A method for accurately diagnosing fan blade coupling faults under variable working conditions

Abstract

This invention discloses a precise diagnostic method for coupled faults in wind turbine blades under varying operating conditions, belonging to the field of blade turbine fault diagnosis technology. The invention includes synchronously acquiring and preprocessing strain, pressure, angular acceleration, and operating parameters; achieving signal synchronization and mode decomposition through phase compensation and variational mode decomposition; employing condition-adaptive sparse Bayesian learning to decouple leading-edge corrosion and bolt loosening characteristics; correcting fault coefficients and quantifying severity using pressure data; and then verifying the results using angular acceleration. This invention, by synchronously acquiring blade strain, leading-edge pressure, and generator angular acceleration, and employing cross-domain phase synchronization and condition-adaptive sparse Bayesian learning algorithms, can decouple and quantify the coupled characteristics of leading-edge corrosion and bolt loosening from the aliased signals under varying operating conditions.

Description

Technical Field

[0001] This invention belongs to the field of blade machine fault diagnosis technology, and in particular relates to a method for accurate diagnosis of wind turbine blade coupling faults under variable operating conditions. Background Technology

[0002] During long-term operation, wind turbine blades, as the core component for capturing wind energy, are prone to two typical mechanical failures: leading-edge corrosion and bolt loosening. Leading-edge corrosion refers to the gradual erosion and increased roughness of the protective layer or substrate on the leading edge of the blade due to rain erosion, sand and dust impact, and ultraviolet aging. This failure alters the aerodynamic damping characteristics of the blade surface, affecting the blade's transient response in flapping and flaring directions. Bolt loosening refers to the decrease in preload or even gaps in the bolts connecting the blade root and hub due to alternating loads, leading to a reduction in the stiffness of the blade root connection and thus changing the blade's natural frequency and boundary conditions. These two failures may occur individually under normal operating conditions, but under strong non-stationary conditions such as after a typhoon, wind speed and direction change drastically. The wind turbine control system will simultaneously perform pitch (changing the pitch angle) and speed (adjusting the generator torque) actions, subjecting the blades to complex and non-stationary aerodynamic loads. At this point, the effects of leading edge corrosion and bolt loosening are highly coupled in traditional monitoring signals such as vibration and torque. Meanwhile, non-faulty operations such as pitch rate and speed change rate will also generate similar fault excitation characteristics, causing the characteristics of the two faults to overlap and become difficult to separate effectively.

[0003] Current wind turbine blade fault diagnosis methods mainly rely on single-type sensor signals. For example, analyzing the peak shift of the spectrum of blade root vibration acceleration signals to determine bolt loosening, or monitoring fatigue accumulation through strain signals to assess leading-edge corrosion. However, under varying operating conditions, the frequency components of the vibration signal are simultaneously affected by changes in aerodynamic load and structural stiffness. The peak shift may be caused by wind speed fluctuations rather than bolt loosening. Similarly, amplitude changes in the strain signal may originate from the additional load generated by pitch control rather than changes in aerodynamic damping caused by leading-edge corrosion. Existing technologies lack methods to simultaneously separate aerodynamic excitation and structural response under highly non-stationary operating conditions and distinguish between the two coupled fault characteristics. Furthermore, traditional methods typically assume stable operating conditions or that fault characteristics have fixed frequency and amplitude patterns, making it difficult to adapt to dynamic environments with real-time changes in pitch rate and rotational speed. This leads to misdiagnosing non-fault excitations generated by variable operating conditions as faults or missing true coupled fault characteristics. Therefore, the following solutions are proposed to address these issues. Summary of the Invention

[0004] The purpose of this invention is to provide a precise diagnostic method for coupled faults in wind turbine blades under varying operating conditions. By synchronously acquiring blade strain, leading-edge pressure, and generator angular acceleration, and employing cross-domain phase synchronization and operating condition adaptive sparse Bayesian learning algorithms, the coupling characteristics of leading-edge corrosion and bolt loosening can be decoupled and quantified from the mixed signals under varying operating conditions. This solves the problem that existing technologies are difficult to separate the two faults and are easily affected by pitch and speed change operations.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0006] This invention provides a method for accurate diagnosis of coupling faults in wind turbine blades under varying operating conditions, comprising the following steps:

