A wind turbine gearbox fault diagnosis method based on digital test
By constructing digital models and generating enhanced datasets using generative adversarial networks, the problems of data imbalance and high cost in traditional wind turbine gearbox fault diagnosis have been solved, achieving efficient and accurate fault diagnosis and ensuring the safe and stable operation of wind turbines.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2026-02-04
- Publication Date
- 2026-06-02
AI Technical Summary
Traditional wind turbine gearbox fault diagnosis methods rely on actual collected data, which suffers from data imbalance, leading to inapplicable diagnostic results. Furthermore, physical testing is costly and time-consuming, affecting the safe and stable operation of the unit.
A high-fidelity multiphysics wind turbine gearbox digital model was constructed. By combining generative adversarial networks (GANs) and deep learning, an enhanced dataset was generated through fault injection to train a fault diagnosis model and achieve fault diagnosis in a virtual environment.
It significantly improves the accuracy and efficiency of fault diagnosis, reduces the cost of physical testing, and ensures the safe and stable operation of wind turbine gearboxes.
Smart Images
Figure CN122132877A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of electronic engineering and computer science, and specifically relates to a method for diagnosing faults in wind turbine gearboxes based on digital testing. Background Technology
[0002] Wind energy, as an important clean energy source, is playing an increasingly vital role in the global energy structure. As a core power generation unit, the safe and stable operation of wind turbines is crucial to the power system. Gearboxes play a critical role in the transmission of wind turbines and operate under variable loads, high speeds, and diverse conditions, making them highly susceptible to fatigue, wear, tooth breakage, and bearing spalling. Gearbox failure not only incurs high repair costs but also leads to turbine downtime, resulting in significant economic losses. Furthermore, traditional fault diagnosis methods are typically based solely on actual collected data, and in practical applications, data imbalance becomes a significant issue, rendering the results of traditional methods potentially inapplicable. Summary of the Invention
[0003] This invention provides a fault diagnosis method for wind turbine gearboxes based on digital testing. The method covers the design of a digital modeling module, a data enhancement module for data fusion and simulation, and a fault diagnosis model module for wind turbine gearboxes. To a certain extent, it can ensure the safety of wind turbine gearboxes, reduce physical testing costs, improve testing efficiency, and adjust the operating conditions of equipment in real time.
[0004] The technical solution of this invention: A method for fault diagnosis of wind turbine gearbox based on digital testing, comprising:
[0005] Step S110: Construct a high-fidelity multiphysics wind turbine gearbox digital model, and use the relevant features and historical data of the wind turbine gearbox to complete the iteration and verification of the wind turbine gearbox digital model;
[0006] Step S120: Using a digital model, simulated operating condition data under normal conditions is obtained. Fault injection technology is used to input faults into the digital model to obtain simulated operating condition data under fault conditions. A generative adversarial network is constructed and trained using the historical data and simulated operating condition data under fault conditions from step S110. The input of the network is the simulated operating condition data under fault conditions, and the output is the enhanced fault data. Physical consistency constraints and data statistical feature constraints are introduced into its loss function. The enhanced fault data is integrated with historical data and real fault data to form an enhanced dataset.
[0007] Step S130: Based on the augmented dataset obtained in step S120, supervised learning training is performed on the fault diagnosis model. Data features are extracted and classified through the fault diagnosis model to realize the fault diagnosis function. At the same time, K-fold cross-validation is introduced to improve the fault diagnosis model.
[0008] A computing device includes: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device performs the method.
[0009] A readable storage medium storing program instructions that, when read and executed by a computing device, cause the computing device to perform the method.
