Method and device for detecting fan bearing of wind power generation equipment

Abstract

The invention discloses a method and device for detecting a fan bearing of wind power generation equipment, and the method comprises the steps: reversely inverting physical parameters from a small number of real observation signals, generating a high-fidelity synthesis fault data set constrained by a kinetic equation based on the real parameters, and carrying out the detection of the fan bearing of the wind power generation equipment. Therefore, the problems of model overfitting and poor generalization caused by scarcity of real fault data are solved under the small sample condition, physical distortion of a pure data driving generation method and excessive dependence of traditional simulation on prior parameters are overcome, and high-precision and high-robustness fan bearing fault grading detection is achieved.

Description

Technical Field

[0001] This invention relates to the field of wind turbine bearing testing technology for wind power generation equipment, and in particular to a method for testing wind turbine bearings of wind power generation equipment, a device for testing wind turbine bearings of wind power generation equipment, an electronic device, and a readable storage medium. Background Technology

[0002] With the continuous and rapid growth of wind power installed capacity, wind turbine bearings, as core components of the transmission chain, are subjected to multiple adverse factors such as complex variable loads, variable speeds, temperature fluctuations, and lubrication degradation over a long period of time. This makes them the weakest link in the entire machine most prone to fatigue spalling, wear, cracking, and other failures. Once the bearings suffer early damage or failure, it will not only lead to unplanned shutdowns of the unit and a sharp decline in power generation efficiency, but may also cause major safety accidents such as main shaft fracture, gearbox damage, or even the overturning of the entire machine, seriously threatening operation and maintenance costs and grid stability.

[0003] In recent years, deep learning-based methods for wind turbine bearing fault diagnosis have made significant progress, reducing human intervention to some extent through end-to-end feature learning. However, these methods heavily rely on a large number of high-quality labeled fault samples. In contrast, actual wind farm operation data primarily reflects healthy conditions, making real fault samples (especially severe mid-to-late stage faults) extremely scarce, and even making it difficult to collect enough samples for training. This leads to models being prone to overfitting under small sample conditions, exhibiting poor generalization ability and high false alarm / false negative rates, failing to meet the demands of engineering sites for high-precision, robust early fault warnings. Summary of the Invention

[0004] The present invention provides a method, apparatus, electronic device, and readable storage medium for testing wind turbine bearings in wind power generation equipment, in order to overcome or at least partially solve the above-mentioned problems.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for testing wind turbine bearings in wind power generation equipment, including: Collect the observation signal sequence of the wind turbine bearing, and perform standardization and time-series slicing on the observation signal sequence to generate a standardized sample sequence set; The actual engineering operating condition parameters of the wind turbine bearing are specified; the actual engineering operating condition parameters include: operating condition information characterizing the operating conditions of the wind turbine bearing in the real operating environment, geometric parameters characterizing the geometric characteristics of the wind turbine bearing, and the material mechanical properties of the wind turbine bearing; Based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties, the constraints of the dynamic equations are determined; Using the standardized sample sequence set and the constraints of the dynamic equation, an inverse physical information neural network is used to invert the inverse physical parameter vector implicit in the observed signal sequence, which is constrained by the physical feasible region. A multi-condition fault parameter set constrained by the dynamic equation is generated based on the inverted physical parameter vector. Based on the extended multi-condition fault parameter set, a wind turbine bearing fault data set is constructed, and the wind turbine bearing fault classification target model is generated using the wind turbine bearing fault data set. The target model for classifying the faults of the wind turbine bearing is controlled, and based on the target data for the target wind turbine bearing, the detection results for the target wind turbine bearing are generated.

[0006] Optionally, the step of determining the constraints of the dynamic equations based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties includes: Based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties, a degree-of-freedom coupled nonlinear dynamic equation is constructed. The time-varying meshing stiffness and bearing support stiffness of the wind turbine bearing are calculated using the aforementioned degree-of-freedom coupled nonlinear dynamic equations. The bending, shearing and axial compression deformation energies are calculated by integrating using the potential energy method, thereby generating a nonlinear time-varying stiffness matrix. The gear eccentricity error, multi-order machining error harmonic components, and bearing geometric dimension deviation are determined, and the nonlinear time-varying stiffness matrix is ​​adjusted using the gear eccentricity error, the multi-order machining error harmonic components, and the bearing geometric dimension deviation to generate the motion equation of the enhanced dynamic system. The motion equations of the enhanced dynamic system are solved numerically by time integration using the backward difference formula method, generating time-domain acceleration response curves of the input axis, intermediate axis and output axis in multiple directions; Perform a fast Fourier transform on the time-domain acceleration response curve to generate a one-sided amplitude spectrum; Frequency domain gain weighting is performed on the single-sided amplitude spectrum to generate an enhanced spectrum signal that highlights the characteristics of the fault impact. An input feature set is constructed based on the enhanced spectral signal and the observed signal sequence.

[0007] Optionally, the step of using the standardized sample sequence set and the constraints of the dynamic equations to invert the implicit inversion physical parameter vector constrained by the physical feasible region from the observed signal sequence using an inverse physical information neural network includes: By inputting the input feature set in parallel into a shared asymmetric convolutional residual block containing response head branches and physical head branches, a temporal feature representation that integrates short-term local impact details and long-term global periodic dynamics is generated. An adaptive normalization layer and activation function are applied to the temporal feature representation. Learnable scale and translation parameters are preserved through batch normalization, and the statistics within the time window are weighted to generate a normalized deep feature vector. The normalized deep feature vector is used to regress physical parameters through a fully connected layer to generate an inverted physical parameter vector constrained by the physical feasible region.

[0008] Optionally, the step of generating a multi-condition fault parameter set based on the inverted physical parameter vector and subject to the constraints of the dynamic equation includes: The inverted physical parameter vectors are encoded into the feature space. By constructing an inverse parameter encoding mechanism, the damping coefficient, linear stiffness, nonlinear stiffness parameters, impact amplitude, and fault characteristic frequency are used as trainable vectors to generate implicit physical quantity encodings capable of expressing complex fault behaviors. Based on the standardized sample sequence set and the input feature set, the data consistency loss is calculated. The difference between the predicted response and the actual measurement result is quantified by the data consistency loss quantification model, and preliminary optimized dynamic system parameters are generated. A regularization term is introduced to impose soft constraints on the parameters of the initially optimized dynamic system. By utilizing prior structural information to shrink the solution space, a set of inverse dynamic parameters with reasonable physical state is generated. The inverse dynamic parameter set is determined as the basic parameter set. Multiple sets of typical speed disturbance values, multiple load random disturbance coefficients, multiple fault level adjustment attenuation coefficients and impact period coefficients are determined. The multiple sets of typical speed disturbance values, multiple load random disturbance coefficients, multiple fault level adjustment attenuation coefficients and impact period coefficients are added to the basic parameter set to generate an extended multi-condition fault parameter set.

[0009] Optionally, the step of performing frequency domain gain weighting processing on the single-sided amplitude spectrum to generate an enhanced spectral signal for highlighting fault impulse characteristics includes: Identify the fundamental frequency position of the meshing frequency and the positions of each harmonic of the fundamental frequency position in the single-sided amplitude spectrum, and generate a set of key fault focus frequencies based on the fundamental frequency position and the positions of each harmonic. Based on the set of key fault focus frequencies, construct a Gaussian or rectangular window-shaped weighting factor function, and calculate a gain weighting vector with the same unilateral amplitude spectrum length as the set of key fault focus frequencies. The gain weighting vector is multiplied element-wise with the single-sided amplitude spectrum to generate a weighted spectrum with an amplitude distribution that highlights the characteristics of the fault impact. An inverse fast Fourier transform is performed on the weighted spectrum to generate an enhanced spectral signal that highlights the characteristics of the fault impact.

[0010] Optionally, the step of constructing a wind turbine bearing fault data set based on the extended multi-condition fault parameter set includes: Based on the extended multi-condition fault parameter set, the generation model is driven to perform forward time integration simulation to obtain a time-domain vibration response sequence that meets the constraints of the nonlinear dynamic equation. The time-domain vibration response sequence is standardized and sliced ​​using a sliding window to obtain a set of standardized fault sample sequences under multiple operating conditions. The standardized fault sample sequence set is subjected to frequency domain transformation and gain weighting to obtain an enhanced fault spectrum signal set that highlights the early damage impact and modulation characteristics. The enhanced fault spectrum signal set and the standardized fault sample sequence set are superimposed by channel dimension or feature splicing to obtain a multimodal fault data sample set that integrates time domain and frequency domain information; Controllable noise disturbances and operating condition labels are added to the multimodal fault data sample set to obtain the wind turbine bearing fault data set.

[0011] Optionally, the step of controlling the wind turbine bearing fault classification target model and generating detection results for the target wind turbine bearing based on target data for the target wind turbine bearing includes: The real-time vibration acceleration signal sequence of the target wind turbine bearing is collected, and the real-time vibration acceleration signal sequence is determined as the target observation signal sequence of the target wind turbine bearing; wherein the target observation signal sequence is the raw time-domain vibration acceleration data continuously collected at a preset location of the target wind power generation equipment according to a preset sampling frequency; The target observation signal sequence of the target wind turbine bearing is subjected to standardization processing and time-series slicing to generate a set of standardized sample sequences for the target model of fault classification of the wind turbine bearing; The standardized target sample sequence set is input into the wind turbine bearing fault classification target model to generate a predicted probability distribution for fault type and fault severity. Based on the predicted probability distribution, the detection results for the fault type and severity of the target wind turbine bearing are determined.

[0012] Secondly, embodiments of this application provide a wind turbine bearing testing device for wind power generation equipment, characterized in that it includes: The standardized sample sequence set generation module is used to collect the observation signal sequence of wind turbine bearings, and perform standardization processing and time-series slicing on the observation signal sequence to generate a standardized sample sequence set. The actual engineering operating condition parameter determination module is used to determine the actual engineering operating condition parameters of the wind turbine bearing; the actual engineering operating condition parameters include: operating condition information characterizing the operating conditions of the wind turbine bearing in the real operating environment, geometric parameters characterizing the geometric features of the wind turbine bearing, and the material mechanical properties of the wind turbine bearing. The dynamic equation constraint determination module is used to determine the dynamic equation constraint conditions based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties. The inversion physical parameter vector inversion module is used to invert the implicit inversion physical parameter vector constrained by the physical feasible region in the observation signal sequence from the observation signal sequence using the standardized sample sequence set and the constraints of the dynamic equation, and employing an inverse physical information neural network. A multi-condition fault parameter set generation module is used to generate a multi-condition fault parameter set constrained by the dynamic equation based on the inverted physical parameter vector. The wind turbine bearing fault classification target model generation module is used to construct a wind turbine bearing fault data set based on the extended multi-condition fault parameter set, and to generate a wind turbine bearing fault classification target model using the wind turbine bearing fault data set. The detection result generation module is used to control the target model for fault classification of the wind turbine bearing, and generate detection results for the target wind turbine bearing based on the target data for the target wind turbine bearing.

