta-BOHB optimizer is used to automatically adjust the hyperparameters of the CNN-Transformer model. After VMD-FFT processing, the generated dataset is used as input and sent to the CNN-Transformer hybrid model optimized by the Meta-BOHB optimizer. The Meta-BOHB optimizer automatically adjusts the hyperparameters of the convolutional neural network and the Transformer network through meta-learning methods, thereby improving the diagnostic accuracy of the model. This optimization process significantly reduces the workload of manual parameter tuning and speeds up the training process, providing important support for improving the accuracy of the diagnostic model.

[0062] The hyperparameters of the CNN-Transformer hybrid model optimized by the meta-learning-HyperBand Bayesian optimizer include: the hyperparameters of the convolutional neural network and the Transformer network respectively, as well as the common hyperparameters of the convolutional neural network and the Transformer network.

[0063] The optimization objective of the Meta-BOHB optimizer is to minimize the following loss function:

[0064]

[0065] where L(θ) is the loss function for hyperparameter optimization, p(θ) is the probability distribution of the hyperparameters, and L(f(x i ,θ)) is the loss calculated based on the hyperparameters θ and the input data x i . By repeatedly adjusting the hyperparameters of the model, Meta-BOHB can quickly find the optimal combination of hyperparameters, thereby improving the accuracy of the model and reducing the computational cost.

[0066] In another exemplary embodiment of the present application, step 105 above can be replaced by the following steps 201 to 202:

[0067] Step 201: Divide the dataset into a training set, a validation set, and a test set.

[0068] Step 202: According to the training set and the validation set, the meta-learning-HyperBand Bayesian optimizer automatically optimizes the hyperparameters of the CNN-Transformer hybrid model through meta-learning methods.

[0069] The specific process of the above step 202 is as Figure 5Part of it: The training data saved by meta-learning is used to train the CNN-Transformer model with the BOHB optimizer, optimizing the hyperparameters of CNN, the hyperparameters of Transformer, and the common hyperparameters of both. The model is evaluated according to the test data. When convergence is achieved or the number of iterations is reached, the optimal hyperparameters are output. Otherwise, the BOHB optimizer is used to continue training the CNN-Transformer hybrid model. Among them, the training data and test data are obtained by dividing the training set and the validation set.

[0070] In another exemplary embodiment of the present application, the above step 106 can be replaced by the following steps 301 to 303:

[0071] Step 301: Use the training set and the validation set to train the optimized CNN-Transformer hybrid model.

[0072] Step 302: If the trained CNN-Transformer hybrid model converges or reaches the maximum number of iterations, use the test set to evaluate the performance of the trained CNN-Transformer hybrid model, and use the trained CNN-Transformer hybrid model with satisfactory performance as the wind turbine bearing fault diagnosis model.

[0073] Step 303: If the trained CNN-Transformer hybrid model does not converge and does not reach the maximum number of iterations, return to the step "Use the training set and the validation set to train the optimized CNN-Transformer hybrid model".

[0074] The optimized CNN-Transformer hybrid model is used to train and classify the data set, and the model can accurately identify and classify various types of wind turbine bearing faults. The experimental results show that in the final test set task, the model achieved correct classification of all fault types, verifying the powerful performance of the model. Its confusion matrix is as Figure 6 shown.

[0075] In another exemplary embodiment of the present application, referring to Figure 5 , after step 108, the method may further include: According to the multiple intrinsic mode functions and frequency domain signals corresponding to the real-time vibration signal, and the fault diagnosis result of the wind turbine bearing, use the meta-learning-HyperBand Bayesian optimizer to optimize the hyperparameters of the wind turbine bearing fault diagnosis model.

[0076] Still referring to Figure 5 , the general process of the method of the present application is as follows:

[0077] Step 1: Input signal, collect vibration signals from the bearing.

[0078] Step 2: Resampling, resample the signal to ensure the consistency of data format or sampling rate.

[0079] Step 3: Signal processing, perform two types of processing on the resampled signal respectively: perform frequency-domain analysis on the signal using the fast Fourier transform; perform time-domain analysis using variational mode decomposition. Overlap and integrate the processed signals in the time domain.

[0080] Step 4: Data partitioning, divide the data into three parts: Training set: used for training the model. Validation set: used for tuning and validating the model. Test set: used to evaluate the performance of the final model.