[0010] Beneficial effects:
[0011] This invention effectively solves the problems of high cost and long cycle of physical testing by constructing a test environment that combines virtual and real elements. It can significantly improve the accuracy of fault diagnosis and test efficiency, and ensure the safe and stable operation of wind turbine gearboxes. Attached Figure Description
[0012] Figure 1 This is a flowchart of a wind turbine gearbox fault diagnosis method based on digital testing according to the present invention. Detailed Implementation
[0013] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0014] This invention provides a fault diagnosis method for wind turbine gearboxes based on digital testing, including the design of a digital modeling module, a data enhancement module for data fusion and a fault diagnosis model module. It can ensure the safety of wind turbine gearboxes to a certain extent, reduce physical testing costs, improve fault diagnosis efficiency, and adjust the operating conditions of equipment in real time. Figure 1 A flowchart illustrating this diagnostic method is provided. Figure 1 As shown, the method includes the following steps:
[0015] Step S110: Construct a high-fidelity multiphysics wind turbine gearbox digital model, and use relevant features and historical data of the wind turbine gearbox to complete the iteration and verification of the digital model of the wind turbine gearbox.
[0016] Step S110-1: Using analytical methods, extract the following from the CAD model: geometric parameters (module, number of teeth, pressure angle), material properties (elastic modulus, Poisson's ratio, density), thermophysical parameters (specific heat capacity, thermal conductivity), and transmission chain topology characteristics of the wind turbine gearbox. Construct a digital model of the wind turbine gearbox based on a physical mechanism-based digital modeling method. In the digital modeling, based on Newton's second law and Lagrange's equations, establish a system vibration differential equation set to describe the rigid body motion and elastic deformation of the gearbox. This is done using the thermal network method. Method: Establish nodal thermal balance equations to describe the conduction and convection of heat between gear oil, gearbox housing, and the environment. Simplify the complex continuum of gears, bearings, etc. in the gearbox into idealized units with lumped mass, lumped stiffness, and lumped damping. Based on this, construct the differential equations of the system, i.e., construct the "white box part," to handle the known mechanism of the transmission chain and material properties. At the same time, utilize a large amount of historical data and a data-driven digital modeling method to construct the "black box part" of the system with the help of neural networks to handle uncertain nonlinear factors.
[0017] Step S110-2: Based on the relevant characteristics and historical data of the wind turbine gearbox, complete the iteration and verification of the digital model of the wind turbine gearbox. That is, use the relevant characteristics of the wind turbine gearbox as the attributes of the digital model of the wind turbine gearbox; use the interfaces and data inputs of each component in the historical data as the input of the digital model; compare the output of the wind turbine gearbox in the historical data with the output of the digital model. If the average accuracy of the state curves corresponding to the mathematical and physical characteristics simulated by the digital model under normal temperature and pressure is within the design allowable range, that is, the average error is less than m%, the peak error is less than n%, where n and m are both preset thresholds and n>m>0, the correctness of the digital model can be determined; otherwise, the digital model of the wind turbine gearbox is corrected.
[0018] The construction process of step S110-1 may include: extracting the physical characteristics of the gearbox material, thermodynamic and kinetic properties, combining the physical mechanism and using digital modeling methods such as neural networks and principal component analysis to realize the digital model.
[0019] Furthermore, in a specific example, the process may include: extracting the actual physical characteristics of the gearbox and constructing a digital model of the gearbox using digital experimentation methods. The gearbox module can be divided into n = 3 functional modules, including the planetary transmission stage module, the high-speed parallel shaft stage module, and the gearbox housing / shaft system module. The planetary transmission stage module is further subdivided into 3 sub-modules: the sun gear, planet gears, and the internal gear ring; the high-speed parallel shaft stage module is subdivided into 2 sub-modules: the high-speed shaft and the high-speed stage gear. The aforementioned physical characteristics include the dynamic characteristics, transmission characteristics, reliability and life characteristics, and material characteristics of each sub-module. Specifically, this includes the material characteristics of the gears and bearings being made of alloy steel; the dynamic characteristics of the gearbox under varying speed and torque conditions, such as vibration and noise; the thermodynamic characteristics of the gearbox during operation, including frictional heat generation, lubricating oil flow, and heat dissipation; the transmission characteristics of the gear transmission system, such as the speed ratio, transmission efficiency, and torque carrying capacity; and the reliability and life characteristics of key components, such as design life, fatigue strength, and failure modes. Subsequently, the physical characteristics of the module were digitally modeled based on digital experiments, constructing material strength equations, multibody dynamics equations, vibroacoustic multimodal equations, thermodynamic equations, and governing equations for the digital module. For the digital modeling method, a mechanism- and data-driven hybrid approach was adopted, rapidly establishing a mechanistic model using data, capable of handling nonlinear and multivariable coupled problems. Neural networks and principal component analysis were used to correct nonlinear parameters (such as time-varying meshing stiffness and damping) in the mechanistic model and establish a digital model.