[0007] Strain measurement points and dynamic pressure sensors are arranged on the wind turbine blades, and an angular acceleration measurement device is installed at the generator output shaft end. Operating parameters of the wind turbine main control system are collected synchronously. Cross-domain phase synchronization is performed on the collected multi-source heterogeneous signals to compensate for phase jumps caused by varying operating conditions, obtaining phase-aligned strain, pressure, and angular acceleration signals. Variational mode decomposition is used to extract the modal components of the signals. An overcomplete dictionary containing a leading-edge corrosion dictionary and a bolt loosening dictionary is constructed. An operating condition-adaptive sparse Bayesian learning method is used to solve for the sparse coefficient vector, achieving decoupling of the two fault characteristics. Correction factors are calculated using dynamic pressure data to quantify and correct the decoupled fault coefficients, obtaining leading-edge corrosion severity indicators and bolt loosening severity indicators. Residual verification is performed using angular acceleration signals to ensure the reliability of the diagnostic results.

[0008] Furthermore, the strain measurement points are arranged according to the vibration mode or mode along the blade root to blade tip direction. Each measurement point is equipped with a triaxial resistance strain gauge, which simultaneously collects the strain in the blade flapping direction and the strain in the oscillation direction. The dynamic pressure sensor is embedded in the easily corroded surface of the blade leading edge. Each sensor is connected to a multi-channel dynamic pressure acquisition system via an independent shielded cable. The number of channels, sampling frequency, and sampling accuracy of the system are configured according to actual testing requirements and are compatible with both current and voltage signals.

[0009] Furthermore, the cross-domain phase synchronization specifically includes: performing bandpass filtering on each signal channel, with the passband range covering the first three flapping and yaw mode frequencies of the blade; extracting the instantaneous phase of each signal using Hilbert transform; calculating the phase difference of other signals relative to the reference signal using blade root flapping strain as the reference signal; establishing a phase compensation model based on the pitch rate to compensate for non-fault phase jumps caused by pitching action; and using the compensated phase difference to resample and interpolate the signal to achieve complete phase synchronization of cross-domain signals.

[0010] Furthermore, in the condition-adaptive sparse Bayesian learning method, the synchronized flapping strain, oscillation strain, pressure signal, and angular acceleration signal are concatenated into a joint observation vector. Each column of the overcomplete dictionary represents a parameterized basis function for a single fault source under different operating conditions. The basis functions of the leading-edge corrosion dictionary include aerodynamic damping ratio and wind speed fluctuation shape function, while the basis functions of the bolt loosening dictionary include relaxation time constant and pulse sequence triggered by pitch rate. A condition-adaptive prior variance is introduced, which is dynamically adjusted according to pitch rate and wind speed change rate, allowing for significant changes in the sparsity coefficients during transients under varying operating conditions, thus avoiding misjudging operating condition disturbances as faults.

[0011] Furthermore, the condition-adaptive sparse Bayesian learning employs an expectation-maximization framework to iteratively solve for the sparse coefficients. In the expectation step, the posterior mean and covariance are calculated based on the current noise variance and prior variance matrix. In the maximization step, the noise variance and prior variance basis values ​​are updated. The iteration continues until the logarithmic marginal likelihood function converges. The converged posterior mean is the activation coefficient of each fault basis function, where the coefficients corresponding to the leading edge corrosion dictionary constitute the leading edge corrosion coefficient vector, and the coefficients corresponding to the bolt loosening dictionary constitute the bolt loosening coefficient vector.

[0012] Furthermore, the calculation of correction factors using dynamic pressure data specifically includes: calculating the pressure pulsation turbulence intensity based on all pressure signals, where the turbulence intensity is the ratio of the root mean square of the pressure pulsation to the mean pressure; calculating the coherence function between the pressure signal and the flapping strain, extracting coherence function values ​​at the blade flapping natural frequency and the oscillation natural frequency respectively; constructing a pressure correction factor based on the two coherence function values, where the leading-edge corrosion correction factor is the ratio of the flapping frequency coherence function value to the sum of the two coherence function values, and the bolt loosening correction factor is one minus the leading-edge corrosion correction factor.