[0081] Step 5: Meta-learning and hyperparameter optimization, the stored training data will be used for meta-learning. Hyperparameters include three categories: hyperparameters specific to convolutional neural networks, common hyperparameters of convolutional neural networks and Transformers, and hyperparameters specific to Transformers. Use Bayesian optimization combined with Hyperband (BOHB) to optimize the hyperparameters.

[0082] Step 6: Model training, train the CNN-Transformer hybrid model using the optimized hyperparameters.

[0083] Step 7: Model evaluation, evaluate the performance of the model based on the test data.

[0084] Step 8: Iterative check, if the desired accuracy or the maximum number of iterations is reached, continue; otherwise, return for further optimization.

[0085] Step 9: Final evaluation and model saving, calculate the accuracy and loss metrics. Select the final CNN-Transformer hybrid model, save the model with the best hyperparameters, and output the diagnostic results.

[0086] Step 10: Result output, store the trained model and diagnostic results to complete the whole process.

[0087] To verify the functionality of each module, the following experiments are conducted:

[0088] 1. VMD-FFT ablation experiment, the experimental results are as Figure 7 shown, where Figure 7 part (a) in it is using only variational mode decomposition; Figure 7 part (b) in it is using only the fast Fourier transform; Figure 7Part (c) in it is to use variational mode decomposition and fast Fourier transform simultaneously. It can be seen from the experimental results that the combined use of VMD-FFT plays an ideal role in the in-depth mining of fault features by the model.

[0089] 2. Comparative experiments were carried out on different traditional models with and without adding optimizers, and the experimental results are as Figure 8 shown. It is not difficult to see that BOHB has obvious improvements to varying degrees for each model, and the CNN-Transformer hybrid model proposed in this application has achieved the best results.

[0090] 3. BOHB was compared with other traditional optimizers, and the experimental results are shown in Table 2. The BOHB proposed in this application not only has better loss values than other models, but also requires relatively less time.

[0091] Table 2 Comparative experiments of optimizers

[0092] Optimizer Best Loss Required Time BOHB 0.0421 12m11s Bayesian Optimization 0.0453 18m30s Hyperband 0.0487 10m07s Sparrow Search Algorithm 0.0528 24m23s

[0093] 4. The dataset was reconstructed with different step sizes, and the performance of the model on the new dataset was observed. The experimental results are shown in Table 3. The Meta-BOHB optimizer proposed in this application achieves the best loss with the minimum number of iterations, which is sufficient to illustrate the superiority of the method proposed in this paper.

[0094] Table 3 Performance of optimizers on the new dataset

[0095] Optimizer Required Iteration Times Best Loss Meta-BOHB 5 0.0418 BOHB 13 0.0421 Bayesian Optimization 18 0.0450

[0096] The method of this application combines the time-frequency domain feature extraction of variational mode decomposition and fast Fourier transform, the convolutional neural network and Transformer model improved by meta-learning-Bayesian optimization combined with HyperBand (Meta-BOHB), effectively solving the problems of dealing with non-stationary signals and noise interference in traditional methods. The combination of variational mode decomposition and fast Fourier transform is used for non-stationary signal decomposition and frequency domain feature extraction, and the combination of convolutional neural network and Transformer enhances the local feature extraction ability and long-term dependence modeling, thus improving the accuracy of fault diagnosis and the robustness of the model. At the same time, meta-learning-HyperBand Bayesian optimization (Meta-BOHB) accelerates the hyperparameter optimization through meta-learning, reduces the time and resources required for manual adjustment, and significantly improves the efficiency and application reliability of the diagnostic model. This application is applicable to the fault diagnosis of rotating machinery such as fan bearings and has high practical application value.

[0097] Based on time-frequency domain analysis and Meta-BOHB optimization of CNN-Transformer, this application proposes a fault diagnosis method for wind turbine bearings, effectively solving the challenges faced by existing technologies in wind turbine bearing fault diagnosis, especially the problems of dealing with non-stationary signals and noise interference under complex working conditions. In addition, this method also overcomes the limitations of manual adjustment of hyperparameters, which is both time-consuming and error-prone, and significantly improves the efficiency and reliability of the model in practical applications.

[0098] Based on the same inventive concept, the embodiment of this application also provides a wind turbine bearing fault diagnosis device for implementing the above-mentioned wind turbine bearing fault diagnosis method. The solution provided by this device to solve problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the wind turbine bearing fault diagnosis device provided below can refer to the limitations on the wind turbine bearing fault diagnosis method in the above text, and will not be elaborated here.