[0020] The iterative and verification process of step S110-2 includes ensuring that the error between the simulation output of the digital model and the physical entity response converges to the preset design tolerance range under extreme wind speed changes caused by low temperature freezing environment, low air density high altitude environment, and gusts and turbulence.
[0021] Furthermore, this process may include: continuously optimizing and validating the digital model of the gearbox based on its inherent correlation attributes and historical operating data. The inherent correlation attributes encompass its geometric parameters and the physical properties of its constituent materials. Input variables provided by various equipment interfaces in historical operating records (e.g., wind speed, transmission torque, rotational speed) are used as excitation conditions for the digital model, and its output results (e.g., vibration response, temperature rise change) are compared and analyzed with actual historical observations. If, under normal operating conditions, the state evolution curves corresponding to the aforementioned mathematical and physical behaviors reproduced by the digital model have an average degree of fit within the allowable range of engineering design—that is, an average deviation of no more than 5% (denoted as m=5), and a maximum instantaneous deviation controlled within 10% (denoted as n=10)—then the digital model can be deemed effective; otherwise, the digital model of the gearbox needs to be corrected and recalibrated.
[0022] Step S120: Using a digital model, simulated operating condition data under normal conditions is obtained as a clean base for the augmented dataset; fault injection technology is used to input faults into the digital model to obtain simulated operating condition data under fault conditions; a generative adversarial network (GAN) is constructed and trained using the historical data and simulated operating condition data under fault conditions from step S110. The input of the network is historical data and simulated operating condition data under fault conditions, and the output is augmented fault data. Physical consistency constraints and data statistical feature constraints are introduced into its loss function to integrate fault data with historical data and real fault data to form an augmented dataset.
[0023] Among them, operating condition data refers to wind speed, transmission torque, rotation speed, vibration response, temperature rise, noise, torque, temperature and air density, etc.
[0024] Obtaining simulated operating condition data under fault conditions can include inputting gear set scratches, downwind bearing overheating, lubricating oil viscosity failure, gear set fretting corrosion, scratches, polishing wear, ring gear / sun pinion scratches, polishing wear, and fretting corrosion as fault excitations into the digital model to obtain simulated operating condition data under fault conditions.
[0025] Physical consistency constraints and data statistical feature constraints are introduced into the loss function to obtain gearbox fault information that is closer to reality. Finally, a class-balanced augmented dataset is constructed as the input of the fault diagnosis model.
[0026] In a specific example, the above process may include: conducting digital experiments using the digital model constructed in step S110, obtaining normal data through long-term digital test runs under normal operating conditions; and referencing potential faults encountered in actual wind turbine gearbox operation, such as gear set scratches, downwind bearing overheating, lubricating oil viscosity failure, gear set fretting corrosion, scratches, polishing wear, and ring gear / sun pinion scratches, polishing wear, and fretting corrosion, injecting these faults as faults to execute digital experiments, thereby obtaining fault data under simulated operating conditions. To simulate bearing inner ring damage, the support force function in the bearing dynamics model... A periodic impact pulse sequence with a frequency of 180Hz (the characteristic frequency of a certain type of high-speed shaft bearing) is applied; to simulate pitting on the tooth surface, the time-varying meshing stiffness function in the gear meshing model is adjusted. A 15% periodic stiffness decrease is applied to simulate transient load changes; to simulate lubricant failure, a predetermined simulation time point is applied. The contact friction coefficient between the gear and the bearing in the model A sudden increase of 50% to 100% simulates the rupture of the lubricating film, leading to direct metal-to-metal contact and a dramatic increase in friction.