[0013] Furthermore, the severity index of leading-edge corrosion is obtained by multiplying the leading-edge corrosion correction factor, the L1 norm of the leading-edge corrosion coefficient vector, and the ratio of the pressure pulsation turbulence intensity to the healthy state reference turbulence intensity; the severity index of bolt loosening is obtained by multiplying the bolt loosening correction factor, the L1 norm of the bolt loosening coefficient vector, and the relative offset of the natural frequency of the oscillation; when the two severity indices exceed the corresponding preset thresholds, the corresponding fault is determined to reach the level requiring maintenance.

[0014] Furthermore, residual verification using angular acceleration signals specifically includes: calculating the fluctuation variance of the acquired angular acceleration signals; reconstructing the angular acceleration signals from the decoupled leading-edge corrosion coefficient vector and bolt loosening coefficient vector using the transfer matrix that maps the fault coefficients of the strain domain to the torque domain; calculating the residual between the measured angular acceleration signals and the reconstructed angular acceleration signals; if the residual is greater than a set multiple of the noise standard deviation, a re-diagnosis process is triggered, returning to adjust the phase synchronization parameters or operating condition sensitivity factor until the residual converges.

[0015] Furthermore, the operating parameters of the wind turbine main control system include wind speed, wind turbine speed or generator speed, pitch angle, pitch rate and generator torque. All operating parameters are synchronized with strain signal, pressure signal and angular acceleration signal using a precision time protocol to ensure that the time alignment error between different physical quantity signals is less than a preset value.

[0016] Furthermore, the variational mode decomposition decomposes the phase-synchronized strain signal into several intrinsic mode functions and extracts the instantaneous amplitude and instantaneous frequency of each intrinsic mode function; the same variational mode decomposition process is performed on the pressure signal and the angular acceleration signal, and the decomposed modal components are used for subsequent construction of an overcomplete dictionary and extraction of fault features.

[0017] The present invention has the following beneficial effects:

[0018] 1. This invention synchronously acquires multiple physical quantity signals such as blade strain, leading-edge pressure, and generator angular acceleration, and uses cross-domain phase synchronization technology to eliminate phase shifts caused by non-stationary operations such as pitch and speed changes, ensuring that signals from different sources are strictly aligned in time. Based on this, an overcomplete dictionary containing two fault features, leading-edge corrosion and bolt loosening, is constructed using an operating condition adaptive sparse Bayesian learning algorithm. The dictionary atoms are dynamically generated according to real-time operating condition parameters. The sparse solution process can decompose the mixed observation vector into sparse coefficients corresponding to different fault sources, thereby achieving the separation of the two coupled fault features under complex operating conditions such as variable speed and pitch speed, and avoiding interference from non-fault excitations on the diagnostic results.

[0019] 2. This invention incorporates instantaneous pressure data measured by a leading-edge pressure sensor array into the fault quantification process. By calculating the pressure pulsation turbulence intensity and the pressure strain coherence function, dynamic changes in aerodynamic loads on the blade surface are obtained. This information is used to correct the fault coefficients output by sparse Bayesian learning, enabling the assessment of the severity of leading-edge corrosion to reflect the actual impact of surface roughness changes on aerodynamic damping. Simultaneously, the pressure data also helps to determine the coherence relationship between aerodynamic excitation and structural response, thereby adjusting the confidence weights of different fault types. This makes the diagnostic results adaptive to changes in aerodynamic boundary conditions such as wind speed pulsation and turbulence intensity.

[0020] 3. This invention employs a data-driven sparse Bayesian learning framework. The fault dictionary atoms are dynamically generated based on real-time operating parameters, and the prior variance is adaptively adjusted with the pitch rate and wind speed change rate, eliminating the need for preset fixed fault waveform templates. The expectation-maximization iterative process automatically learns the posterior distribution of noise variance and sparse coefficients, and the correction factor in the fault severity index is calculated in real time based on the coherence function of pressure and strain. The entire diagnostic process is automatically completed based on the inherent statistical characteristics of sensor data, reducing reliance on manually set empirical parameters or static discrimination rules, making the diagnostic results more objective and repeatable.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a method for accurate diagnosis of wind turbine blade coupling faults under varying operating conditions, as described in this invention.