[0013] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0014] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0015] Fifthly, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the method as described in the first aspect.

[0016] The embodiments of the present invention have the following advantages: This invention addresses the problems of model overfitting and poor generalization caused by the scarcity of real fault data under small sample conditions by inversely retrieving physical parameters from a small number of real observation signals and generating a high-fidelity synthetic fault data set constrained by dynamic equations based on these real parameters. At the same time, it overcomes the physical distortion of pure data-driven generation methods and the excessive dependence of traditional simulations on prior parameters, achieving high-precision and robust wind turbine bearing fault classification detection. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of a wind turbine bearing testing method for wind power generation equipment provided in an embodiment of the present invention; Figure 2 This is a structural block diagram of a wind turbine bearing testing device for wind power generation equipment provided in an embodiment of the present invention; Figure 3 This is a hardware structure block diagram of an electronic device provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Reference Figure 1 The diagram illustrates a flowchart of a wind turbine bearing testing method for a wind power generation device provided in an embodiment of the present invention, which may specifically include the following steps: Step 101: Collect the observation signal sequence of the wind turbine bearing, and perform standardization processing and time-series slicing on the observation signal sequence to generate a standardized sample sequence set; Step 102: Determine the actual engineering operating condition parameters of the wind turbine bearing; the actual engineering operating condition parameters include: operating condition information characterizing the operating conditions of the wind turbine bearing in the real operating environment, geometric parameters characterizing the geometric features of the wind turbine bearing, and the material mechanical properties of the wind turbine bearing. Step 103: Based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties, determine the constraints of the dynamic equations; Step 104: Using the standardized sample sequence set and the constraints of the dynamic equation, an inverse physical information neural network is used to invert the inverse physical parameter vector implicit in the observed signal sequence, which is constrained by the physical feasible region. Step 105: Generate a multi-condition fault parameter set based on the inverted physical parameter vector, which is constrained by the dynamic equation constraints. Step 106: Construct a wind turbine bearing fault data set based on the extended multi-condition fault parameter set, and use the wind turbine bearing fault data set to generate a wind turbine bearing fault classification target model. Step 107: Control the target model for fault classification of the wind turbine bearing, and generate detection results for the target wind turbine bearing based on the target data for the target wind turbine bearing.

[0022] In this embodiment of the invention, observation signal sequences of wind turbine bearings can be collected, and the observation signal sequences can be standardized and time-series sliced ​​to generate a set of standardized sample sequences. This allows for the acquisition of the original characterization of the bearing's operating state from a small amount of real observation data. Preprocessing eliminates dimensional differences and unifies the scale, while slicing increases the number of usable samples, facilitating subsequent determination of physical constraints and inverse parameter inversion.

[0023] The observed signal sequence refers to the raw time-domain vibration acceleration data continuously collected by a high-sensitivity accelerometer at a high sampling frequency at locations such as bearing housings, high-speed shafts of gearboxes, or generator drive ends. It includes the dynamic response information of the bearing under real working conditions.

[0024] Standardization refers to Z-Score standardization, which subtracts the mean from the signal and divides it by the standard deviation, so that signals with different amplitudes and operating conditions have zero mean and unit variance.

[0025] Timing slicing refers to the process of sliding a fixed-length window across a timing signal to capture segments, forming multiple overlapping or non-overlapping short-time samples to characterize local dynamic responses.

[0026] The standardized sample sequence set refers to the training / test sample set obtained after standardization and slicing, which can be directly used for determining dynamic constraints and inverse PINN input.

[0027] The embodiments of the present invention can determine the actual engineering operating parameters of the wind turbine bearing, including operating condition information, geometric parameters and material mechanical properties, so as to provide the prior input required for physical modeling of the bearing system and ensure the accuracy and engineering applicability of the subsequent dynamic equation constraints.

[0028] Actual engineering operating condition parameters refer to a comprehensive set of parameters characterizing wind turbine bearings in a real operating environment, including macroscopic operating conditions, microscopic geometric features, and material properties, used to drive high-fidelity dynamic simulations.

[0029] Operating condition information refers to macroscopic operating condition data that characterizes the wind turbine bearing in a real operating environment, such as high-speed shaft speed, load torque, simulation duration, temperature fluctuations, and changes in lubrication status.

[0030] Geometric parameters refer to data that characterize the microscopic geometric features of wind turbine bearings, such as the number of gear teeth, module, pressure angle, shaft length, inner / outer ring dimensional deviation, and radial clearance.

[0031] Material mechanical properties refer to the mechanical performance data that characterize the materials of wind turbine bearings, such as mass, moment of inertia, shaft torsional stiffness, elastic modulus, and damping coefficient.

[0032] In this embodiment of the invention, the constraints of the dynamic equations can be determined based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties, so as to establish the physical and mathematical framework of the bearing / gearbox system, provide reliable constraint anchors for subsequent inverse PINN inversion, and ensure the physical rationality of parameter solution.

[0033] The constraints of the dynamic equations refer to the constraints of the nonlinear dynamic model constructed based on the operating parameters and observation data. These constraints include coupled motion equations, time-varying stiffness matrices, nonlinear error terms, etc., and are used to quantify the deviation between the system response and the physical laws.

[0034] In this embodiment of the invention, the standardized sample sequence set and the constraints of the dynamic equations can be used to employ an inverse physical information neural network to invert the inverse physical parameter vector implicitly constrained by the physical feasible region in the observed signal sequence. By utilizing the end-to-end learning and automatic differentiation mechanism of the network, the implicit parameters of the system can be inferred from a small amount of real data, thereby achieving inverse solution under physical constraints.

[0035] Inverse physical information neural networks are a framework for solving inverse problems by embedding physical equations into a loss function. They are used to invert system parameters from observed responses while ensuring that the predictions conform to the laws of dynamics.

[0036] The inverted physical parameter vector constrained by the physical feasible region refers to the set of parameters obtained through network regression, which is restricted by physical boundaries (such as parameter positivity and energy conservation), including damping coefficient, stiffness parameter, impact amplitude, etc.

[0037] For example, based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties, the constraints of the dynamic equation can be quickly approximated by using the finite element analysis method combined with empirical formulas. This reduces computational complexity and improves adaptability to complex working conditions, while ensuring the physical accuracy of the constraints, which facilitates subsequent inverse parameter inversion.

[0038] The finite element method refers to the method of discretizing the bearing / gearbox system into a finite element mesh, solving the local stiffness matrix and assembling the global matrix to simulate the dynamic response of the system.

[0039] Empirical formulas are simplified formulas based on engineering test data and statistical models, used to quickly estimate time-varying meshing stiffness, bearing support stiffness, and nonlinear error terms, rather than relying entirely on potential energy integrals.

[0040] The constraints of the dynamic equations refer to the constraints of the simplified nonlinear dynamic model obtained through finite element analysis and empirical formulas. These constraints include the global stiffness matrix and error adjustment terms, and are used to quantify the deviation of the system response from the physical laws.

[0041] By combining finite element analysis with empirical formulas, the constraints of the dynamic equations can be quickly determined from a standardized sample sequence set and operating parameters. This solves the problem of the long processing time of traditional numerical integration methods in practical engineering scenarios with limited computing resources, ensuring the physical rationality and efficiency of the constraints, providing a reliable anchor point for inverse PINN inversion, and improving the overall data generation efficiency.

[0042] Furthermore, by using the standardized sample sequence set and the constraints of the dynamic equation, an inverse physical information neural network based on the Transformer architecture can be employed to invert the inverse physical parameter vector implicitly constrained by the physical feasible region in the observed signal sequence. This allows for the use of a self-attention mechanism to capture long-distance temporal dependencies, thereby improving the inversion accuracy for non-stationary signals while maintaining the strictness of physical constraints.

[0043] The Transformer architecture refers to a neural network structure based on a self-attention mechanism, used to process temporal input feature sets. It captures global correlations through multi-head attention layers, rather than relying on the local receptive fields of convolutions.

[0044] Self-attention mechanism refers to calculating the association weights between each position in the input sequence and other positions, which is used to fuse the global contextual information of the observed signal.

[0045] The inverted physical parameter vector refers to the set of parameters obtained by the final regression through the network. It is constrained by the physical feasible domain (such as parameter boundaries and equation residuals) and includes implicit contents such as damping and stiffness.

[0046] By using a Transformer-based inverse physical information neural network, physical parameter vectors are inverted from a standardized set of sample sequences and dynamic constraints. This addresses the problem of traditional convolutional networks' inadequate capture of long-term dependencies when the observed signals are non-stationary or noisy, ensuring the physical reliability and accuracy of the inverted parameters and providing a high-quality foundation for subsequent multi-condition synthetic data generation.

[0047] Of course, the above examples are merely illustrative. Those skilled in the art can use other methods to determine the constraints of the dynamic equations and to inversely derive the physical parameter vectors. In this regard, the embodiments of the present invention do not impose any limitations.

[0048] In this embodiment of the invention, a multi-condition fault parameter set constrained by the dynamic equation can be generated based on the inverted physical parameter vector. The real parameters obtained by inversion can be used as seeds to generate diverse parameters through controllable expansion, ensuring high physical consistency of subsequent synthetic data.

[0049] The multi-condition fault parameter set refers to the parameter set obtained by expanding the inversion parameters. It is constrained by the dynamic equation and includes multiple speed, load and fault level variants to simulate real and variable environments.

[0050] In this embodiment of the invention, a wind turbine bearing fault data set can be constructed based on the extended multi-condition fault parameter set, and a wind turbine bearing fault classification target model can be generated using the wind turbine bearing fault data set to achieve the generation of massive high-fidelity synthetic data, which can then be used as an enhancement source to train downstream models and improve diagnostic performance.

[0051] The wind turbine bearing fault dataset refers to a diverse and highly physically consistent synthetic vibration response dataset generated through forward simulation, including time-domain sequences, spectral characteristics, and labels.