[0027] By combining the real data (historical data) obtained in step S110 with the simulated fault data obtained through digital experiments using a digital model achieved through data-model fusion, a Transformer-based improved GAN network is trained. Furthermore, by incorporating actual physical constraints and data characteristics, fault data that more closely resembles real data is generated. This fault data is then integrated with historical data and real fault data to obtain an augmented dataset that balances fault and health data. "Balance" refers to the process of data augmentation to ensure that the number of samples in each category (such as health status and various fault types) in the training data is balanced, thereby improving the performance of subsequent diagnostic models.
[0028] A Transformer-based GAN network, hereinafter referred to as the T-GAN architecture, is proposed for processing time-series signals.
[0029] The Generator G uses a Transformer architecture to process time-series data, such as vibration, collected by sensors in the wind turbine gearbox. The time-series signal is sliced at the input layer and transformed into a Transformer embedding vector through a linear mapping layer.
[0030] The encoder part is composed of Composed of several Transformer blocks It is a natural number greater than or equal to 1, in one implementation method Each block contains a multi-head self-attention mechanism and a feedforward network, enabling it to capture long-term temporal dependencies of fault features; the decoder is also composed of... Composed of several Transformer blocks, it adds cross attention after self-attention to fuse the features of the encoder output and the target domain, and reconstructs a time-series signal with real noise patterns.
[0031] The discriminator D is implemented using a classification network corresponding to the generator G. Its task is to distinguish whether the input data comes from the real sensor or is generated by the generator. Its input is historical data and generated augmented data, and its output is a scalar value representing the probability and distance between the input data and the real data distribution.
[0032] In order to fool the discriminator, the generator G continuously adjusts its own parameters so that the generated signal approaches the real data in terms of noise distribution and statistical characteristics.
[0033] To incorporate physical constraints and data feature constraints, the following combined loss function is adopted. :
[0034] ,
[0035] In the formula, Let GAN be the loss function. These are the loss weights and loss function for cycle consistency, respectively. The value is set to 10.0 to ensure the continuity and reversibility of the data over time. These are the loss weights and loss function for physical constraints, respectively. The value is set to 5.0 to ensure that the generated fault features (such as BPFI (Ball Pass Frequency, Inner race) frequency energy) are valid, and to prevent GAN from generating physically impossible signals; These are the loss coefficient and loss function, respectively, for data feature constraints. The value is set to 1.0 to flexibly adjust the domain gap and prevent it from excessively affecting physical characteristics.
[0036] Among them, the loss function of GAN network for:
[0037] ,
[0038] In the formula, Represents real data. Represents enhanced data, Represents interpolated data. Represents the gradient penalty weights. Represents the gradient. Represents the discriminator. Represents the gradient operator. It represents the mathematical expectation.
[0039] Cyclic consistency loss function for:
[0040] ,
[0041] In the formula, Represents a virtual fault signal. Represents the reverse engineer, responsible for converting augmented data back into the virtual domain; Represents a generator. This represents the L1 norm.
[0042] Physical constraint loss for:
[0043] ,
[0044] In the formula, The energy spectrum representing the fast Fourier transform. Represents the set of fault characteristic frequencies. Indicates frequency index.
[0045] Data feature constraints for:
[0046] ,
[0047] In the formula, Representing the maximum mean difference, this is a non-parametric distance metric commonly used to measure the difference between two data distributions. This represents the feature mapping function, which is used in neural networks to map the original signal to a high-dimensional feature space.