[0024] Figure 2 This is a schematic diagram of the sensor position according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figure 1-2 As shown, this invention provides a precise diagnostic method for wind turbine blade coupling faults under varying operating conditions. The diagnostic method includes the following steps:

[0027] Strain measurement points and dynamic pressure sensors are arranged on the wind turbine blades, and an angular acceleration measurement device is installed at the generator output shaft end. Operating parameters of the wind turbine main control system are collected synchronously. Cross-domain phase synchronization is performed on the collected multi-source heterogeneous signals to compensate for phase jumps caused by varying operating conditions, obtaining phase-aligned strain, pressure, and angular acceleration signals. Variational mode decomposition is used to extract the modal components of the signals. An overcomplete dictionary containing a leading-edge corrosion dictionary and a bolt loosening dictionary is constructed. An operating condition-adaptive sparse Bayesian learning method is used to solve for the sparse coefficient vector, achieving decoupling of the two fault characteristics. Correction factors are calculated using dynamic pressure data to quantify and correct the decoupled fault coefficients, obtaining leading-edge corrosion severity indicators and bolt loosening severity indicators. Residual verification is performed using angular acceleration signals to ensure the reliability of the diagnostic results.

[0028] The strain measurement points are arranged along the blade root to blade tip direction according to the vibration mode or mode. Each measurement point is equipped with a triaxial resistance strain gauge, which simultaneously collects the strain in the blade flapping direction and the strain in the oscillation direction. The dynamic pressure sensor is embedded in the easily corroded surface of the blade leading edge. Each sensor is connected to the multi-channel dynamic pressure acquisition system via an independent shielded cable. The number of channels, sampling frequency and sampling accuracy of the system are configured according to the actual test requirements, and it is compatible with current signals and voltage signals.

[0029] Cross-domain phase synchronization specifically includes: performing bandpass filtering on each signal channel, with the passband range covering the first three flapping and yaw mode frequencies of the blade; extracting the instantaneous phase of each signal using Hilbert transform; calculating the phase difference of other signals relative to the reference signal using blade root flapping strain as the reference signal; establishing a phase compensation model based on the pitch rate to compensate for non-fault phase jumps caused by pitching action; and using the compensated phase difference to resample and interpolate the signals to achieve complete phase synchronization of cross-domain signals.

[0030] In the condition-adaptive sparse Bayesian learning method, the synchronized flapping strain, oscillation strain, pressure signal, and angular acceleration signal are concatenated into a joint observation vector. Each column of the overcomplete dictionary represents a parameterized basis function for a single fault source under different operating conditions. The basis functions of the leading-edge corrosion dictionary include aerodynamic damping ratio and wind speed fluctuation shape function, while the basis functions of the bolt loosening dictionary include relaxation time constant and pulse sequence triggered by pitch rate. A condition-adaptive prior variance is introduced, which is dynamically adjusted according to pitch rate and wind speed change rate, allowing for large changes in sparsity coefficients in transient conditions under varying operating conditions, thus avoiding misjudging operating condition disturbances as faults.

[0031] The working condition adaptive sparse Bayesian learning uses an expectation-maximization framework to iteratively solve for sparse coefficients. In the expectation step, the posterior mean and covariance are calculated based on the current noise variance and prior variance matrix. In the maximization step, the noise variance and prior variance basis values ​​are updated. The iteration continues until the log-marginal likelihood function converges. The converged posterior mean is the activation coefficient of each fault basis function. The coefficients corresponding to the leading edge corrosion dictionary form the leading edge corrosion coefficient vector, and the coefficients corresponding to the bolt loosening dictionary form the bolt loosening coefficient vector.

[0032] The calculation of correction factors using dynamic pressure data specifically includes: calculating the pressure pulsation turbulence intensity based on all pressure signals, where the turbulence intensity is the ratio of the root mean square of the pressure pulsation to the mean pressure; calculating the coherence function between the pressure signal and the flapping strain, extracting coherence function values ​​at the blade flapping natural frequency and the oscillation natural frequency respectively; constructing a pressure correction factor based on the two coherence function values; the leading edge corrosion correction factor is the ratio of the flapping frequency coherence function value to the sum of the two coherence function values; and the bolt loosening correction factor is one minus the leading edge corrosion correction factor.

[0033] The severity index of leading-edge corrosion is obtained by multiplying the leading-edge corrosion correction factor, the L1 norm of the leading-edge corrosion coefficient vector, and the ratio of the pressure pulsation turbulence intensity to the healthy state reference turbulence intensity; the severity index of bolt loosening is obtained by multiplying the bolt loosening correction factor, the L1 norm of the bolt loosening coefficient vector, and the relative offset of the natural frequency of the oscillation; when the two severity indices exceed the corresponding preset thresholds, the corresponding fault is determined to reach the level requiring maintenance.