[0052] The wind turbine bearing fault classification target model refers to a deep learning model trained using synthetic data, used for the classification and prediction of fault types and severity.

[0053] In this embodiment of the invention, the target model for fault classification of wind turbine bearings can be controlled to generate detection results for the target wind turbine bearings based on target data for the target wind turbine bearings. In actual deployment scenarios, end-to-end reasoning can be performed on newly collected real-time data to directly output diagnostic conclusions.

[0054] Target data refers to the real-time observation signal sequence of the target wind turbine bearing, which is usually vibration acceleration data collected by sensors.

[0055] The test results refer to the diagnostic conclusions of the fault type (such as inner ring, outer ring, rolling element fault) and severity (such as minor, moderate, severe) output by the model.

[0056] This invention addresses the problems of model overfitting and poor generalization caused by the scarcity of real fault data under small sample conditions by inversely retrieving physical parameters from a small number of real observation signals and generating a high-fidelity synthetic fault data set constrained by dynamic equations based on these real parameters. At the same time, it overcomes the physical distortion of pure data-driven generation methods and the excessive dependence of traditional simulations on prior parameters, achieving high-precision and robust wind turbine bearing fault classification detection.

[0057] Based on the above embodiments, modified embodiments of the above embodiments are proposed. It should be noted that, in order to keep the description brief, only the differences from the above embodiments are described in the modified embodiments.

[0058] In an optional embodiment of the present invention, observation signal sequences of wind turbine bearings can be collected, and the observation signal sequences can be standardized and time-series sliced ​​to generate a set of standardized sample sequences for unifying dimensions and characterizing local dynamic responses. The actual engineering operating condition parameters of the wind turbine bearing are specified; the actual engineering operating condition parameters include: operating condition information characterizing the operating conditions of the wind turbine bearing in the real operating environment, geometric parameters characterizing the geometric characteristics of the wind turbine bearing, and the material mechanical properties of the wind turbine bearing; Based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties, a degree-of-freedom coupled nonlinear dynamic equation is constructed. The time-varying meshing stiffness and bearing support stiffness of the wind turbine bearing are calculated using the aforementioned degree-of-freedom coupled nonlinear dynamic equations. The bending, shearing and axial compression deformation energies are calculated by integrating using the potential energy method to generate a real nonlinear time-varying stiffness matrix. The gear eccentricity error, multi-order machining error harmonic components, and bearing geometric dimension deviation are determined, and the nonlinear time-varying stiffness matrix is ​​adjusted using the gear eccentricity error, the multi-order machining error harmonic components, and the bearing geometric dimension deviation to generate motion equations for a dynamic system that enhance realism. The motion equations of the enhanced dynamic system are solved numerically by time integration using the backward difference formula method. The convergence and accuracy of the rigid system are ensured by implicit integration, and the time-domain acceleration response curves of the input axis, intermediate axis and output axis in multiple directions are generated. Perform a fast Fourier transform on the time-domain acceleration response curve to generate a one-sided amplitude spectrum; Frequency domain gain weighting is performed on the single-sided amplitude spectrum to generate an enhanced spectrum signal that highlights the characteristics of the fault impact. An input feature set is constructed based on the enhanced spectral signal and the observed signal sequence; By inputting the input feature set in parallel into a shared asymmetric convolutional residual block containing response head branches and physical head branches, a temporal feature representation that integrates short-term local impact details and long-term global periodic dynamics is generated. An adaptive normalization layer and activation function are applied to the temporal feature representation. Learnable scale and translation parameters are preserved through batch normalization, and the statistics within the time window are weighted to generate a normalized deep feature vector. The normalized deep feature vector is used to regress physical parameters through a fully connected layer to generate an inverted physical parameter vector constrained by the physical feasible region. The inverted physical parameter vectors are encoded into the feature space. By constructing an inverse parameter encoding mechanism, the damping coefficient, linear stiffness, nonlinear stiffness parameters, impact amplitude, and fault characteristic frequency are used as trainable vectors to generate implicit physical quantity encodings capable of expressing complex fault behaviors. Based on the standardized sample sequence set and the input feature set, the data consistency loss is calculated. The difference between the predicted response and the actual measurement result is quantified by the data consistency loss quantification model, and preliminary optimized dynamic system parameters are generated. A regularization term is introduced to impose soft constraints on the parameters of the initially optimized dynamic system. By utilizing prior structural information to shrink the solution space, a set of inverse dynamic parameters with reasonable physical state is generated. The inverse dynamic parameter set is determined as the basic parameters. Determine multiple sets of typical speed disturbance values, multiple load random disturbance coefficients, multiple fault level adjustment attenuation coefficients, and impact period coefficients, and add the multiple sets of typical speed disturbance values, the multiple load random disturbance coefficients, the multiple fault level adjustment attenuation coefficients, and the impact period coefficients to the basic parameters to generate an extended set of multi-condition fault parameters; A set of wind turbine bearing fault data is constructed based on the extended multi-condition fault parameter set. A target model for classifying wind turbine bearing faults is generated using the aforementioned wind turbine bearing fault data set; The target model for classifying the faults of the wind turbine bearing is controlled, and based on the target data for the target wind turbine bearing, the detection results for the target wind turbine bearing are generated.

[0059] In an optional embodiment of the present invention, the step of determining the constraints of the dynamic equations based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties includes: Based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties, a degree-of-freedom coupled nonlinear dynamic equation is constructed. The time-varying meshing stiffness and bearing support stiffness of the wind turbine bearing are calculated using the aforementioned degree-of-freedom coupled nonlinear dynamic equations. The bending, shearing and axial compression deformation energies are calculated by integrating using the potential energy method, thereby generating a nonlinear time-varying stiffness matrix. The gear eccentricity error, multi-order machining error harmonic components, and bearing geometric dimension deviation are determined, and the nonlinear time-varying stiffness matrix is ​​adjusted using the gear eccentricity error, the multi-order machining error harmonic components, and the bearing geometric dimension deviation to generate the motion equation of the enhanced dynamic system. The motion equations of the enhanced dynamic system are solved numerically by time integration using the backward difference formula method, generating time-domain acceleration response curves of the input axis, intermediate axis and output axis in multiple directions; Perform a fast Fourier transform on the time-domain acceleration response curve to generate a one-sided amplitude spectrum; Frequency domain gain weighting is performed on the single-sided amplitude spectrum to generate an enhanced spectrum signal that highlights the characteristics of the fault impact. An input feature set is constructed based on the enhanced spectral signal and the observed signal sequence.

[0060] In this embodiment of the invention, a degree-of-freedom coupled nonlinear dynamic equation can be constructed based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties. This transforms the macroscopic operating conditions, microscopic geometric features, and material properties of the bearing / gearbox system into a mathematically solvable dynamic model, providing a unified physical basis for subsequent stiffness calculations and constraint conditions, and ensuring that the model can fully reflect the coupled vibration characteristics of the real system.

[0061] The coupled nonlinear dynamic equations of freedom refer to simplifying the wind turbine bearing system into a multi-degree-of-freedom model that includes the interaction of torsional vibration and lateral vibration. Its equations of motion are a combination of mass matrix, damping matrix, time-varying stiffness matrix and nonlinear excitation terms, which are used to describe the dynamic response of the system under complex working conditions.

[0062] In this embodiment of the invention, the time-varying meshing stiffness and bearing support stiffness of the wind turbine bearing can be calculated using the coupled nonlinear dynamic equations of the degrees of freedom. The bending, shearing and axial compression deformation energies are calculated by integrating the potential energy method to generate a nonlinear time-varying stiffness matrix. This accurately simulates the stiffness characteristics that change periodically with time during gear meshing and bearing raceway contact, providing a real and nonlinear stiffness input for the dynamic equations and avoiding the accuracy loss of traditional linear or approximate stiffness models.

[0063] Time-varying meshing stiffness refers to the stiffness value of gears that fluctuates periodically with the meshing angle due to tooth profile deformation and changes in contact position during gear meshing.

[0064] Bearing support stiffness refers to the fluctuation of support stiffness caused by the periodic rolling of rolling elements on the raceway, which is usually related to the number of rolling elements, the cage speed and the journal position.

[0065] Potential energy integral calculation refers to the method of calculating bending, shear and axial compression deformation energy by integrating based on tooth profile geometric parameters, which is used to obtain the true nonlinear time-varying meshing stiffness.

[0066] The nonlinear time-varying stiffness matrix refers to the global matrix assembled from the calculated time-varying meshing stiffness and bearing support stiffness. Its elements change significantly with time and are the core time-varying coefficients in the dynamic equations.

[0067] In this embodiment of the invention, gear eccentricity error, multi-order machining error harmonic components, and bearing geometric dimension deviation can be determined. The nonlinear time-varying stiffness matrix can be adjusted using the gear eccentricity error, the multi-order machining error harmonic components, and the bearing geometric dimension deviation to generate enhanced dynamic system motion equations. This introduces unavoidable non-ideal factors in actual manufacturing and assembly processes, making the dynamic model closer to the vibration behavior of real engineering systems and improving the simulation's engineering applicability and fault feature reproduction capability.

[0068] Gear eccentricity error refers to the offset between the gear mounting center and the geometric center, which causes low-frequency modulation of the meshing stiffness.

[0069] Multi-order machining error harmonic components refer to the periodic errors (such as cumulative pitch error and tooth profile error) generated during gear machining, which are superimposed on the meshing stiffness in the form of multi-order harmonics.

[0070] Bearing geometric deviations refer to microscopic deviations such as the roundness error of the inner / outer ring raceway and the change in radial clearance, which have a periodic impact on the support stiffness.

[0071] The enhanced dynamic system equations of motion refer to the final equations of motion obtained by superimposing error terms on the basic nonlinear equations, which are used to simulate the system response more realistically.

[0072] In this embodiment of the invention, the backward difference formula method can be used to solve the motion equations of the enhanced dynamic system by time integration, generating time-domain acceleration response curves of the input axis, intermediate axis and output axis in multiple directions. This is to deal with the rigid system problems caused by drastic stiffness changes and strong nonlinearity, ensure the stability and high accuracy of long-term numerical integration, and obtain real vibration response data in multiple axes and directions.

[0073] The backward difference formula (BDF) method is an implicit multi-step integration method that solves differential equations through backward difference. It is suitable for rigid systems and has good numerical stability.