[0048] The hyperparameters of the GAN are optimized by combining the above loss function, and finally the virtual fault signal obtained from the digitization model is used. As input to the trained GAN generator G, a virtual enhanced fault signal with similar temporal patterns, noise distribution, and statistical characteristics to real-world fault signals is obtained. By integrating a large amount of real health status data and a small amount of real fault data from a real wind turbine monitoring system, an enhanced dataset balancing fault and health status is obtained, which serves as the input for subsequent fault diagnosis models.
[0049] In the training system of GAN, simulated data is fed into the generator to transform into enhanced signals. Through physical constraint loss and cycle consistency loss, it is ensured that the generated results retain the core fault frequency characteristics and have reversibility. Historical data is used to train the discriminator to identify the real and fake data, and through data feature constraints, it forces the generated signals to infinitely approximate the real environment in terms of statistical distribution. The two are deeply coupled through a combined loss function that includes physical, statistical and adversarial components, and jointly drive the model to generate enhanced data that contains both clear physical fault mechanisms and real working condition texture.
[0050] In step S130, based on the augmented dataset obtained in step S120, supervised learning training is performed on the fault diagnosis model. Data features are extracted and classified through the fault diagnosis model to realize the fault diagnosis function. At the same time, K-fold cross-validation is introduced to improve the fault diagnosis model.
[0051] The specific implementation of the fault diagnosis model's training and feature classification is as follows:
[0052] A fault diagnosis model based on a deep learning architecture is selected, specifically the Transformer architecture, which utilizes a self-attention mechanism to capture long-range dependencies in time-series data and fuse multimodal data. The augmented balanced dataset obtained in step S120 is used as the training set to train the selected deep learning-based fault diagnosis model. The fault diagnosis model extracts fault-sensitive features from the augmented dataset through its network layers and classifies them through a classification layer to achieve fault diagnosis.
[0053] By introducing cross-validation techniques such as 5-fold cross-validation, the augmented dataset is divided into 5 subsets for 5 training and validation cycles to evaluate the robustness and generalization ability of the model. Finally, the hyperparameters of the model are iteratively optimized and adjusted using three performance metrics: mean accuracy, recall, and F1 score. The final result is a fault diagnosis model for wind turbine gearboxes.
[0054] The fault diagnosis process may include: combining the above digital models, GANs, and fault diagnosis models to form a virtual wind turbine gearbox fault diagnosis system based on digital testing; introducing the digital twin method to deeply couple the fault diagnosis system with the actual wind turbine gearbox; realizing data and command connection between the digital models, GANs, and fault diagnosis models and the actual equipment; obtaining enhanced fault data through GANs; training the fault diagnosis model; and achieving closed-loop operation and continuous optimization of the system through real-time iterative updates.
[0055] In summary, this invention discloses a wind turbine gearbox fault diagnosis method based on digital testing, including the design of a wind turbine gearbox digital modeling module, a data enhancement module for data fusion and real-data integration, and a fault diagnosis model module. This method can, to a certain extent, ensure the safety of wind turbine gearboxes, reduce physical testing costs, improve testing efficiency, adjust equipment operating conditions in real time, and improve the fault diagnosis process.
[0056] The present invention also provides a computing device, comprising: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device performs the method described thereon.
[0057] The present invention also provides a readable storage medium storing program instructions, which, when read and executed by a computing device, cause the computing device to perform the method described thereon.
[0058] The contents not described in detail in this specification are existing technologies known to those skilled in the art.
[0059] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered to fall within the protection scope of the present invention.