[0034] The residual verification using angular acceleration signals specifically includes: calculating the fluctuation variance of the acquired angular acceleration signals; reconstructing the angular acceleration signals from the decoupled leading-edge corrosion coefficient vector and bolt loosening coefficient vector using the transfer matrix that maps the fault coefficients of the strain domain to the torque domain; calculating the residual between the measured angular acceleration signals and the reconstructed angular acceleration signals; if the residual is greater than a set multiple of the noise standard deviation, triggering a re-diagnosis process, returning to adjust the phase synchronization parameters or operating condition sensitivity factor until the residual converges.

[0035] The operating parameters of the wind turbine main control system include wind speed, wind turbine speed or generator speed, pitch angle, pitch rate and generator torque. All operating parameters are synchronized with strain signal, pressure signal and angular acceleration signal using a precision time protocol to ensure that the time alignment error between different physical quantity signals is less than the preset value.

[0036] Variational mode decomposition decomposes the phase-synchronized strain signal into several intrinsic mode functions and extracts the instantaneous amplitude and instantaneous frequency of each intrinsic mode function. The same variational mode decomposition process is applied to the pressure signal and angular acceleration signal. The decomposed modal components are used for the subsequent construction of an overcomplete dictionary and the extraction of fault features.

[0037] The specific application of this embodiment is as follows:

[0038] Step S1: Synchronous acquisition and preprocessing of multi-source heterogeneous sensor data

[0039] Step S11: Arrange N strain measurement points at equal intervals from the blade root to the blade tip of the wind turbine blade under test. Install a triaxial resistance strain gauge at each measurement point to collect the dynamic strain response signals in the blade flapping and oscillation directions. Denote the strain vector of the i-th measurement point at time t as... ,in, In order to respond to changes, For oscillation strain, The sampling frequency is not less than 10 times the first-order bending natural frequency of the blade, and is set to 2kHz in this step;

[0040] Step S12: Embed M miniature dynamic pressure sensors on the easily corroded surface of the blade leading edge. The sensor probe diameter is no greater than 3mm. Each sensor is connected to a multi-channel dynamic pressure acquisition system via an independent shielded cable. This system has 32 channels, a maximum sampling frequency of 62.5kHz, a sampling accuracy of 12 bits, and is compatible with 4-20mA current signals and ±10VDC voltage signals. Let the instantaneous pressure value measured by the j-th pressure sensor at time t be... , This pressure signal directly reflects the pulsating changes in aerodynamic load on the blade surface, providing a basis for subsequent separation of aerodynamic excitation and structural response.

[0041] Step S13: Install an incremental encoder at the output shaft end of the wind turbine generator to measure the shaft angular displacement at a sampling frequency of 100kHz. Instantaneous angular acceleration is calculated by performing second-order central difference on the angular displacement sequence:

[0042]

[0043] In the formula, Sampling time interval; instantaneous angular acceleration It reflects the instantaneous fluctuation of the transmission chain torque, including the torque oscillation component generated by the coupling of blade aerodynamic imbalance and structural stiffness abnormality;

[0044] Step S14: Synchronously collect operating parameters of the wind turbine main control system: wind speed Generator speed Pitch angle Pitch rate Generator torque The sampling frequency is kept consistent with the strain signal (2kHz); all the above sensors and data acquisition systems use the IEEE 1588 precision time protocol for clock synchronization to ensure that the time alignment error between different physical quantity signals is less than 1 microsecond.

[0045] Step S2: Cross-domain phase synchronization and signal mode decomposition

[0046] Step S21: Due to the different physical bandwidths of the strain signal, pressure signal, and angular acceleration signal, direct time-domain correlation analysis will produce phase deviation; a cross-domain phase synchronization technique is adopted: first, bandpass filtering is performed on each signal channel, with the passband range covering the first three flapping / swaying mode frequencies of the blade (usually 0.5Hz~20Hz); then, the instantaneous phase of each signal component is extracted using Hilbert transform:

[0047]

[0048] In the formula, Represents the time-domain signal of any channel. Use the Hilbert transform operator; calculate the phase difference of each signal relative to the reference signal (selected as leaf root flapping strain). When the phase difference changes abruptly, it indicates the presence of a non-stationary excitation source or a change in system parameters. To eliminate non-fault phase jumps caused by variable operating conditions (such as pitch control), a pitch rate-related phase compensation model is established.