[0099] In an exemplary embodiment, a wind turbine bearing fault diagnosis device is provided, including: a data collection module, a decomposition module, a conversion module, a merging module, an optimization module, a training module, a real-time acquisition module, and a diagnosis module.

[0100] The data collection module is used to collect the vibration signals of the wind turbine bearing in the normal state and different fault types by means of resampling, and label the normal state or fault type as the label of the vibration signal.

[0101] The decomposition module is used to decompose each vibration signal into multiple intrinsic mode functions using variational mode decomposition.

[0102] The conversion module is used to convert each vibration signal from the time domain to the frequency domain using the fast Fourier transform to obtain the frequency domain signal.

[0103] The merging module is used to merge the multiple intrinsic mode functions and frequency domain signals corresponding to each vibration signal as inputs, and together with the label of each vibration signal, form a data set.

[0104] The optimization module is used to optimize the hyperparameters of the CNN-Transformer hybrid model according to the data set using the meta-learning-HyperBand Bayesian optimizer.

[0105] The training module is used to train the optimized CNN-Transformer hybrid model using the data set to obtain a wind turbine bearing fault diagnosis model.

[0106] The real-time acquisition module is used to collect the real-time vibration signals of the wind turbine bearing during operation, and through variational mode decomposition and fast Fourier transform, obtain the multiple intrinsic mode functions and frequency domain signals corresponding to the real-time vibration signals.

[0107] A diagnosis module, configured to input a plurality of intrinsic mode functions and frequency domain signals corresponding to real-time vibration signals into the wind turbine bearing fault diagnosis model to obtain a fault diagnosis result of the wind turbine bearing; the fault diagnosis result is a normal state or a fault type.

[0108] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0109] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A wind turbine bearing fault diagnosis method, characterized in that: include: By using the resampling method, the vibration signals of the wind turbine bearings in normal state and different fault types are collected, and the normal state or fault type is marked as the label of the vibration signal; Each vibration signal is decomposed into multiple intrinsic mode functions using variational mode decomposition; Each vibration signal is converted from the time domain to the frequency domain using a fast Fourier transform to obtain a frequency domain signal; The multiple intrinsic mode functions and frequency domain signals corresponding to each vibration signal are combined as input, and together with the label of each vibration signal, a data set is formed; Based on the dataset, the meta-learning-HyperBand Bayesian optimizer is used to optimize the hyperparameters of the CNN-Transformer hybrid model; The optimized CNN-Transformer hybrid model is trained using the data set to obtain a wind turbine bearing fault diagnosis model; Collect the real-time vibration signal of the wind turbine bearing under operation, and obtain multiple inherent mode functions and frequency domain signals corresponding to the real-time vibration signal through variational mode decomposition and fast Fourier transform; Multiple natural mode functions and frequency domain signals corresponding to the real-time vibration signal are input into the wind turbine bearing fault diagnosis model to obtain a wind turbine bearing fault diagnosis result; the fault diagnosis result is a normal state or a fault type.

2. The wind turbine bearing fault diagnosis method according to claim 1, characterized in that: The number of samples of vibration signals of wind turbine bearings in normal state and under different fault types collected by resampling is: Where N is the number of samples, L is the total length of the vibration signal, T is the length of a single sample signal, and S is the step size.

3. The wind turbine bearing fault diagnosis method according to claim 1, characterized in that: Labels that indicate normal status or fault type as vibration signals, including: If the fault type is: inner race fault with a fault size of 0.007 inches, the label of the vibration signal is 1; If the fault type is: inner race fault with a fault size of 0.014 inches, the label for the vibration signal is 2; If the fault type is: inner race fault with a fault size of 0.021 inches, the label for the vibration signal is 3; If the fault type is: outer race fault with a fault size of 0.007 inches, the label for the vibration signal is 4; If the fault type is: outer race fault with a fault size of 0.014 inches, the label for the vibration signal is 5; If the fault type is: outer race fault with a fault size of 0.021 inches, the label for the vibration signal is 6; If the fault type is: ball fault with a fault size of 0.007 inches, the label for the vibration signal is 7; If the fault type is: ball fault with a fault size of 0.014 inches, the label for the vibration signal is 8; If the fault type is: ball fault with a fault size of 0.021 inches, the label for the vibration signal is 9; If it is in normal state, the label of the vibration signal is 10.