Claims
1. A method for fault diagnosis of wind turbine gearboxes based on digital testing, characterized in that, include: Step S110: Construct a high-fidelity multiphysics wind turbine gearbox digital model, and use the relevant features and historical data of the wind turbine gearbox to complete the iteration and verification of the wind turbine gearbox digital model; Step S120: Using a digital model, simulated operating condition data under normal conditions is obtained as a clean base for the augmented dataset. Fault injection technology is used to input faults into the digital model to obtain simulated operating condition data under fault conditions. A generative adversarial network is constructed and trained using the historical data and simulated operating condition data under fault conditions from step S110. The input of the network is historical data and simulated operating condition data under fault conditions, and the output is augmented fault data. Physical consistency constraints and data statistical feature constraints are introduced into its loss function to integrate the augmented fault data with historical data and real fault data to form an augmented dataset. Step S130: Based on the augmented dataset obtained in step S120, supervised learning training is performed on the fault diagnosis model. Data features are extracted and classified through the fault diagnosis model to realize the fault diagnosis function. At the same time, K-fold cross-validation is introduced to improve the fault diagnosis model.
2. The method for fault diagnosis of wind turbine gearbox based on digital testing according to claim 1, characterized in that, In step S110, constructing a high-fidelity multiphysics wind turbine gearbox digital model includes: Step S110-1: Using analytical methods, the following are extracted from the CAD model: geometric parameters, material properties, thermophysical parameters, and transmission chain topological characteristics of the gears and bearings of the wind turbine gearbox. A digital model of the wind turbine gearbox is constructed based on a physical mechanism-based digital modeling method. In digital modeling, a system vibration differential equation set is established based on Newton's second law and Lagrange's equations to describe the rigid body motion and elastic deformation of the gearbox. Based on the thermal network method, nodal thermal balance equations are established to describe the conduction and convection of heat between gear oil, gearbox housing, and the environment. The gear and bearing continuum in the gearbox is simplified into an idealized unit with lumped mass, lumped stiffness, and lumped damping. Based on this, the differential equations of the system are constructed, i.e., the white box part is constructed to handle the transmission chain and material properties with known mechanisms. At the same time, a large amount of historical data is used, and a data-driven digital modeling method is employed to construct the black box part of the system using neural networks to handle uncertain nonlinear factors. Step S110-2 involves iterating and verifying the digital model of the wind turbine gearbox based on its relevant characteristics and historical data. This includes: using the relevant characteristics of the wind turbine gearbox as attributes of the digital model; using the interfaces and data inputs of each component in the historical data as inputs to the digital model; comparing the outputs of the wind turbine gearbox in the historical data with the outputs of the digital model; and determining the correctness of the digital model if the average accuracy of the state curves corresponding to the mathematical and physical characteristics simulated by the digital model under normal temperature and pressure is within the design allowable range. Otherwise, the digital model of the wind turbine gearbox is corrected.
3. The method for fault diagnosis of wind turbine gearbox based on digital testing according to claim 2, characterized in that, Step S120 includes: Digital experiments are conducted using the digital model constructed in step S110. Normal data is obtained through long-term digital test operation under normal operating conditions. Faults encountered by the actual wind turbine gearbox are referenced and input into the digital model through fault injection technology and digital experiments are conducted to obtain fault data under simulated operating conditions. By combining historical data from step S110 with simulated fault data obtained through digital experiments, a generative adversarial network based on Transformer is trained. Fault data is generated by combining actual physical constraints and data features. Then, historical data and real fault data are combined with the fault data to obtain an enhanced dataset that balances fault data and healthy data.
4. The method for fault diagnosis of wind turbine gearbox based on digital testing according to claim 3, characterized in that, In the generative adversarial network based on the improved Transformer, the generator G uses the Transformer architecture to process the time-series data collected by the sensors in the gearbox of the wind turbine. The time-series signal is sliced in the input layer and transformed into the embedding vector of Transformer through the linear mapping layer. The encoder part is composed of Composed of several Transformer blocks It is a natural number greater than or equal to 1. Each block contains a multi-head self-attention mechanism and a feedforward network, which enables the capture of long-term temporal dependencies of fault features. The decoder is also made by It consists of several Transformer blocks, with cross attention added after self attention to fuse the features of the encoder output and the target domain, and reconstruct a time-series signal with real noise patterns; The discriminator D is implemented using a classification network corresponding to the generator G. Its task is to distinguish whether the input data comes from the real sensor or is generated by the generator. Its input is historical data and generated augmented data, and its output is a scalar value representing the probability and distance between the input data and the real data distribution. In order to fool the discriminator, the generator G continuously adjusts its own parameters so that the generated signal approaches the real data in terms of noise distribution and statistical characteristics.