[0049]

[0050] In the formula, This is an empirical proportionality coefficient (obtained through offline calibration, typical value 0.05 rad·s / °). The time delay from pitch control to blade aerodynamic response (value ranges from 0.1 to 0.3 seconds); the compensated phase difference is... The compensated signals are resampled and interpolated according to the phase of the reference signal to achieve complete phase synchronization of the cross-domain signals, resulting in a synchronized signal set. ;

[0051] Step S22: Synchronize the strain signal Perform variational mode decomposition (VMD) to decompose it into several eigenmode functions. , ,in The center frequency was determined to be 5 using the center frequency observation method; simultaneously, the instantaneous amplitude of each IMF was extracted. With instantaneous frequency ; for pressure signals Perform the same process to obtain Its amplitude and frequency characteristics; this decomposition process does not require preset basis functions and can adaptively separate aeroelastic response components at different time scales.

[0052] Step S3: Adaptive Sparse Bayesian Learning for Feature Decoupling

[0053] Step S31: Construct the overcomplete dictionary matrix of coupled faults Each column represents a basis function for a single fault source under different operating conditions; the dictionary consists of two parts: the leading edge erosion dictionary. Dictionary of loose bolts Each basis function is not a fixed waveform, but a parameterized waveform dynamically generated based on real-time operating parameters; specifically, the strain response characteristics caused by leading-edge corrosion can be modeled as follows:

[0054]

[0055] In the formula, For amplitude coefficient, The aerodynamic damping ratio (which varies with surface roughness). The natural frequency of the blade's waving. The shape function of wind speed fluctuations (derived from measured wind speeds) (Determined by the autocorrelation function); the strain response characteristics caused by bolt loosening are:

[0056]

[0057] In the formula, For amplitude coefficient, The relaxation time constant after the connection stiffness is reduced. The natural frequency of the oscillation. A pulse sequence triggered by the pitch rate (when (Take 1 if the parameter is set, otherwise take 0); Discretize the above parameterized waveform according to different parameter values ​​to form dictionary atoms;

[0058] Step S32: Synchronize the strain signal With pressure signal and angular acceleration Concatenate them into a joint observation vector:

[0059]

[0060] The dimension is ; Assumption It can be represented as a dictionary matrix. With sparse coefficient vector Linear combination plus noise:

[0061]

[0062] In the formula, This is a sparse coefficient vector, where the positions of the non-zero elements correspond to the activated fault types. Gaussian white noise; traditional sparse Bayesian learning assumption The prior distribution is an independent zero-mean Gaussian distribution, and the variance is controlled by the hyperparameter; however, under varying operating conditions, the activation modes of different faults change with the operating conditions, therefore, an adaptive prior variance for the operating conditions is introduced:

[0063]

[0064] In the formula, For the first Each dictionary atom corresponds to a coefficient The prior variance, As the base value, These are operating condition sensitive factors (obtained through learning from a small amount of offline data). The wind speed change rate; this adaptive mechanism increases the prior variance when the pitch changes rapidly or the wind speed changes abruptly, allowing the coefficient to change significantly to track the transient operating conditions, while avoiding misjudging operating condition disturbances as faults.

[0065] Step S33: Iteratively solve for the sparse coefficients using the Expectation-Maximization (EM) framework. In step E, given the current noise variance and prior variance matrix Calculate the posterior mean and covariance:

[0066]

[0067] In the formula, The total number of atoms in the dictionary; in step M, update the noise variance. and prior base values :

[0068]

[0069]

[0070] Iterate until the logarithmic marginal likelihood function converges; after convergence These are the activation coefficients of each fault basis function, where the coefficients corresponding to the leading-edge erosion dictionary form a vector. The coefficients of the corresponding bolt loosening dictionary are composed of .