[0074] The time-domain acceleration response curve refers to the curve of acceleration in the x, y, and z directions of each axis as a function of time, obtained through numerical solution, which intuitively reflects the vibration evolution process of the system.

[0075] In this embodiment of the invention, a fast Fourier transform can be performed on the time-domain acceleration response curve to generate a single-sided amplitude spectrum, thereby converting the time-domain vibration signal to the frequency domain and extracting the periodic fault characteristic frequency and its sideband information, which facilitates subsequent spectrum enhancement and feature construction.

[0076] The Fast Fourier Transform (FFT) is a discrete Fourier transform algorithm that transforms a time-domain signal into a frequency-domain signal to obtain the single-sided amplitude spectrum.

[0077] The single-sided amplitude spectrum refers to the amplitude spectrum of the positive frequency part in the FFT result, highlighting the meshing frequency, fault characteristic frequency and its harmonics.

[0078] In this embodiment of the invention, frequency domain gain weighting processing can be performed on the single-sided amplitude spectrum to generate an enhanced spectrum signal for highlighting fault impact characteristics. This selectively amplifies weak modulation components and early fault impact signals in key frequency bands (such as the meshing frequency and its harmonics), improves the identifiability of fault characteristics under low signal-to-noise ratio conditions, and provides more significant fault information for subsequent input feature sets.

[0079] Frequency domain gain weighting refers to the process of amplifying the amplitude by applying a weighting factor function within a specific frequency band of the spectrum, while keeping other frequency bands basically unchanged.

[0080] Enhanced spectrum signals refer to the spectrum results after gain weighting, in which fault impulse characteristics (such as modulation sidebands and impulse harmonics) are significantly highlighted.

[0081] In this embodiment of the invention, an input feature set can be constructed based on the enhanced spectral signal and the observed signal sequence to fuse the original time-domain information and frequency-domain enhanced features, forming a multimodal, high-information-content network input. This provides comprehensive observational evidence for the inversion process of the reverse physical information neural network, ensuring the accuracy and robustness of parameter inversion.

[0082] The input feature set refers to the set of features formed by combining the enhanced spectral signal with the observed signal sequence (time domain data). It is usually constructed by channel superposition or feature splicing and is used for parallel input of the network.

[0083] In an optional embodiment of the present invention, the step of using the standardized sample sequence set and the constraints of the dynamic equations to invert the implicit inversion physical parameter vector constrained by the physical feasible region in the observed signal sequence using an inverse physical information neural network includes: By inputting the input feature set in parallel into a shared asymmetric convolutional residual block containing response head branches and physical head branches, a temporal feature representation that integrates short-term local impact details and long-term global periodic dynamics is generated. An adaptive normalization layer and activation function are applied to the temporal feature representation. Learnable scale and translation parameters are preserved through batch normalization, and the statistics within the time window are weighted to generate a normalized deep feature vector. The normalized deep feature vector is used to regress physical parameters through a fully connected layer to generate an inverted physical parameter vector constrained by the physical feasible region.

[0084] In this embodiment of the invention, the input feature set can be fed in parallel into a shared asymmetric convolutional residual block containing a response head branch and a physical head branch to generate a temporal feature representation that integrates short-term local impact details and long-term global periodic dynamics. This allows the use of convolutional branches with different receptive fields to capture multi-scale temporal correlations, ensuring that the feature representation simultaneously includes local fault impacts and global dynamic evolution, which is convenient for subsequent physical parameter regression.

[0085] The input feature set refers to the set of features that are fused from the enhanced spectral signal and the observed signal sequence, and is used as the initial input to the network.

[0086] The response head branch refers to the part of the network used for time-series response reconstruction, which assists physical constraints by predicting vibration responses.

[0087] The physical head branch refers to the part of the network used for parameter regression, which extracts hidden physical quantities from features through fully connected layers.

[0088] Shared asymmetric convolutional residual blocks refer to residual structures containing parallel narrow and wide kernel branches, used to extract multi-scale temporal features.

[0089] The temporal feature representation that integrates short-term local impact details with long-term global periodic dynamics refers to the deep vector obtained by branching and fusion, which integrates local high-frequency impact and global low-frequency periodic information.

[0090] In this embodiment of the invention, an adaptive normalization layer and activation function can be applied to the temporal feature representation. By batch normalization, learnable scale and translation parameters are preserved, and the statistics within the time window are weighted to generate a normalized deep feature vector. This stabilizes the network training process, enhances the nonlinear expressive ability, adapts to the statistical variability of the observed signal, and ensures that the feature vector is suitable for subsequent parameter regression.

[0091] Temporal feature representation refers to the intermediate multi-scale vector obtained from asymmetric convolutional blocks, which characterizes the dynamic pattern of the observed signal.

[0092] An adaptive normalization layer is a batch normalization (BN) layer that preserves learnable parameters to adapt to the data distribution.

[0093] Activation functions are nonlinear functions such as GELU or SiLU, used to introduce nonlinear mappings.

[0094] The normalized deep feature vector refers to the final vector after normalization and activation processing, which is used as input to the fully connected layer.

[0095] In this embodiment of the invention, the normalized deep feature vector can be used to regress physical parameters through a fully connected layer to generate an inverted physical parameter vector constrained by the physical feasible region, so as to realize the mapping from features to parameters, and ensure the rationality of the inversion result through physical boundary constraints, providing the final output for inverse solution.

[0096] A fully connected layer is a dense layer in a network used for parameter regression, which maps features to the parameter space through a weight matrix.

[0097] The inverted physical parameter vector constrained by the physical feasible region refers to the set of parameters with boundary constraints applied after regression, such as damping coefficients and stiffness parameters.

[0098] This invention addresses the shortcomings of traditional methods in physical parameter inversion, such as insufficient accuracy and poor generalization, by extracting multi-scale features through shared asymmetric convolutional residual blocks, adaptive normalization for stable training, and regression of parameter vectors constrained by the physical feasible region using fully connected layers, under conditions of limited real observation data. This ensures that the generated physical parameters conform to the laws of dynamics, providing a reliable foundation for generating high-fidelity synthetic fault data and improving the accuracy and robustness of wind turbine bearing fault diagnosis.

[0099] In an optional embodiment of the present invention, the step of generating a multi-condition fault parameter set based on the inverted physical parameter vector and subject to the constraints of the dynamic equation includes: The inverted physical parameter vectors are encoded into the feature space. By constructing an inverse parameter encoding mechanism, the damping coefficient, linear stiffness, nonlinear stiffness parameters, impact amplitude, and fault characteristic frequency are used as trainable vectors to generate implicit physical quantity encodings capable of expressing complex fault behaviors. Based on the standardized sample sequence set and the input feature set, the data consistency loss is calculated. The difference between the predicted response and the actual measurement result is quantified by the data consistency loss quantification model, and preliminary optimized dynamic system parameters are generated. A regularization term is introduced to impose soft constraints on the parameters of the initially optimized dynamic system. By utilizing prior structural information to shrink the solution space, a set of inverse dynamic parameters with reasonable physical state is generated. The inverse dynamic parameter set is determined as the basic parameter set. Multiple sets of typical speed disturbance values, multiple load random disturbance coefficients, multiple fault level adjustment attenuation coefficients and impact period coefficients are determined. The multiple sets of typical speed disturbance values, multiple load random disturbance coefficients, multiple fault level adjustment attenuation coefficients and impact period coefficients are added to the basic parameter set to generate an extended multi-condition fault parameter set.

[0100] In this embodiment of the invention, the inverted physical parameter vector can be encoded into the feature space. By constructing an inverse parameter encoding mechanism, the damping coefficient, linear stiffness, nonlinear stiffness parameter, impact amplitude, and fault characteristic frequency are used as trainable vectors to generate implicit physical quantity encodings capable of expressing complex fault behaviors. This transforms the inverted parameters into a network-optimizable form, enabling them to capture the nonlinear dynamic characteristics of bearing faults and providing a flexible implicit representation basis for subsequent loss-driven optimization.

[0101] The inverted physical parameter vector refers to the set of parameters obtained by regressing from the observed signal through an inverse physical information neural network, which is constrained by the physical feasible region.

[0102] Inverse parameter encoding refers to a network structure that treats physical parameters as trainable vectors and embeds them into the feature space to express the implicit dynamic characteristics of faults.

[0103] The damping coefficient is a parameter that refers to the energy dissipation of a system and is used to simulate the damping effect in bearing vibration.

[0104] The linear stiffness parameter refers to the elastic stiffness value of the system within a small deformation range.

[0105] Nonlinear stiffness parameters refer to stiffness values ​​that take into account nonlinear factors such as clearance and error, and vary with the amount of deformation.

[0106] Impact amplitude refers to the amplitude of the periodic excitation force caused by the fault, reflecting the severity of the damage.

[0107] Fault characteristic frequencies refer to specific frequencies associated with the type of fault, such as the passing frequency of the inner and outer rings of a bearing.

[0108] Implicit physical quantity encoding refers to encoding the above parameters into high-dimensional feature vectors, which are implicit representations capable of expressing complex fault evolution.

[0109] In this embodiment of the invention, data consistency loss can be calculated based on the standardized sample sequence set and the input feature set. The difference between the predicted response and the actual measurement result is quantified by the data consistency loss, and preliminary optimized dynamic system parameters are generated. Data-driven constraints are used to guide the iteration of network parameters, ensuring that the predicted response is consistent with the observation, thereby initially converging the dynamic system parameters and realizing the preliminary optimization of the inverse solution.

[0110] A standardized sample sequence set refers to a set of observation data that has been standardized and sliced, serving as a true measurement benchmark for loss calculation.

[0111] The input feature set refers to the set of features that fuse the enhanced spectrum and the observed signal, and is used as the predictive input for the network.

[0112] Data consistency loss refers to the loss function that quantifies the difference between the vibration response predicted by the model and the actual observed data, and is usually in the form of mean square error.

[0113] Model-predicted response refers to the simulated vibration signal generated by the network based on the current parameters.

[0114] The preliminary optimized dynamic system parameters refer to the initial convergence parameters obtained after driving the data consistency loss, which are used for subsequent regularization refinement.

[0115] In this embodiment of the invention, a regularization term can be introduced to impose soft constraints on the initially optimized dynamic system parameters. By utilizing prior structural information to shrink the solution space, a set of physically reasonable inverse dynamic parameters can be generated to prevent physically unreasonable drift of parameters when data is scarce. This ensures that the optimization results conform to engineering prior knowledge and improves the reliability and interpretability of the inverse parameters.