4. The wind turbine bearing fault diagnosis method according to claim 1, characterized in that: The optimization objective function of the variational mode decomposition is: Among them, u k (t) represents the kth intrinsic mode function, for u k (t), α represents the balance factor, K represents the number of intrinsic mode functions, t represents time, and || ||2 represents the L2 norm.

5. The wind turbine bearing fault diagnosis method according to claim 1, characterized in that: The CNN-Transformer hybrid model includes: a convolutional neural network and a Transformer network; The convolutional neural network is used to input multiple intrinsic mode functions and frequency domain signals corresponding to the vibration signal and extract local features; The Transformer network is used to output the fault diagnosis results of wind turbine bearings based on the local features extracted by the convolutional neural network.

6. The wind turbine bearing fault diagnosis method according to claim 5, characterized in that: The hyperparameters of the CNN-Transformer hybrid model optimized by Meta-Learning-HyperBand Bayesian Optimizer include: the hyperparameters of the convolutional neural network and the transformer network respectively, as well as the common hyperparameters of the convolutional neural network and the transformer network.

7. The wind turbine bearing fault diagnosis method according to claim 1, characterized in that: Based on the dataset, the meta-learning-HyperBand Bayesian optimizer is used to optimize the hyperparameters of the CNN-Transformer hybrid model, including: Dividing the data set into a training set, a validation set, and a test set; According to the training set and the validation set, the meta-learning-HyperBand Bayesian optimizer automatically optimizes the hyperparameters of the CNN-Transformer hybrid model through a meta-learning method.

8. The wind turbine bearing fault diagnosis method according to claim 7, characterized in that: The optimized CNN-Transformer hybrid model is trained using the data set to obtain a wind turbine bearing fault diagnosis model, which specifically includes: Using the training set and the validation set to train the optimized CNN-Transformer hybrid model; If the trained CNN-Transformer hybrid model converges or reaches the maximum number of iterations, the performance of the trained CNN-Transformer hybrid model is evaluated using the test set, and the trained CNN-Transformer hybrid model that meets the performance requirements is used as a wind turbine bearing fault diagnosis model; If the trained CNN-Transformer hybrid model does not converge and the maximum number of iterations is not reached, return to step "training the optimized CNN-Transformer hybrid model using the training set and the validation set".

9. The wind turbine bearing fault diagnosis method according to claim 1, characterized in that: Inputting a plurality of natural mode functions and frequency domain signals corresponding to the real-time vibration signal into the wind turbine bearing fault diagnosis model to obtain a wind turbine bearing fault diagnosis result, and then further comprising: According to multiple intrinsic mode functions and frequency domain signals corresponding to real-time vibration signals and fault diagnosis results of wind turbine bearings, meta-learning-HyperBand Bayesian optimizer is used to optimize the hyperparameters of the wind turbine bearing fault diagnosis model.

10. A wind turbine bearing fault diagnosis device, characterized in that: The wind turbine bearing fault diagnosis device comprises: A data collection module is used to collect vibration signals of wind turbine bearings in normal state and different fault types by resampling, and mark the normal state or fault type as a label of the vibration signal; a decomposition module for decomposing each vibration signal into a plurality of intrinsic mode functions using variational mode decomposition; A conversion module, used for converting each vibration signal from the time domain to the frequency domain using a fast Fourier transform to obtain a frequency domain signal; A merging module, used for merging multiple intrinsic mode functions and frequency domain signals corresponding to each vibration signal as input, and forming a data set together with a label of each vibration signal; An optimization module, for optimizing hyperparameters of a CNN-Transformer hybrid model using a meta-learning-HyperBand Bayesian optimizer according to the data set; A training module, used to train the optimized CNN-Transformer hybrid model using the data set to obtain a wind turbine bearing fault diagnosis model; The real-time acquisition module is used to collect the real-time vibration signal of the wind turbine bearing under operation, and obtain multiple inherent mode functions and frequency domain signals corresponding to the real-time vibration signal through variational mode decomposition and fast Fourier transform; The diagnosis module is used to input multiple natural mode functions and frequency domain signals corresponding to the real-time vibration signal into the wind turbine bearing fault diagnosis model to obtain the fault diagnosis result of the wind turbine bearing; the fault diagnosis result is a normal state or a fault type.

Images