[0071] Step S4: Fault Quantification and Diagnostic Decision Based on Pressure Correction

[0072] Step S41: Correct the decoupled fault coefficient using synchronously acquired dynamic pressure data; since leading-edge corrosion changes the blade surface roughness, thus affecting the statistical characteristics of pressure pulsation, define the pressure pulsation turbulence intensity:

[0073]

[0074] In the formula, The mean of the pressure signal within a 2-second sliding window is given; simultaneously, the coherence function between the pressure signal and the swing strain is calculated:

[0075]

[0076] In the formula, For cross power spectral density, and For self-power spectral density, For frequency; at the natural frequency of blade flapping If the coherence function value is greater than 0.6, it indicates a strong correlation between aerodynamic excitation and structural response. In this case, the effect of leading-edge corrosion is more significant, and the coherence function should be increased. The confidence weight; conversely, if at the oscillation frequency A low coherence function value indicates that structural nonlinearity caused by bolt loosening is dominant; based on this, a pressure correction factor is constructed:

[0077]

[0078] The revised fault severity index is as follows:

[0079]

[0080] In the formula, It is an L1 norm. This is the reference turbulence intensity under healthy blade conditions (given by factory calibration). This is the offset of the natural frequency of the oscillation. The inherent frequency in a healthy state; when Exceeding the threshold When the value is 0.35, the leading edge corrosion is determined to have reached the level requiring maintenance; when Exceeding the threshold (A value of 0.40) indicates a significant risk of bolt loosening;

[0081] Step S42: Instantaneous angular acceleration As an independent verification signal; calculate the fluctuation variance of the angular acceleration signal. And compare it with the torque fluctuation reconstructed from the decoupled fault coefficient:

[0082]

[0083] In the formula, To map the failure coefficients in the strain domain to the transfer matrix in the torque domain (pre-calculated using the wind turbine drivetrain dynamics model); if the residuals If the error exceeds 3 times the noise standard deviation, a re-diagnosis process is triggered, returning to step S2 to adjust the phase synchronization parameters or step S3 to adjust the operating condition sensitivity factor until the residual converges.

[0084] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0085] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for accurately diagnosing coupling faults of fan blades under variable operating conditions, characterized in that, The diagnostic method includes the following steps: Strain measurement points and dynamic pressure sensors are arranged on the wind turbine blades, and an angular acceleration measurement device is installed at the generator output shaft end. Operating parameters of the wind turbine main control system are collected synchronously. Cross-domain phase synchronization is performed on the collected multi-source heterogeneous signals to compensate for phase jumps caused by varying operating conditions, obtaining phase-aligned strain, pressure, and angular acceleration signals. Variational mode decomposition is used to extract the modal components of the signals. An overcomplete dictionary containing a leading-edge corrosion dictionary and a bolt loosening dictionary is constructed. An operating condition-adaptive sparse Bayesian learning method is used to solve for the sparse coefficient vector, achieving decoupling of the two fault characteristics. Correction factors are calculated using dynamic pressure data to quantify and correct the decoupled fault coefficients, obtaining leading-edge corrosion severity indicators and bolt loosening severity indicators. Residual verification is performed using angular acceleration signals to ensure the reliability of the diagnostic results.

2. The method according to claim 1, characterized in that, The strain measurement points are arranged according to the vibration mode or mode along the blade root to blade tip direction. Each measurement point is equipped with a triaxial resistance strain gauge, which simultaneously collects the strain in the blade flapping direction and the strain in the oscillation direction. The dynamic pressure sensor is embedded in the easily corroded surface of the blade leading edge. Each sensor is connected to a multi-channel dynamic pressure acquisition system via an independent shielded cable. The number of channels, sampling frequency and sampling accuracy of the system are configured according to actual test requirements, and it is compatible with both current and voltage signals.

3. The method for accurate diagnosis of wind turbine blade coupling faults under variable operating conditions according to claim 1, characterized in that, The cross-domain phase synchronization specifically includes: performing bandpass filtering on each signal channel, with the passband range covering the first three flapping and twitching mode frequencies of the blade; extracting the instantaneous phase of each signal using Hilbert transform; calculating the phase difference of other signals relative to the reference signal using blade root flapping strain as the reference signal; establishing a phase compensation model based on the pitch rate to compensate for non-fault phase jumps caused by pitching action; and using the compensated phase difference to resample and interpolate the signals to achieve complete phase synchronization of cross-domain signals.