[0116] The regularization term refers to the loss regularization function based on prior structural information, which is used to shrink the parameter solution space.

[0117] Prior structural information refers to the parameter boundaries and relational constraints based on physical knowledge or empirical data, such as positive stiffness and non-negative damping.

[0118] Soft constraints refer to parameter restrictions implemented through loss weighting rather than hard boundaries.

[0119] A physically reasonable set of inverse dynamic parameters refers to the final set of parameters generated after regularization, which conforms to physical laws and engineering realities.

[0120] In this embodiment of the invention, the inverse dynamic parameter set can be determined as the basic parameters. Multiple sets of typical speed disturbance values, multiple load random disturbance coefficients, multiple fault level adjustment attenuation coefficients, and impact period coefficients are determined. The multiple sets of typical speed disturbance values, multiple load random disturbance coefficients, multiple fault level adjustment attenuation coefficients, and impact period coefficients are added to the basic parameters to generate an extended multi-condition fault parameter set. This allows for the expansion of diverse disturbances based on the inverse parameters, achieving coverage of multiple operating conditions and multiple fault levels, and ensuring that the generated data set has the diversity of actual engineering applications.

[0121] The inverse dynamic parameter set refers to the set of physically reasonable parameters obtained through inverse optimization, which serves as an extended basis.

[0122] Multiple sets of typical speed disturbance values ​​refer to multiple speed variations that are preset or randomly generated, used to simulate variable speed conditions.

[0123] The multi-load random disturbance coefficient refers to a random coefficient added to the base load to simulate load fluctuations.

[0124] The multi-fault level adjustment attenuation coefficient and impact period coefficient refer to the attenuation and period parameters adjusted according to the severity of the fault, which are used to simulate the fault evolution.

[0125] The extended multi-condition fault parameter set refers to the parameter set generated after adding disturbances, which is used to drive the generation of synthetic data.

[0126] This invention generates a multi-condition fault parameter set by using parameter encoding, data consistency loss-driven optimization, regularized soft constraints, and perturbation extension based on inverse parameter calculation. This solves the problems of unstable physical parameter inversion and insufficient diversity of generated data under the condition of scarce real fault samples, ensuring that the parameter set conforms to dynamic constraints, providing high-fidelity and diversified basic parameters, and improving the physical rationality and engineering applicability of wind turbine bearing fault data generation.

[0127] In an optional embodiment of the present invention, the step of performing frequency domain gain weighting processing on the single-sided amplitude spectrum to generate an enhanced spectrum signal for highlighting fault impulse characteristics includes: Identify the fundamental frequency position of the meshing frequency and the positions of each harmonic of the fundamental frequency position in the single-sided amplitude spectrum, and generate a set of key fault focus frequencies based on the fundamental frequency position and the positions of each harmonic. Based on the set of key fault focus frequencies, construct a Gaussian or rectangular window-shaped weighting factor function, and calculate a gain weighting vector with the same unilateral amplitude spectrum length as the set of key fault focus frequencies. The gain weighting vector is multiplied element-wise with the single-sided amplitude spectrum to generate a weighted spectrum with an amplitude distribution that highlights the characteristics of the fault impact. An inverse fast Fourier transform is performed on the weighted spectrum to generate an enhanced spectral signal that highlights the characteristics of the fault impact.

[0128] In this embodiment of the invention, the fundamental frequency position of the meshing frequency and the positions of each harmonic of the fundamental frequency position in the single-sided amplitude spectrum can be identified. Based on the fundamental frequency position and the positions of each harmonic, a set of key fault-related frequency points can be generated to automatically locate the core frequency points and their harmonics related to gear meshing and bearing faults, providing accurate target frequency bands for subsequent weighted processing and ensuring that gain operation is focused on the fault-sensitive area.

[0129] The fundamental frequency position of the meshing frequency refers to the center frequency point of the gear meshing frequency (number of teeth × rotational speed), which is represented by a significant peak in the frequency spectrum.

[0130] The harmonic positions of each order refer to the integer multiples of the meshing frequency, such as the second harmonic, third harmonic, etc., which are often accompanied by fault modulation to generate sidebands.

[0131] The set of critical fault focus frequencies refers to a list of frequencies automatically extracted based on the fundamental frequency and harmonics, used to guide the center positioning of the weighting function.

[0132] In this embodiment of the invention, a Gaussian or rectangular window-shaped weighting factor function can be constructed based on the set of key fault focus frequencies. The gain weighting vector with the same unilateral amplitude spectrum length can be calculated to apply a smooth or sharp amplitude boost near the fault focus frequencies, while keeping the amplitude unchanged or slightly attenuated in non-focus regions, thereby achieving selective enhancement of fault characteristics.

[0133] A Gaussian weighted factor function is a window function constructed in the form of a Gaussian distribution. The central peak corresponds to the frequency of interest, and the width is controlled by the standard deviation σ. It is used to enhance the amplitude of smooth transitions.

[0134] The rectangular window weighting factor function is a function that forms a rectangular window of fixed width near the frequency of interest, used for uniform enhancement gain processing.

[0135] A gain weighting vector is a vector with the same length as the spectrum, where each element corresponds to a gain coefficient at a frequency point (>1 for amplification, =1 for keeping it unchanged).

[0136] In this embodiment of the invention, the gain weighting vector can be multiplied element by element with the single-sided amplitude spectrum to generate a weighted spectrum with amplitude distribution to highlight the characteristics of fault impact, so as to achieve targeted amplitude amplification of key frequency bands, making the harmonics, modulation sidebands and early weak impacts of the fault impact pulse more prominent in the spectrum, which is convenient for subsequent inverse transformation and feature extraction.

[0137] Weighted spectrum refers to the spectrum result after element-wise multiplication, in which the amplitude of the fault-related frequency components is significantly increased.

[0138] In this embodiment of the invention, an inverse fast Fourier transform can be performed on the weighted spectrum to generate an enhanced spectral signal that highlights the characteristics of the fault impact. This transforms the frequency-domain enhanced amplitude spectrum back to the time domain or retains it as an enhanced spectral form, providing a signal containing more obvious fault impact characteristics for constructing a multimodal input feature set.

[0139] The Inverse Fast Fourier Transform (IFFT) is an algorithm that inversely transforms the frequency-domain weighted spectrum back to the time domain, and is used to generate time-domain enhanced vibration signals.

[0140] Enhanced spectrum signals refer to the final frequency or time domain enhanced results, in which fault impulse characteristics (such as pulse impulses and modulation stripes) are highlighted, making them easier to input into subsequent networks.

[0141] This invention addresses the difficulty in identifying fault impacts in real wind turbine bearing signals with strong noise interference and weak early fault characteristics by identifying key frequency points, constructing weighting functions, performing element-wise multiplication, and inverse transformation in a complete frequency domain gain processing chain. This ensures that the generated enhanced spectrum signal has clear physical meaning and significant features, providing high-quality input for the inversion accuracy and synthetic data quality of the inverse physical information neural network, and improving the robustness and accuracy of overall fault diagnosis.

[0142] In an optional embodiment of the present invention, the step of constructing a wind turbine bearing fault data set based on the extended multi-condition fault parameter set includes: Based on the extended multi-condition fault parameter set, the generation model is driven to perform forward time integration simulation to obtain a time-domain vibration response sequence that meets the constraints of the nonlinear dynamic equation. The time-domain vibration response sequence is standardized and sliced ​​using a sliding window to obtain a set of standardized fault sample sequences under multiple operating conditions. The standardized fault sample sequence set is subjected to frequency domain transformation and gain weighting to obtain an enhanced fault spectrum signal set that highlights the early damage impact and modulation characteristics. The enhanced fault spectrum signal set and the standardized fault sample sequence set are superimposed by channel dimension or feature splicing to obtain a multimodal fault data sample set that integrates time domain and frequency domain information; Controllable noise disturbances and operating condition labels are added to the multimodal fault data sample set to obtain the wind turbine bearing fault data set.

[0143] In this embodiment of the invention, a forward time integral simulation can be performed based on the extended multi-condition fault parameter set to drive the generation model, thereby obtaining a time-domain vibration response sequence that conforms to the constraints of the nonlinear dynamic equation. This strictly follows the nonlinear dynamic equation of the bearing / gearbox system. By simulating the real vibration response process through numerical integration, the generated time-domain signal is highly consistent with physical laws in terms of energy conservation, impact structure, and spectral characteristics, providing physically faithful original synthetic data.

[0144] Forward time integration simulation refers to the process of using numerical methods (such as backward difference formulas) to drive the dynamic equations forward from the initial state and parameters to generate time-domain vibration signals.

[0145] The time-domain vibration response sequence refers to the sequence of accelerations in multiple directions along each axis that changes over time, obtained through simulation. It contains real characteristics such as fault impact, modulation, and periodic response.

[0146] In this embodiment of the invention, the time-domain vibration response sequence can be standardized and sliced ​​using a sliding window to obtain a set of standardized fault sample sequences under multiple operating conditions. This eliminates differences in amplitude dimensions under different operating conditions, unifies the scale, and increases the number of samples through slicing, which facilitates subsequent multimodal feature construction and model training.

[0147] Standardization refers to Z-Score standardization, which subtracts the mean from the sequence and divides by the standard deviation to give the signal zero mean and unit variance.

[0148] Sliding window slicing refers to using a fixed-length window to slide and extract segments on the time-series response, forming multiple short-time samples to characterize local fault dynamics.

[0149] The standardized fault sample sequence set under multiple operating conditions refers to the synthetic sample set after standardization and slicing, covering multiple speeds, loads and fault levels.

[0150] In this embodiment of the invention, the standardized fault sample sequence set can be subjected to frequency domain transformation and gain weighting to obtain an enhanced fault spectrum signal set that highlights the early damage impact and modulation characteristics. This converts the time domain signal to the frequency domain and applies selective gain to key fault frequency bands (such as characteristic frequencies and their harmonics) to significantly amplify early weak impact pulses and modulation sidebands, improve the identifiability of fault characteristics, and facilitate multimodal fusion.

[0151] Frequency domain transformation refers to performing a Fast Fourier Transform (FFT) on a time-domain sequence to obtain a single-sided amplitude spectrum.

[0152] Gain-weighted processing refers to applying Gaussian or rectangular window weighting factors near the frequency of concern in the fault area to amplify the amplitude in a targeted manner.