5. The method for fault diagnosis of wind turbine gearbox based on digital testing according to claim 4, characterized in that, For generative adversarial networks, the following combined loss function is adopted. : ， In the formula, To generate the loss function for the adversarial network, These are the loss weights and loss function for cycle consistency, respectively. These are the loss weights and loss function for physical constraints, respectively. These are the loss coefficient and loss function for data feature constraints, respectively; Among them, the generative adversarial network loss function for: ， In the formula, Represents real fault data. Represents enhanced data, Represents interpolated data. Represents the gradient penalty weights. Represents gradient, Represents the discriminator. Represents the gradient operator. Represents mathematical expectation; Cyclic consistency loss function for: ， In the formula, This represents simulated operating condition data under fault conditions obtained from digital experiments. Represents the reverse engineer, responsible for converting augmented data back into the virtual domain; Represents a generator. Represents the L1 norm; Physical constraint loss for: ， In the formula, The energy spectrum representing the fast Fourier transform. Represents the set of fault characteristic frequencies. Indicates frequency index; Data feature constraints for: ， In the formula, Represents the maximum mean difference. The representative feature mapping function is used to map the original signal to a high-dimensional feature space; The hyperparameters of the generative adversarial network are optimized by combining the loss function, and finally the virtual fault signal obtained from the digitization model is used. As input to the generator G of the trained generative adversarial network, a virtual enhanced fault signal with similar temporal patterns, noise distribution, and statistical characteristics to the real signal is obtained. By integrating historical health data and real fault data from a real wind turbine monitoring system, an enhanced dataset balancing fault and health status is obtained, which serves as input for subsequent fault diagnosis models.
6. The method for fault diagnosis of wind turbine gearbox based on digital testing according to claim 3, characterized in that, In step S130, a fault diagnosis model based on the Transformer architecture is selected. This architecture uses a self-attention mechanism to capture long-distance dependencies and multimodal data fusion in time-series data. The augmented dataset obtained in step S120 is used as the training set to train the selected fault diagnosis model based on the deep learning architecture. The fault diagnosis model extracts fault-sensitive features from the enhanced dataset through its network layers and classifies them through a classification layer to achieve fault diagnosis.
7. The method for fault diagnosis of wind turbine gearbox based on digital testing according to claim 6, characterized in that, Step S130 also includes: introducing K-fold cross-validation, dividing the augmented dataset into K subsets for K training and validation cycles to evaluate the robustness and generalization ability of the fault diagnosis model; finally, iteratively optimizing and adjusting the hyperparameters of the fault diagnosis model using average accuracy, recall, and F1 performance metrics; and finally obtaining the fault diagnosis model for wind turbine gearboxes.
8. The method for fault diagnosis of wind turbine gearbox based on digital testing according to claim 7, characterized in that, The fault diagnosis process includes: combining the above digital models, generative adversarial networks, and fault diagnosis models to form a virtual wind turbine gearbox fault diagnosis system based on digital experiments; introducing the digital twin method to couple the fault diagnosis system with the actual wind turbine gearbox, realizing data and command connection between the digital models, generative adversarial networks, and fault diagnosis models and the actual equipment; obtaining enhanced fault data based on generative adversarial networks to train the fault diagnosis model; and achieving closed-loop operation and continuous optimization of the system through real-time iterative updates.
9. A computing device, characterized in that, include: At least one processor and a memory storing program instructions; When the program instructions are read and executed by the processor, the computing device performs the method as described in any one of claims 1-8.
10. A readable storage medium storing program instructions, characterized in that, When the program instructions are read and executed by the computing device, the computing device performs the method as described in any one of claims 1-8.