4. The method for accurate diagnosis of wind turbine blade coupling faults under variable operating conditions according to claim 1, characterized in that, In the aforementioned condition-adaptive sparse Bayesian learning method, synchronized flapping strain, oscillation strain, pressure signal, and angular acceleration signal are concatenated into a joint observation vector. Each column of the overcomplete dictionary represents a parameterized basis function for a single fault source under different operating conditions. The basis functions of the leading-edge corrosion dictionary include aerodynamic damping ratio and wind speed fluctuation shape function, while the basis functions of the bolt loosening dictionary include relaxation time constant and pulse sequence triggered by pitch rate. A condition-adaptive prior variance is introduced, which is dynamically adjusted according to pitch rate and wind speed change rate, allowing for significant changes in the sparsity coefficients during transients under varying operating conditions, thus avoiding misjudging operating condition disturbances as faults.

5. The method for accurate diagnosis of wind turbine blade coupling faults under variable operating conditions according to claim 4, characterized in that, The condition-adaptive sparse Bayesian learning employs an expectation-maximization framework to iteratively solve for sparse coefficients. In the expectation step, the posterior mean and covariance are calculated based on the current noise variance and prior variance matrix. In the maximization step, the noise variance and prior variance base values ​​are updated. The iteration continues until the logarithmic marginal likelihood function converges. The converged posterior mean is the activation coefficient of each fault basis function, where the coefficients corresponding to the leading edge corrosion dictionary constitute the leading edge corrosion coefficient vector, and the coefficients corresponding to the bolt loosening dictionary constitute the bolt loosening coefficient vector.

6. The method for accurate diagnosis of wind turbine blade coupling faults under variable operating conditions according to claim 1, characterized in that, The calculation of correction factors using dynamic pressure data specifically includes: calculating the pressure pulsation turbulence intensity based on all pressure signals, where the turbulence intensity is the ratio of the root mean square of the pressure pulsation to the mean pressure; calculating the coherence function between the pressure signal and the flapping strain, extracting coherence function values ​​at the blade flapping natural frequency and the oscillation natural frequency respectively; constructing a pressure correction factor based on the two coherence function values; the leading edge corrosion correction factor is the ratio of the flapping frequency coherence function value to the sum of the two coherence function values; and the bolt loosening correction factor is one minus the leading edge corrosion correction factor.

7. The method for accurate diagnosis of wind turbine blade coupling faults under variable operating conditions according to claim 6, characterized in that, The severity index of leading-edge corrosion is obtained by multiplying the leading-edge corrosion correction factor, the L1 norm of the leading-edge corrosion coefficient vector, and the ratio of the pressure pulsation turbulence intensity to the healthy state reference turbulence intensity; the severity index of bolt loosening is obtained by multiplying the bolt loosening correction factor, the L1 norm of the bolt loosening coefficient vector, and the relative offset of the natural frequency of the oscillation; when the two severity indices exceed the corresponding preset thresholds, the corresponding fault is determined to reach the level requiring maintenance.

8. The method for accurate diagnosis of wind turbine blade coupling faults under variable operating conditions according to claim 1, characterized in that, The residual verification using angular acceleration signals specifically includes: calculating the fluctuation variance of the acquired angular acceleration signals; reconstructing the angular acceleration signals from the decoupled leading-edge corrosion coefficient vector and bolt loosening coefficient vector using the transfer matrix that maps the fault coefficients of the strain domain to the torque domain; calculating the residual between the measured angular acceleration signals and the reconstructed angular acceleration signals; if the residual is greater than a set multiple of the noise standard deviation, triggering a re-diagnosis process, returning to adjust the phase synchronization parameters or operating condition sensitivity factor until the residual converges.

9. The method for accurate diagnosis of wind turbine blade coupling faults under variable operating conditions according to claim 1, characterized in that, The operating parameters of the wind turbine main control system include wind speed, wind turbine speed or generator speed, pitch angle, pitch rate and generator torque. All operating parameters are synchronized with strain signal, pressure signal and angular acceleration signal using a precision time protocol to ensure that the time alignment error between different physical quantity signals is less than the preset value.

10. The method for accurate diagnosis of wind turbine blade coupling faults under variable operating conditions according to claim 1, characterized in that, The variational mode decomposition decomposes the phase-synchronized strain signal into several intrinsic mode functions and extracts the instantaneous amplitude and instantaneous frequency of each intrinsic mode function; the pressure signal and angular acceleration signal are subjected to the same variational mode decomposition process, and the decomposed modal components are used for subsequent construction of an overcomplete dictionary and extraction of fault features.

Images