[0153] The enhanced fault spectrum signal set refers to the spectrum result after frequency domain gain weighting, which highlights the early damage impact and modulation characteristics.

[0154] In this embodiment of the invention, the enhanced fault spectrum signal set and the standardized fault sample sequence set can be superimposed by channel dimension or feature splicing to obtain a multimodal fault data sample set that integrates time-domain and frequency-domain information. This combines the original dynamic details in the time domain and the prominent fault features in the frequency domain to form complementary multimodal inputs, thereby improving the model's ability to represent complex fault modes.

[0155] Channel dimension overlay refers to stacking time-domain sequences and spectral signals as different channels to form a multi-channel image or tensor.

[0156] Feature concatenation refers to connecting time-domain and frequency-domain vectors along the feature dimension to form a high-dimensional fused feature.

[0157] Multimodal fault data sample set refers to a synthetic sample set that integrates time and frequency domains, and has more comprehensive fault characterization information.

[0158] In this embodiment of the invention, controllable noise disturbances and operating condition labels can be added to the multimodal fault data sample set to obtain a wind turbine bearing fault data set, so as to simulate real-world noise interference, increase data robustness, and assign accurate operating condition and fault labels to each sample, which facilitates supervised learning training and improves the generalization performance of the model in noisy environments.

[0159] Controllable noise perturbation refers to adding Gaussian white noise, ambient noise, or other real noise types to synthetic samples to simulate interference in actual data acquisition.

[0160] Operating condition labeling refers to labeling each sample with tags such as rotational speed, load, fault type, and severity, forming a complete labeled dataset.

[0161] The wind turbine bearing fault dataset refers to the final generated high-fidelity synthetic fault dataset that includes multimodal features, noise disturbances, and complete labels.

[0162] This invention addresses the problems of model overfitting and poor generalization caused by insufficient data under conditions of extremely scarce real fault samples by generating physically faithful time-domain sequences through forward time integration simulation, increasing sample size through standardized slicing, highlighting fault features through frequency domain gain, improving characterization capabilities through multimodal fusion, and enhancing robustness through noise perturbation and label annotation. At the same time, it overcomes the physical distortion of pure data-driven generation methods, realizing a high-fidelity, diverse, and fully labeled wind turbine bearing fault data set, providing a reliable foundation for training high-precision fault classification target models, and improving the overall accuracy and practicality of intelligent operation and maintenance of wind power equipment.

[0163] In an optional embodiment of the present invention, the step of controlling the wind turbine bearing fault classification target model and generating detection results for the target wind turbine bearing based on target data for the target wind turbine bearing includes: The real-time vibration acceleration signal sequence of the target wind turbine bearing is collected, and the real-time vibration acceleration signal sequence is determined as the target observation signal sequence of the target wind turbine bearing; wherein the target observation signal sequence is the raw time-domain vibration acceleration data continuously collected at a preset location of the target wind power generation equipment according to a preset sampling frequency; The target observation signal sequence of the target wind turbine bearing is subjected to standardization processing and time-series slicing to generate a set of standardized sample sequences for the target model of fault classification of the wind turbine bearing; The standardized target sample sequence set is input into the wind turbine bearing fault classification target model to generate a predicted probability distribution for fault type and fault severity. Based on the predicted probability distribution, the detection results for the fault type and severity of the target wind turbine bearing are determined.

[0164] In this embodiment of the invention, a real-time vibration acceleration signal sequence of a target wind turbine bearing can be collected and determined as the target observation signal sequence of the target wind turbine bearing, so as to obtain the original dynamic response data of the target wind turbine under the current operating state, providing the most direct input basis for subsequent model inference.

[0165] Real-time vibration acceleration signal sequence refers to the raw time-domain vibration acceleration data continuously collected by a high-sensitivity acceleration sensor installed at a preset position such as the bearing housing of the target wind turbine, the high-speed shaft of the gearbox, or the drive end of the generator, at a preset sampling frequency (such as above 10kHz).

[0166] The target observation signal sequence refers to the vibration acceleration signal sequence acquired in real time, which is used to characterize the current operating status and potential fault characteristics of the target wind turbine bearing.

[0167] In this embodiment of the invention, the target observation signal sequence of the target wind turbine bearing can be standardized and time-series sliced ​​to generate a standardized sample sequence set for the target model of the wind turbine bearing fault classification. This eliminates the difference in amplitude dimensions under different operating conditions, unifies the data scale, and transforms the continuous long sequence into short-time samples suitable for model input through slicing, which facilitates efficient model processing and feature extraction.

[0168] Standardization refers to Z-Score standardization, which involves subtracting the mean from the signal and then dividing by the standard deviation to give the signal zero mean and unit variance.

[0169] Time slicing refers to the process of sliding a fixed-length window across the target observed signal sequence to extract segments, forming multiple overlapping or non-overlapping short-time samples to characterize the local dynamic response.

[0170] The standardized sample sequence set refers to the sample set that has been standardized and sliced ​​and can be directly input into the fault classification target model.

[0171] In this embodiment of the invention, the standardized target sample sequence set can be input into the wind turbine bearing fault classification target model to generate a predicted probability distribution for fault type and fault severity. By utilizing the multimodal feature extraction, attention fusion and forward computation of the physical information neural network built into the model, fault representations can be automatically extracted from the input samples, and the confidence probability of each fault category (combination of type and severity) can be output, providing a quantitative basis for the final diagnosis.

[0172] The wind turbine bearing fault classification target model refers to a deep learning model trained using high-fidelity synthetic fault data, which includes feature extraction, fusion, and classification prediction layers.

[0173] The predicted probability distribution refers to the softmax probability vector output by the model, which represents the confidence level of the current sample belonging to each combination of fault type (such as inner race, outer race, rolling element) and severity (such as minor, moderate, severe).

[0174] According to the embodiments of the present invention, the detection results of the fault type and severity of the target wind turbine bearing can be determined based on the predicted probability distribution. By taking the category with the highest probability as the judgment basis, the current state of the target wind turbine bearing can be accurately diagnosed, which facilitates quick decision-making by operation and maintenance personnel.

[0175] The detection result refers to the fault diagnosis conclusion finally output by the model, including the specific fault type (such as inner ring fault) and severity level (such as moderate), usually accompanied by a confidence score.

[0176] This invention addresses the challenge of real-time and accurate early identification of bearing faults in real-time wind power operation environments by employing a complete inference chain, including real-time acquisition of target observation signal sequences, standardized slice preprocessing, model forward inference, and determination of the maximum value of the probability distribution. This ensures high accuracy, robustness, and physical interpretability of the detection results, providing reliable technical support for intelligent operation and maintenance and prevention of unplanned shutdowns of wind power equipment.

[0177] To enable those skilled in the art to better understand the embodiments of the present invention, a complete example is used below to illustrate the embodiments of the present invention.

[0178] In practical applications, wind turbine generators, as key equipment in the new energy system, play a vital role in promoting the achievement of "dual carbon" goals and the transformation of the energy structure. However, with the continuous growth of global wind power installed capacity, the reliability and operation and maintenance costs of wind turbines during long-term operation are becoming increasingly prominent issues. Among them, key rotating components such as main shaft bearings, gearbox bearings, and generator bearings, due to long-term exposure to complex loads such as variable speed operation, random wind loads, temperature fluctuations, and lubrication degradation, have become the weakest links in the entire machine most prone to mechanical failure. Once early damage or failure occurs, it may not only trigger a chain of failures and lead to unplanned downtime, but also result in high maintenance costs and power generation losses.

[0179] Against this backdrop, real-time monitoring of wind turbine bearing health and early fault identification has become one of the core challenges of intelligent wind power operation and maintenance. In recent years, the rapid development of sensor technology, big data analytics, and artificial intelligence has driven the widespread application of data-driven fault diagnosis and predictive maintenance methods. These methods, with their powerful nonlinear feature learning capabilities, can automatically extract fault modes from massive amounts of operational data, demonstrating significant potential in early warning, condition identification, and lifespan prediction. However, their performance is highly dependent on a large number of high-quality fault samples, while fault data is extremely scarce in actual wind power scenarios, with operational data primarily based on healthy operating conditions. Imbalanced samples not only limit the model training effect but also lead to insufficient generalization ability and increased false alarm rates.

[0180] To address this bottleneck, fault data generation technology has gradually become a research hotspot, aiming to synthesize high-fidelity vibration responses under different health conditions using limited measured data, thereby improving the robustness and accuracy of intelligent diagnostic models. Although existing data augmentation techniques (such as GANs and VAEs) have made some progress in fault sample synthesis, the generated data often lacks clear physical consistency and is difficult to accurately reflect the dynamic characteristics in the bearing fault evolution process.

[0181] In view of this, this invention proposes a method for generating wind turbine bearing fault data based on Inverse Physics-Informed Neural Networks (Inverse PINN). This method is based on a "data-physics dual-drive" paradigm, integrating the dynamic mechanism of the bearing system with measured vibration data to construct an Inverse Physics-Informed Neural Network (Inverse PINN) framework. By inverting implicit excitations or structural parameter changes under fault conditions, it generates high-fidelity and diverse fault vibration responses while strictly satisfying physical laws. This method avoids the generation distortion caused by the lack of physical constraints in purely data-driven models and overcomes the dependence of traditional physical simulations on accurate fault modeling and a large number of prior parameters, providing a reliable data foundation and technical support for intelligent fault diagnosis of wind turbine bearings under small sample conditions.

[0182] First, a simulation network structure is constructed based on the bearing dynamics equations and actual engineering conditions. Next, an asymmetric convolutional inverse solution deep sensing network structure is constructed for the inverse solution of physical parameters. Then, the inverse parameters are encoded, and limited real vibration data is used as constraints to solve for the dynamic system parameters in reverse. Finally, the obtained real dynamic characteristics are used to generate data for multiple operating conditions and multiple fault levels. Specific steps include: Step 1: Construct a simulation network structure based on the bearing dynamics equations and actual engineering conditions.

[0183] The simulation network in this study consists of a two-stage gearbox dynamics simulation system, which integrates dynamic modeling, time-varying stiffness calculation under complex actual operating conditions, and customized signal processing algorithms. This simulation system constructs a high-fidelity multibody dynamics model for a two-stage reduction gearbox, aiming to accurately characterize its nonlinear vibration response under complex operating conditions. The system input consists of a complete set of physical parameters, covering macroscopic operating conditions (such as high-speed shaft speed). , load torque and simulation duration Microscopic geometric parameters (including the number of teeth of each gear) Modulus Pressure angle and shaft segment length ) and material mechanical properties (such as mass) Moment of inertia Shaft segment torsional stiffness and elastic modulus In addition, the system introduces a series of non-ideal factors to improve the realism of the modeling: gear eccentricity error. Multi-order machining error harmonic components Geometric deviations of bearing inner / outer rings and radial clearance And set the damping coefficient This reflects the energy dissipation mechanism of the structure, thus providing a solid physical basis for the dynamic behavior of real mechanical systems.

[0184] The following describes the dynamic modeling and solution, numerical solution process, output results, and signal enhancement strategy of this simulation system: In terms of dynamic modeling, the system adopts a 24-degree-of-freedom coupled nonlinear dynamic model, comprehensively considering the interaction between torsional and lateral vibrations. This model includes not only the torsional displacements of each axis. It also introduced lateral displacement in two directions. This forms a comprehensive three-dimensional motion description framework. The system's motion equations can be expressed as:

[0185] in, For generalized coordinate vectors, For the quality matrix, Here is the damping matrix. It is a time-varying stiffness matrix, whose elements change significantly over time, originating from the periodic fluctuations of meshing stiffness and bearing support stiffness.

[0186] To achieve high-precision stiffness modeling, the system employs an integral calculation method based on the potential energy method. By analyzing the tooth profile geometric parameters (base circle, addendum circle, dedendum circle, etc.), it calculates the bending, shear, and axial compressive deformation energies, thereby obtaining the true nonlinear time-varying meshing stiffness. Specifically, the meshing stiffness function Depends on the relative rotation angle between the two gears, the expression is:

[0187] in The dynamic energy of inter-tooth contact. This represents the actual meshing clearance. This method overcomes the limitations of traditional sine or square wave approximations and can accurately capture the modulation effect caused by tooth profile errors. Simultaneously, the system introduces a time-varying bearing support stiffness model (TVBS) to consider the stiffness fluctuations caused by the periodic rolling of the rolling elements on the raceway. This model is based on the number of rolling elements. Cage speed and dynamic calculation of bearing stiffness at journal position This effectively simulates the modulation effect of bearing pass frequency (BPO / BPI) on the system response.

[0188] During the dynamic evolution process, the system introduces a nonlinear function of tooth backlash and a multi-harmonic transmission error (TE) term to reproduce the chaotic jump behavior caused by the backlash and the sideband structure generated by error modulation. The transmission error function can be expressed as:

[0189] in The meshing frequency, For the processing error amplitude, This represents the magnitude of the eccentricity error. These nonlinear factors work together to cause the system to exhibit complex transient response characteristics even under steady-state excitation.

[0190] For numerical solutions, the system employs the backward difference formula method to perform time integration on the aforementioned rigid system. Due to the drastic changes in stiffness within the gear system (up to...),... In rigid systems with strong nonlinear coupling, traditional explicit integrators are prone to numerical divergence. The BDF method, due to its implicit characteristics and good stability, exhibits significant advantages in handling such rigid systems, ensuring convergence and accuracy in long-term simulations.

[0191] The system generates rich dynamic response data based on the output results and signal gain strategy, which is used to reveal the vibration mechanism and fault characteristics of the system. Firstly, the output includes the input shaft, intermediate shaft, and output shaft... , Time-domain acceleration response curve in the direction This visually demonstrates the vibration evolution of the system at different spatial locations. Secondly, by performing a Fast Fourier Transform (FFT) on the acceleration signal, the single-sided amplitude spectrum is obtained. Furthermore, a frequency domain gain weighting strategy is implemented, which involves applying a controllable gain in key frequency bands (such as the meshing frequency and its harmonics) to highlight potential weak modulation components or early fault impulse signals. For example, a weighting factor can be defined as follows:

[0192] in The meshing frequency, This is the gain coefficient. Controlled bandwidth. This strategy significantly improves the identifiability of fault characteristics under low signal-to-noise ratio conditions.

[0193] Furthermore, the system synchronously outputs the time-varying parameters of each gear stage, directly reflecting the system parameter excitation source. The phase relationship between these curves and the vibration response can be used to verify the fault propagation path and resonance mechanism. The overall modeling process realizes a closed-loop simulation system from physical parameter input, high-fidelity stiffness modeling, coupled nonlinear dynamic evolution, customized signal enhancement, and multi-dimensional dynamic output, possessing high engineering applicability and scientific research value.

[0194] Step 2: Construct an asymmetric convolutional inverse solution for the deep perception network.

[0195] The inverse PINN model aims to invert system physical parameters from observed structural responses (displacement / velocity / acceleration + external load / excitation) while ensuring that the predicted time-domain response satisfies the constraints of the physical equations. The model consists of a 1D time-series network (naturally adapted to vibration signals and sensor time series), with the core component being an asymmetric convolutional residual block. Asymmetric convolution refers to using asymmetric convolution kernels or parallel branches on the 1D time series. and (In practice, this is a one-dimensional convolution with different widths), utilizing parallel branches (narrow kernel + wide kernel) with different receptive fields and fusing them to capture short-term and long-term correlations.

[0196] Dual-head encoder section: Input: Observed signal sequence It includes any measurable quantity (such as acceleration). (If there is displacement / velocity, it can also be included) and external excitation. Input dimensions .

[0197] Response Head (Temporal Reconstruction / Prediction): Direct Network Prediction And obtained by automatic differentiation As an auxiliary constraint / loss term.

[0198] Physics Head (Physical Parameters): from Regression through several fully connected layers The output is constrained by the physical feasible region. ).

[0199] Asymmetric convolutional residual blocks, each block including: parallel branches: Conv1D(kernel=3, dilation=1) and Conv1D(kernel=7, dilation=2) (a branch with a larger receptive field based on the input features); the branch outputs are concatenated and then... Conv performs channel fusion; residual connections (input added to output). Asymmetric convolutional residual blocks + skip connections are used to preserve local details.

[0200] Adaptive Normalization Layer (BN): This layer preserves learnable scale and translation and weights the statistics within the time window.

[0201] Activation: GELU or SiLU.

[0202] The final pooling / downsampling layer yields a fixed-length representation. .

[0203] Step 3: Encode the inverse parameters and use the limited real vibration data as constraints to solve the dynamic system parameters in reverse.

[0204] Within the framework of Inverse PINN, the core objective is to identify key structural parameters and potential fault characteristic factors of a dynamic system using only a limited number of real vibration response samples. To achieve this, an "inverse parameter encoding mechanism" needs to be constructed within the network, enabling the model to automatically infer damping coefficients. Linear stiffness Nonlinear stiffness parameters Impact amplitude Fault characteristic frequency This includes implicit quantities that are not visible to the naked eye. Solving these implicit physical quantities is essentially an inverse problem under the constraints of partial differential equations. The inverse PINN network achieves the joint constraint solution of physical laws and data through automatic differentiation and end-to-end training.

[0205] In practical implementation, the physical parameters to be identified are first treated as trainable vectors and then inversely encoded into the feature space through the network structure in step two, giving them the nonlinear ability to express complex fault behaviors and structural dynamic characteristics. The output of the parameter network is not directly embedded in the prediction layer but participates in the network as a physical residual. During this process, the data loss can stably reflect whether the current parameter configuration meets the operating conditions. Meanwhile, based on a limited number of real samples... Networks suffer from data consistency losses:

[0206] By using a loss function, the network drives the alignment of the predicted response with the measurement results, thereby shrinking the solution space of the inverse problem to within the feasible region.

[0207] To avoid physically unreasonable drift in the inversion parameters, a regularization term is introduced.

[0208] Its function is to apply soft constraints to the parameter search space using prior structural information, so that the deduction results can still maintain mechanical interpretability even when data is insufficient.

[0209] The overall objective function is as follows:

[0210] The entire process forms a dual coupling mechanism of "data-prior". Ultimately, the gradient descent network simultaneously adjusts the response prediction module and the parameter encoding module, enabling the system parameters to automatically converge to the optimal solution within a physically feasible range. This inverse solution strategy allows the model to accurately recover fault characteristics and structural parameters under limited observation conditions, providing a reliable foundation for health monitoring and fault diagnosis of real engineering equipment.

[0211] Step four: Use the real dynamic characteristics obtained by reverse engineering to generate data for multiple operating conditions and multiple fault levels.

[0212] Using the inversely derived true dynamic parameters as the basic parameters, controllable disturbances are added to generate signals for different operating conditions: Multiple speeds: Add different rotation speeds

[0213] Multiple loads: random disturbances

[0214] Multiple fault levels: Adjustment Attenuation coefficient, impact period, etc. Generative models are defined as follows:

[0215] The parameter set This includes structural dynamic parameters, failure mechanism parameters, and initial state variables. Dynamic parameters The fault parameters determine the intrinsic response characteristics of the system. This allows the model to flexibly generate different types of abnormal signals during inner ring failure, outer ring failure, rolling element defects, multi-point spalling, and crack propagation by controlling the impact amplitude, fault characteristic frequency, damage scale, and propagation state. Initial condition parameters. Provide the generator with a realistic system startup state.

[0216] By reversing the physical constraint solution mechanism within PINN, the generative model strictly adheres to nonlinear dynamic equations in each forward simulation, ensuring high physical consistency in energy coupling, impact structures, and spectral characteristics of the generated sequence. The final output... It can be directly used for fault classification, remaining lifetime prediction, digital twin-driven simulation, and data augmentation tasks.

[0217] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0218] Reference Figure 2 The diagram shows a structural block diagram of a wind turbine bearing testing device for wind power generation equipment provided in an embodiment of the present invention, which may specifically include the following modules: The standardized sample sequence set generation module 201 is used to collect the observation signal sequence of the wind turbine bearing, and perform standardization processing and time-series slicing on the observation signal sequence to generate a standardized sample sequence set. The actual engineering operating condition parameter determination module 202 is used to determine the actual engineering operating condition parameters of the wind turbine bearing; the actual engineering operating condition parameters include: operating condition information characterizing the operating conditions of the wind turbine bearing in the real operating environment, geometric parameters characterizing the geometric features of the wind turbine bearing, and the material mechanical properties of the wind turbine bearing. The dynamic equation constraint determination module 203 is used to determine the dynamic equation constraint conditions based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties. The inversion physical parameter vector inversion module 204 is used to invert the implicit inversion physical parameter vector constrained by the physical feasible region in the observation signal sequence from the observation signal sequence using the standardized sample sequence set and the constraints of the dynamic equation, and employing an inverse physical information neural network. The multi-condition fault parameter set generation module 205 is used to generate a multi-condition fault parameter set constrained by the dynamic equation based on the inverted physical parameter vector. The wind turbine bearing fault classification target model generation module 206 is used to construct a wind turbine bearing fault data set based on the extended multi-condition fault parameter set, and to generate a wind turbine bearing fault classification target model using the wind turbine bearing fault data set. The detection result generation module 207 is used to control the target model for fault classification of the wind turbine bearing and generate detection results for the target wind turbine bearing based on the target data for the target wind turbine bearing.

[0219] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0220] In addition, embodiments of the present invention also provide an electronic device, such as... Figure 3 As shown, it includes a processor 301, a communication interface 302, a memory 303, and a communication bus 304, wherein the processor 301, the communication interface 302, and the memory 303 communicate with each other through the communication bus 304. Memory 303 is used to store computer programs; When the processor 301 executes the program stored in the memory 303, it implements the wind turbine bearing detection method of any of the wind power generation equipment described in the above embodiments: The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0221] The communication interface is used for communication between the aforementioned terminal and other devices.

[0222] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0223] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0224] like Figure 4As shown, in another embodiment of the present invention, a computer-readable storage medium 401 is also provided, which stores instructions that, when run on a computer, cause the computer to execute the wind turbine bearing detection method for wind power generation equipment described in the above embodiment.

[0225] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described wind turbine bearing detection method embodiment and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0226] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0227] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0228] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0229] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for testing wind turbine bearings in wind power generation equipment, characterized in that, include: Collect the observation signal sequence of the wind turbine bearing, and perform standardization and time-series slicing on the observation signal sequence to generate a standardized sample sequence set; The actual engineering operating condition parameters of the wind turbine bearing are specified; the actual engineering operating condition parameters include: operating condition information characterizing the operating conditions of the wind turbine bearing in the real operating environment, geometric parameters characterizing the geometric characteristics of the wind turbine bearing, and the material mechanical properties of the wind turbine bearing; Based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties, the constraints of the dynamic equations are determined; Using the standardized sample sequence set and the constraints of the dynamic equation, an inverse physical information neural network is used to invert the inverse physical parameter vector implicit in the observed signal sequence, which is constrained by the physical feasible region. A multi-condition fault parameter set constrained by the dynamic equation is generated based on the inverted physical parameter vector. Based on the extended multi-condition fault parameter set, a wind turbine bearing fault data set is constructed, and the wind turbine bearing fault classification target model is generated using the wind turbine bearing fault data set. The target model for classifying the faults of the wind turbine bearing is controlled, and based on the target data for the target wind turbine bearing, the detection results for the target wind turbine bearing are generated.

2. The method according to claim 1, characterized in that, The step of determining the constraints of the dynamic equations based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties includes: Based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties, a degree-of-freedom coupled nonlinear dynamic equation is constructed. The time-varying meshing stiffness and bearing support stiffness of the wind turbine bearing are calculated using the aforementioned degree-of-freedom coupled nonlinear dynamic equations. The bending, shearing and axial compression deformation energies are calculated by integrating using the potential energy method, thereby generating a nonlinear time-varying stiffness matrix. The gear eccentricity error, multi-order machining error harmonic components, and bearing geometric dimension deviation are determined, and the nonlinear time-varying stiffness matrix is ​​adjusted using the gear eccentricity error, the multi-order machining error harmonic components, and the bearing geometric dimension deviation to generate the motion equation of the enhanced dynamic system. The motion equations of the enhanced dynamic system are solved numerically by time integration using the backward difference formula method, generating time-domain acceleration response curves of the input axis, intermediate axis and output axis in multiple directions; Perform a fast Fourier transform on the time-domain acceleration response curve to generate a one-sided amplitude spectrum; Frequency domain gain weighting is performed on the single-sided amplitude spectrum to generate an enhanced spectrum signal that highlights the characteristics of the fault impact. An input feature set is constructed based on the enhanced spectral signal and the observed signal sequence.

3. The method according to claim 2, characterized in that, The step of using the standardized sample sequence set and the constraints of the dynamic equations to invert the implicit inversion physical parameter vector constrained by the physical feasible region from the observed signal sequence using an inverse physical information neural network includes: By inputting the input feature set in parallel into a shared asymmetric convolutional residual block containing response head branches and physical head branches, a temporal feature representation that integrates short-term local impact details and long-term global periodic dynamics is generated. An adaptive normalization layer and activation function are applied to the temporal feature representation. Learnable scale and translation parameters are preserved through batch normalization, and the statistics within the time window are weighted to generate a normalized deep feature vector. The normalized deep feature vector is used to regress physical parameters through a fully connected layer to generate an inverted physical parameter vector constrained by the physical feasible region.

4. The method according to claim 3, characterized in that, The step of generating a multi-condition fault parameter set based on the inverted physical parameter vector and constrained by the dynamic equation includes: The inverted physical parameter vectors are encoded into the feature space. By constructing an inverse parameter encoding mechanism, the damping coefficient, linear stiffness, nonlinear stiffness parameters, impact amplitude, and fault characteristic frequency are used as trainable vectors to generate implicit physical quantity encodings capable of expressing complex fault behaviors. Based on the standardized sample sequence set and the input feature set, the data consistency loss is calculated. The difference between the predicted response and the actual measurement result is quantified by the data consistency loss quantification model, and preliminary optimized dynamic system parameters are generated. A regularization term is introduced to impose soft constraints on the parameters of the initially optimized dynamic system. By utilizing prior structural information to shrink the solution space, a set of inverse dynamic parameters with reasonable physical state is generated. The inverse dynamic parameter set is determined as the basic parameter set. Multiple sets of typical speed disturbance values, multiple load random disturbance coefficients, multiple fault level adjustment attenuation coefficients and impact period coefficients are determined. The multiple sets of typical speed disturbance values, multiple load random disturbance coefficients, multiple fault level adjustment attenuation coefficients and impact period coefficients are added to the basic parameter set to generate an extended multi-condition fault parameter set.

5. The method according to claim 2, characterized in that, The step of performing frequency domain gain weighting processing on the single-sided amplitude spectrum to generate an enhanced spectrum signal for highlighting fault impact characteristics includes: Identify the fundamental frequency position of the meshing frequency and the positions of each harmonic of the fundamental frequency position in the single-sided amplitude spectrum, and generate a set of key fault focus frequencies based on the fundamental frequency position and the positions of each harmonic. Based on the set of key fault focus frequencies, construct a Gaussian or rectangular window-shaped weighting factor function, and calculate a gain weighting vector with the same unilateral amplitude spectrum length as the set of key fault focus frequencies. The gain weighting vector is multiplied element-wise with the single-sided amplitude spectrum to generate a weighted spectrum with an amplitude distribution that highlights the characteristics of the fault impact. An inverse fast Fourier transform is performed on the weighted spectrum to generate an enhanced spectral signal that highlights the characteristics of the fault impact.

6. The method according to claim 4, characterized in that, The steps for constructing a wind turbine bearing fault data set based on the extended multi-condition fault parameter set include: Based on the extended multi-condition fault parameter set, the generation model is driven to perform forward time integration simulation to obtain a time-domain vibration response sequence that meets the constraints of the nonlinear dynamic equation. The time-domain vibration response sequence is standardized and sliced ​​using a sliding window to obtain a set of standardized fault sample sequences under multiple operating conditions. The standardized fault sample sequence set is subjected to frequency domain transformation and gain weighting to obtain an enhanced fault spectrum signal set that highlights the early damage impact and modulation characteristics. The enhanced fault spectrum signal set and the standardized fault sample sequence set are superimposed by channel dimension or feature splicing to obtain a multimodal fault data sample set that integrates time domain and frequency domain information; Controllable noise disturbances and operating condition labels are added to the multimodal fault data sample set to obtain the wind turbine bearing fault data set.

7. The method according to claim 6, characterized in that, The steps of controlling the wind turbine bearing fault classification target model and generating detection results for the target wind turbine bearing based on target data for the target wind turbine bearing include: The real-time vibration acceleration signal sequence of the target wind turbine bearing is collected, and the real-time vibration acceleration signal sequence is determined as the target observation signal sequence of the target wind turbine bearing; wherein the target observation signal sequence is the raw time-domain vibration acceleration data continuously collected at a preset location of the target wind power generation equipment according to a preset sampling frequency; The target observation signal sequence of the target wind turbine bearing is subjected to standardization processing and time-series slicing to generate a set of standardized sample sequences for the target model of fault classification of the wind turbine bearing; The standardized target sample sequence set is input into the wind turbine bearing fault classification target model to generate a predicted probability distribution for fault type and fault severity. Based on the predicted probability distribution, the detection results for the fault type and severity of the target wind turbine bearing are determined.

8. A wind turbine bearing testing device for wind power generation equipment, characterized in that, include: The standardized sample sequence set generation module is used to collect the observation signal sequence of wind turbine bearings, and perform standardization processing and time-series slicing on the observation signal sequence to generate a standardized sample sequence set. The actual engineering operating condition parameter determination module is used to determine the actual engineering operating condition parameters of the wind turbine bearing; the actual engineering operating condition parameters include: operating condition information characterizing the operating conditions of the wind turbine bearing in the real operating environment, geometric parameters characterizing the geometric features of the wind turbine bearing, and the material mechanical properties of the wind turbine bearing. The dynamic equation constraint determination module is used to determine the dynamic equation constraint conditions based on the standardized sample sequence set, the operating condition information, the geometric parameters, and the material mechanical properties. The inversion physical parameter vector inversion module is used to invert the implicit inversion physical parameter vector constrained by the physical feasible region in the observation signal sequence from the observation signal sequence using the standardized sample sequence set and the constraints of the dynamic equation, and employing an inverse physical information neural network. A multi-condition fault parameter set generation module is used to generate a multi-condition fault parameter set constrained by the dynamic equation based on the inverted physical parameter vector. The wind turbine bearing fault classification target model generation module is used to construct a wind turbine bearing fault data set based on the extended multi-condition fault parameter set, and to generate a wind turbine bearing fault classification target model using the wind turbine bearing fault data set. The detection result generation module is used to control the target model for fault classification of the wind turbine bearing, and generate detection results for the target wind turbine bearing based on the target data for the target wind turbine bearing.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as described in claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the method as described in claims 1-7.